diff --git a/share/man/man4/tcp.4 b/share/man/man4/tcp.4 index cbb8021226fe..873cfe4b822a 100644 --- a/share/man/man4/tcp.4 +++ b/share/man/man4/tcp.4 @@ -1,850 +1,863 @@ .\" Copyright (c) 1983, 1991, 1993 .\" The Regents of the University of California. .\" Copyright (c) 2010-2011 The FreeBSD Foundation .\" All rights reserved. .\" .\" Portions of this documentation were written at the Centre for Advanced .\" Internet Architectures, Swinburne University of Technology, Melbourne, .\" Australia by David Hayes under sponsorship from the FreeBSD Foundation. .\" .\" Redistribution and use in source and binary forms, with or without .\" modification, are permitted provided that the following conditions .\" are met: .\" 1. Redistributions of source code must retain the above copyright .\" notice, this list of conditions and the following disclaimer. .\" 2. Redistributions in binary form must reproduce the above copyright .\" notice, this list of conditions and the following disclaimer in the .\" documentation and/or other materials provided with the distribution. .\" 3. Neither the name of the University nor the names of its contributors .\" may be used to endorse or promote products derived from this software .\" without specific prior written permission. .\" .\" THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND .\" ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE .\" IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE .\" ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE .\" FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL .\" DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS .\" OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) .\" HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT .\" LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY .\" OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF .\" SUCH DAMAGE. .\" .\" From: @(#)tcp.4 8.1 (Berkeley) 6/5/93 .\" $FreeBSD$ .\" -.Dd April 17, 2021 +.Dd April 18, 2021 .Dt TCP 4 .Os .Sh NAME .Nm tcp .Nd Internet Transmission Control Protocol .Sh SYNOPSIS .In sys/types.h .In sys/socket.h .In netinet/in.h .In netinet/tcp.h .Ft int .Fn socket AF_INET SOCK_STREAM 0 .Sh DESCRIPTION The .Tn TCP protocol provides reliable, flow-controlled, two-way transmission of data. It is a byte-stream protocol used to support the .Dv SOCK_STREAM abstraction. .Tn TCP uses the standard Internet address format and, in addition, provides a per-host collection of .Dq "port addresses" . Thus, each address is composed of an Internet address specifying the host and network, with a specific .Tn TCP port on the host identifying the peer entity. .Pp Sockets utilizing the .Tn TCP protocol are either .Dq active or .Dq passive . Active sockets initiate connections to passive sockets. By default, .Tn TCP sockets are created active; to create a passive socket, the .Xr listen 2 system call must be used after binding the socket with the .Xr bind 2 system call. Only passive sockets may use the .Xr accept 2 call to accept incoming connections. Only active sockets may use the .Xr connect 2 call to initiate connections. .Pp Passive sockets may .Dq underspecify their location to match incoming connection requests from multiple networks. This technique, termed .Dq "wildcard addressing" , allows a single server to provide service to clients on multiple networks. To create a socket which listens on all networks, the Internet address .Dv INADDR_ANY must be bound. The .Tn TCP port may still be specified at this time; if the port is not specified, the system will assign one. Once a connection has been established, the socket's address is fixed by the peer entity's location. The address assigned to the socket is the address associated with the network interface through which packets are being transmitted and received. Normally, this address corresponds to the peer entity's network. .Pp .Tn TCP supports a number of socket options which can be set with .Xr setsockopt 2 and tested with .Xr getsockopt 2 : .Bl -tag -width ".Dv TCP_FUNCTION_BLK" .It Dv TCP_INFO Information about a socket's underlying TCP session may be retrieved by passing the read-only option .Dv TCP_INFO to .Xr getsockopt 2 . It accepts a single argument: a pointer to an instance of .Vt "struct tcp_info" . .Pp This API is subject to change; consult the source to determine which fields are currently filled out by this option. .Fx specific additions include send window size, receive window size, and bandwidth-controlled window space. .It Dv TCP_CCALGOOPT Set or query congestion control algorithm specific parameters. See .Xr mod_cc 4 for details. .It Dv TCP_CONGESTION Select or query the congestion control algorithm that TCP will use for the connection. See .Xr mod_cc 4 for details. .It Dv TCP_FUNCTION_BLK Select or query the set of functions that TCP will use for this connection. This allows a user to select an alternate TCP stack. The alternate TCP stack must already be loaded in the kernel. To list the available TCP stacks, see .Va functions_available in the .Sx MIB Variables section further down. To list the default TCP stack, see .Va functions_default in the .Sx MIB Variables section. .It Dv TCP_KEEPINIT This .Xr setsockopt 2 option accepts a per-socket timeout argument of .Vt "u_int" in seconds, for new, non-established .Tn TCP connections. For the global default in milliseconds see .Va keepinit in the .Sx MIB Variables section further down. .It Dv TCP_KEEPIDLE This .Xr setsockopt 2 option accepts an argument of .Vt "u_int" for the amount of time, in seconds, that the connection must be idle before keepalive probes (if enabled) are sent for the connection of this socket. If set on a listening socket, the value is inherited by the newly created socket upon .Xr accept 2 . For the global default in milliseconds see .Va keepidle in the .Sx MIB Variables section further down. .It Dv TCP_KEEPINTVL This .Xr setsockopt 2 option accepts an argument of .Vt "u_int" to set the per-socket interval, in seconds, between keepalive probes sent to a peer. If set on a listening socket, the value is inherited by the newly created socket upon .Xr accept 2 . For the global default in milliseconds see .Va keepintvl in the .Sx MIB Variables section further down. .It Dv TCP_KEEPCNT This .Xr setsockopt 2 option accepts an argument of .Vt "u_int" and allows a per-socket tuning of the number of probes sent, with no response, before the connection will be dropped. If set on a listening socket, the value is inherited by the newly created socket upon .Xr accept 2 . For the global default see the .Va keepcnt in the .Sx MIB Variables section further down. .It Dv TCP_NODELAY Under most circumstances, .Tn TCP sends data when it is presented; when outstanding data has not yet been acknowledged, it gathers small amounts of output to be sent in a single packet once an acknowledgement is received. For a small number of clients, such as window systems that send a stream of mouse events which receive no replies, this packetization may cause significant delays. The boolean option .Dv TCP_NODELAY defeats this algorithm. .It Dv TCP_MAXSEG By default, a sender- and .No receiver- Ns Tn TCP will negotiate among themselves to determine the maximum segment size to be used for each connection. The .Dv TCP_MAXSEG option allows the user to determine the result of this negotiation, and to reduce it if desired. .It Dv TCP_NOOPT .Tn TCP usually sends a number of options in each packet, corresponding to various .Tn TCP extensions which are provided in this implementation. The boolean option .Dv TCP_NOOPT is provided to disable .Tn TCP option use on a per-connection basis. .It Dv TCP_NOPUSH By convention, the .No sender- Ns Tn TCP will set the .Dq push bit, and begin transmission immediately (if permitted) at the end of every user call to .Xr write 2 or .Xr writev 2 . When this option is set to a non-zero value, .Tn TCP will delay sending any data at all until either the socket is closed, or the internal send buffer is filled. .It Dv TCP_MD5SIG This option enables the use of MD5 digests (also known as TCP-MD5) on writes to the specified socket. Outgoing traffic is digested; digests on incoming traffic are verified. When this option is enabled on a socket, all inbound and outgoing TCP segments must be signed with MD5 digests. .Pp One common use for this in a .Fx router deployment is to enable based routers to interwork with Cisco equipment at peering points. Support for this feature conforms to RFC 2385. .Pp In order for this option to function correctly, it is necessary for the administrator to add a tcp-md5 key entry to the system's security associations database (SADB) using the .Xr setkey 8 utility. This entry can only be specified on a per-host basis at this time. .Pp If an SADB entry cannot be found for the destination, the system does not send any outgoing segments and drops any inbound segments. .It Dv TCP_STATS Manage collection of connection level statistics using the .Xr stats 3 framework. .Pp Each dropped segment is taken into account in the TCP protocol statistics. .It Dv TCP_TXTLS_ENABLE Enable in-kernel Transport Layer Security (TLS) for data written to this socket. See .Xr ktls 4 for more details. .It Dv TCP_TXTLS_MODE The integer argument can be used to get or set the current TLS transmit mode of a socket. See .Xr ktls 4 for more details. .It Dv TCP_RXTLS_ENABLE Enable in-kernel TLS for data read from this socket. See .Xr ktls 4 for more details. .It Dv TCP_REUSPORT_LB_NUMA Changes NUMA affinity filtering for an established TCP listen socket. This option takes a single integer argument which specifies the NUMA domain to filter on for this listen socket. The argument can also have the follwing special values: .Bl -tag -width "Dv TCP_REUSPORT_LB_NUMA" .It Dv TCP_REUSPORT_LB_NUMA_NODOM Remove NUMA filtering for this listen socket. .It Dv TCP_REUSPORT_LB_NUMA_CURDOM Filter traffic associated with the domain where the calling thread is currently executing. This is typically used after a process or thread inherits a listen socket from its parent, and sets its CPU affinity to a particular core. .El +.It Dv TCP_REMOTE_UDP_ENCAPS_PORT +Set and get the remote UDP encapsulation port. +It can only be set on a closed TCP socket. .El .Pp The option level for the .Xr setsockopt 2 call is the protocol number for .Tn TCP , available from .Xr getprotobyname 3 , or .Dv IPPROTO_TCP . All options are declared in .In netinet/tcp.h . .Pp Options at the .Tn IP transport level may be used with .Tn TCP ; see .Xr ip 4 . Incoming connection requests that are source-routed are noted, and the reverse source route is used in responding. .Pp The default congestion control algorithm for .Tn TCP is .Xr cc_newreno 4 . Other congestion control algorithms can be made available using the .Xr mod_cc 4 framework. .Ss MIB Variables The .Tn TCP protocol implements a number of variables in the .Va net.inet.tcp branch of the .Xr sysctl 3 MIB. .Bl -tag -width ".Va TCPCTL_DO_RFC1323" .It Dv TCPCTL_DO_RFC1323 .Pq Va rfc1323 Implement the window scaling and timestamp options of RFC 1323/RFC 7323 (default is true). .It Va tolerate_missing_ts Tolerate the missing of timestamps (RFC 1323/RFC 7323) for .Tn TCP segments belonging to .Tn TCP connections for which support of .Tn TCP timestamps has been negotiated. (default is 0, i.e., the missing of timestamps is not tolerated). .It Dv TCPCTL_MSSDFLT .Pq Va mssdflt The default value used for the maximum segment size .Pq Dq MSS when no advice to the contrary is received from MSS negotiation. .It Dv TCPCTL_SENDSPACE .Pq Va sendspace Maximum .Tn TCP send window. .It Dv TCPCTL_RECVSPACE .Pq Va recvspace Maximum .Tn TCP receive window. .It Va log_in_vain Log any connection attempts to ports where there is not a socket accepting connections. The value of 1 limits the logging to .Tn SYN (connection establishment) packets only. That of 2 results in any .Tn TCP packets to closed ports being logged. Any value unlisted above disables the logging (default is 0, i.e., the logging is disabled). .It Va msl The Maximum Segment Lifetime, in milliseconds, for a packet. .It Va keepinit Timeout, in milliseconds, for new, non-established .Tn TCP connections. The default is 75000 msec. .It Va keepidle Amount of time, in milliseconds, that the connection must be idle before keepalive probes (if enabled) are sent. The default is 7200000 msec (2 hours). .It Va keepintvl The interval, in milliseconds, between keepalive probes sent to remote machines, when no response is received on a .Va keepidle probe. The default is 75000 msec. .It Va keepcnt Number of probes sent, with no response, before a connection is dropped. The default is 8 packets. .It Va always_keepalive Assume that .Dv SO_KEEPALIVE is set on all .Tn TCP connections, the kernel will periodically send a packet to the remote host to verify the connection is still up. .It Va icmp_may_rst Certain .Tn ICMP unreachable messages may abort connections in .Tn SYN-SENT state. .It Va do_tcpdrain Flush packets in the .Tn TCP reassembly queue if the system is low on mbufs. .It Va blackhole If enabled, disable sending of RST when a connection is attempted to a port where there is not a socket accepting connections. See .Xr blackhole 4 . .It Va delayed_ack Delay ACK to try and piggyback it onto a data packet. .It Va delacktime Maximum amount of time, in milliseconds, before a delayed ACK is sent. .It Va path_mtu_discovery Enable Path MTU Discovery. .It Va tcbhashsize Size of the .Tn TCP control-block hash table (read-only). This may be tuned using the kernel option .Dv TCBHASHSIZE or by setting .Va net.inet.tcp.tcbhashsize in the .Xr loader 8 . .It Va pcbcount Number of active process control blocks (read-only). .It Va syncookies Determines whether or not .Tn SYN cookies should be generated for outbound .Tn SYN-ACK packets. .Tn SYN cookies are a great help during .Tn SYN flood attacks, and are enabled by default. (See .Xr syncookies 4 . ) .It Va isn_reseed_interval The interval (in seconds) specifying how often the secret data used in RFC 1948 initial sequence number calculations should be reseeded. By default, this variable is set to zero, indicating that no reseeding will occur. Reseeding should not be necessary, and will break .Dv TIME_WAIT recycling for a few minutes. .It Va reass.cursegments The current total number of segments present in all reassembly queues. .It Va reass.maxsegments The maximum limit on the total number of segments across all reassembly queues. The limit can be adjusted as a tunable. .It Va reass.maxqueuelen The maximum number of segments allowed in each reassembly queue. By default, the system chooses a limit based on each TCP connection's receive buffer size and maximum segment size (MSS). The actual limit applied to a session's reassembly queue will be the lower of the system-calculated automatic limit and the user-specified .Va reass.maxqueuelen limit. .It Va rexmit_initial , rexmit_min , rexmit_slop Adjust the retransmit timer calculation for .Tn TCP . The slop is typically added to the raw calculation to take into account occasional variances that the .Tn SRTT (smoothed round-trip time) is unable to accommodate, while the minimum specifies an absolute minimum. While a number of .Tn TCP RFCs suggest a 1 second minimum, these RFCs tend to focus on streaming behavior, and fail to deal with the fact that a 1 second minimum has severe detrimental effects over lossy interactive connections, such as a 802.11b wireless link, and over very fast but lossy connections for those cases not covered by the fast retransmit code. For this reason, we use 200ms of slop and a near-0 minimum, which gives us an effective minimum of 200ms (similar to .Tn Linux ) . The initial value is used before an RTT measurement has been performed. .It Va initcwnd_segments Enable the ability to specify initial congestion window in number of segments. The default value is 10 as suggested by RFC 6928. Changing the value on fly would not affect connections using congestion window from the hostcache. Caution: This regulates the burst of packets allowed to be sent in the first RTT. The value should be relative to the link capacity. Start with small values for lower-capacity links. Large bursts can cause buffer overruns and packet drops if routers have small buffers or the link is experiencing congestion. .It Va newcwd Enable the New Congestion Window Validation mechanism as described in RFC 7661. This gently reduces the congestion window during periods, where TCP is application limited and the network bandwidth is not utilized completely. That prevents self-inflicted packet losses once the application starts to transmit data at a higher speed. .It Va do_prr Perform SACK loss recovery using the Proportional Rate Reduction (PRR) algorithm described in RFC6937. This improves the effectiveness of retransmissions particular in environments with ACK thinning or burst loss events, as chances to run out of the ACK clock are reduced, preventing lengthy and performance reducing RTO based loss recovery (default is true). .It Va do_prr_conservative While doing Proportional Rate Reduction, remain strictly in a packet conserving mode, sending only one new packet for each ACK received. Helpful when a misconfigured token bucket traffic policer causes persistent high losses leading to RTO, but reduces PRR effectiveness in more common settings (default is false). .It Va rfc6675_pipe Deprecated and superseded by .Va sack.revised .It Va rfc3042 Enable the Limited Transmit algorithm as described in RFC 3042. It helps avoid timeouts on lossy links and also when the congestion window is small, as happens on short transfers. .It Va rfc3390 Enable support for RFC 3390, which allows for a variable-sized starting congestion window on new connections, depending on the maximum segment size. This helps throughput in general, but particularly affects short transfers and high-bandwidth large propagation-delay connections. .It Va sack.enable Enable support for RFC 2018, TCP Selective Acknowledgment option, which allows the receiver to inform the sender about all successfully arrived segments, allowing the sender to retransmit the missing segments only. .It Va sack.revised Enables three updated mechanisms from RFC6675 (default is true). Calculate the bytes in flight using the algorithm described in RFC 6675, and is also an improvement when Proportional Rate Reduction is enabled. Next, Rescue Retransmission helps timely loss recovery, when the trailing segments of a transmission are lost, while no additional data is ready to be sent. In case a partial ACK without a SACK block is received during SACK loss recovery, the trailing segment is immediately resent, rather than waiting for a Retransmission timeout. Finally, SACK loss recovery is also engaged, once two segments plus one byte are SACKed - even if no traditional duplicate ACKs were observed. .It Va sack.maxholes Maximum number of SACK holes per connection. Defaults to 128. .It Va sack.globalmaxholes Maximum number of SACK holes per system, across all connections. Defaults to 65536. .It Va maxtcptw When a TCP connection enters the .Dv TIME_WAIT state, its associated socket structure is freed, since it is of negligible size and use, and a new structure is allocated to contain a minimal amount of information necessary for sustaining a connection in this state, called the compressed TCP TIME_WAIT state. Since this structure is smaller than a socket structure, it can save a significant amount of system memory. The .Va net.inet.tcp.maxtcptw MIB variable controls the maximum number of these structures allocated. By default, it is initialized to .Va kern.ipc.maxsockets / 5. .It Va nolocaltimewait Suppress creating of compressed TCP TIME_WAIT states for connections in which both endpoints are local. .It Va fast_finwait2_recycle Recycle .Tn TCP .Dv FIN_WAIT_2 connections faster when the socket is marked as .Dv SBS_CANTRCVMORE (no user process has the socket open, data received on the socket cannot be read). The timeout used here is .Va finwait2_timeout . .It Va finwait2_timeout Timeout to use for fast recycling of .Tn TCP .Dv FIN_WAIT_2 connections. Defaults to 60 seconds. .It Va ecn.enable Enable support for TCP Explicit Congestion Notification (ECN). ECN allows a TCP sender to reduce the transmission rate in order to avoid packet drops. .Bl -tag -compact .It 0 Disable ECN. .It 1 Allow incoming connections to request ECN. Outgoing connections will request ECN. .It 2 Allow incoming connections to request ECN. Outgoing connections will not request ECN. (default) .El .It Va ecn.maxretries Number of retries (SYN or SYN/ACK retransmits) before disabling ECN on a specific connection. This is needed to help with connection establishment when a broken firewall is in the network path. .It Va pmtud_blackhole_detection Enable automatic path MTU blackhole detection. In case of retransmits of MSS sized segments, the OS will lower the MSS to check if it's an MTU problem. If the current MSS is greater than the configured value to try .Po Va net.inet.tcp.pmtud_blackhole_mss and .Va net.inet.tcp.v6pmtud_blackhole_mss .Pc , it will be set to this value, otherwise, the MSS will be set to the default values .Po Va net.inet.tcp.mssdflt and .Va net.inet.tcp.v6mssdflt .Pc . Settings: .Bl -tag -compact .It 0 Disable path MTU blackhole detection. .It 1 Enable path MTU blackhole detection for IPv4 and IPv6. .It 2 Enable path MTU blackhole detection only for IPv4. .It 3 Enable path MTU blackhole detection only for IPv6. .El .It Va pmtud_blackhole_mss MSS to try for IPv4 if PMTU blackhole detection is turned on. .It Va v6pmtud_blackhole_mss MSS to try for IPv6 if PMTU blackhole detection is turned on. .It Va hostcache.enable The TCP host cache is used to cache connection details and metrics to improve future performance of connections between the same hosts. At the completion of a TCP connection, a host will cache information for the connection for some defined period of time. .Bl -tag -compact .It 0 Disable the host cache. .It 1 Enable the host cache. (default) .It Va hostcache.purgenow Immediately purge all entries once set to any value. Setting this to 2 will also reseed the hash salt. .It Va hostcache.purge Expire all entires on next pruning of host cache entries. Any non-zero setting will be reset to zero, once the pruge is running. .Bl -tag -compact .It 0 Do not purge all entries when pruning the host cache. (default) .It 1 Purge all entries when doing the next pruning. .It 2 Purge all entries, and also reseed the hash salt. .It Va hostcache.prune Time in seconds between pruning expired host cache entries. Defaults to 300 (5 minutes). .It Va hostcache.expire Time in seconds, how long a entry should be kept in the host cache since last accessed. Defaults to 3600 (1 hour). .It Va hostcache.count The current number of entries in the host cache. .It Va hostcache.bucketlimit The maximum number of entries for the same hash. Defaults to 30. .It Va hostcache.hashsize Size of TCP hostcache hashtable. This number has to be a power of two, or will be rejected. Defaults to 512. .It Va hostcache.cachelimit Overall entry limit for hostcache. Defaults to hashsize * bucketlimit. .It Va hostcache.histo Provide a Histogram of the hostcache hash utilization. .It Va hostcache.list Provide a complete list of all current entries in the host cache. .It Va functions_available List of available TCP function blocks (TCP stacks). .It Va functions_default The default TCP function block (TCP stack). .It Va functions_inherit_listen_socket_stack Determines whether to inherit listen socket's tcp stack or use the current system default tcp stack, as defined by .Va functions_default . Default is true. .It Va insecure_rst Use criteria defined in RFC793 instead of RFC5961 for accepting RST segments. Default is false. .It Va insecure_syn Use criteria defined in RFC793 instead of RFC5961 for accepting SYN segments. Default is false. .It Va ts_offset_per_conn When initializing the TCP timestamps, use a per connection offset instead of a per host pair offset. Default is to use per connection offsets as recommended in RFC 7323. .It Va perconn_stats_enable Controls the default collection of statistics for all connections using the .Xr stats 3 framework. 0 disables, 1 enables, 2 enables random sampling across log id connection groups with all connections in a group receiving the same setting. .It Va perconn_stats_sample_rates A CSV list of template_spec=percent key-value pairs which controls the per template sampling rates when .Xr stats 3 sampling is enabled. +.It Va udp_tunneling_port +The local UDP encapsulation port. +A value of 0 indicates that UDP encapsulation is disabled. +The default is 0. +.It Va udp_tunneling_overhead +The overhead taken into account when using UDP encapsulation. +Since MSS clamping by middleboxes will most likely not work, values larger than +8 (the size of the UDP header) are also supported. +Supported values are between 8 and 1024. +The default is 8. .El .Sh ERRORS A socket operation may fail with one of the following errors returned: .Bl -tag -width Er .It Bq Er EISCONN when trying to establish a connection on a socket which already has one; .It Bo Er ENOBUFS Bc or Bo Er ENOMEM Bc when the system runs out of memory for an internal data structure; .It Bq Er ETIMEDOUT when a connection was dropped due to excessive retransmissions; .It Bq Er ECONNRESET when the remote peer forces the connection to be closed; .It Bq Er ECONNREFUSED when the remote peer actively refuses connection establishment (usually because no process is listening to the port); .It Bq Er EADDRINUSE when an attempt is made to create a socket with a port which has already been allocated; .It Bq Er EADDRNOTAVAIL when an attempt is made to create a socket with a network address for which no network interface exists; .It Bq Er EAFNOSUPPORT when an attempt is made to bind or connect a socket to a multicast address. .It Bq Er EINVAL when trying to change TCP function blocks at an invalid point in the session; .It Bq Er ENOENT when trying to use a TCP function block that is not available; .El .Sh SEE ALSO .Xr getsockopt 2 , .Xr socket 2 , .Xr stats 3 , .Xr sysctl 3 , .Xr blackhole 4 , .Xr inet 4 , .Xr intro 4 , .Xr ip 4 , .Xr ktls 4 , .Xr mod_cc 4 , .Xr siftr 4 , .Xr syncache 4 , .Xr tcp_bbr 4 , .Xr setkey 8 , .Xr tcp_functions 9 .Rs .%A "V. Jacobson" .%A "B. Braden" .%A "D. Borman" .%T "TCP Extensions for High Performance" .%O "RFC 1323" .Re .Rs .%A "D. Borman" .%A "B. Braden" .%A "V. Jacobson" .%A "R. Scheffenegger" .%T "TCP Extensions for High Performance" .%O "RFC 7323" .Re .Rs .%A "A. Heffernan" .%T "Protection of BGP Sessions via the TCP MD5 Signature Option" .%O "RFC 2385" .Re .Rs .%A "K. Ramakrishnan" .%A "S. Floyd" .%A "D. Black" .%T "The Addition of Explicit Congestion Notification (ECN) to IP" .%O "RFC 3168" .Re .Sh HISTORY The .Tn TCP protocol appeared in .Bx 4.2 . The RFC 1323 extensions for window scaling and timestamps were added in .Bx 4.4 . The .Dv TCP_INFO option was introduced in .Tn Linux 2.6 and is .Em subject to change . diff --git a/sys/netinet/tcp.h b/sys/netinet/tcp.h index 0b71bd4658f8..d2bf1f8431fd 100644 --- a/sys/netinet/tcp.h +++ b/sys/netinet/tcp.h @@ -1,413 +1,414 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1993 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)tcp.h 8.1 (Berkeley) 6/10/93 * $FreeBSD$ */ #ifndef _NETINET_TCP_H_ #define _NETINET_TCP_H_ #include #include #if __BSD_VISIBLE typedef u_int32_t tcp_seq; #define tcp6_seq tcp_seq /* for KAME src sync over BSD*'s */ #define tcp6hdr tcphdr /* for KAME src sync over BSD*'s */ /* * TCP header. * Per RFC 793, September, 1981. */ struct tcphdr { u_short th_sport; /* source port */ u_short th_dport; /* destination port */ tcp_seq th_seq; /* sequence number */ tcp_seq th_ack; /* acknowledgement number */ #if BYTE_ORDER == LITTLE_ENDIAN u_char th_x2:4, /* (unused) */ th_off:4; /* data offset */ #endif #if BYTE_ORDER == BIG_ENDIAN u_char th_off:4, /* data offset */ th_x2:4; /* (unused) */ #endif u_char th_flags; #define TH_FIN 0x01 #define TH_SYN 0x02 #define TH_RST 0x04 #define TH_PUSH 0x08 #define TH_ACK 0x10 #define TH_URG 0x20 #define TH_ECE 0x40 #define TH_CWR 0x80 #define TH_AE 0x100 /* maps into th_x2 */ #define TH_FLAGS (TH_FIN|TH_SYN|TH_RST|TH_PUSH|TH_ACK|TH_URG|TH_ECE|TH_CWR) #define PRINT_TH_FLAGS "\20\1FIN\2SYN\3RST\4PUSH\5ACK\6URG\7ECE\10CWR\11AE" u_short th_win; /* window */ u_short th_sum; /* checksum */ u_short th_urp; /* urgent pointer */ }; #define PADTCPOLEN(len) ((((len) / 4) + !!((len) % 4)) * 4) #define TCPOPT_EOL 0 #define TCPOLEN_EOL 1 #define TCPOPT_PAD 0 /* padding after EOL */ #define TCPOLEN_PAD 1 #define TCPOPT_NOP 1 #define TCPOLEN_NOP 1 #define TCPOPT_MAXSEG 2 #define TCPOLEN_MAXSEG 4 #define TCPOPT_WINDOW 3 #define TCPOLEN_WINDOW 3 #define TCPOPT_SACK_PERMITTED 4 #define TCPOLEN_SACK_PERMITTED 2 #define TCPOPT_SACK 5 #define TCPOLEN_SACKHDR 2 #define TCPOLEN_SACK 8 /* 2*sizeof(tcp_seq) */ #define TCPOPT_TIMESTAMP 8 #define TCPOLEN_TIMESTAMP 10 #define TCPOLEN_TSTAMP_APPA (TCPOLEN_TIMESTAMP+2) /* appendix A */ #define TCPOPT_SIGNATURE 19 /* Keyed MD5: RFC 2385 */ #define TCPOLEN_SIGNATURE 18 #define TCPOPT_FAST_OPEN 34 #define TCPOLEN_FAST_OPEN_EMPTY 2 /* Miscellaneous constants */ #define MAX_SACK_BLKS 6 /* Max # SACK blocks stored at receiver side */ #define TCP_MAX_SACK 4 /* MAX # SACKs sent in any segment */ /* * The default maximum segment size (MSS) to be used for new TCP connections * when path MTU discovery is not enabled. * * RFC879 derives the default MSS from the largest datagram size hosts are * minimally required to handle directly or through IP reassembly minus the * size of the IP and TCP header. With IPv6 the minimum MTU is specified * in RFC2460. * * For IPv4 the MSS is 576 - sizeof(struct tcpiphdr) * For IPv6 the MSS is IPV6_MMTU - sizeof(struct ip6_hdr) - sizeof(struct tcphdr) * * We use explicit numerical definition here to avoid header pollution. */ #define TCP_MSS 536 #define TCP6_MSS 1220 /* * Limit the lowest MSS we accept for path MTU discovery and the TCP SYN MSS * option. Allowing low values of MSS can consume significant resources and * be used to mount a resource exhaustion attack. * Connections requesting lower MSS values will be rounded up to this value * and the IP_DF flag will be cleared to allow fragmentation along the path. * * See tcp_subr.c tcp_minmss SYSCTL declaration for more comments. Setting * it to "0" disables the minmss check. * * The default value is fine for TCP across the Internet's smallest official * link MTU (256 bytes for AX.25 packet radio). However, a connection is very * unlikely to come across such low MTU interfaces these days (anno domini 2003). */ #define TCP_MINMSS 216 #define TCP_MAXWIN 65535 /* largest value for (unscaled) window */ #define TTCP_CLIENT_SND_WND 4096 /* dflt send window for T/TCP client */ #define TCP_MAX_WINSHIFT 14 /* maximum window shift */ #define TCP_MAXBURST 4 /* maximum segments in a burst */ #define TCP_MAXHLEN (0xf<<2) /* max length of header in bytes */ #define TCP_MAXOLEN (TCP_MAXHLEN - sizeof(struct tcphdr)) /* max space left for options */ #define TCP_FASTOPEN_MIN_COOKIE_LEN 4 /* Per RFC7413 */ #define TCP_FASTOPEN_MAX_COOKIE_LEN 16 /* Per RFC7413 */ #define TCP_FASTOPEN_PSK_LEN 16 /* Same as TCP_FASTOPEN_KEY_LEN */ #endif /* __BSD_VISIBLE */ /* * User-settable options (used with setsockopt). These are discrete * values and are not masked together. Some values appear to be * bitmasks for historical reasons. */ #define TCP_NODELAY 1 /* don't delay send to coalesce packets */ #if __BSD_VISIBLE #define TCP_MAXSEG 2 /* set maximum segment size */ #define TCP_NOPUSH 4 /* don't push last block of write */ #define TCP_NOOPT 8 /* don't use TCP options */ #define TCP_MD5SIG 16 /* use MD5 digests (RFC2385) */ #define TCP_INFO 32 /* retrieve tcp_info structure */ #define TCP_STATS 33 /* retrieve stats blob structure */ #define TCP_LOG 34 /* configure event logging for connection */ #define TCP_LOGBUF 35 /* retrieve event log for connection */ #define TCP_LOGID 36 /* configure log ID to correlate connections */ #define TCP_LOGDUMP 37 /* dump connection log events to device */ #define TCP_LOGDUMPID 38 /* dump events from connections with same ID to device */ #define TCP_TXTLS_ENABLE 39 /* TLS framing and encryption for transmit */ #define TCP_TXTLS_MODE 40 /* Transmit TLS mode */ #define TCP_RXTLS_ENABLE 41 /* TLS framing and encryption for receive */ #define TCP_RXTLS_MODE 42 /* Receive TLS mode */ #define TCP_CONGESTION 64 /* get/set congestion control algorithm */ #define TCP_CCALGOOPT 65 /* get/set cc algorithm specific options */ +#define TCP_REMOTE_UDP_ENCAPS_PORT 71 /* Enable TCP over UDP tunneling via the specified port */ #define TCP_DELACK 72 /* socket option for delayed ack */ #define TCP_FIN_IS_RST 73 /* A fin from the peer is treated has a RST */ #define TCP_LOG_LIMIT 74 /* Limit to number of records in tcp-log */ #define TCP_SHARED_CWND_ALLOWED 75 /* Use of a shared cwnd is allowed */ #define TCP_KEEPINIT 128 /* N, time to establish connection */ #define TCP_KEEPIDLE 256 /* L,N,X start keeplives after this period */ #define TCP_KEEPINTVL 512 /* L,N interval between keepalives */ #define TCP_KEEPCNT 1024 /* L,N number of keepalives before close */ #define TCP_FASTOPEN 1025 /* enable TFO / was created via TFO */ #define TCP_PCAP_OUT 2048 /* number of output packets to keep */ #define TCP_PCAP_IN 4096 /* number of input packets to keep */ #define TCP_FUNCTION_BLK 8192 /* Set the tcp function pointers to the specified stack */ /* Options for Rack and BBR */ #define TCP_REUSPORT_LB_NUMA 1026 /* set listen socket numa domain */ #define TCP_RACK_MBUF_QUEUE 1050 /* Do we allow mbuf queuing if supported */ #define TCP_RACK_PROP 1051 /* RACK proportional rate reduction (bool) */ #define TCP_RACK_TLP_REDUCE 1052 /* RACK TLP cwnd reduction (bool) */ #define TCP_RACK_PACE_REDUCE 1053 /* RACK Pacing reduction factor (divisor) */ #define TCP_RACK_PACE_MAX_SEG 1054 /* Max TSO size we will send */ #define TCP_RACK_PACE_ALWAYS 1055 /* Use the always pace method */ #define TCP_RACK_PROP_RATE 1056 /* The proportional reduction rate */ #define TCP_RACK_PRR_SENDALOT 1057 /* Allow PRR to send more than one seg */ #define TCP_RACK_MIN_TO 1058 /* Minimum time between rack t-o's in ms */ #define TCP_RACK_EARLY_RECOV 1059 /* Should recovery happen early (bool) */ #define TCP_RACK_EARLY_SEG 1060 /* If early recovery max segments */ #define TCP_RACK_REORD_THRESH 1061 /* RACK reorder threshold (shift amount) */ #define TCP_RACK_REORD_FADE 1062 /* Does reordering fade after ms time */ #define TCP_RACK_TLP_THRESH 1063 /* RACK TLP theshold i.e. srtt+(srtt/N) */ #define TCP_RACK_PKT_DELAY 1064 /* RACK added ms i.e. rack-rtt + reord + N */ #define TCP_RACK_TLP_INC_VAR 1065 /* Does TLP include rtt variance in t-o */ #define TCP_BBR_IWINTSO 1067 /* Initial TSO window for BBRs first sends */ #define TCP_BBR_RECFORCE 1068 /* Enter recovery force out a segment disregard pacer no longer valid */ #define TCP_BBR_STARTUP_PG 1069 /* Startup pacing gain */ #define TCP_BBR_DRAIN_PG 1070 /* Drain pacing gain */ #define TCP_BBR_RWND_IS_APP 1071 /* Rwnd limited is considered app limited */ #define TCP_BBR_PROBE_RTT_INT 1072 /* How long in useconds between probe-rtt */ #define TCP_BBR_ONE_RETRAN 1073 /* Is only one segment allowed out during retran */ #define TCP_BBR_STARTUP_LOSS_EXIT 1074 /* Do we exit a loss during startup if not 20% incr */ #define TCP_BBR_USE_LOWGAIN 1075 /* lower the gain in PROBE_BW enable */ #define TCP_BBR_LOWGAIN_THRESH 1076 /* Unused after 2.3 morphs to TSLIMITS >= 2.3 */ #define TCP_BBR_TSLIMITS 1076 /* Do we use experimental Timestamp limiting for our algo */ #define TCP_BBR_LOWGAIN_HALF 1077 /* Unused after 2.3 */ #define TCP_BBR_PACE_OH 1077 /* Reused in 4.2 for pacing overhead setting */ #define TCP_BBR_LOWGAIN_FD 1078 /* Unused after 2.3 */ #define TCP_BBR_HOLD_TARGET 1078 /* For 4.3 on */ #define TCP_BBR_USEDEL_RATE 1079 /* Enable use of delivery rate for loss recovery */ #define TCP_BBR_MIN_RTO 1080 /* Min RTO in milliseconds */ #define TCP_BBR_MAX_RTO 1081 /* Max RTO in milliseconds */ #define TCP_BBR_REC_OVER_HPTS 1082 /* Recovery override htps settings 0/1/3 */ #define TCP_BBR_UNLIMITED 1083 /* Not used before 2.3 and morphs to algorithm >= 2.3 */ #define TCP_BBR_ALGORITHM 1083 /* What measurement algo does BBR use netflix=0, google=1 */ #define TCP_BBR_DRAIN_INC_EXTRA 1084 /* Does the 3/4 drain target include the extra gain */ #define TCP_BBR_STARTUP_EXIT_EPOCH 1085 /* what epoch gets us out of startup */ #define TCP_BBR_PACE_PER_SEC 1086 #define TCP_BBR_PACE_DEL_TAR 1087 #define TCP_BBR_PACE_SEG_MAX 1088 #define TCP_BBR_PACE_SEG_MIN 1089 #define TCP_BBR_PACE_CROSS 1090 #define TCP_RACK_IDLE_REDUCE_HIGH 1092 /* Reduce the highest cwnd seen to IW on idle */ #define TCP_RACK_MIN_PACE 1093 /* Do we enforce rack min pace time */ #define TCP_RACK_MIN_PACE_SEG 1094 /* If so what is the seg threshould */ #define TCP_RACK_GP_INCREASE 1094 /* After 4.1 its the GP increase in older rack */ #define TCP_RACK_TLP_USE 1095 #define TCP_BBR_ACK_COMP_ALG 1096 /* Not used */ #define TCP_BBR_TMR_PACE_OH 1096 /* Recycled in 4.2 */ #define TCP_BBR_EXTRA_GAIN 1097 #define TCP_RACK_DO_DETECTION 1097 /* Recycle of extra gain for rack, attack detection */ #define TCP_BBR_RACK_RTT_USE 1098 /* what RTT should we use 0, 1, or 2? */ #define TCP_BBR_RETRAN_WTSO 1099 #define TCP_DATA_AFTER_CLOSE 1100 #define TCP_BBR_PROBE_RTT_GAIN 1101 #define TCP_BBR_PROBE_RTT_LEN 1102 #define TCP_BBR_SEND_IWND_IN_TSO 1103 /* Do we burst out whole iwin size chunks at start? */ #define TCP_BBR_USE_RACK_RR 1104 /* Do we use the rack rapid recovery for pacing rxt's */ #define TCP_BBR_USE_RACK_CHEAT TCP_BBR_USE_RACK_RR /* Compat. */ #define TCP_BBR_HDWR_PACE 1105 /* Enable/disable hardware pacing */ #define TCP_BBR_UTTER_MAX_TSO 1106 /* Do we enforce an utter max TSO size */ #define TCP_BBR_EXTRA_STATE 1107 /* Special exit-persist catch up */ #define TCP_BBR_FLOOR_MIN_TSO 1108 /* The min tso size */ #define TCP_BBR_MIN_TOPACEOUT 1109 /* Do we suspend pacing until */ #define TCP_BBR_TSTMP_RAISES 1110 /* Can a timestamp measurement raise the b/w */ #define TCP_BBR_POLICER_DETECT 1111 /* Turn on/off google mode policer detection */ #define TCP_BBR_RACK_INIT_RATE 1112 /* Set an initial pacing rate for when we have no b/w in kbits per sec */ #define TCP_RACK_RR_CONF 1113 /* Rack rapid recovery configuration control*/ #define TCP_RACK_CHEAT_NOT_CONF_RATE TCP_RACK_RR_CONF #define TCP_RACK_GP_INCREASE_CA 1114 /* GP increase for Congestion Avoidance */ #define TCP_RACK_GP_INCREASE_SS 1115 /* GP increase for Slow Start */ #define TCP_RACK_GP_INCREASE_REC 1116 /* GP increase for Recovery */ #define TCP_RACK_FORCE_MSEG 1117 /* Override to use the user set max-seg value */ #define TCP_RACK_PACE_RATE_CA 1118 /* Pacing rate for Congestion Avoidance */ #define TCP_RACK_PACE_RATE_SS 1119 /* Pacing rate for Slow Start */ #define TCP_RACK_PACE_RATE_REC 1120 /* Pacing rate for Recovery */ #define TCP_NO_PRR 1122 /* If pacing, don't use prr */ #define TCP_RACK_NONRXT_CFG_RATE 1123 /* In recovery does a non-rxt use the cfg rate */ #define TCP_SHARED_CWND_ENABLE 1124 /* Use a shared cwnd if allowed */ #define TCP_TIMELY_DYN_ADJ 1125 /* Do we attempt dynamic multipler adjustment with timely. */ #define TCP_RACK_NO_PUSH_AT_MAX 1126 /* For timely do not push if we are over max rtt */ #define TCP_RACK_PACE_TO_FILL 1127 /* If we are not in recovery, always pace to fill the cwnd in 1 RTT */ #define TCP_SHARED_CWND_TIME_LIMIT 1128 /* we should limit to low time values the scwnd life */ #define TCP_RACK_PROFILE 1129 /* Select a profile that sets multiple options */ /* Start of reserved space for third-party user-settable options. */ #define TCP_VENDOR SO_VENDOR #define TCP_CA_NAME_MAX 16 /* max congestion control name length */ #define TCPI_OPT_TIMESTAMPS 0x01 #define TCPI_OPT_SACK 0x02 #define TCPI_OPT_WSCALE 0x04 #define TCPI_OPT_ECN 0x08 #define TCPI_OPT_TOE 0x10 /* Maximum length of log ID. */ #define TCP_LOG_ID_LEN 64 /* * The TCP_INFO socket option comes from the Linux 2.6 TCP API, and permits * the caller to query certain information about the state of a TCP * connection. We provide an overlapping set of fields with the Linux * implementation, but since this is a fixed size structure, room has been * left for growth. In order to maximize potential future compatibility with * the Linux API, the same variable names and order have been adopted, and * padding left to make room for omitted fields in case they are added later. * * XXX: This is currently an unstable ABI/API, in that it is expected to * change. */ struct tcp_info { u_int8_t tcpi_state; /* TCP FSM state. */ u_int8_t __tcpi_ca_state; u_int8_t __tcpi_retransmits; u_int8_t __tcpi_probes; u_int8_t __tcpi_backoff; u_int8_t tcpi_options; /* Options enabled on conn. */ u_int8_t tcpi_snd_wscale:4, /* RFC1323 send shift value. */ tcpi_rcv_wscale:4; /* RFC1323 recv shift value. */ u_int32_t tcpi_rto; /* Retransmission timeout (usec). */ u_int32_t __tcpi_ato; u_int32_t tcpi_snd_mss; /* Max segment size for send. */ u_int32_t tcpi_rcv_mss; /* Max segment size for receive. */ u_int32_t __tcpi_unacked; u_int32_t __tcpi_sacked; u_int32_t __tcpi_lost; u_int32_t __tcpi_retrans; u_int32_t __tcpi_fackets; /* Times; measurements in usecs. */ u_int32_t __tcpi_last_data_sent; u_int32_t __tcpi_last_ack_sent; /* Also unimpl. on Linux? */ u_int32_t tcpi_last_data_recv; /* Time since last recv data. */ u_int32_t __tcpi_last_ack_recv; /* Metrics; variable units. */ u_int32_t __tcpi_pmtu; u_int32_t __tcpi_rcv_ssthresh; u_int32_t tcpi_rtt; /* Smoothed RTT in usecs. */ u_int32_t tcpi_rttvar; /* RTT variance in usecs. */ u_int32_t tcpi_snd_ssthresh; /* Slow start threshold. */ u_int32_t tcpi_snd_cwnd; /* Send congestion window. */ u_int32_t __tcpi_advmss; u_int32_t __tcpi_reordering; u_int32_t __tcpi_rcv_rtt; u_int32_t tcpi_rcv_space; /* Advertised recv window. */ /* FreeBSD extensions to tcp_info. */ u_int32_t tcpi_snd_wnd; /* Advertised send window. */ u_int32_t tcpi_snd_bwnd; /* No longer used. */ u_int32_t tcpi_snd_nxt; /* Next egress seqno */ u_int32_t tcpi_rcv_nxt; /* Next ingress seqno */ u_int32_t tcpi_toe_tid; /* HWTID for TOE endpoints */ u_int32_t tcpi_snd_rexmitpack; /* Retransmitted packets */ u_int32_t tcpi_rcv_ooopack; /* Out-of-order packets */ u_int32_t tcpi_snd_zerowin; /* Zero-sized windows sent */ /* Padding to grow without breaking ABI. */ u_int32_t __tcpi_pad[26]; /* Padding. */ }; /* * If this structure is provided when setting the TCP_FASTOPEN socket * option, and the enable member is non-zero, a subsequent connect will use * pre-shared key (PSK) mode using the provided key. */ struct tcp_fastopen { int enable; uint8_t psk[TCP_FASTOPEN_PSK_LEN]; }; #endif #define TCP_FUNCTION_NAME_LEN_MAX 32 struct tcp_function_set { char function_set_name[TCP_FUNCTION_NAME_LEN_MAX]; uint32_t pcbcnt; }; /* TLS modes for TCP_TXTLS_MODE */ #define TCP_TLS_MODE_NONE 0 #define TCP_TLS_MODE_SW 1 #define TCP_TLS_MODE_IFNET 2 #define TCP_TLS_MODE_TOE 3 /* * TCP Control message types */ #define TLS_SET_RECORD_TYPE 1 #define TLS_GET_RECORD 2 /* * TCP specific variables of interest for tp->t_stats stats(9) accounting. */ #define VOI_TCP_TXPB 0 /* Transmit payload bytes */ #define VOI_TCP_RETXPB 1 /* Retransmit payload bytes */ #define VOI_TCP_FRWIN 2 /* Foreign receive window */ #define VOI_TCP_LCWIN 3 /* Local congesiton window */ #define VOI_TCP_RTT 4 /* Round trip time */ #define VOI_TCP_CSIG 5 /* Congestion signal */ #define VOI_TCP_GPUT 6 /* Goodput */ #define VOI_TCP_CALCFRWINDIFF 7 /* Congestion avoidance LCWIN - FRWIN */ #define VOI_TCP_GPUT_ND 8 /* Goodput normalised delta */ #define VOI_TCP_ACKLEN 9 /* Average ACKed bytes per ACK */ #define TCP_REUSPORT_LB_NUMA_NODOM (-2) /* remove numa binding */ #define TCP_REUSPORT_LB_NUMA_CURDOM (-1) /* bind to current domain */ #endif /* !_NETINET_TCP_H_ */ diff --git a/sys/netinet/tcp_input.c b/sys/netinet/tcp_input.c index ed184de4a4bf..8592f3313725 100644 --- a/sys/netinet/tcp_input.c +++ b/sys/netinet/tcp_input.c @@ -1,4049 +1,4074 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995 * The Regents of the University of California. All rights reserved. * Copyright (c) 2007-2008,2010 * Swinburne University of Technology, Melbourne, Australia. * Copyright (c) 2009-2010 Lawrence Stewart * Copyright (c) 2010 The FreeBSD Foundation * Copyright (c) 2010-2011 Juniper Networks, Inc. * All rights reserved. * * Portions of this software were developed at the Centre for Advanced Internet * Architectures, Swinburne University of Technology, by Lawrence Stewart, * James Healy and David Hayes, made possible in part by a grant from the Cisco * University Research Program Fund at Community Foundation Silicon Valley. * * Portions of this software were developed at the Centre for Advanced * Internet Architectures, Swinburne University of Technology, Melbourne, * Australia by David Hayes under sponsorship from the FreeBSD Foundation. * * Portions of this software were developed by Robert N. M. Watson under * contract to Juniper Networks, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)tcp_input.c 8.12 (Berkeley) 5/24/95 */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_tcpdebug.h" #include #include #include #ifdef TCP_HHOOK #include #endif #include #include #include /* for proc0 declaration */ #include #include #include #include #include #include #include #include #include #include #include /* before tcp_seq.h, for tcp_random18() */ #include #include #include #include #include #define TCPSTATES /* for logging */ #include #include #include #include #include #include /* required for icmp_var.h */ #include /* for ICMP_BANDLIM */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef TCPPCAP #include #endif #include #ifdef TCPDEBUG #include #endif /* TCPDEBUG */ #ifdef TCP_OFFLOAD #include #endif +#include #include #include #include const int tcprexmtthresh = 3; VNET_DEFINE(int, tcp_log_in_vain) = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, log_in_vain, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_log_in_vain), 0, "Log all incoming TCP segments to closed ports"); VNET_DEFINE(int, blackhole) = 0; #define V_blackhole VNET(blackhole) SYSCTL_INT(_net_inet_tcp, OID_AUTO, blackhole, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(blackhole), 0, "Do not send RST on segments to closed ports"); VNET_DEFINE(int, tcp_delack_enabled) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, delayed_ack, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_delack_enabled), 0, "Delay ACK to try and piggyback it onto a data packet"); VNET_DEFINE(int, drop_synfin) = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, drop_synfin, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(drop_synfin), 0, "Drop TCP packets with SYN+FIN set"); VNET_DEFINE(int, tcp_do_prr_conservative) = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_prr_conservative, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_prr_conservative), 0, "Do conservative Proportional Rate Reduction"); VNET_DEFINE(int, tcp_do_prr) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_prr, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_prr), 1, "Enable Proportional Rate Reduction per RFC 6937"); VNET_DEFINE(int, tcp_do_newcwv) = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, newcwv, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_newcwv), 0, "Enable New Congestion Window Validation per RFC7661"); VNET_DEFINE(int, tcp_do_rfc3042) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3042, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_rfc3042), 0, "Enable RFC 3042 (Limited Transmit)"); VNET_DEFINE(int, tcp_do_rfc3390) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_rfc3390), 0, "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)"); VNET_DEFINE(int, tcp_initcwnd_segments) = 10; SYSCTL_INT(_net_inet_tcp, OID_AUTO, initcwnd_segments, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_initcwnd_segments), 0, "Slow-start flight size (initial congestion window) in number of segments"); VNET_DEFINE(int, tcp_do_rfc3465) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3465, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_rfc3465), 0, "Enable RFC 3465 (Appropriate Byte Counting)"); VNET_DEFINE(int, tcp_abc_l_var) = 2; SYSCTL_INT(_net_inet_tcp, OID_AUTO, abc_l_var, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_abc_l_var), 2, "Cap the max cwnd increment during slow-start to this number of segments"); static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, ecn, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "TCP ECN"); VNET_DEFINE(int, tcp_do_ecn) = 2; SYSCTL_INT(_net_inet_tcp_ecn, OID_AUTO, enable, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_ecn), 0, "TCP ECN support"); VNET_DEFINE(int, tcp_ecn_maxretries) = 1; SYSCTL_INT(_net_inet_tcp_ecn, OID_AUTO, maxretries, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_ecn_maxretries), 0, "Max retries before giving up on ECN"); VNET_DEFINE(int, tcp_insecure_syn) = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, insecure_syn, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_insecure_syn), 0, "Follow RFC793 instead of RFC5961 criteria for accepting SYN packets"); VNET_DEFINE(int, tcp_insecure_rst) = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, insecure_rst, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_insecure_rst), 0, "Follow RFC793 instead of RFC5961 criteria for accepting RST packets"); VNET_DEFINE(int, tcp_recvspace) = 1024*64; #define V_tcp_recvspace VNET(tcp_recvspace) SYSCTL_INT(_net_inet_tcp, TCPCTL_RECVSPACE, recvspace, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_recvspace), 0, "Initial receive socket buffer size"); VNET_DEFINE(int, tcp_do_autorcvbuf) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_auto, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_autorcvbuf), 0, "Enable automatic receive buffer sizing"); VNET_DEFINE(int, tcp_autorcvbuf_max) = 2*1024*1024; SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_max, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_autorcvbuf_max), 0, "Max size of automatic receive buffer"); VNET_DEFINE(struct inpcbhead, tcb); #define tcb6 tcb /* for KAME src sync over BSD*'s */ VNET_DEFINE(struct inpcbinfo, tcbinfo); /* * TCP statistics are stored in an array of counter(9)s, which size matches * size of struct tcpstat. TCP running connection count is a regular array. */ VNET_PCPUSTAT_DEFINE(struct tcpstat, tcpstat); SYSCTL_VNET_PCPUSTAT(_net_inet_tcp, TCPCTL_STATS, stats, struct tcpstat, tcpstat, "TCP statistics (struct tcpstat, netinet/tcp_var.h)"); VNET_DEFINE(counter_u64_t, tcps_states[TCP_NSTATES]); SYSCTL_COUNTER_U64_ARRAY(_net_inet_tcp, TCPCTL_STATES, states, CTLFLAG_RD | CTLFLAG_VNET, &VNET_NAME(tcps_states)[0], TCP_NSTATES, "TCP connection counts by TCP state"); static void tcp_vnet_init(const void *unused) { COUNTER_ARRAY_ALLOC(V_tcps_states, TCP_NSTATES, M_WAITOK); VNET_PCPUSTAT_ALLOC(tcpstat, M_WAITOK); } VNET_SYSINIT(tcp_vnet_init, SI_SUB_PROTO_IFATTACHDOMAIN, SI_ORDER_ANY, tcp_vnet_init, NULL); #ifdef VIMAGE static void tcp_vnet_uninit(const void *unused) { COUNTER_ARRAY_FREE(V_tcps_states, TCP_NSTATES); VNET_PCPUSTAT_FREE(tcpstat); } VNET_SYSUNINIT(tcp_vnet_uninit, SI_SUB_PROTO_IFATTACHDOMAIN, SI_ORDER_ANY, tcp_vnet_uninit, NULL); #endif /* VIMAGE */ /* * Kernel module interface for updating tcpstat. The first argument is an index * into tcpstat treated as an array. */ void kmod_tcpstat_add(int statnum, int val) { counter_u64_add(VNET(tcpstat)[statnum], val); } #ifdef TCP_HHOOK /* * Wrapper for the TCP established input helper hook. */ void hhook_run_tcp_est_in(struct tcpcb *tp, struct tcphdr *th, struct tcpopt *to) { struct tcp_hhook_data hhook_data; if (V_tcp_hhh[HHOOK_TCP_EST_IN]->hhh_nhooks > 0) { hhook_data.tp = tp; hhook_data.th = th; hhook_data.to = to; hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_IN], &hhook_data, tp->osd); } } #endif /* * CC wrapper hook functions */ void cc_ack_received(struct tcpcb *tp, struct tcphdr *th, uint16_t nsegs, uint16_t type) { #ifdef STATS int32_t gput; #endif INP_WLOCK_ASSERT(tp->t_inpcb); tp->ccv->nsegs = nsegs; tp->ccv->bytes_this_ack = BYTES_THIS_ACK(tp, th); if ((!V_tcp_do_newcwv && (tp->snd_cwnd <= tp->snd_wnd)) || (V_tcp_do_newcwv && (tp->snd_cwnd <= tp->snd_wnd) && (tp->snd_cwnd < (tcp_compute_pipe(tp) * 2)))) tp->ccv->flags |= CCF_CWND_LIMITED; else tp->ccv->flags &= ~CCF_CWND_LIMITED; if (type == CC_ACK) { #ifdef STATS stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF, ((int32_t)tp->snd_cwnd) - tp->snd_wnd); if (!IN_RECOVERY(tp->t_flags)) stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_ACKLEN, tp->ccv->bytes_this_ack / (tcp_maxseg(tp) * nsegs)); if ((tp->t_flags & TF_GPUTINPROG) && SEQ_GEQ(th->th_ack, tp->gput_ack)) { /* * Compute goodput in bits per millisecond. */ gput = (((int64_t)(th->th_ack - tp->gput_seq)) << 3) / max(1, tcp_ts_getticks() - tp->gput_ts); stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT, gput); /* * XXXLAS: This is a temporary hack, and should be * chained off VOI_TCP_GPUT when stats(9) grows an API * to deal with chained VOIs. */ if (tp->t_stats_gput_prev > 0) stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_GPUT_ND, ((gput - tp->t_stats_gput_prev) * 100) / tp->t_stats_gput_prev); tp->t_flags &= ~TF_GPUTINPROG; tp->t_stats_gput_prev = gput; } #endif /* STATS */ if (tp->snd_cwnd > tp->snd_ssthresh) { tp->t_bytes_acked += tp->ccv->bytes_this_ack; if (tp->t_bytes_acked >= tp->snd_cwnd) { tp->t_bytes_acked -= tp->snd_cwnd; tp->ccv->flags |= CCF_ABC_SENTAWND; } } else { tp->ccv->flags &= ~CCF_ABC_SENTAWND; tp->t_bytes_acked = 0; } } if (CC_ALGO(tp)->ack_received != NULL) { /* XXXLAS: Find a way to live without this */ tp->ccv->curack = th->th_ack; CC_ALGO(tp)->ack_received(tp->ccv, type); } #ifdef STATS stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, tp->snd_cwnd); #endif } void cc_conn_init(struct tcpcb *tp) { struct hc_metrics_lite metrics; struct inpcb *inp = tp->t_inpcb; u_int maxseg; int rtt; INP_WLOCK_ASSERT(tp->t_inpcb); tcp_hc_get(&inp->inp_inc, &metrics); maxseg = tcp_maxseg(tp); if (tp->t_srtt == 0 && (rtt = metrics.rmx_rtt)) { tp->t_srtt = rtt; tp->t_rttbest = tp->t_srtt + TCP_RTT_SCALE; TCPSTAT_INC(tcps_usedrtt); if (metrics.rmx_rttvar) { tp->t_rttvar = metrics.rmx_rttvar; TCPSTAT_INC(tcps_usedrttvar); } else { /* default variation is +- 1 rtt */ tp->t_rttvar = tp->t_srtt * TCP_RTTVAR_SCALE / TCP_RTT_SCALE; } TCPT_RANGESET(tp->t_rxtcur, ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1, tp->t_rttmin, TCPTV_REXMTMAX); } if (metrics.rmx_ssthresh) { /* * There's some sort of gateway or interface * buffer limit on the path. Use this to set * the slow start threshold, but set the * threshold to no less than 2*mss. */ tp->snd_ssthresh = max(2 * maxseg, metrics.rmx_ssthresh); TCPSTAT_INC(tcps_usedssthresh); } /* * Set the initial slow-start flight size. * * If a SYN or SYN/ACK was lost and retransmitted, we have to * reduce the initial CWND to one segment as congestion is likely * requiring us to be cautious. */ if (tp->snd_cwnd == 1) tp->snd_cwnd = maxseg; /* SYN(-ACK) lost */ else tp->snd_cwnd = tcp_compute_initwnd(maxseg); if (CC_ALGO(tp)->conn_init != NULL) CC_ALGO(tp)->conn_init(tp->ccv); } void inline cc_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type) { INP_WLOCK_ASSERT(tp->t_inpcb); #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type); #endif switch(type) { case CC_NDUPACK: if (!IN_FASTRECOVERY(tp->t_flags)) { tp->snd_recover = tp->snd_max; if (tp->t_flags2 & TF2_ECN_PERMIT) tp->t_flags2 |= TF2_ECN_SND_CWR; } break; case CC_ECN: if (!IN_CONGRECOVERY(tp->t_flags) || /* * Allow ECN reaction on ACK to CWR, if * that data segment was also CE marked. */ SEQ_GEQ(th->th_ack, tp->snd_recover)) { EXIT_CONGRECOVERY(tp->t_flags); TCPSTAT_INC(tcps_ecn_rcwnd); tp->snd_recover = tp->snd_max + 1; if (tp->t_flags2 & TF2_ECN_PERMIT) tp->t_flags2 |= TF2_ECN_SND_CWR; } break; case CC_RTO: tp->t_dupacks = 0; tp->t_bytes_acked = 0; EXIT_RECOVERY(tp->t_flags); if (tp->t_flags2 & TF2_ECN_PERMIT) tp->t_flags2 |= TF2_ECN_SND_CWR; break; case CC_RTO_ERR: TCPSTAT_INC(tcps_sndrexmitbad); /* RTO was unnecessary, so reset everything. */ tp->snd_cwnd = tp->snd_cwnd_prev; tp->snd_ssthresh = tp->snd_ssthresh_prev; tp->snd_recover = tp->snd_recover_prev; if (tp->t_flags & TF_WASFRECOVERY) ENTER_FASTRECOVERY(tp->t_flags); if (tp->t_flags & TF_WASCRECOVERY) ENTER_CONGRECOVERY(tp->t_flags); tp->snd_nxt = tp->snd_max; tp->t_flags &= ~TF_PREVVALID; tp->t_badrxtwin = 0; break; } if (CC_ALGO(tp)->cong_signal != NULL) { if (th != NULL) tp->ccv->curack = th->th_ack; CC_ALGO(tp)->cong_signal(tp->ccv, type); } } void inline cc_post_recovery(struct tcpcb *tp, struct tcphdr *th) { INP_WLOCK_ASSERT(tp->t_inpcb); /* XXXLAS: KASSERT that we're in recovery? */ if (CC_ALGO(tp)->post_recovery != NULL) { tp->ccv->curack = th->th_ack; CC_ALGO(tp)->post_recovery(tp->ccv); } /* XXXLAS: EXIT_RECOVERY ? */ tp->t_bytes_acked = 0; tp->sackhint.prr_out = 0; } /* * Indicate whether this ack should be delayed. We can delay the ack if * following conditions are met: * - There is no delayed ack timer in progress. * - Our last ack wasn't a 0-sized window. We never want to delay * the ack that opens up a 0-sized window. * - LRO wasn't used for this segment. We make sure by checking that the * segment size is not larger than the MSS. */ #define DELAY_ACK(tp, tlen) \ ((!tcp_timer_active(tp, TT_DELACK) && \ (tp->t_flags & TF_RXWIN0SENT) == 0) && \ (tlen <= tp->t_maxseg) && \ (V_tcp_delack_enabled || (tp->t_flags & TF_NEEDSYN))) void inline cc_ecnpkt_handler(struct tcpcb *tp, struct tcphdr *th, uint8_t iptos) { INP_WLOCK_ASSERT(tp->t_inpcb); if (CC_ALGO(tp)->ecnpkt_handler != NULL) { switch (iptos & IPTOS_ECN_MASK) { case IPTOS_ECN_CE: tp->ccv->flags |= CCF_IPHDR_CE; break; case IPTOS_ECN_ECT0: /* FALLTHROUGH */ case IPTOS_ECN_ECT1: /* FALLTHROUGH */ case IPTOS_ECN_NOTECT: tp->ccv->flags &= ~CCF_IPHDR_CE; break; } if (th->th_flags & TH_CWR) tp->ccv->flags |= CCF_TCPHDR_CWR; else tp->ccv->flags &= ~CCF_TCPHDR_CWR; CC_ALGO(tp)->ecnpkt_handler(tp->ccv); if (tp->ccv->flags & CCF_ACKNOW) { tcp_timer_activate(tp, TT_DELACK, tcp_delacktime); tp->t_flags |= TF_ACKNOW; } } } /* * TCP input handling is split into multiple parts: * tcp6_input is a thin wrapper around tcp_input for the extended * ip6_protox[] call format in ip6_input * tcp_input handles primary segment validation, inpcb lookup and * SYN processing on listen sockets * tcp_do_segment processes the ACK and text of the segment for * establishing, established and closing connections */ #ifdef INET6 int -tcp6_input(struct mbuf **mp, int *offp, int proto) +tcp6_input_with_port(struct mbuf **mp, int *offp, int proto, uint16_t port) { struct mbuf *m; struct in6_ifaddr *ia6; struct ip6_hdr *ip6; m = *mp; if (m->m_len < *offp + sizeof(struct tcphdr)) { m = m_pullup(m, *offp + sizeof(struct tcphdr)); if (m == NULL) { *mp = m; TCPSTAT_INC(tcps_rcvshort); return (IPPROTO_DONE); } } /* * draft-itojun-ipv6-tcp-to-anycast * better place to put this in? */ ip6 = mtod(m, struct ip6_hdr *); ia6 = in6ifa_ifwithaddr(&ip6->ip6_dst, 0 /* XXX */, false); if (ia6 && (ia6->ia6_flags & IN6_IFF_ANYCAST)) { icmp6_error(m, ICMP6_DST_UNREACH, ICMP6_DST_UNREACH_ADDR, (caddr_t)&ip6->ip6_dst - (caddr_t)ip6); *mp = NULL; return (IPPROTO_DONE); } *mp = m; - return (tcp_input(mp, offp, proto)); + return (tcp_input_with_port(mp, offp, proto, port)); +} + +int +tcp6_input(struct mbuf **mp, int *offp, int proto) +{ + + return(tcp6_input_with_port(mp, offp, proto, 0)); } #endif /* INET6 */ int -tcp_input(struct mbuf **mp, int *offp, int proto) +tcp_input_with_port(struct mbuf **mp, int *offp, int proto, uint16_t port) { struct mbuf *m = *mp; struct tcphdr *th = NULL; struct ip *ip = NULL; struct inpcb *inp = NULL; struct tcpcb *tp = NULL; struct socket *so = NULL; u_char *optp = NULL; int off0; int optlen = 0; #ifdef INET int len; uint8_t ipttl; #endif int tlen = 0, off; int drop_hdrlen; int thflags; int rstreason = 0; /* For badport_bandlim accounting purposes */ int lookupflag; uint8_t iptos; struct m_tag *fwd_tag = NULL; #ifdef INET6 struct ip6_hdr *ip6 = NULL; int isipv6; #else const void *ip6 = NULL; #endif /* INET6 */ struct tcpopt to; /* options in this segment */ char *s = NULL; /* address and port logging */ #ifdef TCPDEBUG /* * The size of tcp_saveipgen must be the size of the max ip header, * now IPv6. */ u_char tcp_saveipgen[IP6_HDR_LEN]; struct tcphdr tcp_savetcp; short ostate = 0; #endif NET_EPOCH_ASSERT(); #ifdef INET6 isipv6 = (mtod(m, struct ip *)->ip_v == 6) ? 1 : 0; #endif off0 = *offp; m = *mp; *mp = NULL; to.to_flags = 0; TCPSTAT_INC(tcps_rcvtotal); #ifdef INET6 if (isipv6) { ip6 = mtod(m, struct ip6_hdr *); th = (struct tcphdr *)((caddr_t)ip6 + off0); tlen = sizeof(*ip6) + ntohs(ip6->ip6_plen) - off0; + if (port) + goto skip6_csum; if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID_IPV6) { if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) th->th_sum = m->m_pkthdr.csum_data; else th->th_sum = in6_cksum_pseudo(ip6, tlen, IPPROTO_TCP, m->m_pkthdr.csum_data); th->th_sum ^= 0xffff; } else th->th_sum = in6_cksum(m, IPPROTO_TCP, off0, tlen); if (th->th_sum) { TCPSTAT_INC(tcps_rcvbadsum); goto drop; } - + skip6_csum: /* * Be proactive about unspecified IPv6 address in source. * As we use all-zero to indicate unbounded/unconnected pcb, * unspecified IPv6 address can be used to confuse us. * * Note that packets with unspecified IPv6 destination is * already dropped in ip6_input. */ if (IN6_IS_ADDR_UNSPECIFIED(&ip6->ip6_src)) { /* XXX stat */ goto drop; } iptos = IPV6_TRAFFIC_CLASS(ip6); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { /* * Get IP and TCP header together in first mbuf. * Note: IP leaves IP header in first mbuf. */ if (off0 > sizeof (struct ip)) { ip_stripoptions(m); off0 = sizeof(struct ip); } if (m->m_len < sizeof (struct tcpiphdr)) { if ((m = m_pullup(m, sizeof (struct tcpiphdr))) == NULL) { TCPSTAT_INC(tcps_rcvshort); return (IPPROTO_DONE); } } ip = mtod(m, struct ip *); th = (struct tcphdr *)((caddr_t)ip + off0); tlen = ntohs(ip->ip_len) - off0; iptos = ip->ip_tos; + if (port) + goto skip_csum; if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) { if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) th->th_sum = m->m_pkthdr.csum_data; else th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htonl(m->m_pkthdr.csum_data + tlen + IPPROTO_TCP)); th->th_sum ^= 0xffff; } else { struct ipovly *ipov = (struct ipovly *)ip; /* * Checksum extended TCP header and data. */ len = off0 + tlen; ipttl = ip->ip_ttl; bzero(ipov->ih_x1, sizeof(ipov->ih_x1)); ipov->ih_len = htons(tlen); th->th_sum = in_cksum(m, len); /* Reset length for SDT probes. */ ip->ip_len = htons(len); /* Reset TOS bits */ ip->ip_tos = iptos; /* Re-initialization for later version check */ ip->ip_ttl = ipttl; ip->ip_v = IPVERSION; ip->ip_hl = off0 >> 2; } - - if (th->th_sum) { + skip_csum: + if (th->th_sum && (port == 0)) { TCPSTAT_INC(tcps_rcvbadsum); goto drop; } } #endif /* INET */ /* * Check that TCP offset makes sense, * pull out TCP options and adjust length. XXX */ off = th->th_off << 2; if (off < sizeof (struct tcphdr) || off > tlen) { TCPSTAT_INC(tcps_rcvbadoff); goto drop; } tlen -= off; /* tlen is used instead of ti->ti_len */ if (off > sizeof (struct tcphdr)) { #ifdef INET6 if (isipv6) { if (m->m_len < off0 + off) { m = m_pullup(m, off0 + off); if (m == NULL) { TCPSTAT_INC(tcps_rcvshort); return (IPPROTO_DONE); } } ip6 = mtod(m, struct ip6_hdr *); th = (struct tcphdr *)((caddr_t)ip6 + off0); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { if (m->m_len < sizeof(struct ip) + off) { if ((m = m_pullup(m, sizeof (struct ip) + off)) == NULL) { TCPSTAT_INC(tcps_rcvshort); return (IPPROTO_DONE); } ip = mtod(m, struct ip *); th = (struct tcphdr *)((caddr_t)ip + off0); } } #endif optlen = off - sizeof (struct tcphdr); optp = (u_char *)(th + 1); } thflags = th->th_flags; /* * Convert TCP protocol specific fields to host format. */ tcp_fields_to_host(th); /* * Delay dropping TCP, IP headers, IPv6 ext headers, and TCP options. */ drop_hdrlen = off0 + off; /* * Grab info from PACKET_TAG_IPFORWARD tag prepended to the chain. */ if ( #ifdef INET6 (isipv6 && (m->m_flags & M_IP6_NEXTHOP)) #ifdef INET || (!isipv6 && (m->m_flags & M_IP_NEXTHOP)) #endif #endif #if defined(INET) && !defined(INET6) (m->m_flags & M_IP_NEXTHOP) #endif ) fwd_tag = m_tag_find(m, PACKET_TAG_IPFORWARD, NULL); /* * For initial SYN packets arriving on listening socket, * we don't need write lock. */ lookupflag = (thflags & (TH_ACK|TH_SYN)) == TH_SYN ? INPLOOKUP_RLOCKLISTEN : INPLOOKUP_WLOCKPCB; findpcb: #ifdef INET6 if (isipv6 && fwd_tag != NULL) { struct sockaddr_in6 *next_hop6; next_hop6 = (struct sockaddr_in6 *)(fwd_tag + 1); /* * Transparently forwarded. Pretend to be the destination. * Already got one like this? */ inp = in6_pcblookup_mbuf(&V_tcbinfo, &ip6->ip6_src, th->th_sport, &ip6->ip6_dst, th->th_dport, lookupflag, m->m_pkthdr.rcvif, m); if (!inp) { /* * It's new. Try to find the ambushing socket. * Because we've rewritten the destination address, * any hardware-generated hash is ignored. */ inp = in6_pcblookup(&V_tcbinfo, &ip6->ip6_src, th->th_sport, &next_hop6->sin6_addr, next_hop6->sin6_port ? ntohs(next_hop6->sin6_port) : th->th_dport, INPLOOKUP_WILDCARD | lookupflag, m->m_pkthdr.rcvif); } } else if (isipv6) { inp = in6_pcblookup_mbuf(&V_tcbinfo, &ip6->ip6_src, th->th_sport, &ip6->ip6_dst, th->th_dport, INPLOOKUP_WILDCARD | lookupflag, m->m_pkthdr.rcvif, m); } #endif /* INET6 */ #if defined(INET6) && defined(INET) else #endif #ifdef INET if (fwd_tag != NULL) { struct sockaddr_in *next_hop; next_hop = (struct sockaddr_in *)(fwd_tag+1); /* * Transparently forwarded. Pretend to be the destination. * already got one like this? */ inp = in_pcblookup_mbuf(&V_tcbinfo, ip->ip_src, th->th_sport, ip->ip_dst, th->th_dport, lookupflag, m->m_pkthdr.rcvif, m); if (!inp) { /* * It's new. Try to find the ambushing socket. * Because we've rewritten the destination address, * any hardware-generated hash is ignored. */ inp = in_pcblookup(&V_tcbinfo, ip->ip_src, th->th_sport, next_hop->sin_addr, next_hop->sin_port ? ntohs(next_hop->sin_port) : th->th_dport, INPLOOKUP_WILDCARD | lookupflag, m->m_pkthdr.rcvif); } } else inp = in_pcblookup_mbuf(&V_tcbinfo, ip->ip_src, th->th_sport, ip->ip_dst, th->th_dport, INPLOOKUP_WILDCARD | lookupflag, m->m_pkthdr.rcvif, m); #endif /* INET */ /* * If the INPCB does not exist then all data in the incoming * segment is discarded and an appropriate RST is sent back. * XXX MRT Send RST using which routing table? */ if (inp == NULL) { /* * Log communication attempts to ports that are not * in use. */ if ((V_tcp_log_in_vain == 1 && (thflags & TH_SYN)) || V_tcp_log_in_vain == 2) { if ((s = tcp_log_vain(NULL, th, (void *)ip, ip6))) log(LOG_INFO, "%s; %s: Connection attempt " "to closed port\n", s, __func__); } /* * When blackholing do not respond with a RST but * completely ignore the segment and drop it. */ if ((V_blackhole == 1 && (thflags & TH_SYN)) || V_blackhole == 2) goto dropunlock; rstreason = BANDLIM_RST_CLOSEDPORT; goto dropwithreset; } INP_LOCK_ASSERT(inp); /* * While waiting for inp lock during the lookup, another thread * can have dropped the inpcb, in which case we need to loop back * and try to find a new inpcb to deliver to. */ if (inp->inp_flags & INP_DROPPED) { INP_UNLOCK(inp); inp = NULL; goto findpcb; } if ((inp->inp_flowtype == M_HASHTYPE_NONE) && (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) && ((inp->inp_socket == NULL) || (inp->inp_socket->so_options & SO_ACCEPTCONN) == 0)) { inp->inp_flowid = m->m_pkthdr.flowid; inp->inp_flowtype = M_HASHTYPE_GET(m); } #if defined(IPSEC) || defined(IPSEC_SUPPORT) #ifdef INET6 if (isipv6 && IPSEC_ENABLED(ipv6) && IPSEC_CHECK_POLICY(ipv6, m, inp) != 0) { goto dropunlock; } #ifdef INET else #endif #endif /* INET6 */ #ifdef INET if (IPSEC_ENABLED(ipv4) && IPSEC_CHECK_POLICY(ipv4, m, inp) != 0) { goto dropunlock; } #endif /* INET */ #endif /* IPSEC */ /* * Check the minimum TTL for socket. */ if (inp->inp_ip_minttl != 0) { #ifdef INET6 if (isipv6) { if (inp->inp_ip_minttl > ip6->ip6_hlim) goto dropunlock; } else #endif if (inp->inp_ip_minttl > ip->ip_ttl) goto dropunlock; } /* * A previous connection in TIMEWAIT state is supposed to catch stray * or duplicate segments arriving late. If this segment was a * legitimate new connection attempt, the old INPCB gets removed and * we can try again to find a listening socket. * * At this point, due to earlier optimism, we may hold only an inpcb * lock, and not the inpcbinfo write lock. If so, we need to try to * acquire it, or if that fails, acquire a reference on the inpcb, * drop all locks, acquire a global write lock, and then re-acquire * the inpcb lock. We may at that point discover that another thread * has tried to free the inpcb, in which case we need to loop back * and try to find a new inpcb to deliver to. * * XXXRW: It may be time to rethink timewait locking. */ if (inp->inp_flags & INP_TIMEWAIT) { tcp_dooptions(&to, optp, optlen, (thflags & TH_SYN) ? TO_SYN : 0); /* * NB: tcp_twcheck unlocks the INP and frees the mbuf. */ if (tcp_twcheck(inp, &to, th, m, tlen)) goto findpcb; return (IPPROTO_DONE); } /* * The TCPCB may no longer exist if the connection is winding * down or it is in the CLOSED state. Either way we drop the * segment and send an appropriate response. */ tp = intotcpcb(inp); if (tp == NULL || tp->t_state == TCPS_CLOSED) { rstreason = BANDLIM_RST_CLOSEDPORT; goto dropwithreset; } + if ((tp->t_port != port) && (tp->t_state > TCPS_LISTEN)) { + rstreason = BANDLIM_RST_CLOSEDPORT; + goto dropwithreset; + } + #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) { tcp_offload_input(tp, m); m = NULL; /* consumed by the TOE driver */ goto dropunlock; } #endif #ifdef MAC if (mac_inpcb_check_deliver(inp, m)) goto dropunlock; #endif so = inp->inp_socket; KASSERT(so != NULL, ("%s: so == NULL", __func__)); #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) { ostate = tp->t_state; #ifdef INET6 if (isipv6) { bcopy((char *)ip6, (char *)tcp_saveipgen, sizeof(*ip6)); } else #endif bcopy((char *)ip, (char *)tcp_saveipgen, sizeof(*ip)); tcp_savetcp = *th; } #endif /* TCPDEBUG */ /* * When the socket is accepting connections (the INPCB is in LISTEN * state) we look into the SYN cache if this is a new connection * attempt or the completion of a previous one. */ KASSERT(tp->t_state == TCPS_LISTEN || !(so->so_options & SO_ACCEPTCONN), ("%s: so accepting but tp %p not listening", __func__, tp)); if (tp->t_state == TCPS_LISTEN && (so->so_options & SO_ACCEPTCONN)) { struct in_conninfo inc; bzero(&inc, sizeof(inc)); #ifdef INET6 if (isipv6) { inc.inc_flags |= INC_ISIPV6; if (inp->inp_inc.inc_flags & INC_IPV6MINMTU) inc.inc_flags |= INC_IPV6MINMTU; inc.inc6_faddr = ip6->ip6_src; inc.inc6_laddr = ip6->ip6_dst; } else #endif { inc.inc_faddr = ip->ip_src; inc.inc_laddr = ip->ip_dst; } inc.inc_fport = th->th_sport; inc.inc_lport = th->th_dport; inc.inc_fibnum = so->so_fibnum; /* * Check for an existing connection attempt in syncache if * the flag is only ACK. A successful lookup creates a new * socket appended to the listen queue in SYN_RECEIVED state. */ if ((thflags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK) { /* * Parse the TCP options here because * syncookies need access to the reflected * timestamp. */ tcp_dooptions(&to, optp, optlen, 0); /* * NB: syncache_expand() doesn't unlock * inp and tcpinfo locks. */ - rstreason = syncache_expand(&inc, &to, th, &so, m); + rstreason = syncache_expand(&inc, &to, th, &so, m, port); if (rstreason < 0) { /* * A failing TCP MD5 signature comparison * must result in the segment being dropped * and must not produce any response back * to the sender. */ goto dropunlock; } else if (rstreason == 0) { /* * No syncache entry or ACK was not * for our SYN/ACK. Send a RST. * NB: syncache did its own logging * of the failure cause. */ rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } tfo_socket_result: if (so == NULL) { /* * We completed the 3-way handshake * but could not allocate a socket * either due to memory shortage, * listen queue length limits or * global socket limits. Send RST * or wait and have the remote end * retransmit the ACK for another * try. */ if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Socket allocation failed due to " "limits or memory shortage, %s\n", s, __func__, V_tcp_sc_rst_sock_fail ? "sending RST" : "try again"); if (V_tcp_sc_rst_sock_fail) { rstreason = BANDLIM_UNLIMITED; goto dropwithreset; } else goto dropunlock; } /* * Socket is created in state SYN_RECEIVED. * Unlock the listen socket, lock the newly * created socket and update the tp variable. * If we came here via jump to tfo_socket_result, * then listening socket is read-locked. */ INP_UNLOCK(inp); /* listen socket */ inp = sotoinpcb(so); /* * New connection inpcb is already locked by * syncache_expand(). */ INP_WLOCK_ASSERT(inp); tp = intotcpcb(inp); KASSERT(tp->t_state == TCPS_SYN_RECEIVED, ("%s: ", __func__)); /* * Process the segment and the data it * contains. tcp_do_segment() consumes * the mbuf chain and unlocks the inpcb. */ TCP_PROBE5(receive, NULL, tp, m, tp, th); tp->t_fb->tfb_tcp_do_segment(m, th, so, tp, drop_hdrlen, tlen, iptos); return (IPPROTO_DONE); } /* * Segment flag validation for new connection attempts: * * Our (SYN|ACK) response was rejected. * Check with syncache and remove entry to prevent * retransmits. * * NB: syncache_chkrst does its own logging of failure * causes. */ if (thflags & TH_RST) { - syncache_chkrst(&inc, th, m); + syncache_chkrst(&inc, th, m, port); goto dropunlock; } /* * We can't do anything without SYN. */ if ((thflags & TH_SYN) == 0) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "SYN is missing, segment ignored\n", s, __func__); TCPSTAT_INC(tcps_badsyn); goto dropunlock; } /* * (SYN|ACK) is bogus on a listen socket. */ if (thflags & TH_ACK) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "SYN|ACK invalid, segment rejected\n", s, __func__); - syncache_badack(&inc); /* XXX: Not needed! */ + syncache_badack(&inc, port); /* XXX: Not needed! */ TCPSTAT_INC(tcps_badsyn); rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } /* * If the drop_synfin option is enabled, drop all * segments with both the SYN and FIN bits set. * This prevents e.g. nmap from identifying the * TCP/IP stack. * XXX: Poor reasoning. nmap has other methods * and is constantly refining its stack detection * strategies. * XXX: This is a violation of the TCP specification * and was used by RFC1644. */ if ((thflags & TH_FIN) && V_drop_synfin) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "SYN|FIN segment ignored (based on " "sysctl setting)\n", s, __func__); TCPSTAT_INC(tcps_badsyn); goto dropunlock; } /* * Segment's flags are (SYN) or (SYN|FIN). * * TH_PUSH, TH_URG, TH_ECE, TH_CWR are ignored * as they do not affect the state of the TCP FSM. * The data pointed to by TH_URG and th_urp is ignored. */ KASSERT((thflags & (TH_RST|TH_ACK)) == 0, ("%s: Listen socket: TH_RST or TH_ACK set", __func__)); KASSERT(thflags & (TH_SYN), ("%s: Listen socket: TH_SYN not set", __func__)); INP_RLOCK_ASSERT(inp); #ifdef INET6 /* * If deprecated address is forbidden, * we do not accept SYN to deprecated interface * address to prevent any new inbound connection from * getting established. * When we do not accept SYN, we send a TCP RST, * with deprecated source address (instead of dropping * it). We compromise it as it is much better for peer * to send a RST, and RST will be the final packet * for the exchange. * * If we do not forbid deprecated addresses, we accept * the SYN packet. RFC2462 does not suggest dropping * SYN in this case. * If we decipher RFC2462 5.5.4, it says like this: * 1. use of deprecated addr with existing * communication is okay - "SHOULD continue to be * used" * 2. use of it with new communication: * (2a) "SHOULD NOT be used if alternate address * with sufficient scope is available" * (2b) nothing mentioned otherwise. * Here we fall into (2b) case as we have no choice in * our source address selection - we must obey the peer. * * The wording in RFC2462 is confusing, and there are * multiple description text for deprecated address * handling - worse, they are not exactly the same. * I believe 5.5.4 is the best one, so we follow 5.5.4. */ if (isipv6 && !V_ip6_use_deprecated) { struct in6_ifaddr *ia6; ia6 = in6ifa_ifwithaddr(&ip6->ip6_dst, 0 /* XXX */, false); if (ia6 != NULL && (ia6->ia6_flags & IN6_IFF_DEPRECATED)) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Connection attempt to deprecated " "IPv6 address rejected\n", s, __func__); rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } } #endif /* INET6 */ /* * Basic sanity checks on incoming SYN requests: * Don't respond if the destination is a link layer * broadcast according to RFC1122 4.2.3.10, p. 104. * If it is from this socket it must be forged. * Don't respond if the source or destination is a * global or subnet broad- or multicast address. * Note that it is quite possible to receive unicast * link-layer packets with a broadcast IP address. Use * in_broadcast() to find them. */ if (m->m_flags & (M_BCAST|M_MCAST)) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Connection attempt from broad- or multicast " "link layer address ignored\n", s, __func__); goto dropunlock; } #ifdef INET6 if (isipv6) { if (th->th_dport == th->th_sport && IN6_ARE_ADDR_EQUAL(&ip6->ip6_dst, &ip6->ip6_src)) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Connection attempt to/from self " "ignored\n", s, __func__); goto dropunlock; } if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) || IN6_IS_ADDR_MULTICAST(&ip6->ip6_src)) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Connection attempt from/to multicast " "address ignored\n", s, __func__); goto dropunlock; } } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { if (th->th_dport == th->th_sport && ip->ip_dst.s_addr == ip->ip_src.s_addr) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Connection attempt from/to self " "ignored\n", s, __func__); goto dropunlock; } if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) || IN_MULTICAST(ntohl(ip->ip_src.s_addr)) || ip->ip_src.s_addr == htonl(INADDR_BROADCAST) || in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif)) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Connection attempt from/to broad- " "or multicast address ignored\n", s, __func__); goto dropunlock; } } #endif /* * SYN appears to be valid. Create compressed TCP state * for syncache. */ #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif TCP_PROBE3(debug__input, tp, th, m); tcp_dooptions(&to, optp, optlen, TO_SYN); if ((so = syncache_add(&inc, &to, th, inp, so, m, NULL, NULL, - iptos)) != NULL) + iptos, port)) != NULL) goto tfo_socket_result; /* * Entry added to syncache and mbuf consumed. * Only the listen socket is unlocked by syncache_add(). */ INP_INFO_WUNLOCK_ASSERT(&V_tcbinfo); return (IPPROTO_DONE); } else if (tp->t_state == TCPS_LISTEN) { /* * When a listen socket is torn down the SO_ACCEPTCONN * flag is removed first while connections are drained * from the accept queue in a unlock/lock cycle of the * ACCEPT_LOCK, opening a race condition allowing a SYN * attempt go through unhandled. */ goto dropunlock; } #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (tp->t_flags & TF_SIGNATURE) { tcp_dooptions(&to, optp, optlen, thflags); if ((to.to_flags & TOF_SIGNATURE) == 0) { TCPSTAT_INC(tcps_sig_err_nosigopt); goto dropunlock; } if (!TCPMD5_ENABLED() || TCPMD5_INPUT(m, th, to.to_signature) != 0) goto dropunlock; } #endif TCP_PROBE5(receive, NULL, tp, m, tp, th); /* * Segment belongs to a connection in SYN_SENT, ESTABLISHED or later * state. tcp_do_segment() always consumes the mbuf chain, unlocks * the inpcb, and unlocks pcbinfo. */ tp->t_fb->tfb_tcp_do_segment(m, th, so, tp, drop_hdrlen, tlen, iptos); return (IPPROTO_DONE); dropwithreset: TCP_PROBE5(receive, NULL, tp, m, tp, th); if (inp != NULL) { tcp_dropwithreset(m, th, tp, tlen, rstreason); INP_UNLOCK(inp); } else tcp_dropwithreset(m, th, NULL, tlen, rstreason); m = NULL; /* mbuf chain got consumed. */ goto drop; dropunlock: if (m != NULL) TCP_PROBE5(receive, NULL, tp, m, tp, th); if (inp != NULL) INP_UNLOCK(inp); drop: INP_INFO_WUNLOCK_ASSERT(&V_tcbinfo); if (s != NULL) free(s, M_TCPLOG); if (m != NULL) m_freem(m); return (IPPROTO_DONE); } /* * Automatic sizing of receive socket buffer. Often the send * buffer size is not optimally adjusted to the actual network * conditions at hand (delay bandwidth product). Setting the * buffer size too small limits throughput on links with high * bandwidth and high delay (eg. trans-continental/oceanic links). * * On the receive side the socket buffer memory is only rarely * used to any significant extent. This allows us to be much * more aggressive in scaling the receive socket buffer. For * the case that the buffer space is actually used to a large * extent and we run out of kernel memory we can simply drop * the new segments; TCP on the sender will just retransmit it * later. Setting the buffer size too big may only consume too * much kernel memory if the application doesn't read() from * the socket or packet loss or reordering makes use of the * reassembly queue. * * The criteria to step up the receive buffer one notch are: * 1. Application has not set receive buffer size with * SO_RCVBUF. Setting SO_RCVBUF clears SB_AUTOSIZE. * 2. the number of bytes received during 1/2 of an sRTT * is at least 3/8 of the current socket buffer size. * 3. receive buffer size has not hit maximal automatic size; * * If all of the criteria are met we increaset the socket buffer * by a 1/2 (bounded by the max). This allows us to keep ahead * of slow-start but also makes it so our peer never gets limited * by our rwnd which we then open up causing a burst. * * This algorithm does two steps per RTT at most and only if * we receive a bulk stream w/o packet losses or reorderings. * Shrinking the buffer during idle times is not necessary as * it doesn't consume any memory when idle. * * TODO: Only step up if the application is actually serving * the buffer to better manage the socket buffer resources. */ int tcp_autorcvbuf(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int tlen) { int newsize = 0; if (V_tcp_do_autorcvbuf && (so->so_rcv.sb_flags & SB_AUTOSIZE) && tp->t_srtt != 0 && tp->rfbuf_ts != 0 && TCP_TS_TO_TICKS(tcp_ts_getticks() - tp->rfbuf_ts) > ((tp->t_srtt >> TCP_RTT_SHIFT)/2)) { if (tp->rfbuf_cnt > ((so->so_rcv.sb_hiwat / 2)/ 4 * 3) && so->so_rcv.sb_hiwat < V_tcp_autorcvbuf_max) { newsize = min((so->so_rcv.sb_hiwat + (so->so_rcv.sb_hiwat/2)), V_tcp_autorcvbuf_max); } TCP_PROBE6(receive__autoresize, NULL, tp, m, tp, th, newsize); /* Start over with next RTT. */ tp->rfbuf_ts = 0; tp->rfbuf_cnt = 0; } else { tp->rfbuf_cnt += tlen; /* add up */ } return (newsize); } +int +tcp_input(struct mbuf **mp, int *offp, int proto) +{ + return(tcp_input_with_port(mp, offp, proto, 0)); +} + void tcp_handle_wakeup(struct tcpcb *tp, struct socket *so) { /* * Since tp might be gone if the session entered * the TIME_WAIT state before coming here, we need * to check if the socket is still connected. */ if ((so->so_state & SS_ISCONNECTED) == 0) return; INP_LOCK_ASSERT(tp->t_inpcb); if (tp->t_flags & TF_WAKESOR) { tp->t_flags &= ~TF_WAKESOR; SOCKBUF_UNLOCK_ASSERT(&so->so_rcv); sorwakeup(so); } if (tp->t_flags & TF_WAKESOW) { tp->t_flags &= ~TF_WAKESOW; SOCKBUF_UNLOCK_ASSERT(&so->so_snd); sowwakeup(so); } } void tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int drop_hdrlen, int tlen, uint8_t iptos) { int thflags, acked, ourfinisacked, needoutput = 0, sack_changed; int rstreason, todrop, win, incforsyn = 0; uint32_t tiwin; uint16_t nsegs; char *s; struct in_conninfo *inc; struct mbuf *mfree; struct tcpopt to; int tfo_syn; u_int maxseg; #ifdef TCPDEBUG /* * The size of tcp_saveipgen must be the size of the max ip header, * now IPv6. */ u_char tcp_saveipgen[IP6_HDR_LEN]; struct tcphdr tcp_savetcp; short ostate = 0; #endif thflags = th->th_flags; inc = &tp->t_inpcb->inp_inc; tp->sackhint.last_sack_ack = 0; sack_changed = 0; nsegs = max(1, m->m_pkthdr.lro_nsegs); NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(tp->t_inpcb); KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", __func__)); KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", __func__)); #ifdef TCPPCAP /* Save segment, if requested. */ tcp_pcap_add(th, m, &(tp->t_inpkts)); #endif TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0, tlen, NULL, true); if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: " "SYN|FIN segment ignored (based on " "sysctl setting)\n", s, __func__); free(s, M_TCPLOG); } goto drop; } /* * If a segment with the ACK-bit set arrives in the SYN-SENT state * check SEQ.ACK first. */ if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { rstreason = BANDLIM_UNLIMITED; goto dropwithreset; } /* * Segment received on connection. * Reset idle time and keep-alive timer. * XXX: This should be done after segment * validation to ignore broken/spoofed segs. */ tp->t_rcvtime = ticks; /* * Scale up the window into a 32-bit value. * For the SYN_SENT state the scale is zero. */ tiwin = th->th_win << tp->snd_scale; #ifdef STATS stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin); #endif /* * TCP ECN processing. */ if (tp->t_flags2 & TF2_ECN_PERMIT) { if (thflags & TH_CWR) { tp->t_flags2 &= ~TF2_ECN_SND_ECE; tp->t_flags |= TF_ACKNOW; } switch (iptos & IPTOS_ECN_MASK) { case IPTOS_ECN_CE: tp->t_flags2 |= TF2_ECN_SND_ECE; TCPSTAT_INC(tcps_ecn_ce); break; case IPTOS_ECN_ECT0: TCPSTAT_INC(tcps_ecn_ect0); break; case IPTOS_ECN_ECT1: TCPSTAT_INC(tcps_ecn_ect1); break; } /* Process a packet differently from RFC3168. */ cc_ecnpkt_handler(tp, th, iptos); /* Congestion experienced. */ if (thflags & TH_ECE) { cc_cong_signal(tp, th, CC_ECN); } } /* * Parse options on any incoming segment. */ tcp_dooptions(&to, (u_char *)(th + 1), (th->th_off << 2) - sizeof(struct tcphdr), (thflags & TH_SYN) ? TO_SYN : 0); #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if ((tp->t_flags & TF_SIGNATURE) != 0 && (to.to_flags & TOF_SIGNATURE) == 0) { TCPSTAT_INC(tcps_sig_err_sigopt); /* XXX: should drop? */ } #endif /* * If echoed timestamp is later than the current time, * fall back to non RFC1323 RTT calculation. Normalize * timestamp if syncookies were used when this connection * was established. */ if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { to.to_tsecr -= tp->ts_offset; if (TSTMP_GT(to.to_tsecr, tcp_ts_getticks())) to.to_tsecr = 0; else if (tp->t_flags & TF_PREVVALID && tp->t_badrxtwin != 0 && SEQ_LT(to.to_tsecr, tp->t_badrxtwin)) cc_cong_signal(tp, th, CC_RTO_ERR); } /* * Process options only when we get SYN/ACK back. The SYN case * for incoming connections is handled in tcp_syncache. * According to RFC1323 the window field in a SYN (i.e., a * or ) segment itself is never scaled. * XXX this is traditional behavior, may need to be cleaned up. */ if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { /* Handle parallel SYN for ECN */ if (!(thflags & TH_ACK) && ((thflags & (TH_CWR | TH_ECE)) == (TH_CWR | TH_ECE)) && ((V_tcp_do_ecn == 1) || (V_tcp_do_ecn == 2))) { tp->t_flags2 |= TF2_ECN_PERMIT; tp->t_flags2 |= TF2_ECN_SND_ECE; TCPSTAT_INC(tcps_ecn_shs); } if ((to.to_flags & TOF_SCALE) && (tp->t_flags & TF_REQ_SCALE) && !(tp->t_flags & TF_NOOPT)) { tp->t_flags |= TF_RCVD_SCALE; tp->snd_scale = to.to_wscale; } else tp->t_flags &= ~TF_REQ_SCALE; /* * Initial send window. It will be updated with * the next incoming segment to the scaled value. */ tp->snd_wnd = th->th_win; if ((to.to_flags & TOF_TS) && (tp->t_flags & TF_REQ_TSTMP) && !(tp->t_flags & TF_NOOPT)) { tp->t_flags |= TF_RCVD_TSTMP; tp->ts_recent = to.to_tsval; tp->ts_recent_age = tcp_ts_getticks(); } else tp->t_flags &= ~TF_REQ_TSTMP; if (to.to_flags & TOF_MSS) tcp_mss(tp, to.to_mss); if ((tp->t_flags & TF_SACK_PERMIT) && (!(to.to_flags & TOF_SACKPERM) || (tp->t_flags & TF_NOOPT))) tp->t_flags &= ~TF_SACK_PERMIT; if (IS_FASTOPEN(tp->t_flags)) { if ((to.to_flags & TOF_FASTOPEN) && !(tp->t_flags & TF_NOOPT)) { uint16_t mss; if (to.to_flags & TOF_MSS) mss = to.to_mss; else if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) mss = TCP6_MSS; else mss = TCP_MSS; tcp_fastopen_update_cache(tp, mss, to.to_tfo_len, to.to_tfo_cookie); } else tcp_fastopen_disable_path(tp); } } /* * If timestamps were negotiated during SYN/ACK and a * segment without a timestamp is received, silently drop * the segment, unless it is a RST segment or missing timestamps are * tolerated. * See section 3.2 of RFC 7323. */ if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS)) { if (((thflags & TH_RST) != 0) || V_tcp_tolerate_missing_ts) { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: Timestamp missing, " "segment processed normally\n", s, __func__); free(s, M_TCPLOG); } } else { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: Timestamp missing, " "segment silently dropped\n", s, __func__); free(s, M_TCPLOG); } goto drop; } } /* * If timestamps were not negotiated during SYN/ACK and a * segment with a timestamp is received, ignore the * timestamp and process the packet normally. * See section 3.2 of RFC 7323. */ if (!(tp->t_flags & TF_RCVD_TSTMP) && (to.to_flags & TOF_TS)) { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: Timestamp not expected, " "segment processed normally\n", s, __func__); free(s, M_TCPLOG); } } /* * Header prediction: check for the two common cases * of a uni-directional data xfer. If the packet has * no control flags, is in-sequence, the window didn't * change and we're not retransmitting, it's a * candidate. If the length is zero and the ack moved * forward, we're the sender side of the xfer. Just * free the data acked & wake any higher level process * that was blocked waiting for space. If the length * is non-zero and the ack didn't move, we're the * receiver side. If we're getting packets in-order * (the reassembly queue is empty), add the data to * the socket buffer and note that we need a delayed ack. * Make sure that the hidden state-flags are also off. * Since we check for TCPS_ESTABLISHED first, it can only * be TH_NEEDSYN. */ if (tp->t_state == TCPS_ESTABLISHED && th->th_seq == tp->rcv_nxt && (thflags & (TH_SYN|TH_FIN|TH_RST|TH_URG|TH_ACK)) == TH_ACK && tp->snd_nxt == tp->snd_max && tiwin && tiwin == tp->snd_wnd && ((tp->t_flags & (TF_NEEDSYN|TF_NEEDFIN)) == 0) && SEGQ_EMPTY(tp) && ((to.to_flags & TOF_TS) == 0 || TSTMP_GEQ(to.to_tsval, tp->ts_recent)) ) { /* * If last ACK falls within this segment's sequence numbers, * record the timestamp. * NOTE that the test is modified according to the latest * proposal of the tcplw@cray.com list (Braden 1993/04/26). */ if ((to.to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to.to_tsval; } if (tlen == 0) { if (SEQ_GT(th->th_ack, tp->snd_una) && SEQ_LEQ(th->th_ack, tp->snd_max) && !IN_RECOVERY(tp->t_flags) && (to.to_flags & TOF_SACK) == 0 && TAILQ_EMPTY(&tp->snd_holes)) { /* * This is a pure ack for outstanding data. */ TCPSTAT_INC(tcps_predack); /* * "bad retransmit" recovery without timestamps. */ if ((to.to_flags & TOF_TS) == 0 && tp->t_rxtshift == 1 && tp->t_flags & TF_PREVVALID && (int)(ticks - tp->t_badrxtwin) < 0) { cc_cong_signal(tp, th, CC_RTO_ERR); } /* * Recalculate the transmit timer / rtt. * * Some boxes send broken timestamp replies * during the SYN+ACK phase, ignore * timestamps of 0 or we could calculate a * huge RTT and blow up the retransmit timer. */ if ((to.to_flags & TOF_TS) != 0 && to.to_tsecr) { uint32_t t; t = tcp_ts_getticks() - to.to_tsecr; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; tcp_xmit_timer(tp, TCP_TS_TO_TICKS(t) + 1); } else if (tp->t_rtttime && SEQ_GT(th->th_ack, tp->t_rtseq)) { if (!tp->t_rttlow || tp->t_rttlow > ticks - tp->t_rtttime) tp->t_rttlow = ticks - tp->t_rtttime; tcp_xmit_timer(tp, ticks - tp->t_rtttime); } acked = BYTES_THIS_ACK(tp, th); #ifdef TCP_HHOOK /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ hhook_run_tcp_est_in(tp, th, &to); #endif TCPSTAT_ADD(tcps_rcvackpack, nsegs); TCPSTAT_ADD(tcps_rcvackbyte, acked); sbdrop(&so->so_snd, acked); if (SEQ_GT(tp->snd_una, tp->snd_recover) && SEQ_LEQ(th->th_ack, tp->snd_recover)) tp->snd_recover = th->th_ack - 1; /* * Let the congestion control algorithm update * congestion control related information. This * typically means increasing the congestion * window. */ cc_ack_received(tp, th, nsegs, CC_ACK); tp->snd_una = th->th_ack; /* * Pull snd_wl2 up to prevent seq wrap relative * to th_ack. */ tp->snd_wl2 = th->th_ack; tp->t_dupacks = 0; m_freem(m); /* * If all outstanding data are acked, stop * retransmit timer, otherwise restart timer * using current (possibly backed-off) value. * If process is waiting for space, * wakeup/selwakeup/signal. If data * are ready to send, let tcp_output * decide between more output or persist. */ #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif TCP_PROBE3(debug__input, tp, th, m); if (tp->snd_una == tp->snd_max) tcp_timer_activate(tp, TT_REXMT, 0); else if (!tcp_timer_active(tp, TT_PERSIST)) tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); tp->t_flags |= TF_WAKESOW; if (sbavail(&so->so_snd)) (void) tp->t_fb->tfb_tcp_output(tp); goto check_delack; } } else if (th->th_ack == tp->snd_una && tlen <= sbspace(&so->so_rcv)) { int newsize = 0; /* automatic sockbuf scaling */ /* * This is a pure, in-sequence data packet with * nothing on the reassembly queue and we have enough * buffer space to take it. */ /* Clean receiver SACK report if present */ if ((tp->t_flags & TF_SACK_PERMIT) && tp->rcv_numsacks) tcp_clean_sackreport(tp); TCPSTAT_INC(tcps_preddat); tp->rcv_nxt += tlen; if (tlen && ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && (tp->t_fbyte_in == 0)) { tp->t_fbyte_in = ticks; if (tp->t_fbyte_in == 0) tp->t_fbyte_in = 1; if (tp->t_fbyte_out && tp->t_fbyte_in) tp->t_flags2 |= TF2_FBYTES_COMPLETE; } /* * Pull snd_wl1 up to prevent seq wrap relative to * th_seq. */ tp->snd_wl1 = th->th_seq; /* * Pull rcv_up up to prevent seq wrap relative to * rcv_nxt. */ tp->rcv_up = tp->rcv_nxt; TCPSTAT_ADD(tcps_rcvpack, nsegs); TCPSTAT_ADD(tcps_rcvbyte, tlen); #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif TCP_PROBE3(debug__input, tp, th, m); newsize = tcp_autorcvbuf(m, th, so, tp, tlen); /* Add data to socket buffer. */ SOCKBUF_LOCK(&so->so_rcv); if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { m_freem(m); } else { /* * Set new socket buffer size. * Give up when limit is reached. */ if (newsize) if (!sbreserve_locked(&so->so_rcv, newsize, so, NULL)) so->so_rcv.sb_flags &= ~SB_AUTOSIZE; m_adj(m, drop_hdrlen); /* delayed header drop */ sbappendstream_locked(&so->so_rcv, m, 0); } SOCKBUF_UNLOCK(&so->so_rcv); tp->t_flags |= TF_WAKESOR; if (DELAY_ACK(tp, tlen)) { tp->t_flags |= TF_DELACK; } else { tp->t_flags |= TF_ACKNOW; tp->t_fb->tfb_tcp_output(tp); } goto check_delack; } } /* * Calculate amount of space in receive window, * and then do TCP input processing. * Receive window is amount of space in rcv queue, * but not less than advertised window. */ win = sbspace(&so->so_rcv); if (win < 0) win = 0; tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt)); switch (tp->t_state) { /* * If the state is SYN_RECEIVED: * if seg contains an ACK, but not for our SYN/ACK, send a RST. */ case TCPS_SYN_RECEIVED: if ((thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->snd_una) || SEQ_GT(th->th_ack, tp->snd_max))) { rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } if (IS_FASTOPEN(tp->t_flags)) { /* * When a TFO connection is in SYN_RECEIVED, the * only valid packets are the initial SYN, a * retransmit/copy of the initial SYN (possibly with * a subset of the original data), a valid ACK, a * FIN, or a RST. */ if ((thflags & (TH_SYN|TH_ACK)) == (TH_SYN|TH_ACK)) { rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } else if (thflags & TH_SYN) { /* non-initial SYN is ignored */ if ((tcp_timer_active(tp, TT_DELACK) || tcp_timer_active(tp, TT_REXMT))) goto drop; } else if (!(thflags & (TH_ACK|TH_FIN|TH_RST))) { goto drop; } } break; /* * If the state is SYN_SENT: * if seg contains a RST with valid ACK (SEQ.ACK has already * been verified), then drop the connection. * if seg contains a RST without an ACK, drop the seg. * if seg does not contain SYN, then drop the seg. * Otherwise this is an acceptable SYN segment * initialize tp->rcv_nxt and tp->irs * if seg contains ack then advance tp->snd_una * if seg contains an ECE and ECN support is enabled, the stream * is ECN capable. * if SYN has been acked change to ESTABLISHED else SYN_RCVD state * arrange for segment to be acked (eventually) * continue processing rest of data/controls, beginning with URG */ case TCPS_SYN_SENT: if ((thflags & (TH_ACK|TH_RST)) == (TH_ACK|TH_RST)) { TCP_PROBE5(connect__refused, NULL, tp, m, tp, th); tp = tcp_drop(tp, ECONNREFUSED); } if (thflags & TH_RST) goto drop; if (!(thflags & TH_SYN)) goto drop; tp->irs = th->th_seq; tcp_rcvseqinit(tp); if (thflags & TH_ACK) { int tfo_partial_ack = 0; TCPSTAT_INC(tcps_connects); soisconnected(so); #ifdef MAC mac_socketpeer_set_from_mbuf(m, so); #endif /* Do window scaling on this connection? */ if ((tp->t_flags & (TF_RCVD_SCALE|TF_REQ_SCALE)) == (TF_RCVD_SCALE|TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; } tp->rcv_adv += min(tp->rcv_wnd, TCP_MAXWIN << tp->rcv_scale); tp->snd_una++; /* SYN is acked */ /* * If not all the data that was sent in the TFO SYN * has been acked, resend the remainder right away. */ if (IS_FASTOPEN(tp->t_flags) && (tp->snd_una != tp->snd_max)) { tp->snd_nxt = th->th_ack; tfo_partial_ack = 1; } /* * If there's data, delay ACK; if there's also a FIN * ACKNOW will be turned on later. */ if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial_ack) tcp_timer_activate(tp, TT_DELACK, tcp_delacktime); else tp->t_flags |= TF_ACKNOW; if (((thflags & (TH_CWR | TH_ECE)) == TH_ECE) && (V_tcp_do_ecn == 1)) { tp->t_flags2 |= TF2_ECN_PERMIT; TCPSTAT_INC(tcps_ecn_shs); } /* * Received in SYN_SENT[*] state. * Transitions: * SYN_SENT --> ESTABLISHED * SYN_SENT* --> FIN_WAIT_1 */ tp->t_starttime = ticks; if (tp->t_flags & TF_NEEDFIN) { tcp_state_change(tp, TCPS_FIN_WAIT_1); tp->t_flags &= ~TF_NEEDFIN; thflags &= ~TH_SYN; } else { tcp_state_change(tp, TCPS_ESTABLISHED); TCP_PROBE5(connect__established, NULL, tp, m, tp, th); cc_conn_init(tp); tcp_timer_activate(tp, TT_KEEP, TP_KEEPIDLE(tp)); } } else { /* * Received initial SYN in SYN-SENT[*] state => * simultaneous open. * If it succeeds, connection is * half-synchronized. * Otherwise, do 3-way handshake: * SYN-SENT -> SYN-RECEIVED * SYN-SENT* -> SYN-RECEIVED* */ tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN); tcp_timer_activate(tp, TT_REXMT, 0); tcp_state_change(tp, TCPS_SYN_RECEIVED); } INP_WLOCK_ASSERT(tp->t_inpcb); /* * Advance th->th_seq to correspond to first data byte. * If data, trim to stay within window, * dropping FIN if necessary. */ th->th_seq++; if (tlen > tp->rcv_wnd) { todrop = tlen - tp->rcv_wnd; m_adj(m, -todrop); tlen = tp->rcv_wnd; thflags &= ~TH_FIN; TCPSTAT_INC(tcps_rcvpackafterwin); TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); } tp->snd_wl1 = th->th_seq - 1; tp->rcv_up = th->th_seq; /* * Client side of transaction: already sent SYN and data. * If the remote host used T/TCP to validate the SYN, * our data will be ACK'd; if so, enter normal data segment * processing in the middle of step 5, ack processing. * Otherwise, goto step 6. */ if (thflags & TH_ACK) goto process_ACK; goto step6; /* * If the state is LAST_ACK or CLOSING or TIME_WAIT: * do normal processing. * * NB: Leftover from RFC1644 T/TCP. Cases to be reused later. */ case TCPS_LAST_ACK: case TCPS_CLOSING: break; /* continue normal processing */ } /* * States other than LISTEN or SYN_SENT. * First check the RST flag and sequence number since reset segments * are exempt from the timestamp and connection count tests. This * fixes a bug introduced by the Stevens, vol. 2, p. 960 bugfix * below which allowed reset segments in half the sequence space * to fall though and be processed (which gives forged reset * segments with a random sequence number a 50 percent chance of * killing a connection). * Then check timestamp, if present. * Then check the connection count, if present. * Then check that at least some bytes of segment are within * receive window. If segment begins before rcv_nxt, * drop leading data (and SYN); if nothing left, just ack. */ if (thflags & TH_RST) { /* * RFC5961 Section 3.2 * * - RST drops connection only if SEG.SEQ == RCV.NXT. * - If RST is in window, we send challenge ACK. * * Note: to take into account delayed ACKs, we should * test against last_ack_sent instead of rcv_nxt. * Note 2: we handle special case of closed window, not * covered by the RFC. */ if ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) && SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) || (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq)) { KASSERT(tp->t_state != TCPS_SYN_SENT, ("%s: TH_RST for TCPS_SYN_SENT th %p tp %p", __func__, th, tp)); if (V_tcp_insecure_rst || tp->last_ack_sent == th->th_seq) { TCPSTAT_INC(tcps_drops); /* Drop the connection. */ switch (tp->t_state) { case TCPS_SYN_RECEIVED: so->so_error = ECONNREFUSED; goto close; case TCPS_ESTABLISHED: case TCPS_FIN_WAIT_1: case TCPS_FIN_WAIT_2: case TCPS_CLOSE_WAIT: case TCPS_CLOSING: case TCPS_LAST_ACK: so->so_error = ECONNRESET; close: /* FALLTHROUGH */ default: tp = tcp_close(tp); } } else { TCPSTAT_INC(tcps_badrst); /* Send challenge ACK. */ tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt, tp->snd_nxt, TH_ACK); tp->last_ack_sent = tp->rcv_nxt; m = NULL; } } goto drop; } /* * RFC5961 Section 4.2 * Send challenge ACK for any SYN in synchronized state. */ if ((thflags & TH_SYN) && tp->t_state != TCPS_SYN_SENT && tp->t_state != TCPS_SYN_RECEIVED) { TCPSTAT_INC(tcps_badsyn); if (V_tcp_insecure_syn && SEQ_GEQ(th->th_seq, tp->last_ack_sent) && SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) { tp = tcp_drop(tp, ECONNRESET); rstreason = BANDLIM_UNLIMITED; } else { /* Send challenge ACK. */ tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt, tp->snd_nxt, TH_ACK); tp->last_ack_sent = tp->rcv_nxt; m = NULL; } goto drop; } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment * and it's less than ts_recent, drop it. */ if ((to.to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to.to_tsval, tp->ts_recent)) { /* Check to see if ts_recent is over 24 days old. */ if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { /* * Invalidate ts_recent. If this segment updates * ts_recent, the age will be reset later and ts_recent * will get a valid value. If it does not, setting * ts_recent to zero will at least satisfy the * requirement that zero be placed in the timestamp * echo reply when ts_recent isn't valid. The * age isn't reset until we get a valid ts_recent * because we don't want out-of-order segments to be * dropped when ts_recent is old. */ tp->ts_recent = 0; } else { TCPSTAT_INC(tcps_rcvduppack); TCPSTAT_ADD(tcps_rcvdupbyte, tlen); TCPSTAT_INC(tcps_pawsdrop); if (tlen) goto dropafterack; goto drop; } } /* * In the SYN-RECEIVED state, validate that the packet belongs to * this connection before trimming the data to fit the receive * window. Check the sequence number versus IRS since we know * the sequence numbers haven't wrapped. This is a partial fix * for the "LAND" DoS attack. */ if (tp->t_state == TCPS_SYN_RECEIVED && SEQ_LT(th->th_seq, tp->irs)) { rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } todrop = tp->rcv_nxt - th->th_seq; if (todrop > 0) { if (thflags & TH_SYN) { thflags &= ~TH_SYN; th->th_seq++; if (th->th_urp > 1) th->th_urp--; else thflags &= ~TH_URG; todrop--; } /* * Following if statement from Stevens, vol. 2, p. 960. */ if (todrop > tlen || (todrop == tlen && (thflags & TH_FIN) == 0)) { /* * Any valid FIN must be to the left of the window. * At this point the FIN must be a duplicate or out * of sequence; drop it. */ thflags &= ~TH_FIN; /* * Send an ACK to resynchronize and drop any data. * But keep on processing for RST or ACK. */ tp->t_flags |= TF_ACKNOW; todrop = tlen; TCPSTAT_INC(tcps_rcvduppack); TCPSTAT_ADD(tcps_rcvdupbyte, todrop); } else { TCPSTAT_INC(tcps_rcvpartduppack); TCPSTAT_ADD(tcps_rcvpartdupbyte, todrop); } /* * DSACK - add SACK block for dropped range */ if ((todrop > 0) && (tp->t_flags & TF_SACK_PERMIT)) { tcp_update_sack_list(tp, th->th_seq, th->th_seq + todrop); /* * ACK now, as the next in-sequence segment * will clear the DSACK block again */ tp->t_flags |= TF_ACKNOW; } drop_hdrlen += todrop; /* drop from the top afterwards */ th->th_seq += todrop; tlen -= todrop; if (th->th_urp > todrop) th->th_urp -= todrop; else { thflags &= ~TH_URG; th->th_urp = 0; } } /* * If new data are received on a connection after the * user processes are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tp->t_state > TCPS_CLOSE_WAIT && tlen) { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: %s: Received %d bytes of data " "after socket was closed, " "sending RST and removing tcpcb\n", s, __func__, tcpstates[tp->t_state], tlen); free(s, M_TCPLOG); } tp = tcp_close(tp); TCPSTAT_INC(tcps_rcvafterclose); rstreason = BANDLIM_UNLIMITED; goto dropwithreset; } /* * If segment ends after window, drop trailing data * (and PUSH and FIN); if nothing left, just ACK. */ todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd); if (todrop > 0) { TCPSTAT_INC(tcps_rcvpackafterwin); if (todrop >= tlen) { TCPSTAT_ADD(tcps_rcvbyteafterwin, tlen); /* * If window is closed can only take segments at * window edge, and have to drop data and PUSH from * incoming segments. Continue processing, but * remember to ack. Otherwise, drop segment * and ack. */ if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) { tp->t_flags |= TF_ACKNOW; TCPSTAT_INC(tcps_rcvwinprobe); } else goto dropafterack; } else TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); m_adj(m, -todrop); tlen -= todrop; thflags &= ~(TH_PUSH|TH_FIN); } /* * If last ACK falls within this segment's sequence numbers, * record its timestamp. * NOTE: * 1) That the test incorporates suggestions from the latest * proposal of the tcplw@cray.com list (Braden 1993/04/26). * 2) That updating only on newer timestamps interferes with * our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. * 3) That we modify the segment boundary check to be * Last.ACK.Sent <= SEG.SEQ + SEG.Len * instead of RFC1323's * Last.ACK.Sent < SEG.SEQ + SEG.Len, * This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated * Vol. 2 p.869. In such cases, we can still calculate the * RTT correctly when RCV.NXT == Last.ACK.Sent. */ if ((to.to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN|TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to.to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN * flag is on (half-synchronized state), then queue data for * later processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_state == TCPS_SYN_RECEIVED || (tp->t_flags & TF_NEEDSYN)) { if (tp->t_state == TCPS_SYN_RECEIVED && IS_FASTOPEN(tp->t_flags)) { tp->snd_wnd = tiwin; cc_conn_init(tp); } goto step6; } else if (tp->t_flags & TF_ACKNOW) goto dropafterack; else goto drop; } /* * Ack processing. */ switch (tp->t_state) { /* * In SYN_RECEIVED state, the ack ACKs our SYN, so enter * ESTABLISHED state and continue processing. * The ACK was checked above. */ case TCPS_SYN_RECEIVED: TCPSTAT_INC(tcps_connects); soisconnected(so); /* Do window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE|TF_REQ_SCALE)) == (TF_RCVD_SCALE|TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; } tp->snd_wnd = tiwin; /* * Make transitions: * SYN-RECEIVED -> ESTABLISHED * SYN-RECEIVED* -> FIN-WAIT-1 */ tp->t_starttime = ticks; if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) { tcp_fastopen_decrement_counter(tp->t_tfo_pending); tp->t_tfo_pending = NULL; } if (tp->t_flags & TF_NEEDFIN) { tcp_state_change(tp, TCPS_FIN_WAIT_1); tp->t_flags &= ~TF_NEEDFIN; } else { tcp_state_change(tp, TCPS_ESTABLISHED); TCP_PROBE5(accept__established, NULL, tp, m, tp, th); /* * TFO connections call cc_conn_init() during SYN * processing. Calling it again here for such * connections is not harmless as it would undo the * snd_cwnd reduction that occurs when a TFO SYN|ACK * is retransmitted. */ if (!IS_FASTOPEN(tp->t_flags)) cc_conn_init(tp); tcp_timer_activate(tp, TT_KEEP, TP_KEEPIDLE(tp)); } /* * Account for the ACK of our SYN prior to * regular ACK processing below, except for * simultaneous SYN, which is handled later. */ if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN)) incforsyn = 1; /* * If segment contains data or ACK, will call tcp_reass() * later; if not, do so now to pass queued data to user. */ if (tlen == 0 && (thflags & TH_FIN) == 0) (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0, (struct mbuf *)0); tp->snd_wl1 = th->th_seq - 1; /* FALLTHROUGH */ /* * In ESTABLISHED state: drop duplicate ACKs; ACK out of range * ACKs. If the ack is in the range * tp->snd_una < th->th_ack <= tp->snd_max * then advance tp->snd_una to th->th_ack and drop * data from the retransmission queue. If this ACK reflects * more up to date window information we update our window information. */ case TCPS_ESTABLISHED: case TCPS_FIN_WAIT_1: case TCPS_FIN_WAIT_2: case TCPS_CLOSE_WAIT: case TCPS_CLOSING: case TCPS_LAST_ACK: if (SEQ_GT(th->th_ack, tp->snd_max)) { TCPSTAT_INC(tcps_rcvacktoomuch); goto dropafterack; } if ((tp->t_flags & TF_SACK_PERMIT) && ((to.to_flags & TOF_SACK) || !TAILQ_EMPTY(&tp->snd_holes))) sack_changed = tcp_sack_doack(tp, &to, th->th_ack); else /* * Reset the value so that previous (valid) value * from the last ack with SACK doesn't get used. */ tp->sackhint.sacked_bytes = 0; #ifdef TCP_HHOOK /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ hhook_run_tcp_est_in(tp, th, &to); #endif if (SEQ_LEQ(th->th_ack, tp->snd_una)) { maxseg = tcp_maxseg(tp); if (tlen == 0 && (tiwin == tp->snd_wnd || (tp->t_flags & TF_SACK_PERMIT))) { /* * If this is the first time we've seen a * FIN from the remote, this is not a * duplicate and it needs to be processed * normally. This happens during a * simultaneous close. */ if ((thflags & TH_FIN) && (TCPS_HAVERCVDFIN(tp->t_state) == 0)) { tp->t_dupacks = 0; break; } TCPSTAT_INC(tcps_rcvdupack); /* * If we have outstanding data (other than * a window probe), this is a completely * duplicate ack (ie, window info didn't * change and FIN isn't set), * the ack is the biggest we've * seen and we've seen exactly our rexmt * threshold of them, assume a packet * has been dropped and retransmit it. * Kludge snd_nxt & the congestion * window so we send only this one * packet. * * We know we're losing at the current * window size so do congestion avoidance * (set ssthresh to half the current window * and pull our congestion window back to * the new ssthresh). * * Dup acks mean that packets have left the * network (they're now cached at the receiver) * so bump cwnd by the amount in the receiver * to keep a constant cwnd packets in the * network. * * When using TCP ECN, notify the peer that * we reduced the cwnd. */ /* * Following 2 kinds of acks should not affect * dupack counting: * 1) Old acks * 2) Acks with SACK but without any new SACK * information in them. These could result from * any anomaly in the network like a switch * duplicating packets or a possible DoS attack. */ if (th->th_ack != tp->snd_una || ((tp->t_flags & TF_SACK_PERMIT) && (to.to_flags & TOF_SACK) && !sack_changed)) break; else if (!tcp_timer_active(tp, TT_REXMT)) tp->t_dupacks = 0; else if (++tp->t_dupacks > tcprexmtthresh || IN_FASTRECOVERY(tp->t_flags)) { cc_ack_received(tp, th, nsegs, CC_DUPACK); if (V_tcp_do_prr && IN_FASTRECOVERY(tp->t_flags) && (tp->t_flags & TF_SACK_PERMIT)) { tcp_do_prr_ack(tp, th, &to); } else if ((tp->t_flags & TF_SACK_PERMIT) && (to.to_flags & TOF_SACK) && IN_FASTRECOVERY(tp->t_flags)) { int awnd; /* * Compute the amount of data in flight first. * We can inject new data into the pipe iff * we have less than 1/2 the original window's * worth of data in flight. */ if (V_tcp_do_newsack) awnd = tcp_compute_pipe(tp); else awnd = (tp->snd_nxt - tp->snd_fack) + tp->sackhint.sack_bytes_rexmit; if (awnd < tp->snd_ssthresh) { tp->snd_cwnd += maxseg; if (tp->snd_cwnd > tp->snd_ssthresh) tp->snd_cwnd = tp->snd_ssthresh; } } else tp->snd_cwnd += maxseg; (void) tp->t_fb->tfb_tcp_output(tp); goto drop; } else if (tp->t_dupacks == tcprexmtthresh || (tp->t_flags & TF_SACK_PERMIT && V_tcp_do_newsack && tp->sackhint.sacked_bytes > (tcprexmtthresh - 1) * maxseg)) { enter_recovery: /* * Above is the RFC6675 trigger condition of * more than (dupthresh-1)*maxseg sacked data. * If the count of holes in the * scoreboard is >= dupthresh, we could * also enter loss recovery, but don't * have that value readily available. */ tp->t_dupacks = tcprexmtthresh; tcp_seq onxt = tp->snd_nxt; /* * If we're doing sack, or prr, check * to see if we're already in sack * recovery. If we're not doing sack, * check to see if we're in newreno * recovery. */ if (V_tcp_do_prr || (tp->t_flags & TF_SACK_PERMIT)) { if (IN_FASTRECOVERY(tp->t_flags)) { tp->t_dupacks = 0; break; } } else { if (SEQ_LEQ(th->th_ack, tp->snd_recover)) { tp->t_dupacks = 0; break; } } /* Congestion signal before ack. */ cc_cong_signal(tp, th, CC_NDUPACK); cc_ack_received(tp, th, nsegs, CC_DUPACK); tcp_timer_activate(tp, TT_REXMT, 0); tp->t_rtttime = 0; if (V_tcp_do_prr) { /* * snd_ssthresh is already updated by * cc_cong_signal. */ tp->sackhint.prr_delivered = tp->sackhint.sacked_bytes; tp->sackhint.recover_fs = max(1, tp->snd_nxt - tp->snd_una); } if ((tp->t_flags & TF_SACK_PERMIT) && (to.to_flags & TOF_SACK)) { TCPSTAT_INC( tcps_sack_recovery_episode); tp->snd_recover = tp->snd_nxt; tp->snd_cwnd = maxseg; (void) tp->t_fb->tfb_tcp_output(tp); if (SEQ_GT(th->th_ack, tp->snd_una)) goto resume_partialack; goto drop; } tp->snd_nxt = th->th_ack; tp->snd_cwnd = maxseg; (void) tp->t_fb->tfb_tcp_output(tp); KASSERT(tp->snd_limited <= 2, ("%s: tp->snd_limited too big", __func__)); tp->snd_cwnd = tp->snd_ssthresh + maxseg * (tp->t_dupacks - tp->snd_limited); if (SEQ_GT(onxt, tp->snd_nxt)) tp->snd_nxt = onxt; goto drop; } else if (V_tcp_do_rfc3042) { /* * Process first and second duplicate * ACKs. Each indicates a segment * leaving the network, creating room * for more. Make sure we can send a * packet on reception of each duplicate * ACK by increasing snd_cwnd by one * segment. Restore the original * snd_cwnd after packet transmission. */ cc_ack_received(tp, th, nsegs, CC_DUPACK); uint32_t oldcwnd = tp->snd_cwnd; tcp_seq oldsndmax = tp->snd_max; u_int sent; int avail; KASSERT(tp->t_dupacks == 1 || tp->t_dupacks == 2, ("%s: dupacks not 1 or 2", __func__)); if (tp->t_dupacks == 1) tp->snd_limited = 0; tp->snd_cwnd = (tp->snd_nxt - tp->snd_una) + (tp->t_dupacks - tp->snd_limited) * maxseg; /* * Only call tcp_output when there * is new data available to be sent. * Otherwise we would send pure ACKs. */ SOCKBUF_LOCK(&so->so_snd); avail = sbavail(&so->so_snd) - (tp->snd_nxt - tp->snd_una); SOCKBUF_UNLOCK(&so->so_snd); if (avail > 0) (void) tp->t_fb->tfb_tcp_output(tp); sent = tp->snd_max - oldsndmax; if (sent > maxseg) { KASSERT((tp->t_dupacks == 2 && tp->snd_limited == 0) || (sent == maxseg + 1 && tp->t_flags & TF_SENTFIN), ("%s: sent too much", __func__)); tp->snd_limited = 2; } else if (sent > 0) ++tp->snd_limited; tp->snd_cwnd = oldcwnd; goto drop; } } break; } else { /* * This ack is advancing the left edge, reset the * counter. */ tp->t_dupacks = 0; /* * If this ack also has new SACK info, increment the * counter as per rfc6675. The variable * sack_changed tracks all changes to the SACK * scoreboard, including when partial ACKs without * SACK options are received, and clear the scoreboard * from the left side. Such partial ACKs should not be * counted as dupacks here. */ if ((tp->t_flags & TF_SACK_PERMIT) && (to.to_flags & TOF_SACK) && sack_changed) { tp->t_dupacks++; /* limit overhead by setting maxseg last */ if (!IN_FASTRECOVERY(tp->t_flags) && (tp->sackhint.sacked_bytes > ((tcprexmtthresh - 1) * (maxseg = tcp_maxseg(tp))))) { goto enter_recovery; } } } resume_partialack: KASSERT(SEQ_GT(th->th_ack, tp->snd_una), ("%s: th_ack <= snd_una", __func__)); /* * If the congestion window was inflated to account * for the other side's cached packets, retract it. */ if (IN_FASTRECOVERY(tp->t_flags)) { if (SEQ_LT(th->th_ack, tp->snd_recover)) { if (tp->t_flags & TF_SACK_PERMIT) if (V_tcp_do_prr && to.to_flags & TOF_SACK) { tcp_timer_activate(tp, TT_REXMT, 0); tp->t_rtttime = 0; tcp_do_prr_ack(tp, th, &to); tp->t_flags |= TF_ACKNOW; (void) tcp_output(tp); } else tcp_sack_partialack(tp, th); else tcp_newreno_partial_ack(tp, th); } else cc_post_recovery(tp, th); } else if (IN_CONGRECOVERY(tp->t_flags)) { if (SEQ_LT(th->th_ack, tp->snd_recover)) { if (V_tcp_do_prr) { tp->sackhint.delivered_data = BYTES_THIS_ACK(tp, th); tp->snd_fack = th->th_ack; tcp_do_prr_ack(tp, th, &to); (void) tcp_output(tp); } } else cc_post_recovery(tp, th); } /* * If we reach this point, ACK is not a duplicate, * i.e., it ACKs something we sent. */ if (tp->t_flags & TF_NEEDSYN) { /* * T/TCP: Connection was half-synchronized, and our * SYN has been ACK'd (so connection is now fully * synchronized). Go to non-starred state, * increment snd_una for ACK of SYN, and check if * we can do window scaling. */ tp->t_flags &= ~TF_NEEDSYN; tp->snd_una++; /* Do window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE|TF_REQ_SCALE)) == (TF_RCVD_SCALE|TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; /* Send window already scaled. */ } } process_ACK: INP_WLOCK_ASSERT(tp->t_inpcb); /* * Adjust for the SYN bit in sequence space, * but don't account for it in cwnd calculations. * This is for the SYN_RECEIVED, non-simultaneous * SYN case. SYN_SENT and simultaneous SYN are * treated elsewhere. */ if (incforsyn) tp->snd_una++; acked = BYTES_THIS_ACK(tp, th); KASSERT(acked >= 0, ("%s: acked unexepectedly negative " "(tp->snd_una=%u, th->th_ack=%u, tp=%p, m=%p)", __func__, tp->snd_una, th->th_ack, tp, m)); TCPSTAT_ADD(tcps_rcvackpack, nsegs); TCPSTAT_ADD(tcps_rcvackbyte, acked); /* * If we just performed our first retransmit, and the ACK * arrives within our recovery window, then it was a mistake * to do the retransmit in the first place. Recover our * original cwnd and ssthresh, and proceed to transmit where * we left off. */ if (tp->t_rxtshift == 1 && tp->t_flags & TF_PREVVALID && tp->t_badrxtwin && SEQ_LT(to.to_tsecr, tp->t_badrxtwin)) cc_cong_signal(tp, th, CC_RTO_ERR); /* * If we have a timestamp reply, update smoothed * round trip time. If no timestamp is present but * transmit timer is running and timed sequence * number was acked, update smoothed round trip time. * Since we now have an rtt measurement, cancel the * timer backoff (cf., Phil Karn's retransmit alg.). * Recompute the initial retransmit timer. * * Some boxes send broken timestamp replies * during the SYN+ACK phase, ignore * timestamps of 0 or we could calculate a * huge RTT and blow up the retransmit timer. */ if ((to.to_flags & TOF_TS) != 0 && to.to_tsecr) { uint32_t t; t = tcp_ts_getticks() - to.to_tsecr; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; tcp_xmit_timer(tp, TCP_TS_TO_TICKS(t) + 1); } else if (tp->t_rtttime && SEQ_GT(th->th_ack, tp->t_rtseq)) { if (!tp->t_rttlow || tp->t_rttlow > ticks - tp->t_rtttime) tp->t_rttlow = ticks - tp->t_rtttime; tcp_xmit_timer(tp, ticks - tp->t_rtttime); } /* * If all outstanding data is acked, stop retransmit * timer and remember to restart (more output or persist). * If there is more data to be acked, restart retransmit * timer, using current (possibly backed-off) value. */ if (th->th_ack == tp->snd_max) { tcp_timer_activate(tp, TT_REXMT, 0); needoutput = 1; } else if (!tcp_timer_active(tp, TT_PERSIST)) tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); /* * If no data (only SYN) was ACK'd, * skip rest of ACK processing. */ if (acked == 0) goto step6; /* * Let the congestion control algorithm update congestion * control related information. This typically means increasing * the congestion window. */ cc_ack_received(tp, th, nsegs, CC_ACK); SOCKBUF_LOCK(&so->so_snd); if (acked > sbavail(&so->so_snd)) { if (tp->snd_wnd >= sbavail(&so->so_snd)) tp->snd_wnd -= sbavail(&so->so_snd); else tp->snd_wnd = 0; mfree = sbcut_locked(&so->so_snd, (int)sbavail(&so->so_snd)); ourfinisacked = 1; } else { mfree = sbcut_locked(&so->so_snd, acked); if (tp->snd_wnd >= (uint32_t) acked) tp->snd_wnd -= acked; else tp->snd_wnd = 0; ourfinisacked = 0; } SOCKBUF_UNLOCK(&so->so_snd); tp->t_flags |= TF_WAKESOW; m_freem(mfree); /* Detect una wraparound. */ if (!IN_RECOVERY(tp->t_flags) && SEQ_GT(tp->snd_una, tp->snd_recover) && SEQ_LEQ(th->th_ack, tp->snd_recover)) tp->snd_recover = th->th_ack - 1; /* XXXLAS: Can this be moved up into cc_post_recovery? */ if (IN_RECOVERY(tp->t_flags) && SEQ_GEQ(th->th_ack, tp->snd_recover)) { EXIT_RECOVERY(tp->t_flags); } tp->snd_una = th->th_ack; if (tp->t_flags & TF_SACK_PERMIT) { if (SEQ_GT(tp->snd_una, tp->snd_recover)) tp->snd_recover = tp->snd_una; } if (SEQ_LT(tp->snd_nxt, tp->snd_una)) tp->snd_nxt = tp->snd_una; switch (tp->t_state) { /* * In FIN_WAIT_1 STATE in addition to the processing * for the ESTABLISHED state if our FIN is now acknowledged * then enter FIN_WAIT_2. */ case TCPS_FIN_WAIT_1: if (ourfinisacked) { /* * If we can't receive any more * data, then closing user can proceed. * Starting the timer is contrary to the * specification, but if we don't get a FIN * we'll hang forever. * * XXXjl: * we should release the tp also, and use a * compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } break; /* * In CLOSING STATE in addition to the processing for * the ESTABLISHED state if the ACK acknowledges our FIN * then enter the TIME-WAIT state, otherwise ignore * the segment. */ case TCPS_CLOSING: if (ourfinisacked) { tcp_twstart(tp); m_freem(m); return; } break; /* * In LAST_ACK, we may still be waiting for data to drain * and/or to be acked, as well as for the ack of our FIN. * If our FIN is now acknowledged, delete the TCB, * enter the closed state and return. */ case TCPS_LAST_ACK: if (ourfinisacked) { tp = tcp_close(tp); goto drop; } break; } } step6: INP_WLOCK_ASSERT(tp->t_inpcb); /* * Update window information. * Don't look at window if no ACK: TAC's send garbage on first SYN. */ if ((thflags & TH_ACK) && (SEQ_LT(tp->snd_wl1, th->th_seq) || (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { /* keep track of pure window updates */ if (tlen == 0 && tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) TCPSTAT_INC(tcps_rcvwinupd); tp->snd_wnd = tiwin; tp->snd_wl1 = th->th_seq; tp->snd_wl2 = th->th_ack; if (tp->snd_wnd > tp->max_sndwnd) tp->max_sndwnd = tp->snd_wnd; needoutput = 1; } /* * Process segments with URG. */ if ((thflags & TH_URG) && th->th_urp && TCPS_HAVERCVDFIN(tp->t_state) == 0) { /* * This is a kludge, but if we receive and accept * random urgent pointers, we'll crash in * soreceive. It's hard to imagine someone * actually wanting to send this much urgent data. */ SOCKBUF_LOCK(&so->so_rcv); if (th->th_urp + sbavail(&so->so_rcv) > sb_max) { th->th_urp = 0; /* XXX */ thflags &= ~TH_URG; /* XXX */ SOCKBUF_UNLOCK(&so->so_rcv); /* XXX */ goto dodata; /* XXX */ } /* * If this segment advances the known urgent pointer, * then mark the data stream. This should not happen * in CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since * a FIN has been received from the remote side. * In these states we ignore the URG. * * According to RFC961 (Assigned Protocols), * the urgent pointer points to the last octet * of urgent data. We continue, however, * to consider it to indicate the first octet * of data past the urgent section as the original * spec states (in one of two places). */ if (SEQ_GT(th->th_seq+th->th_urp, tp->rcv_up)) { tp->rcv_up = th->th_seq + th->th_urp; so->so_oobmark = sbavail(&so->so_rcv) + (tp->rcv_up - tp->rcv_nxt) - 1; if (so->so_oobmark == 0) so->so_rcv.sb_state |= SBS_RCVATMARK; sohasoutofband(so); tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA); } SOCKBUF_UNLOCK(&so->so_rcv); /* * Remove out of band data so doesn't get presented to user. * This can happen independent of advancing the URG pointer, * but if two URG's are pending at once, some out-of-band * data may creep in... ick. */ if (th->th_urp <= (uint32_t)tlen && !(so->so_options & SO_OOBINLINE)) { /* hdr drop is delayed */ tcp_pulloutofband(so, th, m, drop_hdrlen); } } else { /* * If no out of band data is expected, * pull receive urgent pointer along * with the receive window. */ if (SEQ_GT(tp->rcv_nxt, tp->rcv_up)) tp->rcv_up = tp->rcv_nxt; } dodata: /* XXX */ INP_WLOCK_ASSERT(tp->t_inpcb); /* * Process the segment text, merging it into the TCP sequencing queue, * and arranging for acknowledgment of receipt if necessary. * This process logically involves adjusting tp->rcv_wnd as data * is presented to the user (this happens in tcp_usrreq.c, * case PRU_RCVD). If a FIN has already been received on this * connection then we just ignore the text. */ tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && IS_FASTOPEN(tp->t_flags)); if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) && TCPS_HAVERCVDFIN(tp->t_state) == 0) { tcp_seq save_start = th->th_seq; tcp_seq save_rnxt = tp->rcv_nxt; int save_tlen = tlen; m_adj(m, drop_hdrlen); /* delayed header drop */ /* * Insert segment which includes th into TCP reassembly queue * with control block tp. Set thflags to whether reassembly now * includes a segment with FIN. This handles the common case * inline (segment is the next to be received on an established * connection, and the queue is empty), avoiding linkage into * and removal from the queue and repetition of various * conversions. * Set DELACK for segments received in order, but ack * immediately when segments are out of order (so * fast retransmit can work). */ if (th->th_seq == tp->rcv_nxt && SEGQ_EMPTY(tp) && (TCPS_HAVEESTABLISHED(tp->t_state) || tfo_syn)) { if (DELAY_ACK(tp, tlen) || tfo_syn) tp->t_flags |= TF_DELACK; else tp->t_flags |= TF_ACKNOW; tp->rcv_nxt += tlen; if (tlen && ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && (tp->t_fbyte_in == 0)) { tp->t_fbyte_in = ticks; if (tp->t_fbyte_in == 0) tp->t_fbyte_in = 1; if (tp->t_fbyte_out && tp->t_fbyte_in) tp->t_flags2 |= TF2_FBYTES_COMPLETE; } thflags = th->th_flags & TH_FIN; TCPSTAT_INC(tcps_rcvpack); TCPSTAT_ADD(tcps_rcvbyte, tlen); SOCKBUF_LOCK(&so->so_rcv); if (so->so_rcv.sb_state & SBS_CANTRCVMORE) m_freem(m); else sbappendstream_locked(&so->so_rcv, m, 0); SOCKBUF_UNLOCK(&so->so_rcv); tp->t_flags |= TF_WAKESOR; } else { /* * XXX: Due to the header drop above "th" is * theoretically invalid by now. Fortunately * m_adj() doesn't actually frees any mbufs * when trimming from the head. */ tcp_seq temp = save_start; thflags = tcp_reass(tp, th, &temp, &tlen, m); tp->t_flags |= TF_ACKNOW; } if ((tp->t_flags & TF_SACK_PERMIT) && (save_tlen > 0) && TCPS_HAVEESTABLISHED(tp->t_state)) { if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) { /* * DSACK actually handled in the fastpath * above. */ tcp_update_sack_list(tp, save_start, save_start + save_tlen); } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) { if ((tp->rcv_numsacks >= 1) && (tp->sackblks[0].end == save_start)) { /* * Partial overlap, recorded at todrop * above. */ tcp_update_sack_list(tp, tp->sackblks[0].start, tp->sackblks[0].end); } else { tcp_update_dsack_list(tp, save_start, save_start + save_tlen); } } else if (tlen >= save_tlen) { /* Update of sackblks. */ tcp_update_dsack_list(tp, save_start, save_start + save_tlen); } else if (tlen > 0) { tcp_update_dsack_list(tp, save_start, save_start + tlen); } } #if 0 /* * Note the amount of data that peer has sent into * our window, in order to estimate the sender's * buffer size. * XXX: Unused. */ if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) len = so->so_rcv.sb_hiwat - (tp->rcv_adv - tp->rcv_nxt); else len = so->so_rcv.sb_hiwat; #endif } else { m_freem(m); thflags &= ~TH_FIN; } /* * If FIN is received ACK the FIN and let the user know * that the connection is closing. */ if (thflags & TH_FIN) { if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { socantrcvmore(so); /* The socket upcall is handled by socantrcvmore. */ tp->t_flags &= ~TF_WAKESOR; /* * If connection is half-synchronized * (ie NEEDSYN flag on) then delay ACK, * so it may be piggybacked when SYN is sent. * Otherwise, since we received a FIN then no * more input can be expected, send ACK now. */ if (tp->t_flags & TF_NEEDSYN) tp->t_flags |= TF_DELACK; else tp->t_flags |= TF_ACKNOW; tp->rcv_nxt++; } switch (tp->t_state) { /* * In SYN_RECEIVED and ESTABLISHED STATES * enter the CLOSE_WAIT state. */ case TCPS_SYN_RECEIVED: tp->t_starttime = ticks; /* FALLTHROUGH */ case TCPS_ESTABLISHED: tcp_state_change(tp, TCPS_CLOSE_WAIT); break; /* * If still in FIN_WAIT_1 STATE FIN has not been acked so * enter the CLOSING state. */ case TCPS_FIN_WAIT_1: tcp_state_change(tp, TCPS_CLOSING); break; /* * In FIN_WAIT_2 state enter the TIME_WAIT state, * starting the time-wait timer, turning off the other * standard timers. */ case TCPS_FIN_WAIT_2: tcp_twstart(tp); return; } } #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif TCP_PROBE3(debug__input, tp, th, m); /* * Return any desired output. */ if (needoutput || (tp->t_flags & TF_ACKNOW)) (void) tp->t_fb->tfb_tcp_output(tp); check_delack: INP_WLOCK_ASSERT(tp->t_inpcb); if (tp->t_flags & TF_DELACK) { tp->t_flags &= ~TF_DELACK; tcp_timer_activate(tp, TT_DELACK, tcp_delacktime); } tcp_handle_wakeup(tp, so); INP_WUNLOCK(tp->t_inpcb); return; dropafterack: /* * Generate an ACK dropping incoming segment if it occupies * sequence space, where the ACK reflects our state. * * We can now skip the test for the RST flag since all * paths to this code happen after packets containing * RST have been dropped. * * In the SYN-RECEIVED state, don't send an ACK unless the * segment we received passes the SYN-RECEIVED ACK test. * If it fails send a RST. This breaks the loop in the * "LAND" DoS attack, and also prevents an ACK storm * between two listening ports that have been sent forged * SYN segments, each with the source address of the other. */ if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) && (SEQ_GT(tp->snd_una, th->th_ack) || SEQ_GT(th->th_ack, tp->snd_max)) ) { rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_DROP, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif TCP_PROBE3(debug__input, tp, th, m); tp->t_flags |= TF_ACKNOW; (void) tp->t_fb->tfb_tcp_output(tp); tcp_handle_wakeup(tp, so); INP_WUNLOCK(tp->t_inpcb); m_freem(m); return; dropwithreset: if (tp != NULL) { tcp_dropwithreset(m, th, tp, tlen, rstreason); tcp_handle_wakeup(tp, so); INP_WUNLOCK(tp->t_inpcb); } else tcp_dropwithreset(m, th, NULL, tlen, rstreason); return; drop: /* * Drop space held by incoming segment and return. */ #ifdef TCPDEBUG if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) tcp_trace(TA_DROP, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif TCP_PROBE3(debug__input, tp, th, m); if (tp != NULL) { tcp_handle_wakeup(tp, so); INP_WUNLOCK(tp->t_inpcb); } m_freem(m); } /* * Issue RST and make ACK acceptable to originator of segment. * The mbuf must still include the original packet header. * tp may be NULL. */ void tcp_dropwithreset(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int tlen, int rstreason) { #ifdef INET struct ip *ip; #endif #ifdef INET6 struct ip6_hdr *ip6; #endif if (tp != NULL) { INP_LOCK_ASSERT(tp->t_inpcb); } /* Don't bother if destination was broadcast/multicast. */ if ((th->th_flags & TH_RST) || m->m_flags & (M_BCAST|M_MCAST)) goto drop; #ifdef INET6 if (mtod(m, struct ip *)->ip_v == 6) { ip6 = mtod(m, struct ip6_hdr *); if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) || IN6_IS_ADDR_MULTICAST(&ip6->ip6_src)) goto drop; /* IPv6 anycast check is done at tcp6_input() */ } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { ip = mtod(m, struct ip *); if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) || IN_MULTICAST(ntohl(ip->ip_src.s_addr)) || ip->ip_src.s_addr == htonl(INADDR_BROADCAST) || in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif)) goto drop; } #endif /* Perform bandwidth limiting. */ if (badport_bandlim(rstreason) < 0) goto drop; /* tcp_respond consumes the mbuf chain. */ if (th->th_flags & TH_ACK) { tcp_respond(tp, mtod(m, void *), th, m, (tcp_seq)0, th->th_ack, TH_RST); } else { if (th->th_flags & TH_SYN) tlen++; if (th->th_flags & TH_FIN) tlen++; tcp_respond(tp, mtod(m, void *), th, m, th->th_seq+tlen, (tcp_seq)0, TH_RST|TH_ACK); } return; drop: m_freem(m); } /* * Parse TCP options and place in tcpopt. */ void tcp_dooptions(struct tcpopt *to, u_char *cp, int cnt, int flags) { int opt, optlen; to->to_flags = 0; for (; cnt > 0; cnt -= optlen, cp += optlen) { opt = cp[0]; if (opt == TCPOPT_EOL) break; if (opt == TCPOPT_NOP) optlen = 1; else { if (cnt < 2) break; optlen = cp[1]; if (optlen < 2 || optlen > cnt) break; } switch (opt) { case TCPOPT_MAXSEG: if (optlen != TCPOLEN_MAXSEG) continue; if (!(flags & TO_SYN)) continue; to->to_flags |= TOF_MSS; bcopy((char *)cp + 2, (char *)&to->to_mss, sizeof(to->to_mss)); to->to_mss = ntohs(to->to_mss); break; case TCPOPT_WINDOW: if (optlen != TCPOLEN_WINDOW) continue; if (!(flags & TO_SYN)) continue; to->to_flags |= TOF_SCALE; to->to_wscale = min(cp[2], TCP_MAX_WINSHIFT); break; case TCPOPT_TIMESTAMP: if (optlen != TCPOLEN_TIMESTAMP) continue; to->to_flags |= TOF_TS; bcopy((char *)cp + 2, (char *)&to->to_tsval, sizeof(to->to_tsval)); to->to_tsval = ntohl(to->to_tsval); bcopy((char *)cp + 6, (char *)&to->to_tsecr, sizeof(to->to_tsecr)); to->to_tsecr = ntohl(to->to_tsecr); break; case TCPOPT_SIGNATURE: /* * In order to reply to a host which has set the * TCP_SIGNATURE option in its initial SYN, we have * to record the fact that the option was observed * here for the syncache code to perform the correct * response. */ if (optlen != TCPOLEN_SIGNATURE) continue; to->to_flags |= TOF_SIGNATURE; to->to_signature = cp + 2; break; case TCPOPT_SACK_PERMITTED: if (optlen != TCPOLEN_SACK_PERMITTED) continue; if (!(flags & TO_SYN)) continue; if (!V_tcp_do_sack) continue; to->to_flags |= TOF_SACKPERM; break; case TCPOPT_SACK: if (optlen <= 2 || (optlen - 2) % TCPOLEN_SACK != 0) continue; if (flags & TO_SYN) continue; to->to_flags |= TOF_SACK; to->to_nsacks = (optlen - 2) / TCPOLEN_SACK; to->to_sacks = cp + 2; TCPSTAT_INC(tcps_sack_rcv_blocks); break; case TCPOPT_FAST_OPEN: /* * Cookie length validation is performed by the * server side cookie checking code or the client * side cookie cache update code. */ if (!(flags & TO_SYN)) continue; if (!V_tcp_fastopen_client_enable && !V_tcp_fastopen_server_enable) continue; to->to_flags |= TOF_FASTOPEN; to->to_tfo_len = optlen - 2; to->to_tfo_cookie = to->to_tfo_len ? cp + 2 : NULL; break; default: continue; } } } /* * Pull out of band byte out of a segment so * it doesn't appear in the user's data queue. * It is still reflected in the segment length for * sequencing purposes. */ void tcp_pulloutofband(struct socket *so, struct tcphdr *th, struct mbuf *m, int off) { int cnt = off + th->th_urp - 1; while (cnt >= 0) { if (m->m_len > cnt) { char *cp = mtod(m, caddr_t) + cnt; struct tcpcb *tp = sototcpcb(so); INP_WLOCK_ASSERT(tp->t_inpcb); tp->t_iobc = *cp; tp->t_oobflags |= TCPOOB_HAVEDATA; bcopy(cp+1, cp, (unsigned)(m->m_len - cnt - 1)); m->m_len--; if (m->m_flags & M_PKTHDR) m->m_pkthdr.len--; return; } cnt -= m->m_len; m = m->m_next; if (m == NULL) break; } panic("tcp_pulloutofband"); } /* * Collect new round-trip time estimate * and update averages and current timeout. */ void tcp_xmit_timer(struct tcpcb *tp, int rtt) { int delta; INP_WLOCK_ASSERT(tp->t_inpcb); TCPSTAT_INC(tcps_rttupdated); tp->t_rttupdated++; #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt * 1000 / hz)); #endif if ((tp->t_srtt != 0) && (tp->t_rxtshift <= TCP_RTT_INVALIDATE)) { /* * srtt is stored as fixed point with 5 bits after the * binary point (i.e., scaled by 8). The following magic * is equivalent to the smoothing algorithm in rfc793 with * an alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed * point). Adjust rtt to origin 0. */ delta = ((rtt - 1) << TCP_DELTA_SHIFT) - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); if ((tp->t_srtt += delta) <= 0) tp->t_srtt = 1; /* * We accumulate a smoothed rtt variance (actually, a * smoothed mean difference), then set the retransmit * timer to smoothed rtt + 4 times the smoothed variance. * rttvar is stored as fixed point with 4 bits after the * binary point (scaled by 16). The following is * equivalent to rfc793 smoothing with an alpha of .75 * (rttvar = rttvar*3/4 + |delta| / 4). This replaces * rfc793's wired-in beta. */ if (delta < 0) delta = -delta; delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); if ((tp->t_rttvar += delta) <= 0) tp->t_rttvar = 1; if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar) tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } else { /* * No rtt measurement yet - use the unsmoothed rtt. * Set the variance to half the rtt (so our first * retransmit happens at 3*rtt). */ tp->t_srtt = rtt << TCP_RTT_SHIFT; tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1); tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } tp->t_rtttime = 0; tp->t_rxtshift = 0; /* * the retransmit should happen at rtt + 4 * rttvar. * Because of the way we do the smoothing, srtt and rttvar * will each average +1/2 tick of bias. When we compute * the retransmit timer, we want 1/2 tick of rounding and * 1 extra tick because of +-1/2 tick uncertainty in the * firing of the timer. The bias will give us exactly the * 1.5 tick we need. But, because the bias is * statistical, we have to test that we don't drop below * the minimum feasible timer (which is 2 ticks). */ TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), max(tp->t_rttmin, rtt + 2), TCPTV_REXMTMAX); /* * We received an ack for a packet that wasn't retransmitted; * it is probably safe to discard any error indications we've * received recently. This isn't quite right, but close enough * for now (a route might have failed after we sent a segment, * and the return path might not be symmetrical). */ tp->t_softerror = 0; } /* * Determine a reasonable value for maxseg size. * If the route is known, check route for mtu. * If none, use an mss that can be handled on the outgoing interface * without forcing IP to fragment. If no route is found, route has no mtu, * or the destination isn't local, use a default, hopefully conservative * size (usually 512 or the default IP max size, but no more than the mtu * of the interface), as we can't discover anything about intervening * gateways or networks. We also initialize the congestion/slow start * window to be a single segment if the destination isn't local. * While looking at the routing entry, we also initialize other path-dependent * parameters from pre-set or cached values in the routing entry. * * NOTE that resulting t_maxseg doesn't include space for TCP options or * IP options, e.g. IPSEC data, since length of this data may vary, and * thus it is calculated for every segment separately in tcp_output(). * * NOTE that this routine is only called when we process an incoming * segment, or an ICMP need fragmentation datagram. Outgoing SYN/ACK MSS * settings are handled in tcp_mssopt(). */ void tcp_mss_update(struct tcpcb *tp, int offer, int mtuoffer, struct hc_metrics_lite *metricptr, struct tcp_ifcap *cap) { int mss = 0; uint32_t maxmtu = 0; struct inpcb *inp = tp->t_inpcb; struct hc_metrics_lite metrics; #ifdef INET6 int isipv6 = ((inp->inp_vflag & INP_IPV6) != 0) ? 1 : 0; size_t min_protoh = isipv6 ? sizeof (struct ip6_hdr) + sizeof (struct tcphdr) : sizeof (struct tcpiphdr); #else - const size_t min_protoh = sizeof(struct tcpiphdr); + size_t min_protoh = sizeof(struct tcpiphdr); #endif INP_WLOCK_ASSERT(tp->t_inpcb); + if (tp->t_port) + min_protoh += V_tcp_udp_tunneling_overhead; if (mtuoffer != -1) { KASSERT(offer == -1, ("%s: conflict", __func__)); offer = mtuoffer - min_protoh; } /* Initialize. */ #ifdef INET6 if (isipv6) { maxmtu = tcp_maxmtu6(&inp->inp_inc, cap); tp->t_maxseg = V_tcp_v6mssdflt; } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { maxmtu = tcp_maxmtu(&inp->inp_inc, cap); tp->t_maxseg = V_tcp_mssdflt; } #endif /* * No route to sender, stay with default mss and return. */ if (maxmtu == 0) { /* * In case we return early we need to initialize metrics * to a defined state as tcp_hc_get() would do for us * if there was no cache hit. */ if (metricptr != NULL) bzero(metricptr, sizeof(struct hc_metrics_lite)); return; } /* What have we got? */ switch (offer) { case 0: /* * Offer == 0 means that there was no MSS on the SYN * segment, in this case we use tcp_mssdflt as * already assigned to t_maxseg above. */ offer = tp->t_maxseg; break; case -1: /* * Offer == -1 means that we didn't receive SYN yet. */ /* FALLTHROUGH */ default: /* * Prevent DoS attack with too small MSS. Round up * to at least minmss. */ offer = max(offer, V_tcp_minmss); } /* * rmx information is now retrieved from tcp_hostcache. */ tcp_hc_get(&inp->inp_inc, &metrics); if (metricptr != NULL) bcopy(&metrics, metricptr, sizeof(struct hc_metrics_lite)); /* * If there's a discovered mtu in tcp hostcache, use it. * Else, use the link mtu. */ if (metrics.rmx_mtu) mss = min(metrics.rmx_mtu, maxmtu) - min_protoh; else { #ifdef INET6 if (isipv6) { mss = maxmtu - min_protoh; if (!V_path_mtu_discovery && !in6_localaddr(&inp->in6p_faddr)) mss = min(mss, V_tcp_v6mssdflt); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { mss = maxmtu - min_protoh; if (!V_path_mtu_discovery && !in_localaddr(inp->inp_faddr)) mss = min(mss, V_tcp_mssdflt); } #endif /* * XXX - The above conditional (mss = maxmtu - min_protoh) * probably violates the TCP spec. * The problem is that, since we don't know the * other end's MSS, we are supposed to use a conservative * default. But, if we do that, then MTU discovery will * never actually take place, because the conservative * default is much less than the MTUs typically seen * on the Internet today. For the moment, we'll sweep * this under the carpet. * * The conservative default might not actually be a problem * if the only case this occurs is when sending an initial * SYN with options and data to a host we've never talked * to before. Then, they will reply with an MSS value which * will get recorded and the new parameters should get * recomputed. For Further Study. */ } mss = min(mss, offer); /* * Sanity check: make sure that maxseg will be large * enough to allow some data on segments even if the * all the option space is used (40bytes). Otherwise * funny things may happen in tcp_output. * * XXXGL: shouldn't we reserve space for IP/IPv6 options? */ mss = max(mss, 64); tp->t_maxseg = mss; } void tcp_mss(struct tcpcb *tp, int offer) { int mss; uint32_t bufsize; struct inpcb *inp; struct socket *so; struct hc_metrics_lite metrics; struct tcp_ifcap cap; KASSERT(tp != NULL, ("%s: tp == NULL", __func__)); bzero(&cap, sizeof(cap)); tcp_mss_update(tp, offer, -1, &metrics, &cap); mss = tp->t_maxseg; inp = tp->t_inpcb; /* * If there's a pipesize, change the socket buffer to that size, * don't change if sb_hiwat is different than default (then it * has been changed on purpose with setsockopt). * Make the socket buffers an integral number of mss units; * if the mss is larger than the socket buffer, decrease the mss. */ so = inp->inp_socket; SOCKBUF_LOCK(&so->so_snd); if ((so->so_snd.sb_hiwat == V_tcp_sendspace) && metrics.rmx_sendpipe) bufsize = metrics.rmx_sendpipe; else bufsize = so->so_snd.sb_hiwat; if (bufsize < mss) mss = bufsize; else { bufsize = roundup(bufsize, mss); if (bufsize > sb_max) bufsize = sb_max; if (bufsize > so->so_snd.sb_hiwat) (void)sbreserve_locked(&so->so_snd, bufsize, so, NULL); } SOCKBUF_UNLOCK(&so->so_snd); /* * Sanity check: make sure that maxseg will be large * enough to allow some data on segments even if the * all the option space is used (40bytes). Otherwise * funny things may happen in tcp_output. * * XXXGL: shouldn't we reserve space for IP/IPv6 options? */ tp->t_maxseg = max(mss, 64); SOCKBUF_LOCK(&so->so_rcv); if ((so->so_rcv.sb_hiwat == V_tcp_recvspace) && metrics.rmx_recvpipe) bufsize = metrics.rmx_recvpipe; else bufsize = so->so_rcv.sb_hiwat; if (bufsize > mss) { bufsize = roundup(bufsize, mss); if (bufsize > sb_max) bufsize = sb_max; if (bufsize > so->so_rcv.sb_hiwat) (void)sbreserve_locked(&so->so_rcv, bufsize, so, NULL); } SOCKBUF_UNLOCK(&so->so_rcv); /* Check the interface for TSO capabilities. */ if (cap.ifcap & CSUM_TSO) { tp->t_flags |= TF_TSO; tp->t_tsomax = cap.tsomax; tp->t_tsomaxsegcount = cap.tsomaxsegcount; tp->t_tsomaxsegsize = cap.tsomaxsegsize; } } /* * Determine the MSS option to send on an outgoing SYN. */ int tcp_mssopt(struct in_conninfo *inc) { int mss = 0; uint32_t thcmtu = 0; uint32_t maxmtu = 0; size_t min_protoh; KASSERT(inc != NULL, ("tcp_mssopt with NULL in_conninfo pointer")); #ifdef INET6 if (inc->inc_flags & INC_ISIPV6) { mss = V_tcp_v6mssdflt; maxmtu = tcp_maxmtu6(inc, NULL); min_protoh = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { mss = V_tcp_mssdflt; maxmtu = tcp_maxmtu(inc, NULL); min_protoh = sizeof(struct tcpiphdr); } #endif #if defined(INET6) || defined(INET) thcmtu = tcp_hc_getmtu(inc); /* IPv4 and IPv6 */ #endif if (maxmtu && thcmtu) mss = min(maxmtu, thcmtu) - min_protoh; else if (maxmtu || thcmtu) mss = max(maxmtu, thcmtu) - min_protoh; return (mss); } void tcp_do_prr_ack(struct tcpcb *tp, struct tcphdr *th, struct tcpopt *to) { int snd_cnt = 0, limit = 0, del_data = 0, pipe = 0; int maxseg = tcp_maxseg(tp); INP_WLOCK_ASSERT(tp->t_inpcb); /* * Compute the amount of data that this ACK is indicating * (del_data) and an estimate of how many bytes are in the * network. */ del_data = tp->sackhint.delivered_data; if (V_tcp_do_newsack) pipe = tcp_compute_pipe(tp); else pipe = (tp->snd_nxt - tp->snd_fack) + tp->sackhint.sack_bytes_rexmit; tp->sackhint.prr_delivered += del_data; /* * Proportional Rate Reduction */ if (pipe >= tp->snd_ssthresh) { if (tp->sackhint.recover_fs == 0) tp->sackhint.recover_fs = imax(1, tp->snd_nxt - tp->snd_una); snd_cnt = howmany((long)tp->sackhint.prr_delivered * tp->snd_ssthresh, tp->sackhint.recover_fs) - tp->sackhint.prr_out; } else { if (V_tcp_do_prr_conservative) limit = tp->sackhint.prr_delivered - tp->sackhint.prr_out; else limit = imax(tp->sackhint.prr_delivered - tp->sackhint.prr_out, del_data) + maxseg; snd_cnt = imin((tp->snd_ssthresh - pipe), limit); } snd_cnt = imax(snd_cnt, 0) / maxseg; /* * Send snd_cnt new data into the network in response to this ack. * If there is going to be a SACK retransmission, adjust snd_cwnd * accordingly. */ if (IN_FASTRECOVERY(tp->t_flags)) { tp->snd_cwnd = imax(maxseg, tp->snd_nxt - tp->snd_recover + tp->sackhint.sack_bytes_rexmit + (snd_cnt * maxseg)); } else if (IN_CONGRECOVERY(tp->t_flags)) tp->snd_cwnd = imax(maxseg, pipe - del_data + (snd_cnt * maxseg)); } /* * On a partial ack arrives, force the retransmission of the * next unacknowledged segment. Do not clear tp->t_dupacks. * By setting snd_nxt to ti_ack, this forces retransmission timer to * be started again. */ void tcp_newreno_partial_ack(struct tcpcb *tp, struct tcphdr *th) { tcp_seq onxt = tp->snd_nxt; uint32_t ocwnd = tp->snd_cwnd; u_int maxseg = tcp_maxseg(tp); INP_WLOCK_ASSERT(tp->t_inpcb); tcp_timer_activate(tp, TT_REXMT, 0); tp->t_rtttime = 0; tp->snd_nxt = th->th_ack; /* * Set snd_cwnd to one segment beyond acknowledged offset. * (tp->snd_una has not yet been updated when this function is called.) */ tp->snd_cwnd = maxseg + BYTES_THIS_ACK(tp, th); tp->t_flags |= TF_ACKNOW; (void) tp->t_fb->tfb_tcp_output(tp); tp->snd_cwnd = ocwnd; if (SEQ_GT(onxt, tp->snd_nxt)) tp->snd_nxt = onxt; /* * Partial window deflation. Relies on fact that tp->snd_una * not updated yet. */ if (tp->snd_cwnd > BYTES_THIS_ACK(tp, th)) tp->snd_cwnd -= BYTES_THIS_ACK(tp, th); else tp->snd_cwnd = 0; tp->snd_cwnd += maxseg; } int tcp_compute_pipe(struct tcpcb *tp) { return (tp->snd_max - tp->snd_una + tp->sackhint.sack_bytes_rexmit - tp->sackhint.sacked_bytes); } uint32_t tcp_compute_initwnd(uint32_t maxseg) { /* * Calculate the Initial Window, also used as Restart Window * * RFC5681 Section 3.1 specifies the default conservative values. * RFC3390 specifies slightly more aggressive values. * RFC6928 increases it to ten segments. * Support for user specified value for initial flight size. */ if (V_tcp_initcwnd_segments) return min(V_tcp_initcwnd_segments * maxseg, max(2 * maxseg, V_tcp_initcwnd_segments * 1460)); else if (V_tcp_do_rfc3390) return min(4 * maxseg, max(2 * maxseg, 4380)); else { /* Per RFC5681 Section 3.1 */ if (maxseg > 2190) return (2 * maxseg); else if (maxseg > 1095) return (3 * maxseg); else return (4 * maxseg); } } diff --git a/sys/netinet/tcp_output.c b/sys/netinet/tcp_output.c index e23cdc749e98..5bda2be14df0 100644 --- a/sys/netinet/tcp_output.c +++ b/sys/netinet/tcp_output.c @@ -1,2132 +1,2184 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95 */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_kern_tls.h" #include "opt_tcpdebug.h" #include #include #include #include #ifdef TCP_HHOOK #include #endif #include #ifdef KERN_TLS #include #endif #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef INET6 #include #include #include #endif #include #define TCPOUTFLAGS #include #include #include #include #include #include #include #include #ifdef TCPPCAP #include #endif #ifdef TCPDEBUG #include #endif #ifdef TCP_OFFLOAD #include #endif #include +#include +#include #include #include VNET_DEFINE(int, path_mtu_discovery) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(path_mtu_discovery), 1, "Enable Path MTU Discovery"); VNET_DEFINE(int, tcp_do_tso) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_tso), 0, "Enable TCP Segmentation Offload"); VNET_DEFINE(int, tcp_sendspace) = 1024*32; #define V_tcp_sendspace VNET(tcp_sendspace) SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size"); VNET_DEFINE(int, tcp_do_autosndbuf) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_autosndbuf), 0, "Enable automatic send buffer sizing"); VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024; SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_autosndbuf_inc), 0, "Incrementor step size of automatic send buffer"); VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024; SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_autosndbuf_max), 0, "Max size of automatic send buffer"); VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0; #define V_tcp_sendbuf_auto_lowat VNET(tcp_sendbuf_auto_lowat) SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sendbuf_auto_lowat), 0, "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT"); /* * Make sure that either retransmit or persist timer is set for SYN, FIN and * non-ACK. */ #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \ KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\ tcp_timer_active((tp), TT_REXMT) || \ tcp_timer_active((tp), TT_PERSIST), \ ("neither rexmt nor persist timer is set")) static void inline cc_after_idle(struct tcpcb *tp); #ifdef TCP_HHOOK /* * Wrapper for the TCP established output helper hook. */ void hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th, struct tcpopt *to, uint32_t len, int tso) { struct tcp_hhook_data hhook_data; if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) { hhook_data.tp = tp; hhook_data.th = th; hhook_data.to = to; hhook_data.len = len; hhook_data.tso = tso; hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data, tp->osd); } } #endif /* * CC wrapper hook functions */ static void inline cc_after_idle(struct tcpcb *tp) { INP_WLOCK_ASSERT(tp->t_inpcb); if (CC_ALGO(tp)->after_idle != NULL) CC_ALGO(tp)->after_idle(tp->ccv); } /* * Tcp output routine: figure out what should be sent and send it. */ int tcp_output(struct tcpcb *tp) { struct socket *so = tp->t_inpcb->inp_socket; int32_t len; uint32_t recwin, sendwin; int off, flags, error = 0; /* Keep compiler happy */ u_int if_hw_tsomaxsegcount = 0; u_int if_hw_tsomaxsegsize = 0; struct mbuf *m; struct ip *ip = NULL; #ifdef TCPDEBUG struct ipovly *ipov = NULL; #endif struct tcphdr *th; u_char opt[TCP_MAXOLEN]; - unsigned ipoptlen, optlen, hdrlen; + unsigned ipoptlen, optlen, hdrlen, ulen; #if defined(IPSEC) || defined(IPSEC_SUPPORT) unsigned ipsec_optlen = 0; #endif int idle, sendalot, curticks; int sack_rxmit, sack_bytes_rxmt; struct sackhole *p; int tso, mtu; struct tcpopt to; + struct udphdr *udp = NULL; unsigned int wanted_cookie = 0; unsigned int dont_sendalot = 0; #if 0 int maxburst = TCP_MAXBURST; #endif #ifdef INET6 struct ip6_hdr *ip6 = NULL; int isipv6; isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; #endif #ifdef KERN_TLS const bool hw_tls = (so->so_snd.sb_flags & SB_TLS_IFNET) != 0; #else const bool hw_tls = false; #endif NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(tp->t_inpcb); #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) return (tcp_offload_output(tp)); #endif /* * For TFO connections in SYN_SENT or SYN_RECEIVED, * only allow the initial SYN or SYN|ACK and those sent * by the retransmit timer. */ if (IS_FASTOPEN(tp->t_flags) && ((tp->t_state == TCPS_SYN_SENT) || (tp->t_state == TCPS_SYN_RECEIVED)) && SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */ (tp->snd_nxt != tp->snd_una)) /* not a retransmit */ return (0); /* * Determine length of data that should be transmitted, * and flags that will be used. * If there is some data or critical controls (SYN, RST) * to send, then transmit; otherwise, investigate further. */ idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) || (tp->t_sndtime && ((ticks - tp->t_sndtime) >= tp->t_rxtcur)))) cc_after_idle(tp); tp->t_flags &= ~TF_LASTIDLE; if (idle) { if (tp->t_flags & TF_MORETOCOME) { tp->t_flags |= TF_LASTIDLE; idle = 0; } } again: /* * If we've recently taken a timeout, snd_max will be greater than * snd_nxt. There may be SACK information that allows us to avoid * resending already delivered data. Adjust snd_nxt accordingly. */ if ((tp->t_flags & TF_SACK_PERMIT) && SEQ_LT(tp->snd_nxt, tp->snd_max)) tcp_sack_adjust(tp); sendalot = 0; tso = 0; mtu = 0; off = tp->snd_nxt - tp->snd_una; sendwin = min(tp->snd_wnd, tp->snd_cwnd); flags = tcp_outflags[tp->t_state]; /* * Send any SACK-generated retransmissions. If we're explicitly trying * to send out new data (when sendalot is 1), bypass this function. * If we retransmit in fast recovery mode, decrement snd_cwnd, since * we're replacing a (future) new transmission with a retransmission * now, and we previously incremented snd_cwnd in tcp_input(). */ /* * Still in sack recovery , reset rxmit flag to zero. */ sack_rxmit = 0; sack_bytes_rxmt = 0; len = 0; p = NULL; if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) && (p = tcp_sack_output(tp, &sack_bytes_rxmt))) { uint32_t cwin; cwin = imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0); /* Do not retransmit SACK segments beyond snd_recover */ if (SEQ_GT(p->end, tp->snd_recover)) { /* * (At least) part of sack hole extends beyond * snd_recover. Check to see if we can rexmit data * for this hole. */ if (SEQ_GEQ(p->rxmit, tp->snd_recover)) { /* * Can't rexmit any more data for this hole. * That data will be rexmitted in the next * sack recovery episode, when snd_recover * moves past p->rxmit. */ p = NULL; goto after_sack_rexmit; } else /* Can rexmit part of the current hole */ len = ((int32_t)ulmin(cwin, tp->snd_recover - p->rxmit)); } else len = ((int32_t)ulmin(cwin, p->end - p->rxmit)); off = p->rxmit - tp->snd_una; KASSERT(off >= 0,("%s: sack block to the left of una : %d", __func__, off)); if (len > 0) { sack_rxmit = 1; sendalot = 1; TCPSTAT_INC(tcps_sack_rexmits); TCPSTAT_ADD(tcps_sack_rexmit_bytes, min(len, tcp_maxseg(tp))); } } after_sack_rexmit: /* * Get standard flags, and add SYN or FIN if requested by 'hidden' * state flags. */ if (tp->t_flags & TF_NEEDFIN) flags |= TH_FIN; if (tp->t_flags & TF_NEEDSYN) flags |= TH_SYN; SOCKBUF_LOCK(&so->so_snd); /* * If in persist timeout with window of 0, send 1 byte. * Otherwise, if window is small but nonzero * and timer expired, we will send what we can * and go to transmit state. */ if (tp->t_flags & TF_FORCEDATA) { if (sendwin == 0) { /* * If we still have some data to send, then * clear the FIN bit. Usually this would * happen below when it realizes that we * aren't sending all the data. However, * if we have exactly 1 byte of unsent data, * then it won't clear the FIN bit below, * and if we are in persist state, we wind * up sending the packet without recording * that we sent the FIN bit. * * We can't just blindly clear the FIN bit, * because if we don't have any more data * to send then the probe will be the FIN * itself. */ if (off < sbused(&so->so_snd)) flags &= ~TH_FIN; sendwin = 1; } else { tcp_timer_activate(tp, TT_PERSIST, 0); tp->t_rxtshift = 0; } } /* * If snd_nxt == snd_max and we have transmitted a FIN, the * offset will be > 0 even if so_snd.sb_cc is 0, resulting in * a negative length. This can also occur when TCP opens up * its congestion window while receiving additional duplicate * acks after fast-retransmit because TCP will reset snd_nxt * to snd_max after the fast-retransmit. * * In the normal retransmit-FIN-only case, however, snd_nxt will * be set to snd_una, the offset will be 0, and the length may * wind up 0. * * If sack_rxmit is true we are retransmitting from the scoreboard * in which case len is already set. */ if (sack_rxmit == 0) { if (sack_bytes_rxmt == 0) len = ((int32_t)min(sbavail(&so->so_snd), sendwin) - off); else { int32_t cwin; /* * We are inside of a SACK recovery episode and are * sending new data, having retransmitted all the * data possible in the scoreboard. */ len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) - off); /* * Don't remove this (len > 0) check ! * We explicitly check for len > 0 here (although it * isn't really necessary), to work around a gcc * optimization issue - to force gcc to compute * len above. Without this check, the computation * of len is bungled by the optimizer. */ if (len > 0) { cwin = tp->snd_cwnd - (tp->snd_nxt - tp->snd_recover) - sack_bytes_rxmt; if (cwin < 0) cwin = 0; len = imin(len, cwin); } } } /* * Lop off SYN bit if it has already been sent. However, if this * is SYN-SENT state and if segment contains data and if we don't * know that foreign host supports TAO, suppress sending segment. */ if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { if (tp->t_state != TCPS_SYN_RECEIVED) flags &= ~TH_SYN; /* * When sending additional segments following a TFO SYN|ACK, * do not include the SYN bit. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_state == TCPS_SYN_RECEIVED)) flags &= ~TH_SYN; off--, len++; } /* * Be careful not to send data and/or FIN on SYN segments. * This measure is needed to prevent interoperability problems * with not fully conformant TCP implementations. */ if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { len = 0; flags &= ~TH_FIN; } /* * On TFO sockets, ensure no data is sent in the following cases: * * - When retransmitting SYN|ACK on a passively-created socket * * - When retransmitting SYN on an actively created socket * * - When sending a zero-length cookie (cookie request) on an * actively created socket * * - When the socket is in the CLOSED state (RST is being sent) */ if (IS_FASTOPEN(tp->t_flags) && (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || ((tp->t_state == TCPS_SYN_SENT) && (tp->t_tfo_client_cookie_len == 0)) || (flags & TH_RST))) len = 0; if (len <= 0) { /* * If FIN has been sent but not acked, * but we haven't been called to retransmit, * len will be < 0. Otherwise, window shrank * after we sent into it. If window shrank to 0, * cancel pending retransmit, pull snd_nxt back * to (closed) window, and set the persist timer * if it isn't already going. If the window didn't * close completely, just wait for an ACK. * * We also do a general check here to ensure that * we will set the persist timer when we have data * to send, but a 0-byte window. This makes sure * the persist timer is set even if the packet * hits one of the "goto send" lines below. */ len = 0; if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && (off < (int) sbavail(&so->so_snd))) { tcp_timer_activate(tp, TT_REXMT, 0); tp->t_rxtshift = 0; tp->snd_nxt = tp->snd_una; if (!tcp_timer_active(tp, TT_PERSIST)) tcp_setpersist(tp); } } /* len will be >= 0 after this point. */ KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); tcp_sndbuf_autoscale(tp, so, sendwin); /* * Decide if we can use TCP Segmentation Offloading (if supported by * hardware). * * TSO may only be used if we are in a pure bulk sending state. The * presence of TCP-MD5, SACK retransmits, SACK advertizements and * IP options prevent using TSO. With TSO the TCP header is the same * (except for the sequence number) for all generated packets. This * makes it impossible to transmit any options which vary per generated * segment or packet. * * IPv4 handling has a clear separation of ip options and ip header * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does * the right thing below to provide length of just ip options and thus * checking for ipoptlen is enough to decide if ip options are present. */ #if defined(IPSEC) || defined(IPSEC_SUPPORT) /* * Pre-calculate here as we save another lookup into the darknesses * of IPsec that way and can actually decide if TSO is ok. */ #ifdef INET6 if (isipv6 && IPSEC_ENABLED(ipv6)) ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb); #ifdef INET else #endif #endif /* INET6 */ #ifdef INET if (IPSEC_ENABLED(ipv4)) ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb); #endif /* INET */ #endif /* IPSEC */ #ifdef INET6 if (isipv6) ipoptlen = ip6_optlen(tp->t_inpcb); else #endif if (tp->t_inpcb->inp_options) ipoptlen = tp->t_inpcb->inp_options->m_len - offsetof(struct ipoption, ipopt_list); else ipoptlen = 0; #if defined(IPSEC) || defined(IPSEC_SUPPORT) ipoptlen += ipsec_optlen; #endif if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && + (tp->t_port == 0) && ((tp->t_flags & TF_SIGNATURE) == 0) && tp->rcv_numsacks == 0 && sack_rxmit == 0 && ipoptlen == 0 && !(flags & TH_SYN)) tso = 1; if (sack_rxmit) { if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd))) flags &= ~TH_FIN; } else { if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + sbused(&so->so_snd))) flags &= ~TH_FIN; } recwin = lmin(lmax(sbspace(&so->so_rcv), 0), (long)TCP_MAXWIN << tp->rcv_scale); /* * Sender silly window avoidance. We transmit under the following * conditions when len is non-zero: * * - We have a full segment (or more with TSO) * - This is the last buffer in a write()/send() and we are * either idle or running NODELAY * - we've timed out (e.g. persist timer) * - we have more then 1/2 the maximum send window's worth of * data (receiver may be limited the window size) * - we need to retransmit */ if (len) { if (len >= tp->t_maxseg) goto send; /* * As the TCP header options are now * considered when setting up the initial * window, we would not send the last segment * if we skip considering the option length here. * Note: this may not work when tcp headers change * very dynamically in the future. */ if ((((tp->t_flags & TF_SIGNATURE) ? PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) + ((tp->t_flags & TF_RCVD_TSTMP) ? PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) + len) >= tp->t_maxseg) goto send; /* * NOTE! on localhost connections an 'ack' from the remote * end may occur synchronously with the output and cause * us to flush a buffer queued with moretocome. XXX * * note: the len + off check is almost certainly unnecessary. */ if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ (idle || (tp->t_flags & TF_NODELAY)) && (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) && (tp->t_flags & TF_NOPUSH) == 0) { goto send; } if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */ goto send; if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) goto send; if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */ goto send; if (sack_rxmit) goto send; } /* * Sending of standalone window updates. * * Window updates are important when we close our window due to a * full socket buffer and are opening it again after the application * reads data from it. Once the window has opened again and the * remote end starts to send again the ACK clock takes over and * provides the most current window information. * * We must avoid the silly window syndrome whereas every read * from the receive buffer, no matter how small, causes a window * update to be sent. We also should avoid sending a flurry of * window updates when the socket buffer had queued a lot of data * and the application is doing small reads. * * Prevent a flurry of pointless window updates by only sending * an update when we can increase the advertized window by more * than 1/4th of the socket buffer capacity. When the buffer is * getting full or is very small be more aggressive and send an * update whenever we can increase by two mss sized segments. * In all other situations the ACK's to new incoming data will * carry further window increases. * * Don't send an independent window update if a delayed * ACK is pending (it will get piggy-backed on it) or the * remote side already has done a half-close and won't send * more data. Skip this if the connection is in T/TCP * half-open state. */ if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && !(tp->t_flags & TF_DELACK) && !TCPS_HAVERCVDFIN(tp->t_state)) { /* * "adv" is the amount we could increase the window, * taking into account that we are limited by * TCP_MAXWIN << tp->rcv_scale. */ int32_t adv; int oldwin; adv = recwin; if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { oldwin = (tp->rcv_adv - tp->rcv_nxt); if (adv > oldwin) adv -= oldwin; else adv = 0; } else oldwin = 0; /* * If the new window size ends up being the same as or less * than the old size when it is scaled, then don't force * a window update. */ if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) goto dontupdate; if (adv >= (int32_t)(2 * tp->t_maxseg) && (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) || recwin <= (so->so_rcv.sb_hiwat / 8) || so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg || adv >= TCP_MAXWIN << tp->rcv_scale)) goto send; if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat) goto send; } dontupdate: /* * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW * is also a catch-all for the retransmit timer timeout case. */ if (tp->t_flags & TF_ACKNOW) goto send; if ((flags & TH_RST) || ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) goto send; if (SEQ_GT(tp->snd_up, tp->snd_una)) goto send; /* * If our state indicates that FIN should be sent * and we have not yet done so, then we need to send. */ if (flags & TH_FIN && ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) goto send; /* * In SACK, it is possible for tcp_output to fail to send a segment * after the retransmission timer has been turned off. Make sure * that the retransmission timer is set. */ if ((tp->t_flags & TF_SACK_PERMIT) && SEQ_GT(tp->snd_max, tp->snd_una) && !tcp_timer_active(tp, TT_REXMT) && !tcp_timer_active(tp, TT_PERSIST)) { tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); goto just_return; } /* * TCP window updates are not reliable, rather a polling protocol * using ``persist'' packets is used to insure receipt of window * updates. The three ``states'' for the output side are: * idle not doing retransmits or persists * persisting to move a small or zero window * (re)transmitting and thereby not persisting * * tcp_timer_active(tp, TT_PERSIST) * is true when we are in persist state. * (tp->t_flags & TF_FORCEDATA) * is set when we are called to send a persist packet. * tcp_timer_active(tp, TT_REXMT) * is set when we are retransmitting * The output side is idle when both timers are zero. * * If send window is too small, there is data to transmit, and no * retransmit or persist is pending, then go to persist state. * If nothing happens soon, send when timer expires: * if window is nonzero, transmit what we can, * otherwise force out a byte. */ if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) && !tcp_timer_active(tp, TT_PERSIST)) { tp->t_rxtshift = 0; tcp_setpersist(tp); } /* * No reason to send a segment, just return. */ just_return: SOCKBUF_UNLOCK(&so->so_snd); return (0); send: SOCKBUF_LOCK_ASSERT(&so->so_snd); if (len > 0) { if (len >= tp->t_maxseg) tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; else tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; } /* * Before ESTABLISHED, force sending of initial options * unless TCP set not to do any options. * NOTE: we assume that the IP/TCP header plus TCP options * always fit in a single mbuf, leaving room for a maximum * link header, i.e. * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES */ optlen = 0; #ifdef INET6 if (isipv6) hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); else #endif hdrlen = sizeof (struct tcpiphdr); if (flags & TH_SYN) { tp->snd_nxt = tp->iss; } /* * Compute options for segment. * We only have to care about SYN and established connection * segments. Options for SYN-ACK segments are handled in TCP * syncache. */ to.to_flags = 0; if ((tp->t_flags & TF_NOOPT) == 0) { /* Maximum segment size. */ if (flags & TH_SYN) { to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc); + if (tp->t_port) + to.to_mss -= V_tcp_udp_tunneling_overhead; to.to_flags |= TOF_MSS; /* * On SYN or SYN|ACK transmits on TFO connections, * only include the TFO option if it is not a * retransmit, as the presence of the TFO option may * have caused the original SYN or SYN|ACK to have * been dropped by a middlebox. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_rxtshift == 0)) { if (tp->t_state == TCPS_SYN_RECEIVED) { to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; to.to_tfo_cookie = (u_int8_t *)&tp->t_tfo_cookie.server; to.to_flags |= TOF_FASTOPEN; wanted_cookie = 1; } else if (tp->t_state == TCPS_SYN_SENT) { to.to_tfo_len = tp->t_tfo_client_cookie_len; to.to_tfo_cookie = tp->t_tfo_cookie.client; to.to_flags |= TOF_FASTOPEN; wanted_cookie = 1; /* * If we wind up having more data to * send with the SYN than can fit in * one segment, don't send any more * until the SYN|ACK comes back from * the other end. */ dont_sendalot = 1; } } } /* Window scaling. */ if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { to.to_wscale = tp->request_r_scale; to.to_flags |= TOF_SCALE; } /* Timestamps. */ if ((tp->t_flags & TF_RCVD_TSTMP) || ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { curticks = tcp_ts_getticks(); to.to_tsval = curticks + tp->ts_offset; to.to_tsecr = tp->ts_recent; to.to_flags |= TOF_TS; if (tp->t_rxtshift == 1) tp->t_badrxtwin = curticks; } /* Set receive buffer autosizing timestamp. */ if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE)) tp->rfbuf_ts = tcp_ts_getticks(); /* Selective ACK's. */ if (tp->t_flags & TF_SACK_PERMIT) { if (flags & TH_SYN) to.to_flags |= TOF_SACKPERM; else if (TCPS_HAVEESTABLISHED(tp->t_state) && tp->rcv_numsacks > 0) { to.to_flags |= TOF_SACK; to.to_nsacks = tp->rcv_numsacks; to.to_sacks = (u_char *)tp->sackblks; } } #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) /* TCP-MD5 (RFC2385). */ /* * Check that TCP_MD5SIG is enabled in tcpcb to * account the size needed to set this TCP option. */ if (tp->t_flags & TF_SIGNATURE) to.to_flags |= TOF_SIGNATURE; #endif /* TCP_SIGNATURE */ /* Processing the options. */ hdrlen += optlen = tcp_addoptions(&to, opt); /* * If we wanted a TFO option to be added, but it was unable * to fit, ensure no data is sent. */ if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && !(to.to_flags & TOF_FASTOPEN)) len = 0; } - + if (tp->t_port) { + if (V_tcp_udp_tunneling_port == 0) { + /* The port was removed?? */ + SOCKBUF_UNLOCK(&so->so_snd); + return (EHOSTUNREACH); + } + hdrlen += sizeof(struct udphdr); + } /* * Adjust data length if insertion of options will * bump the packet length beyond the t_maxseg length. * Clear the FIN bit because we cut off the tail of * the segment. */ if (len + optlen + ipoptlen > tp->t_maxseg) { flags &= ~TH_FIN; if (tso) { u_int if_hw_tsomax; u_int moff; int max_len; /* extract TSO information */ if_hw_tsomax = tp->t_tsomax; if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; /* * Limit a TSO burst to prevent it from * overflowing or exceeding the maximum length * allowed by the network interface: */ KASSERT(ipoptlen == 0, ("%s: TSO can't do IP options", __func__)); /* * Check if we should limit by maximum payload * length: */ if (if_hw_tsomax != 0) { /* compute maximum TSO length */ max_len = (if_hw_tsomax - hdrlen - max_linkhdr); if (max_len <= 0) { len = 0; } else if (len > max_len) { sendalot = 1; len = max_len; } } /* * Prevent the last segment from being * fractional unless the send sockbuf can be * emptied: */ max_len = (tp->t_maxseg - optlen); if (((uint32_t)off + (uint32_t)len) < sbavail(&so->so_snd)) { moff = len % max_len; if (moff != 0) { len -= moff; sendalot = 1; } } /* * In case there are too many small fragments * don't use TSO: */ if (len <= max_len) { len = max_len; sendalot = 1; tso = 0; } /* * Send the FIN in a separate segment * after the bulk sending is done. * We don't trust the TSO implementations * to clear the FIN flag on all but the * last segment. */ if (tp->t_flags & TF_NEEDFIN) sendalot = 1; } else { if (optlen + ipoptlen >= tp->t_maxseg) { /* * Since we don't have enough space to put * the IP header chain and the TCP header in * one packet as required by RFC 7112, don't * send it. Also ensure that at least one * byte of the payload can be put into the * TCP segment. */ SOCKBUF_UNLOCK(&so->so_snd); error = EMSGSIZE; sack_rxmit = 0; goto out; } len = tp->t_maxseg - optlen - ipoptlen; sendalot = 1; if (dont_sendalot) sendalot = 0; } } else tso = 0; KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, ("%s: len > IP_MAXPACKET", __func__)); /*#ifdef DIAGNOSTIC*/ #ifdef INET6 if (max_linkhdr + hdrlen > MCLBYTES) #else if (max_linkhdr + hdrlen > MHLEN) #endif panic("tcphdr too big"); /*#endif*/ /* * This KASSERT is here to catch edge cases at a well defined place. * Before, those had triggered (random) panic conditions further down. */ KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); /* * Grab a header mbuf, attaching a copy of data to * be transmitted, and initialize the header from * the template for sends on this connection. */ if (len) { struct mbuf *mb; struct sockbuf *msb; u_int moff; if ((tp->t_flags & TF_FORCEDATA) && len == 1) { TCPSTAT_INC(tcps_sndprobe); #ifdef STATS if (SEQ_LT(tp->snd_nxt, tp->snd_max)) stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, len); else stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, len); #endif /* STATS */ } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { tp->t_sndrexmitpack++; TCPSTAT_INC(tcps_sndrexmitpack); TCPSTAT_ADD(tcps_sndrexmitbyte, len); #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, len); #endif /* STATS */ } else { TCPSTAT_INC(tcps_sndpack); TCPSTAT_ADD(tcps_sndbyte, len); #ifdef STATS stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, len); #endif /* STATS */ } #ifdef INET6 if (MHLEN < hdrlen + max_linkhdr) m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); else #endif m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) { SOCKBUF_UNLOCK(&so->so_snd); error = ENOBUFS; sack_rxmit = 0; goto out; } m->m_data += max_linkhdr; m->m_len = hdrlen; /* * Start the m_copy functions from the closest mbuf * to the offset in the socket buffer chain. */ mb = sbsndptr_noadv(&so->so_snd, off, &moff); if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) { m_copydata(mb, moff, len, mtod(m, caddr_t) + hdrlen); if (SEQ_LT(tp->snd_nxt, tp->snd_max)) sbsndptr_adv(&so->so_snd, mb, len); m->m_len += len; } else { if (SEQ_LT(tp->snd_nxt, tp->snd_max)) msb = NULL; else msb = &so->so_snd; m->m_next = tcp_m_copym(mb, moff, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb, hw_tls); if (len <= (tp->t_maxseg - optlen)) { /* * Must have ran out of mbufs for the copy * shorten it to no longer need tso. Lets * not put on sendalot since we are low on * mbufs. */ tso = 0; } if (m->m_next == NULL) { SOCKBUF_UNLOCK(&so->so_snd); (void) m_free(m); error = ENOBUFS; sack_rxmit = 0; goto out; } } /* * If we're sending everything we've got, set PUSH. * (This will keep happy those implementations which only * give data to the user when a buffer fills or * a PUSH comes in.) */ if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) && !(flags & TH_SYN)) flags |= TH_PUSH; SOCKBUF_UNLOCK(&so->so_snd); } else { SOCKBUF_UNLOCK(&so->so_snd); if (tp->t_flags & TF_ACKNOW) TCPSTAT_INC(tcps_sndacks); else if (flags & (TH_SYN|TH_FIN|TH_RST)) TCPSTAT_INC(tcps_sndctrl); else if (SEQ_GT(tp->snd_up, tp->snd_una)) TCPSTAT_INC(tcps_sndurg); else TCPSTAT_INC(tcps_sndwinup); m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) { error = ENOBUFS; sack_rxmit = 0; goto out; } #ifdef INET6 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && MHLEN >= hdrlen) { M_ALIGN(m, hdrlen); } else #endif m->m_data += max_linkhdr; m->m_len = hdrlen; } SOCKBUF_UNLOCK_ASSERT(&so->so_snd); m->m_pkthdr.rcvif = (struct ifnet *)0; #ifdef MAC mac_inpcb_create_mbuf(tp->t_inpcb, m); #endif #ifdef INET6 if (isipv6) { ip6 = mtod(m, struct ip6_hdr *); - th = (struct tcphdr *)(ip6 + 1); - tcpip_fillheaders(tp->t_inpcb, ip6, th); + if (tp->t_port) { + udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + sizeof(struct ip6_hdr)); + udp->uh_sport = htons(V_tcp_udp_tunneling_port); + udp->uh_dport = tp->t_port; + ulen = hdrlen + len - sizeof(struct ip6_hdr); + udp->uh_ulen = htons(ulen); + th = (struct tcphdr *)(udp + 1); + } else { + th = (struct tcphdr *)(ip6 + 1); + } + tcpip_fillheaders(tp->t_inpcb, tp->t_port, ip6, th); } else #endif /* INET6 */ { ip = mtod(m, struct ip *); #ifdef TCPDEBUG ipov = (struct ipovly *)ip; #endif - th = (struct tcphdr *)(ip + 1); - tcpip_fillheaders(tp->t_inpcb, ip, th); + if (tp->t_port) { + udp = (struct udphdr *)((caddr_t)ip + ipoptlen + sizeof(struct ip)); + udp->uh_sport = htons(V_tcp_udp_tunneling_port); + udp->uh_dport = tp->t_port; + ulen = hdrlen + len - sizeof(struct ip); + udp->uh_ulen = htons(ulen); + th = (struct tcphdr *)(udp + 1); + } else + th = (struct tcphdr *)(ip + 1); + tcpip_fillheaders(tp->t_inpcb, tp->t_port, ip, th); } /* * Fill in fields, remembering maximum advertised * window for use in delaying messages about window sizes. * If resending a FIN, be sure not to use a new sequence number. */ if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && tp->snd_nxt == tp->snd_max) tp->snd_nxt--; /* * If we are starting a connection, send ECN setup * SYN packet. If we are on a retransmit, we may * resend those bits a number of times as per * RFC 3168. */ if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) { if (tp->t_rxtshift >= 1) { if (tp->t_rxtshift <= V_tcp_ecn_maxretries) flags |= TH_ECE|TH_CWR; } else flags |= TH_ECE|TH_CWR; } /* Handle parallel SYN for ECN */ if ((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_flags2 & TF2_ECN_SND_ECE)) { flags |= TH_ECE; tp->t_flags2 &= ~TF2_ECN_SND_ECE; } if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->t_flags2 & TF2_ECN_PERMIT)) { /* * If the peer has ECN, mark data packets with * ECN capable transmission (ECT). * Ignore pure ack packets, retransmissions and window probes. */ if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && (sack_rxmit == 0) && !((tp->t_flags & TF_FORCEDATA) && len == 1 && SEQ_LT(tp->snd_una, tp->snd_max))) { #ifdef INET6 if (isipv6) ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); else #endif ip->ip_tos |= IPTOS_ECN_ECT0; TCPSTAT_INC(tcps_ecn_ect0); /* * Reply with proper ECN notifications. * Only set CWR on new data segments. */ if (tp->t_flags2 & TF2_ECN_SND_CWR) { flags |= TH_CWR; tp->t_flags2 &= ~TF2_ECN_SND_CWR; } } if (tp->t_flags2 & TF2_ECN_SND_ECE) flags |= TH_ECE; } /* * If we are doing retransmissions, then snd_nxt will * not reflect the first unsent octet. For ACK only * packets, we do not want the sequence number of the * retransmitted packet, we want the sequence number * of the next unsent octet. So, if there is no data * (and no SYN or FIN), use snd_max instead of snd_nxt * when filling in ti_seq. But if we are in persist * state, snd_max might reflect one byte beyond the * right edge of the window, so use snd_nxt in that * case, since we know we aren't doing a retransmission. * (retransmit and persist are mutually exclusive...) */ if (sack_rxmit == 0) { if (len || (flags & (TH_SYN|TH_FIN)) || tcp_timer_active(tp, TT_PERSIST)) th->th_seq = htonl(tp->snd_nxt); else th->th_seq = htonl(tp->snd_max); } else { th->th_seq = htonl(p->rxmit); p->rxmit += len; tp->sackhint.sack_bytes_rexmit += len; } if (IN_RECOVERY(tp->t_flags)) { /* * Account all bytes transmitted while * IN_RECOVERY, simplifying PRR and * Lost Retransmit Detection */ tp->sackhint.prr_out += len; } th->th_ack = htonl(tp->rcv_nxt); if (optlen) { bcopy(opt, th + 1, optlen); th->th_off = (sizeof (struct tcphdr) + optlen) >> 2; } th->th_flags = flags; /* * Calculate receive window. Don't shrink window, * but avoid silly window syndrome. * If a RST segment is sent, advertise a window of zero. */ if (flags & TH_RST) { recwin = 0; } else { if (recwin < (so->so_rcv.sb_hiwat / 4) && recwin < tp->t_maxseg) recwin = 0; if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && recwin < (tp->rcv_adv - tp->rcv_nxt)) recwin = (tp->rcv_adv - tp->rcv_nxt); } /* * According to RFC1323 the window field in a SYN (i.e., a * or ) segment itself is never scaled. The * case is handled in syncache. */ if (flags & TH_SYN) th->th_win = htons((u_short) (min(sbspace(&so->so_rcv), TCP_MAXWIN))); else { /* Avoid shrinking window with window scaling. */ recwin = roundup2(recwin, 1 << tp->rcv_scale); th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); } /* * Adjust the RXWIN0SENT flag - indicate that we have advertised * a 0 window. This may cause the remote transmitter to stall. This * flag tells soreceive() to disable delayed acknowledgements when * draining the buffer. This can occur if the receiver is attempting * to read more data than can be buffered prior to transmitting on * the connection. */ if (th->th_win == 0) { tp->t_sndzerowin++; tp->t_flags |= TF_RXWIN0SENT; } else tp->t_flags &= ~TF_RXWIN0SENT; if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); th->th_flags |= TH_URG; } else /* * If no urgent pointer to send, then we pull * the urgent pointer to the left edge of the send window * so that it doesn't drift into the send window on sequence * number wraparound. */ tp->snd_up = tp->snd_una; /* drag it along */ /* * Put TCP length in extended header, and then * checksum extended header and data. */ m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ - m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (to.to_flags & TOF_SIGNATURE) { /* * Calculate MD5 signature and put it into the place * determined before. * NOTE: since TCP options buffer doesn't point into * mbuf's data, calculate offset and use it. */ if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th, (u_char *)(th + 1) + (to.to_signature - opt))) != 0) { /* * Do not send segment if the calculation of MD5 * digest has failed. */ m_freem(m); goto out; } } #endif #ifdef INET6 if (isipv6) { /* * There is no need to fill in ip6_plen right now. * It will be filled later by ip6_output. */ - m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; - th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + - optlen + len, IPPROTO_TCP, 0); + if (tp->t_port) { + m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; + m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); + udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); + th->th_sum = htons(0); + UDPSTAT_INC(udps_opackets); + } else { + m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; + m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); + th->th_sum = in6_cksum_pseudo(ip6, + sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, + 0); + } } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { - m->m_pkthdr.csum_flags = CSUM_TCP; - th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, - htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); + if (tp->t_port) { + m->m_pkthdr.csum_flags = CSUM_UDP; + m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); + udp->uh_sum = in_pseudo(ip->ip_src.s_addr, + ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); + th->th_sum = htons(0); + UDPSTAT_INC(udps_opackets); + } else { + m->m_pkthdr.csum_flags = CSUM_TCP; + m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); + th->th_sum = in_pseudo(ip->ip_src.s_addr, + ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + + IPPROTO_TCP + len + optlen)); + } /* IP version must be set here for ipv4/ipv6 checking later */ KASSERT(ip->ip_v == IPVERSION, ("%s: IP version incorrect: %d", __func__, ip->ip_v)); } #endif /* * Enable TSO and specify the size of the segments. * The TCP pseudo header checksum is always provided. */ if (tso) { KASSERT(len > tp->t_maxseg - optlen, ("%s: len <= tso_segsz", __func__)); m->m_pkthdr.csum_flags |= CSUM_TSO; m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; } KASSERT(len + hdrlen == m_length(m, NULL), ("%s: mbuf chain shorter than expected: %d + %u != %u", __func__, len, hdrlen, m_length(m, NULL))); #ifdef TCP_HHOOK /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ hhook_run_tcp_est_out(tp, th, &to, len, tso); #endif #ifdef TCPDEBUG /* * Trace. */ if (so->so_options & SO_DEBUG) { u_short save = 0; #ifdef INET6 if (!isipv6) #endif { save = ipov->ih_len; ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */); } tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); #ifdef INET6 if (!isipv6) #endif ipov->ih_len = save; } #endif /* TCPDEBUG */ TCP_PROBE3(debug__output, tp, th, m); /* We're getting ready to send; log now. */ TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, len, NULL, false); /* * Fill in IP length and desired time to live and * send to IP level. There should be a better way * to handle ttl and tos; we could keep them in * the template, but need a way to checksum without them. */ /* * m->m_pkthdr.len should have been set before checksum calculation, * because in6_cksum() need it. */ #ifdef INET6 if (isipv6) { /* * we separately set hoplimit for every segment, since the * user might want to change the value via setsockopt. * Also, desired default hop limit might be changed via * Neighbor Discovery. */ ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL); /* * Set the packet size here for the benefit of DTrace probes. * ip6_output() will set it properly; it's supposed to include * the option header lengths as well. */ ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) tp->t_flags2 |= TF2_PLPMTU_PMTUD; else tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; if (tp->t_state == TCPS_SYN_SENT) TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); TCP_PROBE5(send, NULL, tp, ip6, tp, th); #ifdef TCPPCAP /* Save packet, if requested. */ tcp_pcap_add(th, m, &(tp->t_outpkts)); #endif /* TODO: IPv6 IP6TOS_ECT bit on */ error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &tp->t_inpcb->inp_route6, ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb); if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_nh != NULL) mtu = tp->t_inpcb->inp_route6.ro_nh->nh_mtu; } #endif /* INET6 */ #if defined(INET) && defined(INET6) else #endif #ifdef INET { ip->ip_len = htons(m->m_pkthdr.len); #ifdef INET6 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO) ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL); #endif /* INET6 */ /* * If we do path MTU discovery, then we set DF on every packet. * This might not be the best thing to do according to RFC3390 * Section 2. However the tcp hostcache migitates the problem * so it affects only the first tcp connection with a host. * * NB: Don't set DF on small MTU/MSS to have a safe fallback. */ if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { - ip->ip_off |= htons(IP_DF); tp->t_flags2 |= TF2_PLPMTU_PMTUD; + if (tp->t_port == 0 || len < V_tcp_minmss) { + ip->ip_off |= htons(IP_DF); + } } else { tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; } if (tp->t_state == TCPS_SYN_SENT) TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); TCP_PROBE5(send, NULL, tp, ip, tp, th); #ifdef TCPPCAP /* Save packet, if requested. */ tcp_pcap_add(th, m, &(tp->t_outpkts)); #endif error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route, ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, tp->t_inpcb); if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_nh != NULL) mtu = tp->t_inpcb->inp_route.ro_nh->nh_mtu; } #endif /* INET */ out: /* * In transmit state, time the transmission and arrange for * the retransmit. In persist state, just set snd_max. */ if ((tp->t_flags & TF_FORCEDATA) == 0 || !tcp_timer_active(tp, TT_PERSIST)) { tcp_seq startseq = tp->snd_nxt; /* * Advance snd_nxt over sequence space of this segment. */ if (flags & (TH_SYN|TH_FIN)) { if (flags & TH_SYN) tp->snd_nxt++; if (flags & TH_FIN) { tp->snd_nxt++; tp->t_flags |= TF_SENTFIN; } } if (sack_rxmit) goto timer; tp->snd_nxt += len; if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { tp->snd_max = tp->snd_nxt; /* * Time this transmission if not a retransmission and * not currently timing anything. */ tp->t_sndtime = ticks; if (tp->t_rtttime == 0) { tp->t_rtttime = ticks; tp->t_rtseq = startseq; TCPSTAT_INC(tcps_segstimed); } #ifdef STATS if (!(tp->t_flags & TF_GPUTINPROG) && len) { tp->t_flags |= TF_GPUTINPROG; tp->gput_seq = startseq; tp->gput_ack = startseq + ulmin(sbavail(&so->so_snd) - off, sendwin); tp->gput_ts = tcp_ts_getticks(); } #endif /* STATS */ } /* * Set retransmit timer if not currently set, * and not doing a pure ack or a keep-alive probe. * Initial value for retransmit timer is smoothed * round-trip time + 2 * round-trip time variance. * Initialize shift counter which is used for backoff * of retransmit time. */ timer: if (!tcp_timer_active(tp, TT_REXMT) && ((sack_rxmit && tp->snd_nxt != tp->snd_max) || (tp->snd_nxt != tp->snd_una))) { if (tcp_timer_active(tp, TT_PERSIST)) { tcp_timer_activate(tp, TT_PERSIST, 0); tp->t_rxtshift = 0; } tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); } else if (len == 0 && sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) && !tcp_timer_active(tp, TT_PERSIST)) { /* * Avoid a situation where we do not set persist timer * after a zero window condition. For example: * 1) A -> B: packet with enough data to fill the window * 2) B -> A: ACK for #1 + new data (0 window * advertisement) * 3) A -> B: ACK for #2, 0 len packet * * In this case, A will not activate the persist timer, * because it chose to send a packet. Unless tcp_output * is called for some other reason (delayed ack timer, * another input packet from B, socket syscall), A will * not send zero window probes. * * So, if you send a 0-length packet, but there is data * in the socket buffer, and neither the rexmt or * persist timer is already set, then activate the * persist timer. */ tp->t_rxtshift = 0; tcp_setpersist(tp); } } else { /* * Persist case, update snd_max but since we are in * persist mode (no window) we do not update snd_nxt. */ int xlen = len; if (flags & TH_SYN) ++xlen; if (flags & TH_FIN) { ++xlen; tp->t_flags |= TF_SENTFIN; } if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) tp->snd_max = tp->snd_nxt + xlen; } if ((error == 0) && (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->t_flags & TF_SACK_PERMIT) && tp->rcv_numsacks > 0)) { /* Clean up any DSACK's sent */ tcp_clean_dsack_blocks(tp); } if (error) { /* Record the error. */ TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, error, 0, NULL, false); /* * We know that the packet was lost, so back out the * sequence number advance, if any. * * If the error is EPERM the packet got blocked by the * local firewall. Normally we should terminate the * connection but the blocking may have been spurious * due to a firewall reconfiguration cycle. So we treat * it like a packet loss and let the retransmit timer and * timeouts do their work over time. * XXX: It is a POLA question whether calling tcp_drop right * away would be the really correct behavior instead. */ if (((tp->t_flags & TF_FORCEDATA) == 0 || !tcp_timer_active(tp, TT_PERSIST)) && ((flags & TH_SYN) == 0) && (error != EPERM)) { if (sack_rxmit) { p->rxmit -= len; tp->sackhint.sack_bytes_rexmit -= len; KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, ("sackhint bytes rtx >= 0")); } else tp->snd_nxt -= len; } SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */ switch (error) { case EACCES: case EPERM: tp->t_softerror = error; return (error); case ENOBUFS: TCP_XMIT_TIMER_ASSERT(tp, len, flags); tp->snd_cwnd = tp->t_maxseg; return (0); case EMSGSIZE: /* * For some reason the interface we used initially * to send segments changed to another or lowered * its MTU. * If TSO was active we either got an interface * without TSO capabilits or TSO was turned off. * If we obtained mtu from ip_output() then update * it and try again. */ if (tso) tp->t_flags &= ~TF_TSO; if (mtu != 0) { tcp_mss_update(tp, -1, mtu, NULL, NULL); goto again; } return (error); case EHOSTDOWN: case EHOSTUNREACH: case ENETDOWN: case ENETUNREACH: if (TCPS_HAVERCVDSYN(tp->t_state)) { tp->t_softerror = error; return (0); } /* FALLTHROUGH */ default: return (error); } } TCPSTAT_INC(tcps_sndtotal); /* * Data sent (as far as we can tell). * If this advertises a larger window than any other segment, * then remember the size of the advertised window. * Any pending ACK has now been sent. */ if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) tp->rcv_adv = tp->rcv_nxt + recwin; tp->last_ack_sent = tp->rcv_nxt; tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); if (tcp_timer_active(tp, TT_DELACK)) tcp_timer_activate(tp, TT_DELACK, 0); #if 0 /* * This completely breaks TCP if newreno is turned on. What happens * is that if delayed-acks are turned on on the receiver, this code * on the transmitter effectively destroys the TCP window, forcing * it to four packets (1.5Kx4 = 6K window). */ if (sendalot && --maxburst) goto again; #endif if (sendalot) goto again; return (0); } void tcp_setpersist(struct tcpcb *tp) { int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; int tt; tp->t_flags &= ~TF_PREVVALID; if (tcp_timer_active(tp, TT_REXMT)) panic("tcp_setpersist: retransmit pending"); /* * Start/restart persistence timer. */ TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], tcp_persmin, tcp_persmax); tcp_timer_activate(tp, TT_PERSIST, tt); if (tp->t_rxtshift < TCP_MAXRXTSHIFT) tp->t_rxtshift++; } /* * Insert TCP options according to the supplied parameters to the place * optp in a consistent way. Can handle unaligned destinations. * * The order of the option processing is crucial for optimal packing and * alignment for the scarce option space. * * The optimal order for a SYN/SYN-ACK segment is: * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) + * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40. * * The SACK options should be last. SACK blocks consume 8*n+2 bytes. * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks). * At minimum we need 10 bytes (to generate 1 SACK block). If both * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present, * we only have 10 bytes for SACK options (40 - (12 + 18)). */ int tcp_addoptions(struct tcpopt *to, u_char *optp) { u_int32_t mask, optlen = 0; for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) { if ((to->to_flags & mask) != mask) continue; if (optlen == TCP_MAXOLEN) break; switch (to->to_flags & mask) { case TOF_MSS: while (optlen % 4) { optlen += TCPOLEN_NOP; *optp++ = TCPOPT_NOP; } if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG) continue; optlen += TCPOLEN_MAXSEG; *optp++ = TCPOPT_MAXSEG; *optp++ = TCPOLEN_MAXSEG; to->to_mss = htons(to->to_mss); bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss)); optp += sizeof(to->to_mss); break; case TOF_SCALE: while (!optlen || optlen % 2 != 1) { optlen += TCPOLEN_NOP; *optp++ = TCPOPT_NOP; } if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW) continue; optlen += TCPOLEN_WINDOW; *optp++ = TCPOPT_WINDOW; *optp++ = TCPOLEN_WINDOW; *optp++ = to->to_wscale; break; case TOF_SACKPERM: while (optlen % 2) { optlen += TCPOLEN_NOP; *optp++ = TCPOPT_NOP; } if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED) continue; optlen += TCPOLEN_SACK_PERMITTED; *optp++ = TCPOPT_SACK_PERMITTED; *optp++ = TCPOLEN_SACK_PERMITTED; break; case TOF_TS: while (!optlen || optlen % 4 != 2) { optlen += TCPOLEN_NOP; *optp++ = TCPOPT_NOP; } if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP) continue; optlen += TCPOLEN_TIMESTAMP; *optp++ = TCPOPT_TIMESTAMP; *optp++ = TCPOLEN_TIMESTAMP; to->to_tsval = htonl(to->to_tsval); to->to_tsecr = htonl(to->to_tsecr); bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval)); optp += sizeof(to->to_tsval); bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr)); optp += sizeof(to->to_tsecr); break; case TOF_SIGNATURE: { int siglen = TCPOLEN_SIGNATURE - 2; while (!optlen || optlen % 4 != 2) { optlen += TCPOLEN_NOP; *optp++ = TCPOPT_NOP; } if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) { to->to_flags &= ~TOF_SIGNATURE; continue; } optlen += TCPOLEN_SIGNATURE; *optp++ = TCPOPT_SIGNATURE; *optp++ = TCPOLEN_SIGNATURE; to->to_signature = optp; while (siglen--) *optp++ = 0; break; } case TOF_SACK: { int sackblks = 0; struct sackblk *sack = (struct sackblk *)to->to_sacks; tcp_seq sack_seq; while (!optlen || optlen % 4 != 2) { optlen += TCPOLEN_NOP; *optp++ = TCPOPT_NOP; } if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK) continue; optlen += TCPOLEN_SACKHDR; *optp++ = TCPOPT_SACK; sackblks = min(to->to_nsacks, (TCP_MAXOLEN - optlen) / TCPOLEN_SACK); *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK; while (sackblks--) { sack_seq = htonl(sack->start); bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); optp += sizeof(sack_seq); sack_seq = htonl(sack->end); bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq)); optp += sizeof(sack_seq); optlen += TCPOLEN_SACK; sack++; } TCPSTAT_INC(tcps_sack_send_blocks); break; } case TOF_FASTOPEN: { int total_len; /* XXX is there any point to aligning this option? */ total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len; if (TCP_MAXOLEN - optlen < total_len) { to->to_flags &= ~TOF_FASTOPEN; continue; } *optp++ = TCPOPT_FAST_OPEN; *optp++ = total_len; if (to->to_tfo_len > 0) { bcopy(to->to_tfo_cookie, optp, to->to_tfo_len); optp += to->to_tfo_len; } optlen += total_len; break; } default: panic("%s: unknown TCP option type", __func__); break; } } /* Terminate and pad TCP options to a 4 byte boundary. */ if (optlen % 4) { optlen += TCPOLEN_EOL; *optp++ = TCPOPT_EOL; } /* * According to RFC 793 (STD0007): * "The content of the header beyond the End-of-Option option * must be header padding (i.e., zero)." * and later: "The padding is composed of zeros." */ while (optlen % 4) { optlen += TCPOLEN_PAD; *optp++ = TCPOPT_PAD; } KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__)); return (optlen); } /* * This is a copy of m_copym(), taking the TSO segment size/limit * constraints into account, and advancing the sndptr as it goes. */ struct mbuf * tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen, int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls) { #ifdef KERN_TLS struct ktls_session *tls, *ntls; struct mbuf *start; #endif struct mbuf *n, **np; struct mbuf *top; int32_t off = off0; int32_t len = *plen; int32_t fragsize; int32_t len_cp = 0; int32_t *pkthdrlen; uint32_t mlen, frags; bool copyhdr; KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off)); KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len)); if (off == 0 && m->m_flags & M_PKTHDR) copyhdr = true; else copyhdr = false; while (off > 0) { KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain")); if (off < m->m_len) break; off -= m->m_len; if ((sb) && (m == sb->sb_sndptr)) { sb->sb_sndptroff += m->m_len; sb->sb_sndptr = m->m_next; } m = m->m_next; } np = ⊤ top = NULL; pkthdrlen = NULL; #ifdef KERN_TLS if (hw_tls && (m->m_flags & M_EXTPG)) tls = m->m_epg_tls; else tls = NULL; start = m; #endif while (len > 0) { if (m == NULL) { KASSERT(len == M_COPYALL, ("tcp_m_copym, length > size of mbuf chain")); *plen = len_cp; if (pkthdrlen != NULL) *pkthdrlen = len_cp; break; } #ifdef KERN_TLS if (hw_tls) { if (m->m_flags & M_EXTPG) ntls = m->m_epg_tls; else ntls = NULL; /* * Avoid mixing TLS records with handshake * data or TLS records from different * sessions. */ if (tls != ntls) { MPASS(m != start); *plen = len_cp; if (pkthdrlen != NULL) *pkthdrlen = len_cp; break; } } #endif mlen = min(len, m->m_len - off); if (seglimit) { /* * For M_EXTPG mbufs, add 3 segments * + 1 in case we are crossing page boundaries * + 2 in case the TLS hdr/trailer are used * It is cheaper to just add the segments * than it is to take the cache miss to look * at the mbuf ext_pgs state in detail. */ if (m->m_flags & M_EXTPG) { fragsize = min(segsize, PAGE_SIZE); frags = 3; } else { fragsize = segsize; frags = 0; } /* Break if we really can't fit anymore. */ if ((frags + 1) >= seglimit) { *plen = len_cp; if (pkthdrlen != NULL) *pkthdrlen = len_cp; break; } /* * Reduce size if you can't copy the whole * mbuf. If we can't copy the whole mbuf, also * adjust len so the loop will end after this * mbuf. */ if ((frags + howmany(mlen, fragsize)) >= seglimit) { mlen = (seglimit - frags - 1) * fragsize; len = mlen; *plen = len_cp + len; if (pkthdrlen != NULL) *pkthdrlen = *plen; } frags += howmany(mlen, fragsize); if (frags == 0) frags++; seglimit -= frags; KASSERT(seglimit > 0, ("%s: seglimit went too low", __func__)); } if (copyhdr) n = m_gethdr(M_NOWAIT, m->m_type); else n = m_get(M_NOWAIT, m->m_type); *np = n; if (n == NULL) goto nospace; if (copyhdr) { if (!m_dup_pkthdr(n, m, M_NOWAIT)) goto nospace; if (len == M_COPYALL) n->m_pkthdr.len -= off0; else n->m_pkthdr.len = len; pkthdrlen = &n->m_pkthdr.len; copyhdr = false; } n->m_len = mlen; len_cp += n->m_len; if (m->m_flags & (M_EXT|M_EXTPG)) { n->m_data = m->m_data + off; mb_dupcl(n, m); } else bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t), (u_int)n->m_len); if (sb && (sb->sb_sndptr == m) && ((n->m_len + off) >= m->m_len) && m->m_next) { sb->sb_sndptroff += m->m_len; sb->sb_sndptr = m->m_next; } off = 0; if (len != M_COPYALL) { len -= n->m_len; } m = m->m_next; np = &n->m_next; } return (top); nospace: m_freem(top); return (NULL); } void tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin) { /* * Automatic sizing of send socket buffer. Often the send buffer * size is not optimally adjusted to the actual network conditions * at hand (delay bandwidth product). Setting the buffer size too * small limits throughput on links with high bandwidth and high * delay (eg. trans-continental/oceanic links). Setting the * buffer size too big consumes too much real kernel memory, * especially with many connections on busy servers. * * The criteria to step up the send buffer one notch are: * 1. receive window of remote host is larger than send buffer * (with a fudge factor of 5/4th); * 2. send buffer is filled to 7/8th with data (so we actually * have data to make use of it); * 3. send buffer fill has not hit maximal automatic size; * 4. our send window (slow start and cogestion controlled) is * larger than sent but unacknowledged data in send buffer. * * The remote host receive window scaling factor may limit the * growing of the send buffer before it reaches its allowed * maximum. * * It scales directly with slow start or congestion window * and does at most one step per received ACK. This fast * scaling has the drawback of growing the send buffer beyond * what is strictly necessary to make full use of a given * delay*bandwidth product. However testing has shown this not * to be much of an problem. At worst we are trading wasting * of available bandwidth (the non-use of it) for wasting some * socket buffer memory. * * TODO: Shrink send buffer during idle periods together * with congestion window. Requires another timer. Has to * wait for upcoming tcp timer rewrite. * * XXXGL: should there be used sbused() or sbavail()? */ if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { int lowat; lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0; if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat && sbused(&so->so_snd) >= (so->so_snd.sb_hiwat / 8 * 7) - lowat && sbused(&so->so_snd) < V_tcp_autosndbuf_max && sendwin >= (sbused(&so->so_snd) - (tp->snd_nxt - tp->snd_una))) { if (!sbreserve_locked(&so->so_snd, min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc, V_tcp_autosndbuf_max), so, curthread)) so->so_snd.sb_flags &= ~SB_AUTOSIZE; } } } diff --git a/sys/netinet/tcp_stacks/bbr.c b/sys/netinet/tcp_stacks/bbr.c index 673dee911c87..febac7ad424c 100644 --- a/sys/netinet/tcp_stacks/bbr.c +++ b/sys/netinet/tcp_stacks/bbr.c @@ -1,14982 +1,14952 @@ /*- * Copyright (c) 2016-2020 Netflix, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * */ /** * Author: Randall Stewart * This work is based on the ACM Queue paper * BBR - Congestion Based Congestion Control * and also numerous discussions with Neal, Yuchung and Van. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_tcpdebug.h" #include "opt_ratelimit.h" #include #include #include #include #include #ifdef TCP_HHOOK #include #endif #include #include #include #include #include #include #include #ifdef STATS #include #include #include /* Must come after qmath.h and tree.h */ #endif #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define TCPSTATES /* for logging */ #include #include #include #include #include /* required for icmp_var.h */ #include /* for ICMP_BANDLIM */ #include #include #include #include #define TCPOUTFLAGS #include #include #include #include #include #include #include #include #include #include #include #ifdef TCPDEBUG #include #endif /* TCPDEBUG */ #ifdef TCP_OFFLOAD #include #endif #ifdef INET6 #include #endif #include #include #include #include #include #if defined(IPSEC) || defined(IPSEC_SUPPORT) #include #include #endif /* IPSEC */ #include #include #include #ifdef MAC #include #endif #include "sack_filter.h" #include "tcp_bbr.h" #include "rack_bbr_common.h" uma_zone_t bbr_zone; uma_zone_t bbr_pcb_zone; struct sysctl_ctx_list bbr_sysctl_ctx; struct sysctl_oid *bbr_sysctl_root; #define TCPT_RANGESET_NOSLOP(tv, value, tvmin, tvmax) do { \ (tv) = (value); \ if ((u_long)(tv) < (u_long)(tvmin)) \ (tv) = (tvmin); \ if ((u_long)(tv) > (u_long)(tvmax)) \ (tv) = (tvmax); \ } while(0) /*#define BBR_INVARIANT 1*/ /* * initial window */ static uint32_t bbr_def_init_win = 10; static int32_t bbr_persist_min = 250000; /* 250ms */ static int32_t bbr_persist_max = 1000000; /* 1 Second */ static int32_t bbr_cwnd_may_shrink = 0; static int32_t bbr_cwndtarget_rtt_touse = BBR_RTT_PROP; static int32_t bbr_num_pktepo_for_del_limit = BBR_NUM_RTTS_FOR_DEL_LIMIT; static int32_t bbr_hardware_pacing_limit = 8000; static int32_t bbr_quanta = 3; /* How much extra quanta do we get? */ static int32_t bbr_no_retran = 0; static int32_t bbr_error_base_paceout = 10000; /* usec to pace */ static int32_t bbr_max_net_error_cnt = 10; /* Should the following be dynamic too -- loss wise */ static int32_t bbr_rtt_gain_thresh = 0; /* Measurement controls */ static int32_t bbr_use_google_algo = 1; static int32_t bbr_ts_limiting = 1; static int32_t bbr_ts_can_raise = 0; static int32_t bbr_do_red = 600; static int32_t bbr_red_scale = 20000; static int32_t bbr_red_mul = 1; static int32_t bbr_red_div = 2; static int32_t bbr_red_growth_restrict = 1; static int32_t bbr_target_is_bbunit = 0; static int32_t bbr_drop_limit = 0; /* * How much gain do we need to see to * stay in startup? */ static int32_t bbr_marks_rxt_sack_passed = 0; static int32_t bbr_start_exit = 25; static int32_t bbr_low_start_exit = 25; /* When we are in reduced gain */ static int32_t bbr_startup_loss_thresh = 2000; /* 20.00% loss */ static int32_t bbr_hptsi_max_mul = 1; /* These two mul/div assure a min pacing */ static int32_t bbr_hptsi_max_div = 2; /* time, 0 means turned off. We need this * if we go back ever to where the pacer * has priority over timers. */ static int32_t bbr_policer_call_from_rack_to = 0; static int32_t bbr_policer_detection_enabled = 1; static int32_t bbr_min_measurements_req = 1; /* We need at least 2 * measurments before we are * "good" note that 2 == 1. * This is because we use a > * comparison. This means if * min_measure was 0, it takes * num-measures > min(0) and * you get 1 measurement and * you are good. Set to 1, you * have to have two * measurements (this is done * to prevent it from being ok * to have no measurements). */ static int32_t bbr_no_pacing_until = 4; static int32_t bbr_min_usec_delta = 20000; /* 20,000 usecs */ static int32_t bbr_min_peer_delta = 20; /* 20 units */ static int32_t bbr_delta_percent = 150; /* 15.0 % */ static int32_t bbr_target_cwnd_mult_limit = 8; /* * bbr_cwnd_min_val is the number of * segments we hold to in the RTT probe * state typically 4. */ static int32_t bbr_cwnd_min_val = BBR_PROBERTT_NUM_MSS; static int32_t bbr_cwnd_min_val_hs = BBR_HIGHSPEED_NUM_MSS; static int32_t bbr_gain_to_target = 1; static int32_t bbr_gain_gets_extra_too = 1; /* * bbr_high_gain is the 2/ln(2) value we need * to double the sending rate in startup. This * is used for both cwnd and hptsi gain's. */ static int32_t bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1; static int32_t bbr_startup_lower = BBR_UNIT * 1500 / 1000 + 1; static int32_t bbr_use_lower_gain_in_startup = 1; /* thresholds for reduction on drain in sub-states/drain */ static int32_t bbr_drain_rtt = BBR_SRTT; static int32_t bbr_drain_floor = 88; static int32_t google_allow_early_out = 1; static int32_t google_consider_lost = 1; static int32_t bbr_drain_drop_mul = 4; static int32_t bbr_drain_drop_div = 5; static int32_t bbr_rand_ot = 50; static int32_t bbr_can_force_probertt = 0; static int32_t bbr_can_adjust_probertt = 1; static int32_t bbr_probertt_sets_rtt = 0; static int32_t bbr_can_use_ts_for_rtt = 1; static int32_t bbr_is_ratio = 0; static int32_t bbr_sub_drain_app_limit = 1; static int32_t bbr_prtt_slam_cwnd = 1; static int32_t bbr_sub_drain_slam_cwnd = 1; static int32_t bbr_slam_cwnd_in_main_drain = 1; static int32_t bbr_filter_len_sec = 6; /* How long does the rttProp filter * hold */ static uint32_t bbr_rtt_probe_limit = (USECS_IN_SECOND * 4); /* * bbr_drain_gain is the reverse of the high_gain * designed to drain back out the standing queue * that is formed in startup by causing a larger * hptsi gain and thus drainging the packets * in flight. */ static int32_t bbr_drain_gain = BBR_UNIT * 1000 / 2885; static int32_t bbr_rttprobe_gain = 192; /* * The cwnd_gain is the default cwnd gain applied when * calculating a target cwnd. Note that the cwnd is * a secondary factor in the way BBR works (see the * paper and think about it, it will take some time). * Basically the hptsi_gain spreads the packets out * so you never get more than BDP to the peer even * if the cwnd is high. In our implemenation that * means in non-recovery/retransmission scenarios * cwnd will never be reached by the flight-size. */ static int32_t bbr_cwnd_gain = BBR_UNIT * 2; static int32_t bbr_tlp_type_to_use = BBR_SRTT; static int32_t bbr_delack_time = 100000; /* 100ms in useconds */ static int32_t bbr_sack_not_required = 0; /* set to one to allow non-sack to use bbr */ static int32_t bbr_initial_bw_bps = 62500; /* 500kbps in bytes ps */ static int32_t bbr_ignore_data_after_close = 1; static int16_t bbr_hptsi_gain[] = { (BBR_UNIT *5 / 4), (BBR_UNIT * 3 / 4), BBR_UNIT, BBR_UNIT, BBR_UNIT, BBR_UNIT, BBR_UNIT, BBR_UNIT }; int32_t bbr_use_rack_resend_cheat = 1; int32_t bbr_sends_full_iwnd = 1; #define BBR_HPTSI_GAIN_MAX 8 /* * The BBR module incorporates a number of * TCP ideas that have been put out into the IETF * over the last few years: * - Yuchung Cheng's RACK TCP (for which its named) that * will stop us using the number of dup acks and instead * use time as the gage of when we retransmit. * - Reorder Detection of RFC4737 and the Tail-Loss probe draft * of Dukkipati et.al. * - Van Jacobson's et.al BBR. * * RACK depends on SACK, so if an endpoint arrives that * cannot do SACK the state machine below will shuttle the * connection back to using the "default" TCP stack that is * in FreeBSD. * * To implement BBR and RACK the original TCP stack was first decomposed * into a functional state machine with individual states * for each of the possible TCP connection states. The do_segement * functions role in life is to mandate the connection supports SACK * initially and then assure that the RACK state matches the conenction * state before calling the states do_segment function. Data processing * of inbound segments also now happens in the hpts_do_segment in general * with only one exception. This is so we can keep the connection on * a single CPU. * * Each state is simplified due to the fact that the original do_segment * has been decomposed and we *know* what state we are in (no * switches on the state) and all tests for SACK are gone. This * greatly simplifies what each state does. * * TCP output is also over-written with a new version since it * must maintain the new rack scoreboard and has had hptsi * integrated as a requirment. Still todo is to eliminate the * use of the callout_() system and use the hpts for all * timers as well. */ static uint32_t bbr_rtt_probe_time = 200000; /* 200ms in micro seconds */ static uint32_t bbr_rtt_probe_cwndtarg = 4; /* How many mss's outstanding */ static const int32_t bbr_min_req_free = 2; /* The min we must have on the * free list */ static int32_t bbr_tlp_thresh = 1; static int32_t bbr_reorder_thresh = 2; static int32_t bbr_reorder_fade = 60000000; /* 0 - never fade, def * 60,000,000 - 60 seconds */ static int32_t bbr_pkt_delay = 1000; static int32_t bbr_min_to = 1000; /* Number of usec's minimum timeout */ static int32_t bbr_incr_timers = 1; static int32_t bbr_tlp_min = 10000; /* 10ms in usecs */ static int32_t bbr_delayed_ack_time = 200000; /* 200ms in usecs */ static int32_t bbr_exit_startup_at_loss = 1; /* * bbr_lt_bw_ratio is 1/8th * bbr_lt_bw_diff is < 4 Kbit/sec */ static uint64_t bbr_lt_bw_diff = 4000 / 8; /* In bytes per second */ static uint64_t bbr_lt_bw_ratio = 8; /* For 1/8th */ static uint32_t bbr_lt_bw_max_rtts = 48; /* How many rtt's do we use * the lt_bw for */ static uint32_t bbr_lt_intvl_min_rtts = 4; /* Min num of RTT's to measure * lt_bw */ static int32_t bbr_lt_intvl_fp = 0; /* False positive epoch diff */ static int32_t bbr_lt_loss_thresh = 196; /* Lost vs delivered % */ static int32_t bbr_lt_fd_thresh = 100; /* false detection % */ static int32_t bbr_verbose_logging = 0; /* * Currently regular tcp has a rto_min of 30ms * the backoff goes 12 times so that ends up * being a total of 122.850 seconds before a * connection is killed. */ static int32_t bbr_rto_min_ms = 30; /* 30ms same as main freebsd */ static int32_t bbr_rto_max_sec = 4; /* 4 seconds */ /****************************************************/ /* DEFAULT TSO SIZING (cpu performance impacting) */ /****************************************************/ /* What amount is our formula using to get TSO size */ static int32_t bbr_hptsi_per_second = 1000; /* * For hptsi under bbr_cross_over connections what is delay * target 7ms (in usec) combined with a seg_max of 2 * gets us close to identical google behavior in * TSO size selection (possibly more 1MSS sends). */ static int32_t bbr_hptsi_segments_delay_tar = 7000; /* Does pacing delay include overhead's in its time calculations? */ static int32_t bbr_include_enet_oh = 0; static int32_t bbr_include_ip_oh = 1; static int32_t bbr_include_tcp_oh = 1; static int32_t bbr_google_discount = 10; /* Do we use (nf mode) pkt-epoch to drive us or rttProp? */ static int32_t bbr_state_is_pkt_epoch = 0; static int32_t bbr_state_drain_2_tar = 1; /* What is the max the 0 - bbr_cross_over MBPS TSO target * can reach using our delay target. Note that this * value becomes the floor for the cross over * algorithm. */ static int32_t bbr_hptsi_segments_max = 2; static int32_t bbr_hptsi_segments_floor = 1; static int32_t bbr_hptsi_utter_max = 0; /* What is the min the 0 - bbr_cross-over MBPS TSO target can be */ static int32_t bbr_hptsi_bytes_min = 1460; static int32_t bbr_all_get_min = 0; /* Cross over point from algo-a to algo-b */ static uint32_t bbr_cross_over = TWENTY_THREE_MBPS; /* Do we deal with our restart state? */ static int32_t bbr_uses_idle_restart = 0; static int32_t bbr_idle_restart_threshold = 100000; /* 100ms in useconds */ /* Do we allow hardware pacing? */ static int32_t bbr_allow_hdwr_pacing = 0; static int32_t bbr_hdwr_pace_adjust = 2; /* multipler when we calc the tso size */ static int32_t bbr_hdwr_pace_floor = 1; static int32_t bbr_hdwr_pacing_delay_cnt = 10; /****************************************************/ static int32_t bbr_resends_use_tso = 0; static int32_t bbr_tlp_max_resend = 2; static int32_t bbr_sack_block_limit = 128; #define BBR_MAX_STAT 19 counter_u64_t bbr_state_time[BBR_MAX_STAT]; counter_u64_t bbr_state_lost[BBR_MAX_STAT]; counter_u64_t bbr_state_resend[BBR_MAX_STAT]; counter_u64_t bbr_stat_arry[BBR_STAT_SIZE]; counter_u64_t bbr_opts_arry[BBR_OPTS_SIZE]; counter_u64_t bbr_out_size[TCP_MSS_ACCT_SIZE]; counter_u64_t bbr_flows_whdwr_pacing; counter_u64_t bbr_flows_nohdwr_pacing; counter_u64_t bbr_nohdwr_pacing_enobuf; counter_u64_t bbr_hdwr_pacing_enobuf; static inline uint64_t bbr_get_bw(struct tcp_bbr *bbr); /* * Static defintions we need for forward declarations. */ static uint32_t bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw); static uint32_t bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain); static void bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win); static void bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses); static void bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int line, int dolog); static uint32_t bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain); static void bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch, uint32_t losses); static uint32_t bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm); static uint32_t bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp); static uint32_t bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t srtt, uint32_t cts); static void bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line); static void bbr_set_state_target(struct tcp_bbr *bbr, int line); static void bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line); static void bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line); static void tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts); static void bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts); static void bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, uint32_t rtt, uint32_t line, uint8_t is_start, uint16_t set); static struct bbr_sendmap * bbr_find_lowest_rsm(struct tcp_bbr *bbr); static __inline uint32_t bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type); static void bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which); static void bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt, uint32_t thresh, uint32_t to); static void bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag); static void bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot, uint32_t del_by, uint32_t cts, uint32_t sloton, uint32_t prev_delay); static void bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line); static void bbr_stop_all_timers(struct tcpcb *tp); static void bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts); static void bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts); static void bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts); static void bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len, uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod); static inline uint8_t bbr_state_val(struct tcp_bbr *bbr) { return(bbr->rc_bbr_substate); } static inline uint32_t get_min_cwnd(struct tcp_bbr *bbr) { int mss; mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); if (bbr_get_rtt(bbr, BBR_RTT_PROP) < BBR_HIGH_SPEED) return (bbr_cwnd_min_val_hs * mss); else return (bbr_cwnd_min_val * mss); } static uint32_t bbr_get_persists_timer_val(struct tcpcb *tp, struct tcp_bbr *bbr) { uint64_t srtt, var; uint64_t ret_val; bbr->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT; if (tp->t_srtt == 0) { srtt = (uint64_t)BBR_INITIAL_RTO; var = 0; } else { srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); var = ((uint64_t)TICKS_2_USEC(tp->t_rttvar) >> TCP_RTT_SHIFT); } TCPT_RANGESET_NOSLOP(ret_val, ((srtt + var) * tcp_backoff[tp->t_rxtshift]), bbr_persist_min, bbr_persist_max); return ((uint32_t)ret_val); } static uint32_t bbr_timer_start(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { /* * Start the FR timer, we do this based on getting the first one in * the rc_tmap. Note that if its NULL we must stop the timer. in all * events we need to stop the running timer (if its running) before * starting the new one. */ uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse; int32_t idx; int32_t is_tlp_timer = 0; struct bbr_sendmap *rsm; if (bbr->rc_all_timers_stopped) { /* All timers have been stopped none are to run */ return (0); } if (bbr->rc_in_persist) { /* We can't start any timer in persists */ return (bbr_get_persists_timer_val(tp, bbr)); } rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); if ((rsm == NULL) || ((tp->t_flags & TF_SACK_PERMIT) == 0) || (tp->t_state < TCPS_ESTABLISHED)) { /* Nothing on the send map */ activate_rxt: if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) { uint64_t tov; time_since_sent = 0; rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); if (rsm) { idx = rsm->r_rtr_cnt - 1; if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time)) tstmp_touse = rsm->r_tim_lastsent[idx]; else tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time; if (TSTMP_GT(tstmp_touse, cts)) time_since_sent = cts - tstmp_touse; } bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RXT; if (tp->t_srtt == 0) tov = BBR_INITIAL_RTO; else tov = ((uint64_t)(TICKS_2_USEC(tp->t_srtt) + ((uint64_t)TICKS_2_USEC(tp->t_rttvar) * (uint64_t)4)) >> TCP_RTT_SHIFT); if (tp->t_rxtshift) tov *= tcp_backoff[tp->t_rxtshift]; if (tov > time_since_sent) tov -= time_since_sent; else tov = bbr->r_ctl.rc_min_to; TCPT_RANGESET_NOSLOP(to, tov, (bbr->r_ctl.rc_min_rto_ms * MS_IN_USEC), (bbr->rc_max_rto_sec * USECS_IN_SECOND)); bbr_log_timer_var(bbr, 2, cts, 0, srtt, 0, to); return (to); } return (0); } if (rsm->r_flags & BBR_ACKED) { rsm = bbr_find_lowest_rsm(bbr); if (rsm == NULL) { /* No lowest? */ goto activate_rxt; } } /* Convert from ms to usecs */ if (rsm->r_flags & BBR_SACK_PASSED) { if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) == 1) && (rsm->r_flags & BBR_HAS_FIN)) { /* * We don't start a bbr rack timer if all we have is * a FIN outstanding. */ goto activate_rxt; } srtt = bbr_get_rtt(bbr, BBR_RTT_RACK); thresh = bbr_calc_thresh_rack(bbr, srtt, cts, rsm); idx = rsm->r_rtr_cnt - 1; exp = rsm->r_tim_lastsent[idx] + thresh; if (SEQ_GEQ(exp, cts)) { to = exp - cts; if (to < bbr->r_ctl.rc_min_to) { to = bbr->r_ctl.rc_min_to; } } else { to = bbr->r_ctl.rc_min_to; } } else { /* Ok we need to do a TLP not RACK */ if (bbr->rc_tlp_in_progress != 0) { /* * The previous send was a TLP. */ goto activate_rxt; } rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext); if (rsm == NULL) { /* We found no rsm to TLP with. */ goto activate_rxt; } if (rsm->r_flags & BBR_HAS_FIN) { /* If its a FIN we don't do TLP */ rsm = NULL; goto activate_rxt; } time_since_sent = 0; idx = rsm->r_rtr_cnt - 1; if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time)) tstmp_touse = rsm->r_tim_lastsent[idx]; else tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time; if (TSTMP_GT(tstmp_touse, cts)) time_since_sent = cts - tstmp_touse; is_tlp_timer = 1; srtt = bbr_get_rtt(bbr, bbr_tlp_type_to_use); thresh = bbr_calc_thresh_tlp(tp, bbr, rsm, srtt, cts); if (thresh > time_since_sent) to = thresh - time_since_sent; else to = bbr->r_ctl.rc_min_to; if (to > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { /* * If the TLP time works out to larger than the max * RTO lets not do TLP.. just RTO. */ goto activate_rxt; } if ((bbr->rc_tlp_rtx_out == 1) && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq)) { /* * Second retransmit of the same TLP * lets not. */ bbr->rc_tlp_rtx_out = 0; goto activate_rxt; } if (rsm->r_start != bbr->r_ctl.rc_last_tlp_seq) { /* * The tail is no longer the last one I did a probe * on */ bbr->r_ctl.rc_tlp_seg_send_cnt = 0; bbr->r_ctl.rc_last_tlp_seq = rsm->r_start; } } if (is_tlp_timer == 0) { BBR_STAT_INC(bbr_to_arm_rack); bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RACK; } else { bbr_log_timer_var(bbr, 1, cts, time_since_sent, srtt, thresh, to); if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) { /* * We have exceeded how many times we can retran the * current TLP timer, switch to the RTO timer. */ goto activate_rxt; } else { BBR_STAT_INC(bbr_to_arm_tlp); bbr->r_ctl.rc_hpts_flags |= PACE_TMR_TLP; } } return (to); } static inline int32_t bbr_minseg(struct tcp_bbr *bbr) { return (bbr->r_ctl.rc_pace_min_segs - bbr->rc_last_options); } static void bbr_start_hpts_timer(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts, int32_t frm, int32_t slot, uint32_t tot_len) { struct inpcb *inp; struct hpts_diag diag; uint32_t delayed_ack = 0; uint32_t left = 0; uint32_t hpts_timeout; uint8_t stopped; int32_t delay_calc = 0; uint32_t prev_delay = 0; inp = tp->t_inpcb; if (inp->inp_in_hpts) { /* A previous call is already set up */ return; } if ((tp->t_state == TCPS_CLOSED) || (tp->t_state == TCPS_LISTEN)) { return; } stopped = bbr->rc_tmr_stopped; if (stopped && TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) { left = bbr->r_ctl.rc_timer_exp - cts; } bbr->r_ctl.rc_hpts_flags = 0; bbr->r_ctl.rc_timer_exp = 0; prev_delay = bbr->r_ctl.rc_last_delay_val; if (bbr->r_ctl.rc_last_delay_val && (slot == 0)) { /* * If a previous pacer delay was in place we * are not coming from the output side (where * we calculate a delay, more likely a timer). */ slot = bbr->r_ctl.rc_last_delay_val; if (TSTMP_GT(cts, bbr->rc_pacer_started)) { /* Compensate for time passed */ delay_calc = cts - bbr->rc_pacer_started; if (delay_calc <= slot) slot -= delay_calc; } } /* Do we have early to make up for by pushing out the pacing time? */ if (bbr->r_agg_early_set) { bbr_log_pacing_delay_calc(bbr, 0, bbr->r_ctl.rc_agg_early, cts, slot, 0, bbr->r_agg_early_set, 2); slot += bbr->r_ctl.rc_agg_early; bbr->r_ctl.rc_agg_early = 0; bbr->r_agg_early_set = 0; } /* Are we running a total debt that needs to be compensated for? */ if (bbr->r_ctl.rc_hptsi_agg_delay) { if (slot > bbr->r_ctl.rc_hptsi_agg_delay) { /* We nuke the delay */ slot -= bbr->r_ctl.rc_hptsi_agg_delay; bbr->r_ctl.rc_hptsi_agg_delay = 0; } else { /* We nuke some of the delay, put in a minimal 100usecs */ bbr->r_ctl.rc_hptsi_agg_delay -= slot; bbr->r_ctl.rc_last_delay_val = slot = 100; } } bbr->r_ctl.rc_last_delay_val = slot; hpts_timeout = bbr_timer_start(tp, bbr, cts); if (tp->t_flags & TF_DELACK) { if (bbr->rc_in_persist == 0) { delayed_ack = bbr_delack_time; } else { /* * We are in persists and have * gotten a new data element. */ if (hpts_timeout > bbr_delack_time) { /* * Lets make the persists timer (which acks) * be the smaller of hpts_timeout and bbr_delack_time. */ hpts_timeout = bbr_delack_time; } } } if (delayed_ack && ((hpts_timeout == 0) || (delayed_ack < hpts_timeout))) { /* We need a Delayed ack timer */ bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; hpts_timeout = delayed_ack; } if (slot) { /* Mark that we have a pacing timer up */ BBR_STAT_INC(bbr_paced_segments); bbr->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT; } /* * If no timers are going to run and we will fall off thfe hptsi * wheel, we resort to a keep-alive timer if its configured. */ if ((hpts_timeout == 0) && (slot == 0)) { if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)) { /* * Ok we have no timer (persists, rack, tlp, rxt or * del-ack), we don't have segments being paced. So * all that is left is the keepalive timer. */ if (TCPS_HAVEESTABLISHED(tp->t_state)) { hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp)); } else { hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp)); } bbr->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP; } } if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) == (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) { /* * RACK, TLP, persists and RXT timers all are restartable * based on actions input .. i.e we received a packet (ack * or sack) and that changes things (rw, or snd_una etc). * Thus we can restart them with a new value. For * keep-alive, delayed_ack we keep track of what was left * and restart the timer with a smaller value. */ if (left < hpts_timeout) hpts_timeout = left; } if (bbr->r_ctl.rc_incr_tmrs && slot && (bbr->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) { /* * If configured to do so, and the timer is either * the TLP or RXT timer, we need to increase the timeout * by the pacing time. Consider the bottleneck at my * machine as an example, we are sending something * to start a TLP on. The last packet won't be emitted * fully until the pacing time (the bottleneck will hold * the data in place). Once the packet is emitted that * is when we want to start waiting for the TLP. This * is most evident with hardware pacing (where the nic * is holding the packet(s) before emitting). But it * can also show up in the network so we do it for all * cases. Technically we would take off one packet from * this extra delay but this is easier and being more * conservative is probably better. */ hpts_timeout += slot; } if (hpts_timeout) { /* * Hack alert for now we can't time-out over 2147 seconds (a * bit more than 35min) */ if (hpts_timeout > 0x7ffffffe) hpts_timeout = 0x7ffffffe; bbr->r_ctl.rc_timer_exp = cts + hpts_timeout; } else bbr->r_ctl.rc_timer_exp = 0; if ((slot) && (bbr->rc_use_google || bbr->output_error_seen || (slot <= hpts_timeout)) ) { /* * Tell LRO that it can queue packets while * we pace. */ bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY; if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && (bbr->rc_cwnd_limited == 0)) { /* * If we are not cwnd limited and we * are running a rack timer we put on * the do not disturbe even for sack. */ inp->inp_flags2 |= INP_DONT_SACK_QUEUE; } else inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; bbr->rc_pacer_started = cts; (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot), __LINE__, &diag); bbr->rc_timer_first = 0; bbr->bbr_timer_src = frm; bbr_log_to_start(bbr, cts, hpts_timeout, slot, 1); bbr_log_hpts_diag(bbr, cts, &diag); } else if (hpts_timeout) { (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout), __LINE__, &diag); /* * We add the flag here as well if the slot is set, * since hpts will call in to clear the queue first before * calling the output routine (which does our timers). * We don't want to set the flag if its just a timer * else the arrival of data might (that causes us * to send more) might get delayed. Imagine being * on a keep-alive timer and a request comes in for * more data. */ if (slot) bbr->rc_pacer_started = cts; if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && (bbr->rc_cwnd_limited == 0)) { /* * For a rack timer, don't wake us even * if a sack arrives as long as we are * not cwnd limited. */ bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY; inp->inp_flags2 |= INP_DONT_SACK_QUEUE; } else { /* All other timers wake us up */ bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY; inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; } bbr->bbr_timer_src = frm; bbr_log_to_start(bbr, cts, hpts_timeout, slot, 0); bbr_log_hpts_diag(bbr, cts, &diag); bbr->rc_timer_first = 1; } bbr->rc_tmr_stopped = 0; bbr_log_type_bbrsnd(bbr, tot_len, slot, delay_calc, cts, frm, prev_delay); } static void bbr_timer_audit(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, struct sockbuf *sb) { /* * We received an ack, and then did not call send or were bounced * out due to the hpts was running. Now a timer is up as well, is it * the right timer? */ struct inpcb *inp; struct bbr_sendmap *rsm; uint32_t hpts_timeout; int tmr_up; tmr_up = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK; if (bbr->rc_in_persist && (tmr_up == PACE_TMR_PERSIT)) return; rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) && (tmr_up == PACE_TMR_RXT)) { /* Should be an RXT */ return; } inp = bbr->rc_inp; if (rsm == NULL) { /* Nothing outstanding? */ if (tp->t_flags & TF_DELACK) { if (tmr_up == PACE_TMR_DELACK) /* * We are supposed to have delayed ack up * and we do */ return; } else if (sbavail(&inp->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) { /* * if we hit enobufs then we would expect the * possiblity of nothing outstanding and the RXT up * (and the hptsi timer). */ return; } else if (((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)) && (tmr_up == PACE_TMR_KEEP) && (tp->snd_max == tp->snd_una)) { /* We should have keep alive up and we do */ return; } } if (rsm && (rsm->r_flags & BBR_SACK_PASSED)) { if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) == 1) && (rsm->r_flags & BBR_HAS_FIN)) { /* needs to be a RXT */ if (tmr_up == PACE_TMR_RXT) return; else goto wrong_timer; } else if (tmr_up == PACE_TMR_RACK) return; else goto wrong_timer; } else if (rsm && (tmr_up == PACE_TMR_RACK)) { /* Rack timer has priority if we have data out */ return; } else if (SEQ_GT(tp->snd_max, tp->snd_una) && ((tmr_up == PACE_TMR_TLP) || (tmr_up == PACE_TMR_RXT))) { /* * Either a TLP or RXT is fine if no sack-passed is in place * and data is outstanding. */ return; } else if (tmr_up == PACE_TMR_DELACK) { /* * If the delayed ack was going to go off before the * rtx/tlp/rack timer were going to expire, then that would * be the timer in control. Note we don't check the time * here trusting the code is correct. */ return; } if (SEQ_GT(tp->snd_max, tp->snd_una) && ((tmr_up == PACE_TMR_RXT) || (tmr_up == PACE_TMR_TLP) || (tmr_up == PACE_TMR_RACK))) { /* * We have outstanding data and * we *do* have a RACK, TLP or RXT * timer running. We won't restart * anything here since thats probably ok we * will get called with some timer here shortly. */ return; } /* * Ok the timer originally started is not what we want now. We will * force the hpts to be stopped if any, and restart with the slot * set to what was in the saved slot. */ wrong_timer: if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) { if (inp->inp_in_hpts) tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT); bbr_timer_cancel(bbr, __LINE__, cts); bbr_start_hpts_timer(bbr, tp, cts, 1, bbr->r_ctl.rc_last_delay_val, 0); } else { /* * Output is hptsi so we just need to switch the type of * timer. We don't bother with keep-alive, since when we * jump through the output, it will start the keep-alive if * nothing is sent. * * We only need a delayed-ack added and or the hpts_timeout. */ hpts_timeout = bbr_timer_start(tp, bbr, cts); if (tp->t_flags & TF_DELACK) { if (hpts_timeout == 0) { hpts_timeout = bbr_delack_time; bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; } else if (hpts_timeout > bbr_delack_time) { hpts_timeout = bbr_delack_time; bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; } } if (hpts_timeout) { if (hpts_timeout > 0x7ffffffe) hpts_timeout = 0x7ffffffe; bbr->r_ctl.rc_timer_exp = cts + hpts_timeout; } } } int32_t bbr_clear_lost = 0; /* * Considers the two time values now (cts) and earlier. * If cts is smaller than earlier, we could have * had a sequence wrap (our counter wraps every * 70 min or so) or it could be just clock skew * getting us two differnt time values. Clock skew * will show up within 10ms or so. So in such * a case (where cts is behind earlier time by * less than 10ms) we return 0. Otherwise we * return the true difference between them. */ static inline uint32_t bbr_calc_time(uint32_t cts, uint32_t earlier_time) { /* * Given two timestamps, the current time stamp cts, and some other * time-stamp taken in theory earlier return the difference. The * trick is here sometimes locking will get the other timestamp * after the cts. If this occurs we need to return 0. */ if (TSTMP_GEQ(cts, earlier_time)) return (cts - earlier_time); /* * cts is behind earlier_time if its less than 10ms consider it 0. * If its more than 10ms difference then we had a time wrap. Else * its just the normal locking foo. I wonder if we should not go to * 64bit TS and get rid of this issue. */ if (TSTMP_GEQ((cts + 10000), earlier_time)) return (0); /* * Ok the time must have wrapped. So we need to answer a large * amount of time, which the normal subtraction should do. */ return (cts - earlier_time); } static int sysctl_bbr_clear_lost(SYSCTL_HANDLER_ARGS) { uint32_t stat; int32_t error; error = SYSCTL_OUT(req, &bbr_clear_lost, sizeof(uint32_t)); if (error || req->newptr == NULL) return error; error = SYSCTL_IN(req, &stat, sizeof(uint32_t)); if (error) return (error); if (stat == 1) { #ifdef BBR_INVARIANTS printf("Clearing BBR lost counters\n"); #endif COUNTER_ARRAY_ZERO(bbr_state_lost, BBR_MAX_STAT); COUNTER_ARRAY_ZERO(bbr_state_time, BBR_MAX_STAT); COUNTER_ARRAY_ZERO(bbr_state_resend, BBR_MAX_STAT); } else if (stat == 2) { #ifdef BBR_INVARIANTS printf("Clearing BBR option counters\n"); #endif COUNTER_ARRAY_ZERO(bbr_opts_arry, BBR_OPTS_SIZE); } else if (stat == 3) { #ifdef BBR_INVARIANTS printf("Clearing BBR stats counters\n"); #endif COUNTER_ARRAY_ZERO(bbr_stat_arry, BBR_STAT_SIZE); } else if (stat == 4) { #ifdef BBR_INVARIANTS printf("Clearing BBR out-size counters\n"); #endif COUNTER_ARRAY_ZERO(bbr_out_size, TCP_MSS_ACCT_SIZE); } bbr_clear_lost = 0; return (0); } static void bbr_init_sysctls(void) { struct sysctl_oid *bbr_probertt; struct sysctl_oid *bbr_hptsi; struct sysctl_oid *bbr_measure; struct sysctl_oid *bbr_cwnd; struct sysctl_oid *bbr_timeout; struct sysctl_oid *bbr_states; struct sysctl_oid *bbr_startup; struct sysctl_oid *bbr_policer; /* Probe rtt controls */ bbr_probertt = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "probertt", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, ""); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "gain", CTLFLAG_RW, &bbr_rttprobe_gain, 192, "What is the filter gain drop in probe_rtt (0=disable)?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "cwnd", CTLFLAG_RW, &bbr_rtt_probe_cwndtarg, 4, "How many mss's are outstanding during probe-rtt"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "int", CTLFLAG_RW, &bbr_rtt_probe_limit, 4000000, "If RTT has not shrank in this many micro-seconds enter probe-rtt"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "mintime", CTLFLAG_RW, &bbr_rtt_probe_time, 200000, "How many microseconds in probe-rtt"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "filter_len_sec", CTLFLAG_RW, &bbr_filter_len_sec, 6, "How long in seconds does the rttProp filter run?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "drain_rtt", CTLFLAG_RW, &bbr_drain_rtt, BBR_SRTT, "What is the drain rtt to use in probeRTT (rtt_prop=0, rtt_rack=1, rtt_pkt=2, rtt_srtt=3?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "can_force", CTLFLAG_RW, &bbr_can_force_probertt, 0, "If we keep setting new low rtt's but delay going in probe-rtt can we force in??"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "enter_sets_force", CTLFLAG_RW, &bbr_probertt_sets_rtt, 0, "In NF mode, do we imitate google_mode and set the rttProp on entry to probe-rtt?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "can_adjust", CTLFLAG_RW, &bbr_can_adjust_probertt, 1, "Can we dynamically adjust the probe-rtt limits and times?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "is_ratio", CTLFLAG_RW, &bbr_is_ratio, 0, "is the limit to filter a ratio?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "use_cwnd", CTLFLAG_RW, &bbr_prtt_slam_cwnd, 0, "Should we set/recover cwnd?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "can_use_ts", CTLFLAG_RW, &bbr_can_use_ts_for_rtt, 1, "Can we use the ms timestamp if available for retransmistted rtt calculations?"); /* Pacing controls */ bbr_hptsi = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "pacing", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, ""); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "hw_pacing", CTLFLAG_RW, &bbr_allow_hdwr_pacing, 1, "Do we allow hardware pacing?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "hw_pacing_limit", CTLFLAG_RW, &bbr_hardware_pacing_limit, 4000, "Do we have a limited number of connections for pacing chelsio (0=no limit)?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "hw_pacing_adj", CTLFLAG_RW, &bbr_hdwr_pace_adjust, 2, "Multiplier to calculated tso size?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "hw_pacing_floor", CTLFLAG_RW, &bbr_hdwr_pace_floor, 1, "Do we invoke the hardware pacing floor?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "hw_pacing_delay_cnt", CTLFLAG_RW, &bbr_hdwr_pacing_delay_cnt, 10, "How many packets must be sent after hdwr pacing is enabled"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "bw_cross", CTLFLAG_RW, &bbr_cross_over, 3000000, "What is the point where we cross over to linux like TSO size set"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "seg_deltarg", CTLFLAG_RW, &bbr_hptsi_segments_delay_tar, 7000, "What is the worse case delay target for hptsi < 48Mbp connections"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "enet_oh", CTLFLAG_RW, &bbr_include_enet_oh, 0, "Do we include the ethernet overhead in calculating pacing delay?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "ip_oh", CTLFLAG_RW, &bbr_include_ip_oh, 1, "Do we include the IP overhead in calculating pacing delay?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "tcp_oh", CTLFLAG_RW, &bbr_include_tcp_oh, 0, "Do we include the TCP overhead in calculating pacing delay?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "google_discount", CTLFLAG_RW, &bbr_google_discount, 10, "What is the default google discount percentage wise for pacing (11 = 1.1%%)?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "all_get_min", CTLFLAG_RW, &bbr_all_get_min, 0, "If you are less than a MSS do you just get the min?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "tso_min", CTLFLAG_RW, &bbr_hptsi_bytes_min, 1460, "For 0 -> 24Mbps what is floor number of segments for TSO"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "seg_tso_max", CTLFLAG_RW, &bbr_hptsi_segments_max, 6, "For 0 -> 24Mbps what is top number of segments for TSO"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "seg_floor", CTLFLAG_RW, &bbr_hptsi_segments_floor, 1, "Minimum TSO size we will fall too in segments"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "utter_max", CTLFLAG_RW, &bbr_hptsi_utter_max, 0, "The absolute maximum that any pacing (outside of hardware) can be"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "seg_divisor", CTLFLAG_RW, &bbr_hptsi_per_second, 100, "What is the divisor in our hptsi TSO calculation 512Mbps < X > 24Mbps "); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "srtt_mul", CTLFLAG_RW, &bbr_hptsi_max_mul, 1, "The multiplier for pace len max"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "srtt_div", CTLFLAG_RW, &bbr_hptsi_max_div, 2, "The divisor for pace len max"); /* Measurement controls */ bbr_measure = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "measure", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Measurement controls"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "min_i_bw", CTLFLAG_RW, &bbr_initial_bw_bps, 62500, "Minimum initial b/w in bytes per second"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "no_sack_needed", CTLFLAG_RW, &bbr_sack_not_required, 0, "Do we allow bbr to run on connections not supporting SACK?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "use_google", CTLFLAG_RW, &bbr_use_google_algo, 0, "Use has close to google V1.0 has possible?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "ts_limiting", CTLFLAG_RW, &bbr_ts_limiting, 1, "Do we attempt to use the peers timestamp to limit b/w caculations?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "ts_can_raise", CTLFLAG_RW, &bbr_ts_can_raise, 0, "Can we raise the b/w via timestamp b/w calculation?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "ts_delta", CTLFLAG_RW, &bbr_min_usec_delta, 20000, "How long in usec between ts of our sends in ts validation code?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "ts_peer_delta", CTLFLAG_RW, &bbr_min_peer_delta, 20, "What min numerical value should be between the peer deltas?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "ts_delta_percent", CTLFLAG_RW, &bbr_delta_percent, 150, "What percentage (150 = 15.0) do we allow variance for?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "min_measure_good_bw", CTLFLAG_RW, &bbr_min_measurements_req, 1, "What is the minimum measurment count we need before we switch to our b/w estimate"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "min_measure_before_pace", CTLFLAG_RW, &bbr_no_pacing_until, 4, "How many pkt-epoch's (0 is off) do we need before pacing is on?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "quanta", CTLFLAG_RW, &bbr_quanta, 2, "Extra quanta to add when calculating the target (ID section 4.2.3.2)."); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "noretran", CTLFLAG_RW, &bbr_no_retran, 0, "Should google mode not use retransmission measurements for the b/w estimation?"); /* State controls */ bbr_states = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "states", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "State controls"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "idle_restart", CTLFLAG_RW, &bbr_uses_idle_restart, 0, "Do we use a new special idle_restart state to ramp back up quickly?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "idle_restart_threshold", CTLFLAG_RW, &bbr_idle_restart_threshold, 100000, "How long must we be idle before we restart??"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "use_pkt_epoch", CTLFLAG_RW, &bbr_state_is_pkt_epoch, 0, "Do we use a pkt-epoch for substate if 0 rttProp?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "startup_rtt_gain", CTLFLAG_RW, &bbr_rtt_gain_thresh, 0, "What increase in RTT triggers us to stop ignoring no-loss and possibly exit startup?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "drain_floor", CTLFLAG_RW, &bbr_drain_floor, 88, "What is the lowest we can drain (pg) too?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "drain_2_target", CTLFLAG_RW, &bbr_state_drain_2_tar, 1, "Do we drain to target in drain substate?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "gain_2_target", CTLFLAG_RW, &bbr_gain_to_target, 1, "Does probe bw gain to target??"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "gain_extra_time", CTLFLAG_RW, &bbr_gain_gets_extra_too, 1, "Does probe bw gain get the extra time too?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "ld_div", CTLFLAG_RW, &bbr_drain_drop_div, 5, "Long drain drop divider?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "ld_mul", CTLFLAG_RW, &bbr_drain_drop_mul, 4, "Long drain drop multiplier?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "rand_ot_disc", CTLFLAG_RW, &bbr_rand_ot, 50, "Random discount of the ot?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "dr_filter_life", CTLFLAG_RW, &bbr_num_pktepo_for_del_limit, BBR_NUM_RTTS_FOR_DEL_LIMIT, "How many packet-epochs does the b/w delivery rate last?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "subdrain_applimited", CTLFLAG_RW, &bbr_sub_drain_app_limit, 0, "Does our sub-state drain invoke app limited if its long?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "use_cwnd_subdrain", CTLFLAG_RW, &bbr_sub_drain_slam_cwnd, 0, "Should we set/recover cwnd for sub-state drain?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "use_cwnd_maindrain", CTLFLAG_RW, &bbr_slam_cwnd_in_main_drain, 0, "Should we set/recover cwnd for main-state drain?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "google_gets_earlyout", CTLFLAG_RW, &google_allow_early_out, 1, "Should we allow google probe-bw/drain to exit early at flight target?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "google_exit_loss", CTLFLAG_RW, &google_consider_lost, 1, "Should we have losses exit gain of probebw in google mode??"); /* Startup controls */ bbr_startup = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "startup", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Startup controls"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_startup), OID_AUTO, "cheat_iwnd", CTLFLAG_RW, &bbr_sends_full_iwnd, 1, "Do we not pace but burst out initial windows has our TSO size?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_startup), OID_AUTO, "loss_threshold", CTLFLAG_RW, &bbr_startup_loss_thresh, 2000, "In startup what is the loss threshold in a pe that will exit us from startup?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_startup), OID_AUTO, "use_lowerpg", CTLFLAG_RW, &bbr_use_lower_gain_in_startup, 1, "Should we use a lower hptsi gain if we see loss in startup?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_startup), OID_AUTO, "gain", CTLFLAG_RW, &bbr_start_exit, 25, "What gain percent do we need to see to stay in startup??"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_startup), OID_AUTO, "low_gain", CTLFLAG_RW, &bbr_low_start_exit, 15, "What gain percent do we need to see to stay in the lower gain startup??"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_startup), OID_AUTO, "loss_exit", CTLFLAG_RW, &bbr_exit_startup_at_loss, 1, "Should we exit startup at loss in an epoch if we are not gaining?"); /* CWND controls */ bbr_cwnd = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "cwnd", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Cwnd controls"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "tar_rtt", CTLFLAG_RW, &bbr_cwndtarget_rtt_touse, 0, "Target cwnd rtt measurment to use (0=rtt_prop, 1=rtt_rack, 2=pkt_rtt, 3=srtt)?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "may_shrink", CTLFLAG_RW, &bbr_cwnd_may_shrink, 0, "Can the cwnd shrink if it would grow to more than the target?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "max_target_limit", CTLFLAG_RW, &bbr_target_cwnd_mult_limit, 8, "Do we limit the cwnd to some multiple of the cwnd target if cwnd can't shrink 0=no?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "highspeed_min", CTLFLAG_RW, &bbr_cwnd_min_val_hs, BBR_HIGHSPEED_NUM_MSS, "What is the high-speed min cwnd (rttProp under 1ms)"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "lowspeed_min", CTLFLAG_RW, &bbr_cwnd_min_val, BBR_PROBERTT_NUM_MSS, "What is the min cwnd (rttProp > 1ms)"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "initwin", CTLFLAG_RW, &bbr_def_init_win, 10, "What is the BBR initial window, if 0 use tcp version"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "do_loss_red", CTLFLAG_RW, &bbr_do_red, 600, "Do we reduce the b/w at exit from recovery based on ratio of prop/srtt (800=80.0, 0=off)?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "red_scale", CTLFLAG_RW, &bbr_red_scale, 20000, "What RTT do we scale with?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "red_growslow", CTLFLAG_RW, &bbr_red_growth_restrict, 1, "Do we restrict cwnd growth for whats in flight?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "red_div", CTLFLAG_RW, &bbr_red_div, 2, "If we reduce whats the divisor?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "red_mul", CTLFLAG_RW, &bbr_red_mul, 1, "If we reduce whats the mulitiplier?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "target_is_unit", CTLFLAG_RW, &bbr_target_is_bbunit, 0, "Is the state target the pacing_gain or BBR_UNIT?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "drop_limit", CTLFLAG_RW, &bbr_drop_limit, 0, "Number of segments limit for drop (0=use min_cwnd w/flight)?"); /* Timeout controls */ bbr_timeout = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "timeout", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Time out controls"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "delack", CTLFLAG_RW, &bbr_delack_time, 100000, "BBR's delayed ack time"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "tlp_uses", CTLFLAG_RW, &bbr_tlp_type_to_use, 3, "RTT that TLP uses in its calculations, 0=rttProp, 1=Rack_rtt, 2=pkt_rtt and 3=srtt"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "persmin", CTLFLAG_RW, &bbr_persist_min, 250000, "What is the minimum time in microseconds between persists"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "persmax", CTLFLAG_RW, &bbr_persist_max, 1000000, "What is the largest delay in microseconds between persists"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "tlp_minto", CTLFLAG_RW, &bbr_tlp_min, 10000, "TLP Min timeout in usecs"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "tlp_dack_time", CTLFLAG_RW, &bbr_delayed_ack_time, 200000, "TLP delayed ack compensation value"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "minrto", CTLFLAG_RW, &bbr_rto_min_ms, 30, "Minimum RTO in ms"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "maxrto", CTLFLAG_RW, &bbr_rto_max_sec, 4, "Maxiumum RTO in seconds -- should be at least as large as min_rto"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "tlp_retry", CTLFLAG_RW, &bbr_tlp_max_resend, 2, "How many times does TLP retry a single segment or multiple with no ACK"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "minto", CTLFLAG_RW, &bbr_min_to, 1000, "Minimum rack timeout in useconds"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "pktdelay", CTLFLAG_RW, &bbr_pkt_delay, 1000, "Extra RACK time (in useconds) besides reordering thresh"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "incr_tmrs", CTLFLAG_RW, &bbr_incr_timers, 1, "Increase the RXT/TLP timer by the pacing time used?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "rxtmark_sackpassed", CTLFLAG_RW, &bbr_marks_rxt_sack_passed, 0, "Mark sack passed on all those not ack'd when a RXT hits?"); /* Policer controls */ bbr_policer = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "policer", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Policer controls"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "detect_enable", CTLFLAG_RW, &bbr_policer_detection_enabled, 1, "Is policer detection enabled??"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "min_pes", CTLFLAG_RW, &bbr_lt_intvl_min_rtts, 4, "Minimum number of PE's?"); SYSCTL_ADD_U64(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "bwdiff", CTLFLAG_RW, &bbr_lt_bw_diff, (4000/8), "Minimal bw diff?"); SYSCTL_ADD_U64(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "bwratio", CTLFLAG_RW, &bbr_lt_bw_ratio, 8, "Minimal bw diff?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "from_rack_rxt", CTLFLAG_RW, &bbr_policer_call_from_rack_to, 0, "Do we call the policer detection code from a rack-timeout?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "false_postive", CTLFLAG_RW, &bbr_lt_intvl_fp, 0, "What packet epoch do we do false-postive detection at (0=no)?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "loss_thresh", CTLFLAG_RW, &bbr_lt_loss_thresh, 196, "Loss threshold 196 = 19.6%?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "false_postive_thresh", CTLFLAG_RW, &bbr_lt_fd_thresh, 100, "What percentage is the false detection threshold (150=15.0)?"); /* All the rest */ SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "cheat_rxt", CTLFLAG_RW, &bbr_use_rack_resend_cheat, 0, "Do we burst 1ms between sends on retransmissions (like rack)?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "error_paceout", CTLFLAG_RW, &bbr_error_base_paceout, 10000, "When we hit an error what is the min to pace out in usec's?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "kill_paceout", CTLFLAG_RW, &bbr_max_net_error_cnt, 10, "When we hit this many errors in a row, kill the session?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "data_after_close", CTLFLAG_RW, &bbr_ignore_data_after_close, 1, "Do we hold off sending a RST until all pending data is ack'd"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "resend_use_tso", CTLFLAG_RW, &bbr_resends_use_tso, 0, "Can resends use TSO?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "sblklimit", CTLFLAG_RW, &bbr_sack_block_limit, 128, "When do we start ignoring small sack blocks"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "bb_verbose", CTLFLAG_RW, &bbr_verbose_logging, 0, "Should BBR black box logging be verbose"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "reorder_thresh", CTLFLAG_RW, &bbr_reorder_thresh, 2, "What factor for rack will be added when seeing reordering (shift right)"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "reorder_fade", CTLFLAG_RW, &bbr_reorder_fade, 0, "Does reorder detection fade, if so how many ms (0 means never)"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW, &bbr_tlp_thresh, 1, "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)"); /* Stats and counters */ /* The pacing counters for hdwr/software can't be in the array */ bbr_nohdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK); bbr_hdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "enob_hdwr_pacing", CTLFLAG_RD, &bbr_hdwr_pacing_enobuf, "Total number of enobufs for hardware paced flows"); SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "enob_no_hdwr_pacing", CTLFLAG_RD, &bbr_nohdwr_pacing_enobuf, "Total number of enobufs for non-hardware paced flows"); bbr_flows_whdwr_pacing = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "hdwr_pacing", CTLFLAG_RD, &bbr_flows_whdwr_pacing, "Total number of hardware paced flows"); bbr_flows_nohdwr_pacing = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "software_pacing", CTLFLAG_RD, &bbr_flows_nohdwr_pacing, "Total number of software paced flows"); COUNTER_ARRAY_ALLOC(bbr_stat_arry, BBR_STAT_SIZE, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "stats", CTLFLAG_RD, bbr_stat_arry, BBR_STAT_SIZE, "BBR Stats"); COUNTER_ARRAY_ALLOC(bbr_opts_arry, BBR_OPTS_SIZE, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "opts", CTLFLAG_RD, bbr_opts_arry, BBR_OPTS_SIZE, "BBR Option Stats"); COUNTER_ARRAY_ALLOC(bbr_state_lost, BBR_MAX_STAT, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "lost", CTLFLAG_RD, bbr_state_lost, BBR_MAX_STAT, "Stats of when losses occur"); COUNTER_ARRAY_ALLOC(bbr_state_resend, BBR_MAX_STAT, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "stateresend", CTLFLAG_RD, bbr_state_resend, BBR_MAX_STAT, "Stats of what states resend"); COUNTER_ARRAY_ALLOC(bbr_state_time, BBR_MAX_STAT, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "statetime", CTLFLAG_RD, bbr_state_time, BBR_MAX_STAT, "Stats of time spent in the states"); COUNTER_ARRAY_ALLOC(bbr_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "outsize", CTLFLAG_RD, bbr_out_size, TCP_MSS_ACCT_SIZE, "Size of output calls"); SYSCTL_ADD_PROC(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "clrlost", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, &bbr_clear_lost, 0, sysctl_bbr_clear_lost, "IU", "Clear lost counters"); } static void bbr_counter_destroy(void) { COUNTER_ARRAY_FREE(bbr_stat_arry, BBR_STAT_SIZE); COUNTER_ARRAY_FREE(bbr_opts_arry, BBR_OPTS_SIZE); COUNTER_ARRAY_FREE(bbr_out_size, TCP_MSS_ACCT_SIZE); COUNTER_ARRAY_FREE(bbr_state_lost, BBR_MAX_STAT); COUNTER_ARRAY_FREE(bbr_state_time, BBR_MAX_STAT); COUNTER_ARRAY_FREE(bbr_state_resend, BBR_MAX_STAT); counter_u64_free(bbr_nohdwr_pacing_enobuf); counter_u64_free(bbr_hdwr_pacing_enobuf); counter_u64_free(bbr_flows_whdwr_pacing); counter_u64_free(bbr_flows_nohdwr_pacing); } static __inline void bbr_fill_in_logging_data(struct tcp_bbr *bbr, struct tcp_log_bbr *l, uint32_t cts) { memset(l, 0, sizeof(union tcp_log_stackspecific)); l->cur_del_rate = bbr->r_ctl.rc_bbr_cur_del_rate; l->delRate = get_filter_value(&bbr->r_ctl.rc_delrate); l->rttProp = get_filter_value_small(&bbr->r_ctl.rc_rttprop); l->bw_inuse = bbr_get_bw(bbr); l->inflight = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); l->applimited = bbr->r_ctl.r_app_limited_until; l->delivered = bbr->r_ctl.rc_delivered; l->timeStamp = cts; l->lost = bbr->r_ctl.rc_lost; l->bbr_state = bbr->rc_bbr_state; l->bbr_substate = bbr_state_val(bbr); l->epoch = bbr->r_ctl.rc_rtt_epoch; l->lt_epoch = bbr->r_ctl.rc_lt_epoch; l->pacing_gain = bbr->r_ctl.rc_bbr_hptsi_gain; l->cwnd_gain = bbr->r_ctl.rc_bbr_cwnd_gain; l->inhpts = bbr->rc_inp->inp_in_hpts; l->ininput = bbr->rc_inp->inp_in_input; l->use_lt_bw = bbr->rc_lt_use_bw; l->pkts_out = bbr->r_ctl.rc_flight_at_input; l->pkt_epoch = bbr->r_ctl.rc_pkt_epoch; } static void bbr_log_type_bw_reduce(struct tcp_bbr *bbr, int reason) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = 0; log.u_bbr.flex2 = 0; log.u_bbr.flex5 = 0; log.u_bbr.flex3 = 0; log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_loss_rate; log.u_bbr.flex7 = reason; log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_enters_probertt; log.u_bbr.flex8 = 0; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_BW_RED_EV, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_rwnd_collapse(struct tcp_bbr *bbr, int seq, int mode, uint32_t count) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = seq; log.u_bbr.flex2 = count; log.u_bbr.flex8 = mode; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_LOWGAIN, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_just_return(struct tcp_bbr *bbr, uint32_t cts, uint32_t tlen, uint8_t hpts_calling, uint8_t reason, uint32_t p_maxseg, int len) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = p_maxseg; log.u_bbr.flex2 = bbr->r_ctl.rc_hpts_flags; log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp; log.u_bbr.flex4 = reason; log.u_bbr.flex5 = bbr->rc_in_persist; log.u_bbr.flex6 = bbr->r_ctl.rc_last_delay_val; log.u_bbr.flex7 = p_maxseg; log.u_bbr.flex8 = bbr->rc_in_persist; log.u_bbr.pkts_out = 0; log.u_bbr.applimited = len; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_JUSTRET, 0, tlen, &log, false, &bbr->rc_tv); } } static void bbr_log_type_enter_rec(struct tcp_bbr *bbr, uint32_t seq) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = seq; log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent; log.u_bbr.flex3 = bbr->r_ctl.rc_recovery_start; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_ENTREC, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_msgsize_fail(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t len, uint32_t maxseg, uint32_t mtu, int32_t csum_flags, int32_t tso, uint32_t cts) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = tso; log.u_bbr.flex2 = maxseg; log.u_bbr.flex3 = mtu; log.u_bbr.flex4 = csum_flags; TCP_LOG_EVENTP(tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_MSGSIZE, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_flowend(struct tcp_bbr *bbr) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct sockbuf *r, *s; struct timeval tv; if (bbr->rc_inp->inp_socket) { r = &bbr->rc_inp->inp_socket->so_rcv; s = &bbr->rc_inp->inp_socket->so_snd; } else { r = s = NULL; } bbr_fill_in_logging_data(bbr, &log.u_bbr, tcp_get_usecs(&tv)); TCP_LOG_EVENTP(bbr->rc_tp, NULL, r, s, TCP_LOG_FLOWEND, 0, 0, &log, false, &tv); } } static void bbr_log_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t lost, uint32_t del) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = lost; log.u_bbr.flex2 = del; log.u_bbr.flex3 = bbr->r_ctl.rc_bbr_lastbtlbw; log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_rtt; log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch; log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; log.u_bbr.flex7 = line; log.u_bbr.flex8 = 0; log.u_bbr.inflight = bbr->r_ctl.r_measurement_count; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_PKT_EPOCH, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_time_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t epoch_time) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->r_ctl.rc_lost; log.u_bbr.flex2 = bbr->rc_inp->inp_socket->so_snd.sb_lowat; log.u_bbr.flex3 = bbr->rc_inp->inp_socket->so_snd.sb_hiwat; log.u_bbr.flex7 = line; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_TIME_EPOCH, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_set_of_state_target(struct tcp_bbr *bbr, uint32_t new_tar, int line, int meth) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex2 = new_tar; log.u_bbr.flex3 = line; log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs; log.u_bbr.flex5 = bbr_quanta; log.u_bbr.flex6 = bbr->r_ctl.rc_pace_min_segs; log.u_bbr.flex7 = bbr->rc_last_options; log.u_bbr.flex8 = meth; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_STATE_TARGET, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_statechange(struct tcp_bbr *bbr, uint32_t cts, int32_t line) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = line; log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int; if (bbr_state_is_pkt_epoch) log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PKTRTT); else log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PROP); log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch; log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; log.u_bbr.flex7 = (bbr->r_ctl.rc_target_at_state/1000); log.u_bbr.lt_epoch = bbr->r_ctl.rc_level_state_extra; log.u_bbr.pkts_out = bbr->r_ctl.rc_target_at_state; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_STATE, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, uint32_t rtt, uint32_t line, uint8_t reas, uint16_t cond) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = line; log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; log.u_bbr.flex3 = bbr->r_ctl.last_in_probertt; log.u_bbr.flex4 = applied; log.u_bbr.flex5 = rtt; log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex7 = cond; log.u_bbr.flex8 = reas; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_RTT_SHRINKS, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_exit_rec(struct tcp_bbr *bbr) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = bbr->r_ctl.rc_recovery_start; log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent; log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_EXITREC, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_cwndupd(struct tcp_bbr *bbr, uint32_t bytes_this_ack, uint32_t chg, uint32_t prev_acked, int32_t meth, uint32_t target, uint32_t th_ack, int32_t line) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = line; log.u_bbr.flex2 = prev_acked; log.u_bbr.flex3 = bytes_this_ack; log.u_bbr.flex4 = chg; log.u_bbr.flex5 = th_ack; log.u_bbr.flex6 = target; log.u_bbr.flex8 = meth; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_CWND, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_rtt_sample(struct tcp_bbr *bbr, uint32_t rtt, uint32_t tsin) { /* * Log the rtt sample we are applying to the srtt algorithm in * useconds. */ if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = rtt; log.u_bbr.flex2 = bbr->r_ctl.rc_bbr_state_time; log.u_bbr.flex3 = bbr->r_ctl.rc_ack_hdwr_delay; log.u_bbr.flex4 = bbr->rc_tp->ts_offset; log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; log.u_bbr.pkts_out = tcp_tv_to_mssectick(&bbr->rc_tv); log.u_bbr.flex6 = tsin; log.u_bbr.flex7 = 0; log.u_bbr.flex8 = bbr->rc_ack_was_delayed; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, TCP_LOG_RTT, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_pesist(struct tcp_bbr *bbr, uint32_t cts, uint32_t time_in, int32_t line, uint8_t enter_exit) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = time_in; log.u_bbr.flex2 = line; log.u_bbr.flex8 = enter_exit; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_PERSIST, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_ack_clear(struct tcp_bbr *bbr, uint32_t cts) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->rc_tp->ts_recent_age; log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int; log.u_bbr.flex4 = bbr->r_ctl.rc_went_idle_time; log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_ACKCLEAR, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_ack_event(struct tcp_bbr *bbr, struct tcphdr *th, struct tcpopt *to, uint32_t tlen, uint16_t nsegs, uint32_t cts, int32_t nxt_pkt, struct mbuf *m) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = nsegs; log.u_bbr.flex2 = bbr->r_ctl.rc_lost_bytes; if (m) { struct timespec ts; log.u_bbr.flex3 = m->m_flags; if (m->m_flags & M_TSTMP) { mbuf_tstmp2timespec(m, &ts); tv.tv_sec = ts.tv_sec; tv.tv_usec = ts.tv_nsec / 1000; log.u_bbr.lt_epoch = tcp_tv_to_usectick(&tv); } else { log.u_bbr.lt_epoch = 0; } if (m->m_flags & M_TSTMP_LRO) { tv.tv_sec = m->m_pkthdr.rcv_tstmp / 1000000000; tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000) / 1000; log.u_bbr.flex5 = tcp_tv_to_usectick(&tv); } else { /* No arrival timestamp */ log.u_bbr.flex5 = 0; } log.u_bbr.pkts_out = tcp_get_usecs(&tv); } else { log.u_bbr.flex3 = 0; log.u_bbr.flex5 = 0; log.u_bbr.flex6 = 0; log.u_bbr.pkts_out = 0; } log.u_bbr.flex4 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex7 = bbr->r_wanted_output; log.u_bbr.flex8 = bbr->rc_in_persist; TCP_LOG_EVENTP(bbr->rc_tp, th, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, TCP_LOG_IN, 0, tlen, &log, true, &bbr->rc_tv); } } static void bbr_log_doseg_done(struct tcp_bbr *bbr, uint32_t cts, int32_t nxt_pkt, int32_t did_out) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = did_out; log.u_bbr.flex2 = nxt_pkt; log.u_bbr.flex3 = bbr->r_ctl.rc_last_delay_val; log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags; log.u_bbr.flex5 = bbr->r_ctl.rc_timer_exp; log.u_bbr.flex6 = bbr->r_ctl.rc_lost_bytes; log.u_bbr.flex7 = bbr->r_wanted_output; log.u_bbr.flex8 = bbr->rc_in_persist; log.u_bbr.pkts_out = bbr->r_ctl.highest_hdwr_delay; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_DOSEG_DONE, 0, 0, &log, true, &bbr->rc_tv); } } static void bbr_log_enobuf_jmp(struct tcp_bbr *bbr, uint32_t len, uint32_t cts, int32_t line, uint32_t o_len, uint32_t segcnt, uint32_t segsiz) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = line; log.u_bbr.flex2 = o_len; log.u_bbr.flex3 = segcnt; log.u_bbr.flex4 = segsiz; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_ENOBUF_JMP, ENOBUFS, len, &log, true, &bbr->rc_tv); } } static void bbr_log_to_processing(struct tcp_bbr *bbr, uint32_t cts, int32_t ret, int32_t timers, uint8_t hpts_calling) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = timers; log.u_bbr.flex2 = ret; log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp; log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags; log.u_bbr.flex5 = cts; log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex8 = hpts_calling; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_TO_PROCESS, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_to_event(struct tcp_bbr *bbr, uint32_t cts, int32_t to_num) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; uint64_t ar; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->bbr_timer_src; log.u_bbr.flex2 = 0; log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; ar = (uint64_t)(bbr->r_ctl.rc_resend); ar >>= 32; ar &= 0x00000000ffffffff; log.u_bbr.flex4 = (uint32_t)ar; ar = (uint64_t)bbr->r_ctl.rc_resend; ar &= 0x00000000ffffffff; log.u_bbr.flex5 = (uint32_t)ar; log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); log.u_bbr.flex8 = to_num; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_RTO, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_startup_event(struct tcp_bbr *bbr, uint32_t cts, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint8_t reason) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = flex1; log.u_bbr.flex2 = flex2; log.u_bbr.flex3 = flex3; log.u_bbr.flex4 = 0; log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; log.u_bbr.flex8 = reason; log.u_bbr.cur_del_rate = bbr->r_ctl.rc_bbr_lastbtlbw; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_REDUCE, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = diag->p_nxt_slot; log.u_bbr.flex2 = diag->p_cur_slot; log.u_bbr.flex3 = diag->slot_req; log.u_bbr.flex4 = diag->inp_hptsslot; log.u_bbr.flex5 = diag->slot_remaining; log.u_bbr.flex6 = diag->need_new_to; log.u_bbr.flex7 = diag->p_hpts_active; log.u_bbr.flex8 = diag->p_on_min_sleep; /* Hijack other fields as needed */ log.u_bbr.epoch = diag->have_slept; log.u_bbr.lt_epoch = diag->yet_to_sleep; log.u_bbr.pkts_out = diag->co_ret; log.u_bbr.applimited = diag->hpts_sleep_time; log.u_bbr.delivered = diag->p_prev_slot; log.u_bbr.inflight = diag->p_runningtick; log.u_bbr.bw_inuse = diag->wheel_tick; log.u_bbr.rttProp = diag->wheel_cts; log.u_bbr.delRate = diag->maxticks; log.u_bbr.cur_del_rate = diag->p_curtick; log.u_bbr.cur_del_rate <<= 32; log.u_bbr.cur_del_rate |= diag->p_lasttick; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_HPTSDIAG, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt, uint32_t thresh, uint32_t to) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->rc_tp->t_rttvar; log.u_bbr.flex2 = time_since_sent; log.u_bbr.flex3 = srtt; log.u_bbr.flex4 = thresh; log.u_bbr.flex5 = to; log.u_bbr.flex6 = bbr->rc_tp->t_srtt; log.u_bbr.flex8 = mode; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERPREP, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len, uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = usecs; log.u_bbr.flex2 = len; log.u_bbr.flex3 = (uint32_t)((bw >> 32) & 0x00000000ffffffff); log.u_bbr.flex4 = (uint32_t)(bw & 0x00000000ffffffff); if (override) log.u_bbr.flex5 = (1 << 2); else log.u_bbr.flex5 = 0; log.u_bbr.flex6 = override; log.u_bbr.flex7 = gain; log.u_bbr.flex8 = mod; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_HPTSI_CALC, 0, len, &log, false, &bbr->rc_tv); } } static void bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->bbr_timer_src; log.u_bbr.flex2 = to; log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; log.u_bbr.flex4 = slot; log.u_bbr.flex5 = bbr->rc_inp->inp_hptsslot; log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); log.u_bbr.pkts_out = bbr->rc_inp->inp_flags2; log.u_bbr.flex8 = which; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERSTAR, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_thresh_choice(struct tcp_bbr *bbr, uint32_t cts, uint32_t thresh, uint32_t lro, uint32_t srtt, struct bbr_sendmap *rsm, uint8_t frm) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = thresh; log.u_bbr.flex2 = lro; log.u_bbr.flex3 = bbr->r_ctl.rc_reorder_ts; log.u_bbr.flex4 = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; log.u_bbr.flex5 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); log.u_bbr.flex6 = srtt; log.u_bbr.flex7 = bbr->r_ctl.rc_reorder_shift; log.u_bbr.flex8 = frm; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_THRESH_CALC, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_to_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts, uint8_t hpts_removed) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = line; log.u_bbr.flex2 = bbr->bbr_timer_src; log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; log.u_bbr.flex4 = bbr->rc_in_persist; log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); log.u_bbr.flex8 = hpts_removed; log.u_bbr.pkts_out = bbr->rc_pacer_started; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERCANC, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_tstmp_validation(struct tcp_bbr *bbr, uint64_t peer_delta, uint64_t delta) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = bbr->r_ctl.bbr_peer_tsratio; log.u_bbr.flex2 = (peer_delta >> 32); log.u_bbr.flex3 = (peer_delta & 0x00000000ffffffff); log.u_bbr.flex4 = (delta >> 32); log.u_bbr.flex5 = (delta & 0x00000000ffffffff); log.u_bbr.flex7 = bbr->rc_ts_clock_set; log.u_bbr.flex8 = bbr->rc_ts_cant_be_used; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_TSTMP_VAL, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_tsosize(struct tcp_bbr *bbr, uint32_t cts, uint32_t tsosz, uint32_t tls, uint32_t old_val, uint32_t maxseg, int hdwr) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = tsosz; log.u_bbr.flex2 = tls; log.u_bbr.flex3 = tcp_min_hptsi_time; log.u_bbr.flex4 = bbr->r_ctl.bbr_hptsi_bytes_min; log.u_bbr.flex5 = old_val; log.u_bbr.flex6 = maxseg; log.u_bbr.flex7 = bbr->rc_no_pacing; log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= bbr->rc_past_init_win; if (hdwr) log.u_bbr.flex8 = 0x80 | bbr->rc_use_google; else log.u_bbr.flex8 = bbr->rc_use_google; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_BBRTSO, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_rsmclear(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, uint32_t flags, uint32_t line) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = line; log.u_bbr.flex2 = rsm->r_start; log.u_bbr.flex3 = rsm->r_end; log.u_bbr.flex4 = rsm->r_delivered; log.u_bbr.flex5 = rsm->r_rtr_cnt; log.u_bbr.flex6 = rsm->r_dupack; log.u_bbr.flex7 = rsm->r_tim_lastsent[0]; log.u_bbr.flex8 = rsm->r_flags; /* Hijack the pkts_out fids */ log.u_bbr.applimited = flags; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_RSM_CLEARED, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_bbrupd(struct tcp_bbr *bbr, uint8_t flex8, uint32_t cts, uint32_t flex3, uint32_t flex2, uint32_t flex5, uint32_t flex6, uint32_t pkts_out, int flex7, uint32_t flex4, uint32_t flex1) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = flex1; log.u_bbr.flex2 = flex2; log.u_bbr.flex3 = flex3; log.u_bbr.flex4 = flex4; log.u_bbr.flex5 = flex5; log.u_bbr.flex6 = flex6; log.u_bbr.flex7 = flex7; /* Hijack the pkts_out fids */ log.u_bbr.pkts_out = pkts_out; log.u_bbr.flex8 = flex8; if (bbr->rc_ack_was_delayed) log.u_bbr.epoch = bbr->r_ctl.rc_ack_hdwr_delay; else log.u_bbr.epoch = 0; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_BBRUPD, 0, flex2, &log, false, &bbr->rc_tv); } } static void bbr_log_type_ltbw(struct tcp_bbr *bbr, uint32_t cts, int32_t reason, uint32_t newbw, uint32_t obw, uint32_t diff, uint32_t tim) { if (/*bbr_verbose_logging && */(bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = reason; log.u_bbr.flex2 = newbw; log.u_bbr.flex3 = obw; log.u_bbr.flex4 = diff; log.u_bbr.flex5 = bbr->r_ctl.rc_lt_lost; log.u_bbr.flex6 = bbr->r_ctl.rc_lt_del; log.u_bbr.flex7 = bbr->rc_lt_is_sampling; log.u_bbr.pkts_out = tim; log.u_bbr.bw_inuse = bbr->r_ctl.rc_lt_bw; if (bbr->rc_lt_use_bw == 0) log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch; else log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_BWSAMP, 0, 0, &log, false, &bbr->rc_tv); } } static inline void bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = line; log.u_bbr.flex2 = tick; log.u_bbr.flex3 = tp->t_maxunacktime; log.u_bbr.flex4 = tp->t_acktime; log.u_bbr.flex8 = event; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_PROGRESS, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_type_log_hdwr_pacing(struct tcp_bbr *bbr, const struct ifnet *ifp, uint64_t rate, uint64_t hw_rate, int line, uint32_t cts, int error) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff); log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff); log.u_bbr.flex3 = (((uint64_t)ifp >> 32) & 0x00000000ffffffff); log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff); log.u_bbr.bw_inuse = rate; log.u_bbr.flex5 = line; log.u_bbr.flex6 = error; log.u_bbr.flex8 = bbr->skip_gain; log.u_bbr.flex8 <<= 1; log.u_bbr.flex8 |= bbr->gain_is_limited; log.u_bbr.flex8 <<= 1; log.u_bbr.flex8 |= bbr->bbr_hdrw_pacing; log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_HDWR_PACE, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot, uint32_t del_by, uint32_t cts, uint32_t line, uint32_t prev_delay) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = slot; log.u_bbr.flex2 = del_by; log.u_bbr.flex3 = prev_delay; log.u_bbr.flex4 = line; log.u_bbr.flex5 = bbr->r_ctl.rc_last_delay_val; log.u_bbr.flex6 = bbr->r_ctl.rc_hptsi_agg_delay; log.u_bbr.flex7 = (0x0000ffff & bbr->r_ctl.rc_hpts_flags); log.u_bbr.flex8 = bbr->rc_in_persist; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_BBRSND, 0, len, &log, false, &bbr->rc_tv); } } static void bbr_log_type_bbrrttprop(struct tcp_bbr *bbr, uint32_t t, uint32_t end, uint32_t tsconv, uint32_t cts, int32_t match, uint32_t seq, uint8_t flags) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->r_ctl.rc_delivered; log.u_bbr.flex2 = 0; log.u_bbr.flex3 = bbr->r_ctl.rc_lowest_rtt; log.u_bbr.flex4 = end; log.u_bbr.flex5 = seq; log.u_bbr.flex6 = t; log.u_bbr.flex7 = match; log.u_bbr.flex8 = flags; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_BBRRTT, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_exit_gain(struct tcp_bbr *bbr, uint32_t cts, int32_t entry_method) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex2 = (bbr->rc_tp->t_maxseg - bbr->rc_last_options); log.u_bbr.flex3 = bbr->r_ctl.gain_epoch; log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs; log.u_bbr.flex5 = bbr->r_ctl.rc_pace_min_segs; log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_state_atflight; log.u_bbr.flex7 = 0; log.u_bbr.flex8 = entry_method; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_EXIT_GAIN, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_settings_change(struct tcp_bbr *bbr, int settings_desired) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); /* R-HU */ log.u_bbr.flex1 = 0; log.u_bbr.flex2 = 0; log.u_bbr.flex3 = 0; log.u_bbr.flex4 = 0; log.u_bbr.flex7 = 0; log.u_bbr.flex8 = settings_desired; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_SETTINGS_CHG, 0, 0, &log, false, &bbr->rc_tv); } } /* * Returns the bw from the our filter. */ static inline uint64_t bbr_get_full_bw(struct tcp_bbr *bbr) { uint64_t bw; bw = get_filter_value(&bbr->r_ctl.rc_delrate); return (bw); } static inline void bbr_set_pktepoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line) { uint64_t calclr; uint32_t lost, del; if (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_pktepoch) lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lost_at_pktepoch; else lost = 0; del = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_pkt_epoch_del; if (lost == 0) { calclr = 0; } else if (del) { calclr = lost; calclr *= (uint64_t)1000; calclr /= (uint64_t)del; } else { /* Nothing delivered? 100.0% loss */ calclr = 1000; } bbr->r_ctl.rc_pkt_epoch_loss_rate = (uint32_t)calclr; if (IN_RECOVERY(bbr->rc_tp->t_flags)) bbr->r_ctl.recovery_lr += (uint32_t)calclr; bbr->r_ctl.rc_pkt_epoch++; if (bbr->rc_no_pacing && (bbr->r_ctl.rc_pkt_epoch >= bbr->no_pacing_until)) { bbr->rc_no_pacing = 0; tcp_bbr_tso_size_check(bbr, cts); } bbr->r_ctl.rc_pkt_epoch_rtt = bbr_calc_time(cts, bbr->r_ctl.rc_pkt_epoch_time); bbr->r_ctl.rc_pkt_epoch_time = cts; /* What was our loss rate */ bbr_log_pkt_epoch(bbr, cts, line, lost, del); bbr->r_ctl.rc_pkt_epoch_del = bbr->r_ctl.rc_delivered; bbr->r_ctl.rc_lost_at_pktepoch = bbr->r_ctl.rc_lost; } static inline void bbr_set_epoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line) { uint32_t epoch_time; /* Tick the RTT clock */ bbr->r_ctl.rc_rtt_epoch++; epoch_time = cts - bbr->r_ctl.rc_rcv_epoch_start; bbr_log_time_epoch(bbr, cts, line, epoch_time); bbr->r_ctl.rc_rcv_epoch_start = cts; } static inline void bbr_isit_a_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, int32_t line, int32_t cum_acked) { if (SEQ_GEQ(rsm->r_delivered, bbr->r_ctl.rc_pkt_epoch_del)) { bbr->rc_is_pkt_epoch_now = 1; } } /* * Returns the bw from either the b/w filter * or from the lt_bw (if the connection is being * policed). */ static inline uint64_t __bbr_get_bw(struct tcp_bbr *bbr) { uint64_t bw, min_bw; uint64_t rtt; int gm_measure_cnt = 1; /* * For startup we make, like google, a * minimum b/w. This is generated from the * IW and the rttProp. We do fall back to srtt * if for some reason (initial handshake) we don't * have a rttProp. We, in the worst case, fall back * to the configured min_bw (rc_initial_hptsi_bw). */ if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { /* Attempt first to use rttProp */ rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop); if (rtt && (rtt < 0xffffffff)) { measure: min_bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) * ((uint64_t)1000000); min_bw /= rtt; if (min_bw < bbr->r_ctl.rc_initial_hptsi_bw) { min_bw = bbr->r_ctl.rc_initial_hptsi_bw; } } else if (bbr->rc_tp->t_srtt != 0) { /* No rttProp, use srtt? */ rtt = bbr_get_rtt(bbr, BBR_SRTT); goto measure; } else { min_bw = bbr->r_ctl.rc_initial_hptsi_bw; } } else min_bw = 0; if ((bbr->rc_past_init_win == 0) && (bbr->r_ctl.rc_delivered > bbr_initial_cwnd(bbr, bbr->rc_tp))) bbr->rc_past_init_win = 1; if ((bbr->rc_use_google) && (bbr->r_ctl.r_measurement_count >= 1)) gm_measure_cnt = 0; if (gm_measure_cnt && ((bbr->r_ctl.r_measurement_count < bbr_min_measurements_req) || (bbr->rc_past_init_win == 0))) { /* For google we use our guess rate until we get 1 measurement */ use_initial_window: rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop); if (rtt && (rtt < 0xffffffff)) { /* * We have an RTT measurment. Use that in * combination with our initial window to calculate * a b/w. */ bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) * ((uint64_t)1000000); bw /= rtt; if (bw < bbr->r_ctl.rc_initial_hptsi_bw) { bw = bbr->r_ctl.rc_initial_hptsi_bw; } } else { /* Drop back to the 40 and punt to a default */ bw = bbr->r_ctl.rc_initial_hptsi_bw; } if (bw < 1) /* Probably should panic */ bw = 1; if (bw > min_bw) return (bw); else return (min_bw); } if (bbr->rc_lt_use_bw) bw = bbr->r_ctl.rc_lt_bw; else if (bbr->r_recovery_bw && (bbr->rc_use_google == 0)) bw = bbr->r_ctl.red_bw; else bw = get_filter_value(&bbr->r_ctl.rc_delrate); if (bbr->rc_tp->t_peakrate_thr && (bbr->rc_use_google == 0)) { /* * Enforce user set rate limit, keep in mind that * t_peakrate_thr is in B/s already */ bw = uqmin((uint64_t)bbr->rc_tp->t_peakrate_thr, bw); } if (bw == 0) { /* We should not be at 0, go to the initial window then */ goto use_initial_window; } if (bw < 1) /* Probably should panic */ bw = 1; if (bw < min_bw) bw = min_bw; return (bw); } static inline uint64_t bbr_get_bw(struct tcp_bbr *bbr) { uint64_t bw; bw = __bbr_get_bw(bbr); return (bw); } static inline void bbr_reset_lt_bw_interval(struct tcp_bbr *bbr, uint32_t cts) { bbr->r_ctl.rc_lt_epoch = bbr->r_ctl.rc_pkt_epoch; bbr->r_ctl.rc_lt_time = bbr->r_ctl.rc_del_time; bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered; bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; } static inline void bbr_reset_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts) { bbr->rc_lt_is_sampling = 0; bbr->rc_lt_use_bw = 0; bbr->r_ctl.rc_lt_bw = 0; bbr_reset_lt_bw_interval(bbr, cts); } static inline void bbr_lt_bw_samp_done(struct tcp_bbr *bbr, uint64_t bw, uint32_t cts, uint32_t timin) { uint64_t diff; /* Do we have a previous sample? */ if (bbr->r_ctl.rc_lt_bw) { /* Get the diff in bytes per second */ if (bbr->r_ctl.rc_lt_bw > bw) diff = bbr->r_ctl.rc_lt_bw - bw; else diff = bw - bbr->r_ctl.rc_lt_bw; if ((diff <= bbr_lt_bw_diff) || (diff <= (bbr->r_ctl.rc_lt_bw / bbr_lt_bw_ratio))) { /* Consider us policed */ uint32_t saved_bw; saved_bw = (uint32_t)bbr->r_ctl.rc_lt_bw; bbr->r_ctl.rc_lt_bw = (bw + bbr->r_ctl.rc_lt_bw) / 2; /* average of two */ bbr->rc_lt_use_bw = 1; bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; /* * Use pkt based epoch for measuring length of * policer up */ bbr->r_ctl.rc_lt_epoch_use = bbr->r_ctl.rc_pkt_epoch; /* * reason 4 is we need to start consider being * policed */ bbr_log_type_ltbw(bbr, cts, 4, (uint32_t)bw, saved_bw, (uint32_t)diff, timin); return; } } bbr->r_ctl.rc_lt_bw = bw; bbr_reset_lt_bw_interval(bbr, cts); bbr_log_type_ltbw(bbr, cts, 5, 0, (uint32_t)bw, 0, timin); } static void bbr_randomize_extra_state_time(struct tcp_bbr *bbr) { uint32_t ran, deduct; ran = arc4random_uniform(bbr_rand_ot); if (ran) { deduct = bbr->r_ctl.rc_level_state_extra / ran; bbr->r_ctl.rc_level_state_extra -= deduct; } } /* * Return randomly the starting state * to use in probebw. */ static uint8_t bbr_pick_probebw_substate(struct tcp_bbr *bbr, uint32_t cts) { uint32_t ran; uint8_t ret_val; /* Initialize the offset to 0 */ bbr->r_ctl.rc_exta_time_gd = 0; bbr->rc_hit_state_1 = 0; bbr->r_ctl.rc_level_state_extra = 0; ran = arc4random_uniform((BBR_SUBSTATE_COUNT-1)); /* * The math works funny here :) the return value is used to set the * substate and then the state change is called which increments by * one. So if we return 1 (DRAIN) we will increment to 2 (LEVEL1) when * we fully enter the state. Note that the (8 - 1 - ran) assures that * we return 1 - 7, so we dont return 0 and end up starting in * state 1 (DRAIN). */ ret_val = BBR_SUBSTATE_COUNT - 1 - ran; /* Set an epoch */ if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) bbr_set_epoch(bbr, cts, __LINE__); bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; return (ret_val); } static void bbr_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts, int32_t loss_detected) { uint32_t diff, d_time; uint64_t del_time, bw, lost, delivered; if (bbr->r_use_policer == 0) return; if (bbr->rc_lt_use_bw) { /* We are using lt bw do we stop yet? */ diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use; if (diff > bbr_lt_bw_max_rtts) { /* Reset it all */ reset_all: bbr_reset_lt_bw_sampling(bbr, cts); if (bbr->rc_filled_pipe) { bbr_set_epoch(bbr, cts, __LINE__); bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); bbr_substate_change(bbr, cts, __LINE__, 0); bbr->rc_bbr_state = BBR_STATE_PROBE_BW; bbr_log_type_statechange(bbr, cts, __LINE__); } else { /* * This should not happen really * unless we remove the startup/drain * restrictions above. */ bbr->rc_bbr_state = BBR_STATE_STARTUP; bbr_set_epoch(bbr, cts, __LINE__); bbr->r_ctl.rc_bbr_state_time = cts; bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; bbr_set_state_target(bbr, __LINE__); bbr_log_type_statechange(bbr, cts, __LINE__); } /* reason 0 is to stop using lt-bw */ bbr_log_type_ltbw(bbr, cts, 0, 0, 0, 0, 0); return; } if (bbr_lt_intvl_fp == 0) { /* Not doing false-postive detection */ return; } /* False positive detection */ if (diff == bbr_lt_intvl_fp) { /* At bbr_lt_intvl_fp we record the lost */ bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered; bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; } else if (diff > (bbr_lt_intvl_min_rtts + bbr_lt_intvl_fp)) { /* Now is our loss rate still high? */ lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost; delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del; if ((delivered == 0) || (((lost * 1000)/delivered) < bbr_lt_fd_thresh)) { /* No still below our threshold */ bbr_log_type_ltbw(bbr, cts, 7, lost, delivered, 0, 0); } else { /* Yikes its still high, it must be a false positive */ bbr_log_type_ltbw(bbr, cts, 8, lost, delivered, 0, 0); goto reset_all; } } return; } /* * Wait for the first loss before sampling, to let the policer * exhaust its tokens and estimate the steady-state rate allowed by * the policer. Starting samples earlier includes bursts that * over-estimate the bw. */ if (bbr->rc_lt_is_sampling == 0) { /* reason 1 is to begin doing the sampling */ if (loss_detected == 0) return; bbr_reset_lt_bw_interval(bbr, cts); bbr->rc_lt_is_sampling = 1; bbr_log_type_ltbw(bbr, cts, 1, 0, 0, 0, 0); return; } /* Now how long were we delivering long term last> */ if (TSTMP_GEQ(bbr->r_ctl.rc_del_time, bbr->r_ctl.rc_lt_time)) d_time = bbr->r_ctl.rc_del_time - bbr->r_ctl.rc_lt_time; else d_time = 0; /* To avoid underestimates, reset sampling if we run out of data. */ if (bbr->r_ctl.r_app_limited_until) { /* Can not measure in app-limited state */ bbr_reset_lt_bw_sampling(bbr, cts); /* reason 2 is to reset sampling due to app limits */ bbr_log_type_ltbw(bbr, cts, 2, 0, 0, 0, d_time); return; } diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch; if (diff < bbr_lt_intvl_min_rtts) { /* * need more samples (we don't * start on a round like linux so * we need 1 more). */ /* 6 is not_enough time or no-loss */ bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); return; } if (diff > (4 * bbr_lt_intvl_min_rtts)) { /* * For now if we wait too long, reset all sampling. We need * to do some research here, its possible that we should * base this on how much loss as occurred.. something like * if its under 10% (or some thresh) reset all otherwise * don't. Thats for phase II I guess. */ bbr_reset_lt_bw_sampling(bbr, cts); /* reason 3 is to reset sampling due too long of sampling */ bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time); return; } /* * End sampling interval when a packet is lost, so we estimate the * policer tokens were exhausted. Stopping the sampling before the * tokens are exhausted under-estimates the policed rate. */ if (loss_detected == 0) { /* 6 is not_enough time or no-loss */ bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); return; } /* Calculate packets lost and delivered in sampling interval. */ lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost; delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del; if ((delivered == 0) || (((lost * 1000)/delivered) < bbr_lt_loss_thresh)) { bbr_log_type_ltbw(bbr, cts, 6, lost, delivered, 0, d_time); return; } if (d_time < 1000) { /* Not enough time. wait */ /* 6 is not_enough time or no-loss */ bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); return; } if (d_time >= (0xffffffff / USECS_IN_MSEC)) { /* Too long */ bbr_reset_lt_bw_sampling(bbr, cts); /* reason 3 is to reset sampling due too long of sampling */ bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time); return; } del_time = d_time; bw = delivered; bw *= (uint64_t)USECS_IN_SECOND; bw /= del_time; bbr_lt_bw_samp_done(bbr, bw, cts, d_time); } /* * Allocate a sendmap from our zone. */ static struct bbr_sendmap * bbr_alloc(struct tcp_bbr *bbr) { struct bbr_sendmap *rsm; BBR_STAT_INC(bbr_to_alloc); rsm = uma_zalloc(bbr_zone, (M_NOWAIT | M_ZERO)); if (rsm) { bbr->r_ctl.rc_num_maps_alloced++; return (rsm); } if (bbr->r_ctl.rc_free_cnt) { BBR_STAT_INC(bbr_to_alloc_emerg); rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next); bbr->r_ctl.rc_free_cnt--; return (rsm); } BBR_STAT_INC(bbr_to_alloc_failed); return (NULL); } static struct bbr_sendmap * bbr_alloc_full_limit(struct tcp_bbr *bbr) { if ((V_tcp_map_entries_limit > 0) && (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { BBR_STAT_INC(bbr_alloc_limited); if (!bbr->alloc_limit_reported) { bbr->alloc_limit_reported = 1; BBR_STAT_INC(bbr_alloc_limited_conns); } return (NULL); } return (bbr_alloc(bbr)); } /* wrapper to allocate a sendmap entry, subject to a specific limit */ static struct bbr_sendmap * bbr_alloc_limit(struct tcp_bbr *bbr, uint8_t limit_type) { struct bbr_sendmap *rsm; if (limit_type) { /* currently there is only one limit type */ if (V_tcp_map_split_limit > 0 && bbr->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) { BBR_STAT_INC(bbr_split_limited); if (!bbr->alloc_limit_reported) { bbr->alloc_limit_reported = 1; BBR_STAT_INC(bbr_alloc_limited_conns); } return (NULL); } } /* allocate and mark in the limit type, if set */ rsm = bbr_alloc(bbr); if (rsm != NULL && limit_type) { rsm->r_limit_type = limit_type; bbr->r_ctl.rc_num_split_allocs++; } return (rsm); } static void bbr_free(struct tcp_bbr *bbr, struct bbr_sendmap *rsm) { if (rsm->r_limit_type) { /* currently there is only one limit type */ bbr->r_ctl.rc_num_split_allocs--; } if (rsm->r_is_smallmap) bbr->r_ctl.rc_num_small_maps_alloced--; if (bbr->r_ctl.rc_tlp_send == rsm) bbr->r_ctl.rc_tlp_send = NULL; if (bbr->r_ctl.rc_resend == rsm) { bbr->r_ctl.rc_resend = NULL; } if (bbr->r_ctl.rc_next == rsm) bbr->r_ctl.rc_next = NULL; if (bbr->r_ctl.rc_sacklast == rsm) bbr->r_ctl.rc_sacklast = NULL; if (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) { memset(rsm, 0, sizeof(struct bbr_sendmap)); TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next); rsm->r_limit_type = 0; bbr->r_ctl.rc_free_cnt++; return; } bbr->r_ctl.rc_num_maps_alloced--; uma_zfree(bbr_zone, rsm); } /* * Returns the BDP. */ static uint64_t bbr_get_bw_delay_prod(uint64_t rtt, uint64_t bw) { /* * Calculate the bytes in flight needed given the bw (in bytes per * second) and the specifyed rtt in useconds. We need to put out the * returned value per RTT to match that rate. Gain will normally * raise it up from there. * * This should not overflow as long as the bandwidth is below 1 * TByte per second (bw < 10**12 = 2**40) and the rtt is smaller * than 1000 seconds (rtt < 10**3 * 10**6 = 10**9 = 2**30). */ uint64_t usec_per_sec; usec_per_sec = USECS_IN_SECOND; return ((rtt * bw) / usec_per_sec); } /* * Return the initial cwnd. */ static uint32_t bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp) { uint32_t i_cwnd; if (bbr->rc_init_win) { i_cwnd = bbr->rc_init_win * tp->t_maxseg; } else if (V_tcp_initcwnd_segments) i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg), max(2 * tp->t_maxseg, 14600)); else if (V_tcp_do_rfc3390) i_cwnd = min(4 * tp->t_maxseg, max(2 * tp->t_maxseg, 4380)); else { /* Per RFC5681 Section 3.1 */ if (tp->t_maxseg > 2190) i_cwnd = 2 * tp->t_maxseg; else if (tp->t_maxseg > 1095) i_cwnd = 3 * tp->t_maxseg; else i_cwnd = 4 * tp->t_maxseg; } return (i_cwnd); } /* * Given a specified gain, return the target * cwnd based on that gain. */ static uint32_t bbr_get_raw_target_cwnd(struct tcp_bbr *bbr, uint32_t gain, uint64_t bw) { uint64_t bdp, rtt; uint32_t cwnd; if ((get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) || (bbr_get_full_bw(bbr) == 0)) { /* No measurements yet */ return (bbr_initial_cwnd(bbr, bbr->rc_tp)); } /* * Get bytes per RTT needed (rttProp is normally in * bbr_cwndtarget_rtt_touse) */ rtt = bbr_get_rtt(bbr, bbr_cwndtarget_rtt_touse); /* Get the bdp from the two values */ bdp = bbr_get_bw_delay_prod(rtt, bw); /* Now apply the gain */ cwnd = (uint32_t)(((bdp * ((uint64_t)gain)) + (uint64_t)(BBR_UNIT - 1)) / ((uint64_t)BBR_UNIT)); return (cwnd); } static uint32_t bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain) { uint32_t cwnd, mss; mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); /* Get the base cwnd with gain rounded to a mss */ cwnd = roundup(bbr_get_raw_target_cwnd(bbr, bw, gain), mss); /* * Add in N (2 default since we do not have a * fq layer to trap packets in) quanta's per the I-D * section 4.2.3.2 quanta adjust. */ cwnd += (bbr_quanta * bbr->r_ctl.rc_pace_max_segs); if (bbr->rc_use_google) { if((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && (bbr_state_val(bbr) == BBR_SUB_GAIN)) { /* * The linux implementation adds * an extra 2 x mss in gain cycle which * is documented no-where except in the code. * so we add more for Neal undocumented feature */ cwnd += 2 * mss; } if ((cwnd / mss) & 0x1) { /* Round up for odd num mss */ cwnd += mss; } } /* Are we below the min cwnd? */ if (cwnd < get_min_cwnd(bbr)) return (get_min_cwnd(bbr)); return (cwnd); } static uint16_t bbr_gain_adjust(struct tcp_bbr *bbr, uint16_t gain) { if (gain < 1) gain = 1; return (gain); } static uint32_t bbr_get_header_oh(struct tcp_bbr *bbr) { int seg_oh; seg_oh = 0; if (bbr->r_ctl.rc_inc_tcp_oh) { /* Do we include TCP overhead? */ seg_oh = (bbr->rc_last_options + sizeof(struct tcphdr)); } if (bbr->r_ctl.rc_inc_ip_oh) { /* Do we include IP overhead? */ #ifdef INET6 if (bbr->r_is_v6) seg_oh += sizeof(struct ip6_hdr); else #endif #ifdef INET seg_oh += sizeof(struct ip); #endif } if (bbr->r_ctl.rc_inc_enet_oh) { /* Do we include the ethernet overhead? */ seg_oh += sizeof(struct ether_header); } return(seg_oh); } static uint32_t bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw) { uint64_t divor, res, tim; if (useconds_time == 0) return (0); gain = bbr_gain_adjust(bbr, gain); divor = (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT; tim = useconds_time; res = (tim * bw * gain) / divor; if (res == 0) res = 1; return ((uint32_t)res); } /* * Given a gain and a length return the delay in useconds that * should be used to evenly space out packets * on the connection (based on the gain factor). */ static uint32_t bbr_get_pacing_delay(struct tcp_bbr *bbr, uint16_t gain, int32_t len, uint32_t cts, int nolog) { uint64_t bw, lentim, res; uint32_t usecs, srtt, over = 0; uint32_t seg_oh, num_segs, maxseg; if (len == 0) return (0); maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; num_segs = (len + maxseg - 1) / maxseg; if (bbr->rc_use_google == 0) { seg_oh = bbr_get_header_oh(bbr); len += (num_segs * seg_oh); } gain = bbr_gain_adjust(bbr, gain); bw = bbr_get_bw(bbr); if (bbr->rc_use_google) { uint64_t cbw; /* * Reduce the b/w by the google discount * factor 10 = 1%. */ cbw = bw * (uint64_t)(1000 - bbr->r_ctl.bbr_google_discount); cbw /= (uint64_t)1000; /* We don't apply a discount if it results in 0 */ if (cbw > 0) bw = cbw; } lentim = ((uint64_t)len * (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT); res = lentim / ((uint64_t)gain * bw); if (res == 0) res = 1; usecs = (uint32_t)res; srtt = bbr_get_rtt(bbr, BBR_SRTT); if (bbr_hptsi_max_mul && bbr_hptsi_max_div && (bbr->rc_use_google == 0) && (usecs > ((srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div))) { /* * We cannot let the delay be more than 1/2 the srtt time. * Otherwise we cannot pace out or send properly. */ over = usecs = (srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div; BBR_STAT_INC(bbr_hpts_min_time); } if (!nolog) bbr_log_pacing_delay_calc(bbr, gain, len, cts, usecs, bw, over, 1); return (usecs); } static void bbr_ack_received(struct tcpcb *tp, struct tcp_bbr *bbr, struct tcphdr *th, uint32_t bytes_this_ack, uint32_t sack_changed, uint32_t prev_acked, int32_t line, uint32_t losses) { INP_WLOCK_ASSERT(tp->t_inpcb); uint64_t bw; uint32_t cwnd, target_cwnd, saved_bytes, maxseg; int32_t meth; #ifdef STATS if ((tp->t_flags & TF_GPUTINPROG) && SEQ_GEQ(th->th_ack, tp->gput_ack)) { /* * Strech acks and compressed acks will cause this to * oscillate but we are doing it the same way as the main * stack so it will be compariable (though possibly not * ideal). */ int32_t cgput; int64_t gput, time_stamp; gput = (int64_t) (th->th_ack - tp->gput_seq) * 8; time_stamp = max(1, ((bbr->r_ctl.rc_rcvtime - tp->gput_ts) / 1000)); cgput = gput / time_stamp; stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT, cgput); if (tp->t_stats_gput_prev > 0) stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_GPUT_ND, ((gput - tp->t_stats_gput_prev) * 100) / tp->t_stats_gput_prev); tp->t_flags &= ~TF_GPUTINPROG; tp->t_stats_gput_prev = cgput; } #endif if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) { /* We don't change anything in probe-rtt */ return; } maxseg = tp->t_maxseg - bbr->rc_last_options; saved_bytes = bytes_this_ack; bytes_this_ack += sack_changed; if (bytes_this_ack > prev_acked) { bytes_this_ack -= prev_acked; /* * A byte ack'd gives us a full mss * to be like linux i.e. they count packets. */ if ((bytes_this_ack < maxseg) && bbr->rc_use_google) bytes_this_ack = maxseg; } else { /* Unlikely */ bytes_this_ack = 0; } cwnd = tp->snd_cwnd; bw = get_filter_value(&bbr->r_ctl.rc_delrate); if (bw) target_cwnd = bbr_get_target_cwnd(bbr, bw, (uint32_t)bbr->r_ctl.rc_bbr_cwnd_gain); else target_cwnd = bbr_initial_cwnd(bbr, bbr->rc_tp); if (IN_RECOVERY(tp->t_flags) && (bbr->bbr_prev_in_rec == 0)) { /* * We are entering recovery and * thus packet conservation. */ bbr->pkt_conservation = 1; bbr->r_ctl.rc_recovery_start = bbr->r_ctl.rc_rcvtime; cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + bytes_this_ack; } if (IN_RECOVERY(tp->t_flags)) { uint32_t flight; bbr->bbr_prev_in_rec = 1; if (cwnd > losses) { cwnd -= losses; if (cwnd < maxseg) cwnd = maxseg; } else cwnd = maxseg; flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); bbr_log_type_cwndupd(bbr, flight, 0, losses, 10, 0, 0, line); if (bbr->pkt_conservation) { uint32_t time_in; if (TSTMP_GEQ(bbr->r_ctl.rc_rcvtime, bbr->r_ctl.rc_recovery_start)) time_in = bbr->r_ctl.rc_rcvtime - bbr->r_ctl.rc_recovery_start; else time_in = 0; if (time_in >= bbr_get_rtt(bbr, BBR_RTT_PROP)) { /* Clear packet conservation after an rttProp */ bbr->pkt_conservation = 0; } else { if ((flight + bytes_this_ack) > cwnd) cwnd = flight + bytes_this_ack; if (cwnd < get_min_cwnd(bbr)) cwnd = get_min_cwnd(bbr); tp->snd_cwnd = cwnd; bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, 1, target_cwnd, th->th_ack, line); return; } } } else bbr->bbr_prev_in_rec = 0; if ((bbr->rc_use_google == 0) && bbr->r_ctl.restrict_growth) { bbr->r_ctl.restrict_growth--; if (bytes_this_ack > maxseg) bytes_this_ack = maxseg; } if (bbr->rc_filled_pipe) { /* * Here we have exited startup and filled the pipe. We will * thus allow the cwnd to shrink to the target. We hit here * mostly. */ uint32_t s_cwnd; meth = 2; s_cwnd = min((cwnd + bytes_this_ack), target_cwnd); if (s_cwnd > cwnd) cwnd = s_cwnd; else if (bbr_cwnd_may_shrink || bbr->rc_use_google || bbr->rc_no_pacing) cwnd = s_cwnd; } else { /* * Here we are still in startup, we increase cwnd by what * has been acked. */ if ((cwnd < target_cwnd) || (bbr->rc_past_init_win == 0)) { meth = 3; cwnd += bytes_this_ack; } else { /* * Method 4 means we are at target so no gain in * startup and past the initial window. */ meth = 4; } } tp->snd_cwnd = max(cwnd, get_min_cwnd(bbr)); bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, meth, target_cwnd, th->th_ack, line); } static void tcp_bbr_partialack(struct tcpcb *tp) { struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= tp->snd_cwnd) { bbr->r_wanted_output = 1; } } static void bbr_post_recovery(struct tcpcb *tp) { struct tcp_bbr *bbr; uint32_t flight; INP_WLOCK_ASSERT(tp->t_inpcb); bbr = (struct tcp_bbr *)tp->t_fb_ptr; /* * Here we just exit recovery. */ EXIT_RECOVERY(tp->t_flags); /* Lock in our b/w reduction for the specified number of pkt-epochs */ bbr->r_recovery_bw = 0; tp->snd_recover = tp->snd_una; tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); bbr->pkt_conservation = 0; if (bbr->rc_use_google == 0) { /* * For non-google mode lets * go ahead and make sure we clear * the recovery state so if we * bounce back in to recovery we * will do PC. */ bbr->bbr_prev_in_rec = 0; } bbr_log_type_exit_rec(bbr); if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent); bbr_log_type_cwndupd(bbr, 0, 0, 0, 15, 0, 0, __LINE__); } else { /* For probe-rtt case lets fix up its saved_cwnd */ if (bbr->r_ctl.rc_saved_cwnd < bbr->r_ctl.rc_cwnd_on_ent) { bbr->r_ctl.rc_saved_cwnd = bbr->r_ctl.rc_cwnd_on_ent; bbr_log_type_cwndupd(bbr, 0, 0, 0, 16, 0, 0, __LINE__); } } flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); if ((bbr->rc_use_google == 0) && bbr_do_red) { uint64_t val, lr2use; uint32_t maxseg, newcwnd, acks_inflight, ratio, cwnd; uint32_t *cwnd_p; if (bbr_get_rtt(bbr, BBR_SRTT)) { val = ((uint64_t)bbr_get_rtt(bbr, BBR_RTT_PROP) * (uint64_t)1000); val /= bbr_get_rtt(bbr, BBR_SRTT); ratio = (uint32_t)val; } else ratio = 1000; bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, bbr->r_ctl.recovery_lr, 21, ratio, bbr->r_ctl.rc_red_cwnd_pe, __LINE__); if ((ratio < bbr_do_red) || (bbr_do_red == 0)) goto done; if (((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && bbr_prtt_slam_cwnd) || (bbr_sub_drain_slam_cwnd && (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && bbr->rc_hit_state_1 && (bbr_state_val(bbr) == BBR_SUB_DRAIN)) || ((bbr->rc_bbr_state == BBR_STATE_DRAIN) && bbr_slam_cwnd_in_main_drain)) { /* * Here we must poke at the saved cwnd * as well as the cwnd. */ cwnd = bbr->r_ctl.rc_saved_cwnd; cwnd_p = &bbr->r_ctl.rc_saved_cwnd; } else { cwnd = tp->snd_cwnd; cwnd_p = &tp->snd_cwnd; } maxseg = tp->t_maxseg - bbr->rc_last_options; /* Add the overall lr with the recovery lr */ if (bbr->r_ctl.rc_lost == 0) lr2use = 0; else if (bbr->r_ctl.rc_delivered == 0) lr2use = 1000; else { lr2use = bbr->r_ctl.rc_lost * 1000; lr2use /= bbr->r_ctl.rc_delivered; } lr2use += bbr->r_ctl.recovery_lr; acks_inflight = (flight / (maxseg * 2)); if (bbr_red_scale) { lr2use *= bbr_get_rtt(bbr, BBR_SRTT); lr2use /= bbr_red_scale; if ((bbr_red_growth_restrict) && ((bbr_get_rtt(bbr, BBR_SRTT)/bbr_red_scale) > 1)) bbr->r_ctl.restrict_growth += acks_inflight; } if (lr2use) { val = (uint64_t)cwnd * lr2use; val /= 1000; if (cwnd > val) newcwnd = roundup((cwnd - val), maxseg); else newcwnd = maxseg; } else { val = (uint64_t)cwnd * (uint64_t)bbr_red_mul; val /= (uint64_t)bbr_red_div; newcwnd = roundup((uint32_t)val, maxseg); } /* with standard delayed acks how many acks can I expect? */ if (bbr_drop_limit == 0) { /* * Anticpate how much we will * raise the cwnd based on the acks. */ if ((newcwnd + (acks_inflight * maxseg)) < get_min_cwnd(bbr)) { /* We do enforce the min (with the acks) */ newcwnd = (get_min_cwnd(bbr) - acks_inflight); } } else { /* * A strict drop limit of N is is inplace */ if (newcwnd < (bbr_drop_limit * maxseg)) { newcwnd = bbr_drop_limit * maxseg; } } /* For the next N acks do we restrict the growth */ *cwnd_p = newcwnd; if (tp->snd_cwnd > newcwnd) tp->snd_cwnd = newcwnd; bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, val, 22, (uint32_t)lr2use, bbr_get_rtt(bbr, BBR_SRTT), __LINE__); bbr->r_ctl.rc_red_cwnd_pe = bbr->r_ctl.rc_pkt_epoch; } done: bbr->r_ctl.recovery_lr = 0; if (flight <= tp->snd_cwnd) { bbr->r_wanted_output = 1; } tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); } static void bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts) { bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate); /* Limit the drop in b/w to 1/2 our current filter. */ if (bbr->r_ctl.red_bw > bbr->r_ctl.rc_bbr_cur_del_rate) bbr->r_ctl.red_bw = bbr->r_ctl.rc_bbr_cur_del_rate; if (bbr->r_ctl.red_bw < (get_filter_value(&bbr->r_ctl.rc_delrate) / 2)) bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate) / 2; tcp_bbr_tso_size_check(bbr, cts); } static void bbr_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type, struct bbr_sendmap *rsm) { struct tcp_bbr *bbr; INP_WLOCK_ASSERT(tp->t_inpcb); bbr = (struct tcp_bbr *)tp->t_fb_ptr; switch (type) { case CC_NDUPACK: if (!IN_RECOVERY(tp->t_flags)) { tp->snd_recover = tp->snd_max; /* Start a new epoch */ bbr_set_pktepoch(bbr, bbr->r_ctl.rc_rcvtime, __LINE__); if (bbr->rc_lt_is_sampling || bbr->rc_lt_use_bw) { /* * Move forward the lt epoch * so it won't count the truncated * epoch. */ bbr->r_ctl.rc_lt_epoch++; } if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { /* * Just like the policer detection code * if we are in startup we must push * forward the last startup epoch * to hide the truncated PE. */ bbr->r_ctl.rc_bbr_last_startup_epoch++; } bbr->r_ctl.rc_cwnd_on_ent = tp->snd_cwnd; ENTER_RECOVERY(tp->t_flags); bbr->rc_tlp_rtx_out = 0; bbr->r_ctl.recovery_lr = bbr->r_ctl.rc_pkt_epoch_loss_rate; tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); if (bbr->rc_inp->inp_in_hpts && ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) == 0)) { /* * When we enter recovery, we need to restart * any timers. This may mean we gain an agg * early, which will be made up for at the last * rxt out. */ bbr->rc_timer_first = 1; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); } /* * Calculate a new cwnd based on to the current * delivery rate with no gain. We get the bdp * without gaining it up like we normally would and * we use the last cur_del_rate. */ if ((bbr->rc_use_google == 0) && (bbr->r_ctl.bbr_rttprobe_gain_val || (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT))) { tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + (tp->t_maxseg - bbr->rc_last_options); if (tp->snd_cwnd < get_min_cwnd(bbr)) { /* We always gate to min cwnd */ tp->snd_cwnd = get_min_cwnd(bbr); } bbr_log_type_cwndupd(bbr, 0, 0, 0, 14, 0, 0, __LINE__); } bbr_log_type_enter_rec(bbr, rsm->r_start); } break; case CC_RTO_ERR: KMOD_TCPSTAT_INC(tcps_sndrexmitbad); /* RTO was unnecessary, so reset everything. */ bbr_reset_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime); if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { tp->snd_cwnd = tp->snd_cwnd_prev; tp->snd_ssthresh = tp->snd_ssthresh_prev; tp->snd_recover = tp->snd_recover_prev; tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent); bbr_log_type_cwndupd(bbr, 0, 0, 0, 13, 0, 0, __LINE__); } tp->t_badrxtwin = 0; break; } } /* * Indicate whether this ack should be delayed. We can delay the ack if * following conditions are met: * - There is no delayed ack timer in progress. * - Our last ack wasn't a 0-sized window. We never want to delay * the ack that opens up a 0-sized window. * - LRO wasn't used for this segment. We make sure by checking that the * segment size is not larger than the MSS. * - Delayed acks are enabled or this is a half-synchronized T/TCP * connection. * - The data being acked is less than a full segment (a stretch ack * of more than a segment we should ack. * - nsegs is 1 (if its more than that we received more than 1 ack). */ #define DELAY_ACK(tp, bbr, nsegs) \ (((tp->t_flags & TF_RXWIN0SENT) == 0) && \ ((tp->t_flags & TF_DELACK) == 0) && \ ((bbr->bbr_segs_rcvd + nsegs) < tp->t_delayed_ack) && \ (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN))) /* * Return the lowest RSM in the map of * packets still in flight that is not acked. * This should normally find on the first one * since we remove packets from the send * map after they are marked ACKED. */ static struct bbr_sendmap * bbr_find_lowest_rsm(struct tcp_bbr *bbr) { struct bbr_sendmap *rsm; /* * Walk the time-order transmitted list looking for an rsm that is * not acked. This will be the one that was sent the longest time * ago that is still outstanding. */ TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_tmap, r_tnext) { if (rsm->r_flags & BBR_ACKED) { continue; } goto finish; } finish: return (rsm); } static struct bbr_sendmap * bbr_find_high_nonack(struct tcp_bbr *bbr, struct bbr_sendmap *rsm) { struct bbr_sendmap *prsm; /* * Walk the sequence order list backward until we hit and arrive at * the highest seq not acked. In theory when this is called it * should be the last segment (which it was not). */ prsm = rsm; TAILQ_FOREACH_REVERSE_FROM(prsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { if (prsm->r_flags & (BBR_ACKED | BBR_HAS_FIN)) { continue; } return (prsm); } return (NULL); } /* * Returns to the caller the number of microseconds that * the packet can be outstanding before we think we * should have had an ack returned. */ static uint32_t bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm) { /* * lro is the flag we use to determine if we have seen reordering. * If it gets set we have seen reordering. The reorder logic either * works in one of two ways: * * If reorder-fade is configured, then we track the last time we saw * re-ordering occur. If we reach the point where enough time as * passed we no longer consider reordering has occuring. * * Or if reorder-face is 0, then once we see reordering we consider * the connection to alway be subject to reordering and just set lro * to 1. * * In the end if lro is non-zero we add the extra time for * reordering in. */ int32_t lro; uint32_t thresh, t_rxtcur; if (srtt == 0) srtt = 1; if (bbr->r_ctl.rc_reorder_ts) { if (bbr->r_ctl.rc_reorder_fade) { if (SEQ_GEQ(cts, bbr->r_ctl.rc_reorder_ts)) { lro = cts - bbr->r_ctl.rc_reorder_ts; if (lro == 0) { /* * No time as passed since the last * reorder, mark it as reordering. */ lro = 1; } } else { /* Negative time? */ lro = 0; } if (lro > bbr->r_ctl.rc_reorder_fade) { /* Turn off reordering seen too */ bbr->r_ctl.rc_reorder_ts = 0; lro = 0; } } else { /* Reodering does not fade */ lro = 1; } } else { lro = 0; } thresh = srtt + bbr->r_ctl.rc_pkt_delay; if (lro) { /* It must be set, if not you get 1/4 rtt */ if (bbr->r_ctl.rc_reorder_shift) thresh += (srtt >> bbr->r_ctl.rc_reorder_shift); else thresh += (srtt >> 2); } else { thresh += 1000; } /* We don't let the rack timeout be above a RTO */ if ((bbr->rc_tp)->t_srtt == 0) t_rxtcur = BBR_INITIAL_RTO; else t_rxtcur = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); if (thresh > t_rxtcur) { thresh = t_rxtcur; } /* And we don't want it above the RTO max either */ if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND); } bbr_log_thresh_choice(bbr, cts, thresh, lro, srtt, rsm, BBR_TO_FRM_RACK); return (thresh); } /* * Return to the caller the amount of time in mico-seconds * that should be used for the TLP timer from the last * send time of this packet. */ static uint32_t bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t srtt, uint32_t cts) { uint32_t thresh, len, maxseg, t_rxtcur; struct bbr_sendmap *prsm; if (srtt == 0) srtt = 1; if (bbr->rc_tlp_threshold) thresh = srtt + (srtt / bbr->rc_tlp_threshold); else thresh = (srtt * 2); maxseg = tp->t_maxseg - bbr->rc_last_options; /* Get the previous sent packet, if any */ len = rsm->r_end - rsm->r_start; /* 2.1 behavior */ prsm = TAILQ_PREV(rsm, bbr_head, r_tnext); if (prsm && (len <= maxseg)) { /* * Two packets outstanding, thresh should be (2*srtt) + * possible inter-packet delay (if any). */ uint32_t inter_gap = 0; int idx, nidx; idx = rsm->r_rtr_cnt - 1; nidx = prsm->r_rtr_cnt - 1; if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) { /* Yes it was sent later (or at the same time) */ inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx]; } thresh += inter_gap; } else if (len <= maxseg) { /* * Possibly compensate for delayed-ack. */ uint32_t alt_thresh; alt_thresh = srtt + (srtt / 2) + bbr_delayed_ack_time; if (alt_thresh > thresh) thresh = alt_thresh; } /* Not above the current RTO */ if (tp->t_srtt == 0) t_rxtcur = BBR_INITIAL_RTO; else t_rxtcur = TICKS_2_USEC(tp->t_rxtcur); bbr_log_thresh_choice(bbr, cts, thresh, t_rxtcur, srtt, rsm, BBR_TO_FRM_TLP); /* Not above an RTO */ if (thresh > t_rxtcur) { thresh = t_rxtcur; } /* Not above a RTO max */ if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND); } /* And now apply the user TLP min */ if (thresh < bbr_tlp_min) { thresh = bbr_tlp_min; } return (thresh); } /* * Return one of three RTTs to use (in microseconds). */ static __inline uint32_t bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type) { uint32_t f_rtt; uint32_t srtt; f_rtt = get_filter_value_small(&bbr->r_ctl.rc_rttprop); if (get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) { /* We have no rtt at all */ if (bbr->rc_tp->t_srtt == 0) f_rtt = BBR_INITIAL_RTO; else f_rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); /* * Since we don't know how good the rtt is apply a * delayed-ack min */ if (f_rtt < bbr_delayed_ack_time) { f_rtt = bbr_delayed_ack_time; } } /* Take the filter version or last measured pkt-rtt */ if (rtt_type == BBR_RTT_PROP) { srtt = f_rtt; } else if (rtt_type == BBR_RTT_PKTRTT) { if (bbr->r_ctl.rc_pkt_epoch_rtt) { srtt = bbr->r_ctl.rc_pkt_epoch_rtt; } else { /* No pkt rtt yet */ srtt = f_rtt; } } else if (rtt_type == BBR_RTT_RACK) { srtt = bbr->r_ctl.rc_last_rtt; /* We need to add in any internal delay for our timer */ if (bbr->rc_ack_was_delayed) srtt += bbr->r_ctl.rc_ack_hdwr_delay; } else if (rtt_type == BBR_SRTT) { srtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); } else { /* TSNH */ srtt = f_rtt; #ifdef BBR_INVARIANTS panic("Unknown rtt request type %d", rtt_type); #endif } return (srtt); } static int bbr_is_lost(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts) { uint32_t thresh; thresh = bbr_calc_thresh_rack(bbr, bbr_get_rtt(bbr, BBR_RTT_RACK), cts, rsm); if ((cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) >= thresh) { /* It is lost (past time) */ return (1); } return (0); } /* * Return a sendmap if we need to retransmit something. */ static struct bbr_sendmap * bbr_check_recovery_mode(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { /* * Check to see that we don't need to fall into recovery. We will * need to do so if our oldest transmit is past the time we should * have had an ack. */ struct bbr_sendmap *rsm; int32_t idx; if (TAILQ_EMPTY(&bbr->r_ctl.rc_map)) { /* Nothing outstanding that we know of */ return (NULL); } rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); if (rsm == NULL) { /* Nothing in the transmit map */ return (NULL); } if (tp->t_flags & TF_SENTFIN) { /* Fin restricted, don't find anything once a fin is sent */ return (NULL); } if (rsm->r_flags & BBR_ACKED) { /* * Ok the first one is acked (this really should not happen * since we remove the from the tmap once they are acked) */ rsm = bbr_find_lowest_rsm(bbr); if (rsm == NULL) return (NULL); } idx = rsm->r_rtr_cnt - 1; if (SEQ_LEQ(cts, rsm->r_tim_lastsent[idx])) { /* Send timestamp is the same or less? can't be ready */ return (NULL); } /* Get our RTT time */ if (bbr_is_lost(bbr, rsm, cts) && ((rsm->r_dupack >= DUP_ACK_THRESHOLD) || (rsm->r_flags & BBR_SACK_PASSED))) { if ((rsm->r_flags & BBR_MARKED_LOST) == 0) { rsm->r_flags |= BBR_MARKED_LOST; bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; } bbr_cong_signal(tp, NULL, CC_NDUPACK, rsm); #ifdef BBR_INVARIANTS if ((rsm->r_end - rsm->r_start) == 0) panic("tp:%p bbr:%p rsm:%p length is 0?", tp, bbr, rsm); #endif return (rsm); } return (NULL); } /* * RACK Timer, here we simply do logging and house keeping. * the normal bbr_output_wtime() function will call the * appropriate thing to check if we need to do a RACK retransmit. * We return 1, saying don't proceed with bbr_output_wtime only * when all timers have been stopped (destroyed PCB?). */ static int bbr_timeout_rack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { /* * This timer simply provides an internal trigger to send out data. * The check_recovery_mode call will see if there are needed * retransmissions, if so we will enter fast-recovery. The output * call may or may not do the same thing depending on sysctl * settings. */ uint32_t lost; if (bbr->rc_all_timers_stopped) { return (1); } if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { /* Its not time yet */ return (0); } BBR_STAT_INC(bbr_to_tot); lost = bbr->r_ctl.rc_lost; if (bbr->r_state && (bbr->r_state != tp->t_state)) bbr_set_state(tp, bbr, 0); bbr_log_to_event(bbr, cts, BBR_TO_FRM_RACK); if (bbr->r_ctl.rc_resend == NULL) { /* Lets do the check here */ bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); } if (bbr_policer_call_from_rack_to) bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost)); bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK; return (0); } static __inline void bbr_clone_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *nrsm, struct bbr_sendmap *rsm, uint32_t start) { int idx; nrsm->r_start = start; nrsm->r_end = rsm->r_end; nrsm->r_rtr_cnt = rsm->r_rtr_cnt; nrsm->r_flags = rsm->r_flags; /* We don't transfer forward the SYN flag */ nrsm->r_flags &= ~BBR_HAS_SYN; /* We move forward the FIN flag, not that this should happen */ rsm->r_flags &= ~BBR_HAS_FIN; nrsm->r_dupack = rsm->r_dupack; nrsm->r_rtr_bytes = 0; nrsm->r_is_gain = rsm->r_is_gain; nrsm->r_is_drain = rsm->r_is_drain; nrsm->r_delivered = rsm->r_delivered; nrsm->r_ts_valid = rsm->r_ts_valid; nrsm->r_del_ack_ts = rsm->r_del_ack_ts; nrsm->r_del_time = rsm->r_del_time; nrsm->r_app_limited = rsm->r_app_limited; nrsm->r_first_sent_time = rsm->r_first_sent_time; nrsm->r_flight_at_send = rsm->r_flight_at_send; /* We split a piece the lower section looses any just_ret flag. */ nrsm->r_bbr_state = rsm->r_bbr_state; for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; } rsm->r_end = nrsm->r_start; idx = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); idx /= 8; /* Check if we got too small */ if ((rsm->r_is_smallmap == 0) && ((rsm->r_end - rsm->r_start) <= idx)) { bbr->r_ctl.rc_num_small_maps_alloced++; rsm->r_is_smallmap = 1; } /* Check the new one as well */ if ((nrsm->r_end - nrsm->r_start) <= idx) { bbr->r_ctl.rc_num_small_maps_alloced++; nrsm->r_is_smallmap = 1; } } static int bbr_sack_mergable(struct bbr_sendmap *at, uint32_t start, uint32_t end) { /* * Given a sack block defined by * start and end, and a current postion * at. Return 1 if either side of at * would show that the block is mergable * to that side. A block to be mergable * must have overlap with the start/end * and be in the SACK'd state. */ struct bbr_sendmap *l_rsm; struct bbr_sendmap *r_rsm; /* first get the either side blocks */ l_rsm = TAILQ_PREV(at, bbr_head, r_next); r_rsm = TAILQ_NEXT(at, r_next); if (l_rsm && (l_rsm->r_flags & BBR_ACKED)) { /* Potentially mergeable */ if ((l_rsm->r_end == start) || (SEQ_LT(start, l_rsm->r_end) && SEQ_GT(end, l_rsm->r_end))) { /* * map blk |------| * sack blk |------| * * map blk |------| * sack blk |------| */ return (1); } } if (r_rsm && (r_rsm->r_flags & BBR_ACKED)) { /* Potentially mergeable */ if ((r_rsm->r_start == end) || (SEQ_LT(start, r_rsm->r_start) && SEQ_GT(end, r_rsm->r_start))) { /* * map blk |---------| * sack blk |----| * * map blk |---------| * sack blk |-------| */ return (1); } } return (0); } static struct bbr_sendmap * bbr_merge_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *l_rsm, struct bbr_sendmap *r_rsm) { /* * We are merging two ack'd RSM's, * the l_rsm is on the left (lower seq * values) and the r_rsm is on the right * (higher seq value). The simplest way * to merge these is to move the right * one into the left. I don't think there * is any reason we need to try to find * the oldest (or last oldest retransmitted). */ l_rsm->r_end = r_rsm->r_end; if (l_rsm->r_dupack < r_rsm->r_dupack) l_rsm->r_dupack = r_rsm->r_dupack; if (r_rsm->r_rtr_bytes) l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes; if (r_rsm->r_in_tmap) { /* This really should not happen */ TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, r_rsm, r_tnext); } if (r_rsm->r_app_limited) l_rsm->r_app_limited = r_rsm->r_app_limited; /* Now the flags */ if (r_rsm->r_flags & BBR_HAS_FIN) l_rsm->r_flags |= BBR_HAS_FIN; if (r_rsm->r_flags & BBR_TLP) l_rsm->r_flags |= BBR_TLP; if (r_rsm->r_flags & BBR_RWND_COLLAPSED) l_rsm->r_flags |= BBR_RWND_COLLAPSED; if (r_rsm->r_flags & BBR_MARKED_LOST) { /* This really should not happen */ bbr->r_ctl.rc_lost_bytes -= r_rsm->r_end - r_rsm->r_start; } TAILQ_REMOVE(&bbr->r_ctl.rc_map, r_rsm, r_next); if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) { /* Transfer the split limit to the map we free */ r_rsm->r_limit_type = l_rsm->r_limit_type; l_rsm->r_limit_type = 0; } bbr_free(bbr, r_rsm); return(l_rsm); } /* * TLP Timer, here we simply setup what segment we want to * have the TLP expire on, the normal bbr_output_wtime() will then * send it out. * * We return 1, saying don't proceed with bbr_output_wtime only * when all timers have been stopped (destroyed PCB?). */ static int bbr_timeout_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { /* * Tail Loss Probe. */ struct bbr_sendmap *rsm = NULL; struct socket *so; uint32_t amm; uint32_t out, avail; uint32_t maxseg; int collapsed_win = 0; if (bbr->rc_all_timers_stopped) { return (1); } if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { /* Its not time yet */ return (0); } if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT); return (1); } /* Did we somehow get into persists? */ if (bbr->rc_in_persist) { return (0); } if (bbr->r_state && (bbr->r_state != tp->t_state)) bbr_set_state(tp, bbr, 0); BBR_STAT_INC(bbr_tlp_tot); maxseg = tp->t_maxseg - bbr->rc_last_options; /* * A TLP timer has expired. We have been idle for 2 rtts. So we now * need to figure out how to force a full MSS segment out. */ so = tp->t_inpcb->inp_socket; avail = sbavail(&so->so_snd); out = ctf_outstanding(tp); if (out > tp->snd_wnd) { /* special case, we need a retransmission */ collapsed_win = 1; goto need_retran; } if (avail > out) { /* New data is available */ amm = avail - out; if (amm > maxseg) { amm = maxseg; } else if ((amm < maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) { /* not enough to fill a MTU and no-delay is off */ goto need_retran; } /* Set the send-new override */ if ((out + amm) <= tp->snd_wnd) { bbr->rc_tlp_new_data = 1; } else { goto need_retran; } bbr->r_ctl.rc_tlp_seg_send_cnt = 0; bbr->r_ctl.rc_last_tlp_seq = tp->snd_max; bbr->r_ctl.rc_tlp_send = NULL; /* cap any slots */ BBR_STAT_INC(bbr_tlp_newdata); goto send; } need_retran: /* * Ok we need to arrange the last un-acked segment to be re-sent, or * optionally the first un-acked segment. */ if (collapsed_win == 0) { rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); if (rsm && (BBR_ACKED | BBR_HAS_FIN)) { rsm = bbr_find_high_nonack(bbr, rsm); } if (rsm == NULL) { goto restore; } } else { /* * We must find the last segment * that was acceptable by the client. */ TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { if ((rsm->r_flags & BBR_RWND_COLLAPSED) == 0) { /* Found one */ break; } } if (rsm == NULL) { /* None? if so send the first */ rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); if (rsm == NULL) goto restore; } } if ((rsm->r_end - rsm->r_start) > maxseg) { /* * We need to split this the last segment in two. */ struct bbr_sendmap *nrsm; nrsm = bbr_alloc_full_limit(bbr); if (nrsm == NULL) { /* * We can't get memory to split, we can either just * not split it. Or retransmit the whole piece, lets * do the large send (BTLP :-) ). */ goto go_for_it; } bbr_clone_rsm(bbr, nrsm, rsm, (rsm->r_end - maxseg)); TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~BBR_HAS_FIN); rsm = nrsm; } go_for_it: bbr->r_ctl.rc_tlp_send = rsm; bbr->rc_tlp_rtx_out = 1; if (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) { bbr->r_ctl.rc_tlp_seg_send_cnt++; tp->t_rxtshift++; } else { bbr->r_ctl.rc_last_tlp_seq = rsm->r_start; bbr->r_ctl.rc_tlp_seg_send_cnt = 1; } send: if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) { /* * Can't [re]/transmit a segment we have retranmitted the * max times. We need the retransmit timer to take over. */ restore: bbr->rc_tlp_new_data = 0; bbr->r_ctl.rc_tlp_send = NULL; if (rsm) rsm->r_flags &= ~BBR_TLP; BBR_STAT_INC(bbr_tlp_retran_fail); return (0); } else if (rsm) { rsm->r_flags |= BBR_TLP; } if (rsm && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) && (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend)) { /* * We have retransmitted to many times for TLP. Switch to * the regular RTO timer */ goto restore; } bbr_log_to_event(bbr, cts, BBR_TO_FRM_TLP); bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; return (0); } /* * Delayed ack Timer, here we simply need to setup the * ACK_NOW flag and remove the DELACK flag. From there * the output routine will send the ack out. * * We only return 1, saying don't proceed, if all timers * are stopped (destroyed PCB?). */ static int bbr_timeout_delack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { if (bbr->rc_all_timers_stopped) { return (1); } bbr_log_to_event(bbr, cts, BBR_TO_FRM_DELACK); tp->t_flags &= ~TF_DELACK; tp->t_flags |= TF_ACKNOW; KMOD_TCPSTAT_INC(tcps_delack); bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; return (0); } /* * Here we send a KEEP-ALIVE like probe to the * peer, we do not send data. * * We only return 1, saying don't proceed, if all timers * are stopped (destroyed PCB?). */ static int bbr_timeout_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { struct tcptemp *t_template; int32_t retval = 1; if (bbr->rc_all_timers_stopped) { return (1); } if (bbr->rc_in_persist == 0) return (0); KASSERT(tp->t_inpcb != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); /* * Persistence timer into zero window. Force a byte to be output, if * possible. */ bbr_log_to_event(bbr, cts, BBR_TO_FRM_PERSIST); bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT; KMOD_TCPSTAT_INC(tcps_persisttimeo); /* * Have we exceeded the user specified progress time? */ if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT); goto out; } /* * Hack: if the peer is dead/unreachable, we do not time out if the * window is closed. After a full backoff, drop the connection if * the idle time (no responses to probes) reaches the maximum * backoff that we would use if retransmitting. */ if (tp->t_rxtshift == TCP_MAXRXTSHIFT && (ticks - tp->t_rcvtime >= tcp_maxpersistidle || ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) { KMOD_TCPSTAT_INC(tcps_persistdrop); tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT); goto out; } if ((sbavail(&bbr->rc_inp->inp_socket->so_snd) == 0) && tp->snd_una == tp->snd_max) { bbr_exit_persist(tp, bbr, cts, __LINE__); retval = 0; goto out; } /* * If the user has closed the socket then drop a persisting * connection after a much reduced timeout. */ if (tp->t_state > TCPS_CLOSE_WAIT && (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) { KMOD_TCPSTAT_INC(tcps_persistdrop); tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT); goto out; } t_template = tcpip_maketemplate(bbr->rc_inp); if (t_template) { tcp_respond(tp, t_template->tt_ipgen, &t_template->tt_t, (struct mbuf *)NULL, tp->rcv_nxt, tp->snd_una - 1, 0); /* This sends an ack */ if (tp->t_flags & TF_DELACK) tp->t_flags &= ~TF_DELACK; free(t_template, M_TEMP); } if (tp->t_rxtshift < TCP_MAXRXTSHIFT) tp->t_rxtshift++; bbr_start_hpts_timer(bbr, tp, cts, 3, 0, 0); out: return (retval); } /* * If a keepalive goes off, we had no other timers * happening. We always return 1 here since this * routine either drops the connection or sends * out a segment with respond. */ static int bbr_timeout_keepalive(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { struct tcptemp *t_template; struct inpcb *inp; if (bbr->rc_all_timers_stopped) { return (1); } bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP; inp = tp->t_inpcb; bbr_log_to_event(bbr, cts, BBR_TO_FRM_KEEP); /* * Keep-alive timer went off; send something or drop connection if * idle for too long. */ KMOD_TCPSTAT_INC(tcps_keeptimeo); if (tp->t_state < TCPS_ESTABLISHED) goto dropit; if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && tp->t_state <= TCPS_CLOSING) { if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp)) goto dropit; /* * Send a packet designed to force a response if the peer is * up and reachable: either an ACK if the connection is * still alive, or an RST if the peer has closed the * connection due to timeout or reboot. Using sequence * number tp->snd_una-1 causes the transmitted zero-length * segment to lie outside the receive window; by the * protocol spec, this requires the correspondent TCP to * respond. */ KMOD_TCPSTAT_INC(tcps_keepprobe); t_template = tcpip_maketemplate(inp); if (t_template) { tcp_respond(tp, t_template->tt_ipgen, &t_template->tt_t, (struct mbuf *)NULL, tp->rcv_nxt, tp->snd_una - 1, 0); free(t_template, M_TEMP); } } bbr_start_hpts_timer(bbr, tp, cts, 4, 0, 0); return (1); dropit: KMOD_TCPSTAT_INC(tcps_keepdrops); tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX); tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT); return (1); } /* * Retransmit helper function, clear up all the ack * flags and take care of important book keeping. */ static void bbr_remxt_tmr(struct tcpcb *tp) { /* * The retransmit timer went off, all sack'd blocks must be * un-acked. */ struct bbr_sendmap *rsm, *trsm = NULL; struct tcp_bbr *bbr; uint32_t cts, lost; bbr = (struct tcp_bbr *)tp->t_fb_ptr; cts = tcp_get_usecs(&bbr->rc_tv); lost = bbr->r_ctl.rc_lost; if (bbr->r_state && (bbr->r_state != tp->t_state)) bbr_set_state(tp, bbr, 0); TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { if (rsm->r_flags & BBR_ACKED) { uint32_t old_flags; rsm->r_dupack = 0; if (rsm->r_in_tmap == 0) { /* We must re-add it back to the tlist */ if (trsm == NULL) { TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext); } else { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, trsm, rsm, r_tnext); } rsm->r_in_tmap = 1; } old_flags = rsm->r_flags; rsm->r_flags |= BBR_RXT_CLEARED; rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS); bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__); } else { if ((tp->t_state < TCPS_ESTABLISHED) && (rsm->r_start == tp->snd_una)) { /* * Special case for TCP FO. Where * we sent more data beyond the snd_max. * We don't mark that as lost and stop here. */ break; } if ((rsm->r_flags & BBR_MARKED_LOST) == 0) { bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; } if (bbr_marks_rxt_sack_passed) { /* * With this option, we will rack out * in 1ms increments the rest of the packets. */ rsm->r_flags |= BBR_SACK_PASSED | BBR_MARKED_LOST; rsm->r_flags &= ~BBR_WAS_SACKPASS; } else { /* * With this option we only mark them lost * and remove all sack'd markings. We will run * another RXT or a TLP. This will cause * us to eventually send more based on what * ack's come in. */ rsm->r_flags |= BBR_MARKED_LOST; rsm->r_flags &= ~BBR_WAS_SACKPASS; rsm->r_flags &= ~BBR_SACK_PASSED; } } trsm = rsm; } bbr->r_ctl.rc_resend = TAILQ_FIRST(&bbr->r_ctl.rc_map); /* Clear the count (we just un-acked them) */ bbr_log_to_event(bbr, cts, BBR_TO_FRM_TMR); bbr->rc_tlp_new_data = 0; bbr->r_ctl.rc_tlp_seg_send_cnt = 0; /* zap the behindness on a rxt */ bbr->r_ctl.rc_hptsi_agg_delay = 0; bbr->r_agg_early_set = 0; bbr->r_ctl.rc_agg_early = 0; bbr->rc_tlp_rtx_out = 0; bbr->r_ctl.rc_sacked = 0; bbr->r_ctl.rc_sacklast = NULL; bbr->r_timer_override = 1; bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost)); } /* * Re-transmit timeout! If we drop the PCB we will return 1, otherwise * we will setup to retransmit the lowest seq number outstanding. */ static int bbr_timeout_rxt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { int32_t rexmt; int32_t retval = 0; bool isipv6; bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; if (bbr->rc_all_timers_stopped) { return (1); } if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_una == tp->snd_max)) { /* Nothing outstanding .. nothing to do */ return (0); } /* * Retransmission timer went off. Message has not been acked within * retransmit interval. Back off to a longer retransmit interval * and retransmit one segment. */ if (ctf_progress_timeout_check(tp, true)) { retval = 1; bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT); goto out; } bbr_remxt_tmr(tp); if ((bbr->r_ctl.rc_resend == NULL) || ((bbr->r_ctl.rc_resend->r_flags & BBR_RWND_COLLAPSED) == 0)) { /* * If the rwnd collapsed on * the one we are retransmitting * it does not count against the * rxt count. */ tp->t_rxtshift++; } if (tp->t_rxtshift > TCP_MAXRXTSHIFT) { tp->t_rxtshift = TCP_MAXRXTSHIFT; KMOD_TCPSTAT_INC(tcps_timeoutdrop); retval = 1; tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN); tcp_set_inp_to_drop(bbr->rc_inp, (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT)); goto out; } if (tp->t_state == TCPS_SYN_SENT) { /* * If the SYN was retransmitted, indicate CWND to be limited * to 1 segment in cc_conn_init(). */ tp->snd_cwnd = 1; } else if (tp->t_rxtshift == 1) { /* * first retransmit; record ssthresh and cwnd so they can be * recovered if this turns out to be a "bad" retransmit. A * retransmit is considered "bad" if an ACK for this segment * is received within RTT/2 interval; the assumption here is * that the ACK was already in flight. See "On Estimating * End-to-End Network Path Properties" by Allman and Paxson * for more details. */ tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; if (!IN_RECOVERY(tp->t_flags)) { tp->snd_cwnd_prev = tp->snd_cwnd; tp->snd_ssthresh_prev = tp->snd_ssthresh; tp->snd_recover_prev = tp->snd_recover; tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); tp->t_flags |= TF_PREVVALID; } else { tp->t_flags &= ~TF_PREVVALID; } tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; } else { tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; tp->t_flags &= ~TF_PREVVALID; } KMOD_TCPSTAT_INC(tcps_rexmttimeo); if ((tp->t_state == TCPS_SYN_SENT) || (tp->t_state == TCPS_SYN_RECEIVED)) rexmt = USEC_2_TICKS(BBR_INITIAL_RTO) * tcp_backoff[tp->t_rxtshift]; else rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; TCPT_RANGESET(tp->t_rxtcur, rexmt, MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000)); /* * We enter the path for PLMTUD if connection is established or, if * connection is FIN_WAIT_1 status, reason for the last is that if * amount of data we send is very small, we could send it in couple * of packets and process straight to FIN. In that case we won't * catch ESTABLISHED state. */ #ifdef INET6 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false; #else isipv6 = false; #endif if (((V_tcp_pmtud_blackhole_detect == 1) || (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) || (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) && ((tp->t_state == TCPS_ESTABLISHED) || (tp->t_state == TCPS_FIN_WAIT_1))) { /* * Idea here is that at each stage of mtu probe (usually, * 1448 -> 1188 -> 524) should be given 2 chances to recover * before further clamping down. 'tp->t_rxtshift % 2 == 0' * should take care of that. */ if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) == (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) && (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 && tp->t_rxtshift % 2 == 0)) { /* * Enter Path MTU Black-hole Detection mechanism: - * Disable Path MTU Discovery (IP "DF" bit). - * Reduce MTU to lower value than what we negotiated * with peer. */ if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) { /* * Record that we may have found a black * hole. */ tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE; /* Keep track of previous MSS. */ tp->t_pmtud_saved_maxseg = tp->t_maxseg; } /* * Reduce the MSS to blackhole value or to the * default in an attempt to retransmit. */ #ifdef INET6 isipv6 = bbr->r_is_v6; if (isipv6 && tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { /* Use the sysctl tuneable blackhole MSS. */ tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); } else if (isipv6) { /* Use the default MSS. */ tp->t_maxseg = V_tcp_v6mssdflt; /* * Disable Path MTU Discovery when we switch * to minmss. */ tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) { /* Use the sysctl tuneable blackhole MSS. */ tp->t_maxseg = V_tcp_pmtud_blackhole_mss; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); } else { /* Use the default MSS. */ tp->t_maxseg = V_tcp_mssdflt; /* * Disable Path MTU Discovery when we switch * to minmss. */ tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); } #endif } else { /* * If further retransmissions are still unsuccessful * with a lowered MTU, maybe this isn't a blackhole * and we restore the previous MSS and blackhole * detection flags. The limit '6' is determined by * giving each probe stage (1448, 1188, 524) 2 * chances to recover. */ if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && (tp->t_rxtshift >= 6)) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; tp->t_maxseg = tp->t_pmtud_saved_maxseg; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed); } } } /* * Disable RFC1323 and SACK if we haven't got any response to our * third SYN to work-around some broken terminal servers (most of * which have hopefully been retired) that have bad VJ header * compression code which trashes TCP segments containing * unknown-to-them TCP options. */ if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) && (tp->t_rxtshift == 3)) tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT); /* * If we backed off this far, our srtt estimate is probably bogus. * Clobber it so we'll take the next rtt measurement as our srtt; * move the current srtt into rttvar to keep the current retransmit * times until then. */ if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) { #ifdef INET6 if (bbr->r_is_v6) in6_losing(tp->t_inpcb); else #endif in_losing(tp->t_inpcb); tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT); tp->t_srtt = 0; } sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); tp->snd_recover = tp->snd_max; tp->t_flags |= TF_ACKNOW; tp->t_rtttime = 0; out: return (retval); } static int bbr_process_timers(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, uint8_t hpts_calling) { int32_t ret = 0; int32_t timers = (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK); if (timers == 0) { return (0); } if (tp->t_state == TCPS_LISTEN) { /* no timers on listen sockets */ if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) return (0); return (1); } if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { uint32_t left; if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { ret = -1; bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling); return (0); } if (hpts_calling == 0) { ret = -2; bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling); return (0); } /* * Ok our timer went off early and we are not paced false * alarm, go back to sleep. */ left = bbr->r_ctl.rc_timer_exp - cts; ret = -3; bbr_log_to_processing(bbr, cts, ret, left, hpts_calling); tcp_hpts_insert(tp->t_inpcb, HPTS_USEC_TO_SLOTS(left)); return (1); } bbr->rc_tmr_stopped = 0; bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK; if (timers & PACE_TMR_DELACK) { ret = bbr_timeout_delack(tp, bbr, cts); } else if (timers & PACE_TMR_PERSIT) { ret = bbr_timeout_persist(tp, bbr, cts); } else if (timers & PACE_TMR_RACK) { bbr->r_ctl.rc_tlp_rxt_last_time = cts; ret = bbr_timeout_rack(tp, bbr, cts); } else if (timers & PACE_TMR_TLP) { bbr->r_ctl.rc_tlp_rxt_last_time = cts; ret = bbr_timeout_tlp(tp, bbr, cts); } else if (timers & PACE_TMR_RXT) { bbr->r_ctl.rc_tlp_rxt_last_time = cts; ret = bbr_timeout_rxt(tp, bbr, cts); } else if (timers & PACE_TMR_KEEP) { ret = bbr_timeout_keepalive(tp, bbr, cts); } bbr_log_to_processing(bbr, cts, ret, timers, hpts_calling); return (ret); } static void bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts) { if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { uint8_t hpts_removed = 0; if (bbr->rc_inp->inp_in_hpts && (bbr->rc_timer_first == 1)) { /* * If we are canceling timer's when we have the * timer ahead of the output being paced. We also * must remove ourselves from the hpts. */ hpts_removed = 1; tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT); if (bbr->r_ctl.rc_last_delay_val) { /* Update the last hptsi delay too */ uint32_t time_since_send; if (TSTMP_GT(cts, bbr->rc_pacer_started)) time_since_send = cts - bbr->rc_pacer_started; else time_since_send = 0; if (bbr->r_ctl.rc_last_delay_val > time_since_send) { /* Cut down our slot time */ bbr->r_ctl.rc_last_delay_val -= time_since_send; } else { bbr->r_ctl.rc_last_delay_val = 0; } bbr->rc_pacer_started = cts; } } bbr->rc_timer_first = 0; bbr_log_to_cancel(bbr, line, cts, hpts_removed); bbr->rc_tmr_stopped = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK; bbr->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK); } } static void bbr_timer_stop(struct tcpcb *tp, uint32_t timer_type) { struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; bbr->rc_all_timers_stopped = 1; return; } /* * stop all timers always returning 0. */ static int bbr_stopall(struct tcpcb *tp) { return (0); } static void bbr_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta) { return; } /* * return true if a bbr timer (rack or tlp) is active. */ static int bbr_timer_active(struct tcpcb *tp, uint32_t timer_type) { return (0); } static uint32_t bbr_get_earliest_send_outstanding(struct tcp_bbr *bbr, struct bbr_sendmap *u_rsm, uint32_t cts) { struct bbr_sendmap *rsm; rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); if ((rsm == NULL) || (u_rsm == rsm)) return (cts); return(rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]); } static void bbr_update_rsm(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts, uint32_t pacing_time) { int32_t idx; rsm->r_rtr_cnt++; rsm->r_dupack = 0; if (rsm->r_rtr_cnt > BBR_NUM_OF_RETRANS) { rsm->r_rtr_cnt = BBR_NUM_OF_RETRANS; rsm->r_flags |= BBR_OVERMAX; } if (rsm->r_flags & BBR_RWND_COLLAPSED) { /* Take off the collapsed flag at rxt */ rsm->r_flags &= ~BBR_RWND_COLLAPSED; } if (rsm->r_flags & BBR_MARKED_LOST) { /* We have retransmitted, its no longer lost */ rsm->r_flags &= ~BBR_MARKED_LOST; bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; } if (rsm->r_flags & BBR_RXT_CLEARED) { /* * We hit a RXT timer on it and * we cleared the "acked" flag. * We now have it going back into * flight, we can remove the cleared * flag and possibly do accounting on * this piece. */ rsm->r_flags &= ~BBR_RXT_CLEARED; } if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & BBR_TLP) == 0)) { bbr->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start); rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start); } idx = rsm->r_rtr_cnt - 1; rsm->r_tim_lastsent[idx] = cts; rsm->r_pacing_delay = pacing_time; rsm->r_delivered = bbr->r_ctl.rc_delivered; rsm->r_ts_valid = bbr->rc_ts_valid; if (bbr->rc_ts_valid) rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts; if (bbr->r_ctl.r_app_limited_until) rsm->r_app_limited = 1; else rsm->r_app_limited = 0; if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) rsm->r_bbr_state = bbr_state_val(bbr); else rsm->r_bbr_state = 8; if (rsm->r_flags & BBR_ACKED) { /* Problably MTU discovery messing with us */ uint32_t old_flags; old_flags = rsm->r_flags; rsm->r_flags &= ~BBR_ACKED; bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__); bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); if (bbr->r_ctl.rc_sacked == 0) bbr->r_ctl.rc_sacklast = NULL; } if (rsm->r_in_tmap) { TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); } TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; if (rsm->r_flags & BBR_SACK_PASSED) { /* We have retransmitted due to the SACK pass */ rsm->r_flags &= ~BBR_SACK_PASSED; rsm->r_flags |= BBR_WAS_SACKPASS; } rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts); rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) { rsm->r_is_gain = 1; rsm->r_is_drain = 0; } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) { rsm->r_is_drain = 1; rsm->r_is_gain = 0; } else { rsm->r_is_drain = 0; rsm->r_is_gain = 0; } rsm->r_del_time = bbr->r_ctl.rc_del_time; /* TEMP GOOGLE CODE */ } /* * Returns 0, or the sequence where we stopped * updating. We also update the lenp to be the amount * of data left. */ static uint32_t bbr_update_entry(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts, int32_t *lenp, uint32_t pacing_time) { /* * We (re-)transmitted starting at rsm->r_start for some length * (possibly less than r_end. */ struct bbr_sendmap *nrsm; uint32_t c_end; int32_t len; len = *lenp; c_end = rsm->r_start + len; if (SEQ_GEQ(c_end, rsm->r_end)) { /* * We retransmitted the whole piece or more than the whole * slopping into the next rsm. */ bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); if (c_end == rsm->r_end) { *lenp = 0; return (0); } else { int32_t act_len; /* Hangs over the end return whats left */ act_len = rsm->r_end - rsm->r_start; *lenp = (len - act_len); return (rsm->r_end); } /* We don't get out of this block. */ } /* * Here we retransmitted less than the whole thing which means we * have to split this into what was transmitted and what was not. */ nrsm = bbr_alloc_full_limit(bbr); if (nrsm == NULL) { *lenp = 0; return (0); } /* * So here we are going to take the original rsm and make it what we * retransmitted. nrsm will be the tail portion we did not * retransmit. For example say the chunk was 1, 11 (10 bytes). And * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to * 1, 6 and the new piece will be 6, 11. */ bbr_clone_rsm(bbr, nrsm, rsm, c_end); TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); nrsm->r_dupack = 0; if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~BBR_HAS_FIN); bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); *lenp = 0; return (0); } static uint64_t bbr_get_hardware_rate(struct tcp_bbr *bbr) { uint64_t bw; bw = bbr_get_bw(bbr); bw *= (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]; bw /= (uint64_t)BBR_UNIT; return(bw); } static void bbr_setup_less_of_rate(struct tcp_bbr *bbr, uint32_t cts, uint64_t act_rate, uint64_t rate_wanted) { /* * We could not get a full gains worth * of rate. */ if (get_filter_value(&bbr->r_ctl.rc_delrate) >= act_rate) { /* we can't even get the real rate */ uint64_t red; bbr->skip_gain = 1; bbr->gain_is_limited = 0; red = get_filter_value(&bbr->r_ctl.rc_delrate) - act_rate; if (red) filter_reduce_by(&bbr->r_ctl.rc_delrate, red, cts); } else { /* We can use a lower gain */ bbr->skip_gain = 0; bbr->gain_is_limited = 1; } } static void bbr_update_hardware_pacing_rate(struct tcp_bbr *bbr, uint32_t cts) { const struct tcp_hwrate_limit_table *nrte; int error, rate = -1; if (bbr->r_ctl.crte == NULL) return; if ((bbr->rc_inp->inp_route.ro_nh == NULL) || (bbr->rc_inp->inp_route.ro_nh->nh_ifp == NULL)) { /* Lost our routes? */ /* Clear the way for a re-attempt */ bbr->bbr_attempt_hdwr_pace = 0; lost_rate: bbr->gain_is_limited = 0; bbr->skip_gain = 0; bbr->bbr_hdrw_pacing = 0; counter_u64_add(bbr_flows_whdwr_pacing, -1); counter_u64_add(bbr_flows_nohdwr_pacing, 1); tcp_bbr_tso_size_check(bbr, cts); return; } rate = bbr_get_hardware_rate(bbr); nrte = tcp_chg_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp, bbr->rc_inp->inp_route.ro_nh->nh_ifp, rate, (RS_PACING_GEQ|RS_PACING_SUB_OK), &error, NULL); if (nrte == NULL) { goto lost_rate; } if (nrte != bbr->r_ctl.crte) { bbr->r_ctl.crte = nrte; if (error == 0) { BBR_STAT_INC(bbr_hdwr_rl_mod_ok); if (bbr->r_ctl.crte->rate < rate) { /* We have a problem */ bbr_setup_less_of_rate(bbr, cts, bbr->r_ctl.crte->rate, rate); } else { /* We are good */ bbr->gain_is_limited = 0; bbr->skip_gain = 0; } } else { /* A failure should release the tag */ BBR_STAT_INC(bbr_hdwr_rl_mod_fail); bbr->gain_is_limited = 0; bbr->skip_gain = 0; bbr->bbr_hdrw_pacing = 0; } bbr_type_log_hdwr_pacing(bbr, bbr->r_ctl.crte->ptbl->rs_ifp, rate, ((bbr->r_ctl.crte == NULL) ? 0 : bbr->r_ctl.crte->rate), __LINE__, cts, error); } } static void bbr_adjust_for_hw_pacing(struct tcp_bbr *bbr, uint32_t cts) { /* * If we have hardware pacing support * we need to factor that in for our * TSO size. */ const struct tcp_hwrate_limit_table *rlp; uint32_t cur_delay, seg_sz, maxseg, new_tso, delta, hdwr_delay; if ((bbr->bbr_hdrw_pacing == 0) || (IN_RECOVERY(bbr->rc_tp->t_flags)) || (bbr->r_ctl.crte == NULL)) return; if (bbr->hw_pacing_set == 0) { /* Not yet by the hdwr pacing count delay */ return; } if (bbr_hdwr_pace_adjust == 0) { /* No adjustment */ return; } rlp = bbr->r_ctl.crte; if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; else maxseg = BBR_MIN_SEG - bbr->rc_last_options; /* * So lets first get the * time we will take between * TSO sized sends currently without * hardware help. */ cur_delay = bbr_get_pacing_delay(bbr, BBR_UNIT, bbr->r_ctl.rc_pace_max_segs, cts, 1); hdwr_delay = bbr->r_ctl.rc_pace_max_segs / maxseg; hdwr_delay *= rlp->time_between; if (cur_delay > hdwr_delay) delta = cur_delay - hdwr_delay; else delta = 0; bbr_log_type_tsosize(bbr, cts, delta, cur_delay, hdwr_delay, (bbr->r_ctl.rc_pace_max_segs / maxseg), 1); if (delta && (delta < (max(rlp->time_between, bbr->r_ctl.bbr_hptsi_segments_delay_tar)))) { /* * Now lets divide by the pacing * time between each segment the * hardware sends rounding up and * derive a bytes from that. We multiply * that by bbr_hdwr_pace_adjust to get * more bang for our buck. * * The goal is to have the software pacer * waiting no more than an additional * pacing delay if we can (without the * compensation i.e. x bbr_hdwr_pace_adjust). */ seg_sz = max(((cur_delay + rlp->time_between)/rlp->time_between), (bbr->r_ctl.rc_pace_max_segs/maxseg)); seg_sz *= bbr_hdwr_pace_adjust; if (bbr_hdwr_pace_floor && (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) { /* Currently hardware paces * out rs_min_seg segments at a time. * We need to make sure we always send at least * a full burst of bbr_hdwr_pace_floor down. */ seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg; } seg_sz *= maxseg; } else if (delta == 0) { /* * The highest pacing rate is * above our b/w gained. This means * we probably are going quite fast at * the hardware highest rate. Lets just multiply * the calculated TSO size by the * multiplier factor (its probably * 4 segments in the default config for * mlx). */ seg_sz = bbr->r_ctl.rc_pace_max_segs * bbr_hdwr_pace_adjust; if (bbr_hdwr_pace_floor && (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) { /* Currently hardware paces * out rs_min_seg segments at a time. * We need to make sure we always send at least * a full burst of bbr_hdwr_pace_floor down. */ seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg; } } else { /* * The pacing time difference is so * big that the hardware will * pace out more rapidly then we * really want and then we * will have a long delay. Lets just keep * the same TSO size so its as if * we were not using hdwr pacing (we * just gain a bit of spacing from the * hardware if seg_sz > 1). */ seg_sz = bbr->r_ctl.rc_pace_max_segs; } if (seg_sz > bbr->r_ctl.rc_pace_max_segs) new_tso = seg_sz; else new_tso = bbr->r_ctl.rc_pace_max_segs; if (new_tso >= (PACE_MAX_IP_BYTES-maxseg)) new_tso = PACE_MAX_IP_BYTES - maxseg; if (new_tso != bbr->r_ctl.rc_pace_max_segs) { bbr_log_type_tsosize(bbr, cts, new_tso, 0, bbr->r_ctl.rc_pace_max_segs, maxseg, 0); bbr->r_ctl.rc_pace_max_segs = new_tso; } } static void tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts) { uint64_t bw; uint32_t old_tso = 0, new_tso; uint32_t maxseg, bytes; uint32_t tls_seg=0; /* * Google/linux uses the following algorithm to determine * the TSO size based on the b/w of the link (from Neal Cardwell email 9/27/18): * * bytes = bw_in_bytes_per_second / 1000 * bytes = min(bytes, 64k) * tso_segs = bytes / MSS * if (bw < 1.2Mbs) * min_tso_segs = 1 * else * min_tso_segs = 2 * tso_segs = max(tso_segs, min_tso_segs) * * * Note apply a device specific limit (we apply this in the * tcp_m_copym). * Note that before the initial measurement is made google bursts out * a full iwnd just like new-reno/cubic. * * We do not use this algorithm. Instead we * use a two phased approach: * * if ( bw <= per-tcb-cross-over) * goal_tso = calculate how much with this bw we * can send in goal-time seconds. * if (goal_tso > mss) * seg = goal_tso / mss * tso = seg * mss * else * tso = mss * if (tso > per-tcb-max) * tso = per-tcb-max * else if ( bw > 512Mbps) * tso = max-tso (64k/mss) * else * goal_tso = bw / per-tcb-divsor * seg = (goal_tso + mss-1)/mss * tso = seg * mss * * if (tso < per-tcb-floor) * tso = per-tcb-floor * if (tso > per-tcb-utter_max) * tso = per-tcb-utter_max * * Note the default per-tcb-divisor is 1000 (same as google). * the goal cross over is 30Mbps however. To recreate googles * algorithm you need to set: * * cross-over = 23,168,000 bps * goal-time = 18000 * per-tcb-max = 2 * per-tcb-divisor = 1000 * per-tcb-floor = 1 * * This will get you "google bbr" behavior with respect to tso size. * * Note we do set anything TSO size until we are past the initial * window. Before that we gnerally use either a single MSS * or we use the full IW size (so we burst a IW at a time) */ if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) { maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; } else { maxseg = BBR_MIN_SEG - bbr->rc_last_options; } old_tso = bbr->r_ctl.rc_pace_max_segs; if (bbr->rc_past_init_win == 0) { /* * Not enough data has been acknowledged to make a * judgement. Set up the initial TSO based on if we * are sending a full IW at once or not. */ if (bbr->rc_use_google) bbr->r_ctl.rc_pace_max_segs = ((bbr->rc_tp->t_maxseg - bbr->rc_last_options) * 2); else if (bbr->bbr_init_win_cheat) bbr->r_ctl.rc_pace_max_segs = bbr_initial_cwnd(bbr, bbr->rc_tp); else bbr->r_ctl.rc_pace_max_segs = bbr->rc_tp->t_maxseg - bbr->rc_last_options; if (bbr->r_ctl.rc_pace_min_segs != bbr->rc_tp->t_maxseg) bbr->r_ctl.rc_pace_min_segs = bbr->rc_tp->t_maxseg; if (bbr->r_ctl.rc_pace_max_segs == 0) { bbr->r_ctl.rc_pace_max_segs = maxseg; } bbr_log_type_tsosize(bbr, cts, bbr->r_ctl.rc_pace_max_segs, tls_seg, old_tso, maxseg, 0); bbr_adjust_for_hw_pacing(bbr, cts); return; } /** * Now lets set the TSO goal based on our delivery rate in * bytes per second. Note we only do this if * we have acked at least the initial cwnd worth of data. */ bw = bbr_get_bw(bbr); if (IN_RECOVERY(bbr->rc_tp->t_flags) && (bbr->rc_use_google == 0)) { /* We clamp to one MSS in recovery */ new_tso = maxseg; } else if (bbr->rc_use_google) { int min_tso_segs; /* Google considers the gain too */ if (bbr->r_ctl.rc_bbr_hptsi_gain != BBR_UNIT) { bw *= bbr->r_ctl.rc_bbr_hptsi_gain; bw /= BBR_UNIT; } bytes = bw / 1024; if (bytes > (64 * 1024)) bytes = 64 * 1024; new_tso = bytes / maxseg; if (bw < ONE_POINT_TWO_MEG) min_tso_segs = 1; else min_tso_segs = 2; if (new_tso < min_tso_segs) new_tso = min_tso_segs; new_tso *= maxseg; } else if (bbr->rc_no_pacing) { new_tso = (PACE_MAX_IP_BYTES / maxseg) * maxseg; } else if (bw <= bbr->r_ctl.bbr_cross_over) { /* * Calculate the worse case b/w TSO if we are inserting no * more than a delay_target number of TSO's. */ uint32_t tso_len, min_tso; tso_len = bbr_get_pacing_length(bbr, BBR_UNIT, bbr->r_ctl.bbr_hptsi_segments_delay_tar, bw); if (tso_len > maxseg) { new_tso = tso_len / maxseg; if (new_tso > bbr->r_ctl.bbr_hptsi_segments_max) new_tso = bbr->r_ctl.bbr_hptsi_segments_max; new_tso *= maxseg; } else { /* * less than a full sized frame yikes.. long rtt or * low bw? */ min_tso = bbr_minseg(bbr); if ((tso_len > min_tso) && (bbr_all_get_min == 0)) new_tso = rounddown(tso_len, min_tso); else new_tso = min_tso; } } else if (bw > FIVETWELVE_MBPS) { /* * This guy is so fast b/w wise that we can TSO as large as * possible of segments that the NIC will allow. */ new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg); } else { /* * This formula is based on attempting to send a segment or * more every bbr_hptsi_per_second. The default is 1000 * which means you are targeting what you can send every 1ms * based on the peers bw. * * If the number drops to say 500, then you are looking more * at 2ms and you will raise how much we send in a single * TSO thus saving CPU (less bbr_output_wtime() calls). The * trade off of course is you will send more at once and * thus tend to clump up the sends into larger "bursts" * building a queue. */ bw /= bbr->r_ctl.bbr_hptsi_per_second; new_tso = roundup(bw, (uint64_t)maxseg); /* * Gate the floor to match what our lower than 48Mbps * algorithm does. The ceiling (bbr_hptsi_segments_max) thus * becomes the floor for this calculation. */ if (new_tso < (bbr->r_ctl.bbr_hptsi_segments_max * maxseg)) new_tso = (bbr->r_ctl.bbr_hptsi_segments_max * maxseg); } if (bbr->r_ctl.bbr_hptsi_segments_floor && (new_tso < (maxseg * bbr->r_ctl.bbr_hptsi_segments_floor))) new_tso = maxseg * bbr->r_ctl.bbr_hptsi_segments_floor; if (new_tso > PACE_MAX_IP_BYTES) new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg); /* Enforce an utter maximum. */ if (bbr->r_ctl.bbr_utter_max && (new_tso > (bbr->r_ctl.bbr_utter_max * maxseg))) { new_tso = bbr->r_ctl.bbr_utter_max * maxseg; } if (old_tso != new_tso) { /* Only log changes */ bbr_log_type_tsosize(bbr, cts, new_tso, tls_seg, old_tso, maxseg, 0); bbr->r_ctl.rc_pace_max_segs = new_tso; } /* We have hardware pacing! */ bbr_adjust_for_hw_pacing(bbr, cts); } static void bbr_log_output(struct tcp_bbr *bbr, struct tcpcb *tp, struct tcpopt *to, int32_t len, uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t cts, struct mbuf *mb, int32_t * abandon, struct bbr_sendmap *hintrsm, uint32_t delay_calc, struct sockbuf *sb) { struct bbr_sendmap *rsm, *nrsm; register uint32_t snd_max, snd_una; uint32_t pacing_time; /* * Add to the RACK log of packets in flight or retransmitted. If * there is a TS option we will use the TS echoed, if not we will * grab a TS. * * Retransmissions will increment the count and move the ts to its * proper place. Note that if options do not include TS's then we * won't be able to effectively use the ACK for an RTT on a retran. * * Notes about r_start and r_end. Lets consider a send starting at * sequence 1 for 10 bytes. In such an example the r_start would be * 1 (starting sequence) but the r_end would be r_start+len i.e. 11. * This means that r_end is actually the first sequence for the next * slot (11). * */ INP_WLOCK_ASSERT(tp->t_inpcb); if (err) { /* * We don't log errors -- we could but snd_max does not * advance in this case either. */ return; } if (th_flags & TH_RST) { /* * We don't log resets and we return immediately from * sending */ *abandon = 1; return; } snd_una = tp->snd_una; if (th_flags & (TH_SYN | TH_FIN) && (hintrsm == NULL)) { /* * The call to bbr_log_output is made before bumping * snd_max. This means we can record one extra byte on a SYN * or FIN if seq_out is adding more on and a FIN is present * (and we are not resending). */ if ((th_flags & TH_SYN) && (tp->iss == seq_out)) len++; if (th_flags & TH_FIN) len++; } if (SEQ_LEQ((seq_out + len), snd_una)) { /* Are sending an old segment to induce an ack (keep-alive)? */ return; } if (SEQ_LT(seq_out, snd_una)) { /* huh? should we panic? */ uint32_t end; end = seq_out + len; seq_out = snd_una; len = end - seq_out; } snd_max = tp->snd_max; if (len == 0) { /* We don't log zero window probes */ return; } pacing_time = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, len, cts, 1); /* First question is it a retransmission? */ if (seq_out == snd_max) { again: rsm = bbr_alloc(bbr); if (rsm == NULL) { return; } rsm->r_flags = 0; if (th_flags & TH_SYN) rsm->r_flags |= BBR_HAS_SYN; if (th_flags & TH_FIN) rsm->r_flags |= BBR_HAS_FIN; rsm->r_tim_lastsent[0] = cts; rsm->r_rtr_cnt = 1; rsm->r_rtr_bytes = 0; rsm->r_start = seq_out; rsm->r_end = rsm->r_start + len; rsm->r_dupack = 0; rsm->r_delivered = bbr->r_ctl.rc_delivered; rsm->r_pacing_delay = pacing_time; rsm->r_ts_valid = bbr->rc_ts_valid; if (bbr->rc_ts_valid) rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts; rsm->r_del_time = bbr->r_ctl.rc_del_time; if (bbr->r_ctl.r_app_limited_until) rsm->r_app_limited = 1; else rsm->r_app_limited = 0; rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts); rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); /* * Here we must also add in this rsm since snd_max * is updated after we return from a new send. */ rsm->r_flight_at_send += len; TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next); TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) rsm->r_bbr_state = bbr_state_val(bbr); else rsm->r_bbr_state = 8; if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) { rsm->r_is_gain = 1; rsm->r_is_drain = 0; } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) { rsm->r_is_drain = 1; rsm->r_is_gain = 0; } else { rsm->r_is_drain = 0; rsm->r_is_gain = 0; } return; } /* * If we reach here its a retransmission and we need to find it. */ more: if (hintrsm && (hintrsm->r_start == seq_out)) { rsm = hintrsm; hintrsm = NULL; } else if (bbr->r_ctl.rc_next) { /* We have a hint from a previous run */ rsm = bbr->r_ctl.rc_next; } else { /* No hints sorry */ rsm = NULL; } if ((rsm) && (rsm->r_start == seq_out)) { /* * We used rc_next or hintrsm to retransmit, hopefully the * likely case. */ seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time); if (len == 0) { return; } else { goto more; } } /* Ok it was not the last pointer go through it the hard way. */ TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { if (rsm->r_start == seq_out) { seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time); bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); if (len == 0) { return; } else { continue; } } if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) { /* Transmitted within this piece */ /* * Ok we must split off the front and then let the * update do the rest */ nrsm = bbr_alloc_full_limit(bbr); if (nrsm == NULL) { bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); return; } /* * copy rsm to nrsm and then trim the front of rsm * to not include this part. */ bbr_clone_rsm(bbr, nrsm, rsm, seq_out); TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~BBR_HAS_FIN); seq_out = bbr_update_entry(tp, bbr, nrsm, cts, &len, pacing_time); if (len == 0) { return; } } } /* * Hmm not found in map did they retransmit both old and on into the * new? */ if (seq_out == tp->snd_max) { goto again; } else if (SEQ_LT(seq_out, tp->snd_max)) { #ifdef BBR_INVARIANTS printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n", seq_out, len, tp->snd_una, tp->snd_max); printf("Starting Dump of all rack entries\n"); TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { printf("rsm:%p start:%u end:%u\n", rsm, rsm->r_start, rsm->r_end); } printf("Dump complete\n"); panic("seq_out not found rack:%p tp:%p", bbr, tp); #endif } else { #ifdef BBR_INVARIANTS /* * Hmm beyond sndmax? (only if we are using the new rtt-pack * flag) */ panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p", seq_out, len, tp->snd_max, tp); #endif } } static void bbr_collapse_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, int32_t rtt) { /* * Collapse timeout back the cum-ack moved. */ tp->t_rxtshift = 0; tp->t_softerror = 0; } static void tcp_bbr_xmit_timer(struct tcp_bbr *bbr, uint32_t rtt_usecs, uint32_t rsm_send_time, uint32_t r_start, uint32_t tsin) { bbr->rtt_valid = 1; bbr->r_ctl.cur_rtt = rtt_usecs; bbr->r_ctl.ts_in = tsin; if (rsm_send_time) bbr->r_ctl.cur_rtt_send_time = rsm_send_time; } static void bbr_make_timestamp_determination(struct tcp_bbr *bbr) { /** * We have in our bbr control: * 1) The timestamp we started observing cum-acks (bbr->r_ctl.bbr_ts_check_tstmp). * 2) Our timestamp indicating when we sent that packet (bbr->r_ctl.rsm->bbr_ts_check_our_cts). * 3) The current timestamp that just came in (bbr->r_ctl.last_inbound_ts) * 4) The time that the packet that generated that ack was sent (bbr->r_ctl.cur_rtt_send_time) * * Now we can calculate the time between the sends by doing: * * delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts * * And the peer's time between receiving them by doing: * * peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp * * We want to figure out if the timestamp values are in msec, 10msec or usec. * We also may find that we can't use the timestamps if say we see * that the peer_delta indicates that though we may have taken 10ms to * pace out the data, it only saw 1ms between the two packets. This would * indicate that somewhere on the path is a batching entity that is giving * out time-slices of the actual b/w. This would mean we could not use * reliably the peers timestamps. * * We expect delta > peer_delta initially. Until we figure out the * timestamp difference which we will store in bbr->r_ctl.bbr_peer_tsratio. * If we place 1000 there then its a ms vs our usec. If we place 10000 there * then its 10ms vs our usec. If the peer is running a usec clock we would * put a 1 there. If the value is faster then ours, we will disable the * use of timestamps (though we could revist this later if we find it to be not * just an isolated one or two flows)). * * To detect the batching middle boxes we will come up with our compensation and * if with it in place, we find the peer is drastically off (by some margin) in * the smaller direction, then we will assume the worst case and disable use of timestamps. * */ uint64_t delta, peer_delta, delta_up; delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts; if (delta < bbr_min_usec_delta) { /* * Have not seen a min amount of time * between our send times so we can * make a determination of the timestamp * yet. */ return; } peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp; if (peer_delta < bbr_min_peer_delta) { /* * We may have enough in the form of * our delta but the peers number * has not changed that much. It could * be its clock ratio is such that * we need more data (10ms tick) or * there may be other compression scenarios * going on. In any event we need the * spread to be larger. */ return; } /* Ok lets first see which way our delta is going */ if (peer_delta > delta) { /* Very unlikely, the peer without * compensation shows that it saw * the two sends arrive further apart * then we saw then in micro-seconds. */ if (peer_delta < (delta + ((delta * (uint64_t)1000)/ (uint64_t)bbr_delta_percent))) { /* well it looks like the peer is a micro-second clock. */ bbr->rc_ts_clock_set = 1; bbr->r_ctl.bbr_peer_tsratio = 1; } else { bbr->rc_ts_cant_be_used = 1; bbr->rc_ts_clock_set = 1; } return; } /* Ok we know that the peer_delta is smaller than our send distance */ bbr->rc_ts_clock_set = 1; /* First question is it within the percentage that they are using usec time? */ delta_up = (peer_delta * 1000) / (uint64_t)bbr_delta_percent; if ((peer_delta + delta_up) >= delta) { /* Its a usec clock */ bbr->r_ctl.bbr_peer_tsratio = 1; bbr_log_tstmp_validation(bbr, peer_delta, delta); return; } /* Ok if not usec, what about 10usec (though unlikely)? */ delta_up = (peer_delta * 1000 * 10) / (uint64_t)bbr_delta_percent; if (((peer_delta * 10) + delta_up) >= delta) { bbr->r_ctl.bbr_peer_tsratio = 10; bbr_log_tstmp_validation(bbr, peer_delta, delta); return; } /* And what about 100usec (though again unlikely)? */ delta_up = (peer_delta * 1000 * 100) / (uint64_t)bbr_delta_percent; if (((peer_delta * 100) + delta_up) >= delta) { bbr->r_ctl.bbr_peer_tsratio = 100; bbr_log_tstmp_validation(bbr, peer_delta, delta); return; } /* And how about 1 msec (the most likely one)? */ delta_up = (peer_delta * 1000 * 1000) / (uint64_t)bbr_delta_percent; if (((peer_delta * 1000) + delta_up) >= delta) { bbr->r_ctl.bbr_peer_tsratio = 1000; bbr_log_tstmp_validation(bbr, peer_delta, delta); return; } /* Ok if not msec could it be 10 msec? */ delta_up = (peer_delta * 1000 * 10000) / (uint64_t)bbr_delta_percent; if (((peer_delta * 10000) + delta_up) >= delta) { bbr->r_ctl.bbr_peer_tsratio = 10000; return; } /* If we fall down here the clock tick so slowly we can't use it */ bbr->rc_ts_cant_be_used = 1; bbr->r_ctl.bbr_peer_tsratio = 0; bbr_log_tstmp_validation(bbr, peer_delta, delta); } /* * Collect new round-trip time estimate * and update averages and current timeout. */ static void tcp_bbr_xmit_timer_commit(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts) { int32_t delta; uint32_t rtt, tsin; int32_t rtt_ticks; if (bbr->rtt_valid == 0) /* No valid sample */ return; rtt = bbr->r_ctl.cur_rtt; tsin = bbr->r_ctl.ts_in; if (bbr->rc_prtt_set_ts) { /* * We are to force feed the rttProp filter due * to an entry into PROBE_RTT. This assures * that the times are sync'd between when we * go into PROBE_RTT and the filter expiration. * * Google does not use a true filter, so they do * this implicitly since they only keep one value * and when they enter probe-rtt they update the * value to the newest rtt. */ uint32_t rtt_prop; bbr->rc_prtt_set_ts = 0; rtt_prop = get_filter_value_small(&bbr->r_ctl.rc_rttprop); if (rtt > rtt_prop) filter_increase_by_small(&bbr->r_ctl.rc_rttprop, (rtt - rtt_prop), cts); else apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); } if (bbr->rc_ack_was_delayed) rtt += bbr->r_ctl.rc_ack_hdwr_delay; if (rtt < bbr->r_ctl.rc_lowest_rtt) bbr->r_ctl.rc_lowest_rtt = rtt; bbr_log_rtt_sample(bbr, rtt, tsin); if (bbr->r_init_rtt) { /* * The initial rtt is not-trusted, nuke it and lets get * our first valid measurement in. */ bbr->r_init_rtt = 0; tp->t_srtt = 0; } if ((bbr->rc_ts_clock_set == 0) && bbr->rc_ts_valid) { /* * So we have not yet figured out * what the peers TSTMP value is * in (most likely ms). We need a * series of cum-ack's to determine * this reliably. */ if (bbr->rc_ack_is_cumack) { if (bbr->rc_ts_data_set) { /* Lets attempt to determine the timestamp granularity. */ bbr_make_timestamp_determination(bbr); } else { bbr->rc_ts_data_set = 1; bbr->r_ctl.bbr_ts_check_tstmp = bbr->r_ctl.last_inbound_ts; bbr->r_ctl.bbr_ts_check_our_cts = bbr->r_ctl.cur_rtt_send_time; } } else { /* * We have to have consecutive acks * reset any "filled" state to none. */ bbr->rc_ts_data_set = 0; } } /* Round it up */ rtt_ticks = USEC_2_TICKS((rtt + (USECS_IN_MSEC - 1))); if (rtt_ticks == 0) rtt_ticks = 1; if (tp->t_srtt != 0) { /* * srtt is stored as fixed point with 5 bits after the * binary point (i.e., scaled by 8). The following magic is * equivalent to the smoothing algorithm in rfc793 with an * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point). * Adjust rtt to origin 0. */ delta = ((rtt_ticks - 1) << TCP_DELTA_SHIFT) - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); tp->t_srtt += delta; if (tp->t_srtt <= 0) tp->t_srtt = 1; /* * We accumulate a smoothed rtt variance (actually, a * smoothed mean difference), then set the retransmit timer * to smoothed rtt + 4 times the smoothed variance. rttvar * is stored as fixed point with 4 bits after the binary * point (scaled by 16). The following is equivalent to * rfc793 smoothing with an alpha of .75 (rttvar = * rttvar*3/4 + |delta| / 4). This replaces rfc793's * wired-in beta. */ if (delta < 0) delta = -delta; delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); tp->t_rttvar += delta; if (tp->t_rttvar <= 0) tp->t_rttvar = 1; if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar) tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } else { /* * No rtt measurement yet - use the unsmoothed rtt. Set the * variance to half the rtt (so our first retransmit happens * at 3*rtt). */ tp->t_srtt = rtt_ticks << TCP_RTT_SHIFT; tp->t_rttvar = rtt_ticks << (TCP_RTTVAR_SHIFT - 1); tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } KMOD_TCPSTAT_INC(tcps_rttupdated); tp->t_rttupdated++; #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt_ticks)); #endif /* * the retransmit should happen at rtt + 4 * rttvar. Because of the * way we do the smoothing, srtt and rttvar will each average +1/2 * tick of bias. When we compute the retransmit timer, we want 1/2 * tick of rounding and 1 extra tick because of +-1/2 tick * uncertainty in the firing of the timer. The bias will give us * exactly the 1.5 tick we need. But, because the bias is * statistical, we have to test that we don't drop below the minimum * feasible timer (which is 2 ticks). */ TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), max(MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), rtt_ticks + 2), MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000)); /* * We received an ack for a packet that wasn't retransmitted; it is * probably safe to discard any error indications we've received * recently. This isn't quite right, but close enough for now (a * route might have failed after we sent a segment, and the return * path might not be symmetrical). */ tp->t_softerror = 0; rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); if (bbr->r_ctl.bbr_smallest_srtt_this_state > rtt) bbr->r_ctl.bbr_smallest_srtt_this_state = rtt; } static void bbr_earlier_retran(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t t, uint32_t cts, int ack_type) { /* * For this RSM, we acknowledged the data from a previous * transmission, not the last one we made. This means we did a false * retransmit. */ if (rsm->r_flags & BBR_HAS_FIN) { /* * The sending of the FIN often is multiple sent when we * have everything outstanding ack'd. We ignore this case * since its over now. */ return; } if (rsm->r_flags & BBR_TLP) { /* * We expect TLP's to have this occur often */ bbr->rc_tlp_rtx_out = 0; return; } if (ack_type != BBR_CUM_ACKED) { /* * If it was not a cum-ack we * don't really know for sure since * the timestamp could be from some * other transmission. */ return; } if (rsm->r_flags & BBR_WAS_SACKPASS) { /* * We retransmitted based on a sack and the earlier * retransmission ack'd it - re-ordering is occuring. */ BBR_STAT_INC(bbr_reorder_seen); bbr->r_ctl.rc_reorder_ts = cts; } /* Back down the loss count */ if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; rsm->r_flags &= ~BBR_MARKED_LOST; if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) /* LT sampling also needs adjustment */ bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; } /***** RRS HERE ************************/ /* Do we need to do this??? */ /* bbr_reset_lt_bw_sampling(bbr, cts); */ /***** RRS HERE ************************/ BBR_STAT_INC(bbr_badfr); BBR_STAT_ADD(bbr_badfr_bytes, (rsm->r_end - rsm->r_start)); } static void bbr_set_reduced_rtt(struct tcp_bbr *bbr, uint32_t cts, uint32_t line) { bbr->r_ctl.rc_rtt_shrinks = cts; if (bbr_can_force_probertt && (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) && ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) { /* * We should enter probe-rtt its been too long * since we have been there. */ bbr_enter_probe_rtt(bbr, cts, __LINE__); } else bbr_check_probe_rtt_limits(bbr, cts); } static void tcp_bbr_commit_bw(struct tcp_bbr *bbr, uint32_t cts) { uint64_t orig_bw; if (bbr->r_ctl.rc_bbr_cur_del_rate == 0) { /* We never apply a zero measurment */ bbr_log_type_bbrupd(bbr, 20, cts, 0, 0, 0, 0, 0, 0, 0, 0); return; } if (bbr->r_ctl.r_measurement_count < 0xffffffff) bbr->r_ctl.r_measurement_count++; orig_bw = get_filter_value(&bbr->r_ctl.rc_delrate); apply_filter_max(&bbr->r_ctl.rc_delrate, bbr->r_ctl.rc_bbr_cur_del_rate, bbr->r_ctl.rc_pkt_epoch); bbr_log_type_bbrupd(bbr, 21, cts, (uint32_t)orig_bw, (uint32_t)get_filter_value(&bbr->r_ctl.rc_delrate), 0, 0, 0, 0, 0, 0); if (orig_bw && (orig_bw != get_filter_value(&bbr->r_ctl.rc_delrate))) { if (bbr->bbr_hdrw_pacing) { /* * Apply a new rate to the hardware * possibly. */ bbr_update_hardware_pacing_rate(bbr, cts); } bbr_set_state_target(bbr, __LINE__); tcp_bbr_tso_size_check(bbr, cts); if (bbr->r_recovery_bw) { bbr_setup_red_bw(bbr, cts); bbr_log_type_bw_reduce(bbr, BBR_RED_BW_USELRBW); } } else if ((orig_bw == 0) && get_filter_value(&bbr->r_ctl.rc_delrate)) tcp_bbr_tso_size_check(bbr, cts); } static void bbr_nf_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts) { if (bbr->rc_in_persist == 0) { /* We log only when not in persist */ /* Translate to a Bytes Per Second */ uint64_t tim, bw, ts_diff, ts_bw; uint32_t upper, lower, delivered; if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time)) tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time); else tim = 1; /* * Now that we have processed the tim (skipping the sample * or possibly updating the time, go ahead and * calculate the cdr. */ delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered); bw = (uint64_t)delivered; bw *= (uint64_t)USECS_IN_SECOND; bw /= tim; if (bw == 0) { /* We must have a calculatable amount */ return; } upper = (bw >> 32) & 0x00000000ffffffff; lower = bw & 0x00000000ffffffff; /* * If we are using this b/w shove it in now so we * can see in the trace viewer if it gets over-ridden. */ if (rsm->r_ts_valid && bbr->rc_ts_valid && bbr->rc_ts_clock_set && (bbr->rc_ts_cant_be_used == 0) && bbr->rc_use_ts_limit) { ts_diff = max((bbr->r_ctl.last_inbound_ts - rsm->r_del_ack_ts), 1); ts_diff *= bbr->r_ctl.bbr_peer_tsratio; if ((delivered == 0) || (rtt < 1000)) { /* Can't use the ts */ bbr_log_type_bbrupd(bbr, 61, cts, ts_diff, bbr->r_ctl.last_inbound_ts, rsm->r_del_ack_ts, 0, 0, 0, 0, delivered); } else { ts_bw = (uint64_t)delivered; ts_bw *= (uint64_t)USECS_IN_SECOND; ts_bw /= ts_diff; bbr_log_type_bbrupd(bbr, 62, cts, (ts_bw >> 32), (ts_bw & 0xffffffff), 0, 0, 0, 0, ts_diff, delivered); if ((bbr->ts_can_raise) && (ts_bw > bw)) { bbr_log_type_bbrupd(bbr, 8, cts, delivered, ts_diff, (bw >> 32), (bw & 0x00000000ffffffff), 0, 0, 0, 0); bw = ts_bw; } else if (ts_bw && (ts_bw < bw)) { bbr_log_type_bbrupd(bbr, 7, cts, delivered, ts_diff, (bw >> 32), (bw & 0x00000000ffffffff), 0, 0, 0, 0); bw = ts_bw; } } } if (rsm->r_first_sent_time && TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) { uint64_t sbw, sti; /* * We use what was in flight at the time of our * send and the size of this send to figure * out what we have been sending at (amount). * For the time we take from the time of * the send of the first send outstanding * until this send plus this sends pacing * time. This gives us a good calculation * as to the rate we have been sending at. */ sbw = (uint64_t)(rsm->r_flight_at_send); sbw *= (uint64_t)USECS_IN_SECOND; sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time; sti += rsm->r_pacing_delay; sbw /= sti; if (sbw < bw) { bbr_log_type_bbrupd(bbr, 6, cts, delivered, (uint32_t)sti, (bw >> 32), (uint32_t)bw, rsm->r_first_sent_time, 0, (sbw >> 32), (uint32_t)sbw); bw = sbw; } } /* Use the google algorithm for b/w measurements */ bbr->r_ctl.rc_bbr_cur_del_rate = bw; if ((rsm->r_app_limited == 0) || (bw > get_filter_value(&bbr->r_ctl.rc_delrate))) { tcp_bbr_commit_bw(bbr, cts); bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered, 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time); } } } static void bbr_google_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts) { if (bbr->rc_in_persist == 0) { /* We log only when not in persist */ /* Translate to a Bytes Per Second */ uint64_t tim, bw; uint32_t upper, lower, delivered; int no_apply = 0; if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time)) tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time); else tim = 1; /* * Now that we have processed the tim (skipping the sample * or possibly updating the time, go ahead and * calculate the cdr. */ delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered); bw = (uint64_t)delivered; bw *= (uint64_t)USECS_IN_SECOND; bw /= tim; if (tim < bbr->r_ctl.rc_lowest_rtt) { bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered, tim, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0); no_apply = 1; } upper = (bw >> 32) & 0x00000000ffffffff; lower = bw & 0x00000000ffffffff; /* * If we are using this b/w shove it in now so we * can see in the trace viewer if it gets over-ridden. */ bbr->r_ctl.rc_bbr_cur_del_rate = bw; /* Gate by the sending rate */ if (rsm->r_first_sent_time && TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) { uint64_t sbw, sti; /* * We use what was in flight at the time of our * send and the size of this send to figure * out what we have been sending at (amount). * For the time we take from the time of * the send of the first send outstanding * until this send plus this sends pacing * time. This gives us a good calculation * as to the rate we have been sending at. */ sbw = (uint64_t)(rsm->r_flight_at_send); sbw *= (uint64_t)USECS_IN_SECOND; sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time; sti += rsm->r_pacing_delay; sbw /= sti; if (sbw < bw) { bbr_log_type_bbrupd(bbr, 6, cts, delivered, (uint32_t)sti, (bw >> 32), (uint32_t)bw, rsm->r_first_sent_time, 0, (sbw >> 32), (uint32_t)sbw); bw = sbw; } if ((sti > tim) && (sti < bbr->r_ctl.rc_lowest_rtt)) { bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered, (uint32_t)sti, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0); no_apply = 1; } else no_apply = 0; } bbr->r_ctl.rc_bbr_cur_del_rate = bw; if ((no_apply == 0) && ((rsm->r_app_limited == 0) || (bw > get_filter_value(&bbr->r_ctl.rc_delrate)))) { tcp_bbr_commit_bw(bbr, cts); bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered, 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time); } } } static void bbr_update_bbr_info(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts, uint32_t tsin, uint32_t uts, int32_t match, uint32_t rsm_send_time, int32_t ack_type, struct tcpopt *to) { uint64_t old_rttprop; /* Update our delivery time and amount */ bbr->r_ctl.rc_delivered += (rsm->r_end - rsm->r_start); bbr->r_ctl.rc_del_time = cts; if (rtt == 0) { /* * 0 means its a retransmit, for now we don't use these for * the rest of BBR. */ return; } if ((bbr->rc_use_google == 0) && (match != BBR_RTT_BY_EXACTMATCH) && (match != BBR_RTT_BY_TIMESTAMP)){ /* * We get a lot of rtt updates, lets not pay attention to * any that are not an exact match. That way we don't have * to worry about timestamps and the whole nonsense of * unsure if its a retransmission etc (if we ever had the * timestamp fixed to always have the last thing sent this * would not be a issue). */ return; } if ((bbr_no_retran && bbr->rc_use_google) && (match != BBR_RTT_BY_EXACTMATCH) && (match != BBR_RTT_BY_TIMESTAMP)){ /* * We only do measurements in google mode * with bbr_no_retran on for sure things. */ return; } /* Only update srtt if we know by exact match */ tcp_bbr_xmit_timer(bbr, rtt, rsm_send_time, rsm->r_start, tsin); if (ack_type == BBR_CUM_ACKED) bbr->rc_ack_is_cumack = 1; else bbr->rc_ack_is_cumack = 0; old_rttprop = bbr_get_rtt(bbr, BBR_RTT_PROP); /* * Note the following code differs to the original * BBR spec. It calls for <= not <. However after a * long discussion in email with Neal, he acknowledged * that it should be < than so that we will have flows * going into probe-rtt (we were seeing cases where that * did not happen and caused ugly things to occur). We * have added this agreed upon fix to our code base. */ if (rtt < old_rttprop) { /* Update when we last saw a rtt drop */ bbr_log_rtt_shrinks(bbr, cts, 0, rtt, __LINE__, BBR_RTTS_NEWRTT, 0); bbr_set_reduced_rtt(bbr, cts, __LINE__); } bbr_log_type_bbrrttprop(bbr, rtt, (rsm ? rsm->r_end : 0), uts, cts, match, rsm->r_start, rsm->r_flags); apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); if (old_rttprop != bbr_get_rtt(bbr, BBR_RTT_PROP)) { /* * The RTT-prop moved, reset the target (may be a * nop for some states). */ bbr_set_state_target(bbr, __LINE__); if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_NEW_TARGET, 0); else if (old_rttprop < bbr_get_rtt(bbr, BBR_RTT_PROP)) /* It went up */ bbr_check_probe_rtt_limits(bbr, cts); } if ((bbr->rc_use_google == 0) && (match == BBR_RTT_BY_TIMESTAMP)) { /* * We don't do b/w update with * these since they are not really * reliable. */ return; } if (bbr->r_ctl.r_app_limited_until && (bbr->r_ctl.rc_delivered >= bbr->r_ctl.r_app_limited_until)) { /* We are no longer app-limited */ bbr->r_ctl.r_app_limited_until = 0; } if (bbr->rc_use_google) { bbr_google_measurement(bbr, rsm, rtt, cts); } else { bbr_nf_measurement(bbr, rsm, rtt, cts); } } /* * Convert a timestamp that the main stack * uses (milliseconds) into one that bbr uses * (microseconds). Return that converted timestamp. */ static uint32_t bbr_ts_convert(uint32_t cts) { uint32_t sec, msec; sec = cts / MS_IN_USEC; msec = cts - (MS_IN_USEC * sec); return ((sec * USECS_IN_SECOND) + (msec * MS_IN_USEC)); } /* * Return 0 if we did not update the RTT time, return * 1 if we did. */ static int bbr_update_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, uint32_t th_ack) { int32_t i; uint32_t t, uts = 0; if ((rsm->r_flags & BBR_ACKED) || (rsm->r_flags & BBR_WAS_RENEGED) || (rsm->r_flags & BBR_RXT_CLEARED)) { /* Already done */ return (0); } if (rsm->r_rtr_cnt == 1) { /* * Only one transmit. Hopefully the normal case. */ if (TSTMP_GT(cts, rsm->r_tim_lastsent[0])) t = cts - rsm->r_tim_lastsent[0]; else t = 1; if ((int)t <= 0) t = 1; bbr->r_ctl.rc_last_rtt = t; bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0, BBR_RTT_BY_EXACTMATCH, rsm->r_tim_lastsent[0], ack_type, to); return (1); } /* Convert to usecs */ if ((bbr_can_use_ts_for_rtt == 1) && (bbr->rc_use_google == 1) && (ack_type == BBR_CUM_ACKED) && (to->to_flags & TOF_TS) && (to->to_tsecr != 0)) { t = tcp_tv_to_mssectick(&bbr->rc_tv) - to->to_tsecr; if (t < 1) t = 1; t *= MS_IN_USEC; bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0, BBR_RTT_BY_TIMESTAMP, rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)], ack_type, to); return (1); } uts = bbr_ts_convert(to->to_tsecr); if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0) && (ack_type == BBR_CUM_ACKED) && ((rsm->r_flags & BBR_OVERMAX) == 0)) { /* * Now which timestamp does it match? In this block the ACK * may be coming from a previous transmission. */ uint32_t fudge; fudge = BBR_TIMER_FUDGE; for (i = 0; i < rsm->r_rtr_cnt; i++) { if ((SEQ_GEQ(uts, (rsm->r_tim_lastsent[i] - fudge))) && (SEQ_LEQ(uts, (rsm->r_tim_lastsent[i] + fudge)))) { if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) t = cts - rsm->r_tim_lastsent[i]; else t = 1; if ((int)t <= 0) t = 1; bbr->r_ctl.rc_last_rtt = t; bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_TSMATCHING, rsm->r_tim_lastsent[i], ack_type, to); if ((i + 1) < rsm->r_rtr_cnt) { /* Likely */ bbr_earlier_retran(tp, bbr, rsm, t, cts, ack_type); } else if (rsm->r_flags & BBR_TLP) { bbr->rc_tlp_rtx_out = 0; } return (1); } } /* Fall through if we can't find a matching timestamp */ } /* * Ok its a SACK block that we retransmitted. or a windows * machine without timestamps. We can tell nothing from the * time-stamp since its not there or the time the peer last * recieved a segment that moved forward its cum-ack point. * * Lets look at the last retransmit and see what we can tell * (with BBR for space we only keep 2 note we have to keep * at least 2 so the map can not be condensed more). */ i = rsm->r_rtr_cnt - 1; if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) t = cts - rsm->r_tim_lastsent[i]; else goto not_sure; if (t < bbr->r_ctl.rc_lowest_rtt) { /* * We retransmitted and the ack came back in less * than the smallest rtt we have observed in the * windowed rtt. We most likey did an improper * retransmit as outlined in 4.2 Step 3 point 2 in * the rack-draft. * * Use the prior transmission to update all the * information as long as there is only one prior * transmission. */ if ((rsm->r_flags & BBR_OVERMAX) == 0) { #ifdef BBR_INVARIANTS if (rsm->r_rtr_cnt == 1) panic("rsm:%p bbr:%p rsm has overmax and only 1 retranmit flags:%x?", rsm, bbr, rsm->r_flags); #endif i = rsm->r_rtr_cnt - 2; if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) t = cts - rsm->r_tim_lastsent[i]; else t = 1; bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_EARLIER_RET, rsm->r_tim_lastsent[i], ack_type, to); bbr_earlier_retran(tp, bbr, rsm, t, cts, ack_type); } else { /* * Too many prior transmissions, just * updated BBR delivered */ not_sure: bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts, BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to); } } else { /* * We retransmitted it and the retransmit did the * job. */ if (rsm->r_flags & BBR_TLP) bbr->rc_tlp_rtx_out = 0; if ((rsm->r_flags & BBR_OVERMAX) == 0) bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_THIS_RETRAN, 0, ack_type, to); else bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts, BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to); return (1); } return (0); } /* * Mark the SACK_PASSED flag on all entries prior to rsm send wise. */ static void bbr_log_sack_passed(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm) { struct bbr_sendmap *nrsm; nrsm = rsm; TAILQ_FOREACH_REVERSE_FROM(nrsm, &bbr->r_ctl.rc_tmap, bbr_head, r_tnext) { if (nrsm == rsm) { /* Skip orginal segment he is acked */ continue; } if (nrsm->r_flags & BBR_ACKED) { /* Skip ack'd segments */ continue; } if (nrsm->r_flags & BBR_SACK_PASSED) { /* * We found one that is already marked * passed, we have been here before and * so all others below this are marked. */ break; } BBR_STAT_INC(bbr_sack_passed); nrsm->r_flags |= BBR_SACK_PASSED; if (((nrsm->r_flags & BBR_MARKED_LOST) == 0) && bbr_is_lost(bbr, nrsm, bbr->r_ctl.rc_rcvtime)) { bbr->r_ctl.rc_lost += nrsm->r_end - nrsm->r_start; bbr->r_ctl.rc_lost_bytes += nrsm->r_end - nrsm->r_start; nrsm->r_flags |= BBR_MARKED_LOST; } nrsm->r_flags &= ~BBR_WAS_SACKPASS; } } /* * Returns the number of bytes that were * newly ack'd by sack blocks. */ static uint32_t bbr_proc_sack_blk(struct tcpcb *tp, struct tcp_bbr *bbr, struct sackblk *sack, struct tcpopt *to, struct bbr_sendmap **prsm, uint32_t cts) { int32_t times = 0; uint32_t start, end, maxseg, changed = 0; struct bbr_sendmap *rsm, *nrsm; int32_t used_ref = 1; uint8_t went_back = 0, went_fwd = 0; maxseg = tp->t_maxseg - bbr->rc_last_options; start = sack->start; end = sack->end; rsm = *prsm; if (rsm == NULL) used_ref = 0; /* Do we locate the block behind where we last were? */ if (rsm && SEQ_LT(start, rsm->r_start)) { went_back = 1; TAILQ_FOREACH_REVERSE_FROM(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { if (SEQ_GEQ(start, rsm->r_start) && SEQ_LT(start, rsm->r_end)) { goto do_rest_ofb; } } } start_at_beginning: went_fwd = 1; /* * Ok lets locate the block where this guy is fwd from rsm (if its * set) */ TAILQ_FOREACH_FROM(rsm, &bbr->r_ctl.rc_map, r_next) { if (SEQ_GEQ(start, rsm->r_start) && SEQ_LT(start, rsm->r_end)) { break; } } do_rest_ofb: if (rsm == NULL) { /* * This happens when we get duplicate sack blocks with the * same end. For example SACK 4: 100 SACK 3: 100 The sort * will not change there location so we would just start at * the end of the first one and get lost. */ if (tp->t_flags & TF_SENTFIN) { /* * Check to see if we have not logged the FIN that * went out. */ nrsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); if (nrsm && (nrsm->r_end + 1) == tp->snd_max) { /* * Ok we did not get the FIN logged. */ nrsm->r_end++; rsm = nrsm; goto do_rest_ofb; } } if (times == 1) { #ifdef BBR_INVARIANTS panic("tp:%p bbr:%p sack:%p to:%p prsm:%p", tp, bbr, sack, to, prsm); #else goto out; #endif } times++; BBR_STAT_INC(bbr_sack_proc_restart); rsm = NULL; goto start_at_beginning; } /* Ok we have an ACK for some piece of rsm */ if (rsm->r_start != start) { /* * Need to split this in two pieces the before and after. */ if (bbr_sack_mergable(rsm, start, end)) nrsm = bbr_alloc_full_limit(bbr); else nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); if (nrsm == NULL) { /* We could not allocate ignore the sack */ struct sackblk blk; blk.start = start; blk.end = end; sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk); goto out; } bbr_clone_rsm(bbr, nrsm, rsm, start); TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~BBR_HAS_FIN); rsm = nrsm; } if (SEQ_GEQ(end, rsm->r_end)) { /* * The end of this block is either beyond this guy or right * at this guy. */ if ((rsm->r_flags & BBR_ACKED) == 0) { bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0); changed += (rsm->r_end - rsm->r_start); bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); bbr_log_sack_passed(tp, bbr, rsm); if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; } /* Is Reordering occuring? */ if (rsm->r_flags & BBR_SACK_PASSED) { BBR_STAT_INC(bbr_reorder_seen); bbr->r_ctl.rc_reorder_ts = cts; if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) /* LT sampling also needs adjustment */ bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; } } rsm->r_flags |= BBR_ACKED; rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST); if (rsm->r_in_tmap) { TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } } bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED); if (end == rsm->r_end) { /* This block only - done */ goto out; } /* There is more not coverend by this rsm move on */ start = rsm->r_end; nrsm = TAILQ_NEXT(rsm, r_next); rsm = nrsm; times = 0; goto do_rest_ofb; } if (rsm->r_flags & BBR_ACKED) { /* Been here done that */ goto out; } /* Ok we need to split off this one at the tail */ if (bbr_sack_mergable(rsm, start, end)) nrsm = bbr_alloc_full_limit(bbr); else nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); if (nrsm == NULL) { /* failed XXXrrs what can we do but loose the sack info? */ struct sackblk blk; blk.start = start; blk.end = end; sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk); goto out; } /* Clone it */ bbr_clone_rsm(bbr, nrsm, rsm, end); /* The sack block does not cover this guy fully */ rsm->r_flags &= (~BBR_HAS_FIN); TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } nrsm->r_dupack = 0; bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0); bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED); changed += (rsm->r_end - rsm->r_start); bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); bbr_log_sack_passed(tp, bbr, rsm); /* Is Reordering occuring? */ if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; } if (rsm->r_flags & BBR_SACK_PASSED) { BBR_STAT_INC(bbr_reorder_seen); bbr->r_ctl.rc_reorder_ts = cts; if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) /* LT sampling also needs adjustment */ bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; } } rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST); rsm->r_flags |= BBR_ACKED; if (rsm->r_in_tmap) { TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } out: if (rsm && (rsm->r_flags & BBR_ACKED)) { /* * Now can we merge this newly acked * block with either the previous or * next block? */ nrsm = TAILQ_NEXT(rsm, r_next); if (nrsm && (nrsm->r_flags & BBR_ACKED)) { /* yep this and next can be merged */ rsm = bbr_merge_rsm(bbr, rsm, nrsm); } /* Now what about the previous? */ nrsm = TAILQ_PREV(rsm, bbr_head, r_next); if (nrsm && (nrsm->r_flags & BBR_ACKED)) { /* yep the previous and this can be merged */ rsm = bbr_merge_rsm(bbr, nrsm, rsm); } } if (used_ref == 0) { BBR_STAT_INC(bbr_sack_proc_all); } else { BBR_STAT_INC(bbr_sack_proc_short); } if (went_fwd && went_back) { BBR_STAT_INC(bbr_sack_search_both); } else if (went_fwd) { BBR_STAT_INC(bbr_sack_search_fwd); } else if (went_back) { BBR_STAT_INC(bbr_sack_search_back); } /* Save off where the next seq is */ if (rsm) bbr->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next); else bbr->r_ctl.rc_sacklast = NULL; *prsm = rsm; return (changed); } static void inline bbr_peer_reneges(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, tcp_seq th_ack) { struct bbr_sendmap *tmap; BBR_STAT_INC(bbr_reneges_seen); tmap = NULL; while (rsm && (rsm->r_flags & BBR_ACKED)) { /* Its no longer sacked, mark it so */ uint32_t oflags; bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); #ifdef BBR_INVARIANTS if (rsm->r_in_tmap) { panic("bbr:%p rsm:%p flags:0x%x in tmap?", bbr, rsm, rsm->r_flags); } #endif oflags = rsm->r_flags; if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) /* LT sampling also needs adjustment */ bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; } rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS | BBR_MARKED_LOST); rsm->r_flags |= BBR_WAS_RENEGED; rsm->r_flags |= BBR_RXT_CLEARED; bbr_log_type_rsmclear(bbr, bbr->r_ctl.rc_rcvtime, rsm, oflags, __LINE__); /* Rebuild it into our tmap */ if (tmap == NULL) { TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext); tmap = rsm; } else { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, tmap, rsm, r_tnext); tmap = rsm; } tmap->r_in_tmap = 1; /* * XXXrrs Delivered? Should we do anything here? * * Of course we don't on a rxt timeout so maybe its ok that * we don't? * * For now lets not. */ rsm = TAILQ_NEXT(rsm, r_next); } /* * Now lets possibly clear the sack filter so we start recognizing * sacks that cover this area. */ sack_filter_clear(&bbr->r_ctl.bbr_sf, th_ack); } static void bbr_log_syn(struct tcpcb *tp, struct tcpopt *to) { struct tcp_bbr *bbr; struct bbr_sendmap *rsm; uint32_t cts; bbr = (struct tcp_bbr *)tp->t_fb_ptr; cts = bbr->r_ctl.rc_rcvtime; rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); if (rsm && (rsm->r_flags & BBR_HAS_SYN)) { if ((rsm->r_end - rsm->r_start) <= 1) { /* Log out the SYN completely */ bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; rsm->r_rtr_bytes = 0; TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); if (rsm->r_in_tmap) { TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } if (bbr->r_ctl.rc_next == rsm) { /* scoot along the marker */ bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map); } if (to != NULL) bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, 0); bbr_free(bbr, rsm); } else { /* There is more (Fast open)? strip out SYN. */ rsm->r_flags &= ~BBR_HAS_SYN; rsm->r_start++; } } } /* * Returns the number of bytes that were * acknowledged by SACK blocks. */ static uint32_t bbr_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, uint32_t *prev_acked) { uint32_t changed, last_seq, entered_recovery = 0; struct tcp_bbr *bbr; struct bbr_sendmap *rsm; struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1]; register uint32_t th_ack; int32_t i, j, k, new_sb, num_sack_blks = 0; uint32_t cts, acked, ack_point, sack_changed = 0; uint32_t p_maxseg, maxseg, p_acked = 0; INP_WLOCK_ASSERT(tp->t_inpcb); if (th->th_flags & TH_RST) { /* We don't log resets */ return (0); } bbr = (struct tcp_bbr *)tp->t_fb_ptr; cts = bbr->r_ctl.rc_rcvtime; rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); changed = 0; maxseg = tp->t_maxseg - bbr->rc_last_options; p_maxseg = min(bbr->r_ctl.rc_pace_max_segs, maxseg); th_ack = th->th_ack; if (SEQ_GT(th_ack, tp->snd_una)) { acked = th_ack - tp->snd_una; bbr_log_progress_event(bbr, tp, ticks, PROGRESS_UPDATE, __LINE__); bbr->rc_tp->t_acktime = ticks; } else acked = 0; if (SEQ_LEQ(th_ack, tp->snd_una)) { /* Only sent here for sack processing */ goto proc_sack; } if (rsm && SEQ_GT(th_ack, rsm->r_start)) { changed = th_ack - rsm->r_start; } else if ((rsm == NULL) && ((th_ack - 1) == tp->iss)) { /* * For the SYN incoming case we will not have called * tcp_output for the sending of the SYN, so there will be * no map. All other cases should probably be a panic. */ if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0)) { /* * We have a timestamp that can be used to generate * an initial RTT. */ uint32_t ts, now, rtt; ts = bbr_ts_convert(to->to_tsecr); now = bbr_ts_convert(tcp_tv_to_mssectick(&bbr->rc_tv)); rtt = now - ts; if (rtt < 1) rtt = 1; bbr_log_type_bbrrttprop(bbr, rtt, tp->iss, 0, cts, BBR_RTT_BY_TIMESTAMP, tp->iss, 0); apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); changed = 1; bbr->r_wanted_output = 1; goto out; } goto proc_sack; } else if (rsm == NULL) { goto out; } if (changed) { /* * The ACK point is advancing to th_ack, we must drop off * the packets in the rack log and calculate any eligble * RTT's. */ bbr->r_wanted_output = 1; more: if (rsm == NULL) { if (tp->t_flags & TF_SENTFIN) { /* if we send a FIN we will not hav a map */ goto proc_sack; } #ifdef BBR_INVARIANTS panic("No rack map tp:%p for th:%p state:%d bbr:%p snd_una:%u snd_max:%u chg:%d\n", tp, th, tp->t_state, bbr, tp->snd_una, tp->snd_max, changed); #endif goto proc_sack; } } if (SEQ_LT(th_ack, rsm->r_start)) { /* Huh map is missing this */ #ifdef BBR_INVARIANTS printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d bbr:%p\n", rsm->r_start, th_ack, tp->t_state, bbr->r_state, bbr); panic("th-ack is bad bbr:%p tp:%p", bbr, tp); #endif goto proc_sack; } else if (th_ack == rsm->r_start) { /* None here to ack */ goto proc_sack; } /* * Clear the dup ack counter, it will * either be freed or if there is some * remaining we need to start it at zero. */ rsm->r_dupack = 0; /* Now do we consume the whole thing? */ if (SEQ_GEQ(th_ack, rsm->r_end)) { /* Its all consumed. */ uint32_t left; if (rsm->r_flags & BBR_ACKED) { /* * It was acked on the scoreboard -- remove it from * total */ p_acked += (rsm->r_end - rsm->r_start); bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); if (bbr->r_ctl.rc_sacked == 0) bbr->r_ctl.rc_sacklast = NULL; } else { bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, th_ack); if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; } if (rsm->r_flags & BBR_SACK_PASSED) { /* * There are acked segments ACKED on the * scoreboard further up. We are seeing * reordering. */ BBR_STAT_INC(bbr_reorder_seen); bbr->r_ctl.rc_reorder_ts = cts; if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) /* LT sampling also needs adjustment */ bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; } } rsm->r_flags &= ~BBR_MARKED_LOST; } bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; rsm->r_rtr_bytes = 0; TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); if (rsm->r_in_tmap) { TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } if (bbr->r_ctl.rc_next == rsm) { /* scoot along the marker */ bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map); } bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED); /* Adjust the packet counts */ left = th_ack - rsm->r_end; /* Free back to zone */ bbr_free(bbr, rsm); if (left) { rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); goto more; } goto proc_sack; } if (rsm->r_flags & BBR_ACKED) { /* * It was acked on the scoreboard -- remove it from total * for the part being cum-acked. */ p_acked += (rsm->r_end - rsm->r_start); bbr->r_ctl.rc_sacked -= (th_ack - rsm->r_start); if (bbr->r_ctl.rc_sacked == 0) bbr->r_ctl.rc_sacklast = NULL; } else { /* * It was acked up to th_ack point for the first time */ struct bbr_sendmap lrsm; memcpy(&lrsm, rsm, sizeof(struct bbr_sendmap)); lrsm.r_end = th_ack; bbr_update_rtt(tp, bbr, &lrsm, to, cts, BBR_CUM_ACKED, th_ack); } if ((rsm->r_flags & BBR_MARKED_LOST) && ((rsm->r_flags & BBR_ACKED) == 0)) { /* * It was marked lost and partly ack'd now * for the first time. We lower the rc_lost_bytes * and still leave it MARKED. */ bbr->r_ctl.rc_lost_bytes -= th_ack - rsm->r_start; } bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED); bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; rsm->r_rtr_bytes = 0; /* adjust packet count */ rsm->r_start = th_ack; proc_sack: /* Check for reneging */ rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); if (rsm && (rsm->r_flags & BBR_ACKED) && (th_ack == rsm->r_start)) { /* * The peer has moved snd_una up to the edge of this send, * i.e. one that it had previously acked. The only way that * can be true if the peer threw away data (space issues) * that it had previously sacked (else it would have given * us snd_una up to (rsm->r_end). We need to undo the acked * markings here. * * Note we have to look to make sure th_ack is our * rsm->r_start in case we get an old ack where th_ack is * behind snd_una. */ bbr_peer_reneges(bbr, rsm, th->th_ack); } if ((to->to_flags & TOF_SACK) == 0) { /* We are done nothing left to log */ goto out; } rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); if (rsm) { last_seq = rsm->r_end; } else { last_seq = tp->snd_max; } /* Sack block processing */ if (SEQ_GT(th_ack, tp->snd_una)) ack_point = th_ack; else ack_point = tp->snd_una; for (i = 0; i < to->to_nsacks; i++) { bcopy((to->to_sacks + i * TCPOLEN_SACK), &sack, sizeof(sack)); sack.start = ntohl(sack.start); sack.end = ntohl(sack.end); if (SEQ_GT(sack.end, sack.start) && SEQ_GT(sack.start, ack_point) && SEQ_LT(sack.start, tp->snd_max) && SEQ_GT(sack.end, ack_point) && SEQ_LEQ(sack.end, tp->snd_max)) { if ((bbr->r_ctl.rc_num_small_maps_alloced > bbr_sack_block_limit) && (SEQ_LT(sack.end, last_seq)) && ((sack.end - sack.start) < (p_maxseg / 8))) { /* * Not the last piece and its smaller than * 1/8th of a p_maxseg. We ignore this. */ BBR_STAT_INC(bbr_runt_sacks); continue; } sack_blocks[num_sack_blks] = sack; num_sack_blks++; #ifdef NETFLIX_STATS } else if (SEQ_LEQ(sack.start, th_ack) && SEQ_LEQ(sack.end, th_ack)) { /* * Its a D-SACK block. */ tcp_record_dsack(sack.start, sack.end); #endif } } if (num_sack_blks == 0) goto out; /* * Sort the SACK blocks so we can update the rack scoreboard with * just one pass. */ new_sb = sack_filter_blks(&bbr->r_ctl.bbr_sf, sack_blocks, num_sack_blks, th->th_ack); ctf_log_sack_filter(bbr->rc_tp, new_sb, sack_blocks); BBR_STAT_ADD(bbr_sack_blocks, num_sack_blks); BBR_STAT_ADD(bbr_sack_blocks_skip, (num_sack_blks - new_sb)); num_sack_blks = new_sb; if (num_sack_blks < 2) { goto do_sack_work; } /* Sort the sacks */ for (i = 0; i < num_sack_blks; i++) { for (j = i + 1; j < num_sack_blks; j++) { if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { sack = sack_blocks[i]; sack_blocks[i] = sack_blocks[j]; sack_blocks[j] = sack; } } } /* * Now are any of the sack block ends the same (yes some * implememtations send these)? */ again: if (num_sack_blks > 1) { for (i = 0; i < num_sack_blks; i++) { for (j = i + 1; j < num_sack_blks; j++) { if (sack_blocks[i].end == sack_blocks[j].end) { /* * Ok these two have the same end we * want the smallest end and then * throw away the larger and start * again. */ if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) { /* * The second block covers * more area use that */ sack_blocks[i].start = sack_blocks[j].start; } /* * Now collapse out the dup-sack and * lower the count */ for (k = (j + 1); k < num_sack_blks; k++) { sack_blocks[j].start = sack_blocks[k].start; sack_blocks[j].end = sack_blocks[k].end; j++; } num_sack_blks--; goto again; } } } } do_sack_work: rsm = bbr->r_ctl.rc_sacklast; for (i = 0; i < num_sack_blks; i++) { acked = bbr_proc_sack_blk(tp, bbr, &sack_blocks[i], to, &rsm, cts); if (acked) { bbr->r_wanted_output = 1; changed += acked; sack_changed += acked; } } out: *prev_acked = p_acked; if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) { /* * Ok we have a high probability that we need to go in to * recovery since we have data sack'd */ struct bbr_sendmap *rsm; rsm = bbr_check_recovery_mode(tp, bbr, cts); if (rsm) { /* Enter recovery */ entered_recovery = 1; bbr->r_wanted_output = 1; /* * When we enter recovery we need to assure we send * one packet. */ if (bbr->r_ctl.rc_resend == NULL) { bbr->r_ctl.rc_resend = rsm; } } } if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) { /* * See if we need to rack-retransmit anything if so set it * up as the thing to resend assuming something else is not * already in that position. */ if (bbr->r_ctl.rc_resend == NULL) { bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); } } /* * We return the amount that changed via sack, this is used by the * ack-received code to augment what was changed between th_ack <-> * snd_una. */ return (sack_changed); } static void bbr_strike_dupack(struct tcp_bbr *bbr) { struct bbr_sendmap *rsm; rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); if (rsm && (rsm->r_dupack < 0xff)) { rsm->r_dupack++; if (rsm->r_dupack >= DUP_ACK_THRESHOLD) bbr->r_wanted_output = 1; } } /* * Return value of 1, we do not need to call bbr_process_data(). * return value of 0, bbr_process_data can be called. * For ret_val if its 0 the TCB is locked and valid, if its non-zero * its unlocked and probably unsafe to touch the TCB. */ static int bbr_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val) { int32_t ourfinisacked = 0; int32_t acked_amount; uint16_t nsegs; int32_t acked; uint32_t lost, sack_changed = 0; struct mbuf *mfree; struct tcp_bbr *bbr; uint32_t prev_acked = 0; bbr = (struct tcp_bbr *)tp->t_fb_ptr; lost = bbr->r_ctl.rc_lost; nsegs = max(1, m->m_pkthdr.lro_nsegs); if (SEQ_GT(th->th_ack, tp->snd_max)) { ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val); bbr->r_wanted_output = 1; return (1); } if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { /* Process the ack */ if (bbr->rc_in_persist) tp->t_rxtshift = 0; if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd)) bbr_strike_dupack(bbr); sack_changed = bbr_log_ack(tp, to, th, &prev_acked); } bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, (bbr->r_ctl.rc_lost > lost)); if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { /* * Old ack, behind the last one rcv'd or a duplicate ack * with SACK info. */ if (th->th_ack == tp->snd_una) { bbr_ack_received(tp, bbr, th, 0, sack_changed, prev_acked, __LINE__, 0); if (bbr->r_state == TCPS_SYN_SENT) { /* * Special case on where we sent SYN. When * the SYN-ACK is processed in syn_sent * state it bumps the snd_una. This causes * us to hit here even though we did ack 1 * byte. * * Go through the nothing left case so we * send data. */ goto nothing_left; } } return (0); } /* * If we reach this point, ACK is not a duplicate, i.e., it ACKs * something we sent. */ if (tp->t_flags & TF_NEEDSYN) { /* * T/TCP: Connection was half-synchronized, and our SYN has * been ACK'd (so connection is now fully synchronized). Go * to non-starred state, increment snd_una for ACK of SYN, * and check if we can do window scaling. */ tp->t_flags &= ~TF_NEEDSYN; tp->snd_una++; /* Do window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; /* Send window already scaled. */ } } INP_WLOCK_ASSERT(tp->t_inpcb); acked = BYTES_THIS_ACK(tp, th); KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); /* * If we just performed our first retransmit, and the ACK arrives * within our recovery window, then it was a mistake to do the * retransmit in the first place. Recover our original cwnd and * ssthresh, and proceed to transmit where we left off. */ if (tp->t_flags & TF_PREVVALID) { tp->t_flags &= ~TF_PREVVALID; if (tp->t_rxtshift == 1 && (int)(ticks - tp->t_badrxtwin) < 0) bbr_cong_signal(tp, th, CC_RTO_ERR, NULL); } SOCKBUF_LOCK(&so->so_snd); acked_amount = min(acked, (int)sbavail(&so->so_snd)); tp->snd_wnd -= acked_amount; mfree = sbcut_locked(&so->so_snd, acked_amount); SOCKBUF_UNLOCK(&so->so_snd); tp->t_flags |= TF_WAKESOW; m_freem(mfree); if (SEQ_GT(th->th_ack, tp->snd_una)) { bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp)); } tp->snd_una = th->th_ack; bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, (bbr->r_ctl.rc_lost - lost)); if (IN_RECOVERY(tp->t_flags)) { if (SEQ_LT(th->th_ack, tp->snd_recover) && (SEQ_LT(th->th_ack, tp->snd_max))) { tcp_bbr_partialack(tp); } else { bbr_post_recovery(tp); } } if (SEQ_GT(tp->snd_una, tp->snd_recover)) { tp->snd_recover = tp->snd_una; } if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { tp->snd_nxt = tp->snd_max; } if (tp->snd_una == tp->snd_max) { /* Nothing left outstanding */ nothing_left: bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__); if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0) bbr->rc_tp->t_acktime = 0; if ((sbused(&so->so_snd) == 0) && (tp->t_flags & TF_SENTFIN)) { ourfinisacked = 1; } bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); if (bbr->rc_in_persist == 0) { bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime; } sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime); /* * We invalidate the last ack here since we * don't want to transfer forward the time * for our sum's calculations. */ if ((tp->t_state >= TCPS_FIN_WAIT_1) && (sbavail(&so->so_snd) == 0) && (tp->t_flags2 & TF2_DROP_AF_DATA)) { /* * The socket was gone and the peer sent data, time * to reset him. */ *ret_val = 1; tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); /* tcp_close will kill the inp pre-log the Reset */ tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); tp = tcp_close(tp); ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen); BBR_STAT_INC(bbr_dropped_af_data); return (1); } /* Set need output so persist might get set */ bbr->r_wanted_output = 1; } if (ofia) *ofia = ourfinisacked; return (0); } static void bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line) { if (bbr->rc_in_persist == 0) { bbr_timer_cancel(bbr, __LINE__, cts); bbr->r_ctl.rc_last_delay_val = 0; tp->t_rxtshift = 0; bbr->rc_in_persist = 1; bbr->r_ctl.rc_went_idle_time = cts; /* We should be capped when rw went to 0 but just in case */ bbr_log_type_pesist(bbr, cts, 0, line, 1); /* Time freezes for the state, so do the accounting now */ if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { uint32_t time_in; time_in = cts - bbr->r_ctl.rc_bbr_state_time; if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { int32_t idx; idx = bbr_state_val(bbr); counter_u64_add(bbr_state_time[(idx + 5)], time_in); } else { counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); } } bbr->r_ctl.rc_bbr_state_time = cts; } } static void bbr_restart_after_idle(struct tcp_bbr *bbr, uint32_t cts, uint32_t idle_time) { /* * Note that if idle time does not exceed our * threshold, we do nothing continuing the state * transitions we were last walking through. */ if (idle_time >= bbr_idle_restart_threshold) { if (bbr->rc_use_idle_restart) { bbr->rc_bbr_state = BBR_STATE_IDLE_EXIT; /* * Set our target using BBR_UNIT, so * we increase at a dramatic rate but * we stop when we get the pipe * full again for our current b/w estimate. */ bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; bbr_set_state_target(bbr, __LINE__); /* Now setup our gains to ramp up */ bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; bbr_log_type_statechange(bbr, cts, __LINE__); } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { bbr_substate_change(bbr, cts, __LINE__, 1); } } } static void bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line) { uint32_t idle_time; if (bbr->rc_in_persist == 0) return; idle_time = bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time); bbr->rc_in_persist = 0; bbr->rc_hit_state_1 = 0; bbr->r_ctl.rc_del_time = cts; /* * We invalidate the last ack here since we * don't want to transfer forward the time * for our sum's calculations. */ if (bbr->rc_inp->inp_in_hpts) { tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT); bbr->rc_timer_first = 0; bbr->r_ctl.rc_hpts_flags = 0; bbr->r_ctl.rc_last_delay_val = 0; bbr->r_ctl.rc_hptsi_agg_delay = 0; bbr->r_agg_early_set = 0; bbr->r_ctl.rc_agg_early = 0; } bbr_log_type_pesist(bbr, cts, idle_time, line, 0); if (idle_time >= bbr_rtt_probe_time) { /* * This qualifies as a RTT_PROBE session since we drop the * data outstanding to nothing and waited more than * bbr_rtt_probe_time. */ bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_PERSIST, 0); bbr->r_ctl.last_in_probertt = bbr->r_ctl.rc_rtt_shrinks = cts; } tp->t_rxtshift = 0; /* * If in probeBW and we have persisted more than an RTT lets do * special handling. */ /* Force a time based epoch */ bbr_set_epoch(bbr, cts, __LINE__); /* * Setup the lost so we don't count anything against the guy * we have been stuck with during persists. */ bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; /* Time un-freezes for the state */ bbr->r_ctl.rc_bbr_state_time = cts; if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) || (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)) { /* * If we are going back to probe-bw * or probe_rtt, we may need to possibly * do a fast restart. */ bbr_restart_after_idle(bbr, cts, idle_time); } } static void bbr_collapsed_window(struct tcp_bbr *bbr) { /* * Now we must walk the * send map and divide the * ones left stranded. These * guys can't cause us to abort * the connection and are really * "unsent". However if a buggy * client actually did keep some * of the data i.e. collapsed the win * and refused to ack and then opened * the win and acked that data. We would * get into an ack war, the simplier * method then of just pretending we * did not send those segments something * won't work. */ struct bbr_sendmap *rsm, *nrsm; tcp_seq max_seq; uint32_t maxseg; int can_split = 0; int fnd = 0; maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; max_seq = bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd; bbr_log_type_rwnd_collapse(bbr, max_seq, 1, 0); TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { /* Find the first seq past or at maxseq */ if (rsm->r_flags & BBR_RWND_COLLAPSED) rsm->r_flags &= ~BBR_RWND_COLLAPSED; if (SEQ_GEQ(max_seq, rsm->r_start) && SEQ_GEQ(rsm->r_end, max_seq)) { fnd = 1; break; } } bbr->rc_has_collapsed = 0; if (!fnd) { /* Nothing to do strange */ return; } /* * Now can we split? * * We don't want to split if splitting * would generate too many small segments * less we let an attacker fragment our * send_map and leave us out of memory. */ if ((max_seq != rsm->r_start) && (max_seq != rsm->r_end)){ /* can we split? */ int res1, res2; res1 = max_seq - rsm->r_start; res2 = rsm->r_end - max_seq; if ((res1 >= (maxseg/8)) && (res2 >= (maxseg/8))) { /* No small pieces here */ can_split = 1; } else if (bbr->r_ctl.rc_num_small_maps_alloced < bbr_sack_block_limit) { /* We are under the limit */ can_split = 1; } } /* Ok do we need to split this rsm? */ if (max_seq == rsm->r_start) { /* It's this guy no split required */ nrsm = rsm; } else if (max_seq == rsm->r_end) { /* It's the next one no split required. */ nrsm = TAILQ_NEXT(rsm, r_next); if (nrsm == NULL) { /* Huh? */ return; } } else if (can_split && SEQ_LT(max_seq, rsm->r_end)) { /* yep we need to split it */ nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); if (nrsm == NULL) { /* failed XXXrrs what can we do mark the whole? */ nrsm = rsm; goto no_split; } /* Clone it */ bbr_log_type_rwnd_collapse(bbr, max_seq, 3, 0); bbr_clone_rsm(bbr, nrsm, rsm, max_seq); TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } } else { /* * Split not allowed just start here just * use this guy. */ nrsm = rsm; } no_split: BBR_STAT_INC(bbr_collapsed_win); /* reuse fnd as a count */ fnd = 0; TAILQ_FOREACH_FROM(nrsm, &bbr->r_ctl.rc_map, r_next) { nrsm->r_flags |= BBR_RWND_COLLAPSED; fnd++; bbr->rc_has_collapsed = 1; } bbr_log_type_rwnd_collapse(bbr, max_seq, 4, fnd); } static void bbr_un_collapse_window(struct tcp_bbr *bbr) { struct bbr_sendmap *rsm; int cleared = 0; TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { if (rsm->r_flags & BBR_RWND_COLLAPSED) { /* Clear the flag */ rsm->r_flags &= ~BBR_RWND_COLLAPSED; cleared++; } else break; } bbr_log_type_rwnd_collapse(bbr, (bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd), 0, cleared); bbr->rc_has_collapsed = 0; } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) { /* * Update window information. Don't look at window if no ACK: TAC's * send garbage on first SYN. */ uint16_t nsegs; int32_t tfo_syn; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; INP_WLOCK_ASSERT(tp->t_inpcb); nsegs = max(1, m->m_pkthdr.lro_nsegs); if ((thflags & TH_ACK) && (SEQ_LT(tp->snd_wl1, th->th_seq) || (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { /* keep track of pure window updates */ if (tlen == 0 && tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) KMOD_TCPSTAT_INC(tcps_rcvwinupd); tp->snd_wnd = tiwin; tp->snd_wl1 = th->th_seq; tp->snd_wl2 = th->th_ack; if (tp->snd_wnd > tp->max_sndwnd) tp->max_sndwnd = tp->snd_wnd; bbr->r_wanted_output = 1; } else if (thflags & TH_ACK) { if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) { tp->snd_wnd = tiwin; tp->snd_wl1 = th->th_seq; tp->snd_wl2 = th->th_ack; } } if (tp->snd_wnd < ctf_outstanding(tp)) /* The peer collapsed its window on us */ bbr_collapsed_window(bbr); else if (bbr->rc_has_collapsed) bbr_un_collapse_window(bbr); /* Was persist timer active and now we have window space? */ if ((bbr->rc_in_persist != 0) && (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr)))) { /* * Make the rate persist at end of persist mode if idle long * enough */ bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); /* Make sure we output to start the timer */ bbr->r_wanted_output = 1; } /* Do we need to enter persist? */ if ((bbr->rc_in_persist == 0) && (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_max == tp->snd_una) && sbavail(&tp->t_inpcb->inp_socket->so_snd) && (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) { /* No send window.. we must enter persist */ bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); } if (tp->t_flags2 & TF2_DROP_AF_DATA) { m_freem(m); return (0); } /* * We don't support urgent data but * drag along the up just to make sure * if there is a stack switch no one * is surprised. */ tp->rcv_up = tp->rcv_nxt; INP_WLOCK_ASSERT(tp->t_inpcb); /* * Process the segment text, merging it into the TCP sequencing * queue, and arranging for acknowledgment of receipt if necessary. * This process logically involves adjusting tp->rcv_wnd as data is * presented to the user (this happens in tcp_usrreq.c, case * PRU_RCVD). If a FIN has already been received on this connection * then we just ignore the text. */ tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && IS_FASTOPEN(tp->t_flags)); if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) && TCPS_HAVERCVDFIN(tp->t_state) == 0) { tcp_seq save_start = th->th_seq; tcp_seq save_rnxt = tp->rcv_nxt; int save_tlen = tlen; m_adj(m, drop_hdrlen); /* delayed header drop */ /* * Insert segment which includes th into TCP reassembly * queue with control block tp. Set thflags to whether * reassembly now includes a segment with FIN. This handles * the common case inline (segment is the next to be * received on an established connection, and the queue is * empty), avoiding linkage into and removal from the queue * and repetition of various conversions. Set DELACK for * segments received in order, but ack immediately when * segments are out of order (so fast retransmit can work). */ if (th->th_seq == tp->rcv_nxt && SEGQ_EMPTY(tp) && (TCPS_HAVEESTABLISHED(tp->t_state) || tfo_syn)) { #ifdef NETFLIX_SB_LIMITS u_int mcnt, appended; if (so->so_rcv.sb_shlim) { mcnt = m_memcnt(m); appended = 0; if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, CFO_NOSLEEP, NULL) == false) { counter_u64_add(tcp_sb_shlim_fails, 1); m_freem(m); return (0); } } #endif if (DELAY_ACK(tp, bbr, nsegs) || tfo_syn) { bbr->bbr_segs_rcvd += max(1, nsegs); tp->t_flags |= TF_DELACK; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); } else { bbr->r_wanted_output = 1; tp->t_flags |= TF_ACKNOW; } tp->rcv_nxt += tlen; if (tlen && ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && (tp->t_fbyte_in == 0)) { tp->t_fbyte_in = ticks; if (tp->t_fbyte_in == 0) tp->t_fbyte_in = 1; if (tp->t_fbyte_out && tp->t_fbyte_in) tp->t_flags2 |= TF2_FBYTES_COMPLETE; } thflags = th->th_flags & TH_FIN; KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); SOCKBUF_LOCK(&so->so_rcv); if (so->so_rcv.sb_state & SBS_CANTRCVMORE) m_freem(m); else #ifdef NETFLIX_SB_LIMITS appended = #endif sbappendstream_locked(&so->so_rcv, m, 0); SOCKBUF_UNLOCK(&so->so_rcv); tp->t_flags |= TF_WAKESOR; #ifdef NETFLIX_SB_LIMITS if (so->so_rcv.sb_shlim && appended != mcnt) counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended); #endif } else { /* * XXX: Due to the header drop above "th" is * theoretically invalid by now. Fortunately * m_adj() doesn't actually frees any mbufs when * trimming from the head. */ tcp_seq temp = save_start; thflags = tcp_reass(tp, th, &temp, &tlen, m); tp->t_flags |= TF_ACKNOW; } if ((tp->t_flags & TF_SACK_PERMIT) && (save_tlen > 0) && TCPS_HAVEESTABLISHED(tp->t_state)) { if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) { /* * DSACK actually handled in the fastpath * above. */ tcp_update_sack_list(tp, save_start, save_start + save_tlen); } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) { if ((tp->rcv_numsacks >= 1) && (tp->sackblks[0].end == save_start)) { /* * Partial overlap, recorded at todrop * above. */ tcp_update_sack_list(tp, tp->sackblks[0].start, tp->sackblks[0].end); } else { tcp_update_dsack_list(tp, save_start, save_start + save_tlen); } } else if (tlen >= save_tlen) { /* Update of sackblks. */ tcp_update_dsack_list(tp, save_start, save_start + save_tlen); } else if (tlen > 0) { tcp_update_dsack_list(tp, save_start, save_start + tlen); } } } else { m_freem(m); thflags &= ~TH_FIN; } /* * If FIN is received ACK the FIN and let the user know that the * connection is closing. */ if (thflags & TH_FIN) { if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { socantrcvmore(so); /* The socket upcall is handled by socantrcvmore. */ tp->t_flags &= ~TF_WAKESOR; /* * If connection is half-synchronized (ie NEEDSYN * flag on) then delay ACK, so it may be piggybacked * when SYN is sent. Otherwise, since we received a * FIN then no more input can be expected, send ACK * now. */ if (tp->t_flags & TF_NEEDSYN) { tp->t_flags |= TF_DELACK; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); } else { tp->t_flags |= TF_ACKNOW; } tp->rcv_nxt++; } switch (tp->t_state) { /* * In SYN_RECEIVED and ESTABLISHED STATES enter the * CLOSE_WAIT state. */ case TCPS_SYN_RECEIVED: tp->t_starttime = ticks; /* FALLTHROUGH */ case TCPS_ESTABLISHED: tcp_state_change(tp, TCPS_CLOSE_WAIT); break; /* * If still in FIN_WAIT_1 STATE FIN has not been * acked so enter the CLOSING state. */ case TCPS_FIN_WAIT_1: tcp_state_change(tp, TCPS_CLOSING); break; /* * In FIN_WAIT_2 state enter the TIME_WAIT state, * starting the time-wait timer, turning off the * other standard timers. */ case TCPS_FIN_WAIT_2: bbr->rc_timer_first = 1; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); INP_WLOCK_ASSERT(tp->t_inpcb); tcp_twstart(tp); return (1); } } /* * Return any desired output. */ if ((tp->t_flags & TF_ACKNOW) || (sbavail(&so->so_snd) > ctf_outstanding(tp))) { bbr->r_wanted_output = 1; } INP_WLOCK_ASSERT(tp->t_inpcb); return (0); } /* * Here nothing is really faster, its just that we * have broken out the fast-data path also just like * the fast-ack. Return 1 if we processed the packet * return 0 if you need to take the "slow-path". */ static int bbr_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t nxt_pkt) { uint16_t nsegs; int32_t newsize = 0; /* automatic sockbuf scaling */ struct tcp_bbr *bbr; #ifdef NETFLIX_SB_LIMITS u_int mcnt, appended; #endif #ifdef TCPDEBUG /* * The size of tcp_saveipgen must be the size of the max ip header, * now IPv6. */ u_char tcp_saveipgen[IP6_HDR_LEN]; struct tcphdr tcp_savetcp; short ostate = 0; #endif /* On the hpts and we would have called output */ bbr = (struct tcp_bbr *)tp->t_fb_ptr; /* * If last ACK falls within this segment's sequence numbers, record * the timestamp. NOTE that the test is modified according to the * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). */ if (bbr->r_ctl.rc_resend != NULL) { return (0); } if (tiwin && tiwin != tp->snd_wnd) { return (0); } if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) { return (0); } if (__predict_false((to->to_flags & TOF_TS) && (TSTMP_LT(to->to_tsval, tp->ts_recent)))) { return (0); } if (__predict_false((th->th_ack != tp->snd_una))) { return (0); } if (__predict_false(tlen > sbspace(&so->so_rcv))) { return (0); } if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * This is a pure, in-sequence data packet with nothing on the * reassembly queue and we have enough buffer space to take it. */ nsegs = max(1, m->m_pkthdr.lro_nsegs); #ifdef NETFLIX_SB_LIMITS if (so->so_rcv.sb_shlim) { mcnt = m_memcnt(m); appended = 0; if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, CFO_NOSLEEP, NULL) == false) { counter_u64_add(tcp_sb_shlim_fails, 1); m_freem(m); return (1); } } #endif /* Clean receiver SACK report if present */ if (tp->rcv_numsacks) tcp_clean_sackreport(tp); KMOD_TCPSTAT_INC(tcps_preddat); tp->rcv_nxt += tlen; if (tlen && ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && (tp->t_fbyte_in == 0)) { tp->t_fbyte_in = ticks; if (tp->t_fbyte_in == 0) tp->t_fbyte_in = 1; if (tp->t_fbyte_out && tp->t_fbyte_in) tp->t_flags2 |= TF2_FBYTES_COMPLETE; } /* * Pull snd_wl1 up to prevent seq wrap relative to th_seq. */ tp->snd_wl1 = th->th_seq; /* * Pull rcv_up up to prevent seq wrap relative to rcv_nxt. */ tp->rcv_up = tp->rcv_nxt; KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif newsize = tcp_autorcvbuf(m, th, so, tp, tlen); /* Add data to socket buffer. */ SOCKBUF_LOCK(&so->so_rcv); if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { m_freem(m); } else { /* * Set new socket buffer size. Give up when limit is * reached. */ if (newsize) if (!sbreserve_locked(&so->so_rcv, newsize, so, NULL)) so->so_rcv.sb_flags &= ~SB_AUTOSIZE; m_adj(m, drop_hdrlen); /* delayed header drop */ #ifdef NETFLIX_SB_LIMITS appended = #endif sbappendstream_locked(&so->so_rcv, m, 0); ctf_calc_rwin(so, tp); } SOCKBUF_UNLOCK(&so->so_rcv); tp->t_flags |= TF_WAKESOR; #ifdef NETFLIX_SB_LIMITS if (so->so_rcv.sb_shlim && mcnt != appended) counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended); #endif if (DELAY_ACK(tp, bbr, nsegs)) { bbr->bbr_segs_rcvd += max(1, nsegs); tp->t_flags |= TF_DELACK; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); } else { bbr->r_wanted_output = 1; tp->t_flags |= TF_ACKNOW; } return (1); } /* * This subfunction is used to try to highly optimize the * fast path. We again allow window updates that are * in sequence to remain in the fast-path. We also add * in the __predict's to attempt to help the compiler. * Note that if we return a 0, then we can *not* process * it and the caller should push the packet into the * slow-path. If we return 1, then all is well and * the packet is fully processed. */ static int bbr_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos) { int32_t acked; uint16_t nsegs; uint32_t sack_changed; #ifdef TCPDEBUG /* * The size of tcp_saveipgen must be the size of the max ip header, * now IPv6. */ u_char tcp_saveipgen[IP6_HDR_LEN]; struct tcphdr tcp_savetcp; short ostate = 0; #endif uint32_t prev_acked = 0; struct tcp_bbr *bbr; if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { /* Old ack, behind (or duplicate to) the last one rcv'd */ return (0); } if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) { /* Above what we have sent? */ return (0); } if (__predict_false(tiwin == 0)) { /* zero window */ return (0); } if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) { /* We need a SYN or a FIN, unlikely.. */ return (0); } if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) { /* Timestamp is behind .. old ack with seq wrap? */ return (0); } if (__predict_false(IN_RECOVERY(tp->t_flags))) { /* Still recovering */ return (0); } bbr = (struct tcp_bbr *)tp->t_fb_ptr; if (__predict_false(bbr->r_ctl.rc_resend != NULL)) { /* We are retransmitting */ return (0); } if (__predict_false(bbr->rc_in_persist != 0)) { /* In persist mode */ return (0); } if (bbr->r_ctl.rc_sacked) { /* We have sack holes on our scoreboard */ return (0); } /* Ok if we reach here, we can process a fast-ack */ nsegs = max(1, m->m_pkthdr.lro_nsegs); sack_changed = bbr_log_ack(tp, to, th, &prev_acked); /* * We never detect loss in fast ack [we can't * have a sack and can't be in recovery so * we always pass 0 (nothing detected)]. */ bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, 0); /* Did the window get updated? */ if (tiwin != tp->snd_wnd) { tp->snd_wnd = tiwin; tp->snd_wl1 = th->th_seq; if (tp->snd_wnd > tp->max_sndwnd) tp->max_sndwnd = tp->snd_wnd; } /* Do we need to exit persists? */ if ((bbr->rc_in_persist != 0) && (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr)))) { bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); bbr->r_wanted_output = 1; } /* Do we need to enter persists? */ if ((bbr->rc_in_persist == 0) && (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_max == tp->snd_una) && sbavail(&tp->t_inpcb->inp_socket->so_snd) && (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) { /* No send window.. we must enter persist */ bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); } /* * If last ACK falls within this segment's sequence numbers, record * the timestamp. NOTE that the test is modified according to the * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { tp->ts_recent_age = bbr->r_ctl.rc_rcvtime; tp->ts_recent = to->to_tsval; } /* * This is a pure ack for outstanding data. */ KMOD_TCPSTAT_INC(tcps_predack); /* * "bad retransmit" recovery. */ if (tp->t_flags & TF_PREVVALID) { tp->t_flags &= ~TF_PREVVALID; if (tp->t_rxtshift == 1 && (int)(ticks - tp->t_badrxtwin) < 0) bbr_cong_signal(tp, th, CC_RTO_ERR, NULL); } /* * Recalculate the transmit timer / rtt. * * Some boxes send broken timestamp replies during the SYN+ACK * phase, ignore timestamps of 0 or we could calculate a huge RTT * and blow up the retransmit timer. */ acked = BYTES_THIS_ACK(tp, th); #ifdef TCP_HHOOK /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ hhook_run_tcp_est_in(tp, th, to); #endif KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); sbdrop(&so->so_snd, acked); if (SEQ_GT(th->th_ack, tp->snd_una)) bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp)); tp->snd_una = th->th_ack; if (tp->snd_wnd < ctf_outstanding(tp)) /* The peer collapsed its window on us */ bbr_collapsed_window(bbr); else if (bbr->rc_has_collapsed) bbr_un_collapse_window(bbr); if (SEQ_GT(tp->snd_una, tp->snd_recover)) { tp->snd_recover = tp->snd_una; } bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, 0); /* * Pull snd_wl2 up to prevent seq wrap relative to th_ack. */ tp->snd_wl2 = th->th_ack; m_freem(m); /* * If all outstanding data are acked, stop retransmit timer, * otherwise restart timer using current (possibly backed-off) * value. If process is waiting for space, wakeup/selwakeup/signal. * If data are ready to send, let tcp_output decide between more * output or persist. */ #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif /* Wake up the socket if we have room to write more */ tp->t_flags |= TF_WAKESOW; if (tp->snd_una == tp->snd_max) { /* Nothing left outstanding */ bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__); if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0) bbr->rc_tp->t_acktime = 0; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); if (bbr->rc_in_persist == 0) { bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime; } sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime); /* * We invalidate the last ack here since we * don't want to transfer forward the time * for our sum's calculations. */ bbr->r_wanted_output = 1; } if (sbavail(&so->so_snd)) { bbr->r_wanted_output = 1; } return (1); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t todrop; int32_t ourfinisacked = 0; struct tcp_bbr *bbr; int32_t ret_val = 0; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); /* * If the state is SYN_SENT: if seg contains an ACK, but not for our * SYN, drop the input. if seg contains a RST, then drop the * connection. if seg does not contain SYN, then drop it. Otherwise * this is an acceptable SYN segment initialize tp->rcv_nxt and * tp->irs if seg contains ack then advance tp->snd_una. BRR does * not support ECN so we will not say we are capable. if SYN has * been acked change to ESTABLISHED else SYN_RCVD state arrange for * segment to be acked (eventually) continue processing rest of * data/controls, beginning with URG */ if ((thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) { TCP_PROBE5(connect__refused, NULL, tp, mtod(m, const char *), tp, th); tp = tcp_drop(tp, ECONNREFUSED); ctf_do_drop(m, tp); return (1); } if (thflags & TH_RST) { ctf_do_drop(m, tp); return (1); } if (!(thflags & TH_SYN)) { ctf_do_drop(m, tp); return (1); } tp->irs = th->th_seq; tcp_rcvseqinit(tp); if (thflags & TH_ACK) { int tfo_partial = 0; KMOD_TCPSTAT_INC(tcps_connects); soisconnected(so); #ifdef MAC mac_socketpeer_set_from_mbuf(m, so); #endif /* Do window scaling on this connection? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; } tp->rcv_adv += min(tp->rcv_wnd, TCP_MAXWIN << tp->rcv_scale); /* * If not all the data that was sent in the TFO SYN * has been acked, resend the remainder right away. */ if (IS_FASTOPEN(tp->t_flags) && (tp->snd_una != tp->snd_max)) { tp->snd_nxt = th->th_ack; tfo_partial = 1; } /* * If there's data, delay ACK; if there's also a FIN ACKNOW * will be turned on later. */ if (DELAY_ACK(tp, bbr, 1) && tlen != 0 && !tfo_partial) { bbr->bbr_segs_rcvd += 1; tp->t_flags |= TF_DELACK; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); } else { bbr->r_wanted_output = 1; tp->t_flags |= TF_ACKNOW; } if (SEQ_GT(th->th_ack, tp->iss)) { /* * The SYN is acked * handle it specially. */ bbr_log_syn(tp, to); } if (SEQ_GT(th->th_ack, tp->snd_una)) { /* * We advance snd_una for the * fast open case. If th_ack is * acknowledging data beyond * snd_una we can't just call * ack-processing since the * data stream in our send-map * will start at snd_una + 1 (one * beyond the SYN). If its just * equal we don't need to do that * and there is no send_map. */ tp->snd_una++; } /* * Received in SYN_SENT[*] state. Transitions: * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1 */ tp->t_starttime = ticks; if (tp->t_flags & TF_NEEDFIN) { tcp_state_change(tp, TCPS_FIN_WAIT_1); tp->t_flags &= ~TF_NEEDFIN; thflags &= ~TH_SYN; } else { tcp_state_change(tp, TCPS_ESTABLISHED); TCP_PROBE5(connect__established, NULL, tp, mtod(m, const char *), tp, th); cc_conn_init(tp); } } else { /* * Received initial SYN in SYN-SENT[*] state => simultaneous * open. If segment contains CC option and there is a * cached CC, apply TAO test. If it succeeds, connection is * * half-synchronized. Otherwise, do 3-way handshake: * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If * there was no CC option, clear cached CC value. */ tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN); tcp_state_change(tp, TCPS_SYN_RECEIVED); } INP_WLOCK_ASSERT(tp->t_inpcb); /* * Advance th->th_seq to correspond to first data byte. If data, * trim to stay within window, dropping FIN if necessary. */ th->th_seq++; if (tlen > tp->rcv_wnd) { todrop = tlen - tp->rcv_wnd; m_adj(m, -todrop); tlen = tp->rcv_wnd; thflags &= ~TH_FIN; KMOD_TCPSTAT_INC(tcps_rcvpackafterwin); KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); } tp->snd_wl1 = th->th_seq - 1; tp->rcv_up = th->th_seq; /* * Client side of transaction: already sent SYN and data. If the * remote host used T/TCP to validate the SYN, our data will be * ACK'd; if so, enter normal data segment processing in the middle * of step 5, ack processing. Otherwise, goto step 6. */ if (thflags & TH_ACK) { if ((to->to_flags & TOF_TS) != 0) { uint32_t t, rtt; t = tcp_tv_to_mssectick(&bbr->rc_tv); if (TSTMP_GEQ(t, to->to_tsecr)) { rtt = t - to->to_tsecr; if (rtt == 0) { rtt = 1; } rtt *= MS_IN_USEC; tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0); apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime); } } if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) return (ret_val); /* We may have changed to FIN_WAIT_1 above */ if (tp->t_state == TCPS_FIN_WAIT_1) { /* * In FIN_WAIT_1 STATE in addition to the processing * for the ESTABLISHED state if our FIN is now * acknowledged then enter FIN_WAIT_2. */ if (ourfinisacked) { /* * If we can't receive any more data, then * closing user can proceed. Starting the * timer is contrary to the specification, * but if we don't get a FIN we'll hang * forever. * * XXXjl: we should release the tp also, and * use a compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } } } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ourfinisacked = 0; int32_t ret_val; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->snd_una) || SEQ_GT(th->th_ack, tp->snd_max))) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } if (IS_FASTOPEN(tp->t_flags)) { /* * When a TFO connection is in SYN_RECEIVED, the only valid * packets are the initial SYN, a retransmit/copy of the * initial SYN (possibly with a subset of the original * data), a valid ACK, a FIN, or a RST. */ if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } else if (thflags & TH_SYN) { /* non-initial SYN is ignored */ if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RXT) || (bbr->r_ctl.rc_hpts_flags & PACE_TMR_TLP) || (bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) { ctf_do_drop(m, NULL); return (0); } } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) { ctf_do_drop(m, NULL); return (0); } } if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } /* * In the SYN-RECEIVED state, validate that the packet belongs to * this connection before trimming the data to fit the receive * window. Check the sequence number versus IRS since we know the * sequence numbers haven't wrapped. This is a partial fix for the * "LAND" DoS attack. */ if (SEQ_LT(th->th_seq, tp->irs)) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } tp->snd_wnd = tiwin; /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (IS_FASTOPEN(tp->t_flags)) { cc_conn_init(tp); } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } KMOD_TCPSTAT_INC(tcps_connects); soisconnected(so); /* Do window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; } /* * ok for the first time in lets see if we can use the ts to figure * out what the initial RTT was. */ if ((to->to_flags & TOF_TS) != 0) { uint32_t t, rtt; t = tcp_tv_to_mssectick(&bbr->rc_tv); if (TSTMP_GEQ(t, to->to_tsecr)) { rtt = t - to->to_tsecr; if (rtt == 0) { rtt = 1; } rtt *= MS_IN_USEC; tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0); apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime); } } /* Drop off any SYN in the send map (probably not there) */ if (thflags & TH_ACK) bbr_log_syn(tp, to); if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) { tcp_fastopen_decrement_counter(tp->t_tfo_pending); tp->t_tfo_pending = NULL; } /* * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* -> * FIN-WAIT-1 */ tp->t_starttime = ticks; if (tp->t_flags & TF_NEEDFIN) { tcp_state_change(tp, TCPS_FIN_WAIT_1); tp->t_flags &= ~TF_NEEDFIN; } else { tcp_state_change(tp, TCPS_ESTABLISHED); TCP_PROBE5(accept__established, NULL, tp, mtod(m, const char *), tp, th); /* * TFO connections call cc_conn_init() during SYN * processing. Calling it again here for such connections * is not harmless as it would undo the snd_cwnd reduction * that occurs when a TFO SYN|ACK is retransmitted. */ if (!IS_FASTOPEN(tp->t_flags)) cc_conn_init(tp); } /* * Account for the ACK of our SYN prior to * regular ACK processing below, except for * simultaneous SYN, which is handled later. */ if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN)) tp->snd_una++; /* * If segment contains data or ACK, will call tcp_reass() later; if * not, do so now to pass queued data to user. */ if (tlen == 0 && (thflags & TH_FIN) == 0) (void)tcp_reass(tp, (struct tcphdr *)0, NULL, 0, (struct mbuf *)0); tp->snd_wl1 = th->th_seq - 1; if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (tp->t_state == TCPS_FIN_WAIT_1) { /* We could have went to FIN_WAIT_1 (or EST) above */ /* * In FIN_WAIT_1 STATE in addition to the processing for the * ESTABLISHED state if our FIN is now acknowledged then * enter FIN_WAIT_2. */ if (ourfinisacked) { /* * If we can't receive any more data, then closing * user can proceed. Starting the timer is contrary * to the specification, but if we don't get a FIN * we'll hang forever. * * XXXjl: we should release the tp also, and use a * compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { struct tcp_bbr *bbr; int32_t ret_val; /* * Header prediction: check for the two common cases of a * uni-directional data xfer. If the packet has no control flags, * is in-sequence, the window didn't change and we're not * retransmitting, it's a candidate. If the length is zero and the * ack moved forward, we're the sender side of the xfer. Just free * the data acked & wake any higher level process that was blocked * waiting for space. If the length is non-zero and the ack didn't * move, we're the receiver side. If we're getting packets in-order * (the reassembly queue is empty), add the data toc The socket * buffer and note that we need a delayed ack. Make sure that the * hidden state-flags are also off. Since we check for * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN. */ bbr = (struct tcp_bbr *)tp->t_fb_ptr; if (bbr->r_ctl.rc_delivered < (4 * tp->t_maxseg)) { /* * If we have delived under 4 segments increase the initial * window if raised by the peer. We use this to determine * dynamic and static rwnd's at the end of a connection. */ bbr->r_ctl.rc_init_rwnd = max(tiwin, tp->snd_wnd); } if (__predict_true(((to->to_flags & TOF_SACK) == 0)) && __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) && __predict_true(SEGQ_EMPTY(tp)) && __predict_true(th->th_seq == tp->rcv_nxt)) { if (tlen == 0) { if (bbr_fastack(m, th, so, tp, to, drop_hdrlen, tlen, tiwin, nxt_pkt, iptos)) { return (0); } } else { if (bbr_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen, tiwin, nxt_pkt)) { return (0); } } } ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); bbr->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } /* State changes only happen in bbr_process_data() */ return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { struct tcp_bbr *bbr; int32_t ret_val; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); bbr->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } static int bbr_check_data_after_close(struct mbuf *m, struct tcp_bbr *bbr, struct tcpcb *tp, int32_t * tlen, struct tcphdr *th, struct socket *so) { if (bbr->rc_allow_data_af_clo == 0) { close_now: tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); /* tcp_close will kill the inp pre-log the Reset */ tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); tp = tcp_close(tp); KMOD_TCPSTAT_INC(tcps_rcvafterclose); ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen)); return (1); } if (sbavail(&so->so_snd) == 0) goto close_now; /* Ok we allow data that is ignored and a followup reset */ tp->rcv_nxt = th->th_seq + *tlen; tp->t_flags2 |= TF2_DROP_AF_DATA; bbr->r_wanted_output = 1; *tlen = 0; return (0); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ourfinisacked = 0; int32_t ret_val; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tlen) { /* * We call a new function now so we might continue and setup * to reset at all data being ack'd. */ if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) return (1); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); bbr->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { /* * If we can't receive any more data, then closing user can * proceed. Starting the timer is contrary to the * specification, but if we don't get a FIN we'll hang * forever. * * XXXjl: we should release the tp also, and use a * compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ourfinisacked = 0; int32_t ret_val; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tlen) { /* * We call a new function now so we might continue and setup * to reset at all data being ack'd. */ if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) return (1); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); bbr->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { tcp_twstart(tp); m_freem(m); return (1); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ourfinisacked = 0; int32_t ret_val; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tlen) { /* * We call a new function now so we might continue and setup * to reset at all data being ack'd. */ if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) return (1); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); bbr->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * case TCPS_LAST_ACK: Ack processing. */ if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { tp = tcp_close(tp); ctf_do_drop(m, tp); return (1); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ourfinisacked = 0; int32_t ret_val; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); /* Reset receive buffer auto scaling when not in bulk receive mode. */ if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then we may RST the other end depending on the outcome * of bbr_check_data_after_close. */ if ((so->so_state & SS_NOFDREF) && tlen) { /* * We call a new function now so we might continue and setup * to reset at all data being ack'd. */ if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) return (1); } INP_WLOCK_ASSERT(tp->t_inpcb); /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ INP_WLOCK_ASSERT(tp->t_inpcb); if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); bbr->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ INP_WLOCK_ASSERT(tp->t_inpcb); if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } INP_WLOCK_ASSERT(tp->t_inpcb); return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } static void bbr_stop_all_timers(struct tcpcb *tp) { struct tcp_bbr *bbr; /* * Assure no timers are running. */ if (tcp_timer_active(tp, TT_PERSIST)) { /* We enter in persists, set the flag appropriately */ bbr = (struct tcp_bbr *)tp->t_fb_ptr; bbr->rc_in_persist = 1; } tcp_timer_suspend(tp, TT_PERSIST); tcp_timer_suspend(tp, TT_REXMT); tcp_timer_suspend(tp, TT_KEEP); tcp_timer_suspend(tp, TT_DELACK); } static void bbr_google_mode_on(struct tcp_bbr *bbr) { bbr->rc_use_google = 1; bbr->rc_no_pacing = 0; bbr->r_ctl.bbr_google_discount = bbr_google_discount; bbr->r_use_policer = bbr_policer_detection_enabled; bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10); bbr->bbr_use_rack_cheat = 0; bbr->r_ctl.rc_incr_tmrs = 0; bbr->r_ctl.rc_inc_tcp_oh = 0; bbr->r_ctl.rc_inc_ip_oh = 0; bbr->r_ctl.rc_inc_enet_oh = 0; reset_time(&bbr->r_ctl.rc_delrate, BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT); reset_time_small(&bbr->r_ctl.rc_rttprop, (11 * USECS_IN_SECOND)); tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv)); } static void bbr_google_mode_off(struct tcp_bbr *bbr) { bbr->rc_use_google = 0; bbr->r_ctl.bbr_google_discount = 0; bbr->no_pacing_until = bbr_no_pacing_until; bbr->r_use_policer = 0; if (bbr->no_pacing_until) bbr->rc_no_pacing = 1; else bbr->rc_no_pacing = 0; if (bbr_use_rack_resend_cheat) bbr->bbr_use_rack_cheat = 1; else bbr->bbr_use_rack_cheat = 0; if (bbr_incr_timers) bbr->r_ctl.rc_incr_tmrs = 1; else bbr->r_ctl.rc_incr_tmrs = 0; if (bbr_include_tcp_oh) bbr->r_ctl.rc_inc_tcp_oh = 1; else bbr->r_ctl.rc_inc_tcp_oh = 0; if (bbr_include_ip_oh) bbr->r_ctl.rc_inc_ip_oh = 1; else bbr->r_ctl.rc_inc_ip_oh = 0; if (bbr_include_enet_oh) bbr->r_ctl.rc_inc_enet_oh = 1; else bbr->r_ctl.rc_inc_enet_oh = 0; bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; reset_time(&bbr->r_ctl.rc_delrate, bbr_num_pktepo_for_del_limit); reset_time_small(&bbr->r_ctl.rc_rttprop, (bbr_filter_len_sec * USECS_IN_SECOND)); tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv)); } /* * Return 0 on success, non-zero on failure * which indicates the error (usually no memory). */ static int bbr_init(struct tcpcb *tp) { struct tcp_bbr *bbr = NULL; struct inpcb *inp; uint32_t cts; tp->t_fb_ptr = uma_zalloc(bbr_pcb_zone, (M_NOWAIT | M_ZERO)); if (tp->t_fb_ptr == NULL) { /* * We need to allocate memory but cant. The INP and INP_INFO * locks and they are recusive (happens during setup. So a * scheme to drop the locks fails :( * */ return (ENOMEM); } bbr = (struct tcp_bbr *)tp->t_fb_ptr; bbr->rtt_valid = 0; inp = tp->t_inpcb; inp->inp_flags2 |= INP_CANNOT_DO_ECN; inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; TAILQ_INIT(&bbr->r_ctl.rc_map); TAILQ_INIT(&bbr->r_ctl.rc_free); TAILQ_INIT(&bbr->r_ctl.rc_tmap); bbr->rc_tp = tp; if (tp->t_inpcb) { bbr->rc_inp = tp->t_inpcb; } cts = tcp_get_usecs(&bbr->rc_tv); tp->t_acktime = 0; bbr->rc_allow_data_af_clo = bbr_ignore_data_after_close; bbr->r_ctl.rc_reorder_fade = bbr_reorder_fade; bbr->rc_tlp_threshold = bbr_tlp_thresh; bbr->r_ctl.rc_reorder_shift = bbr_reorder_thresh; bbr->r_ctl.rc_pkt_delay = bbr_pkt_delay; bbr->r_ctl.rc_min_to = bbr_min_to; bbr->rc_bbr_state = BBR_STATE_STARTUP; bbr->r_ctl.bbr_lost_at_state = 0; bbr->r_ctl.rc_lost_at_startup = 0; bbr->rc_all_timers_stopped = 0; bbr->r_ctl.rc_bbr_lastbtlbw = 0; bbr->r_ctl.rc_pkt_epoch_del = 0; bbr->r_ctl.rc_pkt_epoch = 0; bbr->r_ctl.rc_lowest_rtt = 0xffffffff; bbr->r_ctl.rc_bbr_hptsi_gain = bbr_high_gain; bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain; bbr->r_ctl.rc_went_idle_time = cts; bbr->rc_pacer_started = cts; bbr->r_ctl.rc_pkt_epoch_time = cts; bbr->r_ctl.rc_rcvtime = cts; bbr->r_ctl.rc_bbr_state_time = cts; bbr->r_ctl.rc_del_time = cts; bbr->r_ctl.rc_tlp_rxt_last_time = cts; bbr->r_ctl.last_in_probertt = cts; bbr->skip_gain = 0; bbr->gain_is_limited = 0; bbr->no_pacing_until = bbr_no_pacing_until; if (bbr->no_pacing_until) bbr->rc_no_pacing = 1; if (bbr_use_google_algo) { bbr->rc_no_pacing = 0; bbr->rc_use_google = 1; bbr->r_ctl.bbr_google_discount = bbr_google_discount; bbr->r_use_policer = bbr_policer_detection_enabled; } else { bbr->rc_use_google = 0; bbr->r_ctl.bbr_google_discount = 0; bbr->r_use_policer = 0; } if (bbr_ts_limiting) bbr->rc_use_ts_limit = 1; else bbr->rc_use_ts_limit = 0; if (bbr_ts_can_raise) bbr->ts_can_raise = 1; else bbr->ts_can_raise = 0; if (V_tcp_delack_enabled == 1) tp->t_delayed_ack = 2; else if (V_tcp_delack_enabled == 0) tp->t_delayed_ack = 0; else if (V_tcp_delack_enabled < 100) tp->t_delayed_ack = V_tcp_delack_enabled; else tp->t_delayed_ack = 2; if (bbr->rc_use_google == 0) bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; else bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10); bbr->r_ctl.rc_min_rto_ms = bbr_rto_min_ms; bbr->rc_max_rto_sec = bbr_rto_max_sec; bbr->rc_init_win = bbr_def_init_win; if (tp->t_flags & TF_REQ_TSTMP) bbr->rc_last_options = TCP_TS_OVERHEAD; bbr->r_ctl.rc_pace_max_segs = tp->t_maxseg - bbr->rc_last_options; bbr->r_ctl.rc_high_rwnd = tp->snd_wnd; bbr->r_init_rtt = 1; counter_u64_add(bbr_flows_nohdwr_pacing, 1); if (bbr_allow_hdwr_pacing) bbr->bbr_hdw_pace_ena = 1; else bbr->bbr_hdw_pace_ena = 0; if (bbr_sends_full_iwnd) bbr->bbr_init_win_cheat = 1; else bbr->bbr_init_win_cheat = 0; bbr->r_ctl.bbr_utter_max = bbr_hptsi_utter_max; bbr->r_ctl.rc_drain_pg = bbr_drain_gain; bbr->r_ctl.rc_startup_pg = bbr_high_gain; bbr->rc_loss_exit = bbr_exit_startup_at_loss; bbr->r_ctl.bbr_rttprobe_gain_val = bbr_rttprobe_gain; bbr->r_ctl.bbr_hptsi_per_second = bbr_hptsi_per_second; bbr->r_ctl.bbr_hptsi_segments_delay_tar = bbr_hptsi_segments_delay_tar; bbr->r_ctl.bbr_hptsi_segments_max = bbr_hptsi_segments_max; bbr->r_ctl.bbr_hptsi_segments_floor = bbr_hptsi_segments_floor; bbr->r_ctl.bbr_hptsi_bytes_min = bbr_hptsi_bytes_min; bbr->r_ctl.bbr_cross_over = bbr_cross_over; bbr->r_ctl.rc_rtt_shrinks = cts; if (bbr->rc_use_google) { setup_time_filter(&bbr->r_ctl.rc_delrate, FILTER_TYPE_MAX, BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT); setup_time_filter_small(&bbr->r_ctl.rc_rttprop, FILTER_TYPE_MIN, (11 * USECS_IN_SECOND)); } else { setup_time_filter(&bbr->r_ctl.rc_delrate, FILTER_TYPE_MAX, bbr_num_pktepo_for_del_limit); setup_time_filter_small(&bbr->r_ctl.rc_rttprop, FILTER_TYPE_MIN, (bbr_filter_len_sec * USECS_IN_SECOND)); } bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_INIT, 0); if (bbr_uses_idle_restart) bbr->rc_use_idle_restart = 1; else bbr->rc_use_idle_restart = 0; bbr->r_ctl.rc_bbr_cur_del_rate = 0; bbr->r_ctl.rc_initial_hptsi_bw = bbr_initial_bw_bps; if (bbr_resends_use_tso) bbr->rc_resends_use_tso = 1; #ifdef NETFLIX_PEAKRATE tp->t_peakrate_thr = tp->t_maxpeakrate; #endif if (tp->snd_una != tp->snd_max) { /* Create a send map for the current outstanding data */ struct bbr_sendmap *rsm; rsm = bbr_alloc(bbr); if (rsm == NULL) { uma_zfree(bbr_pcb_zone, tp->t_fb_ptr); tp->t_fb_ptr = NULL; return (ENOMEM); } rsm->r_flags = BBR_OVERMAX; rsm->r_tim_lastsent[0] = cts; rsm->r_rtr_cnt = 1; rsm->r_rtr_bytes = 0; rsm->r_start = tp->snd_una; rsm->r_end = tp->snd_max; rsm->r_dupack = 0; rsm->r_delivered = bbr->r_ctl.rc_delivered; rsm->r_ts_valid = 0; rsm->r_del_ack_ts = tp->ts_recent; rsm->r_del_time = cts; if (bbr->r_ctl.r_app_limited_until) rsm->r_app_limited = 1; else rsm->r_app_limited = 0; TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next); TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) rsm->r_bbr_state = bbr_state_val(bbr); else rsm->r_bbr_state = 8; } if (bbr_use_rack_resend_cheat && (bbr->rc_use_google == 0)) bbr->bbr_use_rack_cheat = 1; if (bbr_incr_timers && (bbr->rc_use_google == 0)) bbr->r_ctl.rc_incr_tmrs = 1; if (bbr_include_tcp_oh && (bbr->rc_use_google == 0)) bbr->r_ctl.rc_inc_tcp_oh = 1; if (bbr_include_ip_oh && (bbr->rc_use_google == 0)) bbr->r_ctl.rc_inc_ip_oh = 1; if (bbr_include_enet_oh && (bbr->rc_use_google == 0)) bbr->r_ctl.rc_inc_enet_oh = 1; bbr_log_type_statechange(bbr, cts, __LINE__); if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->t_srtt)) { uint32_t rtt; rtt = (TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); } /* announce the settings and state */ bbr_log_settings_change(bbr, BBR_RECOVERY_LOWRTT); tcp_bbr_tso_size_check(bbr, cts); /* * Now call the generic function to start a timer. This will place * the TCB on the hptsi wheel if a timer is needed with appropriate * flags. */ bbr_stop_all_timers(tp); bbr_start_hpts_timer(bbr, tp, cts, 5, 0, 0); return (0); } /* * Return 0 if we can accept the connection. Return * non-zero if we can't handle the connection. A EAGAIN * means you need to wait until the connection is up. * a EADDRNOTAVAIL means we can never handle the connection * (no SACK). */ static int bbr_handoff_ok(struct tcpcb *tp) { if ((tp->t_state == TCPS_CLOSED) || (tp->t_state == TCPS_LISTEN)) { /* Sure no problem though it may not stick */ return (0); } if ((tp->t_state == TCPS_SYN_SENT) || (tp->t_state == TCPS_SYN_RECEIVED)) { /* * We really don't know you have to get to ESTAB or beyond * to tell. */ return (EAGAIN); } if (tp->t_flags & TF_SENTFIN) return (EINVAL); if ((tp->t_flags & TF_SACK_PERMIT) || bbr_sack_not_required) { return (0); } /* * If we reach here we don't do SACK on this connection so we can * never do rack. */ return (EINVAL); } static void bbr_fini(struct tcpcb *tp, int32_t tcb_is_purged) { if (tp->t_fb_ptr) { uint32_t calc; struct tcp_bbr *bbr; struct bbr_sendmap *rsm; bbr = (struct tcp_bbr *)tp->t_fb_ptr; if (bbr->r_ctl.crte) tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp); bbr_log_flowend(bbr); bbr->rc_tp = NULL; if (tp->t_inpcb) { /* Backout any flags2 we applied */ tp->t_inpcb->inp_flags2 &= ~INP_CANNOT_DO_ECN; tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY; } if (bbr->bbr_hdrw_pacing) counter_u64_add(bbr_flows_whdwr_pacing, -1); else counter_u64_add(bbr_flows_nohdwr_pacing, -1); rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); while (rsm) { TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); uma_zfree(bbr_zone, rsm); rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); } rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); while (rsm) { TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next); uma_zfree(bbr_zone, rsm); rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); } calc = bbr->r_ctl.rc_high_rwnd - bbr->r_ctl.rc_init_rwnd; if (calc > (bbr->r_ctl.rc_init_rwnd / 10)) BBR_STAT_INC(bbr_dynamic_rwnd); else BBR_STAT_INC(bbr_static_rwnd); bbr->r_ctl.rc_free_cnt = 0; uma_zfree(bbr_pcb_zone, tp->t_fb_ptr); tp->t_fb_ptr = NULL; } /* Make sure snd_nxt is correctly set */ tp->snd_nxt = tp->snd_max; } static void bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win) { switch (tp->t_state) { case TCPS_SYN_SENT: bbr->r_state = TCPS_SYN_SENT; bbr->r_substate = bbr_do_syn_sent; break; case TCPS_SYN_RECEIVED: bbr->r_state = TCPS_SYN_RECEIVED; bbr->r_substate = bbr_do_syn_recv; break; case TCPS_ESTABLISHED: bbr->r_ctl.rc_init_rwnd = max(win, bbr->rc_tp->snd_wnd); bbr->r_state = TCPS_ESTABLISHED; bbr->r_substate = bbr_do_established; break; case TCPS_CLOSE_WAIT: bbr->r_state = TCPS_CLOSE_WAIT; bbr->r_substate = bbr_do_close_wait; break; case TCPS_FIN_WAIT_1: bbr->r_state = TCPS_FIN_WAIT_1; bbr->r_substate = bbr_do_fin_wait_1; break; case TCPS_CLOSING: bbr->r_state = TCPS_CLOSING; bbr->r_substate = bbr_do_closing; break; case TCPS_LAST_ACK: bbr->r_state = TCPS_LAST_ACK; bbr->r_substate = bbr_do_lastack; break; case TCPS_FIN_WAIT_2: bbr->r_state = TCPS_FIN_WAIT_2; bbr->r_substate = bbr_do_fin_wait_2; break; case TCPS_LISTEN: case TCPS_CLOSED: case TCPS_TIME_WAIT: default: break; }; } static void bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int32_t line, int dolog) { /* * Now what state are we going into now? Is there adjustments * needed? */ int32_t old_state, old_gain; old_state = bbr_state_val(bbr); old_gain = bbr->r_ctl.rc_bbr_hptsi_gain; if (bbr_state_val(bbr) == BBR_SUB_LEVEL1) { /* Save the lowest srtt we saw in our end of the sub-state */ bbr->rc_hit_state_1 = 0; if (bbr->r_ctl.bbr_smallest_srtt_this_state != 0xffffffff) bbr->r_ctl.bbr_smallest_srtt_state2 = bbr->r_ctl.bbr_smallest_srtt_this_state; } bbr->rc_bbr_substate++; if (bbr->rc_bbr_substate >= BBR_SUBSTATE_COUNT) { /* Cycle back to first state-> gain */ bbr->rc_bbr_substate = 0; } if (bbr_state_val(bbr) == BBR_SUB_GAIN) { /* * We enter the gain(5/4) cycle (possibly less if * shallow buffer detection is enabled) */ if (bbr->skip_gain) { /* * Hardware pacing has set our rate to * the max and limited our b/w just * do level i.e. no gain. */ bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_LEVEL1]; } else if (bbr->gain_is_limited && bbr->bbr_hdrw_pacing && bbr->r_ctl.crte) { /* * We can't gain above the hardware pacing * rate which is less than our rate + the gain * calculate the gain needed to reach the hardware * pacing rate.. */ uint64_t bw, rate, gain_calc; bw = bbr_get_bw(bbr); rate = bbr->r_ctl.crte->rate; if ((rate > bw) && (((bw * (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]) / (uint64_t)BBR_UNIT) > rate)) { gain_calc = (rate * BBR_UNIT) / bw; if (gain_calc < BBR_UNIT) gain_calc = BBR_UNIT; bbr->r_ctl.rc_bbr_hptsi_gain = (uint16_t)gain_calc; } else { bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN]; } } else bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN]; if ((bbr->rc_use_google == 0) && (bbr_gain_to_target == 0)) { bbr->r_ctl.rc_bbr_state_atflight = cts; } else bbr->r_ctl.rc_bbr_state_atflight = 0; } else if (bbr_state_val(bbr) == BBR_SUB_DRAIN) { bbr->rc_hit_state_1 = 1; bbr->r_ctl.rc_exta_time_gd = 0; bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); if (bbr_state_drain_2_tar) { bbr->r_ctl.rc_bbr_state_atflight = 0; } else bbr->r_ctl.rc_bbr_state_atflight = cts; bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_DRAIN]; } else { /* All other cycles hit here 2-7 */ if ((old_state == BBR_SUB_DRAIN) && bbr->rc_hit_state_1) { if (bbr_sub_drain_slam_cwnd && (bbr->rc_use_google == 0) && (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } if ((cts - bbr->r_ctl.rc_bbr_state_time) > bbr_get_rtt(bbr, BBR_RTT_PROP)) bbr->r_ctl.rc_exta_time_gd += ((cts - bbr->r_ctl.rc_bbr_state_time) - bbr_get_rtt(bbr, BBR_RTT_PROP)); else bbr->r_ctl.rc_exta_time_gd = 0; if (bbr->r_ctl.rc_exta_time_gd) { bbr->r_ctl.rc_level_state_extra = bbr->r_ctl.rc_exta_time_gd; /* Now chop up the time for each state (div by 7) */ bbr->r_ctl.rc_level_state_extra /= 7; if (bbr_rand_ot && bbr->r_ctl.rc_level_state_extra) { /* Add a randomization */ bbr_randomize_extra_state_time(bbr); } } } bbr->r_ctl.rc_bbr_state_atflight = max(1, cts); bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[bbr_state_val(bbr)]; } if (bbr->rc_use_google) { bbr->r_ctl.rc_bbr_state_atflight = max(1, cts); } bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain; if (dolog) bbr_log_type_statechange(bbr, cts, line); if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { uint32_t time_in; time_in = cts - bbr->r_ctl.rc_bbr_state_time; if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { counter_u64_add(bbr_state_time[(old_state + 5)], time_in); } else { counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); } } bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff; bbr_set_state_target(bbr, __LINE__); if (bbr_sub_drain_slam_cwnd && (bbr->rc_use_google == 0) && (bbr_state_val(bbr) == BBR_SUB_DRAIN)) { /* Slam down the cwnd */ bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; if (bbr_sub_drain_app_limit) { /* Go app limited if we are on a long drain */ bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes))); } bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } if (bbr->rc_lt_use_bw) { /* In policed mode we clamp pacing_gain to BBR_UNIT */ bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; } /* Google changes TSO size every cycle */ if (bbr->rc_use_google) tcp_bbr_tso_size_check(bbr, cts); bbr->r_ctl.gain_epoch = cts; bbr->r_ctl.rc_bbr_state_time = cts; bbr->r_ctl.substate_pe = bbr->r_ctl.rc_pkt_epoch; } static void bbr_set_probebw_google_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses) { if ((bbr_state_val(bbr) == BBR_SUB_DRAIN) && (google_allow_early_out == 1) && (bbr->r_ctl.rc_flight_at_input <= bbr->r_ctl.rc_target_at_state)) { /* We have reached out target flight size possibly early */ goto change_state; } if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) { return; } if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_get_rtt(bbr, BBR_RTT_PROP)) { /* * Must be a rttProp movement forward before * we can change states. */ return; } if (bbr_state_val(bbr) == BBR_SUB_GAIN) { /* * The needed time has passed but for * the gain cycle extra rules apply: * 1) If we have seen loss, we exit * 2) If we have not reached the target * we stay in GAIN (gain-to-target). */ if (google_consider_lost && losses) goto change_state; if (bbr->r_ctl.rc_target_at_state > bbr->r_ctl.rc_flight_at_input) { return; } } change_state: /* For gain we must reach our target, all others last 1 rttProp */ bbr_substate_change(bbr, cts, __LINE__, 1); } static void bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses) { uint32_t flight, bbr_cur_cycle_time; if (bbr->rc_use_google) { bbr_set_probebw_google_gains(bbr, cts, losses); return; } if (cts == 0) { /* * Never alow cts to be 0 we * do this so we can judge if * we have set a timestamp. */ cts = 1; } if (bbr_state_is_pkt_epoch) bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PKTRTT); else bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PROP); if (bbr->r_ctl.rc_bbr_state_atflight == 0) { if (bbr_state_val(bbr) == BBR_SUB_DRAIN) { flight = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); if (bbr_sub_drain_slam_cwnd && bbr->rc_hit_state_1) { /* Keep it slam down */ if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state) { bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } if (bbr_sub_drain_app_limit) { /* Go app limited if we are on a long drain */ bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + flight); } } if (TSTMP_GT(cts, bbr->r_ctl.gain_epoch) && (((cts - bbr->r_ctl.gain_epoch) > bbr_get_rtt(bbr, BBR_RTT_PROP)) || (flight >= bbr->r_ctl.flightsize_at_drain))) { /* * Still here after the same time as * the gain. We need to drain harder * for the next srtt. Reduce by a set amount * the gain drop is capped at DRAIN states * value (88). */ bbr->r_ctl.flightsize_at_drain = flight; if (bbr_drain_drop_mul && bbr_drain_drop_div && (bbr_drain_drop_mul < bbr_drain_drop_div)) { /* Use your specific drop value (def 4/5 = 20%) */ bbr->r_ctl.rc_bbr_hptsi_gain *= bbr_drain_drop_mul; bbr->r_ctl.rc_bbr_hptsi_gain /= bbr_drain_drop_div; } else { /* You get drop of 20% */ bbr->r_ctl.rc_bbr_hptsi_gain *= 4; bbr->r_ctl.rc_bbr_hptsi_gain /= 5; } if (bbr->r_ctl.rc_bbr_hptsi_gain <= bbr_drain_floor) { /* Reduce our gain again to the bottom */ bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1); } bbr_log_exit_gain(bbr, cts, 4); /* * Extend out so we wait another * epoch before dropping again. */ bbr->r_ctl.gain_epoch = cts; } if (flight <= bbr->r_ctl.rc_target_at_state) { if (bbr_sub_drain_slam_cwnd && (bbr->rc_use_google == 0) && (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); bbr_log_exit_gain(bbr, cts, 3); } } else { /* Its a gain */ if (bbr->r_ctl.rc_lost > bbr->r_ctl.bbr_lost_at_state) { bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); goto change_state; } if ((ctf_outstanding(bbr->rc_tp) >= bbr->r_ctl.rc_target_at_state) || ((ctf_outstanding(bbr->rc_tp) + bbr->rc_tp->t_maxseg - 1) >= bbr->rc_tp->snd_wnd)) { bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); bbr_log_exit_gain(bbr, cts, 2); } } /** * We fall through and return always one of two things has * occurred. * 1) We are still not at target * * 2) We reached the target and set rc_bbr_state_atflight * which means we no longer hit this block * next time we are called. */ return; } change_state: if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) return; if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_cur_cycle_time) { /* Less than a full time-period has passed */ return; } if (bbr->r_ctl.rc_level_state_extra && (bbr_state_val(bbr) > BBR_SUB_DRAIN) && ((cts - bbr->r_ctl.rc_bbr_state_time) < (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) { /* Less than a full time-period + extra has passed */ return; } if (bbr_gain_gets_extra_too && bbr->r_ctl.rc_level_state_extra && (bbr_state_val(bbr) == BBR_SUB_GAIN) && ((cts - bbr->r_ctl.rc_bbr_state_time) < (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) { /* Less than a full time-period + extra has passed */ return; } bbr_substate_change(bbr, cts, __LINE__, 1); } static uint32_t bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain) { uint32_t mss, tar; if (bbr->rc_use_google) { /* Google just uses the cwnd target */ tar = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), gain); } else { mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); /* Get the base cwnd with gain rounded to a mss */ tar = roundup(bbr_get_raw_target_cwnd(bbr, bbr_get_bw(bbr), gain), mss); /* Make sure it is within our min */ if (tar < get_min_cwnd(bbr)) return (get_min_cwnd(bbr)); } return (tar); } static void bbr_set_state_target(struct tcp_bbr *bbr, int line) { uint32_t tar, meth; if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) { /* Special case using old probe-rtt method */ tar = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); meth = 1; } else { /* Non-probe-rtt case and reduced probe-rtt */ if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT)) { /* For gain cycle we use the hptsi gain */ tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain); meth = 2; } else if ((bbr_target_is_bbunit) || bbr->rc_use_google) { /* * If configured, or for google all other states * get BBR_UNIT. */ tar = bbr_get_a_state_target(bbr, BBR_UNIT); meth = 3; } else { /* * Or we set a target based on the pacing gain * for non-google mode and default (non-configured). * Note we don't set a target goal below drain (192). */ if (bbr->r_ctl.rc_bbr_hptsi_gain < bbr_hptsi_gain[BBR_SUB_DRAIN]) { tar = bbr_get_a_state_target(bbr, bbr_hptsi_gain[BBR_SUB_DRAIN]); meth = 4; } else { tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain); meth = 5; } } } bbr_log_set_of_state_target(bbr, tar, line, meth); bbr->r_ctl.rc_target_at_state = tar; } static void bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line) { /* Change to probe_rtt */ uint32_t time_in; bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.flightsize_at_drain + bbr->r_ctl.rc_delivered); /* Setup so we force feed the filter */ if (bbr->rc_use_google || bbr_probertt_sets_rtt) bbr->rc_prtt_set_ts = 1; if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { time_in = cts - bbr->r_ctl.rc_bbr_state_time; counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); } bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_ENTERPROBE, 0); bbr->r_ctl.rc_rtt_shrinks = cts; bbr->r_ctl.last_in_probertt = cts; bbr->r_ctl.rc_probertt_srttchktim = cts; bbr->r_ctl.rc_bbr_state_time = cts; bbr->rc_bbr_state = BBR_STATE_PROBE_RTT; /* We need to force the filter to update */ if ((bbr_sub_drain_slam_cwnd) && bbr->rc_hit_state_1 && (bbr->rc_use_google == 0) && (bbr_state_val(bbr) == BBR_SUB_DRAIN)) { if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_saved_cwnd) bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; } else bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; /* Update the lost */ bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; if ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google){ /* Set to the non-configurable default of 4 (PROBE_RTT_MIN) */ bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; bbr_log_set_of_state_target(bbr, bbr->rc_tp->snd_cwnd, __LINE__, 6); bbr->r_ctl.rc_target_at_state = bbr->rc_tp->snd_cwnd; } else { /* * We bring it down slowly by using a hptsi gain that is * probably 75%. This will slowly float down our outstanding * without tampering with the cwnd. */ bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val; bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; bbr_set_state_target(bbr, __LINE__); if (bbr_prtt_slam_cwnd && (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } } if (ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= bbr->r_ctl.rc_target_at_state) { /* We are at target */ bbr->r_ctl.rc_bbr_enters_probertt = cts; } else { /* We need to come down to reach target before our time begins */ bbr->r_ctl.rc_bbr_enters_probertt = 0; } bbr->r_ctl.rc_pe_of_prtt = bbr->r_ctl.rc_pkt_epoch; BBR_STAT_INC(bbr_enter_probertt); bbr_log_exit_gain(bbr, cts, 0); bbr_log_type_statechange(bbr, cts, line); } static void bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts) { /* * Sanity check on probe-rtt intervals. * In crazy situations where we are competing * against new-reno flows with huge buffers * our rtt-prop interval could come to dominate * things if we can't get through a full set * of cycles, we need to adjust it. */ if (bbr_can_adjust_probertt && (bbr->rc_use_google == 0)) { uint16_t val = 0; uint32_t cur_rttp, fval, newval, baseval; /* Are we to small and go into probe-rtt to often? */ baseval = (bbr_get_rtt(bbr, BBR_RTT_PROP) * (BBR_SUBSTATE_COUNT + 1)); cur_rttp = roundup(baseval, USECS_IN_SECOND); fval = bbr_filter_len_sec * USECS_IN_SECOND; if (bbr_is_ratio == 0) { if (fval > bbr_rtt_probe_limit) newval = cur_rttp + (fval - bbr_rtt_probe_limit); else newval = cur_rttp; } else { int mul; mul = fval / bbr_rtt_probe_limit; newval = cur_rttp * mul; } if (cur_rttp > bbr->r_ctl.rc_probertt_int) { bbr->r_ctl.rc_probertt_int = cur_rttp; reset_time_small(&bbr->r_ctl.rc_rttprop, newval); val = 1; } else { /* * No adjustments were made * do we need to shrink it? */ if (bbr->r_ctl.rc_probertt_int > bbr_rtt_probe_limit) { if (cur_rttp <= bbr_rtt_probe_limit) { /* * Things have calmed down lets * shrink all the way to default */ bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; reset_time_small(&bbr->r_ctl.rc_rttprop, (bbr_filter_len_sec * USECS_IN_SECOND)); cur_rttp = bbr_rtt_probe_limit; newval = (bbr_filter_len_sec * USECS_IN_SECOND); val = 2; } else { /* * Well does some adjustment make sense? */ if (cur_rttp < bbr->r_ctl.rc_probertt_int) { /* We can reduce interval time some */ bbr->r_ctl.rc_probertt_int = cur_rttp; reset_time_small(&bbr->r_ctl.rc_rttprop, newval); val = 3; } } } } if (val) bbr_log_rtt_shrinks(bbr, cts, cur_rttp, newval, __LINE__, BBR_RTTS_RESETS_VALUES, val); } } static void bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { /* Exit probe-rtt */ if (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd) { tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } bbr_log_exit_gain(bbr, cts, 1); bbr->rc_hit_state_1 = 0; bbr->r_ctl.rc_rtt_shrinks = cts; bbr->r_ctl.last_in_probertt = cts; bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_RTTPROBE, 0); bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_delivered); if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { uint32_t time_in; time_in = cts - bbr->r_ctl.rc_bbr_state_time; counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); } if (bbr->rc_filled_pipe) { /* Switch to probe_bw */ bbr->rc_bbr_state = BBR_STATE_PROBE_BW; bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain; bbr_substate_change(bbr, cts, __LINE__, 0); bbr_log_type_statechange(bbr, cts, __LINE__); } else { /* Back to startup */ bbr->rc_bbr_state = BBR_STATE_STARTUP; bbr->r_ctl.rc_bbr_state_time = cts; /* * We don't want to give a complete free 3 * measurements until we exit, so we use * the number of pe's we were in probe-rtt * to add to the startup_epoch. That way * we will still retain the old state. */ bbr->r_ctl.rc_bbr_last_startup_epoch += (bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_pe_of_prtt); bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; /* Make sure to use the lower pg when shifting back in */ if (bbr->r_ctl.rc_lost && bbr_use_lower_gain_in_startup && (bbr->rc_use_google == 0)) bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower; else bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; /* Probably not needed but set it anyway */ bbr_set_state_target(bbr, __LINE__); bbr_log_type_statechange(bbr, cts, __LINE__); bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 0); } bbr_check_probe_rtt_limits(bbr, cts); } static int32_t inline bbr_should_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts) { if ((bbr->rc_past_init_win == 1) && (bbr->rc_in_persist == 0) && (bbr_calc_time(cts, bbr->r_ctl.rc_rtt_shrinks) >= bbr->r_ctl.rc_probertt_int)) { return (1); } if (bbr_can_force_probertt && (bbr->rc_in_persist == 0) && (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) && ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) { return (1); } return (0); } static int32_t bbr_google_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t pkt_epoch) { uint64_t btlbw, gain; if (pkt_epoch == 0) { /* * Need to be on a pkt-epoch to continue. */ return (0); } btlbw = bbr_get_full_bw(bbr); gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; if (btlbw >= gain) { bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch; bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3); bbr->r_ctl.rc_bbr_lastbtlbw = btlbw; } if ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) return (1); bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8); return(0); } static int32_t inline bbr_state_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch) { /* Have we gained 25% in the last 3 packet based epoch's? */ uint64_t btlbw, gain; int do_exit; int delta, rtt_gain; if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) && (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { /* * This qualifies as a RTT_PROBE session since we drop the * data outstanding to nothing and waited more than * bbr_rtt_probe_time. */ bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); bbr_set_reduced_rtt(bbr, cts, __LINE__); } if (bbr_should_enter_probe_rtt(bbr, cts)) { bbr_enter_probe_rtt(bbr, cts, __LINE__); return (0); } if (bbr->rc_use_google) return (bbr_google_startup(bbr, cts, pkt_epoch)); if ((bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) && (bbr_use_lower_gain_in_startup)) { /* Drop to a lower gain 1.5 x since we saw loss */ bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower; } if (pkt_epoch == 0) { /* * Need to be on a pkt-epoch to continue. */ return (0); } if (bbr_rtt_gain_thresh) { /* * Do we allow a flow to stay * in startup with no loss and no * gain in rtt over a set threshold? */ if (bbr->r_ctl.rc_pkt_epoch_rtt && bbr->r_ctl.startup_last_srtt && (bbr->r_ctl.rc_pkt_epoch_rtt > bbr->r_ctl.startup_last_srtt)) { delta = bbr->r_ctl.rc_pkt_epoch_rtt - bbr->r_ctl.startup_last_srtt; rtt_gain = (delta * 100) / bbr->r_ctl.startup_last_srtt; } else rtt_gain = 0; if ((bbr->r_ctl.startup_last_srtt == 0) || (bbr->r_ctl.rc_pkt_epoch_rtt < bbr->r_ctl.startup_last_srtt)) /* First time or new lower value */ bbr->r_ctl.startup_last_srtt = bbr->r_ctl.rc_pkt_epoch_rtt; if ((bbr->r_ctl.rc_lost == 0) && (rtt_gain < bbr_rtt_gain_thresh)) { /* * No loss, and we are under * our gain threhold for * increasing RTT. */ if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch) bbr->r_ctl.rc_bbr_last_startup_epoch++; bbr_log_startup_event(bbr, cts, rtt_gain, delta, bbr->r_ctl.startup_last_srtt, 10); return (0); } } if ((bbr->r_ctl.r_measurement_count == bbr->r_ctl.last_startup_measure) && (bbr->r_ctl.rc_lost_at_startup == bbr->r_ctl.rc_lost) && (!IN_RECOVERY(bbr->rc_tp->t_flags))) { /* * We only assess if we have a new measurment when * we have no loss and are not in recovery. * Drag up by one our last_startup epoch so we will hold * the number of non-gain we have already accumulated. */ if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch) bbr->r_ctl.rc_bbr_last_startup_epoch++; bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 9); return (0); } /* Case where we reduced the lost (bad retransmit) */ if (bbr->r_ctl.rc_lost_at_startup > bbr->r_ctl.rc_lost) bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; bbr->r_ctl.last_startup_measure = bbr->r_ctl.r_measurement_count; btlbw = bbr_get_full_bw(bbr); if (bbr->r_ctl.rc_bbr_hptsi_gain == bbr_startup_lower) gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * (uint64_t)bbr_low_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; else gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; do_exit = 0; if (btlbw > bbr->r_ctl.rc_bbr_lastbtlbw) bbr->r_ctl.rc_bbr_lastbtlbw = btlbw; if (btlbw >= gain) { bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch; /* Update the lost so we won't exit in next set of tests */ bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3); } if ((bbr->rc_loss_exit && (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) && (bbr->r_ctl.rc_pkt_epoch_loss_rate > bbr_startup_loss_thresh)) && ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)) { /* * If we had no gain, we had loss and that loss was above * our threshould, the rwnd is not constrained, and we have * had at least 3 packet epochs exit. Note that this is * switched off by sysctl. Google does not do this by the * way. */ if ((ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + (2 * max(bbr->r_ctl.rc_pace_max_segs, bbr->rc_tp->t_maxseg))) <= bbr->rc_tp->snd_wnd) { do_exit = 1; bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 4); } else { /* Just record an updated loss value */ bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 5); } } else bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; if (((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) || do_exit) { /* Return 1 to exit the startup state. */ return (1); } /* Stay in startup */ bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8); return (0); } static void bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch, uint32_t losses) { /* * A tick occurred in the rtt epoch do we need to do anything? */ #ifdef BBR_INVARIANTS if ((bbr->rc_bbr_state != BBR_STATE_STARTUP) && (bbr->rc_bbr_state != BBR_STATE_DRAIN) && (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) && (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) && (bbr->rc_bbr_state != BBR_STATE_PROBE_BW)) { /* Debug code? */ panic("Unknown BBR state %d?\n", bbr->rc_bbr_state); } #endif if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { /* Do we exit the startup state? */ if (bbr_state_startup(bbr, cts, epoch, pkt_epoch)) { uint32_t time_in; bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 6); bbr->rc_filled_pipe = 1; bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { time_in = cts - bbr->r_ctl.rc_bbr_state_time; counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); } else time_in = 0; if (bbr->rc_no_pacing) bbr->rc_no_pacing = 0; bbr->r_ctl.rc_bbr_state_time = cts; bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_drain_pg; bbr->rc_bbr_state = BBR_STATE_DRAIN; bbr_set_state_target(bbr, __LINE__); if ((bbr->rc_use_google == 0) && bbr_slam_cwnd_in_main_drain) { /* Here we don't have to worry about probe-rtt */ bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain; bbr_log_type_statechange(bbr, cts, __LINE__); if (ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= bbr->r_ctl.rc_target_at_state) { /* * Switch to probe_bw if we are already * there */ bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); bbr_substate_change(bbr, cts, __LINE__, 0); bbr->rc_bbr_state = BBR_STATE_PROBE_BW; bbr_log_type_statechange(bbr, cts, __LINE__); } } } else if (bbr->rc_bbr_state == BBR_STATE_IDLE_EXIT) { uint32_t inflight; struct tcpcb *tp; tp = bbr->rc_tp; inflight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); if (inflight >= bbr->r_ctl.rc_target_at_state) { /* We have reached a flight of the cwnd target */ bbr->rc_bbr_state = BBR_STATE_PROBE_BW; bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; bbr_set_state_target(bbr, __LINE__); /* * Rig it so we don't do anything crazy and * start fresh with a new randomization. */ bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff; bbr->rc_bbr_substate = BBR_SUB_LEVEL6; bbr_substate_change(bbr, cts, __LINE__, 1); } } else if (bbr->rc_bbr_state == BBR_STATE_DRAIN) { /* Has in-flight reached the bdp (or less)? */ uint32_t inflight; struct tcpcb *tp; tp = bbr->rc_tp; inflight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); if ((bbr->rc_use_google == 0) && bbr_slam_cwnd_in_main_drain && (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { /* * Here we don't have to worry about probe-rtt * re-slam it, but keep it slammed down. */ bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } if (inflight <= bbr->r_ctl.rc_target_at_state) { /* We have drained */ bbr->rc_bbr_state = BBR_STATE_PROBE_BW; bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { uint32_t time_in; time_in = cts - bbr->r_ctl.rc_bbr_state_time; counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); } if ((bbr->rc_use_google == 0) && bbr_slam_cwnd_in_main_drain && (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { /* Restore the cwnd */ tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } /* Setup probe-rtt has being done now RRS-HERE */ bbr->r_ctl.rc_rtt_shrinks = cts; bbr->r_ctl.last_in_probertt = cts; bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_LEAVE_DRAIN, 0); /* Randomly pick a sub-state */ bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); bbr_substate_change(bbr, cts, __LINE__, 0); bbr_log_type_statechange(bbr, cts, __LINE__); } } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) { uint32_t flight; flight = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); bbr->r_ctl.r_app_limited_until = (flight + bbr->r_ctl.rc_delivered); if (((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google) && (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { /* * We must keep cwnd at the desired MSS. */ bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } else if ((bbr_prtt_slam_cwnd) && (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { /* Re-slam it */ bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } if (bbr->r_ctl.rc_bbr_enters_probertt == 0) { /* Has outstanding reached our target? */ if (flight <= bbr->r_ctl.rc_target_at_state) { bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_REACHTAR, 0); bbr->r_ctl.rc_bbr_enters_probertt = cts; /* If time is exactly 0, be 1usec off */ if (bbr->r_ctl.rc_bbr_enters_probertt == 0) bbr->r_ctl.rc_bbr_enters_probertt = 1; if (bbr->rc_use_google == 0) { /* * Restore any lowering that as occurred to * reach here */ if (bbr->r_ctl.bbr_rttprobe_gain_val) bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val; else bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; } } if ((bbr->r_ctl.rc_bbr_enters_probertt == 0) && (bbr->rc_use_google == 0) && bbr->r_ctl.bbr_rttprobe_gain_val && (((cts - bbr->r_ctl.rc_probertt_srttchktim) > bbr_get_rtt(bbr, bbr_drain_rtt)) || (flight >= bbr->r_ctl.flightsize_at_drain))) { /* * We have doddled with our current hptsi * gain an srtt and have still not made it * to target, or we have increased our flight. * Lets reduce the gain by xx% * flooring the reduce at DRAIN (based on * mul/div) */ int red; bbr->r_ctl.flightsize_at_drain = flight; bbr->r_ctl.rc_probertt_srttchktim = cts; red = max((bbr->r_ctl.bbr_rttprobe_gain_val / 10), 1); if ((bbr->r_ctl.rc_bbr_hptsi_gain - red) > max(bbr_drain_floor, 1)) { /* Reduce our gain again */ bbr->r_ctl.rc_bbr_hptsi_gain -= red; bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG, 0); } else if (bbr->r_ctl.rc_bbr_hptsi_gain > max(bbr_drain_floor, 1)) { /* one more chance before we give up */ bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1); bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG_FINAL, 0); } else { /* At the very bottom */ bbr->r_ctl.rc_bbr_hptsi_gain = max((bbr_drain_floor-1), 1); } } } if (bbr->r_ctl.rc_bbr_enters_probertt && (TSTMP_GT(cts, bbr->r_ctl.rc_bbr_enters_probertt)) && ((cts - bbr->r_ctl.rc_bbr_enters_probertt) >= bbr_rtt_probe_time)) { /* Time to exit probe RTT normally */ bbr_exit_probe_rtt(bbr->rc_tp, bbr, cts); } } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) && (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { /* * This qualifies as a RTT_PROBE session since we * drop the data outstanding to nothing and waited * more than bbr_rtt_probe_time. */ bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); bbr_set_reduced_rtt(bbr, cts, __LINE__); } if (bbr_should_enter_probe_rtt(bbr, cts)) { bbr_enter_probe_rtt(bbr, cts, __LINE__); } else { bbr_set_probebw_gains(bbr, cts, losses); } } } static void bbr_check_bbr_for_state(struct tcp_bbr *bbr, uint32_t cts, int32_t line, uint32_t losses) { int32_t epoch = 0; if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) { bbr_set_epoch(bbr, cts, line); /* At each epoch doe lt bw sampling */ epoch = 1; } bbr_state_change(bbr, cts, epoch, bbr->rc_is_pkt_epoch_now, losses); } static int bbr_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt, struct timeval *tv) { int32_t thflags, retval; uint32_t cts, lcts; uint32_t tiwin; struct tcpopt to; struct tcp_bbr *bbr; struct bbr_sendmap *rsm; struct timeval ltv; int32_t did_out = 0; int32_t in_recovery; uint16_t nsegs; int32_t prev_state; uint32_t lost; nsegs = max(1, m->m_pkthdr.lro_nsegs); bbr = (struct tcp_bbr *)tp->t_fb_ptr; /* add in our stats */ kern_prefetch(bbr, &prev_state); prev_state = 0; thflags = th->th_flags; /* * If this is either a state-changing packet or current state isn't * established, we require a write lock on tcbinfo. Otherwise, we * allow the tcbinfo to be in either alocked or unlocked, as the * caller may have unnecessarily acquired a write lock due to a * race. */ INP_WLOCK_ASSERT(tp->t_inpcb); KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", __func__)); KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", __func__)); tp->t_rcvtime = ticks; /* * Unscale the window into a 32-bit value. For the SYN_SENT state * the scale is zero. */ tiwin = th->th_win << tp->snd_scale; #ifdef STATS stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin); #endif if (m->m_flags & M_TSTMP) { /* Prefer the hardware timestamp if present */ struct timespec ts; mbuf_tstmp2timespec(m, &ts); bbr->rc_tv.tv_sec = ts.tv_sec; bbr->rc_tv.tv_usec = ts.tv_nsec / 1000; bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv); } else if (m->m_flags & M_TSTMP_LRO) { /* Next the arrival timestamp */ struct timespec ts; mbuf_tstmp2timespec(m, &ts); bbr->rc_tv.tv_sec = ts.tv_sec; bbr->rc_tv.tv_usec = ts.tv_nsec / 1000; bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv); } else { /* * Ok just get the current time. */ bbr->r_ctl.rc_rcvtime = lcts = cts = tcp_get_usecs(&bbr->rc_tv); } /* * Parse options on any incoming segment. */ tcp_dooptions(&to, (u_char *)(th + 1), (th->th_off << 2) - sizeof(struct tcphdr), (thflags & TH_SYN) ? TO_SYN : 0); /* * If timestamps were negotiated during SYN/ACK and a * segment without a timestamp is received, silently drop * the segment, unless it is a RST segment or missing timestamps are * tolerated. * See section 3.2 of RFC 7323. */ if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) && ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) { retval = 0; goto done_with_input; } /* * If echoed timestamp is later than the current time, fall back to * non RFC1323 RTT calculation. Normalize timestamp if syncookies * were used when this connection was established. */ if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { to.to_tsecr -= tp->ts_offset; if (TSTMP_GT(to.to_tsecr, tcp_tv_to_mssectick(&bbr->rc_tv))) to.to_tsecr = 0; } /* * If its the first time in we need to take care of options and * verify we can do SACK for rack! */ if (bbr->r_state == 0) { /* * Process options only when we get SYN/ACK back. The SYN * case for incoming connections is handled in tcp_syncache. * According to RFC1323 the window field in a SYN (i.e., a * or ) segment itself is never scaled. XXX * this is traditional behavior, may need to be cleaned up. */ if (bbr->rc_inp == NULL) { bbr->rc_inp = tp->t_inpcb; } /* * We need to init rc_inp here since its not init'd when * bbr_init is called */ if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { if ((to.to_flags & TOF_SCALE) && (tp->t_flags & TF_REQ_SCALE)) { tp->t_flags |= TF_RCVD_SCALE; tp->snd_scale = to.to_wscale; } else tp->t_flags &= ~TF_REQ_SCALE; /* * Initial send window. It will be updated with the * next incoming segment to the scaled value. */ tp->snd_wnd = th->th_win; if ((to.to_flags & TOF_TS) && (tp->t_flags & TF_REQ_TSTMP)) { tp->t_flags |= TF_RCVD_TSTMP; tp->ts_recent = to.to_tsval; tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); } else tp->t_flags &= ~TF_REQ_TSTMP; if (to.to_flags & TOF_MSS) tcp_mss(tp, to.to_mss); if ((tp->t_flags & TF_SACK_PERMIT) && (to.to_flags & TOF_SACKPERM) == 0) tp->t_flags &= ~TF_SACK_PERMIT; if (IS_FASTOPEN(tp->t_flags)) { if (to.to_flags & TOF_FASTOPEN) { uint16_t mss; if (to.to_flags & TOF_MSS) mss = to.to_mss; else if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) mss = TCP6_MSS; else mss = TCP_MSS; tcp_fastopen_update_cache(tp, mss, to.to_tfo_len, to.to_tfo_cookie); } else tcp_fastopen_disable_path(tp); } } /* * At this point we are at the initial call. Here we decide * if we are doing RACK or not. We do this by seeing if * TF_SACK_PERMIT is set, if not rack is *not* possible and * we switch to the default code. */ if ((tp->t_flags & TF_SACK_PERMIT) == 0) { /* Bail */ tcp_switch_back_to_default(tp); (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen, tlen, iptos); return (1); } /* Set the flag */ bbr->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; tcp_set_hpts(tp->t_inpcb); sack_filter_clear(&bbr->r_ctl.bbr_sf, th->th_ack); } if (thflags & TH_ACK) { /* Track ack types */ if (to.to_flags & TOF_SACK) BBR_STAT_INC(bbr_acks_with_sacks); else BBR_STAT_INC(bbr_plain_acks); } /* * This is the one exception case where we set the rack state * always. All other times (timers etc) we must have a rack-state * set (so we assure we have done the checks above for SACK). */ if (thflags & TH_FIN) tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN); if (bbr->r_state != tp->t_state) bbr_set_state(tp, bbr, tiwin); if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map)) != NULL) kern_prefetch(rsm, &prev_state); prev_state = bbr->r_state; bbr->rc_ack_was_delayed = 0; lost = bbr->r_ctl.rc_lost; bbr->rc_is_pkt_epoch_now = 0; if (m->m_flags & (M_TSTMP|M_TSTMP_LRO)) { /* Get the real time into lcts and figure the real delay */ lcts = tcp_get_usecs(<v); if (TSTMP_GT(lcts, cts)) { bbr->r_ctl.rc_ack_hdwr_delay = lcts - cts; bbr->rc_ack_was_delayed = 1; if (TSTMP_GT(bbr->r_ctl.rc_ack_hdwr_delay, bbr->r_ctl.highest_hdwr_delay)) bbr->r_ctl.highest_hdwr_delay = bbr->r_ctl.rc_ack_hdwr_delay; } else { bbr->r_ctl.rc_ack_hdwr_delay = 0; bbr->rc_ack_was_delayed = 0; } } else { bbr->r_ctl.rc_ack_hdwr_delay = 0; bbr->rc_ack_was_delayed = 0; } bbr_log_ack_event(bbr, th, &to, tlen, nsegs, cts, nxt_pkt, m); if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) { retval = 0; m_freem(m); goto done_with_input; } /* * If a segment with the ACK-bit set arrives in the SYN-SENT state * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9. */ if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } in_recovery = IN_RECOVERY(tp->t_flags); if (tiwin > bbr->r_ctl.rc_high_rwnd) bbr->r_ctl.rc_high_rwnd = tiwin; #ifdef BBR_INVARIANTS if ((tp->t_inpcb->inp_flags & INP_DROPPED) || (tp->t_inpcb->inp_flags2 & INP_FREED)) { panic("tp:%p bbr:%p given a dropped inp:%p", tp, bbr, tp->t_inpcb); } #endif bbr->r_ctl.rc_flight_at_input = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); bbr->rtt_valid = 0; if (to.to_flags & TOF_TS) { bbr->rc_ts_valid = 1; bbr->r_ctl.last_inbound_ts = to.to_tsval; } else { bbr->rc_ts_valid = 0; bbr->r_ctl.last_inbound_ts = 0; } retval = (*bbr->r_substate) (m, th, so, tp, &to, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt, iptos); #ifdef BBR_INVARIANTS if ((retval == 0) && (tp->t_inpcb == NULL)) { panic("retval:%d tp:%p t_inpcb:NULL state:%d", retval, tp, prev_state); } #endif if (nxt_pkt == 0) BBR_STAT_INC(bbr_rlock_left_ret0); else BBR_STAT_INC(bbr_rlock_left_ret1); if (retval == 0) { /* * If retval is 1 the tcb is unlocked and most likely the tp * is gone. */ INP_WLOCK_ASSERT(tp->t_inpcb); tcp_bbr_xmit_timer_commit(bbr, tp, cts); if (bbr->rc_is_pkt_epoch_now) bbr_set_pktepoch(bbr, cts, __LINE__); bbr_check_bbr_for_state(bbr, cts, __LINE__, (bbr->r_ctl.rc_lost - lost)); if (nxt_pkt == 0) { if (bbr->r_wanted_output != 0) { bbr->rc_output_starts_timer = 0; did_out = 1; (void)tp->t_fb->tfb_tcp_output(tp); } else bbr_start_hpts_timer(bbr, tp, cts, 6, 0, 0); } if ((nxt_pkt == 0) && ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) && (SEQ_GT(tp->snd_max, tp->snd_una) || (tp->t_flags & TF_DELACK) || ((V_tcp_always_keepalive || bbr->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)))) { /* * We could not send (probably in the hpts but * stopped the timer)? */ if ((tp->snd_max == tp->snd_una) && ((tp->t_flags & TF_DELACK) == 0) && (bbr->rc_inp->inp_in_hpts) && (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { /* * keep alive not needed if we are hptsi * output yet */ ; } else { if (bbr->rc_inp->inp_in_hpts) { tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT); if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && (TSTMP_GT(lcts, bbr->rc_pacer_started))) { uint32_t del; del = lcts - bbr->rc_pacer_started; if (bbr->r_ctl.rc_last_delay_val > del) { BBR_STAT_INC(bbr_force_timer_start); bbr->r_ctl.rc_last_delay_val -= del; bbr->rc_pacer_started = lcts; } else { /* We are late */ bbr->r_ctl.rc_last_delay_val = 0; BBR_STAT_INC(bbr_force_output); (void)tp->t_fb->tfb_tcp_output(tp); } } } bbr_start_hpts_timer(bbr, tp, cts, 8, bbr->r_ctl.rc_last_delay_val, 0); } } else if ((bbr->rc_output_starts_timer == 0) && (nxt_pkt == 0)) { /* Do we have the correct timer running? */ bbr_timer_audit(tp, bbr, lcts, &so->so_snd); } /* Do we have a new state */ if (bbr->r_state != tp->t_state) bbr_set_state(tp, bbr, tiwin); done_with_input: bbr_log_doseg_done(bbr, cts, nxt_pkt, did_out); if (did_out) bbr->r_wanted_output = 0; #ifdef BBR_INVARIANTS if (tp->t_inpcb == NULL) { panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d", did_out, retval, tp, prev_state); } #endif } return (retval); } static void bbr_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos) { struct timeval tv; int retval; /* First lets see if we have old packets */ if (tp->t_in_pkt) { if (ctf_do_queued_segments(so, tp, 1)) { m_freem(m); return; } } if (m->m_flags & M_TSTMP_LRO) { tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000; tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000; } else { /* Should not be should we kassert instead? */ tcp_get_usecs(&tv); } retval = bbr_do_segment_nounlock(m, th, so, tp, drop_hdrlen, tlen, iptos, 0, &tv); if (retval == 0) { tcp_handle_wakeup(tp, so); INP_WUNLOCK(tp->t_inpcb); } } /* * Return how much data can be sent without violating the * cwnd or rwnd. */ static inline uint32_t bbr_what_can_we_send(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t sendwin, uint32_t avail, int32_t sb_offset, uint32_t cts) { uint32_t len; if (ctf_outstanding(tp) >= tp->snd_wnd) { /* We never want to go over our peers rcv-window */ len = 0; } else { uint32_t flight; flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); if (flight >= sendwin) { /* * We have in flight what we are allowed by cwnd (if * it was rwnd blocking it would have hit above out * >= tp->snd_wnd). */ return (0); } len = sendwin - flight; if ((len + ctf_outstanding(tp)) > tp->snd_wnd) { /* We would send too much (beyond the rwnd) */ len = tp->snd_wnd - ctf_outstanding(tp); } if ((len + sb_offset) > avail) { /* * We don't have that much in the SB, how much is * there? */ len = avail - sb_offset; } } return (len); } static inline void bbr_do_error_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error) { #ifdef NETFLIX_STATS KMOD_TCPSTAT_INC(tcps_sndpack_error); KMOD_TCPSTAT_ADD(tcps_sndbyte_error, len); #endif } static inline void bbr_do_send_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error) { if (error) { bbr_do_error_accounting(tp, bbr, rsm, len, error); return; } if (rsm) { if (rsm->r_flags & BBR_TLP) { /* * TLP should not count in retran count, but in its * own bin */ #ifdef NETFLIX_STATS tp->t_sndtlppack++; tp->t_sndtlpbyte += len; KMOD_TCPSTAT_INC(tcps_tlpresends); KMOD_TCPSTAT_ADD(tcps_tlpresend_bytes, len); #endif } else { /* Retransmit */ tp->t_sndrexmitpack++; KMOD_TCPSTAT_INC(tcps_sndrexmitpack); KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len); #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, len); #endif } /* * Logs in 0 - 8, 8 is all non probe_bw states 0-7 is * sub-state */ counter_u64_add(bbr_state_lost[rsm->r_bbr_state], len); if (bbr->rc_bbr_state != BBR_STATE_PROBE_BW) { /* Non probe_bw log in 1, 2, or 4. */ counter_u64_add(bbr_state_resend[bbr->rc_bbr_state], len); } else { /* * Log our probe state 3, and log also 5-13 to show * us the recovery sub-state for the send. This * means that 3 == (5+6+7+8+9+10+11+12+13) */ counter_u64_add(bbr_state_resend[BBR_STATE_PROBE_BW], len); counter_u64_add(bbr_state_resend[(bbr_state_val(bbr) + 5)], len); } /* Place in both 16's the totals of retransmitted */ counter_u64_add(bbr_state_lost[16], len); counter_u64_add(bbr_state_resend[16], len); /* Place in 17's the total sent */ counter_u64_add(bbr_state_resend[17], len); counter_u64_add(bbr_state_lost[17], len); } else { /* New sends */ KMOD_TCPSTAT_INC(tcps_sndpack); KMOD_TCPSTAT_ADD(tcps_sndbyte, len); /* Place in 17's the total sent */ counter_u64_add(bbr_state_resend[17], len); counter_u64_add(bbr_state_lost[17], len); #ifdef STATS stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, len); #endif } } static void bbr_cwnd_limiting(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t in_level) { if (bbr->rc_filled_pipe && bbr_target_cwnd_mult_limit && (bbr->rc_use_google == 0)) { /* * Limit the cwnd to not be above N x the target plus whats * is outstanding. The target is based on the current b/w * estimate. */ uint32_t target; target = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), BBR_UNIT); target += ctf_outstanding(tp); target *= bbr_target_cwnd_mult_limit; if (tp->snd_cwnd > target) tp->snd_cwnd = target; bbr_log_type_cwndupd(bbr, 0, 0, 0, 10, 0, 0, __LINE__); } } static int bbr_window_update_needed(struct tcpcb *tp, struct socket *so, uint32_t recwin, int32_t maxseg) { /* * "adv" is the amount we could increase the window, taking into * account that we are limited by TCP_MAXWIN << tp->rcv_scale. */ int32_t adv; int32_t oldwin; adv = recwin; if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { oldwin = (tp->rcv_adv - tp->rcv_nxt); if (adv > oldwin) adv -= oldwin; else { /* We can't increase the window */ adv = 0; } } else oldwin = 0; /* * If the new window size ends up being the same as or less * than the old size when it is scaled, then don't force * a window update. */ if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) return (0); if (adv >= (2 * maxseg) && (adv >= (so->so_rcv.sb_hiwat / 4) || recwin <= (so->so_rcv.sb_hiwat / 8) || so->so_rcv.sb_hiwat <= 8 * maxseg)) { return (1); } if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) return (1); return (0); } /* * Return 0 on success and a errno on failure to send. * Note that a 0 return may not mean we sent anything * if the TCB was on the hpts. A non-zero return * does indicate the error we got from ip[6]_output. */ static int bbr_output_wtime(struct tcpcb *tp, const struct timeval *tv) { struct socket *so; int32_t len; uint32_t cts; uint32_t recwin, sendwin; int32_t sb_offset; int32_t flags, abandon, error = 0; struct tcp_log_buffer *lgb = NULL; struct mbuf *m; struct mbuf *mb; uint32_t if_hw_tsomaxsegcount = 0; uint32_t if_hw_tsomaxsegsize = 0; uint32_t if_hw_tsomax = 0; struct ip *ip = NULL; #ifdef TCPDEBUG struct ipovly *ipov = NULL; #endif struct tcp_bbr *bbr; struct tcphdr *th; -#ifdef NETFLIX_TCPOUDP struct udphdr *udp = NULL; -#endif u_char opt[TCP_MAXOLEN]; unsigned ipoptlen, optlen, hdrlen; -#ifdef NETFLIX_TCPOUDP unsigned ulen; -#endif uint32_t bbr_seq; uint32_t delay_calc=0; uint8_t doing_tlp = 0; uint8_t local_options; #ifdef BBR_INVARIANTS uint8_t doing_retran_from = 0; uint8_t picked_up_retran = 0; #endif uint8_t wanted_cookie = 0; uint8_t more_to_rxt=0; int32_t prefetch_so_done = 0; int32_t prefetch_rsm = 0; uint32_t what_we_can = 0; uint32_t tot_len = 0; uint32_t rtr_cnt = 0; uint32_t maxseg, pace_max_segs, p_maxseg; int32_t csum_flags; int32_t hw_tls; #if defined(IPSEC) || defined(IPSEC_SUPPORT) unsigned ipsec_optlen = 0; #endif volatile int32_t sack_rxmit; struct bbr_sendmap *rsm = NULL; int32_t tso, mtu; struct tcpopt to; int32_t slot = 0; struct inpcb *inp; struct sockbuf *sb; uint32_t hpts_calling; #ifdef INET6 struct ip6_hdr *ip6 = NULL; int32_t isipv6; #endif uint8_t app_limited = BBR_JR_SENT_DATA; uint8_t filled_all = 0; bbr = (struct tcp_bbr *)tp->t_fb_ptr; /* We take a cache hit here */ memcpy(&bbr->rc_tv, tv, sizeof(struct timeval)); cts = tcp_tv_to_usectick(&bbr->rc_tv); inp = bbr->rc_inp; so = inp->inp_socket; sb = &so->so_snd; if (sb->sb_flags & SB_TLS_IFNET) hw_tls = 1; else hw_tls = 0; kern_prefetch(sb, &maxseg); maxseg = tp->t_maxseg - bbr->rc_last_options; if (bbr_minseg(bbr) < maxseg) { tcp_bbr_tso_size_check(bbr, cts); } /* Remove any flags that indicate we are pacing on the inp */ pace_max_segs = bbr->r_ctl.rc_pace_max_segs; p_maxseg = min(maxseg, pace_max_segs); INP_WLOCK_ASSERT(inp); #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) return (tcp_offload_output(tp)); #endif #ifdef INET6 if (bbr->r_state) { /* Use the cache line loaded if possible */ isipv6 = bbr->r_is_v6; } else { isipv6 = (inp->inp_vflag & INP_IPV6) != 0; } #endif if (((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && inp->inp_in_hpts) { /* * We are on the hpts for some timer but not hptsi output. * Possibly remove from the hpts so we can send/recv etc. */ if ((tp->t_flags & TF_ACKNOW) == 0) { /* * No immediate demand right now to send an ack, but * the user may have read, making room for new data * (a window update). If so we may want to cancel * whatever timer is running (KEEP/DEL-ACK?) and * continue to send out a window update. Or we may * have gotten more data into the socket buffer to * send. */ recwin = lmin(lmax(sbspace(&so->so_rcv), 0), (long)TCP_MAXWIN << tp->rcv_scale); if ((bbr_window_update_needed(tp, so, recwin, maxseg) == 0) && ((tcp_outflags[tp->t_state] & TH_RST) == 0) && ((sbavail(sb) + ((tcp_outflags[tp->t_state] & TH_FIN) ? 1 : 0)) <= (tp->snd_max - tp->snd_una))) { /* * Nothing new to send and no window update * is needed to send. Lets just return and * let the timer-run off. */ return (0); } } tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT); bbr_timer_cancel(bbr, __LINE__, cts); } if (bbr->r_ctl.rc_last_delay_val) { /* Calculate a rough delay for early escape to sending */ if (SEQ_GT(cts, bbr->rc_pacer_started)) delay_calc = cts - bbr->rc_pacer_started; if (delay_calc >= bbr->r_ctl.rc_last_delay_val) delay_calc -= bbr->r_ctl.rc_last_delay_val; else delay_calc = 0; } /* Mark that we have called bbr_output(). */ if ((bbr->r_timer_override) || (tp->t_state < TCPS_ESTABLISHED)) { /* Timeouts or early states are exempt */ if (inp->inp_in_hpts) tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT); } else if (inp->inp_in_hpts) { if ((bbr->r_ctl.rc_last_delay_val) && (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && delay_calc) { /* * We were being paced for output and the delay has * already exceeded when we were supposed to be * called, lets go ahead and pull out of the hpts * and call output. */ counter_u64_add(bbr_out_size[TCP_MSS_ACCT_LATE], 1); bbr->r_ctl.rc_last_delay_val = 0; tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT); } else if (tp->t_state == TCPS_CLOSED) { bbr->r_ctl.rc_last_delay_val = 0; tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT); } else { /* * On the hpts, you shall not pass! even if ACKNOW * is on, we will when the hpts fires, unless of * course we are overdue. */ counter_u64_add(bbr_out_size[TCP_MSS_ACCT_INPACE], 1); return (0); } } bbr->rc_cwnd_limited = 0; if (bbr->r_ctl.rc_last_delay_val) { /* recalculate the real delay and deal with over/under */ if (SEQ_GT(cts, bbr->rc_pacer_started)) delay_calc = cts - bbr->rc_pacer_started; else delay_calc = 0; if (delay_calc >= bbr->r_ctl.rc_last_delay_val) /* Setup the delay which will be added in */ delay_calc -= bbr->r_ctl.rc_last_delay_val; else { /* * We are early setup to adjust * our slot time. */ uint64_t merged_val; bbr->r_ctl.rc_agg_early += (bbr->r_ctl.rc_last_delay_val - delay_calc); bbr->r_agg_early_set = 1; if (bbr->r_ctl.rc_hptsi_agg_delay) { if (bbr->r_ctl.rc_hptsi_agg_delay >= bbr->r_ctl.rc_agg_early) { /* Nope our previous late cancels out the early */ bbr->r_ctl.rc_hptsi_agg_delay -= bbr->r_ctl.rc_agg_early; bbr->r_agg_early_set = 0; bbr->r_ctl.rc_agg_early = 0; } else { bbr->r_ctl.rc_agg_early -= bbr->r_ctl.rc_hptsi_agg_delay; bbr->r_ctl.rc_hptsi_agg_delay = 0; } } merged_val = bbr->rc_pacer_started; merged_val <<= 32; merged_val |= bbr->r_ctl.rc_last_delay_val; bbr_log_pacing_delay_calc(bbr, inp->inp_hpts_calls, bbr->r_ctl.rc_agg_early, cts, delay_calc, merged_val, bbr->r_agg_early_set, 3); bbr->r_ctl.rc_last_delay_val = 0; BBR_STAT_INC(bbr_early); delay_calc = 0; } } else { /* We were not delayed due to hptsi */ if (bbr->r_agg_early_set) bbr->r_ctl.rc_agg_early = 0; bbr->r_agg_early_set = 0; delay_calc = 0; } if (delay_calc) { /* * We had a hptsi delay which means we are falling behind on * sending at the expected rate. Calculate an extra amount * of data we can send, if any, to put us back on track. */ if ((bbr->r_ctl.rc_hptsi_agg_delay + delay_calc) < bbr->r_ctl.rc_hptsi_agg_delay) bbr->r_ctl.rc_hptsi_agg_delay = 0xffffffff; else bbr->r_ctl.rc_hptsi_agg_delay += delay_calc; } sendwin = min(tp->snd_wnd, tp->snd_cwnd); if ((tp->snd_una == tp->snd_max) && (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) && (sbavail(sb))) { /* * Ok we have been idle with nothing outstanding * we possibly need to start fresh with either a new * suite of states or a fast-ramp up. */ bbr_restart_after_idle(bbr, cts, bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time)); } /* * Now was there a hptsi delay where we are behind? We only count * being behind if: a) We are not in recovery. b) There was a delay. * c) We had room to send something. * */ hpts_calling = inp->inp_hpts_calls; inp->inp_hpts_calls = 0; if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { if (bbr_process_timers(tp, bbr, cts, hpts_calling)) { counter_u64_add(bbr_out_size[TCP_MSS_ACCT_ATIMER], 1); return (0); } } bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY; if (hpts_calling && (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { bbr->r_ctl.rc_last_delay_val = 0; } bbr->r_timer_override = 0; bbr->r_wanted_output = 0; /* * For TFO connections in SYN_RECEIVED, only allow the initial * SYN|ACK and those sent by the retransmit timer. */ if (IS_FASTOPEN(tp->t_flags) && ((tp->t_state == TCPS_SYN_RECEIVED) || (tp->t_state == TCPS_SYN_SENT)) && SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */ (tp->t_rxtshift == 0)) { /* not a retransmit */ len = 0; goto just_return_nolock; } /* * Before sending anything check for a state update. For hpts * calling without input this is important. If its input calling * then this was already done. */ if (bbr->rc_use_google == 0) bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); again: /* * If we've recently taken a timeout, snd_max will be greater than * snd_max. BBR in general does not pay much attention to snd_nxt * for historic reasons the persist timer still uses it. This means * we have to look at it. All retransmissions that are not persits * use the rsm that needs to be sent so snd_nxt is ignored. At the * end of this routine we pull snd_nxt always up to snd_max. */ doing_tlp = 0; #ifdef BBR_INVARIANTS doing_retran_from = picked_up_retran = 0; #endif error = 0; tso = 0; slot = 0; mtu = 0; sendwin = min(tp->snd_wnd, tp->snd_cwnd); sb_offset = tp->snd_max - tp->snd_una; flags = tcp_outflags[tp->t_state]; sack_rxmit = 0; len = 0; rsm = NULL; if (flags & TH_RST) { SOCKBUF_LOCK(sb); goto send; } recheck_resend: while (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) { /* We need to always have one in reserve */ rsm = bbr_alloc(bbr); if (rsm == NULL) { error = ENOMEM; /* Lie to get on the hpts */ tot_len = tp->t_maxseg; if (hpts_calling) /* Retry in a ms */ slot = 1001; goto just_return_nolock; } TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next); bbr->r_ctl.rc_free_cnt++; rsm = NULL; } /* What do we send, a resend? */ if (bbr->r_ctl.rc_resend == NULL) { /* Check for rack timeout */ bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); if (bbr->r_ctl.rc_resend) { #ifdef BBR_INVARIANTS picked_up_retran = 1; #endif bbr_cong_signal(tp, NULL, CC_NDUPACK, bbr->r_ctl.rc_resend); } } if (bbr->r_ctl.rc_resend) { rsm = bbr->r_ctl.rc_resend; #ifdef BBR_INVARIANTS doing_retran_from = 1; #endif /* Remove any TLP flags its a RACK or T-O */ rsm->r_flags &= ~BBR_TLP; bbr->r_ctl.rc_resend = NULL; if (SEQ_LT(rsm->r_start, tp->snd_una)) { #ifdef BBR_INVARIANTS panic("Huh, tp:%p bbr:%p rsm:%p start:%u < snd_una:%u\n", tp, bbr, rsm, rsm->r_start, tp->snd_una); goto recheck_resend; #else /* TSNH */ rsm = NULL; goto recheck_resend; #endif } rtr_cnt++; if (rsm->r_flags & BBR_HAS_SYN) { /* Only retransmit a SYN by itself */ len = 0; if ((flags & TH_SYN) == 0) { /* Huh something is wrong */ rsm->r_start++; if (rsm->r_start == rsm->r_end) { /* Clean it up, somehow we missed the ack? */ bbr_log_syn(tp, NULL); } else { /* TFO with data? */ rsm->r_flags &= ~BBR_HAS_SYN; len = rsm->r_end - rsm->r_start; } } else { /* Retransmitting SYN */ rsm = NULL; SOCKBUF_LOCK(sb); goto send; } } else len = rsm->r_end - rsm->r_start; if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) { len = maxseg; more_to_rxt = 1; } sb_offset = rsm->r_start - tp->snd_una; if (len > 0) { sack_rxmit = 1; KMOD_TCPSTAT_INC(tcps_sack_rexmits); KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes, min(len, maxseg)); } else { /* I dont think this can happen */ rsm = NULL; goto recheck_resend; } BBR_STAT_INC(bbr_resends_set); } else if (bbr->r_ctl.rc_tlp_send) { /* * Tail loss probe */ doing_tlp = 1; rsm = bbr->r_ctl.rc_tlp_send; bbr->r_ctl.rc_tlp_send = NULL; sack_rxmit = 1; len = rsm->r_end - rsm->r_start; rtr_cnt++; if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) len = maxseg; if (SEQ_GT(tp->snd_una, rsm->r_start)) { #ifdef BBR_INVARIANTS panic("tp:%p bbc:%p snd_una:%u rsm:%p r_start:%u", tp, bbr, tp->snd_una, rsm, rsm->r_start); #else /* TSNH */ rsm = NULL; goto recheck_resend; #endif } sb_offset = rsm->r_start - tp->snd_una; BBR_STAT_INC(bbr_tlp_set); } /* * Enforce a connection sendmap count limit if set * as long as we are not retransmiting. */ if ((rsm == NULL) && (V_tcp_map_entries_limit > 0) && (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { BBR_STAT_INC(bbr_alloc_limited); if (!bbr->alloc_limit_reported) { bbr->alloc_limit_reported = 1; BBR_STAT_INC(bbr_alloc_limited_conns); } goto just_return_nolock; } #ifdef BBR_INVARIANTS if (rsm && SEQ_LT(rsm->r_start, tp->snd_una)) { panic("tp:%p bbr:%p rsm:%p sb_offset:%u len:%u", tp, bbr, rsm, sb_offset, len); } #endif /* * Get standard flags, and add SYN or FIN if requested by 'hidden' * state flags. */ if (tp->t_flags & TF_NEEDFIN && (rsm == NULL)) flags |= TH_FIN; if (tp->t_flags & TF_NEEDSYN) flags |= TH_SYN; if (rsm && (rsm->r_flags & BBR_HAS_FIN)) { /* we are retransmitting the fin */ len--; if (len) { /* * When retransmitting data do *not* include the * FIN. This could happen from a TLP probe if we * allowed data with a FIN. */ flags &= ~TH_FIN; } } else if (rsm) { if (flags & TH_FIN) flags &= ~TH_FIN; } if ((sack_rxmit == 0) && (prefetch_rsm == 0)) { void *end_rsm; end_rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext); if (end_rsm) kern_prefetch(end_rsm, &prefetch_rsm); prefetch_rsm = 1; } SOCKBUF_LOCK(sb); /* * If snd_nxt == snd_max and we have transmitted a FIN, the * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a * negative length. This can also occur when TCP opens up its * congestion window while receiving additional duplicate acks after * fast-retransmit because TCP will reset snd_nxt to snd_max after * the fast-retransmit. * * In the normal retransmit-FIN-only case, however, snd_nxt will be * set to snd_una, the sb_offset will be 0, and the length may wind * up 0. * * If sack_rxmit is true we are retransmitting from the scoreboard * in which case len is already set. */ if (sack_rxmit == 0) { uint32_t avail; avail = sbavail(sb); if (SEQ_GT(tp->snd_max, tp->snd_una)) sb_offset = tp->snd_max - tp->snd_una; else sb_offset = 0; if (bbr->rc_tlp_new_data) { /* TLP is forcing out new data */ uint32_t tlplen; doing_tlp = 1; tlplen = maxseg; if (tlplen > (uint32_t)(avail - sb_offset)) { tlplen = (uint32_t)(avail - sb_offset); } if (tlplen > tp->snd_wnd) { len = tp->snd_wnd; } else { len = tlplen; } bbr->rc_tlp_new_data = 0; } else { what_we_can = len = bbr_what_can_we_send(tp, bbr, sendwin, avail, sb_offset, cts); if ((len < p_maxseg) && (bbr->rc_in_persist == 0) && (ctf_outstanding(tp) >= (2 * p_maxseg)) && ((avail - sb_offset) >= p_maxseg)) { /* * We are not completing whats in the socket * buffer (i.e. there is at least a segment * waiting to send) and we have 2 or more * segments outstanding. There is no sense * of sending a little piece. Lets defer and * and wait until we can send a whole * segment. */ len = 0; } if (bbr->rc_in_persist) { /* * We are in persists, figure out if * a retransmit is available (maybe the previous * persists we sent) or if we have to send new * data. */ rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); if (rsm) { len = rsm->r_end - rsm->r_start; if (rsm->r_flags & BBR_HAS_FIN) len--; if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) len = maxseg; if (len > 1) BBR_STAT_INC(bbr_persist_reneg); /* * XXXrrs we could force the len to * 1 byte here to cause the chunk to * split apart.. but that would then * mean we always retransmit it as * one byte even after the window * opens. */ sack_rxmit = 1; sb_offset = rsm->r_start - tp->snd_una; } else { /* * First time through in persists or peer * acked our one byte. Though we do have * to have something in the sb. */ len = 1; sb_offset = 0; if (avail == 0) len = 0; } } } } if (prefetch_so_done == 0) { kern_prefetch(so, &prefetch_so_done); prefetch_so_done = 1; } /* * Lop off SYN bit if it has already been sent. However, if this is * SYN-SENT state and if segment contains data and if we don't know * that foreign host supports TAO, suppress sending segment. */ if ((flags & TH_SYN) && (rsm == NULL) && SEQ_GT(tp->snd_max, tp->snd_una)) { if (tp->t_state != TCPS_SYN_RECEIVED) flags &= ~TH_SYN; /* * When sending additional segments following a TFO SYN|ACK, * do not include the SYN bit. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_state == TCPS_SYN_RECEIVED)) flags &= ~TH_SYN; sb_offset--, len++; if (sbavail(sb) == 0) len = 0; } else if ((flags & TH_SYN) && rsm) { /* * Subtract one from the len for the SYN being * retransmitted. */ len--; } /* * Be careful not to send data and/or FIN on SYN segments. This * measure is needed to prevent interoperability problems with not * fully conformant TCP implementations. */ if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { len = 0; flags &= ~TH_FIN; } /* * On TFO sockets, ensure no data is sent in the following cases: * * - When retransmitting SYN|ACK on a passively-created socket * - When retransmitting SYN on an actively created socket * - When sending a zero-length cookie (cookie request) on an * actively created socket * - When the socket is in the CLOSED state (RST is being sent) */ if (IS_FASTOPEN(tp->t_flags) && (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || ((tp->t_state == TCPS_SYN_SENT) && (tp->t_tfo_client_cookie_len == 0)) || (flags & TH_RST))) { len = 0; sack_rxmit = 0; rsm = NULL; } /* Without fast-open there should never be data sent on a SYN */ if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) len = 0; if (len <= 0) { /* * If FIN has been sent but not acked, but we haven't been * called to retransmit, len will be < 0. Otherwise, window * shrank after we sent into it. If window shrank to 0, * cancel pending retransmit, pull snd_nxt back to (closed) * window, and set the persist timer if it isn't already * going. If the window didn't close completely, just wait * for an ACK. * * We also do a general check here to ensure that we will * set the persist timer when we have data to send, but a * 0-byte window. This makes sure the persist timer is set * even if the packet hits one of the "goto send" lines * below. */ len = 0; if ((tp->snd_wnd == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && (tp->snd_una == tp->snd_max) && (sb_offset < (int)sbavail(sb))) { /* * Not enough room in the rwnd to send * a paced segment out. */ bbr_enter_persist(tp, bbr, cts, __LINE__); } } else if ((rsm == NULL) && (doing_tlp == 0) && (len < bbr->r_ctl.rc_pace_max_segs)) { /* * We are not sending a full segment for * some reason. Should we not send anything (think * sws or persists)? */ if ((tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && (TCPS_HAVEESTABLISHED(tp->t_state)) && (len < (int)(sbavail(sb) - sb_offset))) { /* * Here the rwnd is less than * the pacing size, this is not a retransmit, * we are established and * the send is not the last in the socket buffer * lets not send, and possibly enter persists. */ len = 0; if (tp->snd_max == tp->snd_una) bbr_enter_persist(tp, bbr, cts, __LINE__); } else if ((tp->snd_cwnd >= bbr->r_ctl.rc_pace_max_segs) && (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) && (len < (int)(sbavail(sb) - sb_offset)) && (len < bbr_minseg(bbr))) { /* * Here we are not retransmitting, and * the cwnd is not so small that we could * not send at least a min size (rxt timer * not having gone off), We have 2 segments or * more already in flight, its not the tail end * of the socket buffer and the cwnd is blocking * us from sending out minimum pacing segment size. * Lets not send anything. */ bbr->rc_cwnd_limited = 1; len = 0; } else if (((tp->snd_wnd - ctf_outstanding(tp)) < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) && (len < (int)(sbavail(sb) - sb_offset)) && (TCPS_HAVEESTABLISHED(tp->t_state))) { /* * Here we have a send window but we have * filled it up and we can't send another pacing segment. * We also have in flight more than 2 segments * and we are not completing the sb i.e. we allow * the last bytes of the sb to go out even if * its not a full pacing segment. */ len = 0; } } /* len will be >= 0 after this point. */ KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); tcp_sndbuf_autoscale(tp, so, sendwin); /* * */ if (bbr->rc_in_persist && len && (rsm == NULL) && (len < min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs))) { /* * We are in persist, not doing a retransmit and don't have enough space * yet to send a full TSO. So is it at the end of the sb * if so we need to send else nuke to 0 and don't send. */ int sbleft; if (sbavail(sb) > sb_offset) sbleft = sbavail(sb) - sb_offset; else sbleft = 0; if (sbleft >= min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs)) { /* not at end of sb lets not send */ len = 0; } } /* * Decide if we can use TCP Segmentation Offloading (if supported by * hardware). * * TSO may only be used if we are in a pure bulk sending state. The * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP * options prevent using TSO. With TSO the TCP header is the same * (except for the sequence number) for all generated packets. This * makes it impossible to transmit any options which vary per * generated segment or packet. * * IPv4 handling has a clear separation of ip options and ip header * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() * does the right thing below to provide length of just ip options * and thus checking for ipoptlen is enough to decide if ip options * are present. */ #ifdef INET6 if (isipv6) ipoptlen = ip6_optlen(inp); else #endif if (inp->inp_options) ipoptlen = inp->inp_options->m_len - offsetof(struct ipoption, ipopt_list); else ipoptlen = 0; #if defined(IPSEC) || defined(IPSEC_SUPPORT) /* * Pre-calculate here as we save another lookup into the darknesses * of IPsec that way and can actually decide if TSO is ok. */ #ifdef INET6 if (isipv6 && IPSEC_ENABLED(ipv6)) ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp); #ifdef INET else #endif #endif /* INET6 */ #ifdef INET if (IPSEC_ENABLED(ipv4)) ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp); #endif /* INET */ #endif /* IPSEC */ #if defined(IPSEC) || defined(IPSEC_SUPPORT) ipoptlen += ipsec_optlen; #endif if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && (len > maxseg) && (tp->t_port == 0) && ((tp->t_flags & TF_SIGNATURE) == 0) && tp->rcv_numsacks == 0 && ipoptlen == 0) tso = 1; recwin = lmin(lmax(sbspace(&so->so_rcv), 0), (long)TCP_MAXWIN << tp->rcv_scale); /* * Sender silly window avoidance. We transmit under the following * conditions when len is non-zero: * * - We have a full segment (or more with TSO) - This is the last * buffer in a write()/send() and we are either idle or running * NODELAY - we've timed out (e.g. persist timer) - we have more * then 1/2 the maximum send window's worth of data (receiver may be * limited the window size) - we need to retransmit */ if (rsm) goto send; if (len) { if (sack_rxmit) goto send; if (len >= p_maxseg) goto send; /* * NOTE! on localhost connections an 'ack' from the remote * end may occur synchronously with the output and cause us * to flush a buffer queued with moretocome. XXX * */ if (((tp->t_flags & TF_MORETOCOME) == 0) && /* normal case */ ((tp->t_flags & TF_NODELAY) || ((uint32_t)len + (uint32_t)sb_offset) >= sbavail(&so->so_snd)) && (tp->t_flags & TF_NOPUSH) == 0) { goto send; } if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */ goto send; } if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) { goto send; } } /* * Sending of standalone window updates. * * Window updates are important when we close our window due to a * full socket buffer and are opening it again after the application * reads data from it. Once the window has opened again and the * remote end starts to send again the ACK clock takes over and * provides the most current window information. * * We must avoid the silly window syndrome whereas every read from * the receive buffer, no matter how small, causes a window update * to be sent. We also should avoid sending a flurry of window * updates when the socket buffer had queued a lot of data and the * application is doing small reads. * * Prevent a flurry of pointless window updates by only sending an * update when we can increase the advertized window by more than * 1/4th of the socket buffer capacity. When the buffer is getting * full or is very small be more aggressive and send an update * whenever we can increase by two mss sized segments. In all other * situations the ACK's to new incoming data will carry further * window increases. * * Don't send an independent window update if a delayed ACK is * pending (it will get piggy-backed on it) or the remote side * already has done a half-close and won't send more data. Skip * this if the connection is in T/TCP half-open state. */ if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && !(tp->t_flags & TF_DELACK) && !TCPS_HAVERCVDFIN(tp->t_state)) { /* Check to see if we should do a window update */ if (bbr_window_update_needed(tp, so, recwin, maxseg)) goto send; } /* * Send if we owe the peer an ACK, RST, SYN. ACKNOW * is also a catch-all for the retransmit timer timeout case. */ if (tp->t_flags & TF_ACKNOW) { goto send; } if (flags & TH_RST) { /* Always send a RST if one is due */ goto send; } if ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0) { goto send; } /* * If our state indicates that FIN should be sent and we have not * yet done so, then we need to send. */ if (flags & TH_FIN && ((tp->t_flags & TF_SENTFIN) == 0)) { goto send; } /* * No reason to send a segment, just return. */ just_return: SOCKBUF_UNLOCK(sb); just_return_nolock: if (tot_len) slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0); if (bbr->rc_no_pacing) slot = 0; if (tot_len == 0) { if ((ctf_outstanding(tp) + min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) >= tp->snd_wnd) { BBR_STAT_INC(bbr_rwnd_limited); app_limited = BBR_JR_RWND_LIMITED; bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); if ((bbr->rc_in_persist == 0) && TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_max == tp->snd_una) && sbavail(&tp->t_inpcb->inp_socket->so_snd)) { /* No send window.. we must enter persist */ bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); } } else if (ctf_outstanding(tp) >= sbavail(sb)) { BBR_STAT_INC(bbr_app_limited); app_limited = BBR_JR_APP_LIMITED; bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); } else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + p_maxseg) >= tp->snd_cwnd) { BBR_STAT_INC(bbr_cwnd_limited); app_limited = BBR_JR_CWND_LIMITED; bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes))); bbr->rc_cwnd_limited = 1; } else { BBR_STAT_INC(bbr_app_limited); app_limited = BBR_JR_APP_LIMITED; bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); } bbr->r_ctl.rc_hptsi_agg_delay = 0; bbr->r_agg_early_set = 0; bbr->r_ctl.rc_agg_early = 0; bbr->r_ctl.rc_last_delay_val = 0; } else if (bbr->rc_use_google == 0) bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); /* Are we app limited? */ if ((app_limited == BBR_JR_APP_LIMITED) || (app_limited == BBR_JR_RWND_LIMITED)) { /** * We are application limited. */ bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_delivered); } if (tot_len == 0) counter_u64_add(bbr_out_size[TCP_MSS_ACCT_JUSTRET], 1); /* Dont update the time if we did not send */ bbr->r_ctl.rc_last_delay_val = 0; bbr->rc_output_starts_timer = 1; bbr_start_hpts_timer(bbr, tp, cts, 9, slot, tot_len); bbr_log_type_just_return(bbr, cts, tot_len, hpts_calling, app_limited, p_maxseg, len); if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* Make sure snd_nxt is drug up */ tp->snd_nxt = tp->snd_max; } return (error); send: if (doing_tlp == 0) { /* * Data not a TLP, and its not the rxt firing. If it is the * rxt firing, we want to leave the tlp_in_progress flag on * so we don't send another TLP. It has to be a rack timer * or normal send (response to acked data) to clear the tlp * in progress flag. */ bbr->rc_tlp_in_progress = 0; bbr->rc_tlp_rtx_out = 0; } else { /* * Its a TLP. */ bbr->rc_tlp_in_progress = 1; } bbr_timer_cancel(bbr, __LINE__, cts); if (rsm == NULL) { if (sbused(sb) > 0) { /* * This is sub-optimal. We only send a stand alone * FIN on its own segment. */ if (flags & TH_FIN) { flags &= ~TH_FIN; if ((len == 0) && ((tp->t_flags & TF_ACKNOW) == 0)) { /* Lets not send this */ slot = 0; goto just_return; } } } } else { /* * We do *not* send a FIN on a retransmit if it has data. * The if clause here where len > 1 should never come true. */ if ((len > 0) && (((rsm->r_flags & BBR_HAS_FIN) == 0) && (flags & TH_FIN))) { flags &= ~TH_FIN; len--; } } SOCKBUF_LOCK_ASSERT(sb); if (len > 0) { if ((tp->snd_una == tp->snd_max) && (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { /* * This qualifies as a RTT_PROBE session since we * drop the data outstanding to nothing and waited * more than bbr_rtt_probe_time. */ bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); bbr_set_reduced_rtt(bbr, cts, __LINE__); } if (len >= maxseg) tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; else tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; } /* * Before ESTABLISHED, force sending of initial options unless TCP * set not to do any options. NOTE: we assume that the IP/TCP header * plus TCP options always fit in a single mbuf, leaving room for a * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr) * + optlen <= MCLBYTES */ optlen = 0; #ifdef INET6 if (isipv6) hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); else #endif hdrlen = sizeof(struct tcpiphdr); /* * Compute options for segment. We only have to care about SYN and * established connection segments. Options for SYN-ACK segments * are handled in TCP syncache. */ to.to_flags = 0; local_options = 0; if ((tp->t_flags & TF_NOOPT) == 0) { /* Maximum segment size. */ if (flags & TH_SYN) { to.to_mss = tcp_mssopt(&inp->inp_inc); -#ifdef NETFLIX_TCPOUDP if (tp->t_port) to.to_mss -= V_tcp_udp_tunneling_overhead; -#endif to.to_flags |= TOF_MSS; /* * On SYN or SYN|ACK transmits on TFO connections, * only include the TFO option if it is not a * retransmit, as the presence of the TFO option may * have caused the original SYN or SYN|ACK to have * been dropped by a middlebox. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_rxtshift == 0)) { if (tp->t_state == TCPS_SYN_RECEIVED) { to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; to.to_tfo_cookie = (u_int8_t *)&tp->t_tfo_cookie.server; to.to_flags |= TOF_FASTOPEN; wanted_cookie = 1; } else if (tp->t_state == TCPS_SYN_SENT) { to.to_tfo_len = tp->t_tfo_client_cookie_len; to.to_tfo_cookie = tp->t_tfo_cookie.client; to.to_flags |= TOF_FASTOPEN; wanted_cookie = 1; } } } /* Window scaling. */ if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { to.to_wscale = tp->request_r_scale; to.to_flags |= TOF_SCALE; } /* Timestamps. */ if ((tp->t_flags & TF_RCVD_TSTMP) || ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { to.to_tsval = tcp_tv_to_mssectick(&bbr->rc_tv) + tp->ts_offset; to.to_tsecr = tp->ts_recent; to.to_flags |= TOF_TS; local_options += TCPOLEN_TIMESTAMP + 2; } /* Set receive buffer autosizing timestamp. */ if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE)) tp->rfbuf_ts = tcp_tv_to_mssectick(&bbr->rc_tv); /* Selective ACK's. */ if (flags & TH_SYN) to.to_flags |= TOF_SACKPERM; else if (TCPS_HAVEESTABLISHED(tp->t_state) && tp->rcv_numsacks > 0) { to.to_flags |= TOF_SACK; to.to_nsacks = tp->rcv_numsacks; to.to_sacks = (u_char *)tp->sackblks; } #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) /* TCP-MD5 (RFC2385). */ if (tp->t_flags & TF_SIGNATURE) to.to_flags |= TOF_SIGNATURE; #endif /* TCP_SIGNATURE */ /* Processing the options. */ hdrlen += (optlen = tcp_addoptions(&to, opt)); /* * If we wanted a TFO option to be added, but it was unable * to fit, ensure no data is sent. */ if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && !(to.to_flags & TOF_FASTOPEN)) len = 0; } -#ifdef NETFLIX_TCPOUDP if (tp->t_port) { if (V_tcp_udp_tunneling_port == 0) { /* The port was removed?? */ SOCKBUF_UNLOCK(&so->so_snd); return (EHOSTUNREACH); } hdrlen += sizeof(struct udphdr); } -#endif #ifdef INET6 if (isipv6) ipoptlen = ip6_optlen(tp->t_inpcb); else #endif if (tp->t_inpcb->inp_options) ipoptlen = tp->t_inpcb->inp_options->m_len - offsetof(struct ipoption, ipopt_list); else ipoptlen = 0; ipoptlen = 0; #if defined(IPSEC) || defined(IPSEC_SUPPORT) ipoptlen += ipsec_optlen; #endif if (bbr->rc_last_options != local_options) { /* * Cache the options length this generally does not change * on a connection. We use this to calculate TSO. */ bbr->rc_last_options = local_options; } maxseg = tp->t_maxseg - (ipoptlen + optlen); p_maxseg = min(maxseg, pace_max_segs); /* * Adjust data length if insertion of options will bump the packet * length beyond the t_maxseg length. Clear the FIN bit because we * cut off the tail of the segment. */ if (len > maxseg) { if (len != 0 && (flags & TH_FIN)) { flags &= ~TH_FIN; } if (tso) { uint32_t moff; int32_t max_len; /* extract TSO information */ if_hw_tsomax = tp->t_tsomax; if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; KASSERT(ipoptlen == 0, ("%s: TSO can't do IP options", __func__)); /* * Check if we should limit by maximum payload * length: */ if (if_hw_tsomax != 0) { /* compute maximum TSO length */ max_len = (if_hw_tsomax - hdrlen - max_linkhdr); if (max_len <= 0) { len = 0; } else if (len > max_len) { len = max_len; } } /* * Prevent the last segment from being fractional * unless the send sockbuf can be emptied: */ if ((sb_offset + len) < sbavail(sb)) { moff = len % (uint32_t)maxseg; if (moff != 0) { len -= moff; } } /* * In case there are too many small fragments don't * use TSO: */ if (len <= maxseg) { len = maxseg; tso = 0; } } else { /* Not doing TSO */ if (optlen + ipoptlen >= tp->t_maxseg) { /* * Since we don't have enough space to put * the IP header chain and the TCP header in * one packet as required by RFC 7112, don't * send it. Also ensure that at least one * byte of the payload can be put into the * TCP segment. */ SOCKBUF_UNLOCK(&so->so_snd); error = EMSGSIZE; sack_rxmit = 0; goto out; } len = maxseg; } } else { /* Not doing TSO */ if_hw_tsomaxsegcount = 0; tso = 0; } KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, ("%s: len > IP_MAXPACKET", __func__)); #ifdef DIAGNOSTIC #ifdef INET6 if (max_linkhdr + hdrlen > MCLBYTES) #else if (max_linkhdr + hdrlen > MHLEN) #endif panic("tcphdr too big"); #endif /* * This KASSERT is here to catch edge cases at a well defined place. * Before, those had triggered (random) panic conditions further * down. */ #ifdef BBR_INVARIANTS if (sack_rxmit) { if (SEQ_LT(rsm->r_start, tp->snd_una)) { panic("RSM:%p TP:%p bbr:%p start:%u is < snd_una:%u", rsm, tp, bbr, rsm->r_start, tp->snd_una); } } #endif KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); if ((len == 0) && (flags & TH_FIN) && (sbused(sb))) { /* * We have outstanding data, don't send a fin by itself!. */ slot = 0; goto just_return; } /* * Grab a header mbuf, attaching a copy of data to be transmitted, * and initialize the header from the template for sends on this * connection. */ if (len) { uint32_t moff; uint32_t orig_len; /* * We place a limit on sending with hptsi. */ if ((rsm == NULL) && len > pace_max_segs) len = pace_max_segs; if (len <= maxseg) tso = 0; #ifdef INET6 if (MHLEN < hdrlen + max_linkhdr) m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); else #endif m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) { BBR_STAT_INC(bbr_failed_mbuf_aloc); bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0); SOCKBUF_UNLOCK(sb); error = ENOBUFS; sack_rxmit = 0; goto out; } m->m_data += max_linkhdr; m->m_len = hdrlen; /* * Start the m_copy functions from the closest mbuf to the * sb_offset in the socket buffer chain. */ if ((sb_offset > sbavail(sb)) || ((len + sb_offset) > sbavail(sb))) { #ifdef BBR_INVARIANTS if ((len + sb_offset) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) panic("tp:%p bbr:%p len:%u sb_offset:%u sbavail:%u rsm:%p %u:%u:%u", tp, bbr, len, sb_offset, sbavail(sb), rsm, doing_retran_from, picked_up_retran, doing_tlp); #endif /* * In this messed up situation we have two choices, * a) pretend the send worked, and just start timers * and what not (not good since that may lead us * back here a lot). b) Send the lowest segment * in the map. c) Drop the connection. Lets do * which if it continues to happen will lead to * via timeouts. */ BBR_STAT_INC(bbr_offset_recovery); rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); sb_offset = 0; if (rsm == NULL) { sack_rxmit = 0; len = sbavail(sb); } else { sack_rxmit = 1; if (rsm->r_start != tp->snd_una) { /* * Things are really messed up, * is the only thing to do. */ BBR_STAT_INC(bbr_offset_drop); tcp_set_inp_to_drop(inp, EFAULT); SOCKBUF_UNLOCK(sb); (void)m_free(m); return (0); } len = rsm->r_end - rsm->r_start; } if (len > sbavail(sb)) len = sbavail(sb); if (len > maxseg) len = maxseg; } mb = sbsndptr_noadv(sb, sb_offset, &moff); if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) { m_copydata(mb, moff, (int)len, mtod(m, caddr_t)+hdrlen); if (rsm == NULL) sbsndptr_adv(sb, mb, len); m->m_len += len; } else { struct sockbuf *msb; if (rsm) msb = NULL; else msb = sb; #ifdef BBR_INVARIANTS if ((len + moff) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) { if (rsm) { panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u rsm:%p snd_una:%u rsm_start:%u flg:%x %u:%u:%u sr:%d ", tp, bbr, len, moff, sbavail(sb), rsm, tp->snd_una, rsm->r_flags, rsm->r_start, doing_retran_from, picked_up_retran, doing_tlp, sack_rxmit); } else { panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u sb_offset:%u snd_una:%u", tp, bbr, len, moff, sbavail(sb), sb_offset, tp->snd_una); } } #endif orig_len = len; m->m_next = tcp_m_copym( mb, moff, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb, ((rsm == NULL) ? hw_tls : 0) #ifdef NETFLIX_COPY_ARGS , &filled_all #endif ); if (len <= maxseg) { /* * Must have ran out of mbufs for the copy * shorten it to no longer need tso. Lets * not put on sendalot since we are low on * mbufs. */ tso = 0; } if (m->m_next == NULL) { SOCKBUF_UNLOCK(sb); (void)m_free(m); error = ENOBUFS; sack_rxmit = 0; goto out; } } #ifdef BBR_INVARIANTS if (tso && len < maxseg) { panic("tp:%p tso on, but len:%d < maxseg:%d", tp, len, maxseg); } if (tso && if_hw_tsomaxsegcount) { int32_t seg_cnt = 0; struct mbuf *foo; foo = m; while (foo) { seg_cnt++; foo = foo->m_next; } if (seg_cnt > if_hw_tsomaxsegcount) { panic("seg_cnt:%d > max:%d", seg_cnt, if_hw_tsomaxsegcount); } } #endif /* * If we're sending everything we've got, set PUSH. (This * will keep happy those implementations which only give * data to the user when a buffer fills or a PUSH comes in.) */ if (sb_offset + len == sbused(sb) && sbused(sb) && !(flags & TH_SYN)) { flags |= TH_PUSH; } SOCKBUF_UNLOCK(sb); } else { SOCKBUF_UNLOCK(sb); if (tp->t_flags & TF_ACKNOW) KMOD_TCPSTAT_INC(tcps_sndacks); else if (flags & (TH_SYN | TH_FIN | TH_RST)) KMOD_TCPSTAT_INC(tcps_sndctrl); else KMOD_TCPSTAT_INC(tcps_sndwinup); m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) { BBR_STAT_INC(bbr_failed_mbuf_aloc); bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0); error = ENOBUFS; /* Fudge the send time since we could not send */ sack_rxmit = 0; goto out; } #ifdef INET6 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && MHLEN >= hdrlen) { M_ALIGN(m, hdrlen); } else #endif m->m_data += max_linkhdr; m->m_len = hdrlen; } SOCKBUF_UNLOCK_ASSERT(sb); m->m_pkthdr.rcvif = (struct ifnet *)0; #ifdef MAC mac_inpcb_create_mbuf(inp, m); #endif #ifdef INET6 if (isipv6) { ip6 = mtod(m, struct ip6_hdr *); -#ifdef NETFLIX_TCPOUDP if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + sizeof(struct ip6_hdr)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; ulen = hdrlen + len - sizeof(struct ip6_hdr); udp->uh_ulen = htons(ulen); th = (struct tcphdr *)(udp + 1); } else { -#endif th = (struct tcphdr *)(ip6 + 1); - -#ifdef NETFLIX_TCPOUDP } -#endif - tcpip_fillheaders(inp, -#ifdef NETFLIX_TCPOUDP - tp->t_port, -#endif - ip6, th); + tcpip_fillheaders(inp, tp->t_port, ip6, th); } else #endif /* INET6 */ { ip = mtod(m, struct ip *); #ifdef TCPDEBUG ipov = (struct ipovly *)ip; #endif -#ifdef NETFLIX_TCPOUDP if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip + ipoptlen + sizeof(struct ip)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; ulen = hdrlen + len - sizeof(struct ip); udp->uh_ulen = htons(ulen); th = (struct tcphdr *)(udp + 1); - } else -#endif + } else { th = (struct tcphdr *)(ip + 1); - tcpip_fillheaders(inp, -#ifdef NETFLIX_TCPOUDP - tp->t_port, -#endif - ip, th); + } + tcpip_fillheaders(inp, tp->t_port, ip, th); } /* * If we are doing retransmissions, then snd_nxt will not reflect * the first unsent octet. For ACK only packets, we do not want the * sequence number of the retransmitted packet, we want the sequence * number of the next unsent octet. So, if there is no data (and no * SYN or FIN), use snd_max instead of snd_nxt when filling in * ti_seq. But if we are in persist state, snd_max might reflect * one byte beyond the right edge of the window, so use snd_nxt in * that case, since we know we aren't doing a retransmission. * (retransmit and persist are mutually exclusive...) */ if (sack_rxmit == 0) { if (len && ((flags & (TH_FIN | TH_SYN | TH_RST)) == 0)) { /* New data (including new persists) */ th->th_seq = htonl(tp->snd_max); bbr_seq = tp->snd_max; } else if (flags & TH_SYN) { /* Syn's always send from iss */ th->th_seq = htonl(tp->iss); bbr_seq = tp->iss; } else if (flags & TH_FIN) { if (flags & TH_FIN && tp->t_flags & TF_SENTFIN) { /* * If we sent the fin already its 1 minus * snd_max */ th->th_seq = (htonl(tp->snd_max - 1)); bbr_seq = (tp->snd_max - 1); } else { /* First time FIN use snd_max */ th->th_seq = htonl(tp->snd_max); bbr_seq = tp->snd_max; } } else if (flags & TH_RST) { /* * For a Reset send the last cum ack in sequence * (this like any other choice may still generate a * challenge ack, if a ack-update packet is in * flight). */ th->th_seq = htonl(tp->snd_una); bbr_seq = tp->snd_una; } else { /* * len == 0 and not persist we use snd_max, sending * an ack unless we have sent the fin then its 1 * minus. */ /* * XXXRRS Question if we are in persists and we have * nothing outstanding to send and we have not sent * a FIN, we will send an ACK. In such a case it * might be better to send (tp->snd_una - 1) which * would force the peer to ack. */ if (tp->t_flags & TF_SENTFIN) { th->th_seq = htonl(tp->snd_max - 1); bbr_seq = (tp->snd_max - 1); } else { th->th_seq = htonl(tp->snd_max); bbr_seq = tp->snd_max; } } } else { /* All retransmits use the rsm to guide the send */ th->th_seq = htonl(rsm->r_start); bbr_seq = rsm->r_start; } th->th_ack = htonl(tp->rcv_nxt); if (optlen) { bcopy(opt, th + 1, optlen); th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; } th->th_flags = flags; /* * Calculate receive window. Don't shrink window, but avoid silly * window syndrome. */ if ((flags & TH_RST) || ((recwin < (so->so_rcv.sb_hiwat / 4) && recwin < maxseg))) recwin = 0; if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && recwin < (tp->rcv_adv - tp->rcv_nxt)) recwin = (tp->rcv_adv - tp->rcv_nxt); if (recwin > TCP_MAXWIN << tp->rcv_scale) recwin = TCP_MAXWIN << tp->rcv_scale; /* * According to RFC1323 the window field in a SYN (i.e., a or * ) segment itself is never scaled. The case is * handled in syncache. */ if (flags & TH_SYN) th->th_win = htons((u_short) (min(sbspace(&so->so_rcv), TCP_MAXWIN))); else { /* Avoid shrinking window with window scaling. */ recwin = roundup2(recwin, 1 << tp->rcv_scale); th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); } /* * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0 * window. This may cause the remote transmitter to stall. This * flag tells soreceive() to disable delayed acknowledgements when * draining the buffer. This can occur if the receiver is * attempting to read more data than can be buffered prior to * transmitting on the connection. */ if (th->th_win == 0) { tp->t_sndzerowin++; tp->t_flags |= TF_RXWIN0SENT; } else tp->t_flags &= ~TF_RXWIN0SENT; /* * We don't support urgent data, but drag along * the pointer in case of a stack switch. */ tp->snd_up = tp->snd_una; #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (to.to_flags & TOF_SIGNATURE) { /* * Calculate MD5 signature and put it into the place * determined before. NOTE: since TCP options buffer doesn't * point into mbuf's data, calculate offset and use it. */ if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, (u_char *)(th + 1) + (to.to_signature - opt)) != 0) { /* * Do not send segment if the calculation of MD5 * digest has failed. */ goto out; } } #endif /* * Put TCP length in extended header, and then checksum extended * header and data. */ m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ #ifdef INET6 if (isipv6) { /* * ip6_plen is not need to be filled now, and will be filled * in ip6_output. */ -#ifdef NETFLIX_TCPOUDP if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { -#endif csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 0); -#ifdef NETFLIX_TCPOUDP } -#endif } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { -#ifdef NETFLIX_TCPOUDP if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { -#endif csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); -#ifdef NETFLIX_TCPOUDP } -#endif /* IP version must be set here for ipv4/ipv6 checking later */ KASSERT(ip->ip_v == IPVERSION, ("%s: IP version incorrect: %d", __func__, ip->ip_v)); } #endif /* * Enable TSO and specify the size of the segments. The TCP pseudo * header checksum is always provided. XXX: Fixme: This is currently * not the case for IPv6. */ if (tso) { KASSERT(len > maxseg, ("%s: len:%d <= tso_segsz:%d", __func__, len, maxseg)); m->m_pkthdr.csum_flags |= CSUM_TSO; csum_flags |= CSUM_TSO; m->m_pkthdr.tso_segsz = maxseg; } KASSERT(len + hdrlen == m_length(m, NULL), ("%s: mbuf chain different than expected: %d + %u != %u", __func__, len, hdrlen, m_length(m, NULL))); #ifdef TCP_HHOOK /* Run HHOOK_TC_ESTABLISHED_OUT helper hooks. */ hhook_run_tcp_est_out(tp, th, &to, len, tso); #endif #ifdef TCPDEBUG /* * Trace. */ if (so->so_options & SO_DEBUG) { u_short save = 0; #ifdef INET6 if (!isipv6) #endif { save = ipov->ih_len; ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + * (th->th_off << 2) */ ); } tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); #ifdef INET6 if (!isipv6) #endif ipov->ih_len = save; } #endif /* TCPDEBUG */ /* Log to the black box */ if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); /* Record info on type of transmission */ log.u_bbr.flex1 = bbr->r_ctl.rc_hptsi_agg_delay; log.u_bbr.flex2 = (bbr->r_recovery_bw << 3); log.u_bbr.flex3 = maxseg; log.u_bbr.flex4 = delay_calc; /* Encode filled_all into the upper flex5 bit */ log.u_bbr.flex5 = bbr->rc_past_init_win; log.u_bbr.flex5 <<= 1; log.u_bbr.flex5 |= bbr->rc_no_pacing; log.u_bbr.flex5 <<= 29; if (filled_all) log.u_bbr.flex5 |= 0x80000000; log.u_bbr.flex5 |= tp->t_maxseg; log.u_bbr.flex6 = bbr->r_ctl.rc_pace_max_segs; log.u_bbr.flex7 = (bbr->rc_bbr_state << 8) | bbr_state_val(bbr); /* lets poke in the low and the high here for debugging */ log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg; if (rsm || sack_rxmit) { if (doing_tlp) log.u_bbr.flex8 = 2; else log.u_bbr.flex8 = 1; } else { log.u_bbr.flex8 = 0; } lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, len, &log, false, NULL, NULL, 0, tv); } else { lgb = NULL; } /* * Fill in IP length and desired time to live and send to IP level. * There should be a better way to handle ttl and tos; we could keep * them in the template, but need a way to checksum without them. */ /* * m->m_pkthdr.len should have been set before cksum calcuration, * because in6_cksum() need it. */ #ifdef INET6 if (isipv6) { /* * we separately set hoplimit for every segment, since the * user might want to change the value via setsockopt. Also, * desired default hop limit might be changed via Neighbor * Discovery. */ ip6->ip6_hlim = in6_selecthlim(inp, NULL); /* * Set the packet size here for the benefit of DTrace * probes. ip6_output() will set it properly; it's supposed * to include the option header lengths as well. */ ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); if (V_path_mtu_discovery && maxseg > V_tcp_minmss) tp->t_flags2 |= TF2_PLPMTU_PMTUD; else tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; if (tp->t_state == TCPS_SYN_SENT) TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); TCP_PROBE5(send, NULL, tp, ip6, tp, th); /* TODO: IPv6 IP6TOS_ECT bit on */ error = ip6_output(m, inp->in6p_outputopts, &inp->inp_route6, ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), NULL, NULL, inp); if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL) mtu = inp->inp_route6.ro_nh->nh_mtu; } #endif /* INET6 */ #if defined(INET) && defined(INET6) else #endif #ifdef INET { ip->ip_len = htons(m->m_pkthdr.len); #ifdef INET6 if (isipv6) ip->ip_ttl = in6_selecthlim(inp, NULL); #endif /* INET6 */ /* * If we do path MTU discovery, then we set DF on every * packet. This might not be the best thing to do according * to RFC3390 Section 2. However the tcp hostcache migitates * the problem so it affects only the first tcp connection * with a host. * * NB: Don't set DF on small MTU/MSS to have a safe * fallback. */ if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; if (tp->t_port == 0 || len < V_tcp_minmss) { ip->ip_off |= htons(IP_DF); } } else { tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; } if (tp->t_state == TCPS_SYN_SENT) TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); TCP_PROBE5(send, NULL, tp, ip, tp, th); error = ip_output(m, inp->inp_options, &inp->inp_route, ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0, inp); if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL) mtu = inp->inp_route.ro_nh->nh_mtu; } #endif /* INET */ out: if (lgb) { lgb->tlb_errno = error; lgb = NULL; } /* * In transmit state, time the transmission and arrange for the * retransmit. In persist state, just set snd_max. */ if (error == 0) { if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->t_flags & TF_SACK_PERMIT) && tp->rcv_numsacks > 0) tcp_clean_dsack_blocks(tp); /* We sent an ack clear the bbr_segs_rcvd count */ bbr->output_error_seen = 0; bbr->oerror_cnt = 0; bbr->bbr_segs_rcvd = 0; if (len == 0) counter_u64_add(bbr_out_size[TCP_MSS_ACCT_SNDACK], 1); /* Do accounting for new sends */ if ((len > 0) && (rsm == NULL)) { int idx; if (tp->snd_una == tp->snd_max) { /* * Special case to match google, when * nothing is in flight the delivered * time does get updated to the current * time (see tcp_rate_bsd.c). */ bbr->r_ctl.rc_del_time = cts; } if (len >= maxseg) { idx = (len / maxseg) + 3; if (idx >= TCP_MSS_ACCT_ATIMER) counter_u64_add(bbr_out_size[(TCP_MSS_ACCT_ATIMER - 1)], 1); else counter_u64_add(bbr_out_size[idx], 1); } else { /* smaller than a MSS */ idx = len / (bbr_hptsi_bytes_min - bbr->rc_last_options); if (idx >= TCP_MSS_SMALL_MAX_SIZE_DIV) idx = (TCP_MSS_SMALL_MAX_SIZE_DIV - 1); counter_u64_add(bbr_out_size[(idx + TCP_MSS_SMALL_SIZE_OFF)], 1); } } } abandon = 0; /* * We must do the send accounting before we log the output, * otherwise the state of the rsm could change and we account to the * wrong bucket. */ if (len > 0) { bbr_do_send_accounting(tp, bbr, rsm, len, error); if (error == 0) { if (tp->snd_una == tp->snd_max) bbr->r_ctl.rc_tlp_rxt_last_time = cts; } } bbr_log_output(bbr, tp, &to, len, bbr_seq, (uint8_t) flags, error, cts, mb, &abandon, rsm, 0, sb); if (abandon) { /* * If bbr_log_output destroys the TCB or sees a TH_RST being * sent we should hit this condition. */ return (0); } if (bbr->rc_in_persist == 0) { /* * Advance snd_nxt over sequence space of this segment. */ if (error) /* We don't log or do anything with errors */ goto skip_upd; if (tp->snd_una == tp->snd_max && (len || (flags & (TH_SYN | TH_FIN)))) { /* * Update the time we just added data since none was * outstanding. */ bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__); bbr->rc_tp->t_acktime = ticks; } if (flags & (TH_SYN | TH_FIN) && (rsm == NULL)) { if (flags & TH_SYN) { /* * Smack the snd_max to iss + 1 * if its a FO we will add len below. */ tp->snd_max = tp->iss + 1; } if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) { tp->snd_max++; tp->t_flags |= TF_SENTFIN; } } if (sack_rxmit == 0) tp->snd_max += len; skip_upd: if ((error == 0) && len) tot_len += len; } else { /* Persists case */ int32_t xlen = len; if (error) goto nomore; if (flags & TH_SYN) ++xlen; if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) { ++xlen; tp->t_flags |= TF_SENTFIN; } if (xlen && (tp->snd_una == tp->snd_max)) { /* * Update the time we just added data since none was * outstanding. */ bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__); bbr->rc_tp->t_acktime = ticks; } if (sack_rxmit == 0) tp->snd_max += xlen; tot_len += (len + optlen + ipoptlen); } nomore: if (error) { /* * Failures do not advance the seq counter above. For the * case of ENOBUFS we will fall out and become ack-clocked. * capping the cwnd at the current flight. * Everything else will just have to retransmit with the timer * (no pacer). */ SOCKBUF_UNLOCK_ASSERT(sb); BBR_STAT_INC(bbr_saw_oerr); /* Clear all delay/early tracks */ bbr->r_ctl.rc_hptsi_agg_delay = 0; bbr->r_ctl.rc_agg_early = 0; bbr->r_agg_early_set = 0; bbr->output_error_seen = 1; if (bbr->oerror_cnt < 0xf) bbr->oerror_cnt++; if (bbr_max_net_error_cnt && (bbr->oerror_cnt >= bbr_max_net_error_cnt)) { /* drop the session */ tcp_set_inp_to_drop(inp, ENETDOWN); } switch (error) { case ENOBUFS: /* * Make this guy have to get ack's to send * more but lets make sure we don't * slam him below a T-O (1MSS). */ if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) - maxseg; if (tp->snd_cwnd < maxseg) tp->snd_cwnd = maxseg; } slot = (bbr_error_base_paceout + 1) << bbr->oerror_cnt; BBR_STAT_INC(bbr_saw_enobuf); if (bbr->bbr_hdrw_pacing) counter_u64_add(bbr_hdwr_pacing_enobuf, 1); else counter_u64_add(bbr_nohdwr_pacing_enobuf, 1); /* * Here even in the enobuf's case we want to do our * state update. The reason being we may have been * called by the input function. If so we have had * things change. */ error = 0; goto enobufs; case EMSGSIZE: /* * For some reason the interface we used initially * to send segments changed to another or lowered * its MTU. If TSO was active we either got an * interface without TSO capabilits or TSO was * turned off. If we obtained mtu from ip_output() * then update it and try again. */ /* Turn on tracing (or try to) */ { int old_maxseg; old_maxseg = tp->t_maxseg; BBR_STAT_INC(bbr_saw_emsgsiz); bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, csum_flags, tso, cts); if (mtu != 0) tcp_mss_update(tp, -1, mtu, NULL, NULL); if (old_maxseg <= tp->t_maxseg) { /* Huh it did not shrink? */ tp->t_maxseg = old_maxseg - 40; bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, 0, tso, cts); } /* * Nuke all other things that can interfere * with slot */ if ((tot_len + len) && (len >= tp->t_maxseg)) { slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, (tot_len + len), cts, 0); if (slot < bbr_error_base_paceout) slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt; } else slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt; bbr->rc_output_starts_timer = 1; bbr_start_hpts_timer(bbr, tp, cts, 10, slot, tot_len); return (error); } case EPERM: tp->t_softerror = error; /* Fall through */ case EHOSTDOWN: case EHOSTUNREACH: case ENETDOWN: case ENETUNREACH: if (TCPS_HAVERCVDSYN(tp->t_state)) { tp->t_softerror = error; } /* FALLTHROUGH */ default: slot = (bbr_error_base_paceout + 3) << bbr->oerror_cnt; bbr->rc_output_starts_timer = 1; bbr_start_hpts_timer(bbr, tp, cts, 11, slot, 0); return (error); } #ifdef STATS } else if (((tp->t_flags & TF_GPUTINPROG) == 0) && len && (rsm == NULL) && (bbr->rc_in_persist == 0)) { tp->gput_seq = bbr_seq; tp->gput_ack = bbr_seq + min(sbavail(&so->so_snd) - sb_offset, sendwin); tp->gput_ts = cts; tp->t_flags |= TF_GPUTINPROG; #endif } KMOD_TCPSTAT_INC(tcps_sndtotal); if ((bbr->bbr_hdw_pace_ena) && (bbr->bbr_attempt_hdwr_pace == 0) && (bbr->rc_past_init_win) && (bbr->rc_bbr_state != BBR_STATE_STARTUP) && (get_filter_value(&bbr->r_ctl.rc_delrate)) && (inp->inp_route.ro_nh && inp->inp_route.ro_nh->nh_ifp)) { /* * We are past the initial window and * have at least one measurement so we * could use hardware pacing if its available. * We have an interface and we have not attempted * to setup hardware pacing, lets try to now. */ uint64_t rate_wanted; int err = 0; rate_wanted = bbr_get_hardware_rate(bbr); bbr->bbr_attempt_hdwr_pace = 1; bbr->r_ctl.crte = tcp_set_pacing_rate(bbr->rc_tp, inp->inp_route.ro_nh->nh_ifp, rate_wanted, (RS_PACING_GEQ|RS_PACING_SUB_OK), &err, NULL); if (bbr->r_ctl.crte) { bbr_type_log_hdwr_pacing(bbr, bbr->r_ctl.crte->ptbl->rs_ifp, rate_wanted, bbr->r_ctl.crte->rate, __LINE__, cts, err); BBR_STAT_INC(bbr_hdwr_rl_add_ok); counter_u64_add(bbr_flows_nohdwr_pacing, -1); counter_u64_add(bbr_flows_whdwr_pacing, 1); bbr->bbr_hdrw_pacing = 1; /* Now what is our gain status? */ if (bbr->r_ctl.crte->rate < rate_wanted) { /* We have a problem */ bbr_setup_less_of_rate(bbr, cts, bbr->r_ctl.crte->rate, rate_wanted); } else { /* We are good */ bbr->gain_is_limited = 0; bbr->skip_gain = 0; } tcp_bbr_tso_size_check(bbr, cts); } else { bbr_type_log_hdwr_pacing(bbr, inp->inp_route.ro_nh->nh_ifp, rate_wanted, 0, __LINE__, cts, err); BBR_STAT_INC(bbr_hdwr_rl_add_fail); } } if (bbr->bbr_hdrw_pacing) { /* * Worry about cases where the route * changes or something happened that we * lost our hardware pacing possibly during * the last ip_output call. */ if (inp->inp_snd_tag == NULL) { /* A change during ip output disabled hw pacing? */ bbr->bbr_hdrw_pacing = 0; } else if ((inp->inp_route.ro_nh == NULL) || (inp->inp_route.ro_nh->nh_ifp != inp->inp_snd_tag->ifp)) { /* * We had an interface or route change, * detach from the current hdwr pacing * and setup to re-attempt next go * round. */ bbr->bbr_hdrw_pacing = 0; bbr->bbr_attempt_hdwr_pace = 0; tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp); tcp_bbr_tso_size_check(bbr, cts); } } /* * Data sent (as far as we can tell). If this advertises a larger * window than any other segment, then remember the size of the * advertised window. Any pending ACK has now been sent. */ if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) tp->rcv_adv = tp->rcv_nxt + recwin; tp->last_ack_sent = tp->rcv_nxt; if ((error == 0) && (bbr->r_ctl.rc_pace_max_segs > tp->t_maxseg) && (doing_tlp == 0) && (tso == 0) && (len > 0) && ((flags & TH_RST) == 0) && ((flags & TH_SYN) == 0) && (IN_RECOVERY(tp->t_flags) == 0) && (bbr->rc_in_persist == 0) && (tot_len < bbr->r_ctl.rc_pace_max_segs)) { /* * For non-tso we need to goto again until we have sent out * enough data to match what we are hptsi out every hptsi * interval. */ if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* Make sure snd_nxt is drug up */ tp->snd_nxt = tp->snd_max; } if (rsm != NULL) { rsm = NULL; goto skip_again; } rsm = NULL; sack_rxmit = 0; tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); goto again; } skip_again: if ((error == 0) && (flags & TH_FIN)) tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN); if ((error == 0) && (flags & TH_RST)) tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); if (((flags & (TH_RST | TH_SYN | TH_FIN)) == 0) && tot_len) { /* * Calculate/Re-Calculate the hptsi slot in usecs based on * what we have sent so far */ slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0); if (bbr->rc_no_pacing) slot = 0; } tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); enobufs: if (bbr->rc_use_google == 0) bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes))); bbr->rc_output_starts_timer = 1; if (bbr->bbr_use_rack_cheat && (more_to_rxt || ((bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts)) != NULL))) { /* Rack cheats and shotguns out all rxt's 1ms apart */ if (slot > 1000) slot = 1000; } if (bbr->bbr_hdrw_pacing && (bbr->hw_pacing_set == 0)) { /* * We don't change the tso size until some number of sends * to give the hardware commands time to get down * to the interface. */ bbr->r_ctl.bbr_hdwr_cnt_noset_snt++; if (bbr->r_ctl.bbr_hdwr_cnt_noset_snt >= bbr_hdwr_pacing_delay_cnt) { bbr->hw_pacing_set = 1; tcp_bbr_tso_size_check(bbr, cts); } } bbr_start_hpts_timer(bbr, tp, cts, 12, slot, tot_len); if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* Make sure snd_nxt is drug up */ tp->snd_nxt = tp->snd_max; } return (error); } /* * See bbr_output_wtime() for return values. */ static int bbr_output(struct tcpcb *tp) { int32_t ret; struct timeval tv; struct tcp_bbr *bbr; NET_EPOCH_ASSERT(); bbr = (struct tcp_bbr *)tp->t_fb_ptr; INP_WLOCK_ASSERT(tp->t_inpcb); (void)tcp_get_usecs(&tv); ret = bbr_output_wtime(tp, &tv); return (ret); } static void bbr_mtu_chg(struct tcpcb *tp) { struct tcp_bbr *bbr; struct bbr_sendmap *rsm, *frsm = NULL; uint32_t maxseg; /* * The MTU has changed. a) Clear the sack filter. b) Mark everything * over the current size as SACK_PASS so a retransmit will occur. */ bbr = (struct tcp_bbr *)tp->t_fb_ptr; maxseg = tp->t_maxseg - bbr->rc_last_options; sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { /* Don't mess with ones acked (by sack?) */ if (rsm->r_flags & BBR_ACKED) continue; if ((rsm->r_end - rsm->r_start) > maxseg) { /* * We mark sack-passed on all the previous large * sends we did. This will force them to retransmit. */ rsm->r_flags |= BBR_SACK_PASSED; if (((rsm->r_flags & BBR_MARKED_LOST) == 0) && bbr_is_lost(bbr, rsm, bbr->r_ctl.rc_rcvtime)) { bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; rsm->r_flags |= BBR_MARKED_LOST; } if (frsm == NULL) frsm = rsm; } } if (frsm) { bbr->r_ctl.rc_resend = frsm; } } /* * bbr_ctloutput() must drop the inpcb lock before performing copyin on * socket option arguments. When it re-acquires the lock after the copy, it * has to revalidate that the connection is still valid for the socket * option. */ static int bbr_set_sockopt(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_bbr *bbr) { struct epoch_tracker et; int32_t error = 0, optval; switch (sopt->sopt_name) { case TCP_RACK_PACE_MAX_SEG: case TCP_RACK_MIN_TO: case TCP_RACK_REORD_THRESH: case TCP_RACK_REORD_FADE: case TCP_RACK_TLP_THRESH: case TCP_RACK_PKT_DELAY: case TCP_BBR_ALGORITHM: case TCP_BBR_TSLIMITS: case TCP_BBR_IWINTSO: case TCP_BBR_RECFORCE: case TCP_BBR_STARTUP_PG: case TCP_BBR_DRAIN_PG: case TCP_BBR_RWND_IS_APP: case TCP_BBR_PROBE_RTT_INT: case TCP_BBR_PROBE_RTT_GAIN: case TCP_BBR_PROBE_RTT_LEN: case TCP_BBR_STARTUP_LOSS_EXIT: case TCP_BBR_USEDEL_RATE: case TCP_BBR_MIN_RTO: case TCP_BBR_MAX_RTO: case TCP_BBR_PACE_PER_SEC: case TCP_DELACK: case TCP_BBR_PACE_DEL_TAR: case TCP_BBR_SEND_IWND_IN_TSO: case TCP_BBR_EXTRA_STATE: case TCP_BBR_UTTER_MAX_TSO: case TCP_BBR_MIN_TOPACEOUT: case TCP_BBR_FLOOR_MIN_TSO: case TCP_BBR_TSTMP_RAISES: case TCP_BBR_POLICER_DETECT: case TCP_BBR_USE_RACK_CHEAT: case TCP_DATA_AFTER_CLOSE: case TCP_BBR_HDWR_PACE: case TCP_BBR_PACE_SEG_MAX: case TCP_BBR_PACE_SEG_MIN: case TCP_BBR_PACE_CROSS: case TCP_BBR_PACE_OH: #ifdef NETFLIX_PEAKRATE case TCP_MAXPEAKRATE: #endif case TCP_BBR_TMR_PACE_OH: case TCP_BBR_RACK_RTT_USE: case TCP_BBR_RETRAN_WTSO: break; default: return (tcp_default_ctloutput(so, sopt, inp, tp)); break; } INP_WUNLOCK(inp); error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); if (error) return (error); INP_WLOCK(inp); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { INP_WUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); bbr = (struct tcp_bbr *)tp->t_fb_ptr; switch (sopt->sopt_name) { case TCP_BBR_PACE_PER_SEC: BBR_OPTS_INC(tcp_bbr_pace_per_sec); bbr->r_ctl.bbr_hptsi_per_second = optval; break; case TCP_BBR_PACE_DEL_TAR: BBR_OPTS_INC(tcp_bbr_pace_del_tar); bbr->r_ctl.bbr_hptsi_segments_delay_tar = optval; break; case TCP_BBR_PACE_SEG_MAX: BBR_OPTS_INC(tcp_bbr_pace_seg_max); bbr->r_ctl.bbr_hptsi_segments_max = optval; break; case TCP_BBR_PACE_SEG_MIN: BBR_OPTS_INC(tcp_bbr_pace_seg_min); bbr->r_ctl.bbr_hptsi_bytes_min = optval; break; case TCP_BBR_PACE_CROSS: BBR_OPTS_INC(tcp_bbr_pace_cross); bbr->r_ctl.bbr_cross_over = optval; break; case TCP_BBR_ALGORITHM: BBR_OPTS_INC(tcp_bbr_algorithm); if (optval && (bbr->rc_use_google == 0)) { /* Turn on the google mode */ bbr_google_mode_on(bbr); if ((optval > 3) && (optval < 500)) { /* * Must be at least greater than .3% * and must be less than 50.0%. */ bbr->r_ctl.bbr_google_discount = optval; } } else if ((optval == 0) && (bbr->rc_use_google == 1)) { /* Turn off the google mode */ bbr_google_mode_off(bbr); } break; case TCP_BBR_TSLIMITS: BBR_OPTS_INC(tcp_bbr_tslimits); if (optval == 1) bbr->rc_use_ts_limit = 1; else if (optval == 0) bbr->rc_use_ts_limit = 0; else error = EINVAL; break; case TCP_BBR_IWINTSO: BBR_OPTS_INC(tcp_bbr_iwintso); if ((optval >= 0) && (optval < 128)) { uint32_t twin; bbr->rc_init_win = optval; twin = bbr_initial_cwnd(bbr, tp); if ((bbr->rc_past_init_win == 0) && (twin > tp->snd_cwnd)) tp->snd_cwnd = twin; else error = EBUSY; } else error = EINVAL; break; case TCP_BBR_STARTUP_PG: BBR_OPTS_INC(tcp_bbr_startup_pg); if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) { bbr->r_ctl.rc_startup_pg = optval; if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { bbr->r_ctl.rc_bbr_hptsi_gain = optval; } } else error = EINVAL; break; case TCP_BBR_DRAIN_PG: BBR_OPTS_INC(tcp_bbr_drain_pg); if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) bbr->r_ctl.rc_drain_pg = optval; else error = EINVAL; break; case TCP_BBR_PROBE_RTT_LEN: BBR_OPTS_INC(tcp_bbr_probertt_len); if (optval <= 1) reset_time_small(&bbr->r_ctl.rc_rttprop, (optval * USECS_IN_SECOND)); else error = EINVAL; break; case TCP_BBR_PROBE_RTT_GAIN: BBR_OPTS_INC(tcp_bbr_probertt_gain); if (optval <= BBR_UNIT) bbr->r_ctl.bbr_rttprobe_gain_val = optval; else error = EINVAL; break; case TCP_BBR_PROBE_RTT_INT: BBR_OPTS_INC(tcp_bbr_probe_rtt_int); if (optval > 1000) bbr->r_ctl.rc_probertt_int = optval; else error = EINVAL; break; case TCP_BBR_MIN_TOPACEOUT: BBR_OPTS_INC(tcp_bbr_topaceout); if (optval == 0) { bbr->no_pacing_until = 0; bbr->rc_no_pacing = 0; } else if (optval <= 0x00ff) { bbr->no_pacing_until = optval; if ((bbr->r_ctl.rc_pkt_epoch < bbr->no_pacing_until) && (bbr->rc_bbr_state == BBR_STATE_STARTUP)){ /* Turn on no pacing */ bbr->rc_no_pacing = 1; } } else error = EINVAL; break; case TCP_BBR_STARTUP_LOSS_EXIT: BBR_OPTS_INC(tcp_bbr_startup_loss_exit); bbr->rc_loss_exit = optval; break; case TCP_BBR_USEDEL_RATE: error = EINVAL; break; case TCP_BBR_MIN_RTO: BBR_OPTS_INC(tcp_bbr_min_rto); bbr->r_ctl.rc_min_rto_ms = optval; break; case TCP_BBR_MAX_RTO: BBR_OPTS_INC(tcp_bbr_max_rto); bbr->rc_max_rto_sec = optval; break; case TCP_RACK_MIN_TO: /* Minimum time between rack t-o's in ms */ BBR_OPTS_INC(tcp_rack_min_to); bbr->r_ctl.rc_min_to = optval; break; case TCP_RACK_REORD_THRESH: /* RACK reorder threshold (shift amount) */ BBR_OPTS_INC(tcp_rack_reord_thresh); if ((optval > 0) && (optval < 31)) bbr->r_ctl.rc_reorder_shift = optval; else error = EINVAL; break; case TCP_RACK_REORD_FADE: /* Does reordering fade after ms time */ BBR_OPTS_INC(tcp_rack_reord_fade); bbr->r_ctl.rc_reorder_fade = optval; break; case TCP_RACK_TLP_THRESH: /* RACK TLP theshold i.e. srtt+(srtt/N) */ BBR_OPTS_INC(tcp_rack_tlp_thresh); if (optval) bbr->rc_tlp_threshold = optval; else error = EINVAL; break; case TCP_BBR_USE_RACK_CHEAT: BBR_OPTS_INC(tcp_use_rackcheat); if (bbr->rc_use_google) { error = EINVAL; break; } BBR_OPTS_INC(tcp_rack_cheat); if (optval) bbr->bbr_use_rack_cheat = 1; else bbr->bbr_use_rack_cheat = 0; break; case TCP_BBR_FLOOR_MIN_TSO: BBR_OPTS_INC(tcp_utter_max_tso); if ((optval >= 0) && (optval < 40)) bbr->r_ctl.bbr_hptsi_segments_floor = optval; else error = EINVAL; break; case TCP_BBR_UTTER_MAX_TSO: BBR_OPTS_INC(tcp_utter_max_tso); if ((optval >= 0) && (optval < 0xffff)) bbr->r_ctl.bbr_utter_max = optval; else error = EINVAL; break; case TCP_BBR_EXTRA_STATE: BBR_OPTS_INC(tcp_extra_state); if (optval) bbr->rc_use_idle_restart = 1; else bbr->rc_use_idle_restart = 0; break; case TCP_BBR_SEND_IWND_IN_TSO: BBR_OPTS_INC(tcp_iwnd_tso); if (optval) { bbr->bbr_init_win_cheat = 1; if (bbr->rc_past_init_win == 0) { uint32_t cts; cts = tcp_get_usecs(&bbr->rc_tv); tcp_bbr_tso_size_check(bbr, cts); } } else bbr->bbr_init_win_cheat = 0; break; case TCP_BBR_HDWR_PACE: BBR_OPTS_INC(tcp_hdwr_pacing); if (optval){ bbr->bbr_hdw_pace_ena = 1; bbr->bbr_attempt_hdwr_pace = 0; } else { bbr->bbr_hdw_pace_ena = 0; #ifdef RATELIMIT if (bbr->bbr_hdrw_pacing) { bbr->bbr_hdrw_pacing = 0; in_pcbdetach_txrtlmt(bbr->rc_inp); } #endif } break; case TCP_DELACK: BBR_OPTS_INC(tcp_delack); if (optval < 100) { if (optval == 0) /* off */ tp->t_delayed_ack = 0; else if (optval == 1) /* on which is 2 */ tp->t_delayed_ack = 2; else /* higher than 2 and less than 100 */ tp->t_delayed_ack = optval; if (tp->t_flags & TF_DELACK) { tp->t_flags &= ~TF_DELACK; tp->t_flags |= TF_ACKNOW; NET_EPOCH_ENTER(et); bbr_output(tp); NET_EPOCH_EXIT(et); } } else error = EINVAL; break; case TCP_RACK_PKT_DELAY: /* RACK added ms i.e. rack-rtt + reord + N */ BBR_OPTS_INC(tcp_rack_pkt_delay); bbr->r_ctl.rc_pkt_delay = optval; break; #ifdef NETFLIX_PEAKRATE case TCP_MAXPEAKRATE: BBR_OPTS_INC(tcp_maxpeak); error = tcp_set_maxpeakrate(tp, optval); if (!error) tp->t_peakrate_thr = tp->t_maxpeakrate; break; #endif case TCP_BBR_RETRAN_WTSO: BBR_OPTS_INC(tcp_retran_wtso); if (optval) bbr->rc_resends_use_tso = 1; else bbr->rc_resends_use_tso = 0; break; case TCP_DATA_AFTER_CLOSE: BBR_OPTS_INC(tcp_data_ac); if (optval) bbr->rc_allow_data_af_clo = 1; else bbr->rc_allow_data_af_clo = 0; break; case TCP_BBR_POLICER_DETECT: BBR_OPTS_INC(tcp_policer_det); if (bbr->rc_use_google == 0) error = EINVAL; else if (optval) bbr->r_use_policer = 1; else bbr->r_use_policer = 0; break; case TCP_BBR_TSTMP_RAISES: BBR_OPTS_INC(tcp_ts_raises); if (optval) bbr->ts_can_raise = 1; else bbr->ts_can_raise = 0; break; case TCP_BBR_TMR_PACE_OH: BBR_OPTS_INC(tcp_pacing_oh_tmr); if (bbr->rc_use_google) { error = EINVAL; } else { if (optval) bbr->r_ctl.rc_incr_tmrs = 1; else bbr->r_ctl.rc_incr_tmrs = 0; } break; case TCP_BBR_PACE_OH: BBR_OPTS_INC(tcp_pacing_oh); if (bbr->rc_use_google) { error = EINVAL; } else { if (optval > (BBR_INCL_TCP_OH| BBR_INCL_IP_OH| BBR_INCL_ENET_OH)) { error = EINVAL; break; } if (optval & BBR_INCL_TCP_OH) bbr->r_ctl.rc_inc_tcp_oh = 1; else bbr->r_ctl.rc_inc_tcp_oh = 0; if (optval & BBR_INCL_IP_OH) bbr->r_ctl.rc_inc_ip_oh = 1; else bbr->r_ctl.rc_inc_ip_oh = 0; if (optval & BBR_INCL_ENET_OH) bbr->r_ctl.rc_inc_enet_oh = 1; else bbr->r_ctl.rc_inc_enet_oh = 0; } break; default: return (tcp_default_ctloutput(so, sopt, inp, tp)); break; } #ifdef NETFLIX_STATS tcp_log_socket_option(tp, sopt->sopt_name, optval, error); #endif INP_WUNLOCK(inp); return (error); } /* * return 0 on success, error-num on failure */ static int bbr_get_sockopt(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_bbr *bbr) { int32_t error, optval; /* * Because all our options are either boolean or an int, we can just * pull everything into optval and then unlock and copy. If we ever * add a option that is not a int, then this will have quite an * impact to this routine. */ switch (sopt->sopt_name) { case TCP_BBR_PACE_PER_SEC: optval = bbr->r_ctl.bbr_hptsi_per_second; break; case TCP_BBR_PACE_DEL_TAR: optval = bbr->r_ctl.bbr_hptsi_segments_delay_tar; break; case TCP_BBR_PACE_SEG_MAX: optval = bbr->r_ctl.bbr_hptsi_segments_max; break; case TCP_BBR_MIN_TOPACEOUT: optval = bbr->no_pacing_until; break; case TCP_BBR_PACE_SEG_MIN: optval = bbr->r_ctl.bbr_hptsi_bytes_min; break; case TCP_BBR_PACE_CROSS: optval = bbr->r_ctl.bbr_cross_over; break; case TCP_BBR_ALGORITHM: optval = bbr->rc_use_google; break; case TCP_BBR_TSLIMITS: optval = bbr->rc_use_ts_limit; break; case TCP_BBR_IWINTSO: optval = bbr->rc_init_win; break; case TCP_BBR_STARTUP_PG: optval = bbr->r_ctl.rc_startup_pg; break; case TCP_BBR_DRAIN_PG: optval = bbr->r_ctl.rc_drain_pg; break; case TCP_BBR_PROBE_RTT_INT: optval = bbr->r_ctl.rc_probertt_int; break; case TCP_BBR_PROBE_RTT_LEN: optval = (bbr->r_ctl.rc_rttprop.cur_time_limit / USECS_IN_SECOND); break; case TCP_BBR_PROBE_RTT_GAIN: optval = bbr->r_ctl.bbr_rttprobe_gain_val; break; case TCP_BBR_STARTUP_LOSS_EXIT: optval = bbr->rc_loss_exit; break; case TCP_BBR_USEDEL_RATE: error = EINVAL; break; case TCP_BBR_MIN_RTO: optval = bbr->r_ctl.rc_min_rto_ms; break; case TCP_BBR_MAX_RTO: optval = bbr->rc_max_rto_sec; break; case TCP_RACK_PACE_MAX_SEG: /* Max segments in a pace */ optval = bbr->r_ctl.rc_pace_max_segs; break; case TCP_RACK_MIN_TO: /* Minimum time between rack t-o's in ms */ optval = bbr->r_ctl.rc_min_to; break; case TCP_RACK_REORD_THRESH: /* RACK reorder threshold (shift amount) */ optval = bbr->r_ctl.rc_reorder_shift; break; case TCP_RACK_REORD_FADE: /* Does reordering fade after ms time */ optval = bbr->r_ctl.rc_reorder_fade; break; case TCP_BBR_USE_RACK_CHEAT: /* Do we use the rack cheat for rxt */ optval = bbr->bbr_use_rack_cheat; break; case TCP_BBR_FLOOR_MIN_TSO: optval = bbr->r_ctl.bbr_hptsi_segments_floor; break; case TCP_BBR_UTTER_MAX_TSO: optval = bbr->r_ctl.bbr_utter_max; break; case TCP_BBR_SEND_IWND_IN_TSO: /* Do we send TSO size segments initially */ optval = bbr->bbr_init_win_cheat; break; case TCP_BBR_EXTRA_STATE: optval = bbr->rc_use_idle_restart; break; case TCP_RACK_TLP_THRESH: /* RACK TLP theshold i.e. srtt+(srtt/N) */ optval = bbr->rc_tlp_threshold; break; case TCP_RACK_PKT_DELAY: /* RACK added ms i.e. rack-rtt + reord + N */ optval = bbr->r_ctl.rc_pkt_delay; break; case TCP_BBR_RETRAN_WTSO: optval = bbr->rc_resends_use_tso; break; case TCP_DATA_AFTER_CLOSE: optval = bbr->rc_allow_data_af_clo; break; case TCP_DELACK: optval = tp->t_delayed_ack; break; case TCP_BBR_HDWR_PACE: optval = bbr->bbr_hdw_pace_ena; break; case TCP_BBR_POLICER_DETECT: optval = bbr->r_use_policer; break; case TCP_BBR_TSTMP_RAISES: optval = bbr->ts_can_raise; break; case TCP_BBR_TMR_PACE_OH: optval = bbr->r_ctl.rc_incr_tmrs; break; case TCP_BBR_PACE_OH: optval = 0; if (bbr->r_ctl.rc_inc_tcp_oh) optval |= BBR_INCL_TCP_OH; if (bbr->r_ctl.rc_inc_ip_oh) optval |= BBR_INCL_IP_OH; if (bbr->r_ctl.rc_inc_enet_oh) optval |= BBR_INCL_ENET_OH; break; default: return (tcp_default_ctloutput(so, sopt, inp, tp)); break; } INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof optval); return (error); } /* * return 0 on success, error-num on failure */ static int bbr_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp) { int32_t error = EINVAL; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; if (bbr == NULL) { /* Huh? */ goto out; } if (sopt->sopt_dir == SOPT_SET) { return (bbr_set_sockopt(so, sopt, inp, tp, bbr)); } else if (sopt->sopt_dir == SOPT_GET) { return (bbr_get_sockopt(so, sopt, inp, tp, bbr)); } out: INP_WUNLOCK(inp); return (error); } static int bbr_pru_options(struct tcpcb *tp, int flags) { if (flags & PRUS_OOB) return (EOPNOTSUPP); return (0); } struct tcp_function_block __tcp_bbr = { .tfb_tcp_block_name = __XSTRING(STACKNAME), .tfb_tcp_output = bbr_output, .tfb_do_queued_segments = ctf_do_queued_segments, .tfb_do_segment_nounlock = bbr_do_segment_nounlock, .tfb_tcp_do_segment = bbr_do_segment, .tfb_tcp_ctloutput = bbr_ctloutput, .tfb_tcp_fb_init = bbr_init, .tfb_tcp_fb_fini = bbr_fini, .tfb_tcp_timer_stop_all = bbr_stopall, .tfb_tcp_timer_activate = bbr_timer_activate, .tfb_tcp_timer_active = bbr_timer_active, .tfb_tcp_timer_stop = bbr_timer_stop, .tfb_tcp_rexmit_tmr = bbr_remxt_tmr, .tfb_tcp_handoff_ok = bbr_handoff_ok, .tfb_tcp_mtu_chg = bbr_mtu_chg, .tfb_pru_options = bbr_pru_options, }; static const char *bbr_stack_names[] = { __XSTRING(STACKNAME), #ifdef STACKALIAS __XSTRING(STACKALIAS), #endif }; static bool bbr_mod_inited = false; static int tcp_addbbr(module_t mod, int32_t type, void *data) { int32_t err = 0; int num_stacks; switch (type) { case MOD_LOAD: printf("Attempting to load " __XSTRING(MODNAME) "\n"); bbr_zone = uma_zcreate(__XSTRING(MODNAME) "_map", sizeof(struct bbr_sendmap), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); bbr_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb", sizeof(struct tcp_bbr), NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); sysctl_ctx_init(&bbr_sysctl_ctx); bbr_sysctl_root = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_STATIC_CHILDREN(_net_inet_tcp), OID_AUTO, #ifdef STACKALIAS __XSTRING(STACKALIAS), #else __XSTRING(STACKNAME), #endif CTLFLAG_RW | CTLFLAG_MPSAFE, 0, ""); if (bbr_sysctl_root == NULL) { printf("Failed to add sysctl node\n"); err = EFAULT; goto free_uma; } bbr_init_sysctls(); num_stacks = nitems(bbr_stack_names); err = register_tcp_functions_as_names(&__tcp_bbr, M_WAITOK, bbr_stack_names, &num_stacks); if (err) { printf("Failed to register %s stack name for " "%s module\n", bbr_stack_names[num_stacks], __XSTRING(MODNAME)); sysctl_ctx_free(&bbr_sysctl_ctx); free_uma: uma_zdestroy(bbr_zone); uma_zdestroy(bbr_pcb_zone); bbr_counter_destroy(); printf("Failed to register " __XSTRING(MODNAME) " module err:%d\n", err); return (err); } tcp_lro_reg_mbufq(); bbr_mod_inited = true; printf(__XSTRING(MODNAME) " is now available\n"); break; case MOD_QUIESCE: err = deregister_tcp_functions(&__tcp_bbr, true, false); break; case MOD_UNLOAD: err = deregister_tcp_functions(&__tcp_bbr, false, true); if (err == EBUSY) break; if (bbr_mod_inited) { uma_zdestroy(bbr_zone); uma_zdestroy(bbr_pcb_zone); sysctl_ctx_free(&bbr_sysctl_ctx); bbr_counter_destroy(); printf(__XSTRING(MODNAME) " is now no longer available\n"); bbr_mod_inited = false; } tcp_lro_dereg_mbufq(); err = 0; break; default: return (EOPNOTSUPP); } return (err); } static moduledata_t tcp_bbr = { .name = __XSTRING(MODNAME), .evhand = tcp_addbbr, .priv = 0 }; MODULE_VERSION(MODNAME, 1); DECLARE_MODULE(MODNAME, tcp_bbr, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1); diff --git a/sys/netinet/tcp_stacks/rack.c b/sys/netinet/tcp_stacks/rack.c index 0079bf8b6400..d2093e1afab7 100644 --- a/sys/netinet/tcp_stacks/rack.c +++ b/sys/netinet/tcp_stacks/rack.c @@ -1,14981 +1,14967 @@ /*- * Copyright (c) 2016-2020 Netflix, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_tcpdebug.h" #include "opt_ratelimit.h" #include #include #include #include #ifdef TCP_HHOOK #include #endif #include #include #include #include #include #include /* for proc0 declaration */ #include #include #include #include #ifdef STATS #include #include #include /* Must come after qmath.h and tree.h */ #else #include #endif #include #include #include #include #include #include #include #include #include #include #include #define TCPSTATES /* for logging */ #include #include #include #include #include /* required for icmp_var.h */ #include /* for ICMP_BANDLIM */ #include #include #include #include #include #define TCPOUTFLAGS #include #include #include #include #include #include #include #include #include #include #include #ifdef NETFLIX_SHARED_CWND #include #endif #ifdef TCPDEBUG #include #endif /* TCPDEBUG */ #ifdef TCP_OFFLOAD #include #endif #ifdef INET6 #include #endif #include #if defined(IPSEC) || defined(IPSEC_SUPPORT) #include #include #endif /* IPSEC */ #include #include #include #ifdef MAC #include #endif #include "sack_filter.h" #include "tcp_rack.h" #include "rack_bbr_common.h" uma_zone_t rack_zone; uma_zone_t rack_pcb_zone; #ifndef TICKS2SBT #define TICKS2SBT(__t) (tick_sbt * ((sbintime_t)(__t))) #endif struct sysctl_ctx_list rack_sysctl_ctx; struct sysctl_oid *rack_sysctl_root; #define CUM_ACKED 1 #define SACKED 2 /* * The RACK module incorporates a number of * TCP ideas that have been put out into the IETF * over the last few years: * - Matt Mathis's Rate Halving which slowly drops * the congestion window so that the ack clock can * be maintained during a recovery. * - Yuchung Cheng's RACK TCP (for which its named) that * will stop us using the number of dup acks and instead * use time as the gage of when we retransmit. * - Reorder Detection of RFC4737 and the Tail-Loss probe draft * of Dukkipati et.al. * RACK depends on SACK, so if an endpoint arrives that * cannot do SACK the state machine below will shuttle the * connection back to using the "default" TCP stack that is * in FreeBSD. * * To implement RACK the original TCP stack was first decomposed * into a functional state machine with individual states * for each of the possible TCP connection states. The do_segement * functions role in life is to mandate the connection supports SACK * initially and then assure that the RACK state matches the conenction * state before calling the states do_segment function. Each * state is simplified due to the fact that the original do_segment * has been decomposed and we *know* what state we are in (no * switches on the state) and all tests for SACK are gone. This * greatly simplifies what each state does. * * TCP output is also over-written with a new version since it * must maintain the new rack scoreboard. * */ static int32_t rack_tlp_thresh = 1; static int32_t rack_tlp_limit = 2; /* No more than 2 TLPs w-out new data */ static int32_t rack_tlp_use_greater = 1; static int32_t rack_reorder_thresh = 2; static int32_t rack_reorder_fade = 60000; /* 0 - never fade, def 60,000 * - 60 seconds */ /* Attack threshold detections */ static uint32_t rack_highest_sack_thresh_seen = 0; static uint32_t rack_highest_move_thresh_seen = 0; static int32_t rack_pkt_delay = 1; static int32_t rack_early_recovery = 1; static int32_t rack_send_a_lot_in_prr = 1; static int32_t rack_min_to = 1; /* Number of ms minimum timeout */ static int32_t rack_verbose_logging = 0; static int32_t rack_ignore_data_after_close = 1; static int32_t rack_enable_shared_cwnd = 0; static int32_t rack_limits_scwnd = 1; static int32_t rack_enable_mqueue_for_nonpaced = 0; static int32_t rack_disable_prr = 0; static int32_t use_rack_rr = 1; static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */ static int32_t rack_persist_min = 250; /* 250ms */ static int32_t rack_persist_max = 2000; /* 2 Second */ static int32_t rack_sack_not_required = 0; /* set to one to allow non-sack to use rack */ static int32_t rack_default_init_window = 0; /* Use system default */ static int32_t rack_limit_time_with_srtt = 0; static int32_t rack_hw_pace_adjust = 0; /* * Currently regular tcp has a rto_min of 30ms * the backoff goes 12 times so that ends up * being a total of 122.850 seconds before a * connection is killed. */ static uint32_t rack_def_data_window = 20; static uint32_t rack_goal_bdp = 2; static uint32_t rack_min_srtts = 1; static uint32_t rack_min_measure_usec = 0; static int32_t rack_tlp_min = 10; static int32_t rack_rto_min = 30; /* 30ms same as main freebsd */ static int32_t rack_rto_max = 4000; /* 4 seconds */ static const int32_t rack_free_cache = 2; static int32_t rack_hptsi_segments = 40; static int32_t rack_rate_sample_method = USE_RTT_LOW; static int32_t rack_pace_every_seg = 0; static int32_t rack_delayed_ack_time = 200; /* 200ms */ static int32_t rack_slot_reduction = 4; static int32_t rack_wma_divisor = 8; /* For WMA calculation */ static int32_t rack_cwnd_block_ends_measure = 0; static int32_t rack_rwnd_block_ends_measure = 0; static int32_t rack_lower_cwnd_at_tlp = 0; static int32_t rack_use_proportional_reduce = 0; static int32_t rack_proportional_rate = 10; static int32_t rack_tlp_max_resend = 2; static int32_t rack_limited_retran = 0; static int32_t rack_always_send_oldest = 0; static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE; static uint16_t rack_per_of_gp_ss = 250; /* 250 % slow-start */ static uint16_t rack_per_of_gp_ca = 200; /* 200 % congestion-avoidance */ static uint16_t rack_per_of_gp_rec = 200; /* 200 % of bw */ /* Probertt */ static uint16_t rack_per_of_gp_probertt = 60; /* 60% of bw */ static uint16_t rack_per_of_gp_lowthresh = 40; /* 40% is bottom */ static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */ static uint16_t rack_atexit_prtt_hbp = 130; /* Clamp to 130% on exit prtt if highly buffered path */ static uint16_t rack_atexit_prtt = 130; /* Clamp to 100% on exit prtt if non highly buffered path */ static uint32_t rack_max_drain_wait = 2; /* How man gp srtt's before we give up draining */ static uint32_t rack_must_drain = 1; /* How many GP srtt's we *must* wait */ static uint32_t rack_probertt_use_min_rtt_entry = 1; /* Use the min to calculate the goal else gp_srtt */ static uint32_t rack_probertt_use_min_rtt_exit = 0; static uint32_t rack_probe_rtt_sets_cwnd = 0; static uint32_t rack_probe_rtt_safety_val = 2000000; /* No more than 2 sec in probe-rtt */ static uint32_t rack_time_between_probertt = 9600000; /* 9.6 sec in us */ static uint32_t rack_probertt_gpsrtt_cnt_mul = 0; /* How many srtt periods does probe-rtt last top fraction */ static uint32_t rack_probertt_gpsrtt_cnt_div = 0; /* How many srtt periods does probe-rtt last bottom fraction */ static uint32_t rack_min_probertt_hold = 200000; /* Equal to delayed ack time */ static uint32_t rack_probertt_filter_life = 10000000; static uint32_t rack_probertt_lower_within = 10; static uint32_t rack_min_rtt_movement = 250; /* Must move at least 250 useconds to count as a lowering */ static int32_t rack_pace_one_seg = 0; /* Shall we pace for less than 1.4Meg 1MSS at a time */ static int32_t rack_probertt_clear_is = 1; static int32_t rack_max_drain_hbp = 1; /* Extra drain times gpsrtt for highly buffered paths */ static int32_t rack_hbp_thresh = 3; /* what is the divisor max_rtt/min_rtt to decided a hbp */ /* Part of pacing */ static int32_t rack_max_per_above = 30; /* When we go to increment stop if above 100+this% */ /* Timely information */ /* Combine these two gives the range of 'no change' to bw */ /* ie the up/down provide the upper and lower bound */ static int32_t rack_gp_per_bw_mul_up = 2; /* 2% */ static int32_t rack_gp_per_bw_mul_down = 4; /* 4% */ static int32_t rack_gp_rtt_maxmul = 3; /* 3 x maxmin */ static int32_t rack_gp_rtt_minmul = 1; /* minrtt + (minrtt/mindiv) is lower rtt */ static int32_t rack_gp_rtt_mindiv = 4; /* minrtt + (minrtt * minmul/mindiv) is lower rtt */ static int32_t rack_gp_decrease_per = 20; /* 20% decrease in multipler */ static int32_t rack_gp_increase_per = 2; /* 2% increase in multipler */ static int32_t rack_per_lower_bound = 50; /* Don't allow to drop below this multiplier */ static int32_t rack_per_upper_bound_ss = 0; /* Don't allow SS to grow above this */ static int32_t rack_per_upper_bound_ca = 0; /* Don't allow CA to grow above this */ static int32_t rack_do_dyn_mul = 0; /* Are the rack gp multipliers dynamic */ static int32_t rack_gp_no_rec_chg = 1; /* Prohibit recovery from reducing it's multiplier */ static int32_t rack_timely_dec_clear = 6; /* Do we clear decrement count at a value (6)? */ static int32_t rack_timely_max_push_rise = 3; /* One round of pushing */ static int32_t rack_timely_max_push_drop = 3; /* Three round of pushing */ static int32_t rack_timely_min_segs = 4; /* 4 segment minimum */ static int32_t rack_use_max_for_nobackoff = 0; static int32_t rack_timely_int_timely_only = 0; /* do interim timely's only use the timely algo (no b/w changes)? */ static int32_t rack_timely_no_stopping = 0; static int32_t rack_down_raise_thresh = 100; static int32_t rack_req_segs = 1; /* Weird delayed ack mode */ static int32_t rack_use_imac_dack = 0; /* Rack specific counters */ counter_u64_t rack_badfr; counter_u64_t rack_badfr_bytes; counter_u64_t rack_rtm_prr_retran; counter_u64_t rack_rtm_prr_newdata; counter_u64_t rack_timestamp_mismatch; counter_u64_t rack_reorder_seen; counter_u64_t rack_paced_segments; counter_u64_t rack_unpaced_segments; counter_u64_t rack_calc_zero; counter_u64_t rack_calc_nonzero; counter_u64_t rack_saw_enobuf; counter_u64_t rack_saw_enetunreach; counter_u64_t rack_per_timer_hole; /* Tail loss probe counters */ counter_u64_t rack_tlp_tot; counter_u64_t rack_tlp_newdata; counter_u64_t rack_tlp_retran; counter_u64_t rack_tlp_retran_bytes; counter_u64_t rack_tlp_retran_fail; counter_u64_t rack_to_tot; counter_u64_t rack_to_arm_rack; counter_u64_t rack_to_arm_tlp; counter_u64_t rack_to_alloc; counter_u64_t rack_to_alloc_hard; counter_u64_t rack_to_alloc_emerg; counter_u64_t rack_to_alloc_limited; counter_u64_t rack_alloc_limited_conns; counter_u64_t rack_split_limited; counter_u64_t rack_sack_proc_all; counter_u64_t rack_sack_proc_short; counter_u64_t rack_sack_proc_restart; counter_u64_t rack_sack_attacks_detected; counter_u64_t rack_sack_attacks_reversed; counter_u64_t rack_sack_used_next_merge; counter_u64_t rack_sack_splits; counter_u64_t rack_sack_used_prev_merge; counter_u64_t rack_sack_skipped_acked; counter_u64_t rack_ack_total; counter_u64_t rack_express_sack; counter_u64_t rack_sack_total; counter_u64_t rack_move_none; counter_u64_t rack_move_some; counter_u64_t rack_used_tlpmethod; counter_u64_t rack_used_tlpmethod2; counter_u64_t rack_enter_tlp_calc; counter_u64_t rack_input_idle_reduces; counter_u64_t rack_collapsed_win; counter_u64_t rack_tlp_does_nada; counter_u64_t rack_try_scwnd; /* Temp CPU counters */ counter_u64_t rack_find_high; counter_u64_t rack_progress_drops; counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE]; counter_u64_t rack_opts_arry[RACK_OPTS_SIZE]; static void rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line); static int rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val); static int rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); static void rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, struct tcphdr *th, uint16_t nsegs, uint16_t type, int32_t recovery); static struct rack_sendmap *rack_alloc(struct tcp_rack *rack); static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type); static struct rack_sendmap * rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused); static void rack_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type); static void rack_counter_destroy(void); static int rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp); static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how); static void rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line); static void rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos); static void rack_dtor(void *mem, int32_t size, void *arg); static void rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm, uint32_t t, uint32_t cts); static void rack_log_alt_to_to_cancel(struct tcp_rack *rack, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint32_t flex4, uint32_t flex5, uint32_t flex6, uint16_t flex7, uint8_t mod); static void rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line, struct rack_sendmap *rsm); static struct rack_sendmap * rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm); static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack); static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm); static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged); static int rack_get_sockopt(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack); static void rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, int line); static uint32_t rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss); static int32_t rack_handoff_ok(struct tcpcb *tp); static int32_t rack_init(struct tcpcb *tp); static void rack_init_sysctls(void); static void rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th); static void rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts, uint8_t pass, struct rack_sendmap *hintrsm, uint32_t us_cts); static void rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm); static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm); static int32_t rack_output(struct tcpcb *tp); static uint32_t rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two); static void rack_post_recovery(struct tcpcb *tp, struct tcphdr *th); static void rack_remxt_tmr(struct tcpcb *tp); static int rack_set_sockopt(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack); static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack); static int32_t rack_stopall(struct tcpcb *tp); static void rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta); static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type); static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line); static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type); static uint32_t rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t ts, int32_t * lenp); static void rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t ts); static int rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack); static int32_t tcp_addrack(module_t mod, int32_t type, void *data); static int rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos); static int rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); struct rack_sendmap * tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused); static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt); static void tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th); int32_t rack_clear_counter=0; static int sysctl_rack_clear(SYSCTL_HANDLER_ARGS) { uint32_t stat; int32_t error; error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t)); if (error || req->newptr == NULL) return error; error = SYSCTL_IN(req, &stat, sizeof(uint32_t)); if (error) return (error); if (stat == 1) { #ifdef INVARIANTS printf("Clearing RACK counters\n"); #endif counter_u64_zero(rack_badfr); counter_u64_zero(rack_badfr_bytes); counter_u64_zero(rack_rtm_prr_retran); counter_u64_zero(rack_rtm_prr_newdata); counter_u64_zero(rack_timestamp_mismatch); counter_u64_zero(rack_reorder_seen); counter_u64_zero(rack_tlp_tot); counter_u64_zero(rack_tlp_newdata); counter_u64_zero(rack_tlp_retran); counter_u64_zero(rack_tlp_retran_bytes); counter_u64_zero(rack_tlp_retran_fail); counter_u64_zero(rack_to_tot); counter_u64_zero(rack_to_arm_rack); counter_u64_zero(rack_to_arm_tlp); counter_u64_zero(rack_paced_segments); counter_u64_zero(rack_calc_zero); counter_u64_zero(rack_calc_nonzero); counter_u64_zero(rack_unpaced_segments); counter_u64_zero(rack_saw_enobuf); counter_u64_zero(rack_saw_enetunreach); counter_u64_zero(rack_per_timer_hole); counter_u64_zero(rack_to_alloc_hard); counter_u64_zero(rack_to_alloc_emerg); counter_u64_zero(rack_sack_proc_all); counter_u64_zero(rack_sack_proc_short); counter_u64_zero(rack_sack_proc_restart); counter_u64_zero(rack_to_alloc); counter_u64_zero(rack_to_alloc_limited); counter_u64_zero(rack_alloc_limited_conns); counter_u64_zero(rack_split_limited); counter_u64_zero(rack_find_high); counter_u64_zero(rack_sack_attacks_detected); counter_u64_zero(rack_sack_attacks_reversed); counter_u64_zero(rack_sack_used_next_merge); counter_u64_zero(rack_sack_used_prev_merge); counter_u64_zero(rack_sack_splits); counter_u64_zero(rack_sack_skipped_acked); counter_u64_zero(rack_ack_total); counter_u64_zero(rack_express_sack); counter_u64_zero(rack_sack_total); counter_u64_zero(rack_move_none); counter_u64_zero(rack_move_some); counter_u64_zero(rack_used_tlpmethod); counter_u64_zero(rack_used_tlpmethod2); counter_u64_zero(rack_enter_tlp_calc); counter_u64_zero(rack_progress_drops); counter_u64_zero(rack_tlp_does_nada); counter_u64_zero(rack_try_scwnd); counter_u64_zero(rack_collapsed_win); } rack_clear_counter = 0; return (0); } static void rack_init_sysctls(void) { struct sysctl_oid *rack_counters; struct sysctl_oid *rack_attack; struct sysctl_oid *rack_pacing; struct sysctl_oid *rack_timely; struct sysctl_oid *rack_timers; struct sysctl_oid *rack_tlp; struct sysctl_oid *rack_misc; struct sysctl_oid *rack_measure; struct sysctl_oid *rack_probertt; rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "sack_attack", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Rack Sack Attack Counters and Controls"); rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "stats", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Rack Counters"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "rate_sample_method", CTLFLAG_RW, &rack_rate_sample_method , USE_RTT_LOW, "What method should we use for rate sampling 0=high, 1=low "); /* Probe rtt related controls */ rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "probertt", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "ProbeRTT related Controls"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "exit_per_hpb", CTLFLAG_RW, &rack_atexit_prtt_hbp, 130, "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW, &rack_atexit_prtt, 130, "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "gp_per_mul", CTLFLAG_RW, &rack_per_of_gp_probertt, 60, "What percentage of goodput do we pace at in probertt"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "gp_per_reduce", CTLFLAG_RW, &rack_per_of_gp_probertt_reduce, 10, "What percentage of goodput do we reduce every gp_srtt"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "gp_per_low", CTLFLAG_RW, &rack_per_of_gp_lowthresh, 40, "What percentage of goodput do we allow the multiplier to fall to"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "time_between", CTLFLAG_RW, & rack_time_between_probertt, 96000000, "How many useconds between the lowest rtt falling must past before we enter probertt"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "safety", CTLFLAG_RW, &rack_probe_rtt_safety_val, 2000000, "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "sets_cwnd", CTLFLAG_RW, &rack_probe_rtt_sets_cwnd, 0, "Do we set the cwnd too (if always_lower is on)"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "maxdrainsrtts", CTLFLAG_RW, &rack_max_drain_wait, 2, "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "mustdrainsrtts", CTLFLAG_RW, &rack_must_drain, 1, "We must drain this many gp_srtt's waiting for flight to reach goal"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "goal_use_min_entry", CTLFLAG_RW, &rack_probertt_use_min_rtt_entry, 1, "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "goal_use_min_exit", CTLFLAG_RW, &rack_probertt_use_min_rtt_exit, 0, "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "length_div", CTLFLAG_RW, &rack_probertt_gpsrtt_cnt_div, 0, "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "length_mul", CTLFLAG_RW, &rack_probertt_gpsrtt_cnt_mul, 0, "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "holdtim_at_target", CTLFLAG_RW, &rack_min_probertt_hold, 200000, "What is the minimum time we hold probertt at target"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "filter_life", CTLFLAG_RW, &rack_probertt_filter_life, 10000000, "What is the time for the filters life in useconds"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "lower_within", CTLFLAG_RW, &rack_probertt_lower_within, 10, "If the rtt goes lower within this percentage of the time, go into probe-rtt"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "must_move", CTLFLAG_RW, &rack_min_rtt_movement, 250, "How much is the minimum movement in rtt to count as a drop for probertt purposes"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "clear_is_cnts", CTLFLAG_RW, &rack_probertt_clear_is, 1, "Do we clear I/S counts on exiting probe-rtt"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "hbp_extra_drain", CTLFLAG_RW, &rack_max_drain_hbp, 1, "How many extra drain gpsrtt's do we get in highly buffered paths"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "hbp_threshold", CTLFLAG_RW, &rack_hbp_thresh, 3, "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold"); /* Pacing related sysctls */ rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "pacing", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Pacing related Controls"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "max_pace_over", CTLFLAG_RW, &rack_max_per_above, 30, "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "pace_to_one", CTLFLAG_RW, &rack_pace_one_seg, 0, "Do we allow low b/w pacing of 1MSS instead of two"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "limit_wsrtt", CTLFLAG_RW, &rack_limit_time_with_srtt, 0, "Do we limit pacing time based on srtt"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "init_win", CTLFLAG_RW, &rack_default_init_window, 0, "Do we have a rack initial window 0 = system default"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "hw_pacing_adjust", CTLFLAG_RW, &rack_hw_pace_adjust, 0, "What percentage do we raise the MSS by (11 = 1.1%)"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "gp_per_ss", CTLFLAG_RW, &rack_per_of_gp_ss, 250, "If non zero, what percentage of goodput to pace at in slow start"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "gp_per_ca", CTLFLAG_RW, &rack_per_of_gp_ca, 150, "If non zero, what percentage of goodput to pace at in congestion avoidance"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "gp_per_rec", CTLFLAG_RW, &rack_per_of_gp_rec, 200, "If non zero, what percentage of goodput to pace at in recovery"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "pace_max_seg", CTLFLAG_RW, &rack_hptsi_segments, 40, "What size is the max for TSO segments in pacing and burst mitigation"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "burst_reduces", CTLFLAG_RW, &rack_slot_reduction, 4, "When doing only burst mitigation what is the reduce divisor"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "use_pacing", CTLFLAG_RW, &rack_pace_every_seg, 0, "If set we use pacing, if clear we use only the original burst mitigation"); rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "timely", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Rack Timely RTT Controls"); /* Timely based GP dynmics */ SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "upper", CTLFLAG_RW, &rack_gp_per_bw_mul_up, 2, "Rack timely upper range for equal b/w (in percentage)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "lower", CTLFLAG_RW, &rack_gp_per_bw_mul_down, 4, "Rack timely lower range for equal b/w (in percentage)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "rtt_max_mul", CTLFLAG_RW, &rack_gp_rtt_maxmul, 3, "Rack timely multipler of lowest rtt for rtt_max"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "rtt_min_div", CTLFLAG_RW, &rack_gp_rtt_mindiv, 4, "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "rtt_min_mul", CTLFLAG_RW, &rack_gp_rtt_minmul, 1, "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "decrease", CTLFLAG_RW, &rack_gp_decrease_per, 20, "Rack timely decrease percentage of our GP multiplication factor"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "increase", CTLFLAG_RW, &rack_gp_increase_per, 2, "Rack timely increase perentage of our GP multiplication factor"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "lowerbound", CTLFLAG_RW, &rack_per_lower_bound, 50, "Rack timely lowest percentage we allow GP multiplier to fall to"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "upperboundss", CTLFLAG_RW, &rack_per_upper_bound_ss, 0, "Rack timely higest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "upperboundca", CTLFLAG_RW, &rack_per_upper_bound_ca, 0, "Rack timely higest percentage we allow GP multiplier to CA raise to (0 is no upperbound)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "dynamicgp", CTLFLAG_RW, &rack_do_dyn_mul, 0, "Rack timely do we enable dynmaic timely goodput by default"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "no_rec_red", CTLFLAG_RW, &rack_gp_no_rec_chg, 1, "Rack timely do we prohibit the recovery multiplier from being lowered"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "red_clear_cnt", CTLFLAG_RW, &rack_timely_dec_clear, 6, "Rack timely what threshold do we count to before another boost during b/w decent"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "max_push_rise", CTLFLAG_RW, &rack_timely_max_push_rise, 3, "Rack timely how many times do we push up with b/w increase"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "max_push_drop", CTLFLAG_RW, &rack_timely_max_push_drop, 3, "Rack timely how many times do we push back on b/w decent"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "min_segs", CTLFLAG_RW, &rack_timely_min_segs, 4, "Rack timely when setting the cwnd what is the min num segments"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "noback_max", CTLFLAG_RW, &rack_use_max_for_nobackoff, 0, "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "interim_timely_only", CTLFLAG_RW, &rack_timely_int_timely_only, 0, "Rack timely when doing interim timely's do we only do timely (no b/w consideration)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "nonstop", CTLFLAG_RW, &rack_timely_no_stopping, 0, "Rack timely don't stop increase"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "dec_raise_thresh", CTLFLAG_RW, &rack_down_raise_thresh, 100, "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "bottom_drag_segs", CTLFLAG_RW, &rack_req_segs, 1, "Bottom dragging if not these many segments outstanding and room"); /* TLP and Rack related parameters */ rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "tlp", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "TLP and Rack related Controls"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "use_rrr", CTLFLAG_RW, &use_rack_rr, 1, "Do we use Rack Rapid Recovery"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW, &rack_non_rxt_use_cr, 0, "Do we use ss/ca rate if in recovery we are transmitting a new data chunk"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "tlpmethod", CTLFLAG_RW, &rack_tlp_threshold_use, TLP_USE_TWO_ONE, "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "limit", CTLFLAG_RW, &rack_tlp_limit, 2, "How many TLP's can be sent without sending new data"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "use_greater", CTLFLAG_RW, &rack_tlp_use_greater, 1, "Should we use the rack_rtt time if its greater than srtt"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "tlpminto", CTLFLAG_RW, &rack_tlp_min, 10, "TLP minimum timeout per the specification (10ms)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "send_oldest", CTLFLAG_RW, &rack_always_send_oldest, 0, "Should we always send the oldest TLP and RACK-TLP"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "rack_tlimit", CTLFLAG_RW, &rack_limited_retran, 0, "How many times can a rack timeout drive out sends"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "tlp_retry", CTLFLAG_RW, &rack_tlp_max_resend, 2, "How many times does TLP retry a single segment or multiple with no ACK"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW, &rack_lower_cwnd_at_tlp, 0, "When a TLP completes a retran should we enter recovery"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "reorder_thresh", CTLFLAG_RW, &rack_reorder_thresh, 2, "What factor for rack will be added when seeing reordering (shift right)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW, &rack_tlp_thresh, 1, "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "reorder_fade", CTLFLAG_RW, &rack_reorder_fade, 0, "Does reorder detection fade, if so how many ms (0 means never)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "pktdelay", CTLFLAG_RW, &rack_pkt_delay, 1, "Extra RACK time (in ms) besides reordering thresh"); /* Timer related controls */ rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "timers", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Timer related controls"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "persmin", CTLFLAG_RW, &rack_persist_min, 250, "What is the minimum time in milliseconds between persists"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "persmax", CTLFLAG_RW, &rack_persist_max, 2000, "What is the largest delay in milliseconds between persists"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "delayed_ack", CTLFLAG_RW, &rack_delayed_ack_time, 200, "Delayed ack time (200ms)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "minrto", CTLFLAG_RW, &rack_rto_min, 0, "Minimum RTO in ms -- set with caution below 1000 due to TLP"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "maxrto", CTLFLAG_RW, &rack_rto_max, 0, "Maxiumum RTO in ms -- should be at least as large as min_rto"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "minto", CTLFLAG_RW, &rack_min_to, 1, "Minimum rack timeout in milliseconds"); /* Measure controls */ rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "measure", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Measure related controls"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "wma_divisor", CTLFLAG_RW, &rack_wma_divisor, 8, "When doing b/w calculation what is the divisor for the WMA"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "end_cwnd", CTLFLAG_RW, &rack_cwnd_block_ends_measure, 0, "Does a cwnd just-return end the measurement window (app limited)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "end_rwnd", CTLFLAG_RW, &rack_rwnd_block_ends_measure, 0, "Does an rwnd just-return end the measurement window (app limited -- not persists)"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "min_target", CTLFLAG_RW, &rack_def_data_window, 20, "What is the minimum target window (in mss) for a GP measurements"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "goal_bdp", CTLFLAG_RW, &rack_goal_bdp, 2, "What is the goal BDP to measure"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "min_srtts", CTLFLAG_RW, &rack_min_srtts, 1, "What is the goal BDP to measure"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "min_measure_tim", CTLFLAG_RW, &rack_min_measure_usec, 0, "What is the Minimum time time for a measurement if 0, this is off"); /* Misc rack controls */ rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "misc", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Misc related controls"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "shared_cwnd", CTLFLAG_RW, &rack_enable_shared_cwnd, 0, "Should RACK try to use the shared cwnd on connections where allowed"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "limits_on_scwnd", CTLFLAG_RW, &rack_limits_scwnd, 1, "Should RACK place low end time limits on the shared cwnd feature"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW, &rack_enable_mqueue_for_nonpaced, 0, "Should RACK use mbuf queuing for non-paced connections"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "iMac_dack", CTLFLAG_RW, &rack_use_imac_dack, 0, "Should RACK try to emulate iMac delayed ack"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "no_prr", CTLFLAG_RW, &rack_disable_prr, 0, "Should RACK not use prr and only pace (must have pacing on)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "bb_verbose", CTLFLAG_RW, &rack_verbose_logging, 0, "Should RACK black box logging be verbose"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "data_after_close", CTLFLAG_RW, &rack_ignore_data_after_close, 1, "Do we hold off sending a RST until all pending data is ack'd"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "no_sack_needed", CTLFLAG_RW, &rack_sack_not_required, 0, "Do we allow rack to run on connections not supporting SACK"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "recovery_loss_prop", CTLFLAG_RW, &rack_use_proportional_reduce, 0, "Should we proportionaly reduce cwnd based on the number of losses "); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "recovery_prop", CTLFLAG_RW, &rack_proportional_rate, 10, "What percent reduction per loss"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "prr_sendalot", CTLFLAG_RW, &rack_send_a_lot_in_prr, 1, "Send a lot in prr"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "earlyrecovery", CTLFLAG_RW, &rack_early_recovery, 1, "Do we do early recovery with rack"); /* Sack Attacker detection stuff */ SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "detect_highsackratio", CTLFLAG_RW, &rack_highest_sack_thresh_seen, 0, "Highest sack to ack ratio seen"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "detect_highmoveratio", CTLFLAG_RW, &rack_highest_move_thresh_seen, 0, "Highest move to non-move ratio seen"); rack_ack_total = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "acktotal", CTLFLAG_RD, &rack_ack_total, "Total number of Ack's"); rack_express_sack = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "exp_sacktotal", CTLFLAG_RD, &rack_express_sack, "Total expresss number of Sack's"); rack_sack_total = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "sacktotal", CTLFLAG_RD, &rack_sack_total, "Total number of SACKs"); rack_move_none = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "move_none", CTLFLAG_RD, &rack_move_none, "Total number of SACK index reuse of postions under threshold"); rack_move_some = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "move_some", CTLFLAG_RD, &rack_move_some, "Total number of SACK index reuse of postions over threshold"); rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "attacks", CTLFLAG_RD, &rack_sack_attacks_detected, "Total number of SACK attackers that had sack disabled"); rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "reversed", CTLFLAG_RD, &rack_sack_attacks_reversed, "Total number of SACK attackers that were later determined false positive"); rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "nextmerge", CTLFLAG_RD, &rack_sack_used_next_merge, "Total number of times we used the next merge"); rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "prevmerge", CTLFLAG_RD, &rack_sack_used_prev_merge, "Total number of times we used the prev merge"); /* Counters */ rack_badfr = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "badfr", CTLFLAG_RD, &rack_badfr, "Total number of bad FRs"); rack_badfr_bytes = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "badfr_bytes", CTLFLAG_RD, &rack_badfr_bytes, "Total number of bad FRs"); rack_rtm_prr_retran = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "prrsndret", CTLFLAG_RD, &rack_rtm_prr_retran, "Total number of prr based retransmits"); rack_rtm_prr_newdata = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "prrsndnew", CTLFLAG_RD, &rack_rtm_prr_newdata, "Total number of prr based new transmits"); rack_timestamp_mismatch = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tsnf", CTLFLAG_RD, &rack_timestamp_mismatch, "Total number of timestamps that we could not find the reported ts"); rack_find_high = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "findhigh", CTLFLAG_RD, &rack_find_high, "Total number of FIN causing find-high"); rack_reorder_seen = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "reordering", CTLFLAG_RD, &rack_reorder_seen, "Total number of times we added delay due to reordering"); rack_tlp_tot = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_to_total", CTLFLAG_RD, &rack_tlp_tot, "Total number of tail loss probe expirations"); rack_tlp_newdata = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_new", CTLFLAG_RD, &rack_tlp_newdata, "Total number of tail loss probe sending new data"); rack_tlp_retran = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_retran", CTLFLAG_RD, &rack_tlp_retran, "Total number of tail loss probe sending retransmitted data"); rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD, &rack_tlp_retran_bytes, "Total bytes of tail loss probe sending retransmitted data"); rack_tlp_retran_fail = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_retran_fail", CTLFLAG_RD, &rack_tlp_retran_fail, "Total number of tail loss probe sending retransmitted data that failed (wait for t3)"); rack_to_tot = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "rack_to_tot", CTLFLAG_RD, &rack_to_tot, "Total number of times the rack to expired"); rack_to_arm_rack = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "arm_rack", CTLFLAG_RD, &rack_to_arm_rack, "Total number of times the rack timer armed"); rack_to_arm_tlp = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "arm_tlp", CTLFLAG_RD, &rack_to_arm_tlp, "Total number of times the tlp timer armed"); rack_calc_zero = counter_u64_alloc(M_WAITOK); rack_calc_nonzero = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "calc_zero", CTLFLAG_RD, &rack_calc_zero, "Total number of times pacing time worked out to zero"); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "calc_nonzero", CTLFLAG_RD, &rack_calc_nonzero, "Total number of times pacing time worked out to non-zero"); rack_paced_segments = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "paced", CTLFLAG_RD, &rack_paced_segments, "Total number of times a segment send caused hptsi"); rack_unpaced_segments = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "unpaced", CTLFLAG_RD, &rack_unpaced_segments, "Total number of times a segment did not cause hptsi"); rack_saw_enobuf = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "saw_enobufs", CTLFLAG_RD, &rack_saw_enobuf, "Total number of times a segment did not cause hptsi"); rack_saw_enetunreach = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "saw_enetunreach", CTLFLAG_RD, &rack_saw_enetunreach, "Total number of times a segment did not cause hptsi"); rack_to_alloc = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "allocs", CTLFLAG_RD, &rack_to_alloc, "Total allocations of tracking structures"); rack_to_alloc_hard = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "allochard", CTLFLAG_RD, &rack_to_alloc_hard, "Total allocations done with sleeping the hard way"); rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "allocemerg", CTLFLAG_RD, &rack_to_alloc_emerg, "Total allocations done from emergency cache"); rack_to_alloc_limited = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "alloc_limited", CTLFLAG_RD, &rack_to_alloc_limited, "Total allocations dropped due to limit"); rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "alloc_limited_conns", CTLFLAG_RD, &rack_alloc_limited_conns, "Connections with allocations dropped due to limit"); rack_split_limited = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "split_limited", CTLFLAG_RD, &rack_split_limited, "Split allocations dropped due to limit"); rack_sack_proc_all = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "sack_long", CTLFLAG_RD, &rack_sack_proc_all, "Total times we had to walk whole list for sack processing"); rack_sack_proc_restart = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "sack_restart", CTLFLAG_RD, &rack_sack_proc_restart, "Total times we had to walk whole list due to a restart"); rack_sack_proc_short = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "sack_short", CTLFLAG_RD, &rack_sack_proc_short, "Total times we took shortcut for sack processing"); rack_enter_tlp_calc = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_calc_entered", CTLFLAG_RD, &rack_enter_tlp_calc, "Total times we called calc-tlp"); rack_used_tlpmethod = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "hit_tlp_method", CTLFLAG_RD, &rack_used_tlpmethod, "Total number of runt sacks"); rack_used_tlpmethod2 = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "hit_tlp_method2", CTLFLAG_RD, &rack_used_tlpmethod2, "Total number of times we hit TLP method 2"); rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "skipacked", CTLFLAG_RD, &rack_sack_skipped_acked, "Total number of times we skipped previously sacked"); rack_sack_splits = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "ofsplit", CTLFLAG_RD, &rack_sack_splits, "Total number of times we did the old fashion tree split"); rack_progress_drops = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "prog_drops", CTLFLAG_RD, &rack_progress_drops, "Total number of progress drops"); rack_input_idle_reduces = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD, &rack_input_idle_reduces, "Total number of idle reductions on input"); rack_collapsed_win = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "collapsed_win", CTLFLAG_RD, &rack_collapsed_win, "Total number of collapsed windows"); rack_tlp_does_nada = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_nada", CTLFLAG_RD, &rack_tlp_does_nada, "Total number of nada tlp calls"); rack_try_scwnd = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tried_scwnd", CTLFLAG_RD, &rack_try_scwnd, "Total number of scwnd attempts"); rack_per_timer_hole = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "timer_hole", CTLFLAG_RD, &rack_per_timer_hole, "Total persists start in timer hole"); COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "outsize", CTLFLAG_RD, rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes"); COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "opts", CTLFLAG_RD, rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats"); SYSCTL_ADD_PROC(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters"); } static __inline int rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a) { if (SEQ_GEQ(b->r_start, a->r_start) && SEQ_LT(b->r_start, a->r_end)) { /* * The entry b is within the * block a. i.e.: * a -- |-------------| * b -- |----| * * b -- |------| * * b -- |-----------| */ return (0); } else if (SEQ_GEQ(b->r_start, a->r_end)) { /* * b falls as either the next * sequence block after a so a * is said to be smaller than b. * i.e: * a -- |------| * b -- |--------| * or * b -- |-----| */ return (1); } /* * Whats left is where a is * larger than b. i.e: * a -- |-------| * b -- |---| * or even possibly * b -- |--------------| */ return (-1); } RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp); RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp); static uint32_t rc_init_window(struct tcp_rack *rack) { uint32_t win; if (rack->rc_init_win == 0) { /* * Nothing set by the user, use the system stack * default. */ return(tcp_compute_initwnd(tcp_maxseg(rack->rc_tp))); } win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win; return(win); } static uint64_t rack_get_fixed_pacing_bw(struct tcp_rack *rack) { if (IN_RECOVERY(rack->rc_tp->t_flags)) return (rack->r_ctl.rc_fixed_pacing_rate_rec); else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) return (rack->r_ctl.rc_fixed_pacing_rate_ss); else return (rack->r_ctl.rc_fixed_pacing_rate_ca); } static uint64_t rack_get_bw(struct tcp_rack *rack) { if (rack->use_fixed_rate) { /* Return the fixed pacing rate */ return (rack_get_fixed_pacing_bw(rack)); } if (rack->r_ctl.gp_bw == 0) { /* * We have yet no b/w measurement, * if we have a user set initial bw * return it. If we don't have that and * we have an srtt, use the tcp IW (10) to * calculate a fictional b/w over the SRTT * which is more or less a guess. Note * we don't use our IW from rack on purpose * so if we have like IW=30, we are not * calculating a "huge" b/w. */ uint64_t bw, srtt; if (rack->r_ctl.init_rate) return (rack->r_ctl.init_rate); /* Has the user set a max peak rate? */ #ifdef NETFLIX_PEAKRATE if (rack->rc_tp->t_maxpeakrate) return (rack->rc_tp->t_maxpeakrate); #endif /* Ok lets come up with the IW guess, if we have a srtt */ if (rack->rc_tp->t_srtt == 0) { /* * Go with old pacing method * i.e. burst mitigation only. */ return (0); } /* Ok lets get the initial TCP win (not racks) */ bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)); srtt = ((uint64_t)TICKS_2_USEC(rack->rc_tp->t_srtt) >> TCP_RTT_SHIFT); bw *= (uint64_t)USECS_IN_SECOND; bw /= srtt; return (bw); } else { uint64_t bw; if(rack->r_ctl.num_avg >= RACK_REQ_AVG) { /* Averaging is done, we can return the value */ bw = rack->r_ctl.gp_bw; } else { /* Still doing initial average must calculate */ bw = rack->r_ctl.gp_bw / rack->r_ctl.num_avg; } #ifdef NETFLIX_PEAKRATE if ((rack->rc_tp->t_maxpeakrate) && (bw > rack->rc_tp->t_maxpeakrate)) { /* The user has set a peak rate to pace at * don't allow us to pace faster than that. */ return (rack->rc_tp->t_maxpeakrate); } #endif return (bw); } } static uint16_t rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm) { if (rack->use_fixed_rate) { return (100); } else if (rack->in_probe_rtt && (rsm == NULL)) return(rack->r_ctl.rack_per_of_gp_probertt); else if ((IN_RECOVERY(rack->rc_tp->t_flags) && rack->r_ctl.rack_per_of_gp_rec)) { if (rsm) { /* a retransmission always use the recovery rate */ return(rack->r_ctl.rack_per_of_gp_rec); } else if (rack->rack_rec_nonrxt_use_cr) { /* Directed to use the configured rate */ goto configured_rate; } else if (rack->rack_no_prr && (rack->r_ctl.rack_per_of_gp_rec > 100)) { /* No PRR, lets just use the b/w estimate only */ return(100); } else { /* * Here we may have a non-retransmit but we * have no overrides, so just use the recovery * rate (prr is in effect). */ return(rack->r_ctl.rack_per_of_gp_rec); } } configured_rate: /* For the configured rate we look at our cwnd vs the ssthresh */ if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) return (rack->r_ctl.rack_per_of_gp_ss); else return(rack->r_ctl.rack_per_of_gp_ca); } static uint64_t rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm) { /* * We allow rack_per_of_gp_xx to dictate our bw rate we want. */ uint64_t bw_est; uint64_t gain; gain = (uint64_t)rack_get_output_gain(rack, rsm); bw_est = bw * gain; bw_est /= (uint64_t)100; /* Never fall below the minimum (def 64kbps) */ if (bw_est < RACK_MIN_BW) bw_est = RACK_MIN_BW; return (bw_est); } static void rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; if ((mod != 1) && (rack_verbose_logging == 0)) { /* * We get 3 values currently for mod * 1 - We are retransmitting and this tells the reason. * 2 - We are clearing a dup-ack count. * 3 - We are incrementing a dup-ack count. * * The clear/increment are only logged * if you have BBverbose on. */ return; } memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = tsused; log.u_bbr.flex2 = thresh; log.u_bbr.flex3 = rsm->r_flags; log.u_bbr.flex4 = rsm->r_dupack; log.u_bbr.flex5 = rsm->r_start; log.u_bbr.flex6 = rsm->r_end; log.u_bbr.flex8 = mod; log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_SETTINGS_CHG, 0, 0, &log, false, &tv); } } static void rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = TICKS_2_MSEC(rack->rc_tp->t_srtt >> TCP_RTT_SHIFT); log.u_bbr.flex2 = to * 1000; log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags; log.u_bbr.flex4 = slot; log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot; log.u_bbr.flex6 = rack->rc_tp->t_rxtcur; log.u_bbr.flex7 = rack->rc_in_persist; log.u_bbr.flex8 = which; if (rack->rack_no_prr) log.u_bbr.pkts_out = 0; else log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERSTAR, 0, 0, &log, false, &tv); } } static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex8 = to_num; log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt; log.u_bbr.flex2 = rack->rc_rack_rtt; if (rsm == NULL) log.u_bbr.flex3 = 0; else log.u_bbr.flex3 = rsm->r_end - rsm->r_start; if (rack->rack_no_prr) log.u_bbr.flex5 = 0; else log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_RTO, 0, 0, &log, false, &tv); } } static void rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len, struct rack_sendmap *rsm, int conf) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = t; log.u_bbr.flex2 = len; log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt * HPTS_USEC_IN_MSEC; log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest * HPTS_USEC_IN_MSEC; log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest * HPTS_USEC_IN_MSEC; log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_rtt_cnt; log.u_bbr.flex7 = conf; log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot * (uint64_t)HPTS_USEC_IN_MSEC; log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method; if (rack->rack_no_prr) log.u_bbr.pkts_out = 0; else log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtt; log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); if (rsm) { log.u_bbr.pkt_epoch = rsm->r_start; log.u_bbr.lost = rsm->r_end; log.u_bbr.cwnd_gain = rsm->r_rtr_cnt; } else { /* Its a SYN */ log.u_bbr.pkt_epoch = rack->rc_tp->iss; log.u_bbr.lost = 0; log.u_bbr.cwnd_gain = 0; } /* Write out general bits of interest rrs here */ log.u_bbr.use_lt_bw = rack->rc_highly_buffered; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->forced_ack; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->in_probe_rtt; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->app_limited_needs_set; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->rc_gp_filled; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom; log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight; log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts; log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered; log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts; log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt; TCP_LOG_EVENTP(tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_BBRRTT, 0, 0, &log, false, &tv); } } static void rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt) { /* * Log the rtt sample we are * applying to the srtt algorithm in * useconds. */ if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; /* Convert our ms to a microsecond */ memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = rtt * 1000; log.u_bbr.flex2 = rack->r_ctl.ack_count; log.u_bbr.flex3 = rack->r_ctl.sack_count; log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move; log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra; log.u_bbr.flex8 = rack->sack_attack_disable; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, TCP_LOG_RTT, 0, 0, &log, false, &tv); } } static inline void rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line) { if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = line; log.u_bbr.flex2 = tick; log.u_bbr.flex3 = tp->t_maxunacktime; log.u_bbr.flex4 = tp->t_acktime; log.u_bbr.flex8 = event; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); TCP_LOG_EVENTP(tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_PROGRESS, 0, 0, &log, false, &tv); } } static void rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = slot; if (rack->rack_no_prr) log.u_bbr.flex2 = 0; else log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags); log.u_bbr.flex8 = rack->rc_in_persist; log.u_bbr.timeStamp = cts; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_BBRSND, 0, 0, &log, false, tv); } } static void rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = did_out; log.u_bbr.flex2 = nxt_pkt; log.u_bbr.flex3 = way_out; log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; if (rack->rack_no_prr) log.u_bbr.flex5 = 0; else log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs; log.u_bbr.flex7 = rack->r_wanted_output; log.u_bbr.flex8 = rack->rc_in_persist; log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_DOSEG_DONE, 0, 0, &log, false, &tv); } } static void rack_log_type_hrdwtso(struct tcpcb *tp, struct tcp_rack *rack, int len, int mod, int32_t orig_len, int frm) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; uint32_t cts; memset(&log, 0, sizeof(log)); cts = tcp_get_usecs(&tv); log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs; log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs; log.u_bbr.flex4 = len; log.u_bbr.flex5 = orig_len; log.u_bbr.flex6 = rack->r_ctl.rc_sacked; log.u_bbr.flex7 = mod; log.u_bbr.flex8 = frm; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); TCP_LOG_EVENTP(tp, NULL, &tp->t_inpcb->inp_socket->so_rcv, &tp->t_inpcb->inp_socket->so_snd, TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv); } } static void rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot, uint8_t hpts_calling, int reason, uint32_t cwnd_to_use) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = slot; log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags; log.u_bbr.flex4 = reason; if (rack->rack_no_prr) log.u_bbr.flex5 = 0; else log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex7 = hpts_calling; log.u_bbr.flex8 = rack->rc_in_persist; log.u_bbr.lt_epoch = cwnd_to_use; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_JUSTRET, 0, tlen, &log, false, &tv); } } static void rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts, struct timeval *tv, uint32_t flags_on_entry) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = line; log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to; log.u_bbr.flex3 = flags_on_entry; log.u_bbr.flex4 = us_cts; if (rack->rack_no_prr) log.u_bbr.flex5 = 0; else log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex6 = rack->rc_tp->t_rxtcur; log.u_bbr.flex7 = hpts_removed; log.u_bbr.flex8 = 1; log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags; log.u_bbr.timeStamp = us_cts; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERCANC, 0, 0, &log, false, tv); } } static void rack_log_alt_to_to_cancel(struct tcp_rack *rack, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint32_t flex4, uint32_t flex5, uint32_t flex6, uint16_t flex7, uint8_t mod) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; if (mod == 1) { /* No you can't use 1, its for the real to cancel */ return; } memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.flex1 = flex1; log.u_bbr.flex2 = flex2; log.u_bbr.flex3 = flex3; log.u_bbr.flex4 = flex4; log.u_bbr.flex5 = flex5; log.u_bbr.flex6 = flex6; log.u_bbr.flex7 = flex7; log.u_bbr.flex8 = mod; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERCANC, 0, 0, &log, false, &tv); } } static void rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = timers; log.u_bbr.flex2 = ret; log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp; log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; log.u_bbr.flex5 = cts; if (rack->rack_no_prr) log.u_bbr.flex6 = 0; else log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_TO_PROCESS, 0, 0, &log, false, &tv); } } static void rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = rack->r_ctl.rc_prr_out; log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs; if (rack->rack_no_prr) log.u_bbr.flex3 = 0; else log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered; log.u_bbr.flex5 = rack->r_ctl.rc_sacked; log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt; log.u_bbr.flex8 = frm; log.u_bbr.pkts_out = orig_cwnd; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_BBRUPD, 0, 0, &log, false, &tv); } } #ifdef NETFLIX_EXP_DETECTION static void rack_log_sad(struct tcp_rack *rack, int event) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = rack->r_ctl.sack_count; log.u_bbr.flex2 = rack->r_ctl.ack_count; log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra; log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move; log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced; log.u_bbr.flex6 = tcp_sack_to_ack_thresh; log.u_bbr.pkts_out = tcp_sack_to_move_thresh; log.u_bbr.lt_epoch = (tcp_force_detection << 8); log.u_bbr.lt_epoch |= rack->do_detection; log.u_bbr.applimited = tcp_map_minimum; log.u_bbr.flex7 = rack->sack_attack_disable; log.u_bbr.flex8 = event; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.delivered = tcp_sad_decay_val; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, TCP_SAD_DETECTION, 0, 0, &log, false, &tv); } } #endif static void rack_counter_destroy(void) { counter_u64_free(rack_ack_total); counter_u64_free(rack_express_sack); counter_u64_free(rack_sack_total); counter_u64_free(rack_move_none); counter_u64_free(rack_move_some); counter_u64_free(rack_sack_attacks_detected); counter_u64_free(rack_sack_attacks_reversed); counter_u64_free(rack_sack_used_next_merge); counter_u64_free(rack_sack_used_prev_merge); counter_u64_free(rack_badfr); counter_u64_free(rack_badfr_bytes); counter_u64_free(rack_rtm_prr_retran); counter_u64_free(rack_rtm_prr_newdata); counter_u64_free(rack_timestamp_mismatch); counter_u64_free(rack_find_high); counter_u64_free(rack_reorder_seen); counter_u64_free(rack_tlp_tot); counter_u64_free(rack_tlp_newdata); counter_u64_free(rack_tlp_retran); counter_u64_free(rack_tlp_retran_bytes); counter_u64_free(rack_tlp_retran_fail); counter_u64_free(rack_to_tot); counter_u64_free(rack_to_arm_rack); counter_u64_free(rack_to_arm_tlp); counter_u64_free(rack_calc_zero); counter_u64_free(rack_calc_nonzero); counter_u64_free(rack_paced_segments); counter_u64_free(rack_unpaced_segments); counter_u64_free(rack_saw_enobuf); counter_u64_free(rack_saw_enetunreach); counter_u64_free(rack_to_alloc); counter_u64_free(rack_to_alloc_hard); counter_u64_free(rack_to_alloc_emerg); counter_u64_free(rack_to_alloc_limited); counter_u64_free(rack_alloc_limited_conns); counter_u64_free(rack_split_limited); counter_u64_free(rack_sack_proc_all); counter_u64_free(rack_sack_proc_restart); counter_u64_free(rack_sack_proc_short); counter_u64_free(rack_enter_tlp_calc); counter_u64_free(rack_used_tlpmethod); counter_u64_free(rack_used_tlpmethod2); counter_u64_free(rack_sack_skipped_acked); counter_u64_free(rack_sack_splits); counter_u64_free(rack_progress_drops); counter_u64_free(rack_input_idle_reduces); counter_u64_free(rack_collapsed_win); counter_u64_free(rack_tlp_does_nada); counter_u64_free(rack_try_scwnd); counter_u64_free(rack_per_timer_hole); COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE); COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE); } static struct rack_sendmap * rack_alloc(struct tcp_rack *rack) { struct rack_sendmap *rsm; rsm = uma_zalloc(rack_zone, M_NOWAIT); if (rsm) { rack->r_ctl.rc_num_maps_alloced++; counter_u64_add(rack_to_alloc, 1); return (rsm); } if (rack->rc_free_cnt) { counter_u64_add(rack_to_alloc_emerg, 1); rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext); rack->rc_free_cnt--; return (rsm); } return (NULL); } static struct rack_sendmap * rack_alloc_full_limit(struct tcp_rack *rack) { if ((V_tcp_map_entries_limit > 0) && (rack->do_detection == 0) && (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { counter_u64_add(rack_to_alloc_limited, 1); if (!rack->alloc_limit_reported) { rack->alloc_limit_reported = 1; counter_u64_add(rack_alloc_limited_conns, 1); } return (NULL); } return (rack_alloc(rack)); } /* wrapper to allocate a sendmap entry, subject to a specific limit */ static struct rack_sendmap * rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type) { struct rack_sendmap *rsm; if (limit_type) { /* currently there is only one limit type */ if (V_tcp_map_split_limit > 0 && (rack->do_detection == 0) && rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) { counter_u64_add(rack_split_limited, 1); if (!rack->alloc_limit_reported) { rack->alloc_limit_reported = 1; counter_u64_add(rack_alloc_limited_conns, 1); } return (NULL); } } /* allocate and mark in the limit type, if set */ rsm = rack_alloc(rack); if (rsm != NULL && limit_type) { rsm->r_limit_type = limit_type; rack->r_ctl.rc_num_split_allocs++; } return (rsm); } static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm) { if (rsm->r_flags & RACK_APP_LIMITED) { if (rack->r_ctl.rc_app_limited_cnt > 0) { rack->r_ctl.rc_app_limited_cnt--; } } if (rsm->r_limit_type) { /* currently there is only one limit type */ rack->r_ctl.rc_num_split_allocs--; } if (rsm == rack->r_ctl.rc_first_appl) { if (rack->r_ctl.rc_app_limited_cnt == 0) rack->r_ctl.rc_first_appl = NULL; else { /* Follow the next one out */ struct rack_sendmap fe; fe.r_start = rsm->r_nseq_appl; rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); } } if (rsm == rack->r_ctl.rc_resend) rack->r_ctl.rc_resend = NULL; if (rsm == rack->r_ctl.rc_rsm_at_retran) rack->r_ctl.rc_rsm_at_retran = NULL; if (rsm == rack->r_ctl.rc_end_appl) rack->r_ctl.rc_end_appl = NULL; if (rack->r_ctl.rc_tlpsend == rsm) rack->r_ctl.rc_tlpsend = NULL; if (rack->r_ctl.rc_sacklast == rsm) rack->r_ctl.rc_sacklast = NULL; if (rack->rc_free_cnt < rack_free_cache) { memset(rsm, 0, sizeof(struct rack_sendmap)); TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext); rsm->r_limit_type = 0; rack->rc_free_cnt++; return; } rack->r_ctl.rc_num_maps_alloced--; uma_zfree(rack_zone, rsm); } static uint32_t rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack) { uint64_t srtt, bw, len, tim; uint32_t segsiz, def_len, minl; segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); def_len = rack_def_data_window * segsiz; if (rack->rc_gp_filled == 0) { /* * We have no measurement (IW is in flight?) so * we can only guess using our data_window sysctl * value (usually 100MSS). */ return (def_len); } /* * Now we have a number of factors to consider. * * 1) We have a desired BDP which is usually * at least 2. * 2) We have a minimum number of rtt's usually 1 SRTT * but we allow it too to be more. * 3) We want to make sure a measurement last N useconds (if * we have set rack_min_measure_usec. * * We handle the first concern here by trying to create a data * window of max(rack_def_data_window, DesiredBDP). The * second concern we handle in not letting the measurement * window end normally until at least the required SRTT's * have gone by which is done further below in * rack_enough_for_measurement(). Finally the third concern * we also handle here by calculating how long that time * would take at the current BW and then return the * max of our first calculation and that length. Note * that if rack_min_measure_usec is 0, we don't deal * with concern 3. Also for both Concern 1 and 3 an * application limited period could end the measurement * earlier. * * So lets calculate the BDP with the "known" b/w using * the SRTT has our rtt and then multiply it by the * goal. */ bw = rack_get_bw(rack); srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); len = bw * srtt; len /= (uint64_t)HPTS_USEC_IN_SEC; len *= max(1, rack_goal_bdp); /* Now we need to round up to the nearest MSS */ len = roundup(len, segsiz); if (rack_min_measure_usec) { /* Now calculate our min length for this b/w */ tim = rack_min_measure_usec; minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC; if (minl == 0) minl = 1; minl = roundup(minl, segsiz); if (len < minl) len = minl; } /* * Now if we have a very small window we want * to attempt to get the window that is * as small as possible. This happens on * low b/w connections and we don't want to * span huge numbers of rtt's between measurements. * * We basically include 2 over our "MIN window" so * that the measurement can be shortened (possibly) by * an ack'ed packet. */ if (len < def_len) return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz))); else return (max((uint32_t)len, def_len)); } static int rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack) { uint32_t tim, srtts, segsiz; /* * Has enough time passed for the GP measurement to be valid? */ if ((tp->snd_max == tp->snd_una) || (th_ack == tp->snd_max)){ /* All is acked */ return (1); } if (SEQ_LT(th_ack, tp->gput_seq)) { /* Not enough bytes yet */ return (0); } segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); if (SEQ_LT(th_ack, tp->gput_ack) && ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) { /* Not enough bytes yet */ return (0); } if (rack->r_ctl.rc_first_appl && (rack->r_ctl.rc_first_appl->r_start == th_ack)) { /* * We are up to the app limited point * we have to measure irrespective of the time.. */ return (1); } /* Now what about time? */ srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts); tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts; if (tim >= srtts) { return (1); } /* Nope not even a full SRTT has passed */ return (0); } static void rack_log_timely(struct tcp_rack *rack, uint32_t logged, uint64_t cur_bw, uint64_t low_bnd, uint64_t up_bnd, int line, uint8_t method) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = logged; log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt; log.u_bbr.flex2 <<= 4; log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt; log.u_bbr.flex2 <<= 4; log.u_bbr.flex2 |= rack->rc_gp_incr; log.u_bbr.flex2 <<= 4; log.u_bbr.flex2 |= rack->rc_gp_bwred; log.u_bbr.flex3 = rack->rc_gp_incr; log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss; log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca; log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec; log.u_bbr.flex7 = rack->rc_gp_bwred; log.u_bbr.flex8 = method; log.u_bbr.cur_del_rate = cur_bw; log.u_bbr.delRate = low_bnd; log.u_bbr.bw_inuse = up_bnd; log.u_bbr.rttProp = rack_get_bw(rack); log.u_bbr.pkt_epoch = line; log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt; log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt; log.u_bbr.cwnd_gain = rack->rc_dragged_bottom; log.u_bbr.cwnd_gain <<= 1; log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec; log.u_bbr.cwnd_gain <<= 1; log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss; log.u_bbr.cwnd_gain <<= 1; log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca; log.u_bbr.lost = rack->r_ctl.rc_loss_count; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, TCP_TIMELY_WORK, 0, 0, &log, false, &tv); } } static int rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult) { /* * Before we increase we need to know if * the estimate just made was less than * our pacing goal (i.e. (cur_bw * mult) > last_bw_est) * * If we already are pacing at a fast enough * rate to push us faster there is no sense of * increasing. * * We first caculate our actual pacing rate (ss or ca multipler * times our cur_bw). * * Then we take the last measured rate and multipy by our * maximum pacing overage to give us a max allowable rate. * * If our act_rate is smaller than our max_allowable rate * then we should increase. Else we should hold steady. * */ uint64_t act_rate, max_allow_rate; if (rack_timely_no_stopping) return (1); if ((cur_bw == 0) || (last_bw_est == 0)) { /* * Initial startup case or * everything is acked case. */ rack_log_timely(rack, mult, cur_bw, 0, 0, __LINE__, 9); return (1); } if (mult <= 100) { /* * We can always pace at or slightly above our rate. */ rack_log_timely(rack, mult, cur_bw, 0, 0, __LINE__, 9); return (1); } act_rate = cur_bw * (uint64_t)mult; act_rate /= 100; max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100); max_allow_rate /= 100; if (act_rate < max_allow_rate) { /* * Here the rate we are actually pacing at * is smaller than 10% above our last measurement. * This means we are pacing below what we would * like to try to achieve (plus some wiggle room). */ rack_log_timely(rack, mult, cur_bw, act_rate, max_allow_rate, __LINE__, 9); return (1); } else { /* * Here we are already pacing at least rack_max_per_above(10%) * what we are getting back. This indicates most likely * that we are being limited (cwnd/rwnd/app) and can't * get any more b/w. There is no sense of trying to * raise up the pacing rate its not speeding us up * and we already are pacing faster than we are getting. */ rack_log_timely(rack, mult, cur_bw, act_rate, max_allow_rate, __LINE__, 8); return (0); } } static void rack_validate_multipliers_at_or_above100(struct tcp_rack *rack) { /* * When we drag bottom, we want to assure * that no multiplier is below 1.0, if so * we want to restore it to at least that. */ if (rack->r_ctl.rack_per_of_gp_rec < 100) { /* This is unlikely we usually do not touch recovery */ rack->r_ctl.rack_per_of_gp_rec = 100; } if (rack->r_ctl.rack_per_of_gp_ca < 100) { rack->r_ctl.rack_per_of_gp_ca = 100; } if (rack->r_ctl.rack_per_of_gp_ss < 100) { rack->r_ctl.rack_per_of_gp_ss = 100; } } static void rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack) { if (rack->r_ctl.rack_per_of_gp_ca > 100) { rack->r_ctl.rack_per_of_gp_ca = 100; } if (rack->r_ctl.rack_per_of_gp_ss > 100) { rack->r_ctl.rack_per_of_gp_ss = 100; } } static void rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override) { int32_t calc, logged, plus; logged = 0; if (override) { /* * override is passed when we are * loosing b/w and making one last * gasp at trying to not loose out * to a new-reno flow. */ goto extra_boost; } /* In classic timely we boost by 5x if we have 5 increases in a row, lets not */ if (rack->rc_gp_incr && ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) { /* * Reset and get 5 strokes more before the boost. Note * that the count is 0 based so we have to add one. */ extra_boost: plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST; rack->rc_gp_timely_inc_cnt = 0; } else plus = (uint32_t)rack_gp_increase_per; /* Must be at least 1% increase for true timely increases */ if ((plus < 1) && ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0))) plus = 1; if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0) && rack_bw_can_be_raised(rack, cur_bw, last_bw_est, rack->r_ctl.rack_per_of_gp_rec)) { /* We have been in recovery ding it too */ calc = rack->r_ctl.rack_per_of_gp_rec + plus; if (calc > 0xffff) calc = 0xffff; logged |= 1; rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc; if (rack_per_upper_bound_ss && (rack->rc_dragged_bottom == 0) && (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss)) rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss; } if (rack->rc_gp_saw_ca && (rack->rc_gp_saw_ss == 0) && rack_bw_can_be_raised(rack, cur_bw, last_bw_est, rack->r_ctl.rack_per_of_gp_ca)) { /* In CA */ calc = rack->r_ctl.rack_per_of_gp_ca + plus; if (calc > 0xffff) calc = 0xffff; logged |= 2; rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc; if (rack_per_upper_bound_ca && (rack->rc_dragged_bottom == 0) && (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca)) rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca; } if (rack->rc_gp_saw_ss && rack_bw_can_be_raised(rack, cur_bw, last_bw_est, rack->r_ctl.rack_per_of_gp_ss)) { /* In SS */ calc = rack->r_ctl.rack_per_of_gp_ss + plus; if (calc > 0xffff) calc = 0xffff; rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc; if (rack_per_upper_bound_ss && (rack->rc_dragged_bottom == 0) && (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss)) rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss; logged |= 4; } if (logged && (rack->rc_gp_incr == 0)){ /* Go into increment mode */ rack->rc_gp_incr = 1; rack->rc_gp_timely_inc_cnt = 0; } if (rack->rc_gp_incr && logged && (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) { rack->rc_gp_timely_inc_cnt++; } rack_log_timely(rack, logged, plus, 0, 0, __LINE__, 1); } static uint32_t rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff) { /* * norm_grad = rtt_diff / minrtt; * new_per = curper * (1 - B * norm_grad) * * B = rack_gp_decrease_per (default 10%) * rtt_dif = input var current rtt-diff * curper = input var current percentage * minrtt = from rack filter * */ uint64_t perf; perf = (((uint64_t)curper * ((uint64_t)1000000 - ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 * (((uint64_t)rtt_diff * (uint64_t)1000000)/ (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/ (uint64_t)1000000)) / (uint64_t)1000000); if (perf > curper) { /* TSNH */ perf = curper - 1; } return ((uint32_t)perf); } static uint32_t rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt) { /* * highrttthresh * result = curper * (1 - (B * ( 1 - ------ )) * gp_srtt * * B = rack_gp_decrease_per (default 10%) * highrttthresh = filter_min * rack_gp_rtt_maxmul */ uint64_t perf; uint32_t highrttthresh; highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul; perf = (((uint64_t)curper * ((uint64_t)1000000 - ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 - ((uint64_t)highrttthresh * (uint64_t)1000000) / (uint64_t)rtt)) / 100)) /(uint64_t)1000000); return (perf); } static void rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff) { uint64_t logvar, logvar2, logvar3; uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val; if (rack->rc_gp_incr) { /* Turn off increment counting */ rack->rc_gp_incr = 0; rack->rc_gp_timely_inc_cnt = 0; } ss_red = ca_red = rec_red = 0; logged = 0; /* Calculate the reduction value */ if (rtt_diff < 0) { rtt_diff *= -1; } /* Must be at least 1% reduction */ if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) { /* We have been in recovery ding it too */ if (timely_says == 2) { new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt); alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff); if (alt < new_per) val = alt; else val = new_per; } else val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff); if (rack->r_ctl.rack_per_of_gp_rec > val) { rec_red = (rack->r_ctl.rack_per_of_gp_rec - val); rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val; } else { rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound; rec_red = 0; } if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec) rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound; logged |= 1; } if (rack->rc_gp_saw_ss) { /* Sent in SS */ if (timely_says == 2) { new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt); alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff); if (alt < new_per) val = alt; else val = new_per; } else val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff); if (rack->r_ctl.rack_per_of_gp_ss > new_per) { ss_red = rack->r_ctl.rack_per_of_gp_ss - val; rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val; } else { ss_red = new_per; rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound; logvar = new_per; logvar <<= 32; logvar |= alt; logvar2 = (uint32_t)rtt; logvar2 <<= 32; logvar2 |= (uint32_t)rtt_diff; logvar3 = rack_gp_rtt_maxmul; logvar3 <<= 32; logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt); rack_log_timely(rack, timely_says, logvar2, logvar3, logvar, __LINE__, 10); } if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss) rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound; logged |= 4; } else if (rack->rc_gp_saw_ca) { /* Sent in CA */ if (timely_says == 2) { new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt); alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff); if (alt < new_per) val = alt; else val = new_per; } else val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff); if (rack->r_ctl.rack_per_of_gp_ca > val) { ca_red = rack->r_ctl.rack_per_of_gp_ca - val; rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val; } else { rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound; ca_red = 0; logvar = new_per; logvar <<= 32; logvar |= alt; logvar2 = (uint32_t)rtt; logvar2 <<= 32; logvar2 |= (uint32_t)rtt_diff; logvar3 = rack_gp_rtt_maxmul; logvar3 <<= 32; logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt); rack_log_timely(rack, timely_says, logvar2, logvar3, logvar, __LINE__, 10); } if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca) rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound; logged |= 2; } if (rack->rc_gp_timely_dec_cnt < 0x7) { rack->rc_gp_timely_dec_cnt++; if (rack_timely_dec_clear && (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear)) rack->rc_gp_timely_dec_cnt = 0; } logvar = ss_red; logvar <<= 32; logvar |= ca_red; rack_log_timely(rack, logged, rec_red, rack_per_lower_bound, logvar, __LINE__, 2); } static void rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts, uint32_t rtt, uint32_t line, uint8_t reas) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = line; log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts; log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts; log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss; log.u_bbr.flex5 = rtt; log.u_bbr.flex6 = rack->rc_highly_buffered; log.u_bbr.flex6 <<= 1; log.u_bbr.flex6 |= rack->forced_ack; log.u_bbr.flex6 <<= 1; log.u_bbr.flex6 |= rack->rc_gp_dyn_mul; log.u_bbr.flex6 <<= 1; log.u_bbr.flex6 |= rack->in_probe_rtt; log.u_bbr.flex6 <<= 1; log.u_bbr.flex6 |= rack->measure_saw_probe_rtt; log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt; log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca; log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec; log.u_bbr.flex8 = reas; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.delRate = rack_get_bw(rack); log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt; log.u_bbr.cur_del_rate <<= 32; log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt; log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered; log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt; log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt; log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts; log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight; log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt); log.u_bbr.rttProp = us_cts; log.u_bbr.rttProp <<= 32; log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_RTT_SHRINKS, 0, 0, &log, false, &rack->r_ctl.act_rcv_time); } } static void rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt) { uint64_t bwdp; bwdp = rack_get_bw(rack); bwdp *= (uint64_t)rtt; bwdp /= (uint64_t)HPTS_USEC_IN_SEC; rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz); if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) { /* * A window protocol must be able to have 4 packets * outstanding as the floor in order to function * (especially considering delayed ack :D). */ rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs); } } static void rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts) { /** * ProbeRTT is a bit different in rack_pacing than in * BBR. It is like BBR in that it uses the lowering of * the RTT as a signal that we saw something new and * counts from there for how long between. But it is * different in that its quite simple. It does not * play with the cwnd and wait until we get down * to N segments outstanding and hold that for * 200ms. Instead it just sets the pacing reduction * rate to a set percentage (70 by default) and hold * that for a number of recent GP Srtt's. */ uint32_t segsiz; if (rack->rc_gp_dyn_mul == 0) return; if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) { /* We are idle */ return; } if ((rack->rc_tp->t_flags & TF_GPUTINPROG) && SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) { /* * Stop the goodput now, the idea here is * that future measurements with in_probe_rtt * won't register if they are not greater so * we want to get what info (if any) is available * now. */ rack_do_goodput_measurement(rack->rc_tp, rack, rack->rc_tp->snd_una, __LINE__); } rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt; rack->r_ctl.rc_time_probertt_entered = us_cts; segsiz = min(ctf_fixed_maxseg(rack->rc_tp), rack->r_ctl.rc_pace_min_segs); rack->in_probe_rtt = 1; rack->measure_saw_probe_rtt = 1; rack->r_ctl.rc_lower_rtt_us_cts = us_cts; rack->r_ctl.rc_time_probertt_starts = 0; rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt; if (rack_probertt_use_min_rtt_entry) rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)); else rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt); rack_log_rtt_shrinks(rack, us_cts, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), __LINE__, RACK_RTTS_ENTERPROBE); } static void rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts) { struct rack_sendmap *rsm; uint32_t segsiz; segsiz = min(ctf_fixed_maxseg(rack->rc_tp), rack->r_ctl.rc_pace_min_segs); rack->in_probe_rtt = 0; if ((rack->rc_tp->t_flags & TF_GPUTINPROG) && SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) { /* * Stop the goodput now, the idea here is * that future measurements with in_probe_rtt * won't register if they are not greater so * we want to get what info (if any) is available * now. */ rack_do_goodput_measurement(rack->rc_tp, rack, rack->rc_tp->snd_una, __LINE__); } else if (rack->rc_tp->t_flags & TF_GPUTINPROG) { /* * We don't have enough data to make a measurement. * So lets just stop and start here after exiting * probe-rtt. We probably are not interested in * the results anyway. */ rack->rc_tp->t_flags &= ~TF_GPUTINPROG; } /* * Measurements through the current snd_max are going * to be limited by the slower pacing rate. * * We need to mark these as app-limited so we * don't collapse the b/w. */ rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) { if (rack->r_ctl.rc_app_limited_cnt == 0) rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm; else { /* * Go out to the end app limited and mark * this new one as next and move the end_appl up * to this guy. */ if (rack->r_ctl.rc_end_appl) rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start; rack->r_ctl.rc_end_appl = rsm; } rsm->r_flags |= RACK_APP_LIMITED; rack->r_ctl.rc_app_limited_cnt++; } /* * Now, we need to examine our pacing rate multipliers. * If its under 100%, we need to kick it back up to * 100%. We also don't let it be over our "max" above * the actual rate i.e. 100% + rack_clamp_atexit_prtt. * Note setting clamp_atexit_prtt to 0 has the effect * of setting CA/SS to 100% always at exit (which is * the default behavior). */ if (rack_probertt_clear_is) { rack->rc_gp_incr = 0; rack->rc_gp_bwred = 0; rack->rc_gp_timely_inc_cnt = 0; rack->rc_gp_timely_dec_cnt = 0; } /* Do we do any clamping at exit? */ if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) { rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp; rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp; } if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) { rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt; rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt; } /* * Lets set rtt_diff to 0, so that we will get a "boost" * after exiting. */ rack->r_ctl.rc_rtt_diff = 0; /* Clear all flags so we start fresh */ rack->rc_tp->t_bytes_acked = 0; rack->rc_tp->ccv->flags &= ~CCF_ABC_SENTAWND; /* * If configured to, set the cwnd and ssthresh to * our targets. */ if (rack_probe_rtt_sets_cwnd) { uint64_t ebdp; uint32_t setto; /* Set ssthresh so we get into CA once we hit our target */ if (rack_probertt_use_min_rtt_exit == 1) { /* Set to min rtt */ rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)); } else if (rack_probertt_use_min_rtt_exit == 2) { /* Set to current gp rtt */ rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt); } else if (rack_probertt_use_min_rtt_exit == 3) { /* Set to entry gp rtt */ rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_entry_gp_rtt); } else { uint64_t sum; uint32_t setval; sum = rack->r_ctl.rc_entry_gp_rtt; sum *= 10; sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt)); if (sum >= 20) { /* * A highly buffered path needs * cwnd space for timely to work. * Lets set things up as if * we are heading back here again. */ setval = rack->r_ctl.rc_entry_gp_rtt; } else if (sum >= 15) { /* * Lets take the smaller of the * two since we are just somewhat * buffered. */ setval = rack->r_ctl.rc_gp_srtt; if (setval > rack->r_ctl.rc_entry_gp_rtt) setval = rack->r_ctl.rc_entry_gp_rtt; } else { /* * Here we are not highly buffered * and should pick the min we can to * keep from causing loss. */ setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt); } rack_set_prtt_target(rack, segsiz, setval); } if (rack_probe_rtt_sets_cwnd > 1) { /* There is a percentage here to boost */ ebdp = rack->r_ctl.rc_target_probertt_flight; ebdp *= rack_probe_rtt_sets_cwnd; ebdp /= 100; setto = rack->r_ctl.rc_target_probertt_flight + ebdp; } else setto = rack->r_ctl.rc_target_probertt_flight; rack->rc_tp->snd_cwnd = roundup(setto, segsiz); if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) { /* Enforce a min */ rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs; } /* If we set in the cwnd also set the ssthresh point so we are in CA */ rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1); } rack_log_rtt_shrinks(rack, us_cts, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), __LINE__, RACK_RTTS_EXITPROBE); /* Clear times last so log has all the info */ rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max; rack->r_ctl.rc_time_probertt_entered = us_cts; rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts; rack->r_ctl.rc_time_of_last_probertt = us_cts; } static void rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts) { /* Check in on probe-rtt */ if (rack->rc_gp_filled == 0) { /* We do not do p-rtt unless we have gp measurements */ return; } if (rack->in_probe_rtt) { uint64_t no_overflow; uint32_t endtime, must_stay; if (rack->r_ctl.rc_went_idle_time && ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) { /* * We went idle during prtt, just exit now. */ rack_exit_probertt(rack, us_cts); } else if (rack_probe_rtt_safety_val && TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) && ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) { /* * Probe RTT safety value triggered! */ rack_log_rtt_shrinks(rack, us_cts, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), __LINE__, RACK_RTTS_SAFETY); rack_exit_probertt(rack, us_cts); } /* Calculate the max we will wait */ endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait); if (rack->rc_highly_buffered) endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp); /* Calculate the min we must wait */ must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain); if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) && TSTMP_LT(us_cts, endtime)) { uint32_t calc; /* Do we lower more? */ no_exit: if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered)) calc = us_cts - rack->r_ctl.rc_time_probertt_entered; else calc = 0; calc /= max(rack->r_ctl.rc_gp_srtt, 1); if (calc) { /* Maybe */ calc *= rack_per_of_gp_probertt_reduce; rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc; /* Limit it too */ if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh) rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh; } /* We must reach target or the time set */ return; } if (rack->r_ctl.rc_time_probertt_starts == 0) { if ((TSTMP_LT(us_cts, must_stay) && rack->rc_highly_buffered) || (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight)) { /* We are not past the must_stay time */ goto no_exit; } rack_log_rtt_shrinks(rack, us_cts, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), __LINE__, RACK_RTTS_REACHTARGET); rack->r_ctl.rc_time_probertt_starts = us_cts; if (rack->r_ctl.rc_time_probertt_starts == 0) rack->r_ctl.rc_time_probertt_starts = 1; /* Restore back to our rate we want to pace at in prtt */ rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt; } /* * Setup our end time, some number of gp_srtts plus 200ms. */ no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt * (uint64_t)rack_probertt_gpsrtt_cnt_mul); if (rack_probertt_gpsrtt_cnt_div) endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div); else endtime = 0; endtime += rack_min_probertt_hold; endtime += rack->r_ctl.rc_time_probertt_starts; if (TSTMP_GEQ(us_cts, endtime)) { /* yes, exit probertt */ rack_exit_probertt(rack, us_cts); } } else if((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) { /* Go into probertt, its been too long since we went lower */ rack_enter_probertt(rack, us_cts); } } static void rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est, uint32_t rtt, int32_t rtt_diff) { uint64_t cur_bw, up_bnd, low_bnd, subfr; uint32_t losses; if ((rack->rc_gp_dyn_mul == 0) || (rack->use_fixed_rate) || (rack->in_probe_rtt) || (rack->rc_always_pace == 0)) { /* No dynamic GP multipler in play */ return; } losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start; cur_bw = rack_get_bw(rack); /* Calculate our up and down range */ up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up; up_bnd /= 100; up_bnd += rack->r_ctl.last_gp_comp_bw; subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down; subfr /= 100; low_bnd = rack->r_ctl.last_gp_comp_bw - subfr; if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) { /* * This is the case where our RTT is above * the max target and we have been configured * to just do timely no bonus up stuff in that case. * * There are two configurations, set to 1, and we * just do timely if we are over our max. If its * set above 1 then we slam the multipliers down * to 100 and then decrement per timely. */ rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd, __LINE__, 3); if (rack->r_ctl.rc_no_push_at_mrtt > 1) rack_validate_multipliers_at_or_below_100(rack); rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff); } else if ((last_bw_est < low_bnd) && !losses) { /* * We are decreasing this is a bit complicated this * means we are loosing ground. This could be * because another flow entered and we are competing * for b/w with it. This will push the RTT up which * makes timely unusable unless we want to get shoved * into a corner and just be backed off (the age * old problem with delay based CC). * * On the other hand if it was a route change we * would like to stay somewhat contained and not * blow out the buffers. */ rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd, __LINE__, 3); rack->r_ctl.last_gp_comp_bw = cur_bw; if (rack->rc_gp_bwred == 0) { /* Go into reduction counting */ rack->rc_gp_bwred = 1; rack->rc_gp_timely_dec_cnt = 0; } if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) || (timely_says == 0)) { /* * Push another time with a faster pacing * to try to gain back (we include override to * get a full raise factor). */ if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) || (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) || (timely_says == 0) || (rack_down_raise_thresh == 0)) { /* * Do an override up in b/w if we were * below the threshold or if the threshold * is zero we always do the raise. */ rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1); } else { /* Log it stays the same */ rack_log_timely(rack, 0, last_bw_est, low_bnd, 0, __LINE__, 11); } rack->rc_gp_timely_dec_cnt++; /* We are not incrementing really no-count */ rack->rc_gp_incr = 0; rack->rc_gp_timely_inc_cnt = 0; } else { /* * Lets just use the RTT * information and give up * pushing. */ goto use_timely; } } else if ((timely_says != 2) && !losses && (last_bw_est > up_bnd)) { /* * We are increasing b/w lets keep going, updating * our b/w and ignoring any timely input, unless * of course we are at our max raise (if there is one). */ rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd, __LINE__, 3); rack->r_ctl.last_gp_comp_bw = cur_bw; if (rack->rc_gp_saw_ss && rack_per_upper_bound_ss && (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) { /* * In cases where we can't go higher * we should just use timely. */ goto use_timely; } if (rack->rc_gp_saw_ca && rack_per_upper_bound_ca && (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) { /* * In cases where we can't go higher * we should just use timely. */ goto use_timely; } rack->rc_gp_bwred = 0; rack->rc_gp_timely_dec_cnt = 0; /* You get a set number of pushes if timely is trying to reduce */ if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) { rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0); } else { /* Log it stays the same */ rack_log_timely(rack, 0, last_bw_est, up_bnd, 0, __LINE__, 12); } return; } else { /* * We are staying between the lower and upper range bounds * so use timely to decide. */ rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd, __LINE__, 3); use_timely: if (timely_says) { rack->rc_gp_incr = 0; rack->rc_gp_timely_inc_cnt = 0; if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) && !losses && (last_bw_est < low_bnd)) { /* We are loosing ground */ rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0); rack->rc_gp_timely_dec_cnt++; /* We are not incrementing really no-count */ rack->rc_gp_incr = 0; rack->rc_gp_timely_inc_cnt = 0; } else rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff); } else { rack->rc_gp_bwred = 0; rack->rc_gp_timely_dec_cnt = 0; rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0); } } } static int32_t rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt) { int32_t timely_says; uint64_t log_mult, log_rtt_a_diff; log_rtt_a_diff = rtt; log_rtt_a_diff <<= 32; log_rtt_a_diff |= (uint32_t)rtt_diff; if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul)) { /* Reduce the b/w multipler */ timely_says = 2; log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul; log_mult <<= 32; log_mult |= prev_rtt; rack_log_timely(rack, timely_says, log_mult, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 4); } else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) + ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) / max(rack_gp_rtt_mindiv , 1)))) { /* Increase the b/w multipler */ log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) + ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) / max(rack_gp_rtt_mindiv , 1)); log_mult <<= 32; log_mult |= prev_rtt; timely_says = 0; rack_log_timely(rack, timely_says, log_mult , get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 5); } else { /* * Use a gradient to find it the timely gradient * is: * grad = rc_rtt_diff / min_rtt; * * anything below or equal to 0 will be * a increase indication. Anything above * zero is a decrease. Note we take care * of the actual gradient calculation * in the reduction (its not needed for * increase). */ log_mult = prev_rtt; if (rtt_diff <= 0) { /* * Rttdiff is less than zero, increase the * b/w multipler (its 0 or negative) */ timely_says = 0; rack_log_timely(rack, timely_says, log_mult, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6); } else { /* Reduce the b/w multipler */ timely_says = 1; rack_log_timely(rack, timely_says, log_mult, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7); } } return (timely_says); } static void rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, int line) { uint64_t tim, bytes_ps, ltim, stim, utim; uint32_t segsiz, bytes, reqbytes, us_cts; int32_t gput, new_rtt_diff, timely_says; us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); if (TSTMP_GEQ(us_cts, tp->gput_ts)) tim = us_cts - tp->gput_ts; else tim = 0; if (TSTMP_GT(rack->r_ctl.rc_gp_cumack_ts, rack->r_ctl.rc_gp_output_ts)) stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts; else stim = 0; /* * Use the larger of the send time or ack time. This prevents us * from being influenced by ack artifacts to come up with too * high of measurement. Note that since we are spanning over many more * bytes in most of our measurements hopefully that is less likely to * occur. */ if (tim > stim) utim = max(tim, 1); else utim = max(stim, 1); /* Lets validate utim */ ltim = max(1, (utim/HPTS_USEC_IN_MSEC)); gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim; reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz)); if ((tim == 0) && (stim == 0)) { /* * Invalid measurement time, maybe * all on one ack/one send? */ bytes = 0; bytes_ps = 0; rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes, 0, 0, 0, 10, __LINE__, NULL); goto skip_measurement; } if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) { /* We never made a us_rtt measurement? */ bytes = 0; bytes_ps = 0; rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes, 0, 0, 0, 10, __LINE__, NULL); goto skip_measurement; } /* * Calculate the maximum possible b/w this connection * could have. We base our calculation on the lowest * rtt we have seen during the measurement and the * largest rwnd the client has given us in that time. This * forms a BDP that is the maximum that we could ever * get to the client. Anything larger is not valid. * * I originally had code here that rejected measurements * where the time was less than 1/2 the latest us_rtt. * But after thinking on that I realized its wrong since * say you had a 150Mbps or even 1Gbps link, and you * were a long way away.. example I am in Europe (100ms rtt) * talking to my 1Gbps link in S.C. Now measuring say 150,000 * bytes my time would be 1.2ms, and yet my rtt would say * the measurement was invalid the time was < 50ms. The * same thing is true for 150Mb (8ms of time). * * A better way I realized is to look at what the maximum * the connection could possibly do. This is gated on * the lowest RTT we have seen and the highest rwnd. * We should in theory never exceed that, if we are * then something on the path is storing up packets * and then feeding them all at once to our endpoint * messing up our measurement. */ rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd; rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC; rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt; if (SEQ_LT(th_ack, tp->gput_seq)) { /* No measurement can be made */ bytes = 0; bytes_ps = 0; rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes, 0, 0, 0, 10, __LINE__, NULL); goto skip_measurement; } else bytes = (th_ack - tp->gput_seq); bytes_ps = (uint64_t)bytes; /* * Don't measure a b/w for pacing unless we have gotten at least * an initial windows worth of data in this measurement interval. * * Small numbers of bytes get badly influenced by delayed ack and * other artifacts. Note we take the initial window or our * defined minimum GP (defaulting to 10 which hopefully is the * IW). */ if (rack->rc_gp_filled == 0) { /* * The initial estimate is special. We * have blasted out an IW worth of packets * without a real valid ack ts results. We * then setup the app_limited_needs_set flag, * this should get the first ack in (probably 2 * MSS worth) to be recorded as the timestamp. * We thus allow a smaller number of bytes i.e. * IW - 2MSS. */ reqbytes -= (2 * segsiz); /* Also lets fill previous for our first measurement to be neutral */ rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt; } if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) { rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes, rack->r_ctl.rc_app_limited_cnt, 0, 0, 10, __LINE__, NULL); goto skip_measurement; } /* * We now need to calculate the Timely like status so * we can update (possibly) the b/w multipliers. */ new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt; if (rack->rc_gp_filled == 0) { /* No previous reading */ rack->r_ctl.rc_rtt_diff = new_rtt_diff; } else { if (rack->measure_saw_probe_rtt == 0) { /* * We don't want a probertt to be counted * since it will be negative incorrectly. We * expect to be reducing the RTT when we * pace at a slower rate. */ rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8); rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8); } } timely_says = rack_make_timely_judgement(rack, rack->r_ctl.rc_gp_srtt, rack->r_ctl.rc_rtt_diff, rack->r_ctl.rc_prev_gp_srtt ); bytes_ps *= HPTS_USEC_IN_SEC; bytes_ps /= utim; if (bytes_ps > rack->r_ctl.last_max_bw) { /* * Something is on path playing * since this b/w is not possible based * on our BDP (highest rwnd and lowest rtt * we saw in the measurement window). * * Another option here would be to * instead skip the measurement. */ rack_log_pacing_delay_calc(rack, bytes, reqbytes, bytes_ps, rack->r_ctl.last_max_bw, 0, 11, __LINE__, NULL); bytes_ps = rack->r_ctl.last_max_bw; } /* We store gp for b/w in bytes per second */ if (rack->rc_gp_filled == 0) { /* Initial measurment */ if (bytes_ps) { rack->r_ctl.gp_bw = bytes_ps; rack->rc_gp_filled = 1; rack->r_ctl.num_avg = 1; rack_set_pace_segments(rack->rc_tp, rack, __LINE__); } else { rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes, rack->r_ctl.rc_app_limited_cnt, 0, 0, 10, __LINE__, NULL); } if (rack->rc_inp->inp_in_hpts && (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { /* * Ok we can't trust the pacer in this case * where we transition from un-paced to paced. * Or for that matter when the burst mitigation * was making a wild guess and got it wrong. * Stop the pacer and clear up all the aggregate * delays etc. */ tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); rack->r_ctl.rc_hpts_flags = 0; rack->r_ctl.rc_last_output_to = 0; } } else if (rack->r_ctl.num_avg < RACK_REQ_AVG) { /* Still a small number run an average */ rack->r_ctl.gp_bw += bytes_ps; rack->r_ctl.num_avg++; if (rack->r_ctl.num_avg >= RACK_REQ_AVG) { /* We have collected enought to move forward */ rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_avg; } } else { /* * We want to take 1/wma of the goodput and add in to 7/8th * of the old value weighted by the srtt. So if your measurement * period is say 2 SRTT's long you would get 1/4 as the * value, if it was like 1/2 SRTT then you would get 1/16th. * * But we must be careful not to take too much i.e. if the * srtt is say 20ms and the measurement is taken over * 400ms our weight would be 400/20 i.e. 20. On the * other hand if we get a measurement over 1ms with a * 10ms rtt we only want to take a much smaller portion. */ uint64_t resid_bw, subpart, addpart, srtt; srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); if (srtt == 0) { /* * Strange why did t_srtt go back to zero? */ if (rack->r_ctl.rc_rack_min_rtt) srtt = (rack->r_ctl.rc_rack_min_rtt * HPTS_USEC_IN_MSEC); else srtt = HPTS_USEC_IN_MSEC; } /* * XXXrrs: Note for reviewers, in playing with * dynamic pacing I discovered this GP calculation * as done originally leads to some undesired results. * Basically you can get longer measurements contributing * too much to the WMA. Thus I changed it if you are doing * dynamic adjustments to only do the aportioned adjustment * if we have a very small (time wise) measurement. Longer * measurements just get there weight (defaulting to 1/8) * add to the WMA. We may want to think about changing * this to always do that for both sides i.e. dynamic * and non-dynamic... but considering lots of folks * were playing with this I did not want to change the * calculation per.se. without your thoughts.. Lawerence? * Peter?? */ if (rack->rc_gp_dyn_mul == 0) { subpart = rack->r_ctl.gp_bw * utim; subpart /= (srtt * 8); if (subpart < (rack->r_ctl.gp_bw / 2)) { /* * The b/w update takes no more * away then 1/2 our running total * so factor it in. */ addpart = bytes_ps * utim; addpart /= (srtt * 8); } else { /* * Don't allow a single measurement * to account for more than 1/2 of the * WMA. This could happen on a retransmission * where utim becomes huge compared to * srtt (multiple retransmissions when using * the sending rate which factors in all the * transmissions from the first one). */ subpart = rack->r_ctl.gp_bw / 2; addpart = bytes_ps / 2; } resid_bw = rack->r_ctl.gp_bw - subpart; rack->r_ctl.gp_bw = resid_bw + addpart; } else { if ((utim / srtt) <= 1) { /* * The b/w update was over a small period * of time. The idea here is to prevent a small * measurement time period from counting * too much. So we scale it based on the * time so it attributes less than 1/rack_wma_divisor * of its measurement. */ subpart = rack->r_ctl.gp_bw * utim; subpart /= (srtt * rack_wma_divisor); addpart = bytes_ps * utim; addpart /= (srtt * rack_wma_divisor); } else { /* * The scaled measurement was long * enough so lets just add in the * portion of the measurment i.e. 1/rack_wma_divisor */ subpart = rack->r_ctl.gp_bw / rack_wma_divisor; addpart = bytes_ps / rack_wma_divisor; } if ((rack->measure_saw_probe_rtt == 0) || (bytes_ps > rack->r_ctl.gp_bw)) { /* * For probe-rtt we only add it in * if its larger, all others we just * add in. */ resid_bw = rack->r_ctl.gp_bw - subpart; rack->r_ctl.gp_bw = resid_bw + addpart; } } } /* We do not update any multipliers if we are in or have seen a probe-rtt */ if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set) rack_update_multiplier(rack, timely_says, bytes_ps, rack->r_ctl.rc_gp_srtt, rack->r_ctl.rc_rtt_diff); rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim, rack_get_bw(rack), 3, line, NULL); /* reset the gp srtt and setup the new prev */ rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt; /* Record the lost count for the next measurement */ rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count; /* * We restart our diffs based on the gpsrtt in the * measurement window. */ rack->rc_gp_rtt_set = 0; rack->rc_gp_saw_rec = 0; rack->rc_gp_saw_ca = 0; rack->rc_gp_saw_ss = 0; rack->rc_dragged_bottom = 0; skip_measurement: #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT, gput); /* * XXXLAS: This is a temporary hack, and should be * chained off VOI_TCP_GPUT when stats(9) grows an * API to deal with chained VOIs. */ if (tp->t_stats_gput_prev > 0) stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_GPUT_ND, ((gput - tp->t_stats_gput_prev) * 100) / tp->t_stats_gput_prev); #endif tp->t_flags &= ~TF_GPUTINPROG; tp->t_stats_gput_prev = gput; /* * Now are we app limited now and there is space from where we * were to where we want to go? * * We don't do the other case i.e. non-applimited here since * the next send will trigger us picking up the missing data. */ if (rack->r_ctl.rc_first_appl && TCPS_HAVEESTABLISHED(tp->t_state) && rack->r_ctl.rc_app_limited_cnt && (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) && ((rack->r_ctl.rc_first_appl->r_start - th_ack) > max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) { /* * Yep there is enough outstanding to make a measurement here. */ struct rack_sendmap *rsm, fe; tp->t_flags |= TF_GPUTINPROG; rack->r_ctl.rc_gp_lowrtt = 0xffffffff; rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd; tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); rack->app_limited_needs_set = 0; tp->gput_seq = th_ack; if (rack->in_probe_rtt) rack->measure_saw_probe_rtt = 1; else if ((rack->measure_saw_probe_rtt) && (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit))) rack->measure_saw_probe_rtt = 0; if ((rack->r_ctl.rc_first_appl->r_start - th_ack) >= rack_get_measure_window(tp, rack)) { /* There is a full window to gain info from */ tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack); } else { /* We can only measure up to the applimited point */ tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_start - th_ack); } /* * Now we need to find the timestamp of the send at tp->gput_seq * for the send based measurement. */ fe.r_start = tp->gput_seq; rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); if (rsm) { /* Ok send-based limit is set */ if (SEQ_LT(rsm->r_start, tp->gput_seq)) { /* * Move back to include the earlier part * so our ack time lines up right (this may * make an overlapping measurement but thats * ok). */ tp->gput_seq = rsm->r_start; } if (rsm->r_flags & RACK_ACKED) tp->gput_ts = rsm->r_ack_arrival; else rack->app_limited_needs_set = 1; rack->r_ctl.rc_gp_output_ts = rsm->usec_orig_send; } else { /* * If we don't find the rsm due to some * send-limit set the current time, which * basically disables the send-limit. */ rack->r_ctl.rc_gp_output_ts = tcp_get_usecs(NULL); } rack_log_pacing_delay_calc(rack, tp->gput_seq, tp->gput_ack, (uint64_t)rsm, tp->gput_ts, rack->r_ctl.rc_app_limited_cnt, 9, __LINE__, NULL); } } /* * CC wrapper hook functions */ static void rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, struct tcphdr *th, uint16_t nsegs, uint16_t type, int32_t recovery) { INP_WLOCK_ASSERT(tp->t_inpcb); tp->ccv->nsegs = nsegs; tp->ccv->bytes_this_ack = BYTES_THIS_ACK(tp, th); if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) { uint32_t max; max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp); if (tp->ccv->bytes_this_ack > max) { tp->ccv->bytes_this_ack = max; } } if (rack->r_ctl.cwnd_to_use <= tp->snd_wnd) tp->ccv->flags |= CCF_CWND_LIMITED; else tp->ccv->flags &= ~CCF_CWND_LIMITED; #ifdef STATS stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF, ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd); #endif if ((tp->t_flags & TF_GPUTINPROG) && rack_enough_for_measurement(tp, rack, th->th_ack)) { /* Measure the Goodput */ rack_do_goodput_measurement(tp, rack, th->th_ack, __LINE__); #ifdef NETFLIX_PEAKRATE if ((type == CC_ACK) && (tp->t_maxpeakrate)) { /* * We update t_peakrate_thr. This gives us roughly * one update per round trip time. Note * it will only be used if pace_always is off i.e * we don't do this for paced flows. */ tcp_update_peakrate_thr(tp); } #endif } if (rack->r_ctl.cwnd_to_use > tp->snd_ssthresh) { tp->t_bytes_acked += tp->ccv->bytes_this_ack; if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) { tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use; tp->ccv->flags |= CCF_ABC_SENTAWND; } } else { tp->ccv->flags &= ~CCF_ABC_SENTAWND; tp->t_bytes_acked = 0; } if (CC_ALGO(tp)->ack_received != NULL) { /* XXXLAS: Find a way to live without this */ tp->ccv->curack = th->th_ack; CC_ALGO(tp)->ack_received(tp->ccv, type); } #ifdef STATS stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use); #endif if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) { rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use; } #ifdef NETFLIX_PEAKRATE /* we enforce max peak rate if it is set and we are not pacing */ if ((rack->rc_always_pace == 0) && tp->t_peakrate_thr && (tp->snd_cwnd > tp->t_peakrate_thr)) { tp->snd_cwnd = tp->t_peakrate_thr; } #endif } static void tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th) { struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; INP_WLOCK_ASSERT(tp->t_inpcb); /* * If we are doing PRR and have enough * room to send we are pacing and prr * is disabled we will want to see if we * can send data (by setting r_wanted_output to * true). */ if ((rack->r_ctl.rc_prr_sndcnt > 0) || rack->rack_no_prr) rack->r_wanted_output = 1; } static void rack_post_recovery(struct tcpcb *tp, struct tcphdr *th) { struct tcp_rack *rack; uint32_t orig_cwnd; orig_cwnd = tp->snd_cwnd; INP_WLOCK_ASSERT(tp->t_inpcb); rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack->rc_not_backing_off == 0) { /* only alert CC if we alerted when we entered */ if (CC_ALGO(tp)->post_recovery != NULL) { tp->ccv->curack = th->th_ack; CC_ALGO(tp)->post_recovery(tp->ccv); } if (tp->snd_cwnd > tp->snd_ssthresh) { /* Drop us down to the ssthresh (1/2 cwnd at loss) */ tp->snd_cwnd = tp->snd_ssthresh; } } if ((rack->rack_no_prr == 0) && (rack->r_ctl.rc_prr_sndcnt > 0)) { /* Suck the next prr cnt back into cwnd */ tp->snd_cwnd += rack->r_ctl.rc_prr_sndcnt; rack->r_ctl.rc_prr_sndcnt = 0; rack_log_to_prr(rack, 1, 0); } rack_log_to_prr(rack, 14, orig_cwnd); tp->snd_recover = tp->snd_una; EXIT_RECOVERY(tp->t_flags); } static void rack_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type) { struct tcp_rack *rack; INP_WLOCK_ASSERT(tp->t_inpcb); rack = (struct tcp_rack *)tp->t_fb_ptr; switch (type) { case CC_NDUPACK: tp->t_flags &= ~TF_WASFRECOVERY; tp->t_flags &= ~TF_WASCRECOVERY; if (!IN_FASTRECOVERY(tp->t_flags)) { rack->r_ctl.rc_prr_delivered = 0; rack->r_ctl.rc_prr_out = 0; if (rack->rack_no_prr == 0) { rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp); rack_log_to_prr(rack, 2, 0); } rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una; tp->snd_recover = tp->snd_max; if (tp->t_flags2 & TF2_ECN_PERMIT) tp->t_flags2 |= TF2_ECN_SND_CWR; } break; case CC_ECN: if (!IN_CONGRECOVERY(tp->t_flags) || /* * Allow ECN reaction on ACK to CWR, if * that data segment was also CE marked. */ SEQ_GEQ(th->th_ack, tp->snd_recover)) { EXIT_CONGRECOVERY(tp->t_flags); KMOD_TCPSTAT_INC(tcps_ecn_rcwnd); tp->snd_recover = tp->snd_max + 1; if (tp->t_flags2 & TF2_ECN_PERMIT) tp->t_flags2 |= TF2_ECN_SND_CWR; } break; case CC_RTO: tp->t_dupacks = 0; tp->t_bytes_acked = 0; EXIT_RECOVERY(tp->t_flags); tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 / ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp); tp->snd_cwnd = ctf_fixed_maxseg(tp); if (tp->t_flags2 & TF2_ECN_PERMIT) tp->t_flags2 |= TF2_ECN_SND_CWR; break; case CC_RTO_ERR: KMOD_TCPSTAT_INC(tcps_sndrexmitbad); /* RTO was unnecessary, so reset everything. */ tp->snd_cwnd = tp->snd_cwnd_prev; tp->snd_ssthresh = tp->snd_ssthresh_prev; tp->snd_recover = tp->snd_recover_prev; if (tp->t_flags & TF_WASFRECOVERY) { ENTER_FASTRECOVERY(tp->t_flags); tp->t_flags &= ~TF_WASFRECOVERY; } if (tp->t_flags & TF_WASCRECOVERY) { ENTER_CONGRECOVERY(tp->t_flags); tp->t_flags &= ~TF_WASCRECOVERY; } tp->snd_nxt = tp->snd_max; tp->t_badrxtwin = 0; break; } /* * If we are below our max rtt, don't * signal the CC control to change things. * instead set it up so that we are in * recovery but not going to back off. */ if (rack->rc_highly_buffered) { /* * Do we use the higher rtt for * our threshold to not backoff (like CDG)? */ uint32_t rtt_mul, rtt_div; if (rack_use_max_for_nobackoff) { rtt_mul = (rack_gp_rtt_maxmul - 1); rtt_div = 1; } else { rtt_mul = rack_gp_rtt_minmul; rtt_div = max(rack_gp_rtt_mindiv , 1); } if (rack->r_ctl.rc_gp_srtt <= (rack->r_ctl.rc_lowest_us_rtt + ((rack->r_ctl.rc_lowest_us_rtt * rtt_mul) / rtt_div))) { /* below our min threshold */ rack->rc_not_backing_off = 1; ENTER_RECOVERY(rack->rc_tp->t_flags); rack_log_rtt_shrinks(rack, 0, rtt_mul, rtt_div, RACK_RTTS_NOBACKOFF); return; } } rack->rc_not_backing_off = 0; if (CC_ALGO(tp)->cong_signal != NULL) { if (th != NULL) tp->ccv->curack = th->th_ack; CC_ALGO(tp)->cong_signal(tp->ccv, type); } } static inline void rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp) { uint32_t i_cwnd; INP_WLOCK_ASSERT(tp->t_inpcb); #ifdef NETFLIX_STATS KMOD_TCPSTAT_INC(tcps_idle_restarts); if (tp->t_state == TCPS_ESTABLISHED) KMOD_TCPSTAT_INC(tcps_idle_estrestarts); #endif if (CC_ALGO(tp)->after_idle != NULL) CC_ALGO(tp)->after_idle(tp->ccv); if (tp->snd_cwnd == 1) i_cwnd = tp->t_maxseg; /* SYN(-ACK) lost */ else i_cwnd = rc_init_window(rack); /* * Being idle is no differnt than the initial window. If the cc * clamps it down below the initial window raise it to the initial * window. */ if (tp->snd_cwnd < i_cwnd) { tp->snd_cwnd = i_cwnd; } } /* * Indicate whether this ack should be delayed. We can delay the ack if * following conditions are met: * - There is no delayed ack timer in progress. * - Our last ack wasn't a 0-sized window. We never want to delay * the ack that opens up a 0-sized window. * - LRO wasn't used for this segment. We make sure by checking that the * segment size is not larger than the MSS. * - Delayed acks are enabled or this is a half-synchronized T/TCP * connection. */ #define DELAY_ACK(tp, tlen) \ (((tp->t_flags & TF_RXWIN0SENT) == 0) && \ ((tp->t_flags & TF_DELACK) == 0) && \ (tlen <= tp->t_maxseg) && \ (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN))) static struct rack_sendmap * rack_find_lowest_rsm(struct tcp_rack *rack) { struct rack_sendmap *rsm; /* * Walk the time-order transmitted list looking for an rsm that is * not acked. This will be the one that was sent the longest time * ago that is still outstanding. */ TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) { if (rsm->r_flags & RACK_ACKED) { continue; } goto finish; } finish: return (rsm); } static struct rack_sendmap * rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm) { struct rack_sendmap *prsm; /* * Walk the sequence order list backward until we hit and arrive at * the highest seq not acked. In theory when this is called it * should be the last segment (which it was not). */ counter_u64_add(rack_find_high, 1); prsm = rsm; RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) { if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) { continue; } return (prsm); } return (NULL); } static uint32_t rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts) { int32_t lro; uint32_t thresh; /* * lro is the flag we use to determine if we have seen reordering. * If it gets set we have seen reordering. The reorder logic either * works in one of two ways: * * If reorder-fade is configured, then we track the last time we saw * re-ordering occur. If we reach the point where enough time as * passed we no longer consider reordering has occuring. * * Or if reorder-face is 0, then once we see reordering we consider * the connection to alway be subject to reordering and just set lro * to 1. * * In the end if lro is non-zero we add the extra time for * reordering in. */ if (srtt == 0) srtt = 1; if (rack->r_ctl.rc_reorder_ts) { if (rack->r_ctl.rc_reorder_fade) { if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) { lro = cts - rack->r_ctl.rc_reorder_ts; if (lro == 0) { /* * No time as passed since the last * reorder, mark it as reordering. */ lro = 1; } } else { /* Negative time? */ lro = 0; } if (lro > rack->r_ctl.rc_reorder_fade) { /* Turn off reordering seen too */ rack->r_ctl.rc_reorder_ts = 0; lro = 0; } } else { /* Reodering does not fade */ lro = 1; } } else { lro = 0; } thresh = srtt + rack->r_ctl.rc_pkt_delay; if (lro) { /* It must be set, if not you get 1/4 rtt */ if (rack->r_ctl.rc_reorder_shift) thresh += (srtt >> rack->r_ctl.rc_reorder_shift); else thresh += (srtt >> 2); } else { thresh += 1; } /* We don't let the rack timeout be above a RTO */ if (thresh > TICKS_2_MSEC(rack->rc_tp->t_rxtcur)) { thresh = TICKS_2_MSEC(rack->rc_tp->t_rxtcur); } /* And we don't want it above the RTO max either */ if (thresh > rack_rto_max) { thresh = rack_rto_max; } return (thresh); } static uint32_t rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t srtt) { struct rack_sendmap *prsm; uint32_t thresh, len; int segsiz; if (srtt == 0) srtt = 1; if (rack->r_ctl.rc_tlp_threshold) thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold); else thresh = (srtt * 2); /* Get the previous sent packet, if any */ segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); counter_u64_add(rack_enter_tlp_calc, 1); len = rsm->r_end - rsm->r_start; if (rack->rack_tlp_threshold_use == TLP_USE_ID) { /* Exactly like the ID */ if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) { uint32_t alt_thresh; /* * Compensate for delayed-ack with the d-ack time. */ counter_u64_add(rack_used_tlpmethod, 1); alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; if (alt_thresh > thresh) thresh = alt_thresh; } } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) { /* 2.1 behavior */ prsm = TAILQ_PREV(rsm, rack_head, r_tnext); if (prsm && (len <= segsiz)) { /* * Two packets outstanding, thresh should be (2*srtt) + * possible inter-packet delay (if any). */ uint32_t inter_gap = 0; int idx, nidx; counter_u64_add(rack_used_tlpmethod, 1); idx = rsm->r_rtr_cnt - 1; nidx = prsm->r_rtr_cnt - 1; if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) { /* Yes it was sent later (or at the same time) */ inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx]; } thresh += inter_gap; } else if (len <= segsiz) { /* * Possibly compensate for delayed-ack. */ uint32_t alt_thresh; counter_u64_add(rack_used_tlpmethod2, 1); alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; if (alt_thresh > thresh) thresh = alt_thresh; } } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) { /* 2.2 behavior */ if (len <= segsiz) { uint32_t alt_thresh; /* * Compensate for delayed-ack with the d-ack time. */ counter_u64_add(rack_used_tlpmethod, 1); alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; if (alt_thresh > thresh) thresh = alt_thresh; } } /* Not above an RTO */ if (thresh > TICKS_2_MSEC(tp->t_rxtcur)) { thresh = TICKS_2_MSEC(tp->t_rxtcur); } /* Not above a RTO max */ if (thresh > rack_rto_max) { thresh = rack_rto_max; } /* Apply user supplied min TLP */ if (thresh < rack_tlp_min) { thresh = rack_tlp_min; } return (thresh); } static uint32_t rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack) { /* * We want the rack_rtt which is the * last rtt we measured. However if that * does not exist we fallback to the srtt (which * we probably will never do) and then as a last * resort we use RACK_INITIAL_RTO if no srtt is * yet set. */ if (rack->rc_rack_rtt) return(rack->rc_rack_rtt); else if (tp->t_srtt == 0) return(RACK_INITIAL_RTO); return (TICKS_2_MSEC(tp->t_srtt >> TCP_RTT_SHIFT)); } static struct rack_sendmap * rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused) { /* * Check to see that we don't need to fall into recovery. We will * need to do so if our oldest transmit is past the time we should * have had an ack. */ struct tcp_rack *rack; struct rack_sendmap *rsm; int32_t idx; uint32_t srtt, thresh; rack = (struct tcp_rack *)tp->t_fb_ptr; if (RB_EMPTY(&rack->r_ctl.rc_mtree)) { return (NULL); } rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rsm == NULL) return (NULL); if (rsm->r_flags & RACK_ACKED) { rsm = rack_find_lowest_rsm(rack); if (rsm == NULL) return (NULL); } idx = rsm->r_rtr_cnt - 1; srtt = rack_grab_rtt(tp, rack); thresh = rack_calc_thresh_rack(rack, srtt, tsused); if (TSTMP_LT(tsused, rsm->r_tim_lastsent[idx])) { return (NULL); } if ((tsused - rsm->r_tim_lastsent[idx]) < thresh) { return (NULL); } /* Ok if we reach here we are over-due and this guy can be sent */ if (IN_RECOVERY(tp->t_flags) == 0) { /* * For the one that enters us into recovery record undo * info. */ rack->r_ctl.rc_rsm_start = rsm->r_start; rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; } rack_cong_signal(tp, NULL, CC_NDUPACK); return (rsm); } static uint32_t rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack) { int32_t t; int32_t tt; uint32_t ret_val; t = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT) + ((tp->t_rttvar * 4) >> TCP_RTT_SHIFT)); TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], rack_persist_min, rack_persist_max); if (tp->t_rxtshift < TCP_MAXRXTSHIFT) tp->t_rxtshift++; rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT; ret_val = (uint32_t)tt; return (ret_val); } static uint32_t rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack) { /* * Start the FR timer, we do this based on getting the first one in * the rc_tmap. Note that if its NULL we must stop the timer. in all * events we need to stop the running timer (if its running) before * starting the new one. */ uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse; uint32_t srtt_cur; int32_t idx; int32_t is_tlp_timer = 0; struct rack_sendmap *rsm; if (rack->t_timers_stopped) { /* All timers have been stopped none are to run */ return (0); } if (rack->rc_in_persist) { /* We can't start any timer in persists */ return (rack_get_persists_timer_val(tp, rack)); } rack->rc_on_min_to = 0; if ((tp->t_state < TCPS_ESTABLISHED) || ((tp->t_flags & TF_SACK_PERMIT) == 0)) goto activate_rxt; rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if ((rsm == NULL) || sup_rack) { /* Nothing on the send map */ activate_rxt: time_since_sent = 0; rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rsm) { idx = rsm->r_rtr_cnt - 1; if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], rack->r_ctl.rc_tlp_rxt_last_time)) tstmp_touse = rsm->r_tim_lastsent[idx]; else tstmp_touse = rack->r_ctl.rc_tlp_rxt_last_time; if (TSTMP_GT(cts, tstmp_touse)) time_since_sent = cts - tstmp_touse; } if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) { rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT; to = TICKS_2_MSEC(tp->t_rxtcur); if (to > time_since_sent) to -= time_since_sent; else to = rack->r_ctl.rc_min_to; if (to == 0) to = 1; return (to); } return (0); } if (rsm->r_flags & RACK_ACKED) { rsm = rack_find_lowest_rsm(rack); if (rsm == NULL) { /* No lowest? */ goto activate_rxt; } } if (rack->sack_attack_disable) { /* * We don't want to do * any TLP's if you are an attacker. * Though if you are doing what * is expected you may still have * SACK-PASSED marks. */ goto activate_rxt; } /* Convert from ms to usecs */ if ((rsm->r_flags & RACK_SACK_PASSED) || (rsm->r_dupack >= DUP_ACK_THRESHOLD)) { if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) == 1) && (rsm->r_flags & RACK_HAS_FIN)) { /* * We don't start a rack timer if all we have is a * FIN outstanding. */ goto activate_rxt; } if ((rack->use_rack_rr == 0) && (IN_RECOVERY(tp->t_flags)) && (rack->rack_no_prr == 0) && (rack->r_ctl.rc_prr_sndcnt < ctf_fixed_maxseg(tp))) { /* * We are not cheating, in recovery and * not enough ack's to yet get our next * retransmission out. * * Note that classified attackers do not * get to use the rack-cheat. */ goto activate_tlp; } srtt = rack_grab_rtt(tp, rack); thresh = rack_calc_thresh_rack(rack, srtt, cts); idx = rsm->r_rtr_cnt - 1; exp = rsm->r_tim_lastsent[idx] + thresh; if (SEQ_GEQ(exp, cts)) { to = exp - cts; if (to < rack->r_ctl.rc_min_to) { to = rack->r_ctl.rc_min_to; if (rack->r_rr_config == 3) rack->rc_on_min_to = 1; } } else { to = rack->r_ctl.rc_min_to; if (rack->r_rr_config == 3) rack->rc_on_min_to = 1; } } else { /* Ok we need to do a TLP not RACK */ activate_tlp: if ((rack->rc_tlp_in_progress != 0) && (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) { /* * The previous send was a TLP and we have sent * N TLP's without sending new data. */ goto activate_rxt; } rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext); if (rsm == NULL) { /* We found no rsm to TLP with. */ goto activate_rxt; } if (rsm->r_flags & RACK_HAS_FIN) { /* If its a FIN we dont do TLP */ rsm = NULL; goto activate_rxt; } idx = rsm->r_rtr_cnt - 1; time_since_sent = 0; if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], rack->r_ctl.rc_tlp_rxt_last_time)) tstmp_touse = rsm->r_tim_lastsent[idx]; else tstmp_touse = rack->r_ctl.rc_tlp_rxt_last_time; if (TSTMP_GT(cts, tstmp_touse)) time_since_sent = cts - tstmp_touse; is_tlp_timer = 1; if (tp->t_srtt) { srtt_cur = (tp->t_srtt >> TCP_RTT_SHIFT); srtt = TICKS_2_MSEC(srtt_cur); } else srtt = RACK_INITIAL_RTO; /* * If the SRTT is not keeping up and the * rack RTT has spiked we want to use * the last RTT not the smoothed one. */ if (rack_tlp_use_greater && (srtt < rack_grab_rtt(tp, rack))) srtt = rack_grab_rtt(tp, rack); thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt); if (thresh > time_since_sent) to = thresh - time_since_sent; else { to = rack->r_ctl.rc_min_to; rack_log_alt_to_to_cancel(rack, thresh, /* flex1 */ time_since_sent, /* flex2 */ tstmp_touse, /* flex3 */ rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */ rsm->r_tim_lastsent[idx], srtt, idx, 99); } if (to > TCPTV_REXMTMAX) { /* * If the TLP time works out to larger than the max * RTO lets not do TLP.. just RTO. */ goto activate_rxt; } } if (is_tlp_timer == 0) { rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK; } else { rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP; } if (to == 0) to = 1; return (to); } static void rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { if (rack->rc_in_persist == 0) { if (tp->t_flags & TF_GPUTINPROG) { /* * Stop the goodput now, the calling of the * measurement function clears the flag. */ rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__); } #ifdef NETFLIX_SHARED_CWND if (rack->r_ctl.rc_scw) { tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index); rack->rack_scwnd_is_idle = 1; } #endif rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL); if (rack->r_ctl.rc_went_idle_time == 0) rack->r_ctl.rc_went_idle_time = 1; rack_timer_cancel(tp, rack, cts, __LINE__); tp->t_rxtshift = 0; rack->rc_in_persist = 1; } } static void rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { if (rack->rc_inp->inp_in_hpts) { tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); rack->r_ctl.rc_hpts_flags = 0; } #ifdef NETFLIX_SHARED_CWND if (rack->r_ctl.rc_scw) { tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index); rack->rack_scwnd_is_idle = 0; } #endif if (rack->rc_gp_dyn_mul && (rack->use_fixed_rate == 0) && (rack->rc_always_pace)) { /* * Do we count this as if a probe-rtt just * finished? */ uint32_t time_idle, idle_min; time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time; idle_min = rack_min_probertt_hold; if (rack_probertt_gpsrtt_cnt_div) { uint64_t extra; extra = (uint64_t)rack->r_ctl.rc_gp_srtt * (uint64_t)rack_probertt_gpsrtt_cnt_mul; extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div; idle_min += (uint32_t)extra; } if (time_idle >= idle_min) { /* Yes, we count it as a probe-rtt. */ uint32_t us_cts; us_cts = tcp_get_usecs(NULL); if (rack->in_probe_rtt == 0) { rack->r_ctl.rc_lower_rtt_us_cts = us_cts; rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts; rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts; rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts; } else { rack_exit_probertt(rack, us_cts); } } } rack->rc_in_persist = 0; rack->r_ctl.rc_went_idle_time = 0; tp->t_rxtshift = 0; rack->r_ctl.rc_agg_delayed = 0; rack->r_early = 0; rack->r_late = 0; rack->r_ctl.rc_agg_early = 0; } static void rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts, struct hpts_diag *diag, struct timeval *tv) { if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = diag->p_nxt_slot; log.u_bbr.flex2 = diag->p_cur_slot; log.u_bbr.flex3 = diag->slot_req; log.u_bbr.flex4 = diag->inp_hptsslot; log.u_bbr.flex5 = diag->slot_remaining; log.u_bbr.flex6 = diag->need_new_to; log.u_bbr.flex7 = diag->p_hpts_active; log.u_bbr.flex8 = diag->p_on_min_sleep; /* Hijack other fields as needed */ log.u_bbr.epoch = diag->have_slept; log.u_bbr.lt_epoch = diag->yet_to_sleep; log.u_bbr.pkts_out = diag->co_ret; log.u_bbr.applimited = diag->hpts_sleep_time; log.u_bbr.delivered = diag->p_prev_slot; log.u_bbr.inflight = diag->p_runningtick; log.u_bbr.bw_inuse = diag->wheel_tick; log.u_bbr.rttProp = diag->wheel_cts; log.u_bbr.timeStamp = cts; log.u_bbr.delRate = diag->maxticks; log.u_bbr.cur_del_rate = diag->p_curtick; log.u_bbr.cur_del_rate <<= 32; log.u_bbr.cur_del_rate |= diag->p_lasttick; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_HPTSDIAG, 0, 0, &log, false, tv); } } static void rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts, int32_t slot, uint32_t tot_len_this_send, int sup_rack) { struct hpts_diag diag; struct inpcb *inp; struct timeval tv; uint32_t delayed_ack = 0; uint32_t hpts_timeout; uint8_t stopped; uint32_t left = 0; uint32_t us_cts; inp = tp->t_inpcb; if ((tp->t_state == TCPS_CLOSED) || (tp->t_state == TCPS_LISTEN)) { return; } if (inp->inp_in_hpts) { /* Already on the pacer */ return; } stopped = rack->rc_tmr_stopped; if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) { left = rack->r_ctl.rc_timer_exp - cts; } rack->r_ctl.rc_timer_exp = 0; rack->r_ctl.rc_hpts_flags = 0; us_cts = tcp_get_usecs(&tv); /* Now early/late accounting */ if (rack->r_early) { /* * We have a early carry over set, * we can always add more time so we * can always make this compensation. */ slot += rack->r_ctl.rc_agg_early; rack->r_early = 0; rack->r_ctl.rc_agg_early = 0; } if (rack->r_late) { /* * This is harder, we can * compensate some but it * really depends on what * the current pacing time is. */ if (rack->r_ctl.rc_agg_delayed >= slot) { /* * We can't compensate for it all. * And we have to have some time * on the clock. We always have a min * 10 slots (10 x 10 i.e. 100 usecs). */ if (slot <= HPTS_TICKS_PER_USEC) { /* We gain delay */ rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_USEC - slot); slot = HPTS_TICKS_PER_USEC; } else { /* We take off some */ rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_USEC); slot = HPTS_TICKS_PER_USEC; } } else { slot -= rack->r_ctl.rc_agg_delayed; rack->r_ctl.rc_agg_delayed = 0; /* Make sure we have 100 useconds at minimum */ if (slot < HPTS_TICKS_PER_USEC) { rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_USEC - slot; slot = HPTS_TICKS_PER_USEC; } if (rack->r_ctl.rc_agg_delayed == 0) rack->r_late = 0; } } if (slot) { /* We are pacing too */ rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT; } hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack); #ifdef NETFLIX_EXP_DETECTION if (rack->sack_attack_disable && (slot < tcp_sad_pacing_interval)) { /* * We have a potential attacker on * the line. We have possibly some * (or now) pacing time set. We want to * slow down the processing of sacks by some * amount (if it is an attacker). Set the default * slot for attackers in place (unless the orginal * interval is longer). Its stored in * micro-seconds, so lets convert to msecs. */ slot = tcp_sad_pacing_interval; } #endif if (tp->t_flags & TF_DELACK) { delayed_ack = TICKS_2_MSEC(tcp_delacktime); rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK; } if (delayed_ack && ((hpts_timeout == 0) || (delayed_ack < hpts_timeout))) hpts_timeout = delayed_ack; else rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; /* * If no timers are going to run and we will fall off the hptsi * wheel, we resort to a keep-alive timer if its configured. */ if ((hpts_timeout == 0) && (slot == 0)) { if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)) { /* * Ok we have no timer (persists, rack, tlp, rxt or * del-ack), we don't have segments being paced. So * all that is left is the keepalive timer. */ if (TCPS_HAVEESTABLISHED(tp->t_state)) { /* Get the established keep-alive time */ hpts_timeout = TP_KEEPIDLE(tp); } else { /* Get the initial setup keep-alive time */ hpts_timeout = TP_KEEPINIT(tp); } rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP; if (rack->in_probe_rtt) { /* * We want to instead not wake up a long time from * now but to wake up about the time we would * exit probe-rtt and initiate a keep-alive ack. * This will get us out of probe-rtt and update * our min-rtt. */ hpts_timeout = (rack_min_probertt_hold / HPTS_USEC_IN_MSEC); } } } if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) == (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) { /* * RACK, TLP, persists and RXT timers all are restartable * based on actions input .. i.e we received a packet (ack * or sack) and that changes things (rw, or snd_una etc). * Thus we can restart them with a new value. For * keep-alive, delayed_ack we keep track of what was left * and restart the timer with a smaller value. */ if (left < hpts_timeout) hpts_timeout = left; } if (hpts_timeout) { /* * Hack alert for now we can't time-out over 2,147,483 * seconds (a bit more than 596 hours), which is probably ok * :). */ if (hpts_timeout > 0x7ffffffe) hpts_timeout = 0x7ffffffe; rack->r_ctl.rc_timer_exp = cts + hpts_timeout; } if ((rack->rc_gp_filled == 0) && (hpts_timeout < slot) && (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) { /* * We have no good estimate yet for the * old clunky burst mitigation or the * real pacing. And the tlp or rxt is smaller * than the pacing calculation. Lets not * pace that long since we know the calculation * so far is not accurate. */ slot = hpts_timeout; } rack->r_ctl.last_pacing_time = slot; if (slot) { rack->r_ctl.rc_last_output_to = us_cts + slot; if (rack->rc_always_pace || rack->r_mbuf_queue) { if ((rack->rc_gp_filled == 0) || rack->pacing_longer_than_rtt) { inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY); } else { inp->inp_flags2 |= INP_MBUF_QUEUE_READY; if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && (rack->r_rr_config != 3)) inp->inp_flags2 |= INP_DONT_SACK_QUEUE; else inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; } } if ((rack->use_rack_rr) && (rack->r_rr_config < 2) && ((hpts_timeout) && ((hpts_timeout * HPTS_USEC_IN_MSEC) < slot))) { /* * Arrange for the hpts to kick back in after the * t-o if the t-o does not cause a send. */ (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout), __LINE__, &diag); rack_log_hpts_diag(rack, us_cts, &diag, &tv); rack_log_to_start(rack, cts, hpts_timeout, slot, 0); } else { (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot), __LINE__, &diag); rack_log_hpts_diag(rack, us_cts, &diag, &tv); rack_log_to_start(rack, cts, hpts_timeout, slot, 1); } } else if (hpts_timeout) { if (rack->rc_always_pace || rack->r_mbuf_queue) { if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) { /* For a rack timer, don't wake us */ inp->inp_flags2 |= INP_MBUF_QUEUE_READY; if (rack->r_rr_config != 3) inp->inp_flags2 |= INP_DONT_SACK_QUEUE; else inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; } else { /* All other timers wake us up */ inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY; inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; } } (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout), __LINE__, &diag); rack_log_hpts_diag(rack, us_cts, &diag, &tv); rack_log_to_start(rack, cts, hpts_timeout, slot, 0); } else { /* No timer starting */ #ifdef INVARIANTS if (SEQ_GT(tp->snd_max, tp->snd_una)) { panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?", tp, rack, tot_len_this_send, cts, slot, hpts_timeout); } #endif } rack->rc_tmr_stopped = 0; if (slot) rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv); } /* * RACK Timer, here we simply do logging and house keeping. * the normal rack_output() function will call the * appropriate thing to check if we need to do a RACK retransmit. * We return 1, saying don't proceed with rack_output only * when all timers have been stopped (destroyed PCB?). */ static int rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { /* * This timer simply provides an internal trigger to send out data. * The check_recovery_mode call will see if there are needed * retransmissions, if so we will enter fast-recovery. The output * call may or may not do the same thing depending on sysctl * settings. */ struct rack_sendmap *rsm; int32_t recovery; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } recovery = IN_RECOVERY(tp->t_flags); counter_u64_add(rack_to_tot, 1); if (rack->r_state && (rack->r_state != tp->t_state)) rack_set_state(tp, rack); rack->rc_on_min_to = 0; rsm = rack_check_recovery_mode(tp, cts); rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm); if (rsm) { uint32_t rtt; rack->r_ctl.rc_resend = rsm; if (rack->use_rack_rr) { /* * Don't accumulate extra pacing delay * we are allowing the rack timer to * over-ride pacing i.e. rrr takes precedence * if the pacing interval is longer than the rrr * time (in other words we get the min pacing * time versus rrr pacing time). */ rack->r_timer_override = 1; rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT; } rtt = rack->rc_rack_rtt; if (rtt == 0) rtt = 1; if (rack->rack_no_prr == 0) { if ((recovery == 0) && (rack->r_ctl.rc_prr_sndcnt < ctf_fixed_maxseg(tp))) { /* * The rack-timeout that enter's us into recovery * will force out one MSS and set us up so that we * can do one more send in 2*rtt (transitioning the * rack timeout into a rack-tlp). */ rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp); rack->r_timer_override = 1; rack_log_to_prr(rack, 3, 0); } else if ((rack->r_ctl.rc_prr_sndcnt < (rsm->r_end - rsm->r_start)) && rack->use_rack_rr) { /* * When a rack timer goes, if the rack rr is * on, arrange it so we can send a full segment * overriding prr (though we pay a price for this * for future new sends). */ rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp); rack_log_to_prr(rack, 4, 0); } } } rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK; if (rsm == NULL) { /* restart a timer and return 1 */ rack_start_hpts_timer(rack, tp, cts, 0, 0, 0); return (1); } return (0); } static __inline void rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm, struct rack_sendmap *rsm, uint32_t start) { int idx; nrsm->r_start = start; nrsm->r_end = rsm->r_end; nrsm->r_rtr_cnt = rsm->r_rtr_cnt; nrsm->r_flags = rsm->r_flags; nrsm->r_dupack = rsm->r_dupack; nrsm->usec_orig_send = rsm->usec_orig_send; nrsm->r_rtr_bytes = 0; rsm->r_end = nrsm->r_start; nrsm->r_just_ret = rsm->r_just_ret; for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; } } static struct rack_sendmap * rack_merge_rsm(struct tcp_rack *rack, struct rack_sendmap *l_rsm, struct rack_sendmap *r_rsm) { /* * We are merging two ack'd RSM's, * the l_rsm is on the left (lower seq * values) and the r_rsm is on the right * (higher seq value). The simplest way * to merge these is to move the right * one into the left. I don't think there * is any reason we need to try to find * the oldest (or last oldest retransmitted). */ struct rack_sendmap *rm; l_rsm->r_end = r_rsm->r_end; if (l_rsm->r_dupack < r_rsm->r_dupack) l_rsm->r_dupack = r_rsm->r_dupack; if (r_rsm->r_rtr_bytes) l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes; if (r_rsm->r_in_tmap) { /* This really should not happen */ TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext); r_rsm->r_in_tmap = 0; } /* Now the flags */ if (r_rsm->r_flags & RACK_HAS_FIN) l_rsm->r_flags |= RACK_HAS_FIN; if (r_rsm->r_flags & RACK_TLP) l_rsm->r_flags |= RACK_TLP; if (r_rsm->r_flags & RACK_RWND_COLLAPSED) l_rsm->r_flags |= RACK_RWND_COLLAPSED; if ((r_rsm->r_flags & RACK_APP_LIMITED) && ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) { /* * If both are app-limited then let the * free lower the count. If right is app * limited and left is not, transfer. */ l_rsm->r_flags |= RACK_APP_LIMITED; r_rsm->r_flags &= ~RACK_APP_LIMITED; if (r_rsm == rack->r_ctl.rc_first_appl) rack->r_ctl.rc_first_appl = l_rsm; } rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm); #ifdef INVARIANTS if (rm != r_rsm) { panic("removing head in rack:%p rsm:%p rm:%p", rack, r_rsm, rm); } #endif if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) { /* Transfer the split limit to the map we free */ r_rsm->r_limit_type = l_rsm->r_limit_type; l_rsm->r_limit_type = 0; } rack_free(rack, r_rsm); return(l_rsm); } /* * TLP Timer, here we simply setup what segment we want to * have the TLP expire on, the normal rack_output() will then * send it out. * * We return 1, saying don't proceed with rack_output only * when all timers have been stopped (destroyed PCB?). */ static int rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { /* * Tail Loss Probe. */ struct rack_sendmap *rsm = NULL; struct rack_sendmap *insret; struct socket *so; uint32_t amm, old_prr_snd = 0; uint32_t out, avail; int collapsed_win = 0; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { /* Its not time yet */ return (0); } if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); return (1); } /* * A TLP timer has expired. We have been idle for 2 rtts. So we now * need to figure out how to force a full MSS segment out. */ rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL); counter_u64_add(rack_tlp_tot, 1); if (rack->r_state && (rack->r_state != tp->t_state)) rack_set_state(tp, rack); so = tp->t_inpcb->inp_socket; avail = sbavail(&so->so_snd); out = tp->snd_max - tp->snd_una; if (out > tp->snd_wnd) { /* special case, we need a retransmission */ collapsed_win = 1; goto need_retran; } /* * Check our send oldest always settings, and if * there is an oldest to send jump to the need_retran. */ if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0)) goto need_retran; if (avail > out) { /* New data is available */ amm = avail - out; if (amm > ctf_fixed_maxseg(tp)) { amm = ctf_fixed_maxseg(tp); if ((amm + out) > tp->snd_wnd) { /* We are rwnd limited */ goto need_retran; } } else if (amm < ctf_fixed_maxseg(tp)) { /* not enough to fill a MTU */ goto need_retran; } if (IN_RECOVERY(tp->t_flags)) { /* Unlikely */ if (rack->rack_no_prr == 0) { old_prr_snd = rack->r_ctl.rc_prr_sndcnt; if (out + amm <= tp->snd_wnd) { rack->r_ctl.rc_prr_sndcnt = amm; rack_log_to_prr(rack, 4, 0); } } else goto need_retran; } else { /* Set the send-new override */ if (out + amm <= tp->snd_wnd) rack->r_ctl.rc_tlp_new_data = amm; else goto need_retran; } rack->r_ctl.rc_tlpsend = NULL; counter_u64_add(rack_tlp_newdata, 1); goto send; } need_retran: /* * Ok we need to arrange the last un-acked segment to be re-sent, or * optionally the first un-acked segment. */ if (collapsed_win == 0) { if (rack_always_send_oldest) rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); else { rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) { rsm = rack_find_high_nonack(rack, rsm); } } if (rsm == NULL) { counter_u64_add(rack_tlp_does_nada, 1); #ifdef TCP_BLACKBOX tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true); #endif goto out; } } else { /* * We must find the last segment * that was acceptable by the client. */ RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) { if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) { /* Found one */ break; } } if (rsm == NULL) { /* None? if so send the first */ rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm == NULL) { counter_u64_add(rack_tlp_does_nada, 1); #ifdef TCP_BLACKBOX tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true); #endif goto out; } } } if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) { /* * We need to split this the last segment in two. */ struct rack_sendmap *nrsm; nrsm = rack_alloc_full_limit(rack); if (nrsm == NULL) { /* * No memory to split, we will just exit and punt * off to the RXT timer. */ counter_u64_add(rack_tlp_does_nada, 1); goto out; } rack_clone_rsm(rack, nrsm, rsm, (rsm->r_end - ctf_fixed_maxseg(tp))); insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #ifdef INVARIANTS if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~RACK_HAS_FIN); rsm = nrsm; } rack->r_ctl.rc_tlpsend = rsm; send: rack->r_timer_override = 1; rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; return (0); out: rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; return (0); } /* * Delayed ack Timer, here we simply need to setup the * ACK_NOW flag and remove the DELACK flag. From there * the output routine will send the ack out. * * We only return 1, saying don't proceed, if all timers * are stopped (destroyed PCB?). */ static int rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL); tp->t_flags &= ~TF_DELACK; tp->t_flags |= TF_ACKNOW; KMOD_TCPSTAT_INC(tcps_delack); rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; return (0); } /* * Persists timer, here we simply send the * same thing as a keepalive will. * the one byte send. * * We only return 1, saying don't proceed, if all timers * are stopped (destroyed PCB?). */ static int rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { struct tcptemp *t_template; struct inpcb *inp; int32_t retval = 1; inp = tp->t_inpcb; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } if (rack->rc_in_persist == 0) return (0); if (ctf_progress_timeout_check(tp, false)) { tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(inp, ETIMEDOUT); return (1); } KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); /* * Persistence timer into zero window. Force a byte to be output, if * possible. */ KMOD_TCPSTAT_INC(tcps_persisttimeo); /* * Hack: if the peer is dead/unreachable, we do not time out if the * window is closed. After a full backoff, drop the connection if * the idle time (no responses to probes) reaches the maximum * backoff that we would use if retransmitting. */ if (tp->t_rxtshift == TCP_MAXRXTSHIFT && (ticks - tp->t_rcvtime >= tcp_maxpersistidle || ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) { KMOD_TCPSTAT_INC(tcps_persistdrop); retval = 1; tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); goto out; } if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) && tp->snd_una == tp->snd_max) rack_exit_persist(tp, rack, cts); rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT; /* * If the user has closed the socket then drop a persisting * connection after a much reduced timeout. */ if (tp->t_state > TCPS_CLOSE_WAIT && (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) { retval = 1; KMOD_TCPSTAT_INC(tcps_persistdrop); tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); goto out; } t_template = tcpip_maketemplate(rack->rc_inp); if (t_template) { /* only set it if we were answered */ if (rack->forced_ack == 0) { rack->forced_ack = 1; rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL); } tcp_respond(tp, t_template->tt_ipgen, &t_template->tt_t, (struct mbuf *)NULL, tp->rcv_nxt, tp->snd_una - 1, 0); /* This sends an ack */ if (tp->t_flags & TF_DELACK) tp->t_flags &= ~TF_DELACK; free(t_template, M_TEMP); } if (tp->t_rxtshift < TCP_MAXRXTSHIFT) tp->t_rxtshift++; out: rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL); rack_start_hpts_timer(rack, tp, cts, 0, 0, 0); return (retval); } /* * If a keepalive goes off, we had no other timers * happening. We always return 1 here since this * routine either drops the connection or sends * out a segment with respond. */ static int rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { struct tcptemp *t_template; struct inpcb *inp; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP; inp = tp->t_inpcb; rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL); /* * Keep-alive timer went off; send something or drop connection if * idle for too long. */ KMOD_TCPSTAT_INC(tcps_keeptimeo); if (tp->t_state < TCPS_ESTABLISHED) goto dropit; if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && tp->t_state <= TCPS_CLOSING) { if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp)) goto dropit; /* * Send a packet designed to force a response if the peer is * up and reachable: either an ACK if the connection is * still alive, or an RST if the peer has closed the * connection due to timeout or reboot. Using sequence * number tp->snd_una-1 causes the transmitted zero-length * segment to lie outside the receive window; by the * protocol spec, this requires the correspondent TCP to * respond. */ KMOD_TCPSTAT_INC(tcps_keepprobe); t_template = tcpip_maketemplate(inp); if (t_template) { if (rack->forced_ack == 0) { rack->forced_ack = 1; rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL); } tcp_respond(tp, t_template->tt_ipgen, &t_template->tt_t, (struct mbuf *)NULL, tp->rcv_nxt, tp->snd_una - 1, 0); free(t_template, M_TEMP); } } rack_start_hpts_timer(rack, tp, cts, 0, 0, 0); return (1); dropit: KMOD_TCPSTAT_INC(tcps_keepdrops); tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX); tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); return (1); } /* * Retransmit helper function, clear up all the ack * flags and take care of important book keeping. */ static void rack_remxt_tmr(struct tcpcb *tp) { /* * The retransmit timer went off, all sack'd blocks must be * un-acked. */ struct rack_sendmap *rsm, *trsm = NULL; struct tcp_rack *rack; int32_t cnt = 0; rack = (struct tcp_rack *)tp->t_fb_ptr; rack_timer_cancel(tp, rack, tcp_ts_getticks(), __LINE__); rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL); if (rack->r_state && (rack->r_state != tp->t_state)) rack_set_state(tp, rack); /* * Ideally we would like to be able to * mark SACK-PASS on anything not acked here. * However, if we do that we would burst out * all that data 1ms apart. This would be unwise, * so for now we will just let the normal rxt timer * and tlp timer take care of it. */ RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) { if (rsm->r_flags & RACK_ACKED) { cnt++; rsm->r_dupack = 0; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); if (rsm->r_in_tmap == 0) { /* We must re-add it back to the tlist */ if (trsm == NULL) { TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext); } else { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext); } rsm->r_in_tmap = 1; } } trsm = rsm; if (rsm->r_flags & RACK_ACKED) rsm->r_flags |= RACK_WAS_ACKED; rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS); } /* Clear the count (we just un-acked them) */ rack->r_ctl.rc_sacked = 0; rack->r_ctl.rc_agg_delayed = 0; rack->r_early = 0; rack->r_ctl.rc_agg_early = 0; rack->r_late = 0; /* Clear the tlp rtx mark */ rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); rack->r_ctl.rc_prr_sndcnt = 0; rack_log_to_prr(rack, 6, 0); rack->r_timer_override = 1; } static void rack_cc_conn_init(struct tcpcb *tp) { struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; cc_conn_init(tp); /* * We want a chance to stay in slowstart as * we create a connection. TCP spec says that * initially ssthresh is infinite. For our * purposes that is the snd_wnd. */ if (tp->snd_ssthresh < tp->snd_wnd) { tp->snd_ssthresh = tp->snd_wnd; } /* * We also want to assure a IW worth of * data can get inflight. */ if (rc_init_window(rack) < tp->snd_cwnd) tp->snd_cwnd = rc_init_window(rack); } /* * Re-transmit timeout! If we drop the PCB we will return 1, otherwise * we will setup to retransmit the lowest seq number outstanding. */ static int rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { int32_t rexmt; struct inpcb *inp; int32_t retval = 0; bool isipv6; inp = tp->t_inpcb; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } if (ctf_progress_timeout_check(tp, false)) { tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN); rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(inp, ETIMEDOUT); return (1); } rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_una == tp->snd_max)) { /* Nothing outstanding .. nothing to do */ return (0); } /* * Retransmission timer went off. Message has not been acked within * retransmit interval. Back off to a longer retransmit interval * and retransmit one segment. */ rack_remxt_tmr(tp); if ((rack->r_ctl.rc_resend == NULL) || ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) { /* * If the rwnd collapsed on * the one we are retransmitting * it does not count against the * rxt count. */ tp->t_rxtshift++; } if (tp->t_rxtshift > TCP_MAXRXTSHIFT) { tp->t_rxtshift = TCP_MAXRXTSHIFT; KMOD_TCPSTAT_INC(tcps_timeoutdrop); retval = 1; tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN); tcp_set_inp_to_drop(rack->rc_inp, (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT)); goto out; } if (tp->t_state == TCPS_SYN_SENT) { /* * If the SYN was retransmitted, indicate CWND to be limited * to 1 segment in cc_conn_init(). */ tp->snd_cwnd = 1; } else if (tp->t_rxtshift == 1) { /* * first retransmit; record ssthresh and cwnd so they can be * recovered if this turns out to be a "bad" retransmit. A * retransmit is considered "bad" if an ACK for this segment * is received within RTT/2 interval; the assumption here is * that the ACK was already in flight. See "On Estimating * End-to-End Network Path Properties" by Allman and Paxson * for more details. */ tp->snd_cwnd_prev = tp->snd_cwnd; tp->snd_ssthresh_prev = tp->snd_ssthresh; tp->snd_recover_prev = tp->snd_recover; if (IN_FASTRECOVERY(tp->t_flags)) tp->t_flags |= TF_WASFRECOVERY; else tp->t_flags &= ~TF_WASFRECOVERY; if (IN_CONGRECOVERY(tp->t_flags)) tp->t_flags |= TF_WASCRECOVERY; else tp->t_flags &= ~TF_WASCRECOVERY; tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); tp->t_flags |= TF_PREVVALID; } else tp->t_flags &= ~TF_PREVVALID; KMOD_TCPSTAT_INC(tcps_rexmttimeo); if ((tp->t_state == TCPS_SYN_SENT) || (tp->t_state == TCPS_SYN_RECEIVED)) rexmt = MSEC_2_TICKS(RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift]); else rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; TCPT_RANGESET(tp->t_rxtcur, rexmt, max(MSEC_2_TICKS(rack_rto_min), rexmt), MSEC_2_TICKS(rack_rto_max)); /* * We enter the path for PLMTUD if connection is established or, if * connection is FIN_WAIT_1 status, reason for the last is that if * amount of data we send is very small, we could send it in couple * of packets and process straight to FIN. In that case we won't * catch ESTABLISHED state. */ #ifdef INET6 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false; #else isipv6 = false; #endif if (((V_tcp_pmtud_blackhole_detect == 1) || (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) || (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) && ((tp->t_state == TCPS_ESTABLISHED) || (tp->t_state == TCPS_FIN_WAIT_1))) { /* * Idea here is that at each stage of mtu probe (usually, * 1448 -> 1188 -> 524) should be given 2 chances to recover * before further clamping down. 'tp->t_rxtshift % 2 == 0' * should take care of that. */ if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) == (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) && (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 && tp->t_rxtshift % 2 == 0)) { /* * Enter Path MTU Black-hole Detection mechanism: - * Disable Path MTU Discovery (IP "DF" bit). - * Reduce MTU to lower value than what we negotiated * with peer. */ if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) { /* Record that we may have found a black hole. */ tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE; /* Keep track of previous MSS. */ tp->t_pmtud_saved_maxseg = tp->t_maxseg; } /* * Reduce the MSS to blackhole value or to the * default in an attempt to retransmit. */ #ifdef INET6 if (isipv6 && tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { /* Use the sysctl tuneable blackhole MSS. */ tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); } else if (isipv6) { /* Use the default MSS. */ tp->t_maxseg = V_tcp_v6mssdflt; /* * Disable Path MTU Discovery when we switch * to minmss. */ tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) { /* Use the sysctl tuneable blackhole MSS. */ tp->t_maxseg = V_tcp_pmtud_blackhole_mss; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); } else { /* Use the default MSS. */ tp->t_maxseg = V_tcp_mssdflt; /* * Disable Path MTU Discovery when we switch * to minmss. */ tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); } #endif } else { /* * If further retransmissions are still unsuccessful * with a lowered MTU, maybe this isn't a blackhole * and we restore the previous MSS and blackhole * detection flags. The limit '6' is determined by * giving each probe stage (1448, 1188, 524) 2 * chances to recover. */ if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && (tp->t_rxtshift >= 6)) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; tp->t_maxseg = tp->t_pmtud_saved_maxseg; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed); } } } /* * If we backed off this far, our srtt estimate is probably bogus. * Clobber it so we'll take the next rtt measurement as our srtt; * move the current srtt into rttvar to keep the current retransmit * times until then. */ if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) { #ifdef INET6 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) in6_losing(tp->t_inpcb); else #endif in_losing(tp->t_inpcb); tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT); tp->t_srtt = 0; } sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); tp->snd_recover = tp->snd_max; tp->t_flags |= TF_ACKNOW; tp->t_rtttime = 0; rack_cong_signal(tp, NULL, CC_RTO); out: return (retval); } static int rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling) { int32_t ret = 0; int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK); if (timers == 0) { return (0); } if (tp->t_state == TCPS_LISTEN) { /* no timers on listen sockets */ if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) return (0); return (1); } if ((timers & PACE_TMR_RACK) && rack->rc_on_min_to) { /* * For the rack timer when we * are on a min-timeout (which means rrr_conf = 3) * we don't want to check the timer. It may * be going off for a pace and thats ok we * want to send the retransmit (if its ready). * * If its on a normal rack timer (non-min) then * we will check if its expired. */ goto skip_time_check; } if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { uint32_t left; if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { ret = -1; rack_log_to_processing(rack, cts, ret, 0); return (0); } if (hpts_calling == 0) { /* * A user send or queued mbuf (sack) has called us? We * return 0 and let the pacing guards * deal with it if they should or * should not cause a send. */ ret = -2; rack_log_to_processing(rack, cts, ret, 0); return (0); } /* * Ok our timer went off early and we are not paced false * alarm, go back to sleep. */ ret = -3; left = rack->r_ctl.rc_timer_exp - cts; tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left)); rack_log_to_processing(rack, cts, ret, left); return (1); } skip_time_check: rack->rc_tmr_stopped = 0; rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK; if (timers & PACE_TMR_DELACK) { ret = rack_timeout_delack(tp, rack, cts); } else if (timers & PACE_TMR_RACK) { rack->r_ctl.rc_tlp_rxt_last_time = cts; ret = rack_timeout_rack(tp, rack, cts); } else if (timers & PACE_TMR_TLP) { rack->r_ctl.rc_tlp_rxt_last_time = cts; ret = rack_timeout_tlp(tp, rack, cts); } else if (timers & PACE_TMR_RXT) { rack->r_ctl.rc_tlp_rxt_last_time = cts; ret = rack_timeout_rxt(tp, rack, cts); } else if (timers & PACE_TMR_PERSIT) { ret = rack_timeout_persist(tp, rack, cts); } else if (timers & PACE_TMR_KEEP) { ret = rack_timeout_keepalive(tp, rack, cts); } rack_log_to_processing(rack, cts, ret, timers); return (ret); } static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line) { struct timeval tv; uint32_t us_cts, flags_on_entry; uint8_t hpts_removed = 0; flags_on_entry = rack->r_ctl.rc_hpts_flags; us_cts = tcp_get_usecs(&tv); if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) || ((tp->snd_max - tp->snd_una) == 0))) { tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); hpts_removed = 1; /* If we were not delayed cancel out the flag. */ if ((tp->snd_max - tp->snd_una) == 0) rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT; rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry); } if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK; if (rack->rc_inp->inp_in_hpts && ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) { /* * Canceling timer's when we have no output being * paced. We also must remove ourselves from the * hpts. */ tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); hpts_removed = 1; } rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK); } if (hpts_removed == 0) rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry); } static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type) { return; } static int rack_stopall(struct tcpcb *tp) { struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; rack->t_timers_stopped = 1; return (0); } static void rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta) { return; } static int rack_timer_active(struct tcpcb *tp, uint32_t timer_type) { return (0); } static void rack_stop_all_timers(struct tcpcb *tp) { struct tcp_rack *rack; /* * Assure no timers are running. */ if (tcp_timer_active(tp, TT_PERSIST)) { /* We enter in persists, set the flag appropriately */ rack = (struct tcp_rack *)tp->t_fb_ptr; rack->rc_in_persist = 1; } tcp_timer_suspend(tp, TT_PERSIST); tcp_timer_suspend(tp, TT_REXMT); tcp_timer_suspend(tp, TT_KEEP); tcp_timer_suspend(tp, TT_DELACK); } static void rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t ts) { int32_t idx; rsm->r_rtr_cnt++; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); rsm->r_dupack = 0; if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) { rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS; rsm->r_flags |= RACK_OVERMAX; } if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) { rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start); rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start); } idx = rsm->r_rtr_cnt - 1; rsm->r_tim_lastsent[idx] = ts; if (rsm->r_flags & RACK_ACKED) { /* Problably MTU discovery messing with us */ rsm->r_flags &= ~RACK_ACKED; rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); } if (rsm->r_in_tmap) { TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; if (rsm->r_flags & RACK_SACK_PASSED) { /* We have retransmitted due to the SACK pass */ rsm->r_flags &= ~RACK_SACK_PASSED; rsm->r_flags |= RACK_WAS_SACKPASS; } } static uint32_t rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t ts, int32_t *lenp) { /* * We (re-)transmitted starting at rsm->r_start for some length * (possibly less than r_end. */ struct rack_sendmap *nrsm, *insret; uint32_t c_end; int32_t len; len = *lenp; c_end = rsm->r_start + len; if (SEQ_GEQ(c_end, rsm->r_end)) { /* * We retransmitted the whole piece or more than the whole * slopping into the next rsm. */ rack_update_rsm(tp, rack, rsm, ts); if (c_end == rsm->r_end) { *lenp = 0; return (0); } else { int32_t act_len; /* Hangs over the end return whats left */ act_len = rsm->r_end - rsm->r_start; *lenp = (len - act_len); return (rsm->r_end); } /* We don't get out of this block. */ } /* * Here we retransmitted less than the whole thing which means we * have to split this into what was transmitted and what was not. */ nrsm = rack_alloc_full_limit(rack); if (nrsm == NULL) { /* * We can't get memory, so lets not proceed. */ *lenp = 0; return (0); } /* * So here we are going to take the original rsm and make it what we * retransmitted. nrsm will be the tail portion we did not * retransmit. For example say the chunk was 1, 11 (10 bytes). And * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to * 1, 6 and the new piece will be 6, 11. */ rack_clone_rsm(rack, nrsm, rsm, c_end); nrsm->r_dupack = 0; rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2); insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #ifdef INVARIANTS if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~RACK_HAS_FIN); rack_update_rsm(tp, rack, rsm, ts); *lenp = 0; return (0); } static void rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts, uint8_t pass, struct rack_sendmap *hintrsm, uint32_t us_cts) { struct tcp_rack *rack; struct rack_sendmap *rsm, *nrsm, *insret, fe; register uint32_t snd_max, snd_una; /* * Add to the RACK log of packets in flight or retransmitted. If * there is a TS option we will use the TS echoed, if not we will * grab a TS. * * Retransmissions will increment the count and move the ts to its * proper place. Note that if options do not include TS's then we * won't be able to effectively use the ACK for an RTT on a retran. * * Notes about r_start and r_end. Lets consider a send starting at * sequence 1 for 10 bytes. In such an example the r_start would be * 1 (starting sequence) but the r_end would be r_start+len i.e. 11. * This means that r_end is actually the first sequence for the next * slot (11). * */ /* * If err is set what do we do XXXrrs? should we not add the thing? * -- i.e. return if err != 0 or should we pretend we sent it? -- * i.e. proceed with add ** do this for now. */ INP_WLOCK_ASSERT(tp->t_inpcb); if (err) /* * We don't log errors -- we could but snd_max does not * advance in this case either. */ return; if (th_flags & TH_RST) { /* * We don't log resets and we return immediately from * sending */ return; } rack = (struct tcp_rack *)tp->t_fb_ptr; snd_una = tp->snd_una; if (SEQ_LEQ((seq_out + len), snd_una)) { /* Are sending an old segment to induce an ack (keep-alive)? */ return; } if (SEQ_LT(seq_out, snd_una)) { /* huh? should we panic? */ uint32_t end; end = seq_out + len; seq_out = snd_una; if (SEQ_GEQ(end, seq_out)) len = end - seq_out; else len = 0; } snd_max = tp->snd_max; if (th_flags & (TH_SYN | TH_FIN)) { /* * The call to rack_log_output is made before bumping * snd_max. This means we can record one extra byte on a SYN * or FIN if seq_out is adding more on and a FIN is present * (and we are not resending). */ if ((th_flags & TH_SYN) && (seq_out == tp->iss)) len++; if (th_flags & TH_FIN) len++; if (SEQ_LT(snd_max, tp->snd_nxt)) { /* * The add/update as not been done for the FIN/SYN * yet. */ snd_max = tp->snd_nxt; } } if (len == 0) { /* We don't log zero window probes */ return; } rack->r_ctl.rc_time_last_sent = ts; if (IN_RECOVERY(tp->t_flags)) { rack->r_ctl.rc_prr_out += len; } /* First question is it a retransmission or new? */ if (seq_out == snd_max) { /* Its new */ again: rsm = rack_alloc(rack); if (rsm == NULL) { /* * Hmm out of memory and the tcb got destroyed while * we tried to wait. */ return; } if (th_flags & TH_FIN) { rsm->r_flags = RACK_HAS_FIN; } else { rsm->r_flags = 0; } rsm->r_tim_lastsent[0] = ts; rsm->r_rtr_cnt = 1; rsm->r_rtr_bytes = 0; rsm->usec_orig_send = us_cts; if (th_flags & TH_SYN) { /* The data space is one beyond snd_una */ rsm->r_flags |= RACK_HAS_SIN; rsm->r_start = seq_out + 1; rsm->r_end = rsm->r_start + (len - 1); } else { /* Normal case */ rsm->r_start = seq_out; rsm->r_end = rsm->r_start + len; } rsm->r_dupack = 0; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #ifdef INVARIANTS if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; /* * Special case detection, is there just a single * packet outstanding when we are not in recovery? * * If this is true mark it so. */ if ((IN_RECOVERY(tp->t_flags) == 0) && (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) { struct rack_sendmap *prsm; prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); if (prsm) prsm->r_one_out_nr = 1; } return; } /* * If we reach here its a retransmission and we need to find it. */ memset(&fe, 0, sizeof(fe)); more: if (hintrsm && (hintrsm->r_start == seq_out)) { rsm = hintrsm; hintrsm = NULL; } else { /* No hints sorry */ rsm = NULL; } if ((rsm) && (rsm->r_start == seq_out)) { seq_out = rack_update_entry(tp, rack, rsm, ts, &len); if (len == 0) { return; } else { goto more; } } /* Ok it was not the last pointer go through it the hard way. */ refind: fe.r_start = seq_out; rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); if (rsm) { if (rsm->r_start == seq_out) { seq_out = rack_update_entry(tp, rack, rsm, ts, &len); if (len == 0) { return; } else { goto refind; } } if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) { /* Transmitted within this piece */ /* * Ok we must split off the front and then let the * update do the rest */ nrsm = rack_alloc_full_limit(rack); if (nrsm == NULL) { rack_update_rsm(tp, rack, rsm, ts); return; } /* * copy rsm to nrsm and then trim the front of rsm * to not include this part. */ rack_clone_rsm(rack, nrsm, rsm, seq_out); insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #ifdef INVARIANTS if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~RACK_HAS_FIN); seq_out = rack_update_entry(tp, rack, nrsm, ts, &len); if (len == 0) { return; } else if (len > 0) goto refind; } } /* * Hmm not found in map did they retransmit both old and on into the * new? */ if (seq_out == tp->snd_max) { goto again; } else if (SEQ_LT(seq_out, tp->snd_max)) { #ifdef INVARIANTS printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n", seq_out, len, tp->snd_una, tp->snd_max); printf("Starting Dump of all rack entries\n"); RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) { printf("rsm:%p start:%u end:%u\n", rsm, rsm->r_start, rsm->r_end); } printf("Dump complete\n"); panic("seq_out not found rack:%p tp:%p", rack, tp); #endif } else { #ifdef INVARIANTS /* * Hmm beyond sndmax? (only if we are using the new rtt-pack * flag) */ panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p", seq_out, len, tp->snd_max, tp); #endif } } /* * Record one of the RTT updates from an ack into * our sample structure. */ static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt) { if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) { rack->r_ctl.rack_rs.rs_rtt_lowest = rtt; } if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) { rack->r_ctl.rack_rs.rs_rtt_highest = rtt; } if (rack->rc_tp->t_flags & TF_GPUTINPROG) { if (us_rtt < rack->r_ctl.rc_gp_lowrtt) rack->r_ctl.rc_gp_lowrtt = us_rtt; if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd) rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd; } if ((confidence == 1) && ((rsm == NULL) || (rsm->r_just_ret) || (rsm->r_one_out_nr && len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) { /* * If the rsm had a just return * hit it then we can't trust the * rtt measurement for buffer deterimination * Note that a confidence of 2, indicates * SACK'd which overrides the r_just_ret or * the r_one_out_nr. If it was a CUM-ACK and * we had only two outstanding, but get an * ack for only 1. Then that also lowers our * confidence. */ confidence = 0; } if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) { if (rack->r_ctl.rack_rs.confidence == 0) { /* * We take anything with no current confidence * saved. */ rack->r_ctl.rack_rs.rs_us_rtt = us_rtt; rack->r_ctl.rack_rs.confidence = confidence; rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt; } else if (confidence || rack->r_ctl.rack_rs.confidence) { /* * Once we have a confident number, * we can update it with a smaller * value since this confident number * may include the DSACK time until * the next segment (the second one) arrived. */ rack->r_ctl.rack_rs.rs_us_rtt = us_rtt; rack->r_ctl.rack_rs.confidence = confidence; rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt; } } rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence); rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID; rack->r_ctl.rack_rs.rs_rtt_tot += rtt; rack->r_ctl.rack_rs.rs_rtt_cnt++; } /* * Collect new round-trip time estimate * and update averages and current timeout. */ static void tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp) { int32_t delta; uint32_t o_srtt, o_var; int32_t hrtt_up = 0; int32_t rtt; if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) /* No valid sample */ return; if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) { /* We are to use the lowest RTT seen in a single ack */ rtt = rack->r_ctl.rack_rs.rs_rtt_lowest; } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) { /* We are to use the highest RTT seen in a single ack */ rtt = rack->r_ctl.rack_rs.rs_rtt_highest; } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) { /* We are to use the average RTT seen in a single ack */ rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot / (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt); } else { #ifdef INVARIANTS panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method); #endif return; } if (rtt == 0) rtt = 1; if (rack->rc_gp_rtt_set == 0) { /* * With no RTT we have to accept * even one we are not confident of. */ rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt; rack->rc_gp_rtt_set = 1; } else if (rack->r_ctl.rack_rs.confidence) { /* update the running gp srtt */ rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8); rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8; } if (rack->r_ctl.rack_rs.confidence) { /* * record the low and high for highly buffered path computation, * we only do this if we are confident (not a retransmission). */ if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) { rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt; hrtt_up = 1; } if (rack->rc_highly_buffered == 0) { /* * Currently once we declare a path has * highly buffered there is no going * back, which may be a problem... */ if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) { rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt, rack->r_ctl.rc_highest_us_rtt, rack->r_ctl.rc_lowest_us_rtt, RACK_RTTS_SEEHBP); rack->rc_highly_buffered = 1; } } } if ((rack->r_ctl.rack_rs.confidence) || (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) { /* * If we are highly confident of it it was * never retransmitted we accept it as the last us_rtt. */ rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt; /* The lowest rtt can be set if its was not retransmited */ if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) { rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt; if (rack->r_ctl.rc_lowest_us_rtt == 0) rack->r_ctl.rc_lowest_us_rtt = 1; } } rack_log_rtt_sample(rack, rtt); o_srtt = tp->t_srtt; o_var = tp->t_rttvar; rack = (struct tcp_rack *)tp->t_fb_ptr; if (tp->t_srtt != 0) { /* * srtt is stored as fixed point with 5 bits after the * binary point (i.e., scaled by 8). The following magic is * equivalent to the smoothing algorithm in rfc793 with an * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point). * Adjust rtt to origin 0. */ delta = ((rtt - 1) << TCP_DELTA_SHIFT) - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); tp->t_srtt += delta; if (tp->t_srtt <= 0) tp->t_srtt = 1; /* * We accumulate a smoothed rtt variance (actually, a * smoothed mean difference), then set the retransmit timer * to smoothed rtt + 4 times the smoothed variance. rttvar * is stored as fixed point with 4 bits after the binary * point (scaled by 16). The following is equivalent to * rfc793 smoothing with an alpha of .75 (rttvar = * rttvar*3/4 + |delta| / 4). This replaces rfc793's * wired-in beta. */ if (delta < 0) delta = -delta; delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); tp->t_rttvar += delta; if (tp->t_rttvar <= 0) tp->t_rttvar = 1; if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar) tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } else { /* * No rtt measurement yet - use the unsmoothed rtt. Set the * variance to half the rtt (so our first retransmit happens * at 3*rtt). */ tp->t_srtt = rtt << TCP_RTT_SHIFT; tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1); tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } KMOD_TCPSTAT_INC(tcps_rttupdated); tp->t_rttupdated++; #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt)); #endif tp->t_rxtshift = 0; /* * the retransmit should happen at rtt + 4 * rttvar. Because of the * way we do the smoothing, srtt and rttvar will each average +1/2 * tick of bias. When we compute the retransmit timer, we want 1/2 * tick of rounding and 1 extra tick because of +-1/2 tick * uncertainty in the firing of the timer. The bias will give us * exactly the 1.5 tick we need. But, because the bias is * statistical, we have to test that we don't drop below the minimum * feasible timer (which is 2 ticks). */ TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), max(MSEC_2_TICKS(rack_rto_min), rtt + 2), MSEC_2_TICKS(rack_rto_max)); tp->t_softerror = 0; } static void rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm, uint32_t t, uint32_t cts) { /* * For this RSM, we acknowledged the data from a previous * transmission, not the last one we made. This means we did a false * retransmit. */ struct tcp_rack *rack; if (rsm->r_flags & RACK_HAS_FIN) { /* * The sending of the FIN often is multiple sent when we * have everything outstanding ack'd. We ignore this case * since its over now. */ return; } if (rsm->r_flags & RACK_TLP) { /* * We expect TLP's to have this occur. */ return; } rack = (struct tcp_rack *)tp->t_fb_ptr; /* should we undo cc changes and exit recovery? */ if (IN_RECOVERY(tp->t_flags)) { if (rack->r_ctl.rc_rsm_start == rsm->r_start) { /* * Undo what we ratched down and exit recovery if * possible */ EXIT_RECOVERY(tp->t_flags); tp->snd_recover = tp->snd_una; if (rack->r_ctl.rc_cwnd_at > tp->snd_cwnd) tp->snd_cwnd = rack->r_ctl.rc_cwnd_at; if (rack->r_ctl.rc_ssthresh_at > tp->snd_ssthresh) tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at; } } if (rsm->r_flags & RACK_WAS_SACKPASS) { /* * We retransmitted based on a sack and the earlier * retransmission ack'd it - re-ordering is occuring. */ counter_u64_add(rack_reorder_seen, 1); rack->r_ctl.rc_reorder_ts = cts; } counter_u64_add(rack_badfr, 1); counter_u64_add(rack_badfr_bytes, (rsm->r_end - rsm->r_start)); } static void rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts) { /* * Apply to filter the inbound us-rtt at us_cts. */ uint32_t old_rtt; old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt); apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt, us_rtt, us_cts); if (rack->r_ctl.last_pacing_time && rack->rc_gp_dyn_mul && (rack->r_ctl.last_pacing_time > us_rtt)) rack->pacing_longer_than_rtt = 1; else rack->pacing_longer_than_rtt = 0; if (old_rtt > us_rtt) { /* We just hit a new lower rtt time */ rack_log_rtt_shrinks(rack, us_cts, old_rtt, __LINE__, RACK_RTTS_NEWRTT); /* * Only count it if its lower than what we saw within our * calculated range. */ if ((old_rtt - us_rtt) > rack_min_rtt_movement) { if (rack_probertt_lower_within && rack->rc_gp_dyn_mul && (rack->use_fixed_rate == 0) && (rack->rc_always_pace)) { /* * We are seeing a new lower rtt very close * to the time that we would have entered probe-rtt. * This is probably due to the fact that a peer flow * has entered probe-rtt. Lets go in now too. */ uint32_t val; val = rack_probertt_lower_within * rack_time_between_probertt; val /= 100; if ((rack->in_probe_rtt == 0) && ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val))) { rack_enter_probertt(rack, us_cts); } } rack->r_ctl.rc_lower_rtt_us_cts = us_cts; } } } static int rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack) { int32_t i; uint32_t t, len_acked; if ((rsm->r_flags & RACK_ACKED) || (rsm->r_flags & RACK_WAS_ACKED)) /* Already done */ return (0); if (ack_type == CUM_ACKED) { if (SEQ_GT(th_ack, rsm->r_end)) len_acked = rsm->r_end - rsm->r_start; else len_acked = th_ack - rsm->r_start; } else len_acked = rsm->r_end - rsm->r_start; if (rsm->r_rtr_cnt == 1) { uint32_t us_rtt; t = cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; if ((int)t <= 0) t = 1; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { rack->r_ctl.rc_rack_min_rtt = t; if (rack->r_ctl.rc_rack_min_rtt == 0) { rack->r_ctl.rc_rack_min_rtt = 1; } } us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - rsm->usec_orig_send; if (us_rtt == 0) us_rtt = 1; rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time)); if (ack_type == SACKED) tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt); else { /* * For cum-ack we are only confident if what * is being acked is included in a measurement. * Otherwise it could be an idle period that * includes Delayed-ack time. */ tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, (rack->app_limited_needs_set ? 0 : 1), rsm, rsm->r_rtr_cnt); } if ((rsm->r_flags & RACK_TLP) && (!IN_RECOVERY(tp->t_flags))) { /* Segment was a TLP and our retrans matched */ if (rack->r_ctl.rc_tlp_cwnd_reduce) { rack->r_ctl.rc_rsm_start = tp->snd_max; rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; rack_cong_signal(tp, NULL, CC_NDUPACK); /* * When we enter recovery we need to assure * we send one packet. */ if (rack->rack_no_prr == 0) { rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp); rack_log_to_prr(rack, 7, 0); } } } if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { /* New more recent rack_tmit_time */ rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; rack->rc_rack_rtt = t; } return (1); } /* * We clear the soft/rxtshift since we got an ack. * There is no assurance we will call the commit() function * so we need to clear these to avoid incorrect handling. */ tp->t_rxtshift = 0; tp->t_softerror = 0; if ((to->to_flags & TOF_TS) && (ack_type == CUM_ACKED) && (to->to_tsecr) && ((rsm->r_flags & RACK_OVERMAX) == 0)) { /* * Now which timestamp does it match? In this block the ACK * must be coming from a previous transmission. */ for (i = 0; i < rsm->r_rtr_cnt; i++) { if (rsm->r_tim_lastsent[i] == to->to_tsecr) { t = cts - rsm->r_tim_lastsent[i]; if ((int)t <= 0) t = 1; if ((i + 1) < rsm->r_rtr_cnt) { /* Likely */ rack_earlier_retran(tp, rsm, t, cts); } if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { rack->r_ctl.rc_rack_min_rtt = t; if (rack->r_ctl.rc_rack_min_rtt == 0) { rack->r_ctl.rc_rack_min_rtt = 1; } } if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { /* New more recent rack_tmit_time */ rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; rack->rc_rack_rtt = t; } tcp_rack_xmit_timer(rack, t + 1, len_acked, (t * HPTS_USEC_IN_MSEC), 0, rsm, rsm->r_rtr_cnt); return (1); } } goto ts_not_found; } else { /* * Ok its a SACK block that we retransmitted. or a windows * machine without timestamps. We can tell nothing from the * time-stamp since its not there or the time the peer last * recieved a segment that moved forward its cum-ack point. */ ts_not_found: i = rsm->r_rtr_cnt - 1; t = cts - rsm->r_tim_lastsent[i]; if ((int)t <= 0) t = 1; if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { /* * We retransmitted and the ack came back in less * than the smallest rtt we have observed. We most * likey did an improper retransmit as outlined in * 4.2 Step 3 point 2 in the rack-draft. */ i = rsm->r_rtr_cnt - 2; t = cts - rsm->r_tim_lastsent[i]; rack_earlier_retran(tp, rsm, t, cts); } else if (rack->r_ctl.rc_rack_min_rtt) { /* * We retransmitted it and the retransmit did the * job. */ if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { rack->r_ctl.rc_rack_min_rtt = t; if (rack->r_ctl.rc_rack_min_rtt == 0) { rack->r_ctl.rc_rack_min_rtt = 1; } } if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[i])) { /* New more recent rack_tmit_time */ rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[i]; rack->rc_rack_rtt = t; } return (1); } } return (0); } /* * Mark the SACK_PASSED flag on all entries prior to rsm send wise. */ static void rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm) { struct rack_sendmap *nrsm; nrsm = rsm; TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap, rack_head, r_tnext) { if (nrsm == rsm) { /* Skip orginal segment he is acked */ continue; } if (nrsm->r_flags & RACK_ACKED) { /* * Skip ack'd segments, though we * should not see these, since tmap * should not have ack'd segments. */ continue; } if (nrsm->r_flags & RACK_SACK_PASSED) { /* * We found one that is already marked * passed, we have been here before and * so all others below this are marked. */ break; } nrsm->r_flags |= RACK_SACK_PASSED; nrsm->r_flags &= ~RACK_WAS_SACKPASS; } } static void rack_need_set_test(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack, int line, int use_which) { if ((tp->t_flags & TF_GPUTINPROG) && SEQ_GEQ(rsm->r_end, tp->gput_seq)) { /* * We were app limited, and this ack * butts up or goes beyond the point where we want * to start our next measurement. We need * to record the new gput_ts as here and * possibly update the start sequence. */ uint32_t seq, ts; if (rsm->r_rtr_cnt > 1) { /* * This is a retransmit, can we * really make any assessment at this * point? We are not really sure of * the timestamp, is it this or the * previous transmission? * * Lets wait for something better that * is not retransmitted. */ return; } seq = tp->gput_seq; ts = tp->gput_ts; rack->app_limited_needs_set = 0; tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); /* Do we start at a new end? */ if ((use_which == RACK_USE_BEG) && SEQ_GEQ(rsm->r_start, tp->gput_seq)) { /* * When we get an ACK that just eats * up some of the rsm, we set RACK_USE_BEG * since whats at r_start (i.e. th_ack) * is left unacked and thats where the * measurement not starts. */ tp->gput_seq = rsm->r_start; rack->r_ctl.rc_gp_output_ts = rsm->usec_orig_send; } if ((use_which == RACK_USE_END) && SEQ_GEQ(rsm->r_end, tp->gput_seq)) { /* * We use the end when the cumack * is moving forward and completely * deleting the rsm passed so basically * r_end holds th_ack. * * For SACK's we also want to use the end * since this piece just got sacked and * we want to target anything after that * in our measurement. */ tp->gput_seq = rsm->r_end; rack->r_ctl.rc_gp_output_ts = rsm->usec_orig_send; } if (use_which == RACK_USE_END_OR_THACK) { /* * special case for ack moving forward, * not a sack, we need to move all the * way up to where this ack cum-ack moves * to. */ if (SEQ_GT(th_ack, rsm->r_end)) tp->gput_seq = th_ack; else tp->gput_seq = rsm->r_end; rack->r_ctl.rc_gp_output_ts = rsm->usec_orig_send; } if (SEQ_GT(tp->gput_seq, tp->gput_ack)) { /* * We moved beyond this guy's range, re-calculate * the new end point. */ if (rack->rc_gp_filled == 0) { tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp))); } else { tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack); } } /* * We are moving the goal post, we may be able to clear the * measure_saw_probe_rtt flag. */ if ((rack->in_probe_rtt == 0) && (rack->measure_saw_probe_rtt) && (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit))) rack->measure_saw_probe_rtt = 0; rack_log_pacing_delay_calc(rack, ts, tp->gput_ts, seq, tp->gput_seq, 0, 5, line, NULL); if (rack->rc_gp_filled && ((tp->gput_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp))))) { /* * There is no sense of continuing this measurement * because its too small to gain us anything we * trust. Skip it and that way we can start a new * measurement quicker. */ rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq, 0, 0, 0, 6, __LINE__, NULL); tp->t_flags &= ~TF_GPUTINPROG; } } } static uint32_t rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two) { uint32_t start, end, changed = 0; struct rack_sendmap stack_map; struct rack_sendmap *rsm, *nrsm, fe, *insret, *prev, *next; int32_t used_ref = 1; int moved = 0; start = sack->start; end = sack->end; rsm = *prsm; memset(&fe, 0, sizeof(fe)); do_rest_ofb: if ((rsm == NULL) || (SEQ_LT(end, rsm->r_start)) || (SEQ_GEQ(start, rsm->r_end)) || (SEQ_LT(start, rsm->r_start))) { /* * We are not in the right spot, * find the correct spot in the tree. */ used_ref = 0; fe.r_start = start; rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); moved++; } if (rsm == NULL) { /* TSNH */ goto out; } /* Ok we have an ACK for some piece of this rsm */ if (rsm->r_start != start) { if ((rsm->r_flags & RACK_ACKED) == 0) { /** * Need to split this in two pieces the before and after, * the before remains in the map, the after must be * added. In other words we have: * rsm |--------------| * sackblk |-------> * rsm will become * rsm |---| * and nrsm will be the sacked piece * nrsm |----------| * * But before we start down that path lets * see if the sack spans over on top of * the next guy and it is already sacked. */ next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); if (next && (next->r_flags & RACK_ACKED) && SEQ_GEQ(end, next->r_start)) { /** * So the next one is already acked, and * we can thus by hookery use our stack_map * to reflect the piece being sacked and * then adjust the two tree entries moving * the start and ends around. So we start like: * rsm |------------| (not-acked) * next |-----------| (acked) * sackblk |--------> * We want to end like so: * rsm |------| (not-acked) * next |-----------------| (acked) * nrsm |-----| * Where nrsm is a temporary stack piece we * use to update all the gizmos. */ /* Copy up our fudge block */ nrsm = &stack_map; memcpy(nrsm, rsm, sizeof(struct rack_sendmap)); /* Now adjust our tree blocks */ rsm->r_end = start; next->r_start = start; /* Clear out the dup ack count of the remainder */ rsm->r_dupack = 0; rsm->r_just_ret = 0; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); /* Now lets make sure our fudge block is right */ nrsm->r_start = start; /* Now lets update all the stats and such */ rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0); if (rack->app_limited_needs_set) rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END); changed += (nrsm->r_end - nrsm->r_start); rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start); if (nrsm->r_flags & RACK_SACK_PASSED) { counter_u64_add(rack_reorder_seen, 1); rack->r_ctl.rc_reorder_ts = cts; } /* * Now we want to go up from rsm (the * one left un-acked) to the next one * in the tmap. We do this so when * we walk backwards we include marking * sack-passed on rsm (The one passed in * is skipped since it is generally called * on something sacked before removing it * from the tmap). */ if (rsm->r_in_tmap) { nrsm = TAILQ_NEXT(rsm, r_tnext); /* * Now that we have the next * one walk backwards from there. */ if (nrsm && nrsm->r_in_tmap) rack_log_sack_passed(tp, rack, nrsm); } /* Now are we done? */ if (SEQ_LT(end, next->r_end) || (end == next->r_end)) { /* Done with block */ goto out; } counter_u64_add(rack_sack_used_next_merge, 1); /* Postion for the next block */ start = next->r_end; rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next); if (rsm == NULL) goto out; } else { /** * We can't use any hookery here, so we * need to split the map. We enter like * so: * rsm |--------| * sackblk |-----> * We will add the new block nrsm and * that will be the new portion, and then * fall through after reseting rsm. So we * split and look like this: * rsm |----| * sackblk |-----> * nrsm |---| * We then fall through reseting * rsm to nrsm, so the next block * picks it up. */ nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT); if (nrsm == NULL) { /* * failed XXXrrs what can we do but loose the sack * info? */ goto out; } counter_u64_add(rack_sack_splits, 1); rack_clone_rsm(rack, nrsm, rsm, start); rsm->r_just_ret = 0; insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #ifdef INVARIANTS if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~RACK_HAS_FIN); /* Position us to point to the new nrsm that starts the sack blk */ rsm = nrsm; } } else { /* Already sacked this piece */ counter_u64_add(rack_sack_skipped_acked, 1); moved++; if (end == rsm->r_end) { /* Done with block */ rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); goto out; } else if (SEQ_LT(end, rsm->r_end)) { /* A partial sack to a already sacked block */ moved++; rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); goto out; } else { /* * The end goes beyond this guy * repostion the start to the * next block. */ start = rsm->r_end; rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); if (rsm == NULL) goto out; } } } if (SEQ_GEQ(end, rsm->r_end)) { /** * The end of this block is either beyond this guy or right * at this guy. I.e.: * rsm --- |-----| * end |-----| * * end |---------| */ if ((rsm->r_flags & RACK_ACKED) == 0) { rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0); changed += (rsm->r_end - rsm->r_start); rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); if (rsm->r_in_tmap) /* should be true */ rack_log_sack_passed(tp, rack, rsm); /* Is Reordering occuring? */ if (rsm->r_flags & RACK_SACK_PASSED) { rsm->r_flags &= ~RACK_SACK_PASSED; counter_u64_add(rack_reorder_seen, 1); rack->r_ctl.rc_reorder_ts = cts; } if (rack->app_limited_needs_set) rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END); rsm->r_ack_arrival = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); rsm->r_flags |= RACK_ACKED; rsm->r_flags &= ~RACK_TLP; if (rsm->r_in_tmap) { TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } } else { counter_u64_add(rack_sack_skipped_acked, 1); moved++; } if (end == rsm->r_end) { /* This block only - done, setup for next */ goto out; } /* * There is more not coverend by this rsm move on * to the next block in the RB tree. */ nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); start = rsm->r_end; rsm = nrsm; if (rsm == NULL) goto out; goto do_rest_ofb; } /** * The end of this sack block is smaller than * our rsm i.e.: * rsm --- |-----| * end |--| */ if ((rsm->r_flags & RACK_ACKED) == 0) { prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); if (prev && (prev->r_flags & RACK_ACKED)) { /** * Goal, we want the right remainder of rsm to shrink * in place and span from (rsm->r_start = end) to rsm->r_end. * We want to expand prev to go all the way * to prev->r_end <- end. * so in the tree we have before: * prev |--------| (acked) * rsm |-------| (non-acked) * sackblk |-| * We churn it so we end up with * prev |----------| (acked) * rsm |-----| (non-acked) * nrsm |-| (temporary) */ nrsm = &stack_map; memcpy(nrsm, rsm, sizeof(struct rack_sendmap)); prev->r_end = end; rsm->r_start = end; /* Now adjust nrsm (stack copy) to be * the one that is the small * piece that was "sacked". */ nrsm->r_end = end; rsm->r_dupack = 0; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); /* * Now nrsm is our new little piece * that is acked (which was merged * to prev). Update the rtt and changed * based on that. Also check for reordering. */ rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0); if (rack->app_limited_needs_set) rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END); changed += (nrsm->r_end - nrsm->r_start); rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start); if (nrsm->r_flags & RACK_SACK_PASSED) { counter_u64_add(rack_reorder_seen, 1); rack->r_ctl.rc_reorder_ts = cts; } rsm = prev; counter_u64_add(rack_sack_used_prev_merge, 1); } else { /** * This is the case where our previous * block is not acked either, so we must * split the block in two. */ nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT); if (nrsm == NULL) { /* failed rrs what can we do but loose the sack info? */ goto out; } /** * In this case nrsm becomes * nrsm->r_start = end; * nrsm->r_end = rsm->r_end; * which is un-acked. * * rsm->r_end = nrsm->r_start; * i.e. the remaining un-acked * piece is left on the left * hand side. * * So we start like this * rsm |----------| (not acked) * sackblk |---| * build it so we have * rsm |---| (acked) * nrsm |------| (not acked) */ counter_u64_add(rack_sack_splits, 1); rack_clone_rsm(rack, nrsm, rsm, end); rsm->r_flags &= (~RACK_HAS_FIN); rsm->r_just_ret = 0; insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #ifdef INVARIANTS if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } nrsm->r_dupack = 0; rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2); rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0); changed += (rsm->r_end - rsm->r_start); rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); if (rsm->r_in_tmap) /* should be true */ rack_log_sack_passed(tp, rack, rsm); /* Is Reordering occuring? */ if (rsm->r_flags & RACK_SACK_PASSED) { rsm->r_flags &= ~RACK_SACK_PASSED; counter_u64_add(rack_reorder_seen, 1); rack->r_ctl.rc_reorder_ts = cts; } if (rack->app_limited_needs_set) rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END); rsm->r_ack_arrival = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); rsm->r_flags |= RACK_ACKED; rsm->r_flags &= ~RACK_TLP; if (rsm->r_in_tmap) { TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } } } else if (start != end){ /* * The block was already acked. */ counter_u64_add(rack_sack_skipped_acked, 1); moved++; } out: if (rsm && (rsm->r_flags & RACK_ACKED)) { /* * Now can we merge where we worked * with either the previous or * next block? */ next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); while (next) { if (next->r_flags & RACK_ACKED) { /* yep this and next can be merged */ rsm = rack_merge_rsm(rack, rsm, next); next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); } else break; } /* Now what about the previous? */ prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); while (prev) { if (prev->r_flags & RACK_ACKED) { /* yep the previous and this can be merged */ rsm = rack_merge_rsm(rack, prev, rsm); prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); } else break; } } if (used_ref == 0) { counter_u64_add(rack_sack_proc_all, 1); } else { counter_u64_add(rack_sack_proc_short, 1); } /* Save off the next one for quick reference. */ if (rsm) nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); else nrsm = NULL; *prsm = rack->r_ctl.rc_sacklast = nrsm; /* Pass back the moved. */ *moved_two = moved; return (changed); } static void inline rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack) { struct rack_sendmap *tmap; tmap = NULL; while (rsm && (rsm->r_flags & RACK_ACKED)) { /* Its no longer sacked, mark it so */ rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); #ifdef INVARIANTS if (rsm->r_in_tmap) { panic("rack:%p rsm:%p flags:0x%x in tmap?", rack, rsm, rsm->r_flags); } #endif rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS); /* Rebuild it into our tmap */ if (tmap == NULL) { TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext); tmap = rsm; } else { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext); tmap = rsm; } tmap->r_in_tmap = 1; rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); } /* * Now lets possibly clear the sack filter so we start * recognizing sacks that cover this area. */ sack_filter_clear(&rack->r_ctl.rack_sf, th_ack); } static void rack_do_decay(struct tcp_rack *rack) { struct timeval res; #define timersub(tvp, uvp, vvp) \ do { \ (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec; \ (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec; \ if ((vvp)->tv_usec < 0) { \ (vvp)->tv_sec--; \ (vvp)->tv_usec += 1000000; \ } \ } while (0) timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res); #undef timersub rack->r_ctl.input_pkt++; if ((rack->rc_in_persist) || (res.tv_sec >= 1) || (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) { /* * Check for decay of non-SAD, * we want all SAD detection metrics to * decay 1/4 per second (or more) passed. */ uint32_t pkt_delta; pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt; /* Update our saved tracking values */ rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt; rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time; /* Now do we escape without decay? */ #ifdef NETFLIX_EXP_DETECTION if (rack->rc_in_persist || (rack->rc_tp->snd_max == rack->rc_tp->snd_una) || (pkt_delta < tcp_sad_low_pps)){ /* * We don't decay idle connections * or ones that have a low input pps. */ return; } /* Decay the counters */ rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count, tcp_sad_decay_val); rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count, tcp_sad_decay_val); rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra, tcp_sad_decay_val); rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move, tcp_sad_decay_val); #endif } } static void rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th) { uint32_t changed, entered_recovery = 0; struct tcp_rack *rack; struct rack_sendmap *rsm, *rm; struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1]; register uint32_t th_ack; int32_t i, j, k, num_sack_blks = 0; uint32_t cts, acked, ack_point, sack_changed = 0; int loop_start = 0, moved_two = 0; uint32_t tsused; INP_WLOCK_ASSERT(tp->t_inpcb); if (th->th_flags & TH_RST) { /* We don't log resets */ return; } rack = (struct tcp_rack *)tp->t_fb_ptr; cts = tcp_ts_getticks(); rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); changed = 0; th_ack = th->th_ack; if (rack->sack_attack_disable == 0) rack_do_decay(rack); if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) { /* * You only get credit for * MSS and greater (and you get extra * credit for larger cum-ack moves). */ int ac; ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp); rack->r_ctl.ack_count += ac; counter_u64_add(rack_ack_total, ac); } if (rack->r_ctl.ack_count > 0xfff00000) { /* * reduce the number to keep us under * a uint32_t. */ rack->r_ctl.ack_count /= 2; rack->r_ctl.sack_count /= 2; } if (SEQ_GT(th_ack, tp->snd_una)) { rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__); tp->t_acktime = ticks; } if (rsm && SEQ_GT(th_ack, rsm->r_start)) changed = th_ack - rsm->r_start; if (changed) { /* * The ACK point is advancing to th_ack, we must drop off * the packets in the rack log and calculate any eligble * RTT's. */ rack->r_wanted_output = 1; more: rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm == NULL) { if ((th_ack - 1) == tp->iss) { /* * For the SYN incoming case we will not * have called tcp_output for the sending of * the SYN, so there will be no map. All * other cases should probably be a panic. */ goto proc_sack; } if (tp->t_flags & TF_SENTFIN) { /* if we send a FIN we will not hav a map */ goto proc_sack; } #ifdef INVARIANTS panic("No rack map tp:%p for th:%p state:%d rack:%p snd_una:%u snd_max:%u snd_nxt:%u chg:%d\n", tp, th, tp->t_state, rack, tp->snd_una, tp->snd_max, tp->snd_nxt, changed); #endif goto proc_sack; } if (SEQ_LT(th_ack, rsm->r_start)) { /* Huh map is missing this */ #ifdef INVARIANTS printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n", rsm->r_start, th_ack, tp->t_state, rack->r_state); #endif goto proc_sack; } rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack); /* Now do we consume the whole thing? */ if (SEQ_GEQ(th_ack, rsm->r_end)) { /* Its all consumed. */ uint32_t left; uint8_t newly_acked; rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; rsm->r_rtr_bytes = 0; /* Record the time of highest cumack sent */ rack->r_ctl.rc_gp_cumack_ts = rsm->usec_orig_send; rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #ifdef INVARIANTS if (rm != rsm) { panic("removing head in rack:%p rsm:%p rm:%p", rack, rsm, rm); } #endif if (rsm->r_in_tmap) { TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } newly_acked = 1; if (rsm->r_flags & RACK_ACKED) { /* * It was acked on the scoreboard -- remove * it from total */ rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); newly_acked = 0; } else if (rsm->r_flags & RACK_SACK_PASSED) { /* * There are segments ACKED on the * scoreboard further up. We are seeing * reordering. */ rsm->r_flags &= ~RACK_SACK_PASSED; counter_u64_add(rack_reorder_seen, 1); rsm->r_ack_arrival = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); rsm->r_flags |= RACK_ACKED; rack->r_ctl.rc_reorder_ts = cts; } left = th_ack - rsm->r_end; if (rack->app_limited_needs_set && newly_acked) rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK); /* Free back to zone */ rack_free(rack, rsm); if (left) { goto more; } goto proc_sack; } if (rsm->r_flags & RACK_ACKED) { /* * It was acked on the scoreboard -- remove it from * total for the part being cum-acked. */ rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start); } /* * Clear the dup ack count for * the piece that remains. */ rsm->r_dupack = 0; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); if (rsm->r_rtr_bytes) { /* * It was retransmitted adjust the * sack holes for what was acked. */ int ack_am; ack_am = (th_ack - rsm->r_start); if (ack_am >= rsm->r_rtr_bytes) { rack->r_ctl.rc_holes_rxt -= ack_am; rsm->r_rtr_bytes -= ack_am; } } /* * Update where the piece starts and record * the time of send of highest cumack sent. */ rack->r_ctl.rc_gp_cumack_ts = rsm->usec_orig_send; rsm->r_start = th_ack; if (rack->app_limited_needs_set) rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG); } proc_sack: /* Check for reneging */ rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) { /* * The peer has moved snd_una up to * the edge of this send, i.e. one * that it had previously acked. The only * way that can be true if the peer threw * away data (space issues) that it had * previously sacked (else it would have * given us snd_una up to (rsm->r_end). * We need to undo the acked markings here. * * Note we have to look to make sure th_ack is * our rsm->r_start in case we get an old ack * where th_ack is behind snd_una. */ rack_peer_reneges(rack, rsm, th->th_ack); } if ((to->to_flags & TOF_SACK) == 0) { /* We are done nothing left */ goto out; } /* Sack block processing */ if (SEQ_GT(th_ack, tp->snd_una)) ack_point = th_ack; else ack_point = tp->snd_una; for (i = 0; i < to->to_nsacks; i++) { bcopy((to->to_sacks + i * TCPOLEN_SACK), &sack, sizeof(sack)); sack.start = ntohl(sack.start); sack.end = ntohl(sack.end); if (SEQ_GT(sack.end, sack.start) && SEQ_GT(sack.start, ack_point) && SEQ_LT(sack.start, tp->snd_max) && SEQ_GT(sack.end, ack_point) && SEQ_LEQ(sack.end, tp->snd_max)) { sack_blocks[num_sack_blks] = sack; num_sack_blks++; #ifdef NETFLIX_STATS } else if (SEQ_LEQ(sack.start, th_ack) && SEQ_LEQ(sack.end, th_ack)) { /* * Its a D-SACK block. */ tcp_record_dsack(sack.start, sack.end); #endif } } /* * Sort the SACK blocks so we can update the rack scoreboard with * just one pass. */ num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks, num_sack_blks, th->th_ack); ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks); if (num_sack_blks == 0) { /* Nothing to sack (DSACKs?) */ goto out_with_totals; } if (num_sack_blks < 2) { /* Only one, we don't need to sort */ goto do_sack_work; } /* Sort the sacks */ for (i = 0; i < num_sack_blks; i++) { for (j = i + 1; j < num_sack_blks; j++) { if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { sack = sack_blocks[i]; sack_blocks[i] = sack_blocks[j]; sack_blocks[j] = sack; } } } /* * Now are any of the sack block ends the same (yes some * implementations send these)? */ again: if (num_sack_blks == 0) goto out_with_totals; if (num_sack_blks > 1) { for (i = 0; i < num_sack_blks; i++) { for (j = i + 1; j < num_sack_blks; j++) { if (sack_blocks[i].end == sack_blocks[j].end) { /* * Ok these two have the same end we * want the smallest end and then * throw away the larger and start * again. */ if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) { /* * The second block covers * more area use that */ sack_blocks[i].start = sack_blocks[j].start; } /* * Now collapse out the dup-sack and * lower the count */ for (k = (j + 1); k < num_sack_blks; k++) { sack_blocks[j].start = sack_blocks[k].start; sack_blocks[j].end = sack_blocks[k].end; j++; } num_sack_blks--; goto again; } } } } do_sack_work: /* * First lets look to see if * we have retransmitted and * can use the transmit next? */ rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rsm && SEQ_GT(sack_blocks[0].end, rsm->r_start) && SEQ_LT(sack_blocks[0].start, rsm->r_end)) { /* * We probably did the FR and the next * SACK in continues as we would expect. */ acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two); if (acked) { rack->r_wanted_output = 1; changed += acked; sack_changed += acked; } if (num_sack_blks == 1) { /* * This is what we would expect from * a normal implementation to happen * after we have retransmitted the FR, * i.e the sack-filter pushes down * to 1 block and the next to be retransmitted * is the sequence in the sack block (has more * are acked). Count this as ACK'd data to boost * up the chances of recovering any false positives. */ rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp)); counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp))); counter_u64_add(rack_express_sack, 1); if (rack->r_ctl.ack_count > 0xfff00000) { /* * reduce the number to keep us under * a uint32_t. */ rack->r_ctl.ack_count /= 2; rack->r_ctl.sack_count /= 2; } goto out_with_totals; } else { /* * Start the loop through the * rest of blocks, past the first block. */ moved_two = 0; loop_start = 1; } } /* Its a sack of some sort */ rack->r_ctl.sack_count++; if (rack->r_ctl.sack_count > 0xfff00000) { /* * reduce the number to keep us under * a uint32_t. */ rack->r_ctl.ack_count /= 2; rack->r_ctl.sack_count /= 2; } counter_u64_add(rack_sack_total, 1); if (rack->sack_attack_disable) { /* An attacker disablement is in place */ if (num_sack_blks > 1) { rack->r_ctl.sack_count += (num_sack_blks - 1); rack->r_ctl.sack_moved_extra++; counter_u64_add(rack_move_some, 1); if (rack->r_ctl.sack_moved_extra > 0xfff00000) { rack->r_ctl.sack_moved_extra /= 2; rack->r_ctl.sack_noextra_move /= 2; } } goto out; } rsm = rack->r_ctl.rc_sacklast; for (i = loop_start; i < num_sack_blks; i++) { acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two); if (acked) { rack->r_wanted_output = 1; changed += acked; sack_changed += acked; } if (moved_two) { /* * If we did not get a SACK for at least a MSS and * had to move at all, or if we moved more than our * threshold, it counts against the "extra" move. */ rack->r_ctl.sack_moved_extra += moved_two; counter_u64_add(rack_move_some, 1); } else { /* * else we did not have to move * any more than we would expect. */ rack->r_ctl.sack_noextra_move++; counter_u64_add(rack_move_none, 1); } if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) { /* * If the SACK was not a full MSS then * we add to sack_count the number of * MSS's (or possibly more than * a MSS if its a TSO send) we had to skip by. */ rack->r_ctl.sack_count += moved_two; counter_u64_add(rack_sack_total, moved_two); } /* * Now we need to setup for the next * round. First we make sure we won't * exceed the size of our uint32_t on * the various counts, and then clear out * moved_two. */ if ((rack->r_ctl.sack_moved_extra > 0xfff00000) || (rack->r_ctl.sack_noextra_move > 0xfff00000)) { rack->r_ctl.sack_moved_extra /= 2; rack->r_ctl.sack_noextra_move /= 2; } if (rack->r_ctl.sack_count > 0xfff00000) { rack->r_ctl.ack_count /= 2; rack->r_ctl.sack_count /= 2; } moved_two = 0; } out_with_totals: if (num_sack_blks > 1) { /* * You get an extra stroke if * you have more than one sack-blk, this * could be where we are skipping forward * and the sack-filter is still working, or * it could be an attacker constantly * moving us. */ rack->r_ctl.sack_moved_extra++; counter_u64_add(rack_move_some, 1); } out: #ifdef NETFLIX_EXP_DETECTION if ((rack->do_detection || tcp_force_detection) && tcp_sack_to_ack_thresh && tcp_sack_to_move_thresh && ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) { /* * We have thresholds set to find * possible attackers and disable sack. * Check them. */ uint64_t ackratio, moveratio, movetotal; /* Log detecting */ rack_log_sad(rack, 1); ackratio = (uint64_t)(rack->r_ctl.sack_count); ackratio *= (uint64_t)(1000); if (rack->r_ctl.ack_count) ackratio /= (uint64_t)(rack->r_ctl.ack_count); else { /* We really should not hit here */ ackratio = 1000; } if ((rack->sack_attack_disable == 0) && (ackratio > rack_highest_sack_thresh_seen)) rack_highest_sack_thresh_seen = (uint32_t)ackratio; movetotal = rack->r_ctl.sack_moved_extra; movetotal += rack->r_ctl.sack_noextra_move; moveratio = rack->r_ctl.sack_moved_extra; moveratio *= (uint64_t)1000; if (movetotal) moveratio /= movetotal; else { /* No moves, thats pretty good */ moveratio = 0; } if ((rack->sack_attack_disable == 0) && (moveratio > rack_highest_move_thresh_seen)) rack_highest_move_thresh_seen = (uint32_t)moveratio; if (rack->sack_attack_disable == 0) { if ((ackratio > tcp_sack_to_ack_thresh) && (moveratio > tcp_sack_to_move_thresh)) { /* Disable sack processing */ rack->sack_attack_disable = 1; if (rack->r_rep_attack == 0) { rack->r_rep_attack = 1; counter_u64_add(rack_sack_attacks_detected, 1); } if (tcp_attack_on_turns_on_logging) { /* * Turn on logging, used for debugging * false positives. */ rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging; } /* Clamp the cwnd at flight size */ rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd; rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); rack_log_sad(rack, 2); } } else { /* We are sack-disabled check for false positives */ if ((ackratio <= tcp_restoral_thresh) || (rack->r_ctl.rc_num_maps_alloced < tcp_map_minimum)) { rack->sack_attack_disable = 0; rack_log_sad(rack, 3); /* Restart counting */ rack->r_ctl.sack_count = 0; rack->r_ctl.sack_moved_extra = 0; rack->r_ctl.sack_noextra_move = 1; rack->r_ctl.ack_count = max(1, (BYTES_THIS_ACK(tp, th)/ctf_fixed_maxseg(rack->rc_tp))); if (rack->r_rep_reverse == 0) { rack->r_rep_reverse = 1; counter_u64_add(rack_sack_attacks_reversed, 1); } /* Restore the cwnd */ if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd) rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd; } } } #endif if (changed) { /* Something changed cancel the rack timer */ rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); } tsused = tcp_ts_getticks(); rsm = tcp_rack_output(tp, rack, tsused); if ((!IN_RECOVERY(tp->t_flags)) && rsm) { /* Enter recovery */ rack->r_ctl.rc_rsm_start = rsm->r_start; rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; entered_recovery = 1; rack_cong_signal(tp, NULL, CC_NDUPACK); /* * When we enter recovery we need to assure we send * one packet. */ if (rack->rack_no_prr == 0) { rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp); rack_log_to_prr(rack, 8, 0); } rack->r_timer_override = 1; rack->r_early = 0; rack->r_ctl.rc_agg_early = 0; } else if (IN_RECOVERY(tp->t_flags) && rsm && (rack->r_rr_config == 3)) { /* * Assure we can output and we get no * remembered pace time except the retransmit. */ rack->r_timer_override = 1; rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT; rack->r_ctl.rc_resend = rsm; } if (IN_RECOVERY(tp->t_flags) && (rack->rack_no_prr == 0) && (entered_recovery == 0)) { /* Deal with PRR here (in recovery only) */ uint32_t pipe, snd_una; rack->r_ctl.rc_prr_delivered += changed; /* Compute prr_sndcnt */ if (SEQ_GT(tp->snd_una, th_ack)) { snd_una = tp->snd_una; } else { snd_una = th_ack; } pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt; if (pipe > tp->snd_ssthresh) { long sndcnt; sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh; if (rack->r_ctl.rc_prr_recovery_fs > 0) sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs; else { rack->r_ctl.rc_prr_sndcnt = 0; rack_log_to_prr(rack, 9, 0); sndcnt = 0; } sndcnt++; if (sndcnt > (long)rack->r_ctl.rc_prr_out) sndcnt -= rack->r_ctl.rc_prr_out; else sndcnt = 0; rack->r_ctl.rc_prr_sndcnt = sndcnt; rack_log_to_prr(rack, 10, 0); } else { uint32_t limit; if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out) limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out); else limit = 0; if (changed > limit) limit = changed; limit += ctf_fixed_maxseg(tp); if (tp->snd_ssthresh > pipe) { rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit); rack_log_to_prr(rack, 11, 0); } else { rack->r_ctl.rc_prr_sndcnt = min(0, limit); rack_log_to_prr(rack, 12, 0); } } if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) && ((rack->rc_inp->inp_in_hpts == 0) && ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) { /* * If you are pacing output you don't want * to override. */ rack->r_early = 0; rack->r_ctl.rc_agg_early = 0; rack->r_timer_override = 1; } } } static void rack_strike_dupack(struct tcp_rack *rack) { struct rack_sendmap *rsm; rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rsm && (rsm->r_dupack < 0xff)) { rsm->r_dupack++; if (rsm->r_dupack >= DUP_ACK_THRESHOLD) { rack->r_wanted_output = 1; rack->r_timer_override = 1; rack_log_retran_reason(rack, rsm, __LINE__, 1, 3); } else { rack_log_retran_reason(rack, rsm, __LINE__, 0, 3); } } } static void rack_check_bottom_drag(struct tcpcb *tp, struct tcp_rack *rack, struct socket *so, int32_t acked) { uint32_t segsiz, minseg; segsiz = ctf_fixed_maxseg(tp); minseg = segsiz; if (tp->snd_max == tp->snd_una) { /* * We are doing dynamic pacing and we are way * under. Basically everything got acked while * we were still waiting on the pacer to expire. * * This means we need to boost the b/w in * addition to any earlier boosting of * the multipler. */ rack->rc_dragged_bottom = 1; rack_validate_multipliers_at_or_above100(rack); /* * Lets use the segment bytes acked plus * the lowest RTT seen as the basis to * form a b/w estimate. This will be off * due to the fact that the true estimate * should be around 1/2 the time of the RTT * but we can settle for that. */ if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) && acked) { uint64_t bw, calc_bw, rtt; rtt = rack->r_ctl.rack_rs.rs_us_rtt; bw = acked; calc_bw = bw * 1000000; calc_bw /= rtt; if (rack->r_ctl.last_max_bw && (rack->r_ctl.last_max_bw < calc_bw)) { /* * If we have a last calculated max bw * enforce it. */ calc_bw = rack->r_ctl.last_max_bw; } /* now plop it in */ if (rack->rc_gp_filled == 0) { if (calc_bw > ONE_POINT_TWO_MEG) { /* * If we have no measurement * don't let us set in more than * 1.2Mbps. If we are still too * low after pacing with this we * will hopefully have a max b/w * available to sanity check things. */ calc_bw = ONE_POINT_TWO_MEG; } rack->r_ctl.rc_rtt_diff = 0; rack->r_ctl.gp_bw = calc_bw; rack->rc_gp_filled = 1; rack->r_ctl.num_avg = RACK_REQ_AVG; rack_set_pace_segments(rack->rc_tp, rack, __LINE__); } else if (calc_bw > rack->r_ctl.gp_bw) { rack->r_ctl.rc_rtt_diff = 0; rack->r_ctl.num_avg = RACK_REQ_AVG; rack->r_ctl.gp_bw = calc_bw; rack_set_pace_segments(rack->rc_tp, rack, __LINE__); } else rack_increase_bw_mul(rack, -1, 0, 0, 1); /* * For acks over 1mss we do a extra boost to simulate * where we would get 2 acks (we want 110 for the mul). */ if (acked > segsiz) rack_increase_bw_mul(rack, -1, 0, 0, 1); } else { /* * Huh, this should not be, settle * for just an old increase. */ rack_increase_bw_mul(rack, -1, 0, 0, 1); } } else if ((IN_RECOVERY(tp->t_flags) == 0) && (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)), minseg)) && (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) && (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) && (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <= (segsiz * rack_req_segs))) { /* * We are doing dynamic GP pacing and * we have everything except 1MSS or less * bytes left out. We are still pacing away. * And there is data that could be sent, This * means we are inserting delayed ack time in * our measurements because we are pacing too slow. */ rack_validate_multipliers_at_or_above100(rack); rack->rc_dragged_bottom = 1; rack_increase_bw_mul(rack, -1, 0, 0, 1); } } /* * Return value of 1, we do not need to call rack_process_data(). * return value of 0, rack_process_data can be called. * For ret_val if its 0 the TCP is locked, if its non-zero * its unlocked and probably unsafe to touch the TCB. */ static int rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val) { int32_t ourfinisacked = 0; int32_t nsegs, acked_amount; int32_t acked; struct mbuf *mfree; struct tcp_rack *rack; int32_t under_pacing = 0; int32_t recovery = 0; rack = (struct tcp_rack *)tp->t_fb_ptr; if (SEQ_GT(th->th_ack, tp->snd_max)) { ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val); rack->r_wanted_output = 1; return (1); } if (rack->rc_gp_filled && (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { under_pacing = 1; } if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { if (rack->rc_in_persist) tp->t_rxtshift = 0; if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd)) rack_strike_dupack(rack); rack_log_ack(tp, to, th); } if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { /* * Old ack, behind (or duplicate to) the last one rcv'd * Note: Should mark reordering is occuring! We should also * look for sack blocks arriving e.g. ack 1, 4-4 then ack 1, * 3-3, 4-4 would be reording. As well as ack 1, 3-3 ack 3 */ return (0); } /* * If we reach this point, ACK is not a duplicate, i.e., it ACKs * something we sent. */ if (tp->t_flags & TF_NEEDSYN) { /* * T/TCP: Connection was half-synchronized, and our SYN has * been ACK'd (so connection is now fully synchronized). Go * to non-starred state, increment snd_una for ACK of SYN, * and check if we can do window scaling. */ tp->t_flags &= ~TF_NEEDSYN; tp->snd_una++; /* Do window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; /* Send window already scaled. */ } } nsegs = max(1, m->m_pkthdr.lro_nsegs); INP_WLOCK_ASSERT(tp->t_inpcb); acked = BYTES_THIS_ACK(tp, th); KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs); KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); /* * If we just performed our first retransmit, and the ACK arrives * within our recovery window, then it was a mistake to do the * retransmit in the first place. Recover our original cwnd and * ssthresh, and proceed to transmit where we left off. */ if (tp->t_flags & TF_PREVVALID) { tp->t_flags &= ~TF_PREVVALID; if (tp->t_rxtshift == 1 && (int)(ticks - tp->t_badrxtwin) < 0) rack_cong_signal(tp, th, CC_RTO_ERR); } if (acked) { /* assure we are not backed off */ tp->t_rxtshift = 0; rack->rc_tlp_in_progress = 0; rack->r_ctl.rc_tlp_cnt_out = 0; /* * If it is the RXT timer we want to * stop it, so we can restart a TLP. */ if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); #ifdef NETFLIX_HTTP_LOGGING tcp_http_check_for_comp(rack->rc_tp, th->th_ack); #endif } /* * If we have a timestamp reply, update smoothed round trip time. If * no timestamp is present but transmit timer is running and timed * sequence number was acked, update smoothed round trip time. Since * we now have an rtt measurement, cancel the timer backoff (cf., * Phil Karn's retransmit alg.). Recompute the initial retransmit * timer. * * Some boxes send broken timestamp replies during the SYN+ACK * phase, ignore timestamps of 0 or we could calculate a huge RTT * and blow up the retransmit timer. */ /* * If all outstanding data is acked, stop retransmit timer and * remember to restart (more output or persist). If there is more * data to be acked, restart retransmit timer, using current * (possibly backed-off) value. */ if (acked == 0) { if (ofia) *ofia = ourfinisacked; return (0); } if (rack->r_ctl.rc_early_recovery) { if (IN_RECOVERY(tp->t_flags)) { if (SEQ_LT(th->th_ack, tp->snd_recover) && (SEQ_LT(th->th_ack, tp->snd_max))) { tcp_rack_partialack(tp, th); } else { rack_post_recovery(tp, th); recovery = 1; } } } /* * Let the congestion control algorithm update congestion control * related information. This typically means increasing the * congestion window. */ rack_ack_received(tp, rack, th, nsegs, CC_ACK, recovery); SOCKBUF_LOCK(&so->so_snd); acked_amount = min(acked, (int)sbavail(&so->so_snd)); tp->snd_wnd -= acked_amount; mfree = sbcut_locked(&so->so_snd, acked_amount); if ((sbused(&so->so_snd) == 0) && (acked > acked_amount) && (tp->t_state >= TCPS_FIN_WAIT_1) && (tp->t_flags & TF_SENTFIN)) { /* * We must be sure our fin * was sent and acked (we can be * in FIN_WAIT_1 without having * sent the fin). */ ourfinisacked = 1; } SOCKBUF_UNLOCK(&so->so_snd); tp->t_flags |= TF_WAKESOW; m_freem(mfree); if (rack->r_ctl.rc_early_recovery == 0) { if (IN_RECOVERY(tp->t_flags)) { if (SEQ_LT(th->th_ack, tp->snd_recover) && (SEQ_LT(th->th_ack, tp->snd_max))) { tcp_rack_partialack(tp, th); } else { rack_post_recovery(tp, th); } } } tp->snd_una = th->th_ack; if (SEQ_GT(tp->snd_una, tp->snd_recover)) tp->snd_recover = tp->snd_una; if (SEQ_LT(tp->snd_nxt, tp->snd_una)) { tp->snd_nxt = tp->snd_una; } if (under_pacing && (rack->use_fixed_rate == 0) && (rack->in_probe_rtt == 0) && rack->rc_gp_dyn_mul && rack->rc_always_pace) { /* Check if we are dragging bottom */ rack_check_bottom_drag(tp, rack, so, acked); } if (tp->snd_una == tp->snd_max) { /* Nothing left outstanding */ rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL); if (rack->r_ctl.rc_went_idle_time == 0) rack->r_ctl.rc_went_idle_time = 1; rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0) tp->t_acktime = 0; rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); /* Set need output so persist might get set */ rack->r_wanted_output = 1; sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); if ((tp->t_state >= TCPS_FIN_WAIT_1) && (sbavail(&so->so_snd) == 0) && (tp->t_flags2 & TF2_DROP_AF_DATA)) { /* * The socket was gone and the * peer sent data, time to * reset him. */ *ret_val = 1; /* tcp_close will kill the inp pre-log the Reset */ tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); tp = tcp_close(tp); ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen); return (1); } } if (ofia) *ofia = ourfinisacked; return (0); } static void rack_collapsed_window(struct tcp_rack *rack) { /* * Now we must walk the * send map and divide the * ones left stranded. These * guys can't cause us to abort * the connection and are really * "unsent". However if a buggy * client actually did keep some * of the data i.e. collapsed the win * and refused to ack and then opened * the win and acked that data. We would * get into an ack war, the simplier * method then of just pretending we * did not send those segments something * won't work. */ struct rack_sendmap *rsm, *nrsm, fe, *insret; tcp_seq max_seq; max_seq = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd; memset(&fe, 0, sizeof(fe)); fe.r_start = max_seq; /* Find the first seq past or at maxseq */ rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); if (rsm == NULL) { /* Nothing to do strange */ rack->rc_has_collapsed = 0; return; } /* * Now do we need to split at * the collapse point? */ if (SEQ_GT(max_seq, rsm->r_start)) { nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT); if (nrsm == NULL) { /* We can't get a rsm, mark all? */ nrsm = rsm; goto no_split; } /* Clone it */ rack_clone_rsm(rack, nrsm, rsm, max_seq); insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #ifdef INVARIANTS if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } /* * Set in the new RSM as the * collapsed starting point */ rsm = nrsm; } no_split: counter_u64_add(rack_collapsed_win, 1); RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) { nrsm->r_flags |= RACK_RWND_COLLAPSED; rack->rc_has_collapsed = 1; } } static void rack_un_collapse_window(struct tcp_rack *rack) { struct rack_sendmap *rsm; RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) { if (rsm->r_flags & RACK_RWND_COLLAPSED) rsm->r_flags &= ~RACK_RWND_COLLAPSED; else break; } rack->rc_has_collapsed = 0; } static void rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack, int32_t tlen, int32_t tfo_syn) { if (DELAY_ACK(tp, tlen) || tfo_syn) { if (rack->rc_dack_mode && (tlen > 500) && (rack->rc_dack_toggle == 1)) { goto no_delayed_ack; } rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tp->t_flags |= TF_DELACK; } else { no_delayed_ack: rack->r_wanted_output = 1; tp->t_flags |= TF_ACKNOW; if (rack->rc_dack_mode) { if (tp->t_flags & TF_DELACK) rack->rc_dack_toggle = 1; else rack->rc_dack_toggle = 0; } } } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) { /* * Update window information. Don't look at window if no ACK: TAC's * send garbage on first SYN. */ int32_t nsegs; int32_t tfo_syn; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; INP_WLOCK_ASSERT(tp->t_inpcb); nsegs = max(1, m->m_pkthdr.lro_nsegs); if ((thflags & TH_ACK) && (SEQ_LT(tp->snd_wl1, th->th_seq) || (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { /* keep track of pure window updates */ if (tlen == 0 && tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) KMOD_TCPSTAT_INC(tcps_rcvwinupd); tp->snd_wnd = tiwin; tp->snd_wl1 = th->th_seq; tp->snd_wl2 = th->th_ack; if (tp->snd_wnd > tp->max_sndwnd) tp->max_sndwnd = tp->snd_wnd; rack->r_wanted_output = 1; } else if (thflags & TH_ACK) { if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) { tp->snd_wnd = tiwin; tp->snd_wl1 = th->th_seq; tp->snd_wl2 = th->th_ack; } } if (tp->snd_wnd < ctf_outstanding(tp)) /* The peer collapsed the window */ rack_collapsed_window(rack); else if (rack->rc_has_collapsed) rack_un_collapse_window(rack); /* Was persist timer active and now we have window space? */ if ((rack->rc_in_persist != 0) && (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs))) { rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime); tp->snd_nxt = tp->snd_max; /* Make sure we output to start the timer */ rack->r_wanted_output = 1; } /* Do we enter persists? */ if ((rack->rc_in_persist == 0) && (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) && TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_max == tp->snd_una) && sbavail(&tp->t_inpcb->inp_socket->so_snd) && (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) { /* * Here the rwnd is less than * the pacing size, we are established, * nothing is outstanding, and there is * data to send. Enter persists. */ tp->snd_nxt = tp->snd_una; rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime); } if (tp->t_flags2 & TF2_DROP_AF_DATA) { m_freem(m); return (0); } /* * don't process the URG bit, ignore them drag * along the up. */ tp->rcv_up = tp->rcv_nxt; INP_WLOCK_ASSERT(tp->t_inpcb); /* * Process the segment text, merging it into the TCP sequencing * queue, and arranging for acknowledgment of receipt if necessary. * This process logically involves adjusting tp->rcv_wnd as data is * presented to the user (this happens in tcp_usrreq.c, case * PRU_RCVD). If a FIN has already been received on this connection * then we just ignore the text. */ tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && IS_FASTOPEN(tp->t_flags)); if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) && TCPS_HAVERCVDFIN(tp->t_state) == 0) { tcp_seq save_start = th->th_seq; tcp_seq save_rnxt = tp->rcv_nxt; int save_tlen = tlen; m_adj(m, drop_hdrlen); /* delayed header drop */ /* * Insert segment which includes th into TCP reassembly * queue with control block tp. Set thflags to whether * reassembly now includes a segment with FIN. This handles * the common case inline (segment is the next to be * received on an established connection, and the queue is * empty), avoiding linkage into and removal from the queue * and repetition of various conversions. Set DELACK for * segments received in order, but ack immediately when * segments are out of order (so fast retransmit can work). */ if (th->th_seq == tp->rcv_nxt && SEGQ_EMPTY(tp) && (TCPS_HAVEESTABLISHED(tp->t_state) || tfo_syn)) { #ifdef NETFLIX_SB_LIMITS u_int mcnt, appended; if (so->so_rcv.sb_shlim) { mcnt = m_memcnt(m); appended = 0; if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, CFO_NOSLEEP, NULL) == false) { counter_u64_add(tcp_sb_shlim_fails, 1); m_freem(m); return (0); } } #endif rack_handle_delayed_ack(tp, rack, tlen, tfo_syn); tp->rcv_nxt += tlen; if (tlen && ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && (tp->t_fbyte_in == 0)) { tp->t_fbyte_in = ticks; if (tp->t_fbyte_in == 0) tp->t_fbyte_in = 1; if (tp->t_fbyte_out && tp->t_fbyte_in) tp->t_flags2 |= TF2_FBYTES_COMPLETE; } thflags = th->th_flags & TH_FIN; KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs); KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); SOCKBUF_LOCK(&so->so_rcv); if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { m_freem(m); } else #ifdef NETFLIX_SB_LIMITS appended = #endif sbappendstream_locked(&so->so_rcv, m, 0); SOCKBUF_UNLOCK(&so->so_rcv); tp->t_flags |= TF_WAKESOR; #ifdef NETFLIX_SB_LIMITS if (so->so_rcv.sb_shlim && appended != mcnt) counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended); #endif } else { /* * XXX: Due to the header drop above "th" is * theoretically invalid by now. Fortunately * m_adj() doesn't actually frees any mbufs when * trimming from the head. */ tcp_seq temp = save_start; thflags = tcp_reass(tp, th, &temp, &tlen, m); tp->t_flags |= TF_ACKNOW; } if ((tp->t_flags & TF_SACK_PERMIT) && (save_tlen > 0) && TCPS_HAVEESTABLISHED(tp->t_state)) { if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) { /* * DSACK actually handled in the fastpath * above. */ RACK_OPTS_INC(tcp_sack_path_1); tcp_update_sack_list(tp, save_start, save_start + save_tlen); } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) { if ((tp->rcv_numsacks >= 1) && (tp->sackblks[0].end == save_start)) { /* * Partial overlap, recorded at todrop * above. */ RACK_OPTS_INC(tcp_sack_path_2a); tcp_update_sack_list(tp, tp->sackblks[0].start, tp->sackblks[0].end); } else { RACK_OPTS_INC(tcp_sack_path_2b); tcp_update_dsack_list(tp, save_start, save_start + save_tlen); } } else if (tlen >= save_tlen) { /* Update of sackblks. */ RACK_OPTS_INC(tcp_sack_path_3); tcp_update_dsack_list(tp, save_start, save_start + save_tlen); } else if (tlen > 0) { RACK_OPTS_INC(tcp_sack_path_4); tcp_update_dsack_list(tp, save_start, save_start + tlen); } } } else { m_freem(m); thflags &= ~TH_FIN; } /* * If FIN is received ACK the FIN and let the user know that the * connection is closing. */ if (thflags & TH_FIN) { if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { socantrcvmore(so); /* The socket upcall is handled by socantrcvmore. */ tp->t_flags &= ~TF_WAKESOR; /* * If connection is half-synchronized (ie NEEDSYN * flag on) then delay ACK, so it may be piggybacked * when SYN is sent. Otherwise, since we received a * FIN then no more input can be expected, send ACK * now. */ if (tp->t_flags & TF_NEEDSYN) { rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tp->t_flags |= TF_DELACK; } else { tp->t_flags |= TF_ACKNOW; } tp->rcv_nxt++; } switch (tp->t_state) { /* * In SYN_RECEIVED and ESTABLISHED STATES enter the * CLOSE_WAIT state. */ case TCPS_SYN_RECEIVED: tp->t_starttime = ticks; /* FALLTHROUGH */ case TCPS_ESTABLISHED: rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tcp_state_change(tp, TCPS_CLOSE_WAIT); break; /* * If still in FIN_WAIT_1 STATE FIN has not been * acked so enter the CLOSING state. */ case TCPS_FIN_WAIT_1: rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tcp_state_change(tp, TCPS_CLOSING); break; /* * In FIN_WAIT_2 state enter the TIME_WAIT state, * starting the time-wait timer, turning off the * other standard timers. */ case TCPS_FIN_WAIT_2: rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tcp_twstart(tp); return (1); } } /* * Return any desired output. */ if ((tp->t_flags & TF_ACKNOW) || (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) { rack->r_wanted_output = 1; } INP_WLOCK_ASSERT(tp->t_inpcb); return (0); } /* * Here nothing is really faster, its just that we * have broken out the fast-data path also just like * the fast-ack. */ static int rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos) { int32_t nsegs; int32_t newsize = 0; /* automatic sockbuf scaling */ struct tcp_rack *rack; #ifdef NETFLIX_SB_LIMITS u_int mcnt, appended; #endif #ifdef TCPDEBUG /* * The size of tcp_saveipgen must be the size of the max ip header, * now IPv6. */ u_char tcp_saveipgen[IP6_HDR_LEN]; struct tcphdr tcp_savetcp; short ostate = 0; #endif /* * If last ACK falls within this segment's sequence numbers, record * the timestamp. NOTE that the test is modified according to the * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). */ if (__predict_false(th->th_seq != tp->rcv_nxt)) { return (0); } if (__predict_false(tp->snd_nxt != tp->snd_max)) { return (0); } if (tiwin && tiwin != tp->snd_wnd) { return (0); } if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) { return (0); } if (__predict_false((to->to_flags & TOF_TS) && (TSTMP_LT(to->to_tsval, tp->ts_recent)))) { return (0); } if (__predict_false((th->th_ack != tp->snd_una))) { return (0); } if (__predict_false(tlen > sbspace(&so->so_rcv))) { return (0); } if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } rack = (struct tcp_rack *)tp->t_fb_ptr; /* * This is a pure, in-sequence data packet with nothing on the * reassembly queue and we have enough buffer space to take it. */ nsegs = max(1, m->m_pkthdr.lro_nsegs); #ifdef NETFLIX_SB_LIMITS if (so->so_rcv.sb_shlim) { mcnt = m_memcnt(m); appended = 0; if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, CFO_NOSLEEP, NULL) == false) { counter_u64_add(tcp_sb_shlim_fails, 1); m_freem(m); return (1); } } #endif /* Clean receiver SACK report if present */ if (tp->rcv_numsacks) tcp_clean_sackreport(tp); KMOD_TCPSTAT_INC(tcps_preddat); tp->rcv_nxt += tlen; if (tlen && ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && (tp->t_fbyte_in == 0)) { tp->t_fbyte_in = ticks; if (tp->t_fbyte_in == 0) tp->t_fbyte_in = 1; if (tp->t_fbyte_out && tp->t_fbyte_in) tp->t_flags2 |= TF2_FBYTES_COMPLETE; } /* * Pull snd_wl1 up to prevent seq wrap relative to th_seq. */ tp->snd_wl1 = th->th_seq; /* * Pull rcv_up up to prevent seq wrap relative to rcv_nxt. */ tp->rcv_up = tp->rcv_nxt; KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs); KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif newsize = tcp_autorcvbuf(m, th, so, tp, tlen); /* Add data to socket buffer. */ SOCKBUF_LOCK(&so->so_rcv); if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { m_freem(m); } else { /* * Set new socket buffer size. Give up when limit is * reached. */ if (newsize) if (!sbreserve_locked(&so->so_rcv, newsize, so, NULL)) so->so_rcv.sb_flags &= ~SB_AUTOSIZE; m_adj(m, drop_hdrlen); /* delayed header drop */ #ifdef NETFLIX_SB_LIMITS appended = #endif sbappendstream_locked(&so->so_rcv, m, 0); ctf_calc_rwin(so, tp); } SOCKBUF_UNLOCK(&so->so_rcv); tp->t_flags |= TF_WAKESOR; #ifdef NETFLIX_SB_LIMITS if (so->so_rcv.sb_shlim && mcnt != appended) counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended); #endif rack_handle_delayed_ack(tp, rack, tlen, 0); if (tp->snd_una == tp->snd_max) sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); return (1); } /* * This subfunction is used to try to highly optimize the * fast path. We again allow window updates that are * in sequence to remain in the fast-path. We also add * in the __predict's to attempt to help the compiler. * Note that if we return a 0, then we can *not* process * it and the caller should push the packet into the * slow-path. */ static int rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint32_t cts) { int32_t acked; int32_t nsegs; #ifdef TCPDEBUG /* * The size of tcp_saveipgen must be the size of the max ip header, * now IPv6. */ u_char tcp_saveipgen[IP6_HDR_LEN]; struct tcphdr tcp_savetcp; short ostate = 0; #endif int32_t under_pacing = 0; struct tcp_rack *rack; if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { /* Old ack, behind (or duplicate to) the last one rcv'd */ return (0); } if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) { /* Above what we have sent? */ return (0); } if (__predict_false(tp->snd_nxt != tp->snd_max)) { /* We are retransmitting */ return (0); } if (__predict_false(tiwin == 0)) { /* zero window */ return (0); } if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) { /* We need a SYN or a FIN, unlikely.. */ return (0); } if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) { /* Timestamp is behind .. old ack with seq wrap? */ return (0); } if (__predict_false(IN_RECOVERY(tp->t_flags))) { /* Still recovering */ return (0); } rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack->r_ctl.rc_sacked) { /* We have sack holes on our scoreboard */ return (0); } /* Ok if we reach here, we can process a fast-ack */ if (rack->rc_gp_filled && (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { under_pacing = 1; } nsegs = max(1, m->m_pkthdr.lro_nsegs); rack_log_ack(tp, to, th); /* Did the window get updated? */ if (tiwin != tp->snd_wnd) { tp->snd_wnd = tiwin; tp->snd_wl1 = th->th_seq; if (tp->snd_wnd > tp->max_sndwnd) tp->max_sndwnd = tp->snd_wnd; } /* Do we exit persists? */ if ((rack->rc_in_persist != 0) && (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs))) { rack_exit_persist(tp, rack, cts); } /* Do we enter persists? */ if ((rack->rc_in_persist == 0) && (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) && TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_max == tp->snd_una) && sbavail(&tp->t_inpcb->inp_socket->so_snd) && (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) { /* * Here the rwnd is less than * the pacing size, we are established, * nothing is outstanding, and there is * data to send. Enter persists. */ tp->snd_nxt = tp->snd_una; rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime); } /* * If last ACK falls within this segment's sequence numbers, record * the timestamp. NOTE that the test is modified according to the * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * This is a pure ack for outstanding data. */ KMOD_TCPSTAT_INC(tcps_predack); /* * "bad retransmit" recovery. */ if (tp->t_flags & TF_PREVVALID) { tp->t_flags &= ~TF_PREVVALID; if (tp->t_rxtshift == 1 && (int)(ticks - tp->t_badrxtwin) < 0) rack_cong_signal(tp, th, CC_RTO_ERR); } /* * Recalculate the transmit timer / rtt. * * Some boxes send broken timestamp replies during the SYN+ACK * phase, ignore timestamps of 0 or we could calculate a huge RTT * and blow up the retransmit timer. */ acked = BYTES_THIS_ACK(tp, th); #ifdef TCP_HHOOK /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ hhook_run_tcp_est_in(tp, th, to); #endif KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs); KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); sbdrop(&so->so_snd, acked); if (acked) { /* assure we are not backed off */ tp->t_rxtshift = 0; rack->rc_tlp_in_progress = 0; rack->r_ctl.rc_tlp_cnt_out = 0; /* * If it is the RXT timer we want to * stop it, so we can restart a TLP. */ if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); #ifdef NETFLIX_HTTP_LOGGING tcp_http_check_for_comp(rack->rc_tp, th->th_ack); #endif } /* * Let the congestion control algorithm update congestion control * related information. This typically means increasing the * congestion window. */ rack_ack_received(tp, rack, th, nsegs, CC_ACK, 0); tp->snd_una = th->th_ack; if (tp->snd_wnd < ctf_outstanding(tp)) { /* The peer collapsed the window */ rack_collapsed_window(rack); } else if (rack->rc_has_collapsed) rack_un_collapse_window(rack); /* * Pull snd_wl2 up to prevent seq wrap relative to th_ack. */ tp->snd_wl2 = th->th_ack; tp->t_dupacks = 0; m_freem(m); /* ND6_HINT(tp); *//* Some progress has been made. */ /* * If all outstanding data are acked, stop retransmit timer, * otherwise restart timer using current (possibly backed-off) * value. If process is waiting for space, wakeup/selwakeup/signal. * If data are ready to send, let tcp_output decide between more * output or persist. */ #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif if (under_pacing && (rack->use_fixed_rate == 0) && (rack->in_probe_rtt == 0) && rack->rc_gp_dyn_mul && rack->rc_always_pace) { /* Check if we are dragging bottom */ rack_check_bottom_drag(tp, rack, so, acked); } if (tp->snd_una == tp->snd_max) { rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL); if (rack->r_ctl.rc_went_idle_time == 0) rack->r_ctl.rc_went_idle_time = 1; rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0) tp->t_acktime = 0; rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); } /* Wake up the socket if we have room to write more */ tp->t_flags |= TF_WAKESOW; if (sbavail(&so->so_snd)) { rack->r_wanted_output = 1; } return (1); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; int32_t todrop; int32_t ourfinisacked = 0; struct tcp_rack *rack; ctf_calc_rwin(so, tp); /* * If the state is SYN_SENT: if seg contains an ACK, but not for our * SYN, drop the input. if seg contains a RST, then drop the * connection. if seg does not contain SYN, then drop it. Otherwise * this is an acceptable SYN segment initialize tp->rcv_nxt and * tp->irs if seg contains ack then advance tp->snd_una if seg * contains an ECE and ECN support is enabled, the stream is ECN * capable. if SYN has been acked change to ESTABLISHED else * SYN_RCVD state arrange for segment to be acked (eventually) * continue processing rest of data/controls. */ if ((thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) { TCP_PROBE5(connect__refused, NULL, tp, mtod(m, const char *), tp, th); tp = tcp_drop(tp, ECONNREFUSED); ctf_do_drop(m, tp); return (1); } if (thflags & TH_RST) { ctf_do_drop(m, tp); return (1); } if (!(thflags & TH_SYN)) { ctf_do_drop(m, tp); return (1); } tp->irs = th->th_seq; tcp_rcvseqinit(tp); rack = (struct tcp_rack *)tp->t_fb_ptr; if (thflags & TH_ACK) { int tfo_partial = 0; KMOD_TCPSTAT_INC(tcps_connects); soisconnected(so); #ifdef MAC mac_socketpeer_set_from_mbuf(m, so); #endif /* Do window scaling on this connection? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; } tp->rcv_adv += min(tp->rcv_wnd, TCP_MAXWIN << tp->rcv_scale); /* * If not all the data that was sent in the TFO SYN * has been acked, resend the remainder right away. */ if (IS_FASTOPEN(tp->t_flags) && (tp->snd_una != tp->snd_max)) { tp->snd_nxt = th->th_ack; tfo_partial = 1; } /* * If there's data, delay ACK; if there's also a FIN ACKNOW * will be turned on later. */ if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) { rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tp->t_flags |= TF_DELACK; } else { rack->r_wanted_output = 1; tp->t_flags |= TF_ACKNOW; rack->rc_dack_toggle = 0; } if (((thflags & (TH_CWR | TH_ECE)) == TH_ECE) && (V_tcp_do_ecn == 1)) { tp->t_flags2 |= TF2_ECN_PERMIT; KMOD_TCPSTAT_INC(tcps_ecn_shs); } if (SEQ_GT(th->th_ack, tp->snd_una)) { /* * We advance snd_una for the * fast open case. If th_ack is * acknowledging data beyond * snd_una we can't just call * ack-processing since the * data stream in our send-map * will start at snd_una + 1 (one * beyond the SYN). If its just * equal we don't need to do that * and there is no send_map. */ tp->snd_una++; } /* * Received in SYN_SENT[*] state. Transitions: * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1 */ tp->t_starttime = ticks; if (tp->t_flags & TF_NEEDFIN) { tcp_state_change(tp, TCPS_FIN_WAIT_1); tp->t_flags &= ~TF_NEEDFIN; thflags &= ~TH_SYN; } else { tcp_state_change(tp, TCPS_ESTABLISHED); TCP_PROBE5(connect__established, NULL, tp, mtod(m, const char *), tp, th); rack_cc_conn_init(tp); } } else { /* * Received initial SYN in SYN-SENT[*] state => simultaneous * open. If segment contains CC option and there is a * cached CC, apply TAO test. If it succeeds, connection is * * half-synchronized. Otherwise, do 3-way handshake: * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If * there was no CC option, clear cached CC value. */ tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN); tcp_state_change(tp, TCPS_SYN_RECEIVED); } INP_WLOCK_ASSERT(tp->t_inpcb); /* * Advance th->th_seq to correspond to first data byte. If data, * trim to stay within window, dropping FIN if necessary. */ th->th_seq++; if (tlen > tp->rcv_wnd) { todrop = tlen - tp->rcv_wnd; m_adj(m, -todrop); tlen = tp->rcv_wnd; thflags &= ~TH_FIN; KMOD_TCPSTAT_INC(tcps_rcvpackafterwin); KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); } tp->snd_wl1 = th->th_seq - 1; tp->rcv_up = th->th_seq; /* * Client side of transaction: already sent SYN and data. If the * remote host used T/TCP to validate the SYN, our data will be * ACK'd; if so, enter normal data segment processing in the middle * of step 5, ack processing. Otherwise, goto step 6. */ if (thflags & TH_ACK) { /* For syn-sent we need to possibly update the rtt */ if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) { uint32_t t; t = tcp_ts_getticks() - to->to_tsecr; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; tcp_rack_xmit_timer(rack, t + 1, 1, (t * HPTS_USEC_IN_MSEC), 0, NULL, 2); tcp_rack_xmit_timer_commit(rack, tp); } if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) return (ret_val); /* We may have changed to FIN_WAIT_1 above */ if (tp->t_state == TCPS_FIN_WAIT_1) { /* * In FIN_WAIT_1 STATE in addition to the processing * for the ESTABLISHED state if our FIN is now * acknowledged then enter FIN_WAIT_2. */ if (ourfinisacked) { /* * If we can't receive any more data, then * closing user can proceed. Starting the * timer is contrary to the specification, * but if we don't get a FIN we'll hang * forever. * * XXXjl: we should release the tp also, and * use a compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { struct tcp_rack *rack; int32_t ret_val = 0; int32_t ourfinisacked = 0; ctf_calc_rwin(so, tp); if ((thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->snd_una) || SEQ_GT(th->th_ack, tp->snd_max))) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } rack = (struct tcp_rack *)tp->t_fb_ptr; if (IS_FASTOPEN(tp->t_flags)) { /* * When a TFO connection is in SYN_RECEIVED, the * only valid packets are the initial SYN, a * retransmit/copy of the initial SYN (possibly with * a subset of the original data), a valid ACK, a * FIN, or a RST. */ if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } else if (thflags & TH_SYN) { /* non-initial SYN is ignored */ if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) || (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) || (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) { ctf_do_drop(m, NULL); return (0); } } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) { ctf_do_drop(m, NULL); return (0); } } if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } /* * In the SYN-RECEIVED state, validate that the packet belongs to * this connection before trimming the data to fit the receive * window. Check the sequence number versus IRS since we know the * sequence numbers haven't wrapped. This is a partial fix for the * "LAND" DoS attack. */ if (SEQ_LT(th->th_seq, tp->irs)) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } tp->snd_wnd = tiwin; /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (IS_FASTOPEN(tp->t_flags)) { rack_cc_conn_init(tp); } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } KMOD_TCPSTAT_INC(tcps_connects); soisconnected(so); /* Do window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; } /* * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* -> * FIN-WAIT-1 */ tp->t_starttime = ticks; if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) { tcp_fastopen_decrement_counter(tp->t_tfo_pending); tp->t_tfo_pending = NULL; } if (tp->t_flags & TF_NEEDFIN) { tcp_state_change(tp, TCPS_FIN_WAIT_1); tp->t_flags &= ~TF_NEEDFIN; } else { tcp_state_change(tp, TCPS_ESTABLISHED); TCP_PROBE5(accept__established, NULL, tp, mtod(m, const char *), tp, th); /* * TFO connections call cc_conn_init() during SYN * processing. Calling it again here for such connections * is not harmless as it would undo the snd_cwnd reduction * that occurs when a TFO SYN|ACK is retransmitted. */ if (!IS_FASTOPEN(tp->t_flags)) rack_cc_conn_init(tp); } /* * Account for the ACK of our SYN prior to * regular ACK processing below, except for * simultaneous SYN, which is handled later. */ if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN)) tp->snd_una++; /* * If segment contains data or ACK, will call tcp_reass() later; if * not, do so now to pass queued data to user. */ if (tlen == 0 && (thflags & TH_FIN) == 0) (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0, (struct mbuf *)0); tp->snd_wl1 = th->th_seq - 1; /* For syn-recv we need to possibly update the rtt */ if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) { uint32_t t; t = tcp_ts_getticks() - to->to_tsecr; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; tcp_rack_xmit_timer(rack, t + 1, 1, (t * HPTS_USEC_IN_MSEC), 0, NULL, 2); tcp_rack_xmit_timer_commit(rack, tp); } if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (tp->t_state == TCPS_FIN_WAIT_1) { /* We could have went to FIN_WAIT_1 (or EST) above */ /* * In FIN_WAIT_1 STATE in addition to the processing for the * ESTABLISHED state if our FIN is now acknowledged then * enter FIN_WAIT_2. */ if (ourfinisacked) { /* * If we can't receive any more data, then closing * user can proceed. Starting the timer is contrary * to the specification, but if we don't get a FIN * we'll hang forever. * * XXXjl: we should release the tp also, and use a * compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; struct tcp_rack *rack; /* * Header prediction: check for the two common cases of a * uni-directional data xfer. If the packet has no control flags, * is in-sequence, the window didn't change and we're not * retransmitting, it's a candidate. If the length is zero and the * ack moved forward, we're the sender side of the xfer. Just free * the data acked & wake any higher level process that was blocked * waiting for space. If the length is non-zero and the ack didn't * move, we're the receiver side. If we're getting packets in-order * (the reassembly queue is empty), add the data toc The socket * buffer and note that we need a delayed ack. Make sure that the * hidden state-flags are also off. Since we check for * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN. */ rack = (struct tcp_rack *)tp->t_fb_ptr; if (__predict_true(((to->to_flags & TOF_SACK) == 0)) && __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) && __predict_true(SEGQ_EMPTY(tp)) && __predict_true(th->th_seq == tp->rcv_nxt)) { if (tlen == 0) { if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen, tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) { return (0); } } else { if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen, tiwin, nxt_pkt, iptos)) { return (0); } } } ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output= 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } /* State changes only happen in rack_process_data() */ return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } static int rack_check_data_after_close(struct mbuf *m, struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so) { struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack->rc_allow_data_af_clo == 0) { close_now: tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); /* tcp_close will kill the inp pre-log the Reset */ tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); tp = tcp_close(tp); KMOD_TCPSTAT_INC(tcps_rcvafterclose); ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen)); return (1); } if (sbavail(&so->so_snd) == 0) goto close_now; /* Ok we allow data that is ignored and a followup reset */ tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); tp->rcv_nxt = th->th_seq + *tlen; tp->t_flags2 |= TF2_DROP_AF_DATA; rack->r_wanted_output = 1; *tlen = 0; return (0); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; int32_t ourfinisacked = 0; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tlen) { if (rack_check_data_after_close(m, tp, &tlen, th, so)) return (1); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { /* * If we can't receive any more data, then closing user can * proceed. Starting the timer is contrary to the * specification, but if we don't get a FIN we'll hang * forever. * * XXXjl: we should release the tp also, and use a * compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; int32_t ourfinisacked = 0; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tlen) { if (rack_check_data_after_close(m, tp, &tlen, th, so)) return (1); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output= 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { tcp_twstart(tp); m_freem(m); return (1); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; int32_t ourfinisacked = 0; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tlen) { if (rack_check_data_after_close(m, tp, &tlen, th, so)) return (1); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * case TCPS_LAST_ACK: Ack processing. */ if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { tp = tcp_close(tp); ctf_do_drop(m, tp); return (1); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; int32_t ourfinisacked = 0; ctf_calc_rwin(so, tp); /* Reset receive buffer auto scaling when not in bulk receive mode. */ if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tlen) { if (rack_check_data_after_close(m, tp, &tlen, th, so)) return (1); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } static void inline rack_clear_rate_sample(struct tcp_rack *rack) { rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY; rack->r_ctl.rack_rs.rs_rtt_cnt = 0; rack->r_ctl.rack_rs.rs_rtt_tot = 0; } static void rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line) { uint64_t bw_est, rate_wanted; int chged = 0; uint32_t user_max; user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs; if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs) chged = 1; rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp); if (rack->use_fixed_rate || rack->rc_force_max_seg) { if (user_max != rack->r_ctl.rc_pace_max_segs) chged = 1; } if (rack->rc_force_max_seg) { rack->r_ctl.rc_pace_max_segs = user_max; } else if (rack->use_fixed_rate) { bw_est = rack_get_bw(rack); if ((rack->r_ctl.crte == NULL) || (bw_est != rack->r_ctl.crte->rate)) { rack->r_ctl.rc_pace_max_segs = user_max; } else { /* We are pacing right at the hardware rate */ uint32_t segsiz; segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size( tp, bw_est, segsiz, 0, rack->r_ctl.crte, NULL); } } else if (rack->rc_always_pace) { if (rack->r_ctl.gp_bw || #ifdef NETFLIX_PEAKRATE rack->rc_tp->t_maxpeakrate || #endif rack->r_ctl.init_rate) { /* We have a rate of some sort set */ uint32_t orig; bw_est = rack_get_bw(rack); orig = rack->r_ctl.rc_pace_max_segs; rate_wanted = rack_get_output_bw(rack, bw_est, NULL); if (rate_wanted) { /* We have something */ rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, rate_wanted, ctf_fixed_maxseg(rack->rc_tp)); } else rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs; if (orig != rack->r_ctl.rc_pace_max_segs) chged = 1; } else if ((rack->r_ctl.gp_bw == 0) && (rack->r_ctl.rc_pace_max_segs == 0)) { /* * If we have nothing limit us to bursting * out IW sized pieces. */ chged = 1; rack->r_ctl.rc_pace_max_segs = rc_init_window(rack); } } if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) { chged = 1; rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES; } if (chged) rack_log_type_hrdwtso(tp, rack, 0, rack->rc_inp->inp_socket->so_snd.sb_flags, line, 2); } static int rack_init(struct tcpcb *tp) { struct tcp_rack *rack = NULL; struct rack_sendmap *insret; uint32_t iwin, snt, us_cts; tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT); if (tp->t_fb_ptr == NULL) { /* * We need to allocate memory but cant. The INP and INP_INFO * locks and they are recusive (happens during setup. So a * scheme to drop the locks fails :( * */ return (ENOMEM); } memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack)); rack = (struct tcp_rack *)tp->t_fb_ptr; RB_INIT(&rack->r_ctl.rc_mtree); TAILQ_INIT(&rack->r_ctl.rc_free); TAILQ_INIT(&rack->r_ctl.rc_tmap); rack->rc_tp = tp; if (tp->t_inpcb) { rack->rc_inp = tp->t_inpcb; } /* Probably not needed but lets be sure */ rack_clear_rate_sample(rack); rack->r_ctl.rc_reorder_fade = rack_reorder_fade; rack->rc_allow_data_af_clo = rack_ignore_data_after_close; rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh; if (use_rack_rr) rack->use_rack_rr = 1; if (V_tcp_delack_enabled) tp->t_delayed_ack = 1; else tp->t_delayed_ack = 0; if (rack_enable_shared_cwnd) rack->rack_enable_scwnd = 1; rack->rc_user_set_max_segs = rack_hptsi_segments; rack->rc_force_max_seg = 0; if (rack_use_imac_dack) rack->rc_dack_mode = 1; rack->r_ctl.rc_reorder_shift = rack_reorder_thresh; rack->r_ctl.rc_pkt_delay = rack_pkt_delay; rack->r_ctl.rc_prop_reduce = rack_use_proportional_reduce; rack->r_ctl.rc_prop_rate = rack_proportional_rate; rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp; rack->r_ctl.rc_early_recovery = rack_early_recovery; rack->r_ctl.rc_lowest_us_rtt = 0xffffffff; rack->r_ctl.rc_highest_us_rtt = 0; if (rack_disable_prr) rack->rack_no_prr = 1; if (rack_gp_no_rec_chg) rack->rc_gp_no_rec_chg = 1; rack->rc_always_pace = rack_pace_every_seg; if (rack_enable_mqueue_for_nonpaced) rack->r_mbuf_queue = 1; else rack->r_mbuf_queue = 0; if (rack->r_mbuf_queue || rack->rc_always_pace) tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ; else tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; rack_set_pace_segments(tp, rack, __LINE__); if (rack_limits_scwnd) rack->r_limit_scw = 1; else rack->r_limit_scw = 0; rack->r_ctl.rc_high_rwnd = tp->snd_wnd; rack->r_ctl.cwnd_to_use = tp->snd_cwnd; rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method; rack->rack_tlp_threshold_use = rack_tlp_threshold_use; rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr; rack->r_ctl.rc_min_to = rack_min_to; microuptime(&rack->r_ctl.act_rcv_time); rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time; rack->r_running_late = 0; rack->r_running_early = 0; rack->rc_init_win = rack_default_init_window; rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss; if (rack_do_dyn_mul) { /* When dynamic adjustment is on CA needs to start at 100% */ rack->rc_gp_dyn_mul = 1; if (rack_do_dyn_mul >= 100) rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul; } else rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca; rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec; rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt; rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time); setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN, rack_probertt_filter_life); us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); rack->r_ctl.rc_lower_rtt_us_cts = us_cts; rack->r_ctl.rc_time_of_last_probertt = us_cts; rack->r_ctl.rc_time_probertt_starts = 0; /* Do we force on detection? */ #ifdef NETFLIX_EXP_DETECTION if (tcp_force_detection) rack->do_detection = 1; else #endif rack->do_detection = 0; if (rack_non_rxt_use_cr) rack->rack_rec_nonrxt_use_cr = 1; if (tp->snd_una != tp->snd_max) { /* Create a send map for the current outstanding data */ struct rack_sendmap *rsm; rsm = rack_alloc(rack); if (rsm == NULL) { uma_zfree(rack_pcb_zone, tp->t_fb_ptr); tp->t_fb_ptr = NULL; return (ENOMEM); } rsm->r_flags = RACK_OVERMAX; rsm->r_tim_lastsent[0] = rack->r_ctl.rc_tlp_rxt_last_time; rsm->r_rtr_cnt = 1; rsm->r_rtr_bytes = 0; rsm->r_start = tp->snd_una; if (tp->t_flags & TF_SENTFIN) { rsm->r_end = tp->snd_max - 1; rsm->r_flags |= RACK_HAS_FIN; } else { rsm->r_end = tp->snd_max; } rsm->usec_orig_send = us_cts; rsm->r_dupack = 0; insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #ifdef INVARIANTS if (insret != NULL) { panic("Insert in rb tree fails ret:%p rack:%p rsm:%p", insret, rack, rsm); } #endif TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; } /* Cancel the GP measurement in progress */ tp->t_flags &= ~TF_GPUTINPROG; if (SEQ_GT(tp->snd_max, tp->iss)) snt = tp->snd_max - tp->iss; else snt = 0; iwin = rc_init_window(rack); if (snt < iwin) { /* We are not past the initial window * so we need to make sure cwnd is * correct. */ if (tp->snd_cwnd < iwin) tp->snd_cwnd = iwin; /* * If we are within the initial window * we want ssthresh to be unlimited. Setting * it to the rwnd (which the default stack does * and older racks) is not really a good idea * since we want to be in SS and grow both the * cwnd and the rwnd (via dynamic rwnd growth). If * we set it to the rwnd then as the peer grows its * rwnd we will be stuck in CA and never hit SS. * * Its far better to raise it up high (this takes the * risk that there as been a loss already, probably * we should have an indicator in all stacks of loss * but we don't), but considering the normal use this * is a risk worth taking. The consequences of not * hitting SS are far worse than going one more time * into it early on (before we have sent even a IW). * It is highly unlikely that we will have had a loss * before getting the IW out. */ tp->snd_ssthresh = 0xffffffff; } rack_stop_all_timers(tp); rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), 0, 0, 0); rack_log_rtt_shrinks(rack, us_cts, 0, __LINE__, RACK_RTTS_INIT); return (0); } static int rack_handoff_ok(struct tcpcb *tp) { if ((tp->t_state == TCPS_CLOSED) || (tp->t_state == TCPS_LISTEN)) { /* Sure no problem though it may not stick */ return (0); } if ((tp->t_state == TCPS_SYN_SENT) || (tp->t_state == TCPS_SYN_RECEIVED)) { /* * We really don't know if you support sack, * you have to get to ESTAB or beyond to tell. */ return (EAGAIN); } if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) { /* * Rack will only send a FIN after all data is acknowledged. * So in this case we have more data outstanding. We can't * switch stacks until either all data and only the FIN * is left (in which case rack_init() now knows how * to deal with that) all is acknowledged and we * are only left with incoming data, though why you * would want to switch to rack after all data is acknowledged * I have no idea (rrs)! */ return (EAGAIN); } if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){ return (0); } /* * If we reach here we don't do SACK on this connection so we can * never do rack. */ return (EINVAL); } static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged) { if (tp->t_fb_ptr) { struct tcp_rack *rack; struct rack_sendmap *rsm, *nrsm, *rm; rack = (struct tcp_rack *)tp->t_fb_ptr; #ifdef NETFLIX_SHARED_CWND if (rack->r_ctl.rc_scw) { uint32_t limit; if (rack->r_limit_scw) limit = max(1, rack->r_ctl.rc_lowest_us_rtt); else limit = 0; tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index, limit); rack->r_ctl.rc_scw = NULL; } #endif /* rack does not use force data but other stacks may clear it */ tp->t_flags &= ~TF_FORCEDATA; if (tp->t_inpcb) { tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY; tp->t_inpcb->inp_flags2 &= ~INP_DONT_SACK_QUEUE; } #ifdef TCP_BLACKBOX tcp_log_flowend(tp); #endif RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) { rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #ifdef INVARIANTS if (rm != rsm) { panic("At fini, rack:%p rsm:%p rm:%p", rack, rsm, rm); } #endif uma_zfree(rack_zone, rsm); } rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); while (rsm) { TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext); uma_zfree(rack_zone, rsm); rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); } rack->rc_free_cnt = 0; uma_zfree(rack_pcb_zone, tp->t_fb_ptr); tp->t_fb_ptr = NULL; } /* Cancel the GP measurement in progress */ tp->t_flags &= ~TF_GPUTINPROG; /* Make sure snd_nxt is correctly set */ tp->snd_nxt = tp->snd_max; } static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack) { switch (tp->t_state) { case TCPS_SYN_SENT: rack->r_state = TCPS_SYN_SENT; rack->r_substate = rack_do_syn_sent; break; case TCPS_SYN_RECEIVED: rack->r_state = TCPS_SYN_RECEIVED; rack->r_substate = rack_do_syn_recv; break; case TCPS_ESTABLISHED: rack_set_pace_segments(tp, rack, __LINE__); rack->r_state = TCPS_ESTABLISHED; rack->r_substate = rack_do_established; break; case TCPS_CLOSE_WAIT: rack->r_state = TCPS_CLOSE_WAIT; rack->r_substate = rack_do_close_wait; break; case TCPS_FIN_WAIT_1: rack->r_state = TCPS_FIN_WAIT_1; rack->r_substate = rack_do_fin_wait_1; break; case TCPS_CLOSING: rack->r_state = TCPS_CLOSING; rack->r_substate = rack_do_closing; break; case TCPS_LAST_ACK: rack->r_state = TCPS_LAST_ACK; rack->r_substate = rack_do_lastack; break; case TCPS_FIN_WAIT_2: rack->r_state = TCPS_FIN_WAIT_2; rack->r_substate = rack_do_fin_wait_2; break; case TCPS_LISTEN: case TCPS_CLOSED: case TCPS_TIME_WAIT: default: break; }; } static void rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb) { /* * We received an ack, and then did not * call send or were bounced out due to the * hpts was running. Now a timer is up as well, is * it the right timer? */ struct rack_sendmap *rsm; int tmr_up; tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK; if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT)) return; rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) && (tmr_up == PACE_TMR_RXT)) { /* Should be an RXT */ return; } if (rsm == NULL) { /* Nothing outstanding? */ if (tp->t_flags & TF_DELACK) { if (tmr_up == PACE_TMR_DELACK) /* We are supposed to have delayed ack up and we do */ return; } else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) { /* * if we hit enobufs then we would expect the possiblity * of nothing outstanding and the RXT up (and the hptsi timer). */ return; } else if (((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)) && (tmr_up == PACE_TMR_KEEP) && (tp->snd_max == tp->snd_una)) { /* We should have keep alive up and we do */ return; } } if (SEQ_GT(tp->snd_max, tp->snd_una) && ((tmr_up == PACE_TMR_TLP) || (tmr_up == PACE_TMR_RACK) || (tmr_up == PACE_TMR_RXT))) { /* * Either a Rack, TLP or RXT is fine if we * have outstanding data. */ return; } else if (tmr_up == PACE_TMR_DELACK) { /* * If the delayed ack was going to go off * before the rtx/tlp/rack timer were going to * expire, then that would be the timer in control. * Note we don't check the time here trusting the * code is correct. */ return; } /* * Ok the timer originally started is not what we want now. * We will force the hpts to be stopped if any, and restart * with the slot set to what was in the saved slot. */ if (rack->rc_inp->inp_in_hpts) { if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { uint32_t us_cts; us_cts = tcp_get_usecs(NULL); if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) { rack->r_early = 1; rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts); } rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT; } tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT); } rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), 0, 0, 0); } static int rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt, struct timeval *tv) { int32_t thflags, retval, did_out = 0; int32_t way_out = 0; uint32_t cts; uint32_t tiwin; struct timespec ts; struct tcpopt to; struct tcp_rack *rack; struct rack_sendmap *rsm; int32_t prev_state = 0; uint32_t us_cts; /* * tv passed from common code is from either M_TSTMP_LRO or * tcp_get_usecs() if no LRO m_pkthdr timestamp is present. The * rack_pacing stack assumes tv always refers to 'now', so we overwrite * tv here to guarantee that. */ if (m->m_flags & M_TSTMP_LRO) tcp_get_usecs(tv); cts = tcp_tv_to_mssectick(tv); rack = (struct tcp_rack *)tp->t_fb_ptr; if ((m->m_flags & M_TSTMP) || (m->m_flags & M_TSTMP_LRO)) { mbuf_tstmp2timespec(m, &ts); rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec; rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000; } else rack->r_ctl.act_rcv_time = *tv; kern_prefetch(rack, &prev_state); prev_state = 0; thflags = th->th_flags; NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(tp->t_inpcb); KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", __func__)); KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", __func__)); if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval ltv; #ifdef NETFLIX_HTTP_LOGGING struct http_sendfile_track *http_req; if (SEQ_GT(th->th_ack, tp->snd_una)) { http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1)); } else { http_req = tcp_http_find_req_for_seq(tp, th->th_ack); } #endif memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; if (rack->rack_no_prr == 0) log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt; else log.u_bbr.flex1 = 0; log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->rc_tp->t_maxseg; log.u_bbr.flex3 = m->m_flags; log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; if (m->m_flags & M_TSTMP) { /* Record the hardware timestamp if present */ mbuf_tstmp2timespec(m, &ts); ltv.tv_sec = ts.tv_sec; ltv.tv_usec = ts.tv_nsec / 1000; log.u_bbr.lt_epoch = tcp_tv_to_usectick(<v); } else if (m->m_flags & M_TSTMP_LRO) { /* Record the LRO the arrival timestamp */ mbuf_tstmp2timespec(m, &ts); ltv.tv_sec = ts.tv_sec; ltv.tv_usec = ts.tv_nsec / 1000; log.u_bbr.flex5 = tcp_tv_to_usectick(<v); } log.u_bbr.timeStamp = tcp_get_usecs(<v); /* Log the rcv time */ log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp; #ifdef NETFLIX_HTTP_LOGGING log.u_bbr.applimited = tp->t_http_closed; log.u_bbr.applimited <<= 8; log.u_bbr.applimited |= tp->t_http_open; log.u_bbr.applimited <<= 8; log.u_bbr.applimited |= tp->t_http_req; if (http_req) { /* Copy out any client req info */ /* seconds */ log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC); /* useconds */ log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC); log.u_bbr.rttProp = http_req->timestamp; log.u_bbr.cur_del_rate = http_req->start; if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) { log.u_bbr.flex8 |= 1; } else { log.u_bbr.flex8 |= 2; log.u_bbr.bw_inuse = http_req->end; } log.u_bbr.flex6 = http_req->start_seq; if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) { log.u_bbr.flex8 |= 4; log.u_bbr.epoch = http_req->end_seq; } } #endif TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0, tlen, &log, true, <v); } if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) { way_out = 4; retval = 0; goto done_with_input; } /* * If a segment with the ACK-bit set arrives in the SYN-SENT state * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9. */ if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return(1); } /* * Parse options on any incoming segment. */ tcp_dooptions(&to, (u_char *)(th + 1), (th->th_off << 2) - sizeof(struct tcphdr), (thflags & TH_SYN) ? TO_SYN : 0); /* * If timestamps were negotiated during SYN/ACK and a * segment without a timestamp is received, silently drop * the segment, unless it is a RST segment or missing timestamps are * tolerated. * See section 3.2 of RFC 7323. */ if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) && ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) { way_out = 5; retval = 0; goto done_with_input; } /* * Segment received on connection. Reset idle time and keep-alive * timer. XXX: This should be done after segment validation to * ignore broken/spoofed segs. */ if (tp->t_idle_reduce && (tp->snd_max == tp->snd_una) && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) { counter_u64_add(rack_input_idle_reduces, 1); rack_cc_after_idle(rack, tp); } tp->t_rcvtime = ticks; /* * Unscale the window into a 32-bit value. For the SYN_SENT state * the scale is zero. */ tiwin = th->th_win << tp->snd_scale; #ifdef STATS stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin); #endif if (tiwin > rack->r_ctl.rc_high_rwnd) rack->r_ctl.rc_high_rwnd = tiwin; /* * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move * this to occur after we've validated the segment. */ if (tp->t_flags2 & TF2_ECN_PERMIT) { if (thflags & TH_CWR) { tp->t_flags2 &= ~TF2_ECN_SND_ECE; tp->t_flags |= TF_ACKNOW; } switch (iptos & IPTOS_ECN_MASK) { case IPTOS_ECN_CE: tp->t_flags2 |= TF2_ECN_SND_ECE; KMOD_TCPSTAT_INC(tcps_ecn_ce); break; case IPTOS_ECN_ECT0: KMOD_TCPSTAT_INC(tcps_ecn_ect0); break; case IPTOS_ECN_ECT1: KMOD_TCPSTAT_INC(tcps_ecn_ect1); break; } /* Process a packet differently from RFC3168. */ cc_ecnpkt_handler(tp, th, iptos); /* Congestion experienced. */ if (thflags & TH_ECE) { rack_cong_signal(tp, th, CC_ECN); } } /* * If echoed timestamp is later than the current time, fall back to * non RFC1323 RTT calculation. Normalize timestamp if syncookies * were used when this connection was established. */ if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { to.to_tsecr -= tp->ts_offset; if (TSTMP_GT(to.to_tsecr, cts)) to.to_tsecr = 0; } /* * If its the first time in we need to take care of options and * verify we can do SACK for rack! */ if (rack->r_state == 0) { /* Should be init'd by rack_init() */ KASSERT(rack->rc_inp != NULL, ("%s: rack->rc_inp unexpectedly NULL", __func__)); if (rack->rc_inp == NULL) { rack->rc_inp = tp->t_inpcb; } /* * Process options only when we get SYN/ACK back. The SYN * case for incoming connections is handled in tcp_syncache. * According to RFC1323 the window field in a SYN (i.e., a * or ) segment itself is never scaled. XXX * this is traditional behavior, may need to be cleaned up. */ if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { /* Handle parallel SYN for ECN */ if (!(thflags & TH_ACK) && ((thflags & (TH_CWR | TH_ECE)) == (TH_CWR | TH_ECE)) && ((V_tcp_do_ecn == 1) || (V_tcp_do_ecn == 2))) { tp->t_flags2 |= TF2_ECN_PERMIT; tp->t_flags2 |= TF2_ECN_SND_ECE; TCPSTAT_INC(tcps_ecn_shs); } if ((to.to_flags & TOF_SCALE) && (tp->t_flags & TF_REQ_SCALE)) { tp->t_flags |= TF_RCVD_SCALE; tp->snd_scale = to.to_wscale; } else tp->t_flags &= ~TF_REQ_SCALE; /* * Initial send window. It will be updated with the * next incoming segment to the scaled value. */ tp->snd_wnd = th->th_win; if ((to.to_flags & TOF_TS) && (tp->t_flags & TF_REQ_TSTMP)) { tp->t_flags |= TF_RCVD_TSTMP; tp->ts_recent = to.to_tsval; tp->ts_recent_age = cts; } else tp->t_flags &= ~TF_REQ_TSTMP; if (to.to_flags & TOF_MSS) tcp_mss(tp, to.to_mss); if ((tp->t_flags & TF_SACK_PERMIT) && (to.to_flags & TOF_SACKPERM) == 0) tp->t_flags &= ~TF_SACK_PERMIT; if (IS_FASTOPEN(tp->t_flags)) { if (to.to_flags & TOF_FASTOPEN) { uint16_t mss; if (to.to_flags & TOF_MSS) mss = to.to_mss; else if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) mss = TCP6_MSS; else mss = TCP_MSS; tcp_fastopen_update_cache(tp, mss, to.to_tfo_len, to.to_tfo_cookie); } else tcp_fastopen_disable_path(tp); } } /* * At this point we are at the initial call. Here we decide * if we are doing RACK or not. We do this by seeing if * TF_SACK_PERMIT is set and the sack-not-required is clear. * The code now does do dup-ack counting so if you don't * switch back you won't get rack & TLP, but you will still * get this stack. */ if ((rack_sack_not_required == 0) && ((tp->t_flags & TF_SACK_PERMIT) == 0)) { tcp_switch_back_to_default(tp); (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen, tlen, iptos); return (1); } /* Set the flag */ rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; tcp_set_hpts(tp->t_inpcb); sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack); } if (thflags & TH_FIN) tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN); us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); if ((rack->rc_gp_dyn_mul) && (rack->use_fixed_rate == 0) && (rack->rc_always_pace)) { /* Check in on probertt */ rack_check_probe_rtt(rack, us_cts); } if (rack->forced_ack) { uint32_t us_rtt; /* * A persist or keep-alive was forced out, update our * min rtt time. Note we do not worry about lost * retransmissions since KEEP-ALIVES and persists * are usually way long on times of sending (though * if we were really paranoid or worried we could * at least use timestamps if available to validate). */ rack->forced_ack = 0; us_rtt = us_cts - rack->r_ctl.forced_ack_ts; if (us_rtt == 0) us_rtt = 1; rack_log_rtt_upd(tp, rack, us_rtt, 0, NULL, 3); rack_apply_updated_usrtt(rack, us_rtt, us_cts); } /* * This is the one exception case where we set the rack state * always. All other times (timers etc) we must have a rack-state * set (so we assure we have done the checks above for SACK). */ rack->r_ctl.rc_rcvtime = cts; if (rack->r_state != tp->t_state) rack_set_state(tp, rack); if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL) kern_prefetch(rsm, &prev_state); prev_state = rack->r_state; rack_clear_rate_sample(rack); retval = (*rack->r_substate) (m, th, so, tp, &to, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt, iptos); #ifdef INVARIANTS if ((retval == 0) && (tp->t_inpcb == NULL)) { panic("retval:%d tp:%p t_inpcb:NULL state:%d", retval, tp, prev_state); } #endif if (retval == 0) { /* * If retval is 1 the tcb is unlocked and most likely the tp * is gone. */ INP_WLOCK_ASSERT(tp->t_inpcb); if ((rack->rc_gp_dyn_mul) && (rack->rc_always_pace) && (rack->use_fixed_rate == 0) && rack->in_probe_rtt && (rack->r_ctl.rc_time_probertt_starts == 0)) { /* * If we are going for target, lets recheck before * we output. */ rack_check_probe_rtt(rack, us_cts); } if (rack->set_pacing_done_a_iw == 0) { /* How much has been acked? */ if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) { /* We have enough to set in the pacing segment size */ rack->set_pacing_done_a_iw = 1; rack_set_pace_segments(tp, rack, __LINE__); } } tcp_rack_xmit_timer_commit(rack, tp); if (nxt_pkt == 0) { if (rack->r_wanted_output != 0) { do_output_now: did_out = 1; (void)tp->t_fb->tfb_tcp_output(tp); } rack_start_hpts_timer(rack, tp, cts, 0, 0, 0); } if ((nxt_pkt == 0) && ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) && (SEQ_GT(tp->snd_max, tp->snd_una) || (tp->t_flags & TF_DELACK) || ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)))) { /* We could not send (probably in the hpts but stopped the timer earlier)? */ if ((tp->snd_max == tp->snd_una) && ((tp->t_flags & TF_DELACK) == 0) && (rack->rc_inp->inp_in_hpts) && (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { /* keep alive not needed if we are hptsi output yet */ ; } else { int late = 0; if (rack->rc_inp->inp_in_hpts) { if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { us_cts = tcp_get_usecs(NULL); if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) { rack->r_early = 1; rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts); } else late = 1; rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT; } tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT); } if (late && (did_out == 0)) { /* * We are late in the sending * and we did not call the output * (this probably should not happen). */ goto do_output_now; } rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), 0, 0, 0); } way_out = 1; } else if (nxt_pkt == 0) { /* Do we have the correct timer running? */ rack_timer_audit(tp, rack, &so->so_snd); way_out = 2; } done_with_input: rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out); if (did_out) rack->r_wanted_output = 0; #ifdef INVARIANTS if (tp->t_inpcb == NULL) { panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d", did_out, retval, tp, prev_state); } #endif } return (retval); } void rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos) { struct timeval tv; /* First lets see if we have old packets */ if (tp->t_in_pkt) { if (ctf_do_queued_segments(so, tp, 1)) { m_freem(m); return; } } if (m->m_flags & M_TSTMP_LRO) { tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000; tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000; } else { /* Should not be should we kassert instead? */ tcp_get_usecs(&tv); } if(rack_do_segment_nounlock(m, th, so, tp, drop_hdrlen, tlen, iptos, 0, &tv) == 0) { tcp_handle_wakeup(tp, so); INP_WUNLOCK(tp->t_inpcb); } } struct rack_sendmap * tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused) { struct rack_sendmap *rsm = NULL; int32_t idx; uint32_t srtt = 0, thresh = 0, ts_low = 0; /* Return the next guy to be re-transmitted */ if (RB_EMPTY(&rack->r_ctl.rc_mtree)) { return (NULL); } if (tp->t_flags & TF_SENTFIN) { /* retran the end FIN? */ return (NULL); } /* ok lets look at this one */ rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) { goto check_it; } rsm = rack_find_lowest_rsm(rack); if (rsm == NULL) { return (NULL); } check_it: if (rsm->r_flags & RACK_ACKED) { return (NULL); } if (((rsm->r_flags & RACK_SACK_PASSED) == 0) && (rsm->r_dupack < DUP_ACK_THRESHOLD)) { /* Its not yet ready */ return (NULL); } srtt = rack_grab_rtt(tp, rack); idx = rsm->r_rtr_cnt - 1; ts_low = rsm->r_tim_lastsent[idx]; thresh = rack_calc_thresh_rack(rack, srtt, tsused); if ((tsused == ts_low) || (TSTMP_LT(tsused, ts_low))) { /* No time since sending */ return (NULL); } if ((tsused - ts_low) < thresh) { /* It has not been long enough yet */ return (NULL); } if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) || ((rsm->r_flags & RACK_SACK_PASSED) && (rack->sack_attack_disable == 0))) { /* * We have passed the dup-ack threshold * a SACK has indicated this is missing. * Note that if you are a declared attacker * it is only the dup-ack threshold that * will cause retransmits. */ /* log retransmit reason */ rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1); return (rsm); } return (NULL); } static void rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line, struct rack_sendmap *rsm) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = slot; log.u_bbr.flex2 = len; log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs; log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs; log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss; log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca; log.u_bbr.use_lt_bw = rack->app_limited_needs_set; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw = rack->rc_gp_filled; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->in_probe_rtt; log.u_bbr.pkt_epoch = line; log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec; log.u_bbr.bw_inuse = bw_est; log.u_bbr.delRate = bw; if (rack->r_ctl.gp_bw == 0) log.u_bbr.cur_del_rate = 0; else log.u_bbr.cur_del_rate = rack_get_bw(rack); log.u_bbr.rttProp = len_time; log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt; log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit; log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm); if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) { /* We are in slow start */ log.u_bbr.flex7 = 1; } else { /* we are on congestion avoidance */ log.u_bbr.flex7 = 0; } log.u_bbr.flex8 = method; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec; log.u_bbr.cwnd_gain <<= 1; log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss; log.u_bbr.cwnd_gain <<= 1; log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_HPTSI_CALC, 0, 0, &log, false, &tv); } } static uint32_t rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss) { uint32_t new_tso, user_max; user_max = rack->rc_user_set_max_segs * mss; if (rack->rc_force_max_seg) { return (user_max); } if (rack->use_fixed_rate && ((rack->r_ctl.crte == NULL) || (bw != rack->r_ctl.crte->rate))) { /* Use the user mss since we are not exactly matched */ return (user_max); } new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL); if (new_tso > user_max) new_tso = user_max; return(new_tso); } static void rack_log_hdwr_pacing(struct tcp_rack *rack, const struct ifnet *ifp, uint64_t rate, uint64_t hw_rate, int line, int error) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff); log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff); log.u_bbr.flex3 = (((uint64_t)ifp >> 32) & 0x00000000ffffffff); log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff); log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.bw_inuse = rate; log.u_bbr.flex5 = line; log.u_bbr.flex6 = error; log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs; log.u_bbr.flex8 = rack->use_fixed_rate; log.u_bbr.flex8 <<= 1; log.u_bbr.flex8 |= rack->rack_hdrw_pacing; log.u_bbr.pkts_out = rack->rc_tp->t_maxseg; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_HDWR_PACE, 0, 0, &log, false, &tv); } } static int32_t pace_to_fill_cwnd(struct tcp_rack *rack, int32_t slot, uint32_t len, uint32_t segsiz) { uint64_t lentim, fill_bw; /* Lets first see if we are full, if so continue with normal rate */ if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use) return (slot); if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd) return (slot); if (rack->r_ctl.rc_last_us_rtt == 0) return (slot); if (rack->rc_pace_fill_if_rttin_range && (rack->r_ctl.rc_last_us_rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) { /* The rtt is huge, N * smallest, lets not fill */ return (slot); } /* * first lets calculate the b/w based on the last us-rtt * and the sndwnd. */ fill_bw = rack->r_ctl.cwnd_to_use; /* Take the rwnd if its smaller */ if (fill_bw > rack->rc_tp->snd_wnd) fill_bw = rack->rc_tp->snd_wnd; fill_bw *= (uint64_t)HPTS_USEC_IN_SEC; fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt; /* We are below the min b/w */ if (fill_bw < RACK_MIN_BW) return (slot); /* * Ok fill_bw holds our mythical b/w to fill the cwnd * in a rtt, what does that time wise equate too? */ lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC; lentim /= fill_bw; if (lentim < slot) { rack_log_pacing_delay_calc(rack, len, slot, fill_bw, 0, lentim, 12, __LINE__, NULL); return ((int32_t)lentim); } else return (slot); } static int32_t rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz) { struct rack_sendmap *lrsm; int32_t slot = 0; int err; if (rack->rc_always_pace == 0) { /* * We use the most optimistic possible cwnd/srtt for * sending calculations. This will make our * calculation anticipate getting more through * quicker then possible. But thats ok we don't want * the peer to have a gap in data sending. */ uint32_t srtt, cwnd, tr_perms = 0; int32_t reduce = 0; old_method: /* * We keep no precise pacing with the old method * instead we use the pacer to mitigate bursts. */ rack->r_ctl.rc_agg_delayed = 0; rack->r_early = 0; rack->r_late = 0; rack->r_ctl.rc_agg_early = 0; if (rack->r_ctl.rc_rack_min_rtt) srtt = rack->r_ctl.rc_rack_min_rtt; else srtt = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT)); if (rack->r_ctl.rc_rack_largest_cwnd) cwnd = rack->r_ctl.rc_rack_largest_cwnd; else cwnd = rack->r_ctl.cwnd_to_use; tr_perms = cwnd / srtt; if (tr_perms == 0) { tr_perms = ctf_fixed_maxseg(tp); } /* * Calculate how long this will take to drain, if * the calculation comes out to zero, thats ok we * will use send_a_lot to possibly spin around for * more increasing tot_len_this_send to the point * that its going to require a pace, or we hit the * cwnd. Which in that case we are just waiting for * a ACK. */ slot = len / tr_perms; /* Now do we reduce the time so we don't run dry? */ if (slot && rack_slot_reduction) { reduce = (slot / rack_slot_reduction); if (reduce < slot) { slot -= reduce; } else slot = 0; } slot *= HPTS_USEC_IN_MSEC; if (rsm == NULL) { /* * We always consider ourselves app limited with old style * that are not retransmits. This could be the initial * measurement, but thats ok its all setup and specially * handled. If another send leaks out, then that too will * be mark app-limited. */ lrsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (lrsm && ((lrsm->r_flags & RACK_APP_LIMITED) == 0)) { rack->r_ctl.rc_first_appl = lrsm; lrsm->r_flags |= RACK_APP_LIMITED; rack->r_ctl.rc_app_limited_cnt++; } } rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL); } else { uint64_t bw_est, res, lentim, rate_wanted; uint32_t orig_val, srtt, segs, oh; if ((rack->r_rr_config == 1) && rsm) { return (rack->r_ctl.rc_min_to * HPTS_USEC_IN_MSEC); } if (rack->use_fixed_rate) { rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack); } else if ((rack->r_ctl.init_rate == 0) && #ifdef NETFLIX_PEAKRATE (rack->rc_tp->t_maxpeakrate == 0) && #endif (rack->r_ctl.gp_bw == 0)) { /* no way to yet do an estimate */ bw_est = rate_wanted = 0; } else { bw_est = rack_get_bw(rack); rate_wanted = rack_get_output_bw(rack, bw_est, rsm); } if ((bw_est == 0) || (rate_wanted == 0)) { /* * No way yet to make a b/w estimate or * our raise is set incorrectly. */ goto old_method; } /* We need to account for all the overheads */ segs = (len + segsiz - 1) / segsiz; /* * We need the diff between 1514 bytes (e-mtu with e-hdr) * and how much data we put in each packet. Yes this * means we may be off if we are larger than 1500 bytes * or smaller. But this just makes us more conservative. */ if (ETHERNET_SEGMENT_SIZE > segsiz) oh = ETHERNET_SEGMENT_SIZE - segsiz; else oh = 0; segs *= oh; lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC; res = lentim / rate_wanted; slot = (uint32_t)res; orig_val = rack->r_ctl.rc_pace_max_segs; rack_set_pace_segments(rack->rc_tp, rack, __LINE__); /* Did we change the TSO size, if so log it */ if (rack->r_ctl.rc_pace_max_segs != orig_val) rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL); if ((rack->rc_pace_to_cwnd) && (rack->in_probe_rtt == 0) && (IN_RECOVERY(rack->rc_tp->t_flags) == 0)) { /* * We want to pace at our rate *or* faster to * fill the cwnd to the max if its not full. */ slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz); } if ((rack->rc_inp->inp_route.ro_nh != NULL) && (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) { if ((rack->rack_hdw_pace_ena) && (rack->rack_hdrw_pacing == 0) && (rack->rack_attempt_hdwr_pace == 0)) { /* * Lets attempt to turn on hardware pacing * if we can. */ rack->rack_attempt_hdwr_pace = 1; rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp, rack->rc_inp->inp_route.ro_nh->nh_ifp, rate_wanted, RS_PACING_GEQ, &err, NULL); if (rack->r_ctl.crte) { rack->rack_hdrw_pacing = 1; rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(rack->rc_tp, rate_wanted, segsiz, 0, rack->r_ctl.crte, NULL); rack_log_hdwr_pacing(rack, rack->rc_inp->inp_route.ro_nh->nh_ifp, rate_wanted, rack->r_ctl.crte->rate, __LINE__, err); } } else if (rack->rack_hdrw_pacing && (rack->r_ctl.crte->rate != rate_wanted)) { /* Do we need to adjust our rate? */ const struct tcp_hwrate_limit_table *nrte; nrte = tcp_chg_pacing_rate(rack->r_ctl.crte, rack->rc_tp, rack->rc_inp->inp_route.ro_nh->nh_ifp, rate_wanted, RS_PACING_GEQ, &err, NULL); if (nrte == NULL) { /* Lost the rate */ rack->rack_hdrw_pacing = 0; rack_set_pace_segments(rack->rc_tp, rack, __LINE__); } else if (nrte != rack->r_ctl.crte) { rack->r_ctl.crte = nrte; rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(rack->rc_tp, rate_wanted, segsiz, 0, rack->r_ctl.crte, NULL); rack_log_hdwr_pacing(rack, rack->rc_inp->inp_route.ro_nh->nh_ifp, rate_wanted, rack->r_ctl.crte->rate, __LINE__, err); } } } if (rack_limit_time_with_srtt && (rack->use_fixed_rate == 0) && #ifdef NETFLIX_PEAKRATE (rack->rc_tp->t_maxpeakrate == 0) && #endif (rack->rack_hdrw_pacing == 0)) { /* * Sanity check, we do not allow the pacing delay * to be longer than the SRTT of the path. If it is * a slow path, then adding a packet should increase * the RTT and compensate for this i.e. the srtt will * be greater so the allowed pacing time will be greater. * * Note this restriction is not for where a peak rate * is set, we are doing fixed pacing or hardware pacing. */ if (rack->rc_tp->t_srtt) srtt = (TICKS_2_USEC(rack->rc_tp->t_srtt) >> TCP_RTT_SHIFT); else srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC; /* its in ms convert */ if (srtt < slot) { rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL); slot = srtt; } } rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm); } if (slot) counter_u64_add(rack_calc_nonzero, 1); else counter_u64_add(rack_calc_zero, 1); return (slot); } static void rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq startseq, uint32_t sb_offset) { struct rack_sendmap *my_rsm = NULL; struct rack_sendmap fe; if (tp->t_state < TCPS_ESTABLISHED) { /* * We don't start any measurements if we are * not at least established. */ return; } tp->t_flags |= TF_GPUTINPROG; rack->r_ctl.rc_gp_lowrtt = 0xffffffff; rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd; tp->gput_seq = startseq; rack->app_limited_needs_set = 0; if (rack->in_probe_rtt) rack->measure_saw_probe_rtt = 1; else if ((rack->measure_saw_probe_rtt) && (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit))) rack->measure_saw_probe_rtt = 0; if (rack->rc_gp_filled) tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); else { /* Special case initial measurement */ rack->r_ctl.rc_gp_output_ts = tp->gput_ts = tcp_get_usecs(NULL); } /* * We take a guess out into the future, * if we have no measurement and no * initial rate, we measure the first * initial-windows worth of data to * speed up getting some GP measurement and * thus start pacing. */ if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) { rack->app_limited_needs_set = 1; tp->gput_ack = startseq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp))); rack_log_pacing_delay_calc(rack, tp->gput_seq, tp->gput_ack, 0, tp->gput_ts, rack->r_ctl.rc_app_limited_cnt, 9, __LINE__, NULL); return; } if (sb_offset) { /* * We are out somewhere in the sb * can we use the already outstanding data? */ if (rack->r_ctl.rc_app_limited_cnt == 0) { /* * Yes first one is good and in this case * the tp->gput_ts is correctly set based on * the last ack that arrived (no need to * set things up when an ack comes in). */ my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if ((my_rsm == NULL) || (my_rsm->r_rtr_cnt != 1)) { /* retransmission? */ goto use_latest; } } else { if (rack->r_ctl.rc_first_appl == NULL) { /* * If rc_first_appl is NULL * then the cnt should be 0. * This is probably an error, maybe * a KASSERT would be approprate. */ goto use_latest; } /* * If we have a marker pointer to the last one that is * app limited we can use that, but we need to set * things up so that when it gets ack'ed we record * the ack time (if its not already acked). */ rack->app_limited_needs_set = 1; /* * We want to get to the rsm that is either * next with space i.e. over 1 MSS or the one * after that (after the app-limited). */ my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rack->r_ctl.rc_first_appl); if (my_rsm) { if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp)) /* Have to use the next one */ my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, my_rsm); else { /* Use after the first MSS of it is acked */ tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp); goto start_set; } } if ((my_rsm == NULL) || (my_rsm->r_rtr_cnt != 1)) { /* * Either its a retransmit or * the last is the app-limited one. */ goto use_latest; } } tp->gput_seq = my_rsm->r_start; start_set: if (my_rsm->r_flags & RACK_ACKED) { /* * This one has been acked use the arrival ack time */ tp->gput_ts = my_rsm->r_ack_arrival; rack->app_limited_needs_set = 0; } rack->r_ctl.rc_gp_output_ts = my_rsm->usec_orig_send; tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack); rack_log_pacing_delay_calc(rack, tp->gput_seq, tp->gput_ack, (uint64_t)my_rsm, tp->gput_ts, rack->r_ctl.rc_app_limited_cnt, 9, __LINE__, NULL); return; } use_latest: /* * We don't know how long we may have been * idle or if this is the first-send. Lets * setup the flag so we will trim off * the first ack'd data so we get a true * measurement. */ rack->app_limited_needs_set = 1; tp->gput_ack = startseq + rack_get_measure_window(tp, rack); /* Find this guy so we can pull the send time */ fe.r_start = startseq; my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); if (my_rsm) { rack->r_ctl.rc_gp_output_ts = my_rsm->usec_orig_send; if (my_rsm->r_flags & RACK_ACKED) { /* * Unlikely since its probably what was * just transmitted (but I am paranoid). */ tp->gput_ts = my_rsm->r_ack_arrival; rack->app_limited_needs_set = 0; } if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) { /* This also is unlikely */ tp->gput_seq = my_rsm->r_start; } } else { /* * TSNH unless we have some send-map limit, * and even at that it should not be hitting * that limit (we should have stopped sending). */ rack->r_ctl.rc_gp_output_ts = tcp_get_usecs(NULL); } rack_log_pacing_delay_calc(rack, tp->gput_seq, tp->gput_ack, (uint64_t)my_rsm, tp->gput_ts, rack->r_ctl.rc_app_limited_cnt, 9, __LINE__, NULL); } static inline uint32_t rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cwnd_to_use, uint32_t avail, int32_t sb_offset) { uint32_t len; uint32_t sendwin; if (tp->snd_wnd > cwnd_to_use) sendwin = cwnd_to_use; else sendwin = tp->snd_wnd; if (ctf_outstanding(tp) >= tp->snd_wnd) { /* We never want to go over our peers rcv-window */ len = 0; } else { uint32_t flight; flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked); if (flight >= sendwin) { /* * We have in flight what we are allowed by cwnd (if * it was rwnd blocking it would have hit above out * >= tp->snd_wnd). */ return (0); } len = sendwin - flight; if ((len + ctf_outstanding(tp)) > tp->snd_wnd) { /* We would send too much (beyond the rwnd) */ len = tp->snd_wnd - ctf_outstanding(tp); } if ((len + sb_offset) > avail) { /* * We don't have that much in the SB, how much is * there? */ len = avail - sb_offset; } } return (len); } static int rack_output(struct tcpcb *tp) { struct socket *so; uint32_t recwin; uint32_t sb_offset; int32_t len, flags, error = 0; struct mbuf *m; struct mbuf *mb; uint32_t if_hw_tsomaxsegcount = 0; uint32_t if_hw_tsomaxsegsize; int32_t segsiz, minseg; long tot_len_this_send = 0; struct ip *ip = NULL; #ifdef TCPDEBUG struct ipovly *ipov = NULL; #endif struct udphdr *udp = NULL; struct tcp_rack *rack; struct tcphdr *th; uint8_t pass = 0; uint8_t mark = 0; uint8_t wanted_cookie = 0; u_char opt[TCP_MAXOLEN]; unsigned ipoptlen, optlen, hdrlen, ulen=0; uint32_t rack_seq; #if defined(IPSEC) || defined(IPSEC_SUPPORT) unsigned ipsec_optlen = 0; #endif int32_t idle, sendalot; int32_t sub_from_prr = 0; volatile int32_t sack_rxmit; struct rack_sendmap *rsm = NULL; int32_t tso, mtu; struct tcpopt to; int32_t slot = 0; int32_t sup_rack = 0; uint32_t cts, us_cts, delayed, early; uint8_t hpts_calling, new_data_tlp = 0, doing_tlp = 0; uint32_t cwnd_to_use; int32_t do_a_prefetch; int32_t prefetch_rsm = 0; int32_t orig_len; struct timeval tv; int32_t prefetch_so_done = 0; struct tcp_log_buffer *lgb = NULL; struct inpcb *inp; struct sockbuf *sb; #ifdef INET6 struct ip6_hdr *ip6 = NULL; int32_t isipv6; #endif uint8_t filled_all = 0; bool hw_tls = false; /* setup and take the cache hits here */ rack = (struct tcp_rack *)tp->t_fb_ptr; inp = rack->rc_inp; so = inp->inp_socket; sb = &so->so_snd; kern_prefetch(sb, &do_a_prefetch); do_a_prefetch = 1; hpts_calling = inp->inp_hpts_calls; hw_tls = (so->so_snd.sb_flags & SB_TLS_IFNET) != 0; NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(inp); #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) return (tcp_offload_output(tp)); #endif /* * For TFO connections in SYN_RECEIVED, only allow the initial * SYN|ACK and those sent by the retransmit timer. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_state == TCPS_SYN_RECEIVED) && SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN|ACK sent */ (rack->r_ctl.rc_resend == NULL)) /* not a retransmit */ return (0); #ifdef INET6 if (rack->r_state) { /* Use the cache line loaded if possible */ isipv6 = rack->r_is_v6; } else { isipv6 = (inp->inp_vflag & INP_IPV6) != 0; } #endif early = 0; us_cts = tcp_get_usecs(&tv); cts = tcp_tv_to_mssectick(&tv); if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && inp->inp_in_hpts) { /* * We are on the hpts for some timer but not hptsi output. * Remove from the hpts unconditionally. */ rack_timer_cancel(tp, rack, cts, __LINE__); } /* Are we pacing and late? */ if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) { /* We are delayed */ delayed = us_cts - rack->r_ctl.rc_last_output_to; } else { delayed = 0; } /* Do the timers, which may override the pacer */ if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { if (rack_process_timers(tp, rack, cts, hpts_calling)) { counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1); return (0); } } if ((rack->r_timer_override) || (delayed) || (tp->t_state < TCPS_ESTABLISHED)) { if (tp->t_inpcb->inp_in_hpts) tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT); } else if (tp->t_inpcb->inp_in_hpts) { /* * On the hpts you can't pass even if ACKNOW is on, we will * when the hpts fires. */ counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1); return (0); } inp->inp_hpts_calls = 0; /* Finish out both pacing early and late accounting */ if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) { early = rack->r_ctl.rc_last_output_to - us_cts; } else early = 0; if (delayed) { rack->r_ctl.rc_agg_delayed += delayed; rack->r_late = 1; } else if (early) { rack->r_ctl.rc_agg_early += early; rack->r_early = 1; } /* Now that early/late accounting is done turn off the flag */ rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT; rack->r_wanted_output = 0; rack->r_timer_override = 0; /* * For TFO connections in SYN_SENT or SYN_RECEIVED, * only allow the initial SYN or SYN|ACK and those sent * by the retransmit timer. */ if (IS_FASTOPEN(tp->t_flags) && ((tp->t_state == TCPS_SYN_RECEIVED) || (tp->t_state == TCPS_SYN_SENT)) && SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */ (tp->t_rxtshift == 0)) { /* not a retransmit */ cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd; goto just_return_nolock; } /* * Determine length of data that should be transmitted, and flags * that will be used. If there is some data or critical controls * (SYN, RST) to send, then transmit; otherwise, investigate * further. */ idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); if (tp->t_idle_reduce) { if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) rack_cc_after_idle(rack, tp); } tp->t_flags &= ~TF_LASTIDLE; if (idle) { if (tp->t_flags & TF_MORETOCOME) { tp->t_flags |= TF_LASTIDLE; idle = 0; } } if ((tp->snd_una == tp->snd_max) && rack->r_ctl.rc_went_idle_time && TSTMP_GT(us_cts, rack->r_ctl.rc_went_idle_time)) { idle = us_cts - rack->r_ctl.rc_went_idle_time; if (idle > rack_min_probertt_hold) { /* Count as a probe rtt */ if (rack->in_probe_rtt == 0) { rack->r_ctl.rc_lower_rtt_us_cts = us_cts; rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts; rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts; rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts; } else { rack_exit_probertt(rack, us_cts); } } idle = 0; } again: /* * If we've recently taken a timeout, snd_max will be greater than * snd_nxt. There may be SACK information that allows us to avoid * resending already delivered data. Adjust snd_nxt accordingly. */ sendalot = 0; us_cts = tcp_get_usecs(&tv); cts = tcp_tv_to_mssectick(&tv); tso = 0; mtu = 0; segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); minseg = segsiz; sb_offset = tp->snd_max - tp->snd_una; cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd; #ifdef NETFLIX_SHARED_CWND if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) && rack->rack_enable_scwnd) { /* We are doing cwnd sharing */ if (rack->rc_gp_filled && (rack->rack_attempted_scwnd == 0) && (rack->r_ctl.rc_scw == NULL) && tp->t_lib) { /* The pcbid is in, lets make an attempt */ counter_u64_add(rack_try_scwnd, 1); rack->rack_attempted_scwnd = 1; rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp, &rack->r_ctl.rc_scw_index, segsiz); } if (rack->r_ctl.rc_scw && (rack->rack_scwnd_is_idle == 1) && (rack->rc_in_persist == 0) && sbavail(sb)) { /* we are no longer out of data */ tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index); rack->rack_scwnd_is_idle = 0; } if (rack->r_ctl.rc_scw) { /* First lets update and get the cwnd */ rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index, tp->snd_cwnd, tp->snd_wnd, segsiz); } } #endif flags = tcp_outflags[tp->t_state]; while (rack->rc_free_cnt < rack_free_cache) { rsm = rack_alloc(rack); if (rsm == NULL) { if (inp->inp_hpts_calls) /* Retry in a ms */ slot = (1 * HPTS_USEC_IN_MSEC); goto just_return_nolock; } TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext); rack->rc_free_cnt++; rsm = NULL; } if (inp->inp_hpts_calls) inp->inp_hpts_calls = 0; sack_rxmit = 0; len = 0; rsm = NULL; if (flags & TH_RST) { SOCKBUF_LOCK(sb); goto send; } if (rack->r_ctl.rc_resend) { /* Retransmit timer */ rsm = rack->r_ctl.rc_resend; rack->r_ctl.rc_resend = NULL; rsm->r_flags &= ~RACK_TLP; len = rsm->r_end - rsm->r_start; sack_rxmit = 1; sendalot = 0; KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start), ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p", __func__, __LINE__, rsm->r_start, tp->snd_una, tp, rack, rsm)); sb_offset = rsm->r_start - tp->snd_una; if (len >= segsiz) len = segsiz; } else if ((rack->rc_in_persist == 0) && ((rsm = tcp_rack_output(tp, rack, cts)) != NULL)) { /* We have a retransmit that takes precedence */ rsm->r_flags &= ~RACK_TLP; if ((!IN_RECOVERY(tp->t_flags)) && ((tp->t_flags & (TF_WASFRECOVERY | TF_WASCRECOVERY)) == 0)) { /* Enter recovery if not induced by a time-out */ rack->r_ctl.rc_rsm_start = rsm->r_start; rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; rack_cong_signal(tp, NULL, CC_NDUPACK); /* * When we enter recovery we need to assure we send * one packet. */ if (rack->rack_no_prr == 0) { rack->r_ctl.rc_prr_sndcnt = segsiz; rack_log_to_prr(rack, 13, 0); } } #ifdef INVARIANTS if (SEQ_LT(rsm->r_start, tp->snd_una)) { panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n", tp, rack, rsm, rsm->r_start, tp->snd_una); } #endif len = rsm->r_end - rsm->r_start; KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start), ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p", __func__, __LINE__, rsm->r_start, tp->snd_una, tp, rack, rsm)); sb_offset = rsm->r_start - tp->snd_una; /* Can we send it within the PRR boundary? */ if (rack->rack_no_prr == 0) { if ((rack->use_rack_rr == 0) && (len > rack->r_ctl.rc_prr_sndcnt)) { /* It does not fit */ if ((ctf_flight_size(tp, rack->r_ctl.rc_sacked) > len) && (rack->r_ctl.rc_prr_sndcnt < segsiz)) { /* * prr is less than a segment, we * have more acks due in besides * what we need to resend. Lets not send * to avoid sending small pieces of * what we need to retransmit. */ len = 0; goto just_return_nolock; } len = rack->r_ctl.rc_prr_sndcnt; } } sendalot = 0; if (len >= segsiz) len = segsiz; if (len > 0) { sub_from_prr = 1; sack_rxmit = 1; KMOD_TCPSTAT_INC(tcps_sack_rexmits); KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes, min(len, segsiz)); counter_u64_add(rack_rtm_prr_retran, 1); } } else if (rack->r_ctl.rc_tlpsend) { /* Tail loss probe */ long cwin; long tlen; doing_tlp = 1; /* * Check if we can do a TLP with a RACK'd packet * this can happen if we are not doing the rack * cheat and we skipped to a TLP and it * went off. */ rsm = rack->r_ctl.rc_tlpsend; rsm->r_flags |= RACK_TLP; rack->r_ctl.rc_tlpsend = NULL; sack_rxmit = 1; tlen = rsm->r_end - rsm->r_start; if (tlen > segsiz) tlen = segsiz; KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start), ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p", __func__, __LINE__, rsm->r_start, tp->snd_una, tp, rack, rsm)); sb_offset = rsm->r_start - tp->snd_una; cwin = min(tp->snd_wnd, tlen); len = cwin; } /* * Enforce a connection sendmap count limit if set * as long as we are not retransmiting. */ if ((rsm == NULL) && (rack->do_detection == 0) && (V_tcp_map_entries_limit > 0) && (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { counter_u64_add(rack_to_alloc_limited, 1); if (!rack->alloc_limit_reported) { rack->alloc_limit_reported = 1; counter_u64_add(rack_alloc_limited_conns, 1); } goto just_return_nolock; } if (rsm && (rsm->r_flags & RACK_HAS_FIN)) { /* we are retransmitting the fin */ len--; if (len) { /* * When retransmitting data do *not* include the * FIN. This could happen from a TLP probe. */ flags &= ~TH_FIN; } } #ifdef INVARIANTS /* For debugging */ rack->r_ctl.rc_rsm_at_retran = rsm; #endif /* * Get standard flags, and add SYN or FIN if requested by 'hidden' * state flags. */ if (tp->t_flags & TF_NEEDFIN) flags |= TH_FIN; if (tp->t_flags & TF_NEEDSYN) flags |= TH_SYN; if ((sack_rxmit == 0) && (prefetch_rsm == 0)) { void *end_rsm; end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext); if (end_rsm) kern_prefetch(end_rsm, &prefetch_rsm); prefetch_rsm = 1; } SOCKBUF_LOCK(sb); /* * If snd_nxt == snd_max and we have transmitted a FIN, the * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a * negative length. This can also occur when TCP opens up its * congestion window while receiving additional duplicate acks after * fast-retransmit because TCP will reset snd_nxt to snd_max after * the fast-retransmit. * * In the normal retransmit-FIN-only case, however, snd_nxt will be * set to snd_una, the sb_offset will be 0, and the length may wind * up 0. * * If sack_rxmit is true we are retransmitting from the scoreboard * in which case len is already set. */ if ((sack_rxmit == 0) && (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) { uint32_t avail; avail = sbavail(sb); if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail) sb_offset = tp->snd_nxt - tp->snd_una; else sb_offset = 0; if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) { if (rack->r_ctl.rc_tlp_new_data) { /* TLP is forcing out new data */ if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) { rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset); } if (rack->r_ctl.rc_tlp_new_data > tp->snd_wnd) len = tp->snd_wnd; else len = rack->r_ctl.rc_tlp_new_data; rack->r_ctl.rc_tlp_new_data = 0; new_data_tlp = doing_tlp = 1; } else len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset); if (IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) { /* * For prr=off, we need to send only 1 MSS * at a time. We do this because another sack could * be arriving that causes us to send retransmits and * we don't want to be on a long pace due to a larger send * that keeps us from sending out the retransmit. */ len = segsiz; } } else { uint32_t outstanding; /* * We are inside of a SACK recovery episode and are * sending new data, having retransmitted all the * data possible so far in the scoreboard. */ outstanding = tp->snd_max - tp->snd_una; if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) { if (tp->snd_wnd > outstanding) { len = tp->snd_wnd - outstanding; /* Check to see if we have the data */ if ((sb_offset + len) > avail) { /* It does not all fit */ if (avail > sb_offset) len = avail - sb_offset; else len = 0; } } else len = 0; } else if (avail > sb_offset) len = avail - sb_offset; else len = 0; if (len > 0) { if (len > rack->r_ctl.rc_prr_sndcnt) len = rack->r_ctl.rc_prr_sndcnt; if (len > 0) { sub_from_prr = 1; counter_u64_add(rack_rtm_prr_newdata, 1); } } if (len > segsiz) { /* * We should never send more than a MSS when * retransmitting or sending new data in prr * mode unless the override flag is on. Most * likely the PRR algorithm is not going to * let us send a lot as well :-) */ if (rack->r_ctl.rc_prr_sendalot == 0) len = segsiz; } else if (len < segsiz) { /* * Do we send any? The idea here is if the * send empty's the socket buffer we want to * do it. However if not then lets just wait * for our prr_sndcnt to get bigger. */ long leftinsb; leftinsb = sbavail(sb) - sb_offset; if (leftinsb > len) { /* This send does not empty the sb */ len = 0; } } } } else if (!TCPS_HAVEESTABLISHED(tp->t_state)) { /* * If you have not established * and are not doing FAST OPEN * no data please. */ if ((sack_rxmit == 0) && (!IS_FASTOPEN(tp->t_flags))){ len = 0; sb_offset = 0; } } if (prefetch_so_done == 0) { kern_prefetch(so, &prefetch_so_done); prefetch_so_done = 1; } /* * Lop off SYN bit if it has already been sent. However, if this is * SYN-SENT state and if segment contains data and if we don't know * that foreign host supports TAO, suppress sending segment. */ if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) && ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) { /* * When sending additional segments following a TFO SYN|ACK, * do not include the SYN bit. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_state == TCPS_SYN_RECEIVED)) flags &= ~TH_SYN; } /* * Be careful not to send data and/or FIN on SYN segments. This * measure is needed to prevent interoperability problems with not * fully conformant TCP implementations. */ if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { len = 0; flags &= ~TH_FIN; } /* * On TFO sockets, ensure no data is sent in the following cases: * * - When retransmitting SYN|ACK on a passively-created socket * * - When retransmitting SYN on an actively created socket * * - When sending a zero-length cookie (cookie request) on an * actively created socket * * - When the socket is in the CLOSED state (RST is being sent) */ if (IS_FASTOPEN(tp->t_flags) && (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || ((tp->t_state == TCPS_SYN_SENT) && (tp->t_tfo_client_cookie_len == 0)) || (flags & TH_RST))) { sack_rxmit = 0; len = 0; } /* Without fast-open there should never be data sent on a SYN */ if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) { tp->snd_nxt = tp->iss; len = 0; } orig_len = len; if (len <= 0) { /* * If FIN has been sent but not acked, but we haven't been * called to retransmit, len will be < 0. Otherwise, window * shrank after we sent into it. If window shrank to 0, * cancel pending retransmit, pull snd_nxt back to (closed) * window, and set the persist timer if it isn't already * going. If the window didn't close completely, just wait * for an ACK. * * We also do a general check here to ensure that we will * set the persist timer when we have data to send, but a * 0-byte window. This makes sure the persist timer is set * even if the packet hits one of the "goto send" lines * below. */ len = 0; if ((tp->snd_wnd == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && (tp->snd_una == tp->snd_max) && (sb_offset < (int)sbavail(sb))) { tp->snd_nxt = tp->snd_una; rack_enter_persist(tp, rack, cts); } } else if ((rsm == NULL) && ((doing_tlp == 0) || (new_data_tlp == 1)) && (len < rack->r_ctl.rc_pace_max_segs)) { /* * We are not sending a maximum sized segment for * some reason. Should we not send anything (think * sws or persists)? */ if ((tp->snd_wnd < min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)), minseg)) && (TCPS_HAVEESTABLISHED(tp->t_state)) && (len < minseg) && (len < (int)(sbavail(sb) - sb_offset))) { /* * Here the rwnd is less than * the minimum pacing size, this is not a retransmit, * we are established and * the send is not the last in the socket buffer * we send nothing, and we may enter persists * if nothing is outstanding. */ len = 0; if (tp->snd_max == tp->snd_una) { /* * Nothing out we can * go into persists. */ rack_enter_persist(tp, rack, cts); tp->snd_nxt = tp->snd_una; } } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) && (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) && (len < (int)(sbavail(sb) - sb_offset)) && (len < minseg)) { /* * Here we are not retransmitting, and * the cwnd is not so small that we could * not send at least a min size (rxt timer * not having gone off), We have 2 segments or * more already in flight, its not the tail end * of the socket buffer and the cwnd is blocking * us from sending out a minimum pacing segment size. * Lets not send anything. */ len = 0; } else if (((tp->snd_wnd - ctf_outstanding(tp)) < min((rack->r_ctl.rc_high_rwnd/2), minseg)) && (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) && (len < (int)(sbavail(sb) - sb_offset)) && (TCPS_HAVEESTABLISHED(tp->t_state))) { /* * Here we have a send window but we have * filled it up and we can't send another pacing segment. * We also have in flight more than 2 segments * and we are not completing the sb i.e. we allow * the last bytes of the sb to go out even if * its not a full pacing segment. */ len = 0; } } /* len will be >= 0 after this point. */ KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); tcp_sndbuf_autoscale(tp, so, min(tp->snd_wnd, cwnd_to_use)); /* * Decide if we can use TCP Segmentation Offloading (if supported by * hardware). * * TSO may only be used if we are in a pure bulk sending state. The * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP * options prevent using TSO. With TSO the TCP header is the same * (except for the sequence number) for all generated packets. This * makes it impossible to transmit any options which vary per * generated segment or packet. * * IPv4 handling has a clear separation of ip options and ip header * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does * the right thing below to provide length of just ip options and thus * checking for ipoptlen is enough to decide if ip options are present. */ #ifdef INET6 if (isipv6) ipoptlen = ip6_optlen(tp->t_inpcb); else #endif if (tp->t_inpcb->inp_options) ipoptlen = tp->t_inpcb->inp_options->m_len - offsetof(struct ipoption, ipopt_list); else ipoptlen = 0; #if defined(IPSEC) || defined(IPSEC_SUPPORT) /* * Pre-calculate here as we save another lookup into the darknesses * of IPsec that way and can actually decide if TSO is ok. */ #ifdef INET6 if (isipv6 && IPSEC_ENABLED(ipv6)) ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb); #ifdef INET else #endif #endif /* INET6 */ #ifdef INET if (IPSEC_ENABLED(ipv4)) ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb); #endif /* INET */ #endif #if defined(IPSEC) || defined(IPSEC_SUPPORT) ipoptlen += ipsec_optlen; #endif if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz && (tp->t_port == 0) && ((tp->t_flags & TF_SIGNATURE) == 0) && tp->rcv_numsacks == 0 && sack_rxmit == 0 && ipoptlen == 0) tso = 1; { uint32_t outstanding; outstanding = tp->snd_max - tp->snd_una; if (tp->t_flags & TF_SENTFIN) { /* * If we sent a fin, snd_max is 1 higher than * snd_una */ outstanding--; } if (sack_rxmit) { if ((rsm->r_flags & RACK_HAS_FIN) == 0) flags &= ~TH_FIN; } else { if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + sbused(sb))) flags &= ~TH_FIN; } } recwin = lmin(lmax(sbspace(&so->so_rcv), 0), (long)TCP_MAXWIN << tp->rcv_scale); /* * Sender silly window avoidance. We transmit under the following * conditions when len is non-zero: * * - We have a full segment (or more with TSO) - This is the last * buffer in a write()/send() and we are either idle or running * NODELAY - we've timed out (e.g. persist timer) - we have more * then 1/2 the maximum send window's worth of data (receiver may be * limited the window size) - we need to retransmit */ if (len) { if (len >= segsiz) { goto send; } /* * NOTE! on localhost connections an 'ack' from the remote * end may occur synchronously with the output and cause us * to flush a buffer queued with moretocome. XXX * */ if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ (idle || (tp->t_flags & TF_NODELAY)) && ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) && (tp->t_flags & TF_NOPUSH) == 0) { pass = 2; goto send; } if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */ pass = 22; goto send; } if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) { pass = 4; goto send; } if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */ pass = 5; goto send; } if (sack_rxmit) { pass = 6; goto send; } if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) && (ctf_outstanding(tp) < (segsiz * 2))) { /* * We have less than two MSS outstanding (delayed ack) * and our rwnd will not let us send a full sized * MSS. Lets go ahead and let this small segment * out because we want to try to have at least two * packets inflight to not be caught by delayed ack. */ pass = 12; goto send; } } /* * Sending of standalone window updates. * * Window updates are important when we close our window due to a * full socket buffer and are opening it again after the application * reads data from it. Once the window has opened again and the * remote end starts to send again the ACK clock takes over and * provides the most current window information. * * We must avoid the silly window syndrome whereas every read from * the receive buffer, no matter how small, causes a window update * to be sent. We also should avoid sending a flurry of window * updates when the socket buffer had queued a lot of data and the * application is doing small reads. * * Prevent a flurry of pointless window updates by only sending an * update when we can increase the advertized window by more than * 1/4th of the socket buffer capacity. When the buffer is getting * full or is very small be more aggressive and send an update * whenever we can increase by two mss sized segments. In all other * situations the ACK's to new incoming data will carry further * window increases. * * Don't send an independent window update if a delayed ACK is * pending (it will get piggy-backed on it) or the remote side * already has done a half-close and won't send more data. Skip * this if the connection is in T/TCP half-open state. */ if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && !(tp->t_flags & TF_DELACK) && !TCPS_HAVERCVDFIN(tp->t_state)) { /* * "adv" is the amount we could increase the window, taking * into account that we are limited by TCP_MAXWIN << * tp->rcv_scale. */ int32_t adv; int oldwin; adv = recwin; if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { oldwin = (tp->rcv_adv - tp->rcv_nxt); if (adv > oldwin) adv -= oldwin; else { /* We can't increase the window */ adv = 0; } } else oldwin = 0; /* * If the new window size ends up being the same as or less * than the old size when it is scaled, then don't force * a window update. */ if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) goto dontupdate; if (adv >= (int32_t)(2 * segsiz) && (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) || recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) || so->so_rcv.sb_hiwat <= 8 * segsiz)) { pass = 7; goto send; } if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) { pass = 23; goto send; } } dontupdate: /* * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW * is also a catch-all for the retransmit timer timeout case. */ if (tp->t_flags & TF_ACKNOW) { pass = 8; goto send; } if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) { pass = 9; goto send; } /* * If our state indicates that FIN should be sent and we have not * yet done so, then we need to send. */ if ((flags & TH_FIN) && (tp->snd_nxt == tp->snd_una)) { pass = 11; goto send; } /* * No reason to send a segment, just return. */ just_return: SOCKBUF_UNLOCK(sb); just_return_nolock: { int app_limited = CTF_JR_SENT_DATA; if (tot_len_this_send > 0) { /* Make sure snd_nxt is up to max */ if (SEQ_GT(tp->snd_max, tp->snd_nxt)) tp->snd_nxt = tp->snd_max; slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz); } else { int end_window = 0; uint32_t seq = tp->gput_ack; rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm) { /* * Mark the last sent that we just-returned (hinting * that delayed ack may play a role in any rtt measurement). */ rsm->r_just_ret = 1; } counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1); rack->r_ctl.rc_agg_delayed = 0; rack->r_early = 0; rack->r_late = 0; rack->r_ctl.rc_agg_early = 0; if ((ctf_outstanding(tp) + min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)), minseg)) >= tp->snd_wnd) { /* We are limited by the rwnd */ app_limited = CTF_JR_RWND_LIMITED; } else if (ctf_outstanding(tp) >= sbavail(sb)) { /* We are limited by whats available -- app limited */ app_limited = CTF_JR_APP_LIMITED; } else if ((idle == 0) && ((tp->t_flags & TF_NODELAY) == 0) && ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) && (len < segsiz)) { /* * No delay is not on and the * user is sending less than 1MSS. This * brings out SWS avoidance so we * don't send. Another app-limited case. */ app_limited = CTF_JR_APP_LIMITED; } else if (tp->t_flags & TF_NOPUSH) { /* * The user has requested no push of * the last segment and we are * at the last segment. Another app * limited case. */ app_limited = CTF_JR_APP_LIMITED; } else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) { /* Its the cwnd */ app_limited = CTF_JR_CWND_LIMITED; } else if (rack->rc_in_persist == 1) { /* We are in persists */ app_limited = CTF_JR_PERSISTS; } else if (IN_RECOVERY(tp->t_flags) && (rack->rack_no_prr == 0) && (rack->r_ctl.rc_prr_sndcnt < segsiz)) { app_limited = CTF_JR_PRR; } else { /* Now why here are we not sending? */ #ifdef NOW #ifdef INVARIANTS panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use); #endif #endif app_limited = CTF_JR_ASSESSING; } /* * App limited in some fashion, for our pacing GP * measurements we don't want any gap (even cwnd). * Close down the measurement window. */ if (rack_cwnd_block_ends_measure && ((app_limited == CTF_JR_CWND_LIMITED) || (app_limited == CTF_JR_PRR))) { /* * The reason we are not sending is * the cwnd (or prr). We have been configured * to end the measurement window in * this case. */ end_window = 1; } else if (app_limited == CTF_JR_PERSISTS) { /* * We never end the measurement window * in persists, though in theory we * should be only entering after everything * is acknowledged (so we will probably * never come here). */ end_window = 0; } else if (rack_rwnd_block_ends_measure && (app_limited == CTF_JR_RWND_LIMITED)) { /* * We are rwnd limited and have been * configured to end the measurement * window in this case. */ end_window = 1; } else if (app_limited == CTF_JR_APP_LIMITED) { /* * A true application limited period, we have * ran out of data. */ end_window = 1; } else if (app_limited == CTF_JR_ASSESSING) { /* * In the assessing case we hit the end of * the if/else and had no known reason * This will panic us under invariants.. * * If we get this out in logs we need to * investagate which reason we missed. */ end_window = 1; } if (end_window) { uint8_t log = 0; if ((tp->t_flags & TF_GPUTINPROG) && SEQ_GT(tp->gput_ack, tp->snd_max)) { /* Mark the last packet has app limited */ tp->gput_ack = tp->snd_max; log = 1; } rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) { if (rack->r_ctl.rc_app_limited_cnt == 0) rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm; else { /* * Go out to the end app limited and mark * this new one as next and move the end_appl up * to this guy. */ if (rack->r_ctl.rc_end_appl) rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start; rack->r_ctl.rc_end_appl = rsm; } rsm->r_flags |= RACK_APP_LIMITED; rack->r_ctl.rc_app_limited_cnt++; } if (log) rack_log_pacing_delay_calc(rack, rack->r_ctl.rc_app_limited_cnt, seq, tp->gput_ack, 0, 0, 4, __LINE__, NULL); } } if (slot) { /* set the rack tcb into the slot N */ counter_u64_add(rack_paced_segments, 1); } else if (tot_len_this_send) { counter_u64_add(rack_unpaced_segments, 1); } /* Check if we need to go into persists or not */ if ((rack->rc_in_persist == 0) && (tp->snd_max == tp->snd_una) && TCPS_HAVEESTABLISHED(tp->t_state) && sbavail(sb) && (sbavail(sb) > tp->snd_wnd) && (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) { /* Yes lets make sure to move to persist before timer-start */ rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime); } rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack); rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use); } #ifdef NETFLIX_SHARED_CWND if ((sbavail(sb) == 0) && rack->r_ctl.rc_scw) { tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index); rack->rack_scwnd_is_idle = 1; } #endif return (0); send: if ((flags & TH_FIN) && sbavail(sb)) { /* * We do not transmit a FIN * with data outstanding. We * need to make it so all data * is acked first. */ flags &= ~TH_FIN; } /* Enforce stack imposed max seg size if we have one */ if (rack->r_ctl.rc_pace_max_segs && (len > rack->r_ctl.rc_pace_max_segs)) { mark = 1; len = rack->r_ctl.rc_pace_max_segs; } SOCKBUF_LOCK_ASSERT(sb); if (len > 0) { if (len >= segsiz) tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; else tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; } /* * Before ESTABLISHED, force sending of initial options unless TCP * set not to do any options. NOTE: we assume that the IP/TCP header * plus TCP options always fit in a single mbuf, leaving room for a * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr) * + optlen <= MCLBYTES */ optlen = 0; #ifdef INET6 if (isipv6) hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); else #endif hdrlen = sizeof(struct tcpiphdr); /* * Compute options for segment. We only have to care about SYN and * established connection segments. Options for SYN-ACK segments * are handled in TCP syncache. */ to.to_flags = 0; if ((tp->t_flags & TF_NOOPT) == 0) { /* Maximum segment size. */ if (flags & TH_SYN) { tp->snd_nxt = tp->iss; to.to_mss = tcp_mssopt(&inp->inp_inc); -#ifdef NETFLIX_TCPOUDP if (tp->t_port) to.to_mss -= V_tcp_udp_tunneling_overhead; -#endif to.to_flags |= TOF_MSS; /* * On SYN or SYN|ACK transmits on TFO connections, * only include the TFO option if it is not a * retransmit, as the presence of the TFO option may * have caused the original SYN or SYN|ACK to have * been dropped by a middlebox. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_rxtshift == 0)) { if (tp->t_state == TCPS_SYN_RECEIVED) { to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; to.to_tfo_cookie = (u_int8_t *)&tp->t_tfo_cookie.server; to.to_flags |= TOF_FASTOPEN; wanted_cookie = 1; } else if (tp->t_state == TCPS_SYN_SENT) { to.to_tfo_len = tp->t_tfo_client_cookie_len; to.to_tfo_cookie = tp->t_tfo_cookie.client; to.to_flags |= TOF_FASTOPEN; wanted_cookie = 1; /* * If we wind up having more data to * send with the SYN than can fit in * one segment, don't send any more * until the SYN|ACK comes back from * the other end. */ sendalot = 0; } } } /* Window scaling. */ if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { to.to_wscale = tp->request_r_scale; to.to_flags |= TOF_SCALE; } /* Timestamps. */ if ((tp->t_flags & TF_RCVD_TSTMP) || ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { to.to_tsval = cts + tp->ts_offset; to.to_tsecr = tp->ts_recent; to.to_flags |= TOF_TS; } /* Set receive buffer autosizing timestamp. */ if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE)) tp->rfbuf_ts = tcp_ts_getticks(); /* Selective ACK's. */ if (flags & TH_SYN) to.to_flags |= TOF_SACKPERM; else if (TCPS_HAVEESTABLISHED(tp->t_state) && tp->rcv_numsacks > 0) { to.to_flags |= TOF_SACK; to.to_nsacks = tp->rcv_numsacks; to.to_sacks = (u_char *)tp->sackblks; } #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) /* TCP-MD5 (RFC2385). */ if (tp->t_flags & TF_SIGNATURE) to.to_flags |= TOF_SIGNATURE; #endif /* TCP_SIGNATURE */ /* Processing the options. */ hdrlen += optlen = tcp_addoptions(&to, opt); /* * If we wanted a TFO option to be added, but it was unable * to fit, ensure no data is sent. */ if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && !(to.to_flags & TOF_FASTOPEN)) len = 0; } -#ifdef NETFLIX_TCPOUDP if (tp->t_port) { if (V_tcp_udp_tunneling_port == 0) { /* The port was removed?? */ SOCKBUF_UNLOCK(&so->so_snd); return (EHOSTUNREACH); } hdrlen += sizeof(struct udphdr); } -#endif #ifdef INET6 if (isipv6) ipoptlen = ip6_optlen(tp->t_inpcb); else #endif if (tp->t_inpcb->inp_options) ipoptlen = tp->t_inpcb->inp_options->m_len - offsetof(struct ipoption, ipopt_list); else ipoptlen = 0; #if defined(IPSEC) || defined(IPSEC_SUPPORT) ipoptlen += ipsec_optlen; #endif /* * Adjust data length if insertion of options will bump the packet * length beyond the t_maxseg length. Clear the FIN bit because we * cut off the tail of the segment. */ if (len + optlen + ipoptlen > tp->t_maxseg) { if (tso) { uint32_t if_hw_tsomax; uint32_t moff; int32_t max_len; /* extract TSO information */ if_hw_tsomax = tp->t_tsomax; if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; KASSERT(ipoptlen == 0, ("%s: TSO can't do IP options", __func__)); /* * Check if we should limit by maximum payload * length: */ if (if_hw_tsomax != 0) { /* compute maximum TSO length */ max_len = (if_hw_tsomax - hdrlen - max_linkhdr); if (max_len <= 0) { len = 0; } else if (len > max_len) { sendalot = 1; len = max_len; mark = 2; } } /* * Prevent the last segment from being fractional * unless the send sockbuf can be emptied: */ max_len = (tp->t_maxseg - optlen); if ((sb_offset + len) < sbavail(sb)) { moff = len % (u_int)max_len; if (moff != 0) { mark = 3; len -= moff; } } /* * In case there are too many small fragments don't * use TSO: */ if (len <= segsiz) { mark = 4; tso = 0; } /* * Send the FIN in a separate segment after the bulk * sending is done. We don't trust the TSO * implementations to clear the FIN flag on all but * the last segment. */ if (tp->t_flags & TF_NEEDFIN) { sendalot = 4; } } else { mark = 5; if (optlen + ipoptlen >= tp->t_maxseg) { /* * Since we don't have enough space to put * the IP header chain and the TCP header in * one packet as required by RFC 7112, don't * send it. Also ensure that at least one * byte of the payload can be put into the * TCP segment. */ SOCKBUF_UNLOCK(&so->so_snd); error = EMSGSIZE; sack_rxmit = 0; goto out; } len = tp->t_maxseg - optlen - ipoptlen; sendalot = 5; } } else { tso = 0; mark = 6; } KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, ("%s: len > IP_MAXPACKET", __func__)); #ifdef DIAGNOSTIC #ifdef INET6 if (max_linkhdr + hdrlen > MCLBYTES) #else if (max_linkhdr + hdrlen > MHLEN) #endif panic("tcphdr too big"); #endif /* * This KASSERT is here to catch edge cases at a well defined place. * Before, those had triggered (random) panic conditions further * down. */ KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); if ((len == 0) && (flags & TH_FIN) && (sbused(sb))) { /* * We have outstanding data, don't send a fin by itself!. */ goto just_return; } /* * Grab a header mbuf, attaching a copy of data to be transmitted, * and initialize the header from the template for sends on this * connection. */ if (len) { uint32_t max_val; uint32_t moff; if (rack->r_ctl.rc_pace_max_segs) max_val = rack->r_ctl.rc_pace_max_segs; else if (rack->rc_user_set_max_segs) max_val = rack->rc_user_set_max_segs * segsiz; else max_val = len; /* * We allow a limit on sending with hptsi. */ if (len > max_val) { mark = 7; len = max_val; } #ifdef INET6 if (MHLEN < hdrlen + max_linkhdr) m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); else #endif m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) { SOCKBUF_UNLOCK(sb); error = ENOBUFS; sack_rxmit = 0; goto out; } m->m_data += max_linkhdr; m->m_len = hdrlen; /* * Start the m_copy functions from the closest mbuf to the * sb_offset in the socket buffer chain. */ mb = sbsndptr_noadv(sb, sb_offset, &moff); if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) { m_copydata(mb, moff, (int)len, mtod(m, caddr_t)+hdrlen); if (SEQ_LT(tp->snd_nxt, tp->snd_max)) sbsndptr_adv(sb, mb, len); m->m_len += len; } else { struct sockbuf *msb; if (SEQ_LT(tp->snd_nxt, tp->snd_max)) msb = NULL; else msb = sb; m->m_next = tcp_m_copym( mb, moff, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb, ((rsm == NULL) ? hw_tls : 0) #ifdef NETFLIX_COPY_ARGS , &filled_all #endif ); if (len <= (tp->t_maxseg - optlen)) { /* * Must have ran out of mbufs for the copy * shorten it to no longer need tso. Lets * not put on sendalot since we are low on * mbufs. */ tso = 0; } if (m->m_next == NULL) { SOCKBUF_UNLOCK(sb); (void)m_free(m); error = ENOBUFS; sack_rxmit = 0; goto out; } } if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { if (rsm && (rsm->r_flags & RACK_TLP)) { /* * TLP should not count in retran count, but * in its own bin */ counter_u64_add(rack_tlp_retran, 1); counter_u64_add(rack_tlp_retran_bytes, len); } else { tp->t_sndrexmitpack++; KMOD_TCPSTAT_INC(tcps_sndrexmitpack); KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len); } #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, len); #endif } else { KMOD_TCPSTAT_INC(tcps_sndpack); KMOD_TCPSTAT_ADD(tcps_sndbyte, len); #ifdef STATS stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, len); #endif } /* * If we're sending everything we've got, set PUSH. (This * will keep happy those implementations which only give * data to the user when a buffer fills or a PUSH comes in.) */ if (sb_offset + len == sbused(sb) && sbused(sb) && !(flags & TH_SYN)) flags |= TH_PUSH; SOCKBUF_UNLOCK(sb); } else { SOCKBUF_UNLOCK(sb); if (tp->t_flags & TF_ACKNOW) KMOD_TCPSTAT_INC(tcps_sndacks); else if (flags & (TH_SYN | TH_FIN | TH_RST)) KMOD_TCPSTAT_INC(tcps_sndctrl); else KMOD_TCPSTAT_INC(tcps_sndwinup); m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) { error = ENOBUFS; sack_rxmit = 0; goto out; } #ifdef INET6 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && MHLEN >= hdrlen) { M_ALIGN(m, hdrlen); } else #endif m->m_data += max_linkhdr; m->m_len = hdrlen; } SOCKBUF_UNLOCK_ASSERT(sb); m->m_pkthdr.rcvif = (struct ifnet *)0; #ifdef MAC mac_inpcb_create_mbuf(inp, m); #endif #ifdef INET6 if (isipv6) { ip6 = mtod(m, struct ip6_hdr *); -#ifdef NETFLIX_TCPOUDP if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + sizeof(struct ip6_hdr)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; ulen = hdrlen + len - sizeof(struct ip6_hdr); udp->uh_ulen = htons(ulen); th = (struct tcphdr *)(udp + 1); - } else -#endif + } else { th = (struct tcphdr *)(ip6 + 1); - tcpip_fillheaders(inp, -#ifdef NETFLIX_TCPOUDP - tp->t_port, -#endif - ip6, th); + } + tcpip_fillheaders(inp, tp->t_port, ip6, th); } else #endif /* INET6 */ { ip = mtod(m, struct ip *); #ifdef TCPDEBUG ipov = (struct ipovly *)ip; #endif -#ifdef NETFLIX_TCPOUDP if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip + ipoptlen + sizeof(struct ip)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; ulen = hdrlen + len - sizeof(struct ip); udp->uh_ulen = htons(ulen); th = (struct tcphdr *)(udp + 1); - } else -#endif + } else { th = (struct tcphdr *)(ip + 1); - tcpip_fillheaders(inp, -#ifdef NETFLIX_TCPOUDP - tp->t_port, -#endif - ip, th); + } + tcpip_fillheaders(inp, tp->t_port, ip, th); } /* * Fill in fields, remembering maximum advertised window for use in * delaying messages about window sizes. If resending a FIN, be sure * not to use a new sequence number. */ if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && tp->snd_nxt == tp->snd_max) tp->snd_nxt--; /* * If we are starting a connection, send ECN setup SYN packet. If we * are on a retransmit, we may resend those bits a number of times * as per RFC 3168. */ if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) { if (tp->t_rxtshift >= 1) { if (tp->t_rxtshift <= V_tcp_ecn_maxretries) flags |= TH_ECE | TH_CWR; } else flags |= TH_ECE | TH_CWR; } /* Handle parallel SYN for ECN */ if ((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_flags2 & TF2_ECN_SND_ECE)) { flags |= TH_ECE; tp->t_flags2 &= ~TF2_ECN_SND_ECE; } if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->t_flags2 & TF2_ECN_PERMIT)) { /* * If the peer has ECN, mark data packets with ECN capable * transmission (ECT). Ignore pure ack packets, * retransmissions. */ if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && (sack_rxmit == 0)) { #ifdef INET6 if (isipv6) ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); else #endif ip->ip_tos |= IPTOS_ECN_ECT0; KMOD_TCPSTAT_INC(tcps_ecn_ect0); /* * Reply with proper ECN notifications. * Only set CWR on new data segments. */ if (tp->t_flags2 & TF2_ECN_SND_CWR) { flags |= TH_CWR; tp->t_flags2 &= ~TF2_ECN_SND_CWR; } } if (tp->t_flags2 & TF2_ECN_SND_ECE) flags |= TH_ECE; } /* * If we are doing retransmissions, then snd_nxt will not reflect * the first unsent octet. For ACK only packets, we do not want the * sequence number of the retransmitted packet, we want the sequence * number of the next unsent octet. So, if there is no data (and no * SYN or FIN), use snd_max instead of snd_nxt when filling in * ti_seq. But if we are in persist state, snd_max might reflect * one byte beyond the right edge of the window, so use snd_nxt in * that case, since we know we aren't doing a retransmission. * (retransmit and persist are mutually exclusive...) */ if (sack_rxmit == 0) { if (len || (flags & (TH_SYN | TH_FIN)) || rack->rc_in_persist) { th->th_seq = htonl(tp->snd_nxt); rack_seq = tp->snd_nxt; } else if (flags & TH_RST) { /* * For a Reset send the last cum ack in sequence * (this like any other choice may still generate a * challenge ack, if a ack-update packet is in * flight). */ th->th_seq = htonl(tp->snd_una); rack_seq = tp->snd_una; } else { th->th_seq = htonl(tp->snd_max); rack_seq = tp->snd_max; } } else { th->th_seq = htonl(rsm->r_start); rack_seq = rsm->r_start; } th->th_ack = htonl(tp->rcv_nxt); if (optlen) { bcopy(opt, th + 1, optlen); th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; } th->th_flags = flags; /* * Calculate receive window. Don't shrink window, but avoid silly * window syndrome. * If a RST segment is sent, advertise a window of zero. */ if (flags & TH_RST) { recwin = 0; } else { if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && recwin < (long)segsiz) recwin = 0; if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) recwin = (long)(tp->rcv_adv - tp->rcv_nxt); } /* * According to RFC1323 the window field in a SYN (i.e., a or * ) segment itself is never scaled. The case is * handled in syncache. */ if (flags & TH_SYN) th->th_win = htons((u_short) (min(sbspace(&so->so_rcv), TCP_MAXWIN))); else { /* Avoid shrinking window with window scaling. */ recwin = roundup2(recwin, 1 << tp->rcv_scale); th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); } /* * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0 * window. This may cause the remote transmitter to stall. This * flag tells soreceive() to disable delayed acknowledgements when * draining the buffer. This can occur if the receiver is * attempting to read more data than can be buffered prior to * transmitting on the connection. */ if (th->th_win == 0) { tp->t_sndzerowin++; tp->t_flags |= TF_RXWIN0SENT; } else tp->t_flags &= ~TF_RXWIN0SENT; tp->snd_up = tp->snd_una; /* drag it along, its deprecated */ #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (to.to_flags & TOF_SIGNATURE) { /* * Calculate MD5 signature and put it into the place * determined before. * NOTE: since TCP options buffer doesn't point into * mbuf's data, calculate offset and use it. */ if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, (u_char *)(th + 1) + (to.to_signature - opt)) != 0) { /* * Do not send segment if the calculation of MD5 * digest has failed. */ goto out; } } #endif /* * Put TCP length in extended header, and then checksum extended * header and data. */ m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ #ifdef INET6 if (isipv6) { /* * ip6_plen is not need to be filled now, and will be filled * in ip6_output. */ if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 0); } } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { m->m_pkthdr.csum_flags = CSUM_TCP; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); } /* IP version must be set here for ipv4/ipv6 checking later */ KASSERT(ip->ip_v == IPVERSION, ("%s: IP version incorrect: %d", __func__, ip->ip_v)); } #endif /* * Enable TSO and specify the size of the segments. The TCP pseudo * header checksum is always provided. XXX: Fixme: This is currently * not the case for IPv6. */ if (tso) { KASSERT(len > tp->t_maxseg - optlen, ("%s: len <= tso_segsz", __func__)); m->m_pkthdr.csum_flags |= CSUM_TSO; m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; } KASSERT(len + hdrlen == m_length(m, NULL), ("%s: mbuf chain different than expected: %d + %u != %u", __func__, len, hdrlen, m_length(m, NULL))); #ifdef TCP_HHOOK /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ hhook_run_tcp_est_out(tp, th, &to, len, tso); #endif #ifdef TCPDEBUG /* * Trace. */ if (so->so_options & SO_DEBUG) { u_short save = 0; #ifdef INET6 if (!isipv6) #endif { save = ipov->ih_len; ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + * (th->th_off << 2) */ ); } tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); #ifdef INET6 if (!isipv6) #endif ipov->ih_len = save; } #endif /* TCPDEBUG */ /* We're getting ready to send; log now. */ if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; if (rack->rack_no_prr) log.u_bbr.flex1 = 0; else log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs; log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs; log.u_bbr.flex4 = orig_len; if (filled_all) log.u_bbr.flex5 = 0x80000000; else log.u_bbr.flex5 = 0; /* Save off the early/late values */ log.u_bbr.flex6 = rack->r_ctl.rc_agg_early; log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed; log.u_bbr.bw_inuse = rack_get_bw(rack); if (rsm || sack_rxmit) { if (doing_tlp) log.u_bbr.flex8 = 2; else log.u_bbr.flex8 = 1; } else { log.u_bbr.flex8 = 0; } log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm); log.u_bbr.flex7 = mark; log.u_bbr.pkts_out = tp->t_maxseg; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.lt_epoch = cwnd_to_use; log.u_bbr.delivered = sendalot; lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, len, &log, false, NULL, NULL, 0, &tv); } else lgb = NULL; /* * Fill in IP length and desired time to live and send to IP level. * There should be a better way to handle ttl and tos; we could keep * them in the template, but need a way to checksum without them. */ /* * m->m_pkthdr.len should have been set before cksum calcuration, * because in6_cksum() need it. */ #ifdef INET6 if (isipv6) { /* * we separately set hoplimit for every segment, since the * user might want to change the value via setsockopt. Also, * desired default hop limit might be changed via Neighbor * Discovery. */ ip6->ip6_hlim = in6_selecthlim(inp, NULL); /* * Set the packet size here for the benefit of DTrace * probes. ip6_output() will set it properly; it's supposed * to include the option header lengths as well. */ ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) tp->t_flags2 |= TF2_PLPMTU_PMTUD; else tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; if (tp->t_state == TCPS_SYN_SENT) TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); TCP_PROBE5(send, NULL, tp, ip6, tp, th); /* TODO: IPv6 IP6TOS_ECT bit on */ error = ip6_output(m, inp->in6p_outputopts, &inp->inp_route6, ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), NULL, NULL, inp); if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL) mtu = inp->inp_route6.ro_nh->nh_mtu; } #endif /* INET6 */ #if defined(INET) && defined(INET6) else #endif #ifdef INET { ip->ip_len = htons(m->m_pkthdr.len); #ifdef INET6 if (inp->inp_vflag & INP_IPV6PROTO) ip->ip_ttl = in6_selecthlim(inp, NULL); #endif /* INET6 */ /* * If we do path MTU discovery, then we set DF on every * packet. This might not be the best thing to do according * to RFC3390 Section 2. However the tcp hostcache migitates * the problem so it affects only the first tcp connection * with a host. * * NB: Don't set DF on small MTU/MSS to have a safe * fallback. */ if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; if (tp->t_port == 0 || len < V_tcp_minmss) { ip->ip_off |= htons(IP_DF); } } else { tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; } if (tp->t_state == TCPS_SYN_SENT) TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); TCP_PROBE5(send, NULL, tp, ip, tp, th); error = ip_output(m, inp->inp_options, &inp->inp_route, ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0, inp); if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL) mtu = inp->inp_route.ro_nh->nh_mtu; } #endif /* INET */ out: if (lgb) { lgb->tlb_errno = error; lgb = NULL; } /* * In transmit state, time the transmission and arrange for the * retransmit. In persist state, just set snd_max. */ if (error == 0) { rack->forced_ack = 0; /* If we send something zap the FA flag */ if (rsm && (doing_tlp == 0)) { /* Set we retransmitted */ rack->rc_gp_saw_rec = 1; } else { if (cwnd_to_use > tp->snd_ssthresh) { /* Set we sent in CA */ rack->rc_gp_saw_ca = 1; } else { /* Set we sent in SS */ rack->rc_gp_saw_ss = 1; } } if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->t_flags & TF_SACK_PERMIT) && tp->rcv_numsacks > 0) tcp_clean_dsack_blocks(tp); tot_len_this_send += len; if (len == 0) counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1); else if (len == 1) { counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1); } else if (len > 1) { int idx; idx = (len / segsiz) + 3; if (idx >= TCP_MSS_ACCT_ATIMER) counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1); else counter_u64_add(rack_out_size[idx], 1); } } if (rack->rack_no_prr == 0) { if (sub_from_prr && (error == 0)) { if (rack->r_ctl.rc_prr_sndcnt >= len) rack->r_ctl.rc_prr_sndcnt -= len; else rack->r_ctl.rc_prr_sndcnt = 0; } } sub_from_prr = 0; rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error, cts, pass, rsm, us_cts); if ((error == 0) && (len > 0) && (tp->snd_una == tp->snd_max)) rack->r_ctl.rc_tlp_rxt_last_time = cts; /* Now are we in persists? */ if (rack->rc_in_persist == 0) { tcp_seq startseq = tp->snd_nxt; /* Track our lost count */ if (rsm && (doing_tlp == 0)) rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start; /* * Advance snd_nxt over sequence space of this segment. */ if (error) /* We don't log or do anything with errors */ goto nomore; if (doing_tlp == 0) { if (rsm == NULL) { /* * Not a retransmission of some * sort, new data is going out so * clear our TLP count and flag. */ rack->rc_tlp_in_progress = 0; rack->r_ctl.rc_tlp_cnt_out = 0; } } else { /* * We have just sent a TLP, mark that it is true * and make sure our in progress is set so we * continue to check the count. */ rack->rc_tlp_in_progress = 1; rack->r_ctl.rc_tlp_cnt_out++; } if (flags & (TH_SYN | TH_FIN)) { if (flags & TH_SYN) tp->snd_nxt++; if (flags & TH_FIN) { tp->snd_nxt++; tp->t_flags |= TF_SENTFIN; } } /* In the ENOBUFS case we do *not* update snd_max */ if (sack_rxmit) goto nomore; tp->snd_nxt += len; if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { if (tp->snd_una == tp->snd_max) { /* * Update the time we just added data since * none was outstanding. */ rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__); tp->t_acktime = ticks; } tp->snd_max = tp->snd_nxt; /* * Time this transmission if not a retransmission and * not currently timing anything. * This is only relevant in case of switching back to * the base stack. */ if (tp->t_rtttime == 0) { tp->t_rtttime = ticks; tp->t_rtseq = startseq; KMOD_TCPSTAT_INC(tcps_segstimed); } if (len && ((tp->t_flags & TF_GPUTINPROG) == 0)) rack_start_gp_measurement(tp, rack, startseq, sb_offset); } } else { /* * Persist case, update snd_max but since we are in persist * mode (no window) we do not update snd_nxt. */ int32_t xlen = len; if (error) goto nomore; if (flags & TH_SYN) ++xlen; if (flags & TH_FIN) { ++xlen; tp->t_flags |= TF_SENTFIN; } /* In the ENOBUFS case we do *not* update snd_max */ if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) { if (tp->snd_una == tp->snd_max) { /* * Update the time we just added data since * none was outstanding. */ rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__); tp->t_acktime = ticks; } tp->snd_max = tp->snd_nxt + len; } } nomore: if (error) { rack->r_ctl.rc_agg_delayed = 0; rack->r_early = 0; rack->r_late = 0; rack->r_ctl.rc_agg_early = 0; SOCKBUF_UNLOCK_ASSERT(sb); /* Check gotos. */ /* * Failures do not advance the seq counter above. For the * case of ENOBUFS we will fall out and retry in 1ms with * the hpts. Everything else will just have to retransmit * with the timer. * * In any case, we do not want to loop around for another * send without a good reason. */ sendalot = 0; switch (error) { case EPERM: tp->t_softerror = error; return (error); case ENOBUFS: if (slot == 0) { /* * Pace us right away to retry in a some * time */ slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC); if (rack->rc_enobuf < 126) rack->rc_enobuf++; if (slot > ((rack->rc_rack_rtt / 2) * HPTS_USEC_IN_MSEC)) { slot = (rack->rc_rack_rtt / 2) * HPTS_USEC_IN_MSEC; } if (slot < (10 * HPTS_USEC_IN_MSEC)) slot = 10 * HPTS_USEC_IN_MSEC; } counter_u64_add(rack_saw_enobuf, 1); error = 0; goto enobufs; case EMSGSIZE: /* * For some reason the interface we used initially * to send segments changed to another or lowered * its MTU. If TSO was active we either got an * interface without TSO capabilits or TSO was * turned off. If we obtained mtu from ip_output() * then update it and try again. */ if (tso) tp->t_flags &= ~TF_TSO; if (mtu != 0) { tcp_mss_update(tp, -1, mtu, NULL, NULL); goto again; } slot = 10 * HPTS_USEC_IN_MSEC; rack_start_hpts_timer(rack, tp, cts, slot, 0, 0); return (error); case ENETUNREACH: counter_u64_add(rack_saw_enetunreach, 1); case EHOSTDOWN: case EHOSTUNREACH: case ENETDOWN: if (TCPS_HAVERCVDSYN(tp->t_state)) { tp->t_softerror = error; } /* FALLTHROUGH */ default: slot = 10 * HPTS_USEC_IN_MSEC; rack_start_hpts_timer(rack, tp, cts, slot, 0, 0); return (error); } } else { rack->rc_enobuf = 0; } KMOD_TCPSTAT_INC(tcps_sndtotal); /* * Data sent (as far as we can tell). If this advertises a larger * window than any other segment, then remember the size of the * advertised window. Any pending ACK has now been sent. */ if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) tp->rcv_adv = tp->rcv_nxt + recwin; tp->last_ack_sent = tp->rcv_nxt; tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); enobufs: /* Assure when we leave that snd_nxt will point to top */ if (SEQ_GT(tp->snd_max, tp->snd_nxt)) tp->snd_nxt = tp->snd_max; if (sendalot) { /* Do we need to turn off sendalot? */ if (rack->r_ctl.rc_pace_max_segs && (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) { /* We hit our max. */ sendalot = 0; } else if ((rack->rc_user_set_max_segs) && (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) { /* We hit the user defined max */ sendalot = 0; } } if ((error == 0) && (flags & TH_FIN)) tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN); if (flags & TH_RST) { /* * We don't send again after sending a RST. */ slot = 0; sendalot = 0; if (error == 0) tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); } else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) { /* * Get our pacing rate, if an error * occurred in sending (ENOBUF) we would * hit the else if with slot preset. Other * errors return. */ slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz); } if (rsm && rack->use_rack_rr) { /* Its a retransmit and we use the rack cheat? */ if ((slot == 0) || (rack->rc_always_pace == 0) || (rack->r_rr_config == 1)) { /* * We have no pacing set or we * are using old-style rack or * we are overriden to use the old 1ms pacing. */ slot = rack->r_ctl.rc_min_to * HPTS_USEC_IN_MSEC; } } if (slot) { /* set the rack tcb into the slot N */ counter_u64_add(rack_paced_segments, 1); } else if (sendalot) { if (len) counter_u64_add(rack_unpaced_segments, 1); sack_rxmit = 0; goto again; } else if (len) { counter_u64_add(rack_unpaced_segments, 1); } rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0); return (error); } static void rack_update_seg(struct tcp_rack *rack) { uint32_t orig_val; orig_val = rack->r_ctl.rc_pace_max_segs; rack_set_pace_segments(rack->rc_tp, rack, __LINE__); if (orig_val != rack->r_ctl.rc_pace_max_segs) rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL); } /* * rack_ctloutput() must drop the inpcb lock before performing copyin on * socket option arguments. When it re-acquires the lock after the copy, it * has to revalidate that the connection is still valid for the socket * option. */ static int rack_set_sockopt(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack) { struct epoch_tracker et; uint64_t val; int32_t error = 0, optval; uint16_t ca, ss; switch (sopt->sopt_name) { case TCP_RACK_PROP_RATE: /* URL:prop_rate */ case TCP_RACK_PROP : /* URL:prop */ case TCP_RACK_TLP_REDUCE: /* URL:tlp_reduce */ case TCP_RACK_EARLY_RECOV: /* URL:early_recov */ case TCP_RACK_PACE_REDUCE: /* Not used */ /* Pacing related ones */ case TCP_RACK_PACE_ALWAYS: /* URL:pace_always */ case TCP_BBR_RACK_INIT_RATE: /* URL:irate */ case TCP_BBR_IWINTSO: /* URL:tso_iwin */ case TCP_RACK_PACE_MAX_SEG: /* URL:pace_max_seg */ case TCP_RACK_FORCE_MSEG: /* URL:force_max_seg */ case TCP_RACK_PACE_RATE_CA: /* URL:pr_ca */ case TCP_RACK_PACE_RATE_SS: /* URL:pr_ss*/ case TCP_RACK_PACE_RATE_REC: /* URL:pr_rec */ case TCP_RACK_GP_INCREASE_CA: /* URL:gp_inc_ca */ case TCP_RACK_GP_INCREASE_SS: /* URL:gp_inc_ss */ case TCP_RACK_GP_INCREASE_REC: /* URL:gp_inc_rec */ case TCP_RACK_RR_CONF: /* URL:rrr_conf */ case TCP_BBR_HDWR_PACE: /* URL:hdwrpace */ /* End pacing related */ case TCP_DELACK: case TCP_RACK_PRR_SENDALOT: /* URL:prr_sendalot */ case TCP_RACK_MIN_TO: /* URL:min_to */ case TCP_RACK_EARLY_SEG: /* URL:early_seg */ case TCP_RACK_REORD_THRESH: /* URL:reord_thresh */ case TCP_RACK_REORD_FADE: /* URL:reord_fade */ case TCP_RACK_TLP_THRESH: /* URL:tlp_thresh */ case TCP_RACK_PKT_DELAY: /* URL:pkt_delay */ case TCP_RACK_TLP_USE: /* URL:tlp_use */ case TCP_RACK_TLP_INC_VAR: /* URL:tlp_inc_var */ case TCP_RACK_IDLE_REDUCE_HIGH: /* URL:idle_reduce_high */ case TCP_BBR_RACK_RTT_USE: /* URL:rttuse */ case TCP_BBR_USE_RACK_RR: /* URL:rackrr */ case TCP_RACK_DO_DETECTION: /* URL:detect */ case TCP_NO_PRR: /* URL:noprr */ case TCP_TIMELY_DYN_ADJ: /* URL:dynamic */ case TCP_DATA_AFTER_CLOSE: case TCP_RACK_NONRXT_CFG_RATE: /* URL:nonrxtcr */ case TCP_SHARED_CWND_ENABLE: /* URL:scwnd */ case TCP_RACK_MBUF_QUEUE: /* URL:mqueue */ case TCP_RACK_NO_PUSH_AT_MAX: /* URL:npush */ case TCP_RACK_PACE_TO_FILL: /* URL:fillcw */ case TCP_SHARED_CWND_TIME_LIMIT: /* URL:lscwnd */ case TCP_RACK_PROFILE: /* URL:profile */ break; default: return (tcp_default_ctloutput(so, sopt, inp, tp)); break; } INP_WUNLOCK(inp); error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); if (error) return (error); INP_WLOCK(inp); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { INP_WUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); rack = (struct tcp_rack *)tp->t_fb_ptr; switch (sopt->sopt_name) { case TCP_RACK_PROFILE: RACK_OPTS_INC(tcp_profile); if (optval == 1) { /* pace_always=1 */ rack->rc_always_pace = 1; tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ; /* scwnd=1 */ rack->rack_enable_scwnd = 1; /* dynamic=100 */ rack->rc_gp_dyn_mul = 1; rack->r_ctl.rack_per_of_gp_ca = 100; /* rrr_conf=3 */ rack->r_rr_config = 3; /* npush=2 */ rack->r_ctl.rc_no_push_at_mrtt = 2; /* fillcw=1 */ rack->rc_pace_to_cwnd = 1; rack->rc_pace_fill_if_rttin_range = 0; rack->rtt_limit_mul = 0; /* noprr=1 */ rack->rack_no_prr = 1; /* lscwnd=1 */ rack->r_limit_scw = 1; } else if (optval == 2) { /* pace_always=1 */ rack->rc_always_pace = 1; tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ; /* scwnd=1 */ rack->rack_enable_scwnd = 1; /* dynamic=100 */ rack->rc_gp_dyn_mul = 1; rack->r_ctl.rack_per_of_gp_ca = 100; /* rrr_conf=3 */ rack->r_rr_config = 3; /* npush=2 */ rack->r_ctl.rc_no_push_at_mrtt = 2; /* fillcw=1 */ rack->rc_pace_to_cwnd = 1; rack->rc_pace_fill_if_rttin_range = 0; rack->rtt_limit_mul = 0; /* noprr=1 */ rack->rack_no_prr = 1; /* lscwnd=0 */ rack->r_limit_scw = 0; } break; case TCP_SHARED_CWND_TIME_LIMIT: RACK_OPTS_INC(tcp_lscwnd); if (optval) rack->r_limit_scw = 1; else rack->r_limit_scw = 0; break; case TCP_RACK_PACE_TO_FILL: RACK_OPTS_INC(tcp_fillcw); if (optval == 0) rack->rc_pace_to_cwnd = 0; else rack->rc_pace_to_cwnd = 1; if ((optval >= rack_gp_rtt_maxmul) && rack_gp_rtt_maxmul && (optval < 0xf)) { rack->rc_pace_fill_if_rttin_range = 1; rack->rtt_limit_mul = optval; } else { rack->rc_pace_fill_if_rttin_range = 0; rack->rtt_limit_mul = 0; } break; case TCP_RACK_NO_PUSH_AT_MAX: RACK_OPTS_INC(tcp_npush); if (optval == 0) rack->r_ctl.rc_no_push_at_mrtt = 0; else if (optval < 0xff) rack->r_ctl.rc_no_push_at_mrtt = optval; else error = EINVAL; break; case TCP_SHARED_CWND_ENABLE: RACK_OPTS_INC(tcp_rack_scwnd); if (optval == 0) rack->rack_enable_scwnd = 0; else rack->rack_enable_scwnd = 1; break; case TCP_RACK_MBUF_QUEUE: /* Now do we use the LRO mbuf-queue feature */ RACK_OPTS_INC(tcp_rack_mbufq); if (optval) rack->r_mbuf_queue = 1; else rack->r_mbuf_queue = 0; if (rack->r_mbuf_queue || rack->rc_always_pace) tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ; else tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; break; case TCP_RACK_NONRXT_CFG_RATE: RACK_OPTS_INC(tcp_rack_cfg_rate); if (optval == 0) rack->rack_rec_nonrxt_use_cr = 0; else rack->rack_rec_nonrxt_use_cr = 1; break; case TCP_NO_PRR: RACK_OPTS_INC(tcp_rack_noprr); if (optval == 0) rack->rack_no_prr = 0; else rack->rack_no_prr = 1; break; case TCP_TIMELY_DYN_ADJ: RACK_OPTS_INC(tcp_timely_dyn); if (optval == 0) rack->rc_gp_dyn_mul = 0; else { rack->rc_gp_dyn_mul = 1; if (optval >= 100) { /* * If the user sets something 100 or more * its the gp_ca value. */ rack->r_ctl.rack_per_of_gp_ca = optval; } } break; case TCP_RACK_DO_DETECTION: RACK_OPTS_INC(tcp_rack_do_detection); if (optval == 0) rack->do_detection = 0; else rack->do_detection = 1; break; case TCP_RACK_PROP_RATE: if ((optval <= 0) || (optval >= 100)) { error = EINVAL; break; } RACK_OPTS_INC(tcp_rack_prop_rate); rack->r_ctl.rc_prop_rate = optval; break; case TCP_RACK_TLP_USE: if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) { error = EINVAL; break; } RACK_OPTS_INC(tcp_tlp_use); rack->rack_tlp_threshold_use = optval; break; case TCP_RACK_PROP: /* RACK proportional rate reduction (bool) */ RACK_OPTS_INC(tcp_rack_prop); rack->r_ctl.rc_prop_reduce = optval; break; case TCP_RACK_TLP_REDUCE: /* RACK TLP cwnd reduction (bool) */ RACK_OPTS_INC(tcp_rack_tlp_reduce); rack->r_ctl.rc_tlp_cwnd_reduce = optval; break; case TCP_RACK_EARLY_RECOV: /* Should recovery happen early (bool) */ RACK_OPTS_INC(tcp_rack_early_recov); rack->r_ctl.rc_early_recovery = optval; break; /* Pacing related ones */ case TCP_RACK_PACE_ALWAYS: /* * zero is old rack method, 1 is new * method using a pacing rate. */ RACK_OPTS_INC(tcp_rack_pace_always); if (optval > 0) rack->rc_always_pace = 1; else rack->rc_always_pace = 0; if (rack->r_mbuf_queue || rack->rc_always_pace) tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ; else tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; /* A rate may be set irate or other, if so set seg size */ rack_update_seg(rack); break; case TCP_BBR_RACK_INIT_RATE: RACK_OPTS_INC(tcp_initial_rate); val = optval; /* Change from kbits per second to bytes per second */ val *= 1000; val /= 8; rack->r_ctl.init_rate = val; if (rack->rc_init_win != rack_default_init_window) { uint32_t win, snt; /* * Options don't always get applied * in the order you think. So in order * to assure we update a cwnd we need * to check and see if we are still * where we should raise the cwnd. */ win = rc_init_window(rack); if (SEQ_GT(tp->snd_max, tp->iss)) snt = tp->snd_max - tp->iss; else snt = 0; if ((snt < win) && (tp->snd_cwnd < win)) tp->snd_cwnd = win; } if (rack->rc_always_pace) rack_update_seg(rack); break; case TCP_BBR_IWINTSO: RACK_OPTS_INC(tcp_initial_win); if (optval && (optval <= 0xff)) { uint32_t win, snt; rack->rc_init_win = optval; win = rc_init_window(rack); if (SEQ_GT(tp->snd_max, tp->iss)) snt = tp->snd_max - tp->iss; else snt = 0; if ((snt < win) && (tp->t_srtt | #ifdef NETFLIX_PEAKRATE tp->t_maxpeakrate | #endif rack->r_ctl.init_rate)) { /* * We are not past the initial window * and we have some bases for pacing, * so we need to possibly adjust up * the cwnd. Note even if we don't set * the cwnd, its still ok to raise the rc_init_win * which can be used coming out of idle when we * would have a rate. */ if (tp->snd_cwnd < win) tp->snd_cwnd = win; } if (rack->rc_always_pace) rack_update_seg(rack); } else error = EINVAL; break; case TCP_RACK_FORCE_MSEG: RACK_OPTS_INC(tcp_rack_force_max_seg); if (optval) rack->rc_force_max_seg = 1; else rack->rc_force_max_seg = 0; break; case TCP_RACK_PACE_MAX_SEG: /* Max segments size in a pace in bytes */ RACK_OPTS_INC(tcp_rack_max_seg); rack->rc_user_set_max_segs = optval; rack_set_pace_segments(tp, rack, __LINE__); break; case TCP_RACK_PACE_RATE_REC: /* Set the fixed pacing rate in Bytes per second ca */ RACK_OPTS_INC(tcp_rack_pace_rate_rec); rack->r_ctl.rc_fixed_pacing_rate_rec = optval; if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0) rack->r_ctl.rc_fixed_pacing_rate_ca = optval; if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0) rack->r_ctl.rc_fixed_pacing_rate_ss = optval; rack->use_fixed_rate = 1; rack_log_pacing_delay_calc(rack, rack->r_ctl.rc_fixed_pacing_rate_ss, rack->r_ctl.rc_fixed_pacing_rate_ca, rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8, __LINE__, NULL); break; case TCP_RACK_PACE_RATE_SS: /* Set the fixed pacing rate in Bytes per second ca */ RACK_OPTS_INC(tcp_rack_pace_rate_ss); rack->r_ctl.rc_fixed_pacing_rate_ss = optval; if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0) rack->r_ctl.rc_fixed_pacing_rate_ca = optval; if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0) rack->r_ctl.rc_fixed_pacing_rate_rec = optval; rack->use_fixed_rate = 1; rack_log_pacing_delay_calc(rack, rack->r_ctl.rc_fixed_pacing_rate_ss, rack->r_ctl.rc_fixed_pacing_rate_ca, rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8, __LINE__, NULL); break; case TCP_RACK_PACE_RATE_CA: /* Set the fixed pacing rate in Bytes per second ca */ RACK_OPTS_INC(tcp_rack_pace_rate_ca); rack->r_ctl.rc_fixed_pacing_rate_ca = optval; if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0) rack->r_ctl.rc_fixed_pacing_rate_ss = optval; if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0) rack->r_ctl.rc_fixed_pacing_rate_rec = optval; rack->use_fixed_rate = 1; rack_log_pacing_delay_calc(rack, rack->r_ctl.rc_fixed_pacing_rate_ss, rack->r_ctl.rc_fixed_pacing_rate_ca, rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8, __LINE__, NULL); break; case TCP_RACK_GP_INCREASE_REC: RACK_OPTS_INC(tcp_gp_inc_rec); rack->r_ctl.rack_per_of_gp_rec = optval; rack_log_pacing_delay_calc(rack, rack->r_ctl.rack_per_of_gp_ss, rack->r_ctl.rack_per_of_gp_ca, rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1, __LINE__, NULL); break; case TCP_RACK_GP_INCREASE_CA: RACK_OPTS_INC(tcp_gp_inc_ca); ca = optval; if (ca < 100) { /* * We don't allow any reduction * over the GP b/w. */ error = EINVAL; break; } rack->r_ctl.rack_per_of_gp_ca = ca; rack_log_pacing_delay_calc(rack, rack->r_ctl.rack_per_of_gp_ss, rack->r_ctl.rack_per_of_gp_ca, rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1, __LINE__, NULL); break; case TCP_RACK_GP_INCREASE_SS: RACK_OPTS_INC(tcp_gp_inc_ss); ss = optval; if (ss < 100) { /* * We don't allow any reduction * over the GP b/w. */ error = EINVAL; break; } rack->r_ctl.rack_per_of_gp_ss = ss; rack_log_pacing_delay_calc(rack, rack->r_ctl.rack_per_of_gp_ss, rack->r_ctl.rack_per_of_gp_ca, rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1, __LINE__, NULL); break; case TCP_RACK_RR_CONF: RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate); if (optval && optval <= 3) rack->r_rr_config = optval; else rack->r_rr_config = 0; break; case TCP_BBR_HDWR_PACE: RACK_OPTS_INC(tcp_hdwr_pacing); if (optval){ if (rack->rack_hdrw_pacing == 0) { rack->rack_hdw_pace_ena = 1; rack->rack_attempt_hdwr_pace = 0; } else error = EALREADY; } else { rack->rack_hdw_pace_ena = 0; #ifdef RATELIMIT if (rack->rack_hdrw_pacing) { rack->rack_hdrw_pacing = 0; in_pcbdetach_txrtlmt(rack->rc_inp); } #endif } break; /* End Pacing related ones */ case TCP_RACK_PRR_SENDALOT: /* Allow PRR to send more than one seg */ RACK_OPTS_INC(tcp_rack_prr_sendalot); rack->r_ctl.rc_prr_sendalot = optval; break; case TCP_RACK_MIN_TO: /* Minimum time between rack t-o's in ms */ RACK_OPTS_INC(tcp_rack_min_to); rack->r_ctl.rc_min_to = optval; break; case TCP_RACK_EARLY_SEG: /* If early recovery max segments */ RACK_OPTS_INC(tcp_rack_early_seg); rack->r_ctl.rc_early_recovery_segs = optval; break; case TCP_RACK_REORD_THRESH: /* RACK reorder threshold (shift amount) */ RACK_OPTS_INC(tcp_rack_reord_thresh); if ((optval > 0) && (optval < 31)) rack->r_ctl.rc_reorder_shift = optval; else error = EINVAL; break; case TCP_RACK_REORD_FADE: /* Does reordering fade after ms time */ RACK_OPTS_INC(tcp_rack_reord_fade); rack->r_ctl.rc_reorder_fade = optval; break; case TCP_RACK_TLP_THRESH: /* RACK TLP theshold i.e. srtt+(srtt/N) */ RACK_OPTS_INC(tcp_rack_tlp_thresh); if (optval) rack->r_ctl.rc_tlp_threshold = optval; else error = EINVAL; break; case TCP_BBR_USE_RACK_RR: RACK_OPTS_INC(tcp_rack_rr); if (optval) rack->use_rack_rr = 1; else rack->use_rack_rr = 0; break; case TCP_RACK_PKT_DELAY: /* RACK added ms i.e. rack-rtt + reord + N */ RACK_OPTS_INC(tcp_rack_pkt_delay); rack->r_ctl.rc_pkt_delay = optval; break; case TCP_RACK_TLP_INC_VAR: /* Does TLP include rtt variance in t-o */ error = EINVAL; break; case TCP_RACK_IDLE_REDUCE_HIGH: error = EINVAL; break; case TCP_DELACK: if (optval == 0) tp->t_delayed_ack = 0; else tp->t_delayed_ack = 1; if (tp->t_flags & TF_DELACK) { tp->t_flags &= ~TF_DELACK; tp->t_flags |= TF_ACKNOW; NET_EPOCH_ENTER(et); rack_output(tp); NET_EPOCH_EXIT(et); } break; case TCP_BBR_RACK_RTT_USE: if ((optval != USE_RTT_HIGH) && (optval != USE_RTT_LOW) && (optval != USE_RTT_AVG)) error = EINVAL; else rack->r_ctl.rc_rate_sample_method = optval; break; case TCP_DATA_AFTER_CLOSE: if (optval) rack->rc_allow_data_af_clo = 1; else rack->rc_allow_data_af_clo = 0; break; case TCP_RACK_PACE_REDUCE: /* sysctl only now */ error = EINVAL; break; default: return (tcp_default_ctloutput(so, sopt, inp, tp)); break; } #ifdef NETFLIX_STATS tcp_log_socket_option(tp, sopt->sopt_name, optval, error); #endif INP_WUNLOCK(inp); return (error); } static int rack_get_sockopt(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack) { int32_t error, optval; uint64_t val; /* * Because all our options are either boolean or an int, we can just * pull everything into optval and then unlock and copy. If we ever * add a option that is not a int, then this will have quite an * impact to this routine. */ error = 0; switch (sopt->sopt_name) { case TCP_RACK_PROFILE: /* You cannot retrieve a profile, its write only */ error = EINVAL; break; case TCP_RACK_PACE_TO_FILL: optval = rack->rc_pace_to_cwnd; break; case TCP_RACK_NO_PUSH_AT_MAX: optval = rack->r_ctl.rc_no_push_at_mrtt; break; case TCP_SHARED_CWND_ENABLE: optval = rack->rack_enable_scwnd; break; case TCP_RACK_NONRXT_CFG_RATE: optval = rack->rack_rec_nonrxt_use_cr; break; case TCP_NO_PRR: optval = rack->rack_no_prr; break; case TCP_RACK_DO_DETECTION: optval = rack->do_detection; break; case TCP_RACK_MBUF_QUEUE: /* Now do we use the LRO mbuf-queue feature */ optval = rack->r_mbuf_queue; break; case TCP_TIMELY_DYN_ADJ: optval = rack->rc_gp_dyn_mul; break; case TCP_BBR_IWINTSO: optval = rack->rc_init_win; break; case TCP_RACK_PROP_RATE: optval = rack->r_ctl.rc_prop_rate; break; case TCP_RACK_PROP: /* RACK proportional rate reduction (bool) */ optval = rack->r_ctl.rc_prop_reduce; break; case TCP_RACK_TLP_REDUCE: /* RACK TLP cwnd reduction (bool) */ optval = rack->r_ctl.rc_tlp_cwnd_reduce; break; case TCP_RACK_EARLY_RECOV: /* Should recovery happen early (bool) */ optval = rack->r_ctl.rc_early_recovery; break; case TCP_RACK_PACE_REDUCE: /* RACK Hptsi reduction factor (divisor) */ error = EINVAL; break; case TCP_BBR_RACK_INIT_RATE: val = rack->r_ctl.init_rate; /* convert to kbits per sec */ val *= 8; val /= 1000; optval = (uint32_t)val; break; case TCP_RACK_FORCE_MSEG: optval = rack->rc_force_max_seg; break; case TCP_RACK_PACE_MAX_SEG: /* Max segments in a pace */ optval = rack->rc_user_set_max_segs; break; case TCP_RACK_PACE_ALWAYS: /* Use the always pace method */ optval = rack->rc_always_pace; break; case TCP_RACK_PRR_SENDALOT: /* Allow PRR to send more than one seg */ optval = rack->r_ctl.rc_prr_sendalot; break; case TCP_RACK_MIN_TO: /* Minimum time between rack t-o's in ms */ optval = rack->r_ctl.rc_min_to; break; case TCP_RACK_EARLY_SEG: /* If early recovery max segments */ optval = rack->r_ctl.rc_early_recovery_segs; break; case TCP_RACK_REORD_THRESH: /* RACK reorder threshold (shift amount) */ optval = rack->r_ctl.rc_reorder_shift; break; case TCP_RACK_REORD_FADE: /* Does reordering fade after ms time */ optval = rack->r_ctl.rc_reorder_fade; break; case TCP_BBR_USE_RACK_RR: /* Do we use the rack cheat for rxt */ optval = rack->use_rack_rr; break; case TCP_RACK_RR_CONF: optval = rack->r_rr_config; break; case TCP_BBR_HDWR_PACE: optval = rack->rack_hdw_pace_ena; break; case TCP_RACK_TLP_THRESH: /* RACK TLP theshold i.e. srtt+(srtt/N) */ optval = rack->r_ctl.rc_tlp_threshold; break; case TCP_RACK_PKT_DELAY: /* RACK added ms i.e. rack-rtt + reord + N */ optval = rack->r_ctl.rc_pkt_delay; break; case TCP_RACK_TLP_USE: optval = rack->rack_tlp_threshold_use; break; case TCP_RACK_TLP_INC_VAR: /* Does TLP include rtt variance in t-o */ error = EINVAL; break; case TCP_RACK_IDLE_REDUCE_HIGH: error = EINVAL; break; case TCP_RACK_PACE_RATE_CA: optval = rack->r_ctl.rc_fixed_pacing_rate_ca; break; case TCP_RACK_PACE_RATE_SS: optval = rack->r_ctl.rc_fixed_pacing_rate_ss; break; case TCP_RACK_PACE_RATE_REC: optval = rack->r_ctl.rc_fixed_pacing_rate_rec; break; case TCP_RACK_GP_INCREASE_SS: optval = rack->r_ctl.rack_per_of_gp_ca; break; case TCP_RACK_GP_INCREASE_CA: optval = rack->r_ctl.rack_per_of_gp_ss; break; case TCP_BBR_RACK_RTT_USE: optval = rack->r_ctl.rc_rate_sample_method; break; case TCP_DELACK: optval = tp->t_delayed_ack; break; case TCP_DATA_AFTER_CLOSE: optval = rack->rc_allow_data_af_clo; break; case TCP_SHARED_CWND_TIME_LIMIT: optval = rack->r_limit_scw; break; default: return (tcp_default_ctloutput(so, sopt, inp, tp)); break; } INP_WUNLOCK(inp); if (error == 0) { error = sooptcopyout(sopt, &optval, sizeof optval); } return (error); } static int rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp) { int32_t error = EINVAL; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack == NULL) { /* Huh? */ goto out; } if (sopt->sopt_dir == SOPT_SET) { return (rack_set_sockopt(so, sopt, inp, tp, rack)); } else if (sopt->sopt_dir == SOPT_GET) { return (rack_get_sockopt(so, sopt, inp, tp, rack)); } out: INP_WUNLOCK(inp); return (error); } static int rack_pru_options(struct tcpcb *tp, int flags) { if (flags & PRUS_OOB) return (EOPNOTSUPP); return (0); } static struct tcp_function_block __tcp_rack = { .tfb_tcp_block_name = __XSTRING(STACKNAME), .tfb_tcp_output = rack_output, .tfb_do_queued_segments = ctf_do_queued_segments, .tfb_do_segment_nounlock = rack_do_segment_nounlock, .tfb_tcp_do_segment = rack_do_segment, .tfb_tcp_ctloutput = rack_ctloutput, .tfb_tcp_fb_init = rack_init, .tfb_tcp_fb_fini = rack_fini, .tfb_tcp_timer_stop_all = rack_stopall, .tfb_tcp_timer_activate = rack_timer_activate, .tfb_tcp_timer_active = rack_timer_active, .tfb_tcp_timer_stop = rack_timer_stop, .tfb_tcp_rexmit_tmr = rack_remxt_tmr, .tfb_tcp_handoff_ok = rack_handoff_ok, .tfb_pru_options = rack_pru_options, }; static const char *rack_stack_names[] = { __XSTRING(STACKNAME), #ifdef STACKALIAS __XSTRING(STACKALIAS), #endif }; static int rack_ctor(void *mem, int32_t size, void *arg, int32_t how) { memset(mem, 0, size); return (0); } static void rack_dtor(void *mem, int32_t size, void *arg) { } static bool rack_mod_inited = false; static int tcp_addrack(module_t mod, int32_t type, void *data) { int32_t err = 0; int num_stacks; switch (type) { case MOD_LOAD: rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map", sizeof(struct rack_sendmap), rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0); rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb", sizeof(struct tcp_rack), rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); sysctl_ctx_init(&rack_sysctl_ctx); rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_STATIC_CHILDREN(_net_inet_tcp), OID_AUTO, #ifdef STACKALIAS __XSTRING(STACKALIAS), #else __XSTRING(STACKNAME), #endif CTLFLAG_RW | CTLFLAG_MPSAFE, 0, ""); if (rack_sysctl_root == NULL) { printf("Failed to add sysctl node\n"); err = EFAULT; goto free_uma; } rack_init_sysctls(); num_stacks = nitems(rack_stack_names); err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK, rack_stack_names, &num_stacks); if (err) { printf("Failed to register %s stack name for " "%s module\n", rack_stack_names[num_stacks], __XSTRING(MODNAME)); sysctl_ctx_free(&rack_sysctl_ctx); free_uma: uma_zdestroy(rack_zone); uma_zdestroy(rack_pcb_zone); rack_counter_destroy(); printf("Failed to register rack module -- err:%d\n", err); return (err); } tcp_lro_reg_mbufq(); rack_mod_inited = true; break; case MOD_QUIESCE: err = deregister_tcp_functions(&__tcp_rack, true, false); break; case MOD_UNLOAD: err = deregister_tcp_functions(&__tcp_rack, false, true); if (err == EBUSY) break; if (rack_mod_inited) { uma_zdestroy(rack_zone); uma_zdestroy(rack_pcb_zone); sysctl_ctx_free(&rack_sysctl_ctx); rack_counter_destroy(); rack_mod_inited = false; } tcp_lro_dereg_mbufq(); err = 0; break; default: return (EOPNOTSUPP); } return (err); } static moduledata_t tcp_rack = { .name = __XSTRING(MODNAME), .evhand = tcp_addrack, .priv = 0 }; MODULE_VERSION(MODNAME, 1); DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1); diff --git a/sys/netinet/tcp_subr.c b/sys/netinet/tcp_subr.c index e973555efbcb..1ebc7357def3 100644 --- a/sys/netinet/tcp_subr.c +++ b/sys/netinet/tcp_subr.c @@ -1,3529 +1,3951 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95 */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_kern_tls.h" #include "opt_tcpdebug.h" #include #include #include #include #include #ifdef TCP_HHOOK #include #endif #include #ifdef TCP_HHOOK #include #endif #ifdef KERN_TLS #include #endif #include #include #include #include #include #include #include #ifdef INET6 #include #endif #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef INET6 #include #include #include #include #include #include #include #endif #include #include #include #include #include #include #include #include #include #ifdef INET6 #include #endif #include #include #ifdef TCPPCAP #include #endif #ifdef TCPDEBUG #include #endif #ifdef INET6 #include #endif #ifdef TCP_OFFLOAD #include #endif +#include +#include #include #include #include #include VNET_DEFINE(int, tcp_mssdflt) = TCP_MSS; #ifdef INET6 VNET_DEFINE(int, tcp_v6mssdflt) = TCP6_MSS; #endif #ifdef NETFLIX_EXP_DETECTION /* Sack attack detection thresholds and such */ SYSCTL_NODE(_net_inet_tcp, OID_AUTO, sack_attack, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Sack Attack detection thresholds"); int32_t tcp_force_detection = 0; SYSCTL_INT(_net_inet_tcp_sack_attack, OID_AUTO, force_detection, CTLFLAG_RW, &tcp_force_detection, 0, "Do we force detection even if the INP has it off?"); int32_t tcp_sack_to_ack_thresh = 700; /* 70 % */ SYSCTL_INT(_net_inet_tcp_sack_attack, OID_AUTO, sack_to_ack_thresh, CTLFLAG_RW, &tcp_sack_to_ack_thresh, 700, "Percentage of sacks to acks we must see above (10.1 percent is 101)?"); int32_t tcp_sack_to_move_thresh = 600; /* 60 % */ SYSCTL_INT(_net_inet_tcp_sack_attack, OID_AUTO, move_thresh, CTLFLAG_RW, &tcp_sack_to_move_thresh, 600, "Percentage of sack moves we must see above (10.1 percent is 101)"); int32_t tcp_restoral_thresh = 650; /* 65 % (sack:2:ack -5%) */ SYSCTL_INT(_net_inet_tcp_sack_attack, OID_AUTO, restore_thresh, CTLFLAG_RW, &tcp_restoral_thresh, 550, "Percentage of sack to ack percentage we must see below to restore(10.1 percent is 101)"); int32_t tcp_sad_decay_val = 800; SYSCTL_INT(_net_inet_tcp_sack_attack, OID_AUTO, decay_per, CTLFLAG_RW, &tcp_sad_decay_val, 800, "The decay percentage (10.1 percent equals 101 )"); int32_t tcp_map_minimum = 500; SYSCTL_INT(_net_inet_tcp_sack_attack, OID_AUTO, nummaps, CTLFLAG_RW, &tcp_map_minimum, 500, "Number of Map enteries before we start detection"); int32_t tcp_attack_on_turns_on_logging = 0; SYSCTL_INT(_net_inet_tcp_sack_attack, OID_AUTO, attacks_logged, CTLFLAG_RW, &tcp_attack_on_turns_on_logging, 0, "When we have a positive hit on attack, do we turn on logging?"); int32_t tcp_sad_pacing_interval = 2000; SYSCTL_INT(_net_inet_tcp_sack_attack, OID_AUTO, sad_pacing_int, CTLFLAG_RW, &tcp_sad_pacing_interval, 2000, "What is the minimum pacing interval for a classified attacker?"); int32_t tcp_sad_low_pps = 100; SYSCTL_INT(_net_inet_tcp_sack_attack, OID_AUTO, sad_low_pps, CTLFLAG_RW, &tcp_sad_low_pps, 100, "What is the input pps that below which we do not decay?"); #endif struct rwlock tcp_function_lock; static int sysctl_net_inet_tcp_mss_check(SYSCTL_HANDLER_ARGS) { int error, new; new = V_tcp_mssdflt; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr) { if (new < TCP_MINMSS) error = EINVAL; else V_tcp_mssdflt = new; } return (error); } SYSCTL_PROC(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(tcp_mssdflt), 0, &sysctl_net_inet_tcp_mss_check, "I", "Default TCP Maximum Segment Size"); #ifdef INET6 static int sysctl_net_inet_tcp_mss_v6_check(SYSCTL_HANDLER_ARGS) { int error, new; new = V_tcp_v6mssdflt; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr) { if (new < TCP_MINMSS) error = EINVAL; else V_tcp_v6mssdflt = new; } return (error); } SYSCTL_PROC(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(tcp_v6mssdflt), 0, &sysctl_net_inet_tcp_mss_v6_check, "I", "Default TCP Maximum Segment Size for IPv6"); #endif /* INET6 */ /* * Minimum MSS we accept and use. This prevents DoS attacks where * we are forced to a ridiculous low MSS like 20 and send hundreds * of packets instead of one. The effect scales with the available * bandwidth and quickly saturates the CPU and network interface * with packet generation and sending. Set to zero to disable MINMSS * checking. This setting prevents us from sending too small packets. */ VNET_DEFINE(int, tcp_minmss) = TCP_MINMSS; SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_minmss), 0, "Minimum TCP Maximum Segment Size"); VNET_DEFINE(int, tcp_do_rfc1323) = 1; SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_rfc1323), 0, "Enable rfc1323 (high performance TCP) extensions"); VNET_DEFINE(int, tcp_tolerate_missing_ts) = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, tolerate_missing_ts, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_tolerate_missing_ts), 0, "Tolerate missing TCP timestamps"); VNET_DEFINE(int, tcp_ts_offset_per_conn) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, ts_offset_per_conn, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_ts_offset_per_conn), 0, "Initialize TCP timestamps per connection instead of per host pair"); static int tcp_log_debug = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, log_debug, CTLFLAG_RW, &tcp_log_debug, 0, "Log errors caused by incoming TCP segments"); static int tcp_tcbhashsize; SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &tcp_tcbhashsize, 0, "Size of TCP control-block hashtable"); static int do_tcpdrain = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0, "Enable tcp_drain routine for extra help when low on mbufs"); SYSCTL_UINT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_VNET | CTLFLAG_RD, &VNET_NAME(tcbinfo.ipi_count), 0, "Number of active PCBs"); VNET_DEFINE_STATIC(int, icmp_may_rst) = 1; #define V_icmp_may_rst VNET(icmp_may_rst) SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(icmp_may_rst), 0, "Certain ICMP unreachable messages may abort connections in SYN_SENT"); VNET_DEFINE_STATIC(int, tcp_isn_reseed_interval) = 0; #define V_tcp_isn_reseed_interval VNET(tcp_isn_reseed_interval) SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_isn_reseed_interval), 0, "Seconds between reseeding of ISN secret"); static int tcp_soreceive_stream; SYSCTL_INT(_net_inet_tcp, OID_AUTO, soreceive_stream, CTLFLAG_RDTUN, &tcp_soreceive_stream, 0, "Using soreceive_stream for TCP sockets"); VNET_DEFINE(uma_zone_t, sack_hole_zone); #define V_sack_hole_zone VNET(sack_hole_zone) VNET_DEFINE(uint32_t, tcp_map_entries_limit) = 0; /* unlimited */ static int sysctl_net_inet_tcp_map_limit_check(SYSCTL_HANDLER_ARGS) { int error; uint32_t new; new = V_tcp_map_entries_limit; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr) { /* only allow "0" and value > minimum */ if (new > 0 && new < TCP_MIN_MAP_ENTRIES_LIMIT) error = EINVAL; else V_tcp_map_entries_limit = new; } return (error); } SYSCTL_PROC(_net_inet_tcp, OID_AUTO, map_limit, CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(tcp_map_entries_limit), 0, &sysctl_net_inet_tcp_map_limit_check, "IU", "Total sendmap entries limit"); VNET_DEFINE(uint32_t, tcp_map_split_limit) = 0; /* unlimited */ SYSCTL_UINT(_net_inet_tcp, OID_AUTO, split_limit, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_map_split_limit), 0, "Total sendmap split entries limit"); #ifdef TCP_HHOOK VNET_DEFINE(struct hhook_head *, tcp_hhh[HHOOK_TCP_LAST+1]); #endif #define TS_OFFSET_SECRET_LENGTH SIPHASH_KEY_LENGTH VNET_DEFINE_STATIC(u_char, ts_offset_secret[TS_OFFSET_SECRET_LENGTH]); #define V_ts_offset_secret VNET(ts_offset_secret) static int tcp_default_fb_init(struct tcpcb *tp); static void tcp_default_fb_fini(struct tcpcb *tp, int tcb_is_purged); static int tcp_default_handoff_ok(struct tcpcb *tp); static struct inpcb *tcp_notify(struct inpcb *, int); static struct inpcb *tcp_mtudisc_notify(struct inpcb *, int); static void tcp_mtudisc(struct inpcb *, int); static char * tcp_log_addr(struct in_conninfo *inc, struct tcphdr *th, void *ip4hdr, const void *ip6hdr); static struct tcp_function_block tcp_def_funcblk = { .tfb_tcp_block_name = "freebsd", .tfb_tcp_output = tcp_output, .tfb_tcp_do_segment = tcp_do_segment, .tfb_tcp_ctloutput = tcp_default_ctloutput, .tfb_tcp_handoff_ok = tcp_default_handoff_ok, .tfb_tcp_fb_init = tcp_default_fb_init, .tfb_tcp_fb_fini = tcp_default_fb_fini, }; static int tcp_fb_cnt = 0; struct tcp_funchead t_functions; static struct tcp_function_block *tcp_func_set_ptr = &tcp_def_funcblk; static struct tcp_function_block * find_tcp_functions_locked(struct tcp_function_set *fs) { struct tcp_function *f; struct tcp_function_block *blk=NULL; TAILQ_FOREACH(f, &t_functions, tf_next) { if (strcmp(f->tf_name, fs->function_set_name) == 0) { blk = f->tf_fb; break; } } return(blk); } static struct tcp_function_block * find_tcp_fb_locked(struct tcp_function_block *blk, struct tcp_function **s) { struct tcp_function_block *rblk=NULL; struct tcp_function *f; TAILQ_FOREACH(f, &t_functions, tf_next) { if (f->tf_fb == blk) { rblk = blk; if (s) { *s = f; } break; } } return (rblk); } struct tcp_function_block * find_and_ref_tcp_functions(struct tcp_function_set *fs) { struct tcp_function_block *blk; rw_rlock(&tcp_function_lock); blk = find_tcp_functions_locked(fs); if (blk) refcount_acquire(&blk->tfb_refcnt); rw_runlock(&tcp_function_lock); return(blk); } struct tcp_function_block * find_and_ref_tcp_fb(struct tcp_function_block *blk) { struct tcp_function_block *rblk; rw_rlock(&tcp_function_lock); rblk = find_tcp_fb_locked(blk, NULL); if (rblk) refcount_acquire(&rblk->tfb_refcnt); rw_runlock(&tcp_function_lock); return(rblk); } static struct tcp_function_block * find_and_ref_tcp_default_fb(void) { struct tcp_function_block *rblk; rw_rlock(&tcp_function_lock); rblk = tcp_func_set_ptr; refcount_acquire(&rblk->tfb_refcnt); rw_runlock(&tcp_function_lock); return (rblk); } void tcp_switch_back_to_default(struct tcpcb *tp) { struct tcp_function_block *tfb; KASSERT(tp->t_fb != &tcp_def_funcblk, ("%s: called by the built-in default stack", __func__)); /* * Release the old stack. This function will either find a new one * or panic. */ if (tp->t_fb->tfb_tcp_fb_fini != NULL) (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); refcount_release(&tp->t_fb->tfb_refcnt); /* * Now, we'll find a new function block to use. * Start by trying the current user-selected * default, unless this stack is the user-selected * default. */ tfb = find_and_ref_tcp_default_fb(); if (tfb == tp->t_fb) { refcount_release(&tfb->tfb_refcnt); tfb = NULL; } /* Does the stack accept this connection? */ if (tfb != NULL && tfb->tfb_tcp_handoff_ok != NULL && (*tfb->tfb_tcp_handoff_ok)(tp)) { refcount_release(&tfb->tfb_refcnt); tfb = NULL; } /* Try to use that stack. */ if (tfb != NULL) { /* Initialize the new stack. If it succeeds, we are done. */ tp->t_fb = tfb; if (tp->t_fb->tfb_tcp_fb_init == NULL || (*tp->t_fb->tfb_tcp_fb_init)(tp) == 0) return; /* * Initialization failed. Release the reference count on * the stack. */ refcount_release(&tfb->tfb_refcnt); } /* * If that wasn't feasible, use the built-in default * stack which is not allowed to reject anyone. */ tfb = find_and_ref_tcp_fb(&tcp_def_funcblk); if (tfb == NULL) { /* there always should be a default */ panic("Can't refer to tcp_def_funcblk"); } if (tfb->tfb_tcp_handoff_ok != NULL) { if ((*tfb->tfb_tcp_handoff_ok) (tp)) { /* The default stack cannot say no */ panic("Default stack rejects a new session?"); } } tp->t_fb = tfb; if (tp->t_fb->tfb_tcp_fb_init != NULL && (*tp->t_fb->tfb_tcp_fb_init)(tp)) { /* The default stack cannot fail */ panic("Default stack initialization failed"); } } +static void +tcp_recv_udp_tunneled_packet(struct mbuf *m, int off, struct inpcb *inp, + const struct sockaddr *sa, void *ctx) +{ + struct ip *iph; +#ifdef INET6 + struct ip6_hdr *ip6; +#endif + struct udphdr *uh; + struct tcphdr *th; + int thlen; + uint16_t port; + + TCPSTAT_INC(tcps_tunneled_pkts); + if ((m->m_flags & M_PKTHDR) == 0) { + /* Can't handle one that is not a pkt hdr */ + TCPSTAT_INC(tcps_tunneled_errs); + goto out; + } + thlen = sizeof(struct tcphdr); + if (m->m_len < off + sizeof(struct udphdr) + thlen && + (m = m_pullup(m, off + sizeof(struct udphdr) + thlen)) == NULL) { + TCPSTAT_INC(tcps_tunneled_errs); + goto out; + } + iph = mtod(m, struct ip *); + uh = (struct udphdr *)((caddr_t)iph + off); + th = (struct tcphdr *)(uh + 1); + thlen = th->th_off << 2; + if (m->m_len < off + sizeof(struct udphdr) + thlen) { + m = m_pullup(m, off + sizeof(struct udphdr) + thlen); + if (m == NULL) { + TCPSTAT_INC(tcps_tunneled_errs); + goto out; + } else { + iph = mtod(m, struct ip *); + uh = (struct udphdr *)((caddr_t)iph + off); + th = (struct tcphdr *)(uh + 1); + } + } + m->m_pkthdr.tcp_tun_port = port = uh->uh_sport; + bcopy(th, uh, m->m_len - off); + m->m_len -= sizeof(struct udphdr); + m->m_pkthdr.len -= sizeof(struct udphdr); + /* + * We use the same algorithm for + * both UDP and TCP for c-sum. So + * the code in tcp_input will skip + * the checksum. So we do nothing + * with the flag (m->m_pkthdr.csum_flags). + */ + switch (iph->ip_v) { +#ifdef INET + case IPVERSION: + iph->ip_len = htons(ntohs(iph->ip_len) - sizeof(struct udphdr)); + tcp_input_with_port(&m, &off, IPPROTO_TCP, port); + break; +#endif +#ifdef INET6 + case IPV6_VERSION >> 4: + ip6 = mtod(m, struct ip6_hdr *); + ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) - sizeof(struct udphdr)); + tcp6_input_with_port(&m, &off, IPPROTO_TCP, port); + break; +#endif + default: + goto out; + break; + } + return; +out: + m_freem(m); +} + static int sysctl_net_inet_default_tcp_functions(SYSCTL_HANDLER_ARGS) { int error=ENOENT; struct tcp_function_set fs; struct tcp_function_block *blk; memset(&fs, 0, sizeof(fs)); rw_rlock(&tcp_function_lock); blk = find_tcp_fb_locked(tcp_func_set_ptr, NULL); if (blk) { /* Found him */ strcpy(fs.function_set_name, blk->tfb_tcp_block_name); fs.pcbcnt = blk->tfb_refcnt; } rw_runlock(&tcp_function_lock); error = sysctl_handle_string(oidp, fs.function_set_name, sizeof(fs.function_set_name), req); /* Check for error or no change */ if (error != 0 || req->newptr == NULL) return(error); rw_wlock(&tcp_function_lock); blk = find_tcp_functions_locked(&fs); if ((blk == NULL) || (blk->tfb_flags & TCP_FUNC_BEING_REMOVED)) { error = ENOENT; goto done; } tcp_func_set_ptr = blk; done: rw_wunlock(&tcp_function_lock); return (error); } SYSCTL_PROC(_net_inet_tcp, OID_AUTO, functions_default, CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_NEEDGIANT, NULL, 0, sysctl_net_inet_default_tcp_functions, "A", "Set/get the default TCP functions"); static int sysctl_net_inet_list_available(SYSCTL_HANDLER_ARGS) { int error, cnt, linesz; struct tcp_function *f; char *buffer, *cp; size_t bufsz, outsz; bool alias; cnt = 0; rw_rlock(&tcp_function_lock); TAILQ_FOREACH(f, &t_functions, tf_next) { cnt++; } rw_runlock(&tcp_function_lock); bufsz = (cnt+2) * ((TCP_FUNCTION_NAME_LEN_MAX * 2) + 13) + 1; buffer = malloc(bufsz, M_TEMP, M_WAITOK); error = 0; cp = buffer; linesz = snprintf(cp, bufsz, "\n%-32s%c %-32s %s\n", "Stack", 'D', "Alias", "PCB count"); cp += linesz; bufsz -= linesz; outsz = linesz; rw_rlock(&tcp_function_lock); TAILQ_FOREACH(f, &t_functions, tf_next) { alias = (f->tf_name != f->tf_fb->tfb_tcp_block_name); linesz = snprintf(cp, bufsz, "%-32s%c %-32s %u\n", f->tf_fb->tfb_tcp_block_name, (f->tf_fb == tcp_func_set_ptr) ? '*' : ' ', alias ? f->tf_name : "-", f->tf_fb->tfb_refcnt); if (linesz >= bufsz) { error = EOVERFLOW; break; } cp += linesz; bufsz -= linesz; outsz += linesz; } rw_runlock(&tcp_function_lock); if (error == 0) error = sysctl_handle_string(oidp, buffer, outsz + 1, req); free(buffer, M_TEMP); return (error); } SYSCTL_PROC(_net_inet_tcp, OID_AUTO, functions_available, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, NULL, 0, sysctl_net_inet_list_available, "A", "list available TCP Function sets"); +VNET_DEFINE(int, tcp_udp_tunneling_port) = TCP_TUNNELING_PORT_DEFAULT; + +#ifdef INET +VNET_DEFINE(struct socket *, udp4_tun_socket) = NULL; +#define V_udp4_tun_socket VNET(udp4_tun_socket) +#endif +#ifdef INET6 +VNET_DEFINE(struct socket *, udp6_tun_socket) = NULL; +#define V_udp6_tun_socket VNET(udp6_tun_socket) +#endif + +static void +tcp_over_udp_stop(void) +{ + /* + * This function assumes sysctl caller holds inp_rinfo_lock() + * for writting! + */ +#ifdef INET + if (V_udp4_tun_socket != NULL) { + soclose(V_udp4_tun_socket); + V_udp4_tun_socket = NULL; + } +#endif +#ifdef INET6 + if (V_udp6_tun_socket != NULL) { + soclose(V_udp6_tun_socket); + V_udp6_tun_socket = NULL; + } +#endif +} + +static int +tcp_over_udp_start(void) +{ + uint16_t port; + int ret; +#ifdef INET + struct sockaddr_in sin; +#endif +#ifdef INET6 + struct sockaddr_in6 sin6; +#endif + /* + * This function assumes sysctl caller holds inp_info_rlock() + * for writting! + */ + port = V_tcp_udp_tunneling_port; + if (ntohs(port) == 0) { + /* Must have a port set */ + return (EINVAL); + } +#ifdef INET + if (V_udp4_tun_socket != NULL) { + /* Already running -- must stop first */ + return (EALREADY); + } +#endif +#ifdef INET6 + if (V_udp6_tun_socket != NULL) { + /* Already running -- must stop first */ + return (EALREADY); + } +#endif +#ifdef INET + if ((ret = socreate(PF_INET, &V_udp4_tun_socket, + SOCK_DGRAM, IPPROTO_UDP, + curthread->td_ucred, curthread))) { + tcp_over_udp_stop(); + return (ret); + } + /* Call the special UDP hook. */ + if ((ret = udp_set_kernel_tunneling(V_udp4_tun_socket, + tcp_recv_udp_tunneled_packet, + tcp_ctlinput_viaudp, + NULL))) { + tcp_over_udp_stop(); + return (ret); + } + /* Ok, we have a socket, bind it to the port. */ + memset(&sin, 0, sizeof(struct sockaddr_in)); + sin.sin_len = sizeof(struct sockaddr_in); + sin.sin_family = AF_INET; + sin.sin_port = htons(port); + if ((ret = sobind(V_udp4_tun_socket, + (struct sockaddr *)&sin, curthread))) { + tcp_over_udp_stop(); + return (ret); + } +#endif +#ifdef INET6 + if ((ret = socreate(PF_INET6, &V_udp6_tun_socket, + SOCK_DGRAM, IPPROTO_UDP, + curthread->td_ucred, curthread))) { + tcp_over_udp_stop(); + return (ret); + } + /* Call the special UDP hook. */ + if ((ret = udp_set_kernel_tunneling(V_udp6_tun_socket, + tcp_recv_udp_tunneled_packet, + tcp6_ctlinput_viaudp, + NULL))) { + tcp_over_udp_stop(); + return (ret); + } + /* Ok, we have a socket, bind it to the port. */ + memset(&sin6, 0, sizeof(struct sockaddr_in6)); + sin6.sin6_len = sizeof(struct sockaddr_in6); + sin6.sin6_family = AF_INET6; + sin6.sin6_port = htons(port); + if ((ret = sobind(V_udp6_tun_socket, + (struct sockaddr *)&sin6, curthread))) { + tcp_over_udp_stop(); + return (ret); + } +#endif + return (0); +} + +static int +sysctl_net_inet_tcp_udp_tunneling_port_check(SYSCTL_HANDLER_ARGS) +{ + int error; + uint32_t old, new; + + old = V_tcp_udp_tunneling_port; + new = old; + error = sysctl_handle_int(oidp, &new, 0, req); + if ((error == 0) && + (req->newptr != NULL)) { + if ((new < TCP_TUNNELING_PORT_MIN) || + (new > TCP_TUNNELING_PORT_MAX)) { + error = EINVAL; + } else { + V_tcp_udp_tunneling_port = new; + if (old != 0) { + tcp_over_udp_stop(); + } + if (new != 0) { + error = tcp_over_udp_start(); + } + } + } + return (error); +} + +SYSCTL_PROC(_net_inet_tcp, OID_AUTO, udp_tunneling_port, + CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, + &VNET_NAME(tcp_udp_tunneling_port), + 0, &sysctl_net_inet_tcp_udp_tunneling_port_check, "IU", + "Tunneling port for tcp over udp"); + +VNET_DEFINE(int, tcp_udp_tunneling_overhead) = TCP_TUNNELING_OVERHEAD_DEFAULT; + +static int +sysctl_net_inet_tcp_udp_tunneling_overhead_check(SYSCTL_HANDLER_ARGS) +{ + int error, new; + + new = V_tcp_udp_tunneling_overhead; + error = sysctl_handle_int(oidp, &new, 0, req); + if (error == 0 && req->newptr) { + if ((new < TCP_TUNNELING_OVERHEAD_MIN) || + (new > TCP_TUNNELING_OVERHEAD_MAX)) + error = EINVAL; + else + V_tcp_udp_tunneling_overhead = new; + } + return (error); +} + +SYSCTL_PROC(_net_inet_tcp, OID_AUTO, udp_tunneling_overhead, + CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, + &VNET_NAME(tcp_udp_tunneling_overhead), + 0, &sysctl_net_inet_tcp_udp_tunneling_overhead_check, "IU", + "MSS reduction when using tcp over udp"); + /* * Exports one (struct tcp_function_info) for each alias/name. */ static int sysctl_net_inet_list_func_info(SYSCTL_HANDLER_ARGS) { int cnt, error; struct tcp_function *f; struct tcp_function_info tfi; /* * We don't allow writes. */ if (req->newptr != NULL) return (EINVAL); /* * Wire the old buffer so we can directly copy the functions to * user space without dropping the lock. */ if (req->oldptr != NULL) { error = sysctl_wire_old_buffer(req, 0); if (error) return (error); } /* * Walk the list and copy out matching entries. If INVARIANTS * is compiled in, also walk the list to verify the length of * the list matches what we have recorded. */ rw_rlock(&tcp_function_lock); cnt = 0; #ifndef INVARIANTS if (req->oldptr == NULL) { cnt = tcp_fb_cnt; goto skip_loop; } #endif TAILQ_FOREACH(f, &t_functions, tf_next) { #ifdef INVARIANTS cnt++; #endif if (req->oldptr != NULL) { bzero(&tfi, sizeof(tfi)); tfi.tfi_refcnt = f->tf_fb->tfb_refcnt; tfi.tfi_id = f->tf_fb->tfb_id; (void)strlcpy(tfi.tfi_alias, f->tf_name, sizeof(tfi.tfi_alias)); (void)strlcpy(tfi.tfi_name, f->tf_fb->tfb_tcp_block_name, sizeof(tfi.tfi_name)); error = SYSCTL_OUT(req, &tfi, sizeof(tfi)); /* * Don't stop on error, as that is the * mechanism we use to accumulate length * information if the buffer was too short. */ } } KASSERT(cnt == tcp_fb_cnt, ("%s: cnt (%d) != tcp_fb_cnt (%d)", __func__, cnt, tcp_fb_cnt)); #ifndef INVARIANTS skip_loop: #endif rw_runlock(&tcp_function_lock); if (req->oldptr == NULL) error = SYSCTL_OUT(req, NULL, (cnt + 1) * sizeof(struct tcp_function_info)); return (error); } SYSCTL_PROC(_net_inet_tcp, OID_AUTO, function_info, CTLTYPE_OPAQUE | CTLFLAG_SKIP | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_net_inet_list_func_info, "S,tcp_function_info", "List TCP function block name-to-ID mappings"); /* * tfb_tcp_handoff_ok() function for the default stack. * Note that we'll basically try to take all comers. */ static int tcp_default_handoff_ok(struct tcpcb *tp) { return (0); } /* * tfb_tcp_fb_init() function for the default stack. * * This handles making sure we have appropriate timers set if you are * transitioning a socket that has some amount of setup done. * * The init() fuction from the default can *never* return non-zero i.e. * it is required to always succeed since it is the stack of last resort! */ static int tcp_default_fb_init(struct tcpcb *tp) { struct socket *so; INP_WLOCK_ASSERT(tp->t_inpcb); KASSERT(tp->t_state >= 0 && tp->t_state < TCPS_TIME_WAIT, ("%s: connection %p in unexpected state %d", __func__, tp, tp->t_state)); /* * Nothing to do for ESTABLISHED or LISTEN states. And, we don't * know what to do for unexpected states (which includes TIME_WAIT). */ if (tp->t_state <= TCPS_LISTEN || tp->t_state >= TCPS_TIME_WAIT) return (0); /* * Make sure some kind of transmission timer is set if there is * outstanding data. */ so = tp->t_inpcb->inp_socket; if ((!TCPS_HAVEESTABLISHED(tp->t_state) || sbavail(&so->so_snd) || tp->snd_una != tp->snd_max) && !(tcp_timer_active(tp, TT_REXMT) || tcp_timer_active(tp, TT_PERSIST))) { /* * If the session has established and it looks like it should * be in the persist state, set the persist timer. Otherwise, * set the retransmit timer. */ if (TCPS_HAVEESTABLISHED(tp->t_state) && tp->snd_wnd == 0 && (int32_t)(tp->snd_nxt - tp->snd_una) < (int32_t)sbavail(&so->so_snd)) tcp_setpersist(tp); else tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); } /* All non-embryonic sessions get a keepalive timer. */ if (!tcp_timer_active(tp, TT_KEEP)) tcp_timer_activate(tp, TT_KEEP, TCPS_HAVEESTABLISHED(tp->t_state) ? TP_KEEPIDLE(tp) : TP_KEEPINIT(tp)); /* * Make sure critical variables are initialized * if transitioning while in Recovery. */ if IN_FASTRECOVERY(tp->t_flags) { if (tp->sackhint.recover_fs == 0) tp->sackhint.recover_fs = max(1, tp->snd_nxt - tp->snd_una); } return (0); } /* * tfb_tcp_fb_fini() function for the default stack. * * This changes state as necessary (or prudent) to prepare for another stack * to assume responsibility for the connection. */ static void tcp_default_fb_fini(struct tcpcb *tp, int tcb_is_purged) { INP_WLOCK_ASSERT(tp->t_inpcb); return; } /* * Target size of TCP PCB hash tables. Must be a power of two. * * Note that this can be overridden by the kernel environment * variable net.inet.tcp.tcbhashsize */ #ifndef TCBHASHSIZE #define TCBHASHSIZE 0 #endif /* * XXX * Callouts should be moved into struct tcp directly. They are currently * separate because the tcpcb structure is exported to userland for sysctl * parsing purposes, which do not know about callouts. */ struct tcpcb_mem { struct tcpcb tcb; struct tcp_timer tt; struct cc_var ccv; #ifdef TCP_HHOOK struct osd osd; #endif }; VNET_DEFINE_STATIC(uma_zone_t, tcpcb_zone); #define V_tcpcb_zone VNET(tcpcb_zone) MALLOC_DEFINE(M_TCPLOG, "tcplog", "TCP address and flags print buffers"); MALLOC_DEFINE(M_TCPFUNCTIONS, "tcpfunc", "TCP function set memory"); static struct mtx isn_mtx; #define ISN_LOCK_INIT() mtx_init(&isn_mtx, "isn_mtx", NULL, MTX_DEF) #define ISN_LOCK() mtx_lock(&isn_mtx) #define ISN_UNLOCK() mtx_unlock(&isn_mtx) /* * TCP initialization. */ static void tcp_zone_change(void *tag) { uma_zone_set_max(V_tcbinfo.ipi_zone, maxsockets); uma_zone_set_max(V_tcpcb_zone, maxsockets); tcp_tw_zone_change(); } static int tcp_inpcb_init(void *mem, int size, int flags) { struct inpcb *inp = mem; INP_LOCK_INIT(inp, "inp", "tcpinp"); return (0); } /* * Take a value and get the next power of 2 that doesn't overflow. * Used to size the tcp_inpcb hash buckets. */ static int maketcp_hashsize(int size) { int hashsize; /* * auto tune. * get the next power of 2 higher than maxsockets. */ hashsize = 1 << fls(size); /* catch overflow, and just go one power of 2 smaller */ if (hashsize < size) { hashsize = 1 << (fls(size) - 1); } return (hashsize); } static volatile int next_tcp_stack_id = 1; /* * Register a TCP function block with the name provided in the names * array. (Note that this function does NOT automatically register * blk->tfb_tcp_block_name as a stack name. Therefore, you should * explicitly include blk->tfb_tcp_block_name in the list of names if * you wish to register the stack with that name.) * * Either all name registrations will succeed or all will fail. If * a name registration fails, the function will update the num_names * argument to point to the array index of the name that encountered * the failure. * * Returns 0 on success, or an error code on failure. */ int register_tcp_functions_as_names(struct tcp_function_block *blk, int wait, const char *names[], int *num_names) { struct tcp_function *n; struct tcp_function_set fs; int error, i; KASSERT(names != NULL && *num_names > 0, ("%s: Called with 0-length name list", __func__)); KASSERT(names != NULL, ("%s: Called with NULL name list", __func__)); KASSERT(rw_initialized(&tcp_function_lock), ("%s: called too early", __func__)); if ((blk->tfb_tcp_output == NULL) || (blk->tfb_tcp_do_segment == NULL) || (blk->tfb_tcp_ctloutput == NULL) || (strlen(blk->tfb_tcp_block_name) == 0)) { /* * These functions are required and you * need a name. */ *num_names = 0; return (EINVAL); } if (blk->tfb_tcp_timer_stop_all || blk->tfb_tcp_timer_activate || blk->tfb_tcp_timer_active || blk->tfb_tcp_timer_stop) { /* * If you define one timer function you * must have them all. */ if ((blk->tfb_tcp_timer_stop_all == NULL) || (blk->tfb_tcp_timer_activate == NULL) || (blk->tfb_tcp_timer_active == NULL) || (blk->tfb_tcp_timer_stop == NULL)) { *num_names = 0; return (EINVAL); } } if (blk->tfb_flags & TCP_FUNC_BEING_REMOVED) { *num_names = 0; return (EINVAL); } refcount_init(&blk->tfb_refcnt, 0); blk->tfb_id = atomic_fetchadd_int(&next_tcp_stack_id, 1); for (i = 0; i < *num_names; i++) { n = malloc(sizeof(struct tcp_function), M_TCPFUNCTIONS, wait); if (n == NULL) { error = ENOMEM; goto cleanup; } n->tf_fb = blk; (void)strlcpy(fs.function_set_name, names[i], sizeof(fs.function_set_name)); rw_wlock(&tcp_function_lock); if (find_tcp_functions_locked(&fs) != NULL) { /* Duplicate name space not allowed */ rw_wunlock(&tcp_function_lock); free(n, M_TCPFUNCTIONS); error = EALREADY; goto cleanup; } (void)strlcpy(n->tf_name, names[i], sizeof(n->tf_name)); TAILQ_INSERT_TAIL(&t_functions, n, tf_next); tcp_fb_cnt++; rw_wunlock(&tcp_function_lock); } return(0); cleanup: /* * Deregister the names we just added. Because registration failed * for names[i], we don't need to deregister that name. */ *num_names = i; rw_wlock(&tcp_function_lock); while (--i >= 0) { TAILQ_FOREACH(n, &t_functions, tf_next) { if (!strncmp(n->tf_name, names[i], TCP_FUNCTION_NAME_LEN_MAX)) { TAILQ_REMOVE(&t_functions, n, tf_next); tcp_fb_cnt--; n->tf_fb = NULL; free(n, M_TCPFUNCTIONS); break; } } } rw_wunlock(&tcp_function_lock); return (error); } /* * Register a TCP function block using the name provided in the name * argument. * * Returns 0 on success, or an error code on failure. */ int register_tcp_functions_as_name(struct tcp_function_block *blk, const char *name, int wait) { const char *name_list[1]; int num_names, rv; num_names = 1; if (name != NULL) name_list[0] = name; else name_list[0] = blk->tfb_tcp_block_name; rv = register_tcp_functions_as_names(blk, wait, name_list, &num_names); return (rv); } /* * Register a TCP function block using the name defined in * blk->tfb_tcp_block_name. * * Returns 0 on success, or an error code on failure. */ int register_tcp_functions(struct tcp_function_block *blk, int wait) { return (register_tcp_functions_as_name(blk, NULL, wait)); } /* * Deregister all names associated with a function block. This * functionally removes the function block from use within the system. * * When called with a true quiesce argument, mark the function block * as being removed so no more stacks will use it and determine * whether the removal would succeed. * * When called with a false quiesce argument, actually attempt the * removal. * * When called with a force argument, attempt to switch all TCBs to * use the default stack instead of returning EBUSY. * * Returns 0 on success (or if the removal would succeed, or an error * code on failure. */ int deregister_tcp_functions(struct tcp_function_block *blk, bool quiesce, bool force) { struct tcp_function *f; if (blk == &tcp_def_funcblk) { /* You can't un-register the default */ return (EPERM); } rw_wlock(&tcp_function_lock); if (blk == tcp_func_set_ptr) { /* You can't free the current default */ rw_wunlock(&tcp_function_lock); return (EBUSY); } /* Mark the block so no more stacks can use it. */ blk->tfb_flags |= TCP_FUNC_BEING_REMOVED; /* * If TCBs are still attached to the stack, attempt to switch them * to the default stack. */ if (force && blk->tfb_refcnt) { struct inpcb *inp; struct tcpcb *tp; VNET_ITERATOR_DECL(vnet_iter); rw_wunlock(&tcp_function_lock); VNET_LIST_RLOCK(); VNET_FOREACH(vnet_iter) { CURVNET_SET(vnet_iter); INP_INFO_WLOCK(&V_tcbinfo); CK_LIST_FOREACH(inp, V_tcbinfo.ipi_listhead, inp_list) { INP_WLOCK(inp); if (inp->inp_flags & INP_TIMEWAIT) { INP_WUNLOCK(inp); continue; } tp = intotcpcb(inp); if (tp == NULL || tp->t_fb != blk) { INP_WUNLOCK(inp); continue; } tcp_switch_back_to_default(tp); INP_WUNLOCK(inp); } INP_INFO_WUNLOCK(&V_tcbinfo); CURVNET_RESTORE(); } VNET_LIST_RUNLOCK(); rw_wlock(&tcp_function_lock); } if (blk->tfb_refcnt) { /* TCBs still attached. */ rw_wunlock(&tcp_function_lock); return (EBUSY); } if (quiesce) { /* Skip removal. */ rw_wunlock(&tcp_function_lock); return (0); } /* Remove any function names that map to this function block. */ while (find_tcp_fb_locked(blk, &f) != NULL) { TAILQ_REMOVE(&t_functions, f, tf_next); tcp_fb_cnt--; f->tf_fb = NULL; free(f, M_TCPFUNCTIONS); } rw_wunlock(&tcp_function_lock); return (0); } void tcp_init(void) { const char *tcbhash_tuneable; int hashsize; tcbhash_tuneable = "net.inet.tcp.tcbhashsize"; #ifdef TCP_HHOOK if (hhook_head_register(HHOOK_TYPE_TCP, HHOOK_TCP_EST_IN, &V_tcp_hhh[HHOOK_TCP_EST_IN], HHOOK_NOWAIT|HHOOK_HEADISINVNET) != 0) printf("%s: WARNING: unable to register helper hook\n", __func__); if (hhook_head_register(HHOOK_TYPE_TCP, HHOOK_TCP_EST_OUT, &V_tcp_hhh[HHOOK_TCP_EST_OUT], HHOOK_NOWAIT|HHOOK_HEADISINVNET) != 0) printf("%s: WARNING: unable to register helper hook\n", __func__); #endif #ifdef STATS if (tcp_stats_init()) printf("%s: WARNING: unable to initialise TCP stats\n", __func__); #endif hashsize = TCBHASHSIZE; TUNABLE_INT_FETCH(tcbhash_tuneable, &hashsize); if (hashsize == 0) { /* * Auto tune the hash size based on maxsockets. * A perfect hash would have a 1:1 mapping * (hashsize = maxsockets) however it's been * suggested that O(2) average is better. */ hashsize = maketcp_hashsize(maxsockets / 4); /* * Our historical default is 512, * do not autotune lower than this. */ if (hashsize < 512) hashsize = 512; if (bootverbose && IS_DEFAULT_VNET(curvnet)) printf("%s: %s auto tuned to %d\n", __func__, tcbhash_tuneable, hashsize); } /* * We require a hashsize to be a power of two. * Previously if it was not a power of two we would just reset it * back to 512, which could be a nasty surprise if you did not notice * the error message. * Instead what we do is clip it to the closest power of two lower * than the specified hash value. */ if (!powerof2(hashsize)) { int oldhashsize = hashsize; hashsize = maketcp_hashsize(hashsize); /* prevent absurdly low value */ if (hashsize < 16) hashsize = 16; printf("%s: WARNING: TCB hash size not a power of 2, " "clipped from %d to %d.\n", __func__, oldhashsize, hashsize); } in_pcbinfo_init(&V_tcbinfo, "tcp", &V_tcb, hashsize, hashsize, "tcp_inpcb", tcp_inpcb_init, IPI_HASHFIELDS_4TUPLE); /* * These have to be type stable for the benefit of the timers. */ V_tcpcb_zone = uma_zcreate("tcpcb", sizeof(struct tcpcb_mem), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); uma_zone_set_max(V_tcpcb_zone, maxsockets); uma_zone_set_warning(V_tcpcb_zone, "kern.ipc.maxsockets limit reached"); tcp_tw_init(); syncache_init(); tcp_hc_init(); TUNABLE_INT_FETCH("net.inet.tcp.sack.enable", &V_tcp_do_sack); V_sack_hole_zone = uma_zcreate("sackhole", sizeof(struct sackhole), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); tcp_fastopen_init(); /* Skip initialization of globals for non-default instances. */ if (!IS_DEFAULT_VNET(curvnet)) return; tcp_reass_global_init(); /* XXX virtualize those bellow? */ tcp_delacktime = TCPTV_DELACK; tcp_keepinit = TCPTV_KEEP_INIT; tcp_keepidle = TCPTV_KEEP_IDLE; tcp_keepintvl = TCPTV_KEEPINTVL; tcp_maxpersistidle = TCPTV_KEEP_IDLE; tcp_msl = TCPTV_MSL; tcp_rexmit_initial = TCPTV_RTOBASE; if (tcp_rexmit_initial < 1) tcp_rexmit_initial = 1; tcp_rexmit_min = TCPTV_MIN; if (tcp_rexmit_min < 1) tcp_rexmit_min = 1; tcp_persmin = TCPTV_PERSMIN; tcp_persmax = TCPTV_PERSMAX; tcp_rexmit_slop = TCPTV_CPU_VAR; tcp_finwait2_timeout = TCPTV_FINWAIT2_TIMEOUT; tcp_tcbhashsize = hashsize; /* Setup the tcp function block list */ TAILQ_INIT(&t_functions); rw_init(&tcp_function_lock, "tcp_func_lock"); register_tcp_functions(&tcp_def_funcblk, M_WAITOK); #ifdef TCP_BLACKBOX /* Initialize the TCP logging data. */ tcp_log_init(); #endif arc4rand(&V_ts_offset_secret, sizeof(V_ts_offset_secret), 0); if (tcp_soreceive_stream) { #ifdef INET tcp_usrreqs.pru_soreceive = soreceive_stream; #endif #ifdef INET6 tcp6_usrreqs.pru_soreceive = soreceive_stream; #endif /* INET6 */ } #ifdef INET6 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) #else /* INET6 */ #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) #endif /* INET6 */ if (max_protohdr < TCP_MINPROTOHDR) max_protohdr = TCP_MINPROTOHDR; if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) panic("tcp_init"); #undef TCP_MINPROTOHDR ISN_LOCK_INIT(); EVENTHANDLER_REGISTER(shutdown_pre_sync, tcp_fini, NULL, SHUTDOWN_PRI_DEFAULT); EVENTHANDLER_REGISTER(maxsockets_change, tcp_zone_change, NULL, EVENTHANDLER_PRI_ANY); tcp_inp_lro_direct_queue = counter_u64_alloc(M_WAITOK); tcp_inp_lro_wokeup_queue = counter_u64_alloc(M_WAITOK); tcp_inp_lro_compressed = counter_u64_alloc(M_WAITOK); tcp_inp_lro_single_push = counter_u64_alloc(M_WAITOK); tcp_inp_lro_locks_taken = counter_u64_alloc(M_WAITOK); tcp_inp_lro_sack_wake = counter_u64_alloc(M_WAITOK); tcp_extra_mbuf = counter_u64_alloc(M_WAITOK); tcp_would_have_but = counter_u64_alloc(M_WAITOK); tcp_comp_total = counter_u64_alloc(M_WAITOK); tcp_uncomp_total = counter_u64_alloc(M_WAITOK); tcp_csum_hardware = counter_u64_alloc(M_WAITOK); tcp_csum_hardware_w_ph = counter_u64_alloc(M_WAITOK); tcp_csum_software = counter_u64_alloc(M_WAITOK); #ifdef TCPPCAP tcp_pcap_init(); #endif } #ifdef VIMAGE static void tcp_destroy(void *unused __unused) { int n; #ifdef TCP_HHOOK int error; #endif /* * All our processes are gone, all our sockets should be cleaned * up, which means, we should be past the tcp_discardcb() calls. * Sleep to let all tcpcb timers really disappear and cleanup. */ for (;;) { INP_LIST_RLOCK(&V_tcbinfo); n = V_tcbinfo.ipi_count; INP_LIST_RUNLOCK(&V_tcbinfo); if (n == 0) break; pause("tcpdes", hz / 10); } tcp_hc_destroy(); syncache_destroy(); tcp_tw_destroy(); in_pcbinfo_destroy(&V_tcbinfo); /* tcp_discardcb() clears the sack_holes up. */ uma_zdestroy(V_sack_hole_zone); uma_zdestroy(V_tcpcb_zone); /* * Cannot free the zone until all tcpcbs are released as we attach * the allocations to them. */ tcp_fastopen_destroy(); #ifdef TCP_HHOOK error = hhook_head_deregister(V_tcp_hhh[HHOOK_TCP_EST_IN]); if (error != 0) { printf("%s: WARNING: unable to deregister helper hook " "type=%d, id=%d: error %d returned\n", __func__, HHOOK_TYPE_TCP, HHOOK_TCP_EST_IN, error); } error = hhook_head_deregister(V_tcp_hhh[HHOOK_TCP_EST_OUT]); if (error != 0) { printf("%s: WARNING: unable to deregister helper hook " "type=%d, id=%d: error %d returned\n", __func__, HHOOK_TYPE_TCP, HHOOK_TCP_EST_OUT, error); } #endif } VNET_SYSUNINIT(tcp, SI_SUB_PROTO_DOMAIN, SI_ORDER_FOURTH, tcp_destroy, NULL); #endif void tcp_fini(void *xtp) { } /* * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. * tcp_template used to store this data in mbufs, but we now recopy it out * of the tcpcb each time to conserve mbufs. */ void -tcpip_fillheaders(struct inpcb *inp, void *ip_ptr, void *tcp_ptr) +tcpip_fillheaders(struct inpcb *inp, uint16_t port, void *ip_ptr, void *tcp_ptr) { struct tcphdr *th = (struct tcphdr *)tcp_ptr; INP_WLOCK_ASSERT(inp); #ifdef INET6 if ((inp->inp_vflag & INP_IPV6) != 0) { struct ip6_hdr *ip6; ip6 = (struct ip6_hdr *)ip_ptr; ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | (inp->inp_flow & IPV6_FLOWINFO_MASK); ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | (IPV6_VERSION & IPV6_VERSION_MASK); - ip6->ip6_nxt = IPPROTO_TCP; + if (port == 0) + ip6->ip6_nxt = IPPROTO_TCP; + else + ip6->ip6_nxt = IPPROTO_UDP; ip6->ip6_plen = htons(sizeof(struct tcphdr)); ip6->ip6_src = inp->in6p_laddr; ip6->ip6_dst = inp->in6p_faddr; } #endif /* INET6 */ #if defined(INET6) && defined(INET) else #endif #ifdef INET { struct ip *ip; ip = (struct ip *)ip_ptr; ip->ip_v = IPVERSION; ip->ip_hl = 5; ip->ip_tos = inp->inp_ip_tos; ip->ip_len = 0; ip->ip_id = 0; ip->ip_off = 0; ip->ip_ttl = inp->inp_ip_ttl; ip->ip_sum = 0; - ip->ip_p = IPPROTO_TCP; + if (port == 0) + ip->ip_p = IPPROTO_TCP; + else + ip->ip_p = IPPROTO_UDP; ip->ip_src = inp->inp_laddr; ip->ip_dst = inp->inp_faddr; } #endif /* INET */ th->th_sport = inp->inp_lport; th->th_dport = inp->inp_fport; th->th_seq = 0; th->th_ack = 0; th->th_x2 = 0; th->th_off = 5; th->th_flags = 0; th->th_win = 0; th->th_urp = 0; th->th_sum = 0; /* in_pseudo() is called later for ipv4 */ } /* * Create template to be used to send tcp packets on a connection. * Allocates an mbuf and fills in a skeletal tcp/ip header. The only * use for this function is in keepalives, which use tcp_respond. */ struct tcptemp * tcpip_maketemplate(struct inpcb *inp) { struct tcptemp *t; t = malloc(sizeof(*t), M_TEMP, M_NOWAIT); if (t == NULL) return (NULL); - tcpip_fillheaders(inp, (void *)&t->tt_ipgen, (void *)&t->tt_t); + tcpip_fillheaders(inp, 0, (void *)&t->tt_ipgen, (void *)&t->tt_t); return (t); } /* * Send a single message to the TCP at address specified by * the given TCP/IP header. If m == NULL, then we make a copy * of the tcpiphdr at th and send directly to the addressed host. * This is used to force keep alive messages out using the TCP * template for a connection. If flags are given then we send * a message back to the TCP which originated the segment th, * and discard the mbuf containing it and any other attached mbufs. * * In any case the ack and sequence number of the transmitted * segment are as specified by the parameters. * * NOTE: If m != NULL, then th must point to *inside* the mbuf. */ void tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, tcp_seq ack, tcp_seq seq, int flags) { struct tcpopt to; struct inpcb *inp; struct ip *ip; struct mbuf *optm; + struct udphdr *uh = NULL; struct tcphdr *nth; u_char *optp; #ifdef INET6 struct ip6_hdr *ip6; int isipv6; #endif /* INET6 */ - int optlen, tlen, win; + int optlen, tlen, win, ulen; bool incl_opts; + uint16_t port; KASSERT(tp != NULL || m != NULL, ("tcp_respond: tp and m both NULL")); NET_EPOCH_ASSERT(); #ifdef INET6 isipv6 = ((struct ip *)ipgen)->ip_v == (IPV6_VERSION >> 4); ip6 = ipgen; #endif /* INET6 */ ip = ipgen; if (tp != NULL) { inp = tp->t_inpcb; KASSERT(inp != NULL, ("tcp control block w/o inpcb")); INP_LOCK_ASSERT(inp); } else inp = NULL; + if (m != NULL) { +#ifdef INET6 + if (isipv6 && ip6 && (ip6->ip6_nxt == IPPROTO_UDP)) + port = m->m_pkthdr.tcp_tun_port; + else +#endif + if (ip && (ip->ip_p == IPPROTO_UDP)) + port = m->m_pkthdr.tcp_tun_port; + else + port = 0; + } else + port = tp->t_port; + incl_opts = false; win = 0; if (tp != NULL) { if (!(flags & TH_RST)) { win = sbspace(&inp->inp_socket->so_rcv); if (win > TCP_MAXWIN << tp->rcv_scale) win = TCP_MAXWIN << tp->rcv_scale; } if ((tp->t_flags & TF_NOOPT) == 0) incl_opts = true; } if (m == NULL) { m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) return; m->m_data += max_linkhdr; #ifdef INET6 if (isipv6) { bcopy((caddr_t)ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); ip6 = mtod(m, struct ip6_hdr *); nth = (struct tcphdr *)(ip6 + 1); + if (port) { + /* Insert a UDP header */ + uh = (struct udphdr *)nth; + uh->uh_sport = htons(V_tcp_udp_tunneling_port); + uh->uh_dport = port; + nth = (struct tcphdr *)(uh + 1); + } } else #endif /* INET6 */ { bcopy((caddr_t)ip, mtod(m, caddr_t), sizeof(struct ip)); ip = mtod(m, struct ip *); nth = (struct tcphdr *)(ip + 1); + if (port) { + /* Insert a UDP header */ + uh = (struct udphdr *)nth; + uh->uh_sport = htons(V_tcp_udp_tunneling_port); + uh->uh_dport = port; + nth = (struct tcphdr *)(uh + 1); + } } bcopy((caddr_t)th, (caddr_t)nth, sizeof(struct tcphdr)); flags = TH_ACK; - } else if (!M_WRITABLE(m)) { + } else if ((!M_WRITABLE(m)) || (port != 0)) { struct mbuf *n; /* Can't reuse 'm', allocate a new mbuf. */ n = m_gethdr(M_NOWAIT, MT_DATA); if (n == NULL) { m_freem(m); return; } if (!m_dup_pkthdr(n, m, M_NOWAIT)) { m_freem(m); m_freem(n); return; } n->m_data += max_linkhdr; /* m_len is set later */ #define xchg(a,b,type) { type t; t=a; a=b; b=t; } #ifdef INET6 if (isipv6) { bcopy((caddr_t)ip6, mtod(n, caddr_t), sizeof(struct ip6_hdr)); ip6 = mtod(n, struct ip6_hdr *); xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); nth = (struct tcphdr *)(ip6 + 1); + if (port) { + /* Insert a UDP header */ + uh = (struct udphdr *)nth; + uh->uh_sport = htons(V_tcp_udp_tunneling_port); + uh->uh_dport = port; + nth = (struct tcphdr *)(uh + 1); + } } else #endif /* INET6 */ { bcopy((caddr_t)ip, mtod(n, caddr_t), sizeof(struct ip)); ip = mtod(n, struct ip *); xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, uint32_t); nth = (struct tcphdr *)(ip + 1); + if (port) { + /* Insert a UDP header */ + uh = (struct udphdr *)nth; + uh->uh_sport = htons(V_tcp_udp_tunneling_port); + uh->uh_dport = port; + nth = (struct tcphdr *)(uh + 1); + } } bcopy((caddr_t)th, (caddr_t)nth, sizeof(struct tcphdr)); xchg(nth->th_dport, nth->th_sport, uint16_t); th = nth; m_freem(m); m = n; } else { /* * reuse the mbuf. * XXX MRT We inherit the FIB, which is lucky. */ m_freem(m->m_next); m->m_next = NULL; m->m_data = (caddr_t)ipgen; /* m_len is set later */ #ifdef INET6 if (isipv6) { xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); nth = (struct tcphdr *)(ip6 + 1); } else #endif /* INET6 */ { xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, uint32_t); nth = (struct tcphdr *)(ip + 1); } if (th != nth) { /* * this is usually a case when an extension header * exists between the IPv6 header and the * TCP header. */ nth->th_sport = th->th_sport; nth->th_dport = th->th_dport; } xchg(nth->th_dport, nth->th_sport, uint16_t); #undef xchg } tlen = 0; #ifdef INET6 if (isipv6) tlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET tlen = sizeof (struct tcpiphdr); #endif + if (port) + tlen += sizeof (struct udphdr); #ifdef INVARIANTS m->m_len = 0; KASSERT(M_TRAILINGSPACE(m) >= tlen, ("Not enough trailing space for message (m=%p, need=%d, have=%ld)", m, tlen, (long)M_TRAILINGSPACE(m))); #endif m->m_len = tlen; to.to_flags = 0; if (incl_opts) { /* Make sure we have room. */ if (M_TRAILINGSPACE(m) < TCP_MAXOLEN) { m->m_next = m_get(M_NOWAIT, MT_DATA); if (m->m_next) { optp = mtod(m->m_next, u_char *); optm = m->m_next; } else incl_opts = false; } else { optp = (u_char *) (nth + 1); optm = m; } } if (incl_opts) { /* Timestamps. */ if (tp->t_flags & TF_RCVD_TSTMP) { to.to_tsval = tcp_ts_getticks() + tp->ts_offset; to.to_tsecr = tp->ts_recent; to.to_flags |= TOF_TS; } #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) /* TCP-MD5 (RFC2385). */ if (tp->t_flags & TF_SIGNATURE) to.to_flags |= TOF_SIGNATURE; #endif /* Add the options. */ tlen += optlen = tcp_addoptions(&to, optp); /* Update m_len in the correct mbuf. */ optm->m_len += optlen; } else optlen = 0; #ifdef INET6 if (isipv6) { + if (uh) { + ulen = tlen - sizeof(struct ip6_hdr); + uh->uh_ulen = htons(ulen); + } ip6->ip6_flow = 0; ip6->ip6_vfc = IPV6_VERSION; - ip6->ip6_nxt = IPPROTO_TCP; + if (port) + ip6->ip6_nxt = IPPROTO_UDP; + else + ip6->ip6_nxt = IPPROTO_TCP; ip6->ip6_plen = htons(tlen - sizeof(*ip6)); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { + if (uh) { + ulen = tlen - sizeof(struct ip); + uh->uh_ulen = htons(ulen); + } ip->ip_len = htons(tlen); ip->ip_ttl = V_ip_defttl; + if (port) { + ip->ip_p = IPPROTO_UDP; + } else { + ip->ip_p = IPPROTO_TCP; + } if (V_path_mtu_discovery) ip->ip_off |= htons(IP_DF); } #endif m->m_pkthdr.len = tlen; m->m_pkthdr.rcvif = NULL; #ifdef MAC if (inp != NULL) { /* * Packet is associated with a socket, so allow the * label of the response to reflect the socket label. */ INP_LOCK_ASSERT(inp); mac_inpcb_create_mbuf(inp, m); } else { /* * Packet is not associated with a socket, so possibly * update the label in place. */ mac_netinet_tcp_reply(m); } #endif nth->th_seq = htonl(seq); nth->th_ack = htonl(ack); nth->th_x2 = 0; nth->th_off = (sizeof (struct tcphdr) + optlen) >> 2; nth->th_flags = flags; if (tp != NULL) nth->th_win = htons((u_short) (win >> tp->rcv_scale)); else nth->th_win = htons((u_short)win); nth->th_urp = 0; #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (to.to_flags & TOF_SIGNATURE) { if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, nth, to.to_signature) != 0) { m_freem(m); return; } } #endif - m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); #ifdef INET6 if (isipv6) { - m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; - nth->th_sum = in6_cksum_pseudo(ip6, - tlen - sizeof(struct ip6_hdr), IPPROTO_TCP, 0); + if (port) { + m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; + m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); + uh->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); + nth->th_sum = 0; + } else { + m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; + m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); + nth->th_sum = in6_cksum_pseudo(ip6, + tlen - sizeof(struct ip6_hdr), IPPROTO_TCP, 0); + } ip6->ip6_hlim = in6_selecthlim(tp != NULL ? tp->t_inpcb : NULL, NULL); } #endif /* INET6 */ #if defined(INET6) && defined(INET) else #endif #ifdef INET { - m->m_pkthdr.csum_flags = CSUM_TCP; - nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, - htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); + if (port) { + uh->uh_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, + htons(ulen + IPPROTO_UDP)); + m->m_pkthdr.csum_flags = CSUM_UDP; + m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); + nth->th_sum = 0; + } else { + m->m_pkthdr.csum_flags = CSUM_TCP; + m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); + nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, + htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); + } } #endif /* INET */ #ifdef TCPDEBUG if (tp == NULL || (inp->inp_socket->so_options & SO_DEBUG)) tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); #endif TCP_PROBE3(debug__output, tp, th, m); if (flags & TH_RST) TCP_PROBE5(accept__refused, NULL, NULL, m, tp, nth); #ifdef INET6 if (isipv6) { TCP_PROBE5(send, NULL, tp, ip6, tp, nth); (void)ip6_output(m, NULL, NULL, 0, NULL, NULL, inp); } #endif /* INET6 */ #if defined(INET) && defined(INET6) else #endif #ifdef INET { TCP_PROBE5(send, NULL, tp, ip, tp, nth); (void)ip_output(m, NULL, NULL, 0, NULL, inp); } #endif } /* * Create a new TCP control block, making an * empty reassembly queue and hooking it to the argument * protocol control block. The `inp' parameter must have * come from the zone allocator set up in tcp_init(). */ struct tcpcb * tcp_newtcpcb(struct inpcb *inp) { struct tcpcb_mem *tm; struct tcpcb *tp; #ifdef INET6 int isipv6 = (inp->inp_vflag & INP_IPV6) != 0; #endif /* INET6 */ tm = uma_zalloc(V_tcpcb_zone, M_NOWAIT | M_ZERO); if (tm == NULL) return (NULL); tp = &tm->tcb; /* Initialise cc_var struct for this tcpcb. */ tp->ccv = &tm->ccv; tp->ccv->type = IPPROTO_TCP; tp->ccv->ccvc.tcp = tp; rw_rlock(&tcp_function_lock); tp->t_fb = tcp_func_set_ptr; refcount_acquire(&tp->t_fb->tfb_refcnt); rw_runlock(&tcp_function_lock); /* * Use the current system default CC algorithm. */ CC_LIST_RLOCK(); KASSERT(!STAILQ_EMPTY(&cc_list), ("cc_list is empty!")); CC_ALGO(tp) = CC_DEFAULT(); CC_LIST_RUNLOCK(); /* * The tcpcb will hold a reference on its inpcb until tcp_discardcb() * is called. */ in_pcbref(inp); /* Reference for tcpcb */ tp->t_inpcb = inp; if (CC_ALGO(tp)->cb_init != NULL) if (CC_ALGO(tp)->cb_init(tp->ccv) > 0) { if (tp->t_fb->tfb_tcp_fb_fini) (*tp->t_fb->tfb_tcp_fb_fini)(tp, 1); in_pcbrele_wlocked(inp); refcount_release(&tp->t_fb->tfb_refcnt); uma_zfree(V_tcpcb_zone, tm); return (NULL); } #ifdef TCP_HHOOK tp->osd = &tm->osd; if (khelp_init_osd(HELPER_CLASS_TCP, tp->osd)) { if (tp->t_fb->tfb_tcp_fb_fini) (*tp->t_fb->tfb_tcp_fb_fini)(tp, 1); in_pcbrele_wlocked(inp); refcount_release(&tp->t_fb->tfb_refcnt); uma_zfree(V_tcpcb_zone, tm); return (NULL); } #endif #ifdef VIMAGE tp->t_vnet = inp->inp_vnet; #endif tp->t_timers = &tm->tt; TAILQ_INIT(&tp->t_segq); tp->t_maxseg = #ifdef INET6 isipv6 ? V_tcp_v6mssdflt : #endif /* INET6 */ V_tcp_mssdflt; /* Set up our timeouts. */ callout_init(&tp->t_timers->tt_rexmt, 1); callout_init(&tp->t_timers->tt_persist, 1); callout_init(&tp->t_timers->tt_keep, 1); callout_init(&tp->t_timers->tt_2msl, 1); callout_init(&tp->t_timers->tt_delack, 1); if (V_tcp_do_rfc1323) tp->t_flags = (TF_REQ_SCALE|TF_REQ_TSTMP); if (V_tcp_do_sack) tp->t_flags |= TF_SACK_PERMIT; TAILQ_INIT(&tp->snd_holes); /* * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives * reasonable initial retransmit time. */ tp->t_srtt = TCPTV_SRTTBASE; tp->t_rttvar = ((tcp_rexmit_initial - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; tp->t_rttmin = tcp_rexmit_min; tp->t_rxtcur = tcp_rexmit_initial; tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; tp->t_rcvtime = ticks; /* * IPv4 TTL initialization is necessary for an IPv6 socket as well, * because the socket may be bound to an IPv6 wildcard address, * which may match an IPv4-mapped IPv6 address. */ inp->inp_ip_ttl = V_ip_defttl; inp->inp_ppcb = tp; #ifdef TCPPCAP /* * Init the TCP PCAP queues. */ tcp_pcap_tcpcb_init(tp); #endif #ifdef TCP_BLACKBOX /* Initialize the per-TCPCB log data. */ tcp_log_tcpcbinit(tp); #endif tp->t_pacing_rate = -1; if (tp->t_fb->tfb_tcp_fb_init) { if ((*tp->t_fb->tfb_tcp_fb_init)(tp)) { refcount_release(&tp->t_fb->tfb_refcnt); in_pcbrele_wlocked(inp); uma_zfree(V_tcpcb_zone, tm); return (NULL); } } #ifdef STATS if (V_tcp_perconn_stats_enable == 1) tp->t_stats = stats_blob_alloc(V_tcp_perconn_stats_dflt_tpl, 0); #endif return (tp); /* XXX */ } /* * Switch the congestion control algorithm back to NewReno for any active * control blocks using an algorithm which is about to go away. * This ensures the CC framework can allow the unload to proceed without leaving * any dangling pointers which would trigger a panic. * Returning non-zero would inform the CC framework that something went wrong * and it would be unsafe to allow the unload to proceed. However, there is no * way for this to occur with this implementation so we always return zero. */ int tcp_ccalgounload(struct cc_algo *unload_algo) { struct cc_algo *tmpalgo; struct inpcb *inp; struct tcpcb *tp; VNET_ITERATOR_DECL(vnet_iter); /* * Check all active control blocks across all network stacks and change * any that are using "unload_algo" back to NewReno. If "unload_algo" * requires cleanup code to be run, call it. */ VNET_LIST_RLOCK(); VNET_FOREACH(vnet_iter) { CURVNET_SET(vnet_iter); INP_INFO_WLOCK(&V_tcbinfo); /* * New connections already part way through being initialised * with the CC algo we're removing will not race with this code * because the INP_INFO_WLOCK is held during initialisation. We * therefore don't enter the loop below until the connection * list has stabilised. */ CK_LIST_FOREACH(inp, &V_tcb, inp_list) { INP_WLOCK(inp); /* Important to skip tcptw structs. */ if (!(inp->inp_flags & INP_TIMEWAIT) && (tp = intotcpcb(inp)) != NULL) { /* * By holding INP_WLOCK here, we are assured * that the connection is not currently * executing inside the CC module's functions * i.e. it is safe to make the switch back to * NewReno. */ if (CC_ALGO(tp) == unload_algo) { tmpalgo = CC_ALGO(tp); if (tmpalgo->cb_destroy != NULL) tmpalgo->cb_destroy(tp->ccv); CC_DATA(tp) = NULL; /* * NewReno may allocate memory on * demand for certain stateful * configuration as needed, but is * coded to never fail on memory * allocation failure so it is a safe * fallback. */ CC_ALGO(tp) = &newreno_cc_algo; } } INP_WUNLOCK(inp); } INP_INFO_WUNLOCK(&V_tcbinfo); CURVNET_RESTORE(); } VNET_LIST_RUNLOCK(); return (0); } /* * Drop a TCP connection, reporting * the specified error. If connection is synchronized, * then send a RST to peer. */ struct tcpcb * tcp_drop(struct tcpcb *tp, int errno) { struct socket *so = tp->t_inpcb->inp_socket; NET_EPOCH_ASSERT(); INP_INFO_LOCK_ASSERT(&V_tcbinfo); INP_WLOCK_ASSERT(tp->t_inpcb); if (TCPS_HAVERCVDSYN(tp->t_state)) { tcp_state_change(tp, TCPS_CLOSED); (void) tp->t_fb->tfb_tcp_output(tp); TCPSTAT_INC(tcps_drops); } else TCPSTAT_INC(tcps_conndrops); if (errno == ETIMEDOUT && tp->t_softerror) errno = tp->t_softerror; so->so_error = errno; return (tcp_close(tp)); } void tcp_discardcb(struct tcpcb *tp) { struct inpcb *inp = tp->t_inpcb; struct socket *so = inp->inp_socket; #ifdef INET6 int isipv6 = (inp->inp_vflag & INP_IPV6) != 0; #endif /* INET6 */ int released __unused; INP_WLOCK_ASSERT(inp); /* * Make sure that all of our timers are stopped before we delete the * PCB. * * If stopping a timer fails, we schedule a discard function in same * callout, and the last discard function called will take care of * deleting the tcpcb. */ tp->t_timers->tt_draincnt = 0; tcp_timer_stop(tp, TT_REXMT); tcp_timer_stop(tp, TT_PERSIST); tcp_timer_stop(tp, TT_KEEP); tcp_timer_stop(tp, TT_2MSL); tcp_timer_stop(tp, TT_DELACK); if (tp->t_fb->tfb_tcp_timer_stop_all) { /* * Call the stop-all function of the methods, * this function should call the tcp_timer_stop() * method with each of the function specific timeouts. * That stop will be called via the tfb_tcp_timer_stop() * which should use the async drain function of the * callout system (see tcp_var.h). */ tp->t_fb->tfb_tcp_timer_stop_all(tp); } /* * If we got enough samples through the srtt filter, * save the rtt and rttvar in the routing entry. * 'Enough' is arbitrarily defined as 4 rtt samples. * 4 samples is enough for the srtt filter to converge * to within enough % of the correct value; fewer samples * and we could save a bogus rtt. The danger is not high * as tcp quickly recovers from everything. * XXX: Works very well but needs some more statistics! */ if (tp->t_rttupdated >= 4) { struct hc_metrics_lite metrics; uint32_t ssthresh; bzero(&metrics, sizeof(metrics)); /* * Update the ssthresh always when the conditions below * are satisfied. This gives us better new start value * for the congestion avoidance for new connections. * ssthresh is only set if packet loss occurred on a session. * * XXXRW: 'so' may be NULL here, and/or socket buffer may be * being torn down. Ideally this code would not use 'so'. */ ssthresh = tp->snd_ssthresh; if (ssthresh != 0 && ssthresh < so->so_snd.sb_hiwat / 2) { /* * convert the limit from user data bytes to * packets then to packet data bytes. */ ssthresh = (ssthresh + tp->t_maxseg / 2) / tp->t_maxseg; if (ssthresh < 2) ssthresh = 2; ssthresh *= (tp->t_maxseg + #ifdef INET6 (isipv6 ? sizeof (struct ip6_hdr) + sizeof (struct tcphdr) : #endif sizeof (struct tcpiphdr) #ifdef INET6 ) #endif ); } else ssthresh = 0; metrics.rmx_ssthresh = ssthresh; metrics.rmx_rtt = tp->t_srtt; metrics.rmx_rttvar = tp->t_rttvar; metrics.rmx_cwnd = tp->snd_cwnd; metrics.rmx_sendpipe = 0; metrics.rmx_recvpipe = 0; tcp_hc_update(&inp->inp_inc, &metrics); } /* free the reassembly queue, if any */ tcp_reass_flush(tp); #ifdef TCP_OFFLOAD /* Disconnect offload device, if any. */ if (tp->t_flags & TF_TOE) tcp_offload_detach(tp); #endif tcp_free_sackholes(tp); #ifdef TCPPCAP /* Free the TCP PCAP queues. */ tcp_pcap_drain(&(tp->t_inpkts)); tcp_pcap_drain(&(tp->t_outpkts)); #endif /* Allow the CC algorithm to clean up after itself. */ if (CC_ALGO(tp)->cb_destroy != NULL) CC_ALGO(tp)->cb_destroy(tp->ccv); CC_DATA(tp) = NULL; #ifdef TCP_HHOOK khelp_destroy_osd(tp->osd); #endif #ifdef STATS stats_blob_destroy(tp->t_stats); #endif CC_ALGO(tp) = NULL; inp->inp_ppcb = NULL; if (tp->t_timers->tt_draincnt == 0) { /* We own the last reference on tcpcb, let's free it. */ #ifdef TCP_BLACKBOX tcp_log_tcpcbfini(tp); #endif TCPSTATES_DEC(tp->t_state); if (tp->t_fb->tfb_tcp_fb_fini) (*tp->t_fb->tfb_tcp_fb_fini)(tp, 1); refcount_release(&tp->t_fb->tfb_refcnt); tp->t_inpcb = NULL; uma_zfree(V_tcpcb_zone, tp); released = in_pcbrele_wlocked(inp); KASSERT(!released, ("%s: inp %p should not have been released " "here", __func__, inp)); } } void tcp_timer_discard(void *ptp) { struct inpcb *inp; struct tcpcb *tp; struct epoch_tracker et; tp = (struct tcpcb *)ptp; CURVNET_SET(tp->t_vnet); NET_EPOCH_ENTER(et); inp = tp->t_inpcb; KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); INP_WLOCK(inp); KASSERT((tp->t_timers->tt_flags & TT_STOPPED) != 0, ("%s: tcpcb has to be stopped here", __func__)); tp->t_timers->tt_draincnt--; if (tp->t_timers->tt_draincnt == 0) { /* We own the last reference on this tcpcb, let's free it. */ #ifdef TCP_BLACKBOX tcp_log_tcpcbfini(tp); #endif TCPSTATES_DEC(tp->t_state); if (tp->t_fb->tfb_tcp_fb_fini) (*tp->t_fb->tfb_tcp_fb_fini)(tp, 1); refcount_release(&tp->t_fb->tfb_refcnt); tp->t_inpcb = NULL; uma_zfree(V_tcpcb_zone, tp); if (in_pcbrele_wlocked(inp)) { NET_EPOCH_EXIT(et); CURVNET_RESTORE(); return; } } INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); CURVNET_RESTORE(); } /* * Attempt to close a TCP control block, marking it as dropped, and freeing * the socket if we hold the only reference. */ struct tcpcb * tcp_close(struct tcpcb *tp) { struct inpcb *inp = tp->t_inpcb; struct socket *so; INP_INFO_LOCK_ASSERT(&V_tcbinfo); INP_WLOCK_ASSERT(inp); #ifdef TCP_OFFLOAD if (tp->t_state == TCPS_LISTEN) tcp_offload_listen_stop(tp); #endif /* * This releases the TFO pending counter resource for TFO listen * sockets as well as passively-created TFO sockets that transition * from SYN_RECEIVED to CLOSED. */ if (tp->t_tfo_pending) { tcp_fastopen_decrement_counter(tp->t_tfo_pending); tp->t_tfo_pending = NULL; } in_pcbdrop(inp); TCPSTAT_INC(tcps_closed); if (tp->t_state != TCPS_CLOSED) tcp_state_change(tp, TCPS_CLOSED); KASSERT(inp->inp_socket != NULL, ("tcp_close: inp_socket NULL")); so = inp->inp_socket; soisdisconnected(so); if (inp->inp_flags & INP_SOCKREF) { KASSERT(so->so_state & SS_PROTOREF, ("tcp_close: !SS_PROTOREF")); inp->inp_flags &= ~INP_SOCKREF; INP_WUNLOCK(inp); SOCK_LOCK(so); so->so_state &= ~SS_PROTOREF; sofree(so); return (NULL); } return (tp); } void tcp_drain(void) { VNET_ITERATOR_DECL(vnet_iter); if (!do_tcpdrain) return; VNET_LIST_RLOCK_NOSLEEP(); VNET_FOREACH(vnet_iter) { CURVNET_SET(vnet_iter); struct inpcb *inpb; struct tcpcb *tcpb; /* * Walk the tcpbs, if existing, and flush the reassembly queue, * if there is one... * XXX: The "Net/3" implementation doesn't imply that the TCP * reassembly queue should be flushed, but in a situation * where we're really low on mbufs, this is potentially * useful. */ INP_INFO_WLOCK(&V_tcbinfo); CK_LIST_FOREACH(inpb, V_tcbinfo.ipi_listhead, inp_list) { INP_WLOCK(inpb); if (inpb->inp_flags & INP_TIMEWAIT) { INP_WUNLOCK(inpb); continue; } if ((tcpb = intotcpcb(inpb)) != NULL) { tcp_reass_flush(tcpb); tcp_clean_sackreport(tcpb); #ifdef TCP_BLACKBOX tcp_log_drain(tcpb); #endif #ifdef TCPPCAP if (tcp_pcap_aggressive_free) { /* Free the TCP PCAP queues. */ tcp_pcap_drain(&(tcpb->t_inpkts)); tcp_pcap_drain(&(tcpb->t_outpkts)); } #endif } INP_WUNLOCK(inpb); } INP_INFO_WUNLOCK(&V_tcbinfo); CURVNET_RESTORE(); } VNET_LIST_RUNLOCK_NOSLEEP(); } /* * Notify a tcp user of an asynchronous error; * store error as soft error, but wake up user * (for now, won't do anything until can select for soft error). * * Do not wake up user since there currently is no mechanism for * reporting soft errors (yet - a kqueue filter may be added). */ static struct inpcb * tcp_notify(struct inpcb *inp, int error) { struct tcpcb *tp; INP_INFO_LOCK_ASSERT(&V_tcbinfo); INP_WLOCK_ASSERT(inp); if ((inp->inp_flags & INP_TIMEWAIT) || (inp->inp_flags & INP_DROPPED)) return (inp); tp = intotcpcb(inp); KASSERT(tp != NULL, ("tcp_notify: tp == NULL")); /* * Ignore some errors if we are hooked up. * If connection hasn't completed, has retransmitted several times, * and receives a second error, give up now. This is better * than waiting a long time to establish a connection that * can never complete. */ if (tp->t_state == TCPS_ESTABLISHED && (error == EHOSTUNREACH || error == ENETUNREACH || error == EHOSTDOWN)) { if (inp->inp_route.ro_nh) { NH_FREE(inp->inp_route.ro_nh); inp->inp_route.ro_nh = (struct nhop_object *)NULL; } return (inp); } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && tp->t_softerror) { tp = tcp_drop(tp, error); if (tp != NULL) return (inp); else return (NULL); } else { tp->t_softerror = error; return (inp); } #if 0 wakeup( &so->so_timeo); sorwakeup(so); sowwakeup(so); #endif } static int tcp_pcblist(SYSCTL_HANDLER_ARGS) { struct epoch_tracker et; struct inpcb *inp; struct xinpgen xig; int error; if (req->newptr != NULL) return (EPERM); if (req->oldptr == NULL) { int n; n = V_tcbinfo.ipi_count + counter_u64_fetch(V_tcps_states[TCPS_SYN_RECEIVED]); n += imax(n / 8, 10); req->oldidx = 2 * (sizeof xig) + n * sizeof(struct xtcpcb); return (0); } if ((error = sysctl_wire_old_buffer(req, 0)) != 0) return (error); bzero(&xig, sizeof(xig)); xig.xig_len = sizeof xig; xig.xig_count = V_tcbinfo.ipi_count + counter_u64_fetch(V_tcps_states[TCPS_SYN_RECEIVED]); xig.xig_gen = V_tcbinfo.ipi_gencnt; xig.xig_sogen = so_gencnt; error = SYSCTL_OUT(req, &xig, sizeof xig); if (error) return (error); error = syncache_pcblist(req); if (error) return (error); NET_EPOCH_ENTER(et); for (inp = CK_LIST_FIRST(V_tcbinfo.ipi_listhead); inp != NULL; inp = CK_LIST_NEXT(inp, inp_list)) { INP_RLOCK(inp); if (inp->inp_gencnt <= xig.xig_gen) { int crerr; /* * XXX: This use of cr_cansee(), introduced with * TCP state changes, is not quite right, but for * now, better than nothing. */ if (inp->inp_flags & INP_TIMEWAIT) { if (intotw(inp) != NULL) crerr = cr_cansee(req->td->td_ucred, intotw(inp)->tw_cred); else crerr = EINVAL; /* Skip this inp. */ } else crerr = cr_canseeinpcb(req->td->td_ucred, inp); if (crerr == 0) { struct xtcpcb xt; tcp_inptoxtp(inp, &xt); INP_RUNLOCK(inp); error = SYSCTL_OUT(req, &xt, sizeof xt); if (error) break; else continue; } } INP_RUNLOCK(inp); } NET_EPOCH_EXIT(et); if (!error) { /* * Give the user an updated idea of our state. * If the generation differs from what we told * her before, she knows that something happened * while we were processing this request, and it * might be necessary to retry. */ xig.xig_gen = V_tcbinfo.ipi_gencnt; xig.xig_sogen = so_gencnt; xig.xig_count = V_tcbinfo.ipi_count + counter_u64_fetch(V_tcps_states[TCPS_SYN_RECEIVED]); error = SYSCTL_OUT(req, &xig, sizeof xig); } return (error); } SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_NEEDGIANT, NULL, 0, tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); #ifdef INET static int tcp_getcred(SYSCTL_HANDLER_ARGS) { struct xucred xuc; struct sockaddr_in addrs[2]; struct epoch_tracker et; struct inpcb *inp; int error; error = priv_check(req->td, PRIV_NETINET_GETCRED); if (error) return (error); error = SYSCTL_IN(req, addrs, sizeof(addrs)); if (error) return (error); NET_EPOCH_ENTER(et); inp = in_pcblookup(&V_tcbinfo, addrs[1].sin_addr, addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, INPLOOKUP_RLOCKPCB, NULL); NET_EPOCH_EXIT(et); if (inp != NULL) { if (inp->inp_socket == NULL) error = ENOENT; if (error == 0) error = cr_canseeinpcb(req->td->td_ucred, inp); if (error == 0) cru2x(inp->inp_cred, &xuc); INP_RUNLOCK(inp); } else error = ENOENT; if (error == 0) error = SYSCTL_OUT(req, &xuc, sizeof(struct xucred)); return (error); } SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, CTLTYPE_OPAQUE | CTLFLAG_RW | CTLFLAG_PRISON | CTLFLAG_NEEDGIANT, 0, 0, tcp_getcred, "S,xucred", "Get the xucred of a TCP connection"); #endif /* INET */ #ifdef INET6 static int tcp6_getcred(SYSCTL_HANDLER_ARGS) { struct epoch_tracker et; struct xucred xuc; struct sockaddr_in6 addrs[2]; struct inpcb *inp; int error; #ifdef INET int mapped = 0; #endif error = priv_check(req->td, PRIV_NETINET_GETCRED); if (error) return (error); error = SYSCTL_IN(req, addrs, sizeof(addrs)); if (error) return (error); if ((error = sa6_embedscope(&addrs[0], V_ip6_use_defzone)) != 0 || (error = sa6_embedscope(&addrs[1], V_ip6_use_defzone)) != 0) { return (error); } if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { #ifdef INET if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) mapped = 1; else #endif return (EINVAL); } NET_EPOCH_ENTER(et); #ifdef INET if (mapped == 1) inp = in_pcblookup(&V_tcbinfo, *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], addrs[1].sin6_port, *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], addrs[0].sin6_port, INPLOOKUP_RLOCKPCB, NULL); else #endif inp = in6_pcblookup(&V_tcbinfo, &addrs[1].sin6_addr, addrs[1].sin6_port, &addrs[0].sin6_addr, addrs[0].sin6_port, INPLOOKUP_RLOCKPCB, NULL); NET_EPOCH_EXIT(et); if (inp != NULL) { if (inp->inp_socket == NULL) error = ENOENT; if (error == 0) error = cr_canseeinpcb(req->td->td_ucred, inp); if (error == 0) cru2x(inp->inp_cred, &xuc); INP_RUNLOCK(inp); } else error = ENOENT; if (error == 0) error = SYSCTL_OUT(req, &xuc, sizeof(struct xucred)); return (error); } SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, CTLTYPE_OPAQUE | CTLFLAG_RW | CTLFLAG_PRISON | CTLFLAG_NEEDGIANT, 0, 0, tcp6_getcred, "S,xucred", "Get the xucred of a TCP6 connection"); #endif /* INET6 */ #ifdef INET -void -tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip) +static void +tcp_ctlinput_with_port(int cmd, struct sockaddr *sa, void *vip, uint16_t port) { struct ip *ip = vip; struct tcphdr *th; struct in_addr faddr; struct inpcb *inp; struct tcpcb *tp; struct inpcb *(*notify)(struct inpcb *, int) = tcp_notify; struct icmp *icp; struct in_conninfo inc; tcp_seq icmp_tcp_seq; int mtu; faddr = ((struct sockaddr_in *)sa)->sin_addr; if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) return; if (cmd == PRC_MSGSIZE) notify = tcp_mtudisc_notify; else if (V_icmp_may_rst && (cmd == PRC_UNREACH_ADMIN_PROHIB || cmd == PRC_UNREACH_PORT || cmd == PRC_UNREACH_PROTOCOL || cmd == PRC_TIMXCEED_INTRANS) && ip) notify = tcp_drop_syn_sent; /* * Hostdead is ugly because it goes linearly through all PCBs. * XXX: We never get this from ICMP, otherwise it makes an * excellent DoS attack on machines with many connections. */ else if (cmd == PRC_HOSTDEAD) ip = NULL; else if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) return; if (ip == NULL) { in_pcbnotifyall(&V_tcbinfo, faddr, inetctlerrmap[cmd], notify); return; } icp = (struct icmp *)((caddr_t)ip - offsetof(struct icmp, icmp_ip)); th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2)); inp = in_pcblookup(&V_tcbinfo, faddr, th->th_dport, ip->ip_src, th->th_sport, INPLOOKUP_WLOCKPCB, NULL); if (inp != NULL && PRC_IS_REDIRECT(cmd)) { /* signal EHOSTDOWN, as it flushes the cached route */ inp = (*notify)(inp, EHOSTDOWN); goto out; } icmp_tcp_seq = th->th_seq; if (inp != NULL) { if (!(inp->inp_flags & INP_TIMEWAIT) && !(inp->inp_flags & INP_DROPPED) && !(inp->inp_socket == NULL)) { tp = intotcpcb(inp); + if (tp->t_port != port) { + goto out; + } if (SEQ_GEQ(ntohl(icmp_tcp_seq), tp->snd_una) && SEQ_LT(ntohl(icmp_tcp_seq), tp->snd_max)) { if (cmd == PRC_MSGSIZE) { /* * MTU discovery: * If we got a needfrag set the MTU * in the route to the suggested new * value (if given) and then notify. */ mtu = ntohs(icp->icmp_nextmtu); /* * If no alternative MTU was * proposed, try the next smaller * one. */ if (!mtu) mtu = ip_next_mtu( ntohs(ip->ip_len), 1); if (mtu < V_tcp_minmss + sizeof(struct tcpiphdr)) mtu = V_tcp_minmss + sizeof(struct tcpiphdr); /* * Only process the offered MTU if it * is smaller than the current one. */ if (mtu < tp->t_maxseg + sizeof(struct tcpiphdr)) { bzero(&inc, sizeof(inc)); inc.inc_faddr = faddr; inc.inc_fibnum = inp->inp_inc.inc_fibnum; tcp_hc_updatemtu(&inc, mtu); tcp_mtudisc(inp, mtu); } } else inp = (*notify)(inp, inetctlerrmap[cmd]); } } } else { bzero(&inc, sizeof(inc)); inc.inc_fport = th->th_dport; inc.inc_lport = th->th_sport; inc.inc_faddr = faddr; inc.inc_laddr = ip->ip_src; - syncache_unreach(&inc, icmp_tcp_seq); + syncache_unreach(&inc, icmp_tcp_seq, port); } out: if (inp != NULL) INP_WUNLOCK(inp); } + +void +tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip) +{ + tcp_ctlinput_with_port(cmd, sa, vip, htons(0)); +} + +void +tcp_ctlinput_viaudp(int cmd, struct sockaddr *sa, void *vip, void *unused) +{ + /* Its a tunneled TCP over UDP icmp */ + struct ip *outer_ip, *inner_ip; + struct icmp *icmp; + struct udphdr *udp; + struct tcphdr *th, ttemp; + int i_hlen, o_len; + uint16_t port; + + inner_ip = (struct ip *)vip; + icmp = (struct icmp *)((caddr_t)inner_ip - + (sizeof(struct icmp) - sizeof(struct ip))); + outer_ip = (struct ip *)((caddr_t)icmp - sizeof(struct ip)); + i_hlen = inner_ip->ip_hl << 2; + o_len = ntohs(outer_ip->ip_len); + if (o_len < + (sizeof(struct ip) + 8 + i_hlen + sizeof(struct udphdr) + offsetof(struct tcphdr, th_ack))) { + /* Not enough data present */ + return; + } + /* Ok lets strip out the inner udphdr header by copying up on top of it the tcp hdr */ + udp = (struct udphdr *)(((caddr_t)inner_ip) + i_hlen); + if (ntohs(udp->uh_sport) != V_tcp_udp_tunneling_port) { + return; + } + port = udp->uh_dport; + th = (struct tcphdr *)(udp + 1); + memcpy(&ttemp, th, sizeof(struct tcphdr)); + memcpy(udp, &ttemp, sizeof(struct tcphdr)); + /* Now adjust down the size of the outer IP header */ + o_len -= sizeof(struct udphdr); + outer_ip->ip_len = htons(o_len); + /* Now call in to the normal handling code */ + tcp_ctlinput_with_port(cmd, sa, vip, port); +} #endif /* INET */ #ifdef INET6 -void -tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d) +static void +tcp6_ctlinput_with_port(int cmd, struct sockaddr *sa, void *d, uint16_t port) { struct in6_addr *dst; struct inpcb *(*notify)(struct inpcb *, int) = tcp_notify; struct ip6_hdr *ip6; struct mbuf *m; struct inpcb *inp; struct tcpcb *tp; struct icmp6_hdr *icmp6; struct ip6ctlparam *ip6cp = NULL; const struct sockaddr_in6 *sa6_src = NULL; struct in_conninfo inc; struct tcp_ports { uint16_t th_sport; uint16_t th_dport; } t_ports; tcp_seq icmp_tcp_seq; unsigned int mtu; unsigned int off; if (sa->sa_family != AF_INET6 || sa->sa_len != sizeof(struct sockaddr_in6)) return; /* if the parameter is from icmp6, decode it. */ if (d != NULL) { ip6cp = (struct ip6ctlparam *)d; icmp6 = ip6cp->ip6c_icmp6; m = ip6cp->ip6c_m; ip6 = ip6cp->ip6c_ip6; off = ip6cp->ip6c_off; sa6_src = ip6cp->ip6c_src; dst = ip6cp->ip6c_finaldst; } else { m = NULL; ip6 = NULL; off = 0; /* fool gcc */ sa6_src = &sa6_any; dst = NULL; } if (cmd == PRC_MSGSIZE) notify = tcp_mtudisc_notify; else if (V_icmp_may_rst && (cmd == PRC_UNREACH_ADMIN_PROHIB || cmd == PRC_UNREACH_PORT || cmd == PRC_UNREACH_PROTOCOL || cmd == PRC_TIMXCEED_INTRANS) && ip6 != NULL) notify = tcp_drop_syn_sent; /* * Hostdead is ugly because it goes linearly through all PCBs. * XXX: We never get this from ICMP, otherwise it makes an * excellent DoS attack on machines with many connections. */ else if (cmd == PRC_HOSTDEAD) ip6 = NULL; else if ((unsigned)cmd >= PRC_NCMDS || inet6ctlerrmap[cmd] == 0) return; if (ip6 == NULL) { in6_pcbnotify(&V_tcbinfo, sa, 0, (const struct sockaddr *)sa6_src, 0, cmd, NULL, notify); return; } /* Check if we can safely get the ports from the tcp hdr */ if (m == NULL || (m->m_pkthdr.len < (int32_t) (off + sizeof(struct tcp_ports)))) { return; } bzero(&t_ports, sizeof(struct tcp_ports)); m_copydata(m, off, sizeof(struct tcp_ports), (caddr_t)&t_ports); inp = in6_pcblookup(&V_tcbinfo, &ip6->ip6_dst, t_ports.th_dport, &ip6->ip6_src, t_ports.th_sport, INPLOOKUP_WLOCKPCB, NULL); if (inp != NULL && PRC_IS_REDIRECT(cmd)) { /* signal EHOSTDOWN, as it flushes the cached route */ inp = (*notify)(inp, EHOSTDOWN); goto out; } off += sizeof(struct tcp_ports); if (m->m_pkthdr.len < (int32_t) (off + sizeof(tcp_seq))) { goto out; } m_copydata(m, off, sizeof(tcp_seq), (caddr_t)&icmp_tcp_seq); if (inp != NULL) { if (!(inp->inp_flags & INP_TIMEWAIT) && !(inp->inp_flags & INP_DROPPED) && !(inp->inp_socket == NULL)) { tp = intotcpcb(inp); + if (tp->t_port != port) { + goto out; + } if (SEQ_GEQ(ntohl(icmp_tcp_seq), tp->snd_una) && SEQ_LT(ntohl(icmp_tcp_seq), tp->snd_max)) { if (cmd == PRC_MSGSIZE) { /* * MTU discovery: * If we got a needfrag set the MTU * in the route to the suggested new * value (if given) and then notify. */ mtu = ntohl(icmp6->icmp6_mtu); /* * If no alternative MTU was * proposed, or the proposed * MTU was too small, set to * the min. */ if (mtu < IPV6_MMTU) mtu = IPV6_MMTU - 8; bzero(&inc, sizeof(inc)); inc.inc_fibnum = M_GETFIB(m); inc.inc_flags |= INC_ISIPV6; inc.inc6_faddr = *dst; if (in6_setscope(&inc.inc6_faddr, m->m_pkthdr.rcvif, NULL)) goto out; /* * Only process the offered MTU if it * is smaller than the current one. */ if (mtu < tp->t_maxseg + sizeof (struct tcphdr) + sizeof (struct ip6_hdr)) { tcp_hc_updatemtu(&inc, mtu); tcp_mtudisc(inp, mtu); ICMP6STAT_INC(icp6s_pmtuchg); } } else inp = (*notify)(inp, inet6ctlerrmap[cmd]); } } } else { bzero(&inc, sizeof(inc)); inc.inc_fibnum = M_GETFIB(m); inc.inc_flags |= INC_ISIPV6; inc.inc_fport = t_ports.th_dport; inc.inc_lport = t_ports.th_sport; inc.inc6_faddr = *dst; inc.inc6_laddr = ip6->ip6_src; - syncache_unreach(&inc, icmp_tcp_seq); + syncache_unreach(&inc, icmp_tcp_seq, port); } out: if (inp != NULL) INP_WUNLOCK(inp); } + +void +tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d) +{ + tcp6_ctlinput_with_port(cmd, sa, d, htons(0)); +} + +void +tcp6_ctlinput_viaudp(int cmd, struct sockaddr *sa, void *d, void *unused) +{ + struct ip6ctlparam *ip6cp; + struct mbuf *m; + struct udphdr *udp; + uint16_t port; + + ip6cp = (struct ip6ctlparam *)d; + m = m_pulldown(ip6cp->ip6c_m, ip6cp->ip6c_off, sizeof(struct udphdr), NULL); + if (m == NULL) { + return; + } + udp = mtod(m, struct udphdr *); + if (ntohs(udp->uh_sport) != V_tcp_udp_tunneling_port) { + return; + } + port = udp->uh_dport; + m_adj(m, sizeof(struct udphdr)); + if ((m->m_flags & M_PKTHDR) == 0) { + ip6cp->ip6c_m->m_pkthdr.len -= sizeof(struct udphdr); + } + /* Now call in to the normal handling code */ + tcp6_ctlinput_with_port(cmd, sa, d, port); +} + #endif /* INET6 */ static uint32_t tcp_keyed_hash(struct in_conninfo *inc, u_char *key, u_int len) { SIPHASH_CTX ctx; uint32_t hash[2]; KASSERT(len >= SIPHASH_KEY_LENGTH, ("%s: keylen %u too short ", __func__, len)); SipHash24_Init(&ctx); SipHash_SetKey(&ctx, (uint8_t *)key); SipHash_Update(&ctx, &inc->inc_fport, sizeof(uint16_t)); SipHash_Update(&ctx, &inc->inc_lport, sizeof(uint16_t)); switch (inc->inc_flags & INC_ISIPV6) { #ifdef INET case 0: SipHash_Update(&ctx, &inc->inc_faddr, sizeof(struct in_addr)); SipHash_Update(&ctx, &inc->inc_laddr, sizeof(struct in_addr)); break; #endif #ifdef INET6 case INC_ISIPV6: SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(struct in6_addr)); SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(struct in6_addr)); break; #endif } SipHash_Final((uint8_t *)hash, &ctx); return (hash[0] ^ hash[1]); } uint32_t tcp_new_ts_offset(struct in_conninfo *inc) { struct in_conninfo inc_store, *local_inc; if (!V_tcp_ts_offset_per_conn) { memcpy(&inc_store, inc, sizeof(struct in_conninfo)); inc_store.inc_lport = 0; inc_store.inc_fport = 0; local_inc = &inc_store; } else { local_inc = inc; } return (tcp_keyed_hash(local_inc, V_ts_offset_secret, sizeof(V_ts_offset_secret))); } /* * Following is where TCP initial sequence number generation occurs. * * There are two places where we must use initial sequence numbers: * 1. In SYN-ACK packets. * 2. In SYN packets. * * All ISNs for SYN-ACK packets are generated by the syncache. See * tcp_syncache.c for details. * * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling * depends on this property. In addition, these ISNs should be * unguessable so as to prevent connection hijacking. To satisfy * the requirements of this situation, the algorithm outlined in * RFC 1948 is used, with only small modifications. * * Implementation details: * * Time is based off the system timer, and is corrected so that it * increases by one megabyte per second. This allows for proper * recycling on high speed LANs while still leaving over an hour * before rollover. * * As reading the *exact* system time is too expensive to be done * whenever setting up a TCP connection, we increment the time * offset in two ways. First, a small random positive increment * is added to isn_offset for each connection that is set up. * Second, the function tcp_isn_tick fires once per clock tick * and increments isn_offset as necessary so that sequence numbers * are incremented at approximately ISN_BYTES_PER_SECOND. The * random positive increments serve only to ensure that the same * exact sequence number is never sent out twice (as could otherwise * happen when a port is recycled in less than the system tick * interval.) * * net.inet.tcp.isn_reseed_interval controls the number of seconds * between seeding of isn_secret. This is normally set to zero, * as reseeding should not be necessary. * * Locking of the global variables isn_secret, isn_last_reseed, isn_offset, * isn_offset_old, and isn_ctx is performed using the ISN lock. In * general, this means holding an exclusive (write) lock. */ #define ISN_BYTES_PER_SECOND 1048576 #define ISN_STATIC_INCREMENT 4096 #define ISN_RANDOM_INCREMENT (4096 - 1) #define ISN_SECRET_LENGTH SIPHASH_KEY_LENGTH VNET_DEFINE_STATIC(u_char, isn_secret[ISN_SECRET_LENGTH]); VNET_DEFINE_STATIC(int, isn_last); VNET_DEFINE_STATIC(int, isn_last_reseed); VNET_DEFINE_STATIC(u_int32_t, isn_offset); VNET_DEFINE_STATIC(u_int32_t, isn_offset_old); #define V_isn_secret VNET(isn_secret) #define V_isn_last VNET(isn_last) #define V_isn_last_reseed VNET(isn_last_reseed) #define V_isn_offset VNET(isn_offset) #define V_isn_offset_old VNET(isn_offset_old) tcp_seq tcp_new_isn(struct in_conninfo *inc) { tcp_seq new_isn; u_int32_t projected_offset; ISN_LOCK(); /* Seed if this is the first use, reseed if requested. */ if ((V_isn_last_reseed == 0) || ((V_tcp_isn_reseed_interval > 0) && (((u_int)V_isn_last_reseed + (u_int)V_tcp_isn_reseed_interval*hz) < (u_int)ticks))) { arc4rand(&V_isn_secret, sizeof(V_isn_secret), 0); V_isn_last_reseed = ticks; } /* Compute the hash and return the ISN. */ new_isn = (tcp_seq)tcp_keyed_hash(inc, V_isn_secret, sizeof(V_isn_secret)); V_isn_offset += ISN_STATIC_INCREMENT + (arc4random() & ISN_RANDOM_INCREMENT); if (ticks != V_isn_last) { projected_offset = V_isn_offset_old + ISN_BYTES_PER_SECOND / hz * (ticks - V_isn_last); if (SEQ_GT(projected_offset, V_isn_offset)) V_isn_offset = projected_offset; V_isn_offset_old = V_isn_offset; V_isn_last = ticks; } new_isn += V_isn_offset; ISN_UNLOCK(); return (new_isn); } /* * When a specific ICMP unreachable message is received and the * connection state is SYN-SENT, drop the connection. This behavior * is controlled by the icmp_may_rst sysctl. */ struct inpcb * tcp_drop_syn_sent(struct inpcb *inp, int errno) { struct tcpcb *tp; NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(inp); if ((inp->inp_flags & INP_TIMEWAIT) || (inp->inp_flags & INP_DROPPED)) return (inp); tp = intotcpcb(inp); if (tp->t_state != TCPS_SYN_SENT) return (inp); if (IS_FASTOPEN(tp->t_flags)) tcp_fastopen_disable_path(tp); tp = tcp_drop(tp, errno); if (tp != NULL) return (inp); else return (NULL); } /* * When `need fragmentation' ICMP is received, update our idea of the MSS * based on the new value. Also nudge TCP to send something, since we * know the packet we just sent was dropped. * This duplicates some code in the tcp_mss() function in tcp_input.c. */ static struct inpcb * tcp_mtudisc_notify(struct inpcb *inp, int error) { tcp_mtudisc(inp, -1); return (inp); } static void tcp_mtudisc(struct inpcb *inp, int mtuoffer) { struct tcpcb *tp; struct socket *so; INP_WLOCK_ASSERT(inp); if ((inp->inp_flags & INP_TIMEWAIT) || (inp->inp_flags & INP_DROPPED)) return; tp = intotcpcb(inp); KASSERT(tp != NULL, ("tcp_mtudisc: tp == NULL")); tcp_mss_update(tp, -1, mtuoffer, NULL, NULL); so = inp->inp_socket; SOCKBUF_LOCK(&so->so_snd); /* If the mss is larger than the socket buffer, decrease the mss. */ if (so->so_snd.sb_hiwat < tp->t_maxseg) tp->t_maxseg = so->so_snd.sb_hiwat; SOCKBUF_UNLOCK(&so->so_snd); TCPSTAT_INC(tcps_mturesent); tp->t_rtttime = 0; tp->snd_nxt = tp->snd_una; tcp_free_sackholes(tp); tp->snd_recover = tp->snd_max; if (tp->t_flags & TF_SACK_PERMIT) EXIT_FASTRECOVERY(tp->t_flags); tp->t_fb->tfb_tcp_output(tp); } #ifdef INET /* * Look-up the routing entry to the peer of this inpcb. If no route * is found and it cannot be allocated, then return 0. This routine * is called by TCP routines that access the rmx structure and by * tcp_mss_update to get the peer/interface MTU. */ uint32_t tcp_maxmtu(struct in_conninfo *inc, struct tcp_ifcap *cap) { struct nhop_object *nh; struct ifnet *ifp; uint32_t maxmtu = 0; KASSERT(inc != NULL, ("tcp_maxmtu with NULL in_conninfo pointer")); if (inc->inc_faddr.s_addr != INADDR_ANY) { nh = fib4_lookup(inc->inc_fibnum, inc->inc_faddr, 0, NHR_NONE, 0); if (nh == NULL) return (0); ifp = nh->nh_ifp; maxmtu = nh->nh_mtu; /* Report additional interface capabilities. */ if (cap != NULL) { if (ifp->if_capenable & IFCAP_TSO4 && ifp->if_hwassist & CSUM_TSO) { cap->ifcap |= CSUM_TSO; cap->tsomax = ifp->if_hw_tsomax; cap->tsomaxsegcount = ifp->if_hw_tsomaxsegcount; cap->tsomaxsegsize = ifp->if_hw_tsomaxsegsize; } } } return (maxmtu); } #endif /* INET */ #ifdef INET6 uint32_t tcp_maxmtu6(struct in_conninfo *inc, struct tcp_ifcap *cap) { struct nhop_object *nh; struct in6_addr dst6; uint32_t scopeid; struct ifnet *ifp; uint32_t maxmtu = 0; KASSERT(inc != NULL, ("tcp_maxmtu6 with NULL in_conninfo pointer")); if (inc->inc_flags & INC_IPV6MINMTU) return (IPV6_MMTU); if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { in6_splitscope(&inc->inc6_faddr, &dst6, &scopeid); nh = fib6_lookup(inc->inc_fibnum, &dst6, scopeid, NHR_NONE, 0); if (nh == NULL) return (0); ifp = nh->nh_ifp; maxmtu = nh->nh_mtu; /* Report additional interface capabilities. */ if (cap != NULL) { if (ifp->if_capenable & IFCAP_TSO6 && ifp->if_hwassist & CSUM_TSO) { cap->ifcap |= CSUM_TSO; cap->tsomax = ifp->if_hw_tsomax; cap->tsomaxsegcount = ifp->if_hw_tsomaxsegcount; cap->tsomaxsegsize = ifp->if_hw_tsomaxsegsize; } } } return (maxmtu); } #endif /* INET6 */ /* * Calculate effective SMSS per RFC5681 definition for a given TCP * connection at its current state, taking into account SACK and etc. */ u_int tcp_maxseg(const struct tcpcb *tp) { u_int optlen; if (tp->t_flags & TF_NOOPT) return (tp->t_maxseg); /* * Here we have a simplified code from tcp_addoptions(), * without a proper loop, and having most of paddings hardcoded. * We might make mistakes with padding here in some edge cases, * but this is harmless, since result of tcp_maxseg() is used * only in cwnd and ssthresh estimations. */ if (TCPS_HAVEESTABLISHED(tp->t_state)) { if (tp->t_flags & TF_RCVD_TSTMP) optlen = TCPOLEN_TSTAMP_APPA; else optlen = 0; #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (tp->t_flags & TF_SIGNATURE) optlen += PADTCPOLEN(TCPOLEN_SIGNATURE); #endif if ((tp->t_flags & TF_SACK_PERMIT) && tp->rcv_numsacks > 0) { optlen += TCPOLEN_SACKHDR; optlen += tp->rcv_numsacks * TCPOLEN_SACK; optlen = PADTCPOLEN(optlen); } } else { if (tp->t_flags & TF_REQ_TSTMP) optlen = TCPOLEN_TSTAMP_APPA; else optlen = PADTCPOLEN(TCPOLEN_MAXSEG); if (tp->t_flags & TF_REQ_SCALE) optlen += PADTCPOLEN(TCPOLEN_WINDOW); #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (tp->t_flags & TF_SIGNATURE) optlen += PADTCPOLEN(TCPOLEN_SIGNATURE); #endif if (tp->t_flags & TF_SACK_PERMIT) optlen += PADTCPOLEN(TCPOLEN_SACK_PERMITTED); } #undef PAD optlen = min(optlen, TCP_MAXOLEN); return (tp->t_maxseg - optlen); } static int sysctl_drop(SYSCTL_HANDLER_ARGS) { /* addrs[0] is a foreign socket, addrs[1] is a local one. */ struct sockaddr_storage addrs[2]; struct inpcb *inp; struct tcpcb *tp; struct tcptw *tw; struct sockaddr_in *fin, *lin; struct epoch_tracker et; #ifdef INET6 struct sockaddr_in6 *fin6, *lin6; #endif int error; inp = NULL; fin = lin = NULL; #ifdef INET6 fin6 = lin6 = NULL; #endif error = 0; if (req->oldptr != NULL || req->oldlen != 0) return (EINVAL); if (req->newptr == NULL) return (EPERM); if (req->newlen < sizeof(addrs)) return (ENOMEM); error = SYSCTL_IN(req, &addrs, sizeof(addrs)); if (error) return (error); switch (addrs[0].ss_family) { #ifdef INET6 case AF_INET6: fin6 = (struct sockaddr_in6 *)&addrs[0]; lin6 = (struct sockaddr_in6 *)&addrs[1]; if (fin6->sin6_len != sizeof(struct sockaddr_in6) || lin6->sin6_len != sizeof(struct sockaddr_in6)) return (EINVAL); if (IN6_IS_ADDR_V4MAPPED(&fin6->sin6_addr)) { if (!IN6_IS_ADDR_V4MAPPED(&lin6->sin6_addr)) return (EINVAL); in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[0]); in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[1]); fin = (struct sockaddr_in *)&addrs[0]; lin = (struct sockaddr_in *)&addrs[1]; break; } error = sa6_embedscope(fin6, V_ip6_use_defzone); if (error) return (error); error = sa6_embedscope(lin6, V_ip6_use_defzone); if (error) return (error); break; #endif #ifdef INET case AF_INET: fin = (struct sockaddr_in *)&addrs[0]; lin = (struct sockaddr_in *)&addrs[1]; if (fin->sin_len != sizeof(struct sockaddr_in) || lin->sin_len != sizeof(struct sockaddr_in)) return (EINVAL); break; #endif default: return (EINVAL); } NET_EPOCH_ENTER(et); switch (addrs[0].ss_family) { #ifdef INET6 case AF_INET6: inp = in6_pcblookup(&V_tcbinfo, &fin6->sin6_addr, fin6->sin6_port, &lin6->sin6_addr, lin6->sin6_port, INPLOOKUP_WLOCKPCB, NULL); break; #endif #ifdef INET case AF_INET: inp = in_pcblookup(&V_tcbinfo, fin->sin_addr, fin->sin_port, lin->sin_addr, lin->sin_port, INPLOOKUP_WLOCKPCB, NULL); break; #endif } if (inp != NULL) { if (inp->inp_flags & INP_TIMEWAIT) { /* * XXXRW: There currently exists a state where an * inpcb is present, but its timewait state has been * discarded. For now, don't allow dropping of this * type of inpcb. */ tw = intotw(inp); if (tw != NULL) tcp_twclose(tw, 0); else INP_WUNLOCK(inp); } else if (!(inp->inp_flags & INP_DROPPED) && !(inp->inp_socket->so_options & SO_ACCEPTCONN)) { tp = intotcpcb(inp); tp = tcp_drop(tp, ECONNABORTED); if (tp != NULL) INP_WUNLOCK(inp); } else INP_WUNLOCK(inp); } else error = ESRCH; NET_EPOCH_EXIT(et); return (error); } SYSCTL_PROC(_net_inet_tcp, TCPCTL_DROP, drop, CTLFLAG_VNET | CTLTYPE_STRUCT | CTLFLAG_WR | CTLFLAG_SKIP | CTLFLAG_NEEDGIANT, NULL, 0, sysctl_drop, "", "Drop TCP connection"); #ifdef KERN_TLS static int sysctl_switch_tls(SYSCTL_HANDLER_ARGS) { /* addrs[0] is a foreign socket, addrs[1] is a local one. */ struct sockaddr_storage addrs[2]; struct inpcb *inp; struct sockaddr_in *fin, *lin; struct epoch_tracker et; #ifdef INET6 struct sockaddr_in6 *fin6, *lin6; #endif int error; inp = NULL; fin = lin = NULL; #ifdef INET6 fin6 = lin6 = NULL; #endif error = 0; if (req->oldptr != NULL || req->oldlen != 0) return (EINVAL); if (req->newptr == NULL) return (EPERM); if (req->newlen < sizeof(addrs)) return (ENOMEM); error = SYSCTL_IN(req, &addrs, sizeof(addrs)); if (error) return (error); switch (addrs[0].ss_family) { #ifdef INET6 case AF_INET6: fin6 = (struct sockaddr_in6 *)&addrs[0]; lin6 = (struct sockaddr_in6 *)&addrs[1]; if (fin6->sin6_len != sizeof(struct sockaddr_in6) || lin6->sin6_len != sizeof(struct sockaddr_in6)) return (EINVAL); if (IN6_IS_ADDR_V4MAPPED(&fin6->sin6_addr)) { if (!IN6_IS_ADDR_V4MAPPED(&lin6->sin6_addr)) return (EINVAL); in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[0]); in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[1]); fin = (struct sockaddr_in *)&addrs[0]; lin = (struct sockaddr_in *)&addrs[1]; break; } error = sa6_embedscope(fin6, V_ip6_use_defzone); if (error) return (error); error = sa6_embedscope(lin6, V_ip6_use_defzone); if (error) return (error); break; #endif #ifdef INET case AF_INET: fin = (struct sockaddr_in *)&addrs[0]; lin = (struct sockaddr_in *)&addrs[1]; if (fin->sin_len != sizeof(struct sockaddr_in) || lin->sin_len != sizeof(struct sockaddr_in)) return (EINVAL); break; #endif default: return (EINVAL); } NET_EPOCH_ENTER(et); switch (addrs[0].ss_family) { #ifdef INET6 case AF_INET6: inp = in6_pcblookup(&V_tcbinfo, &fin6->sin6_addr, fin6->sin6_port, &lin6->sin6_addr, lin6->sin6_port, INPLOOKUP_WLOCKPCB, NULL); break; #endif #ifdef INET case AF_INET: inp = in_pcblookup(&V_tcbinfo, fin->sin_addr, fin->sin_port, lin->sin_addr, lin->sin_port, INPLOOKUP_WLOCKPCB, NULL); break; #endif } NET_EPOCH_EXIT(et); if (inp != NULL) { if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) != 0 || inp->inp_socket == NULL) { error = ECONNRESET; INP_WUNLOCK(inp); } else { struct socket *so; so = inp->inp_socket; soref(so); error = ktls_set_tx_mode(so, arg2 == 0 ? TCP_TLS_MODE_SW : TCP_TLS_MODE_IFNET); INP_WUNLOCK(inp); SOCK_LOCK(so); sorele(so); } } else error = ESRCH; return (error); } SYSCTL_PROC(_net_inet_tcp, OID_AUTO, switch_to_sw_tls, CTLFLAG_VNET | CTLTYPE_STRUCT | CTLFLAG_WR | CTLFLAG_SKIP | CTLFLAG_NEEDGIANT, NULL, 0, sysctl_switch_tls, "", "Switch TCP connection to SW TLS"); SYSCTL_PROC(_net_inet_tcp, OID_AUTO, switch_to_ifnet_tls, CTLFLAG_VNET | CTLTYPE_STRUCT | CTLFLAG_WR | CTLFLAG_SKIP | CTLFLAG_NEEDGIANT, NULL, 1, sysctl_switch_tls, "", "Switch TCP connection to ifnet TLS"); #endif /* * Generate a standardized TCP log line for use throughout the * tcp subsystem. Memory allocation is done with M_NOWAIT to * allow use in the interrupt context. * * NB: The caller MUST free(s, M_TCPLOG) the returned string. * NB: The function may return NULL if memory allocation failed. * * Due to header inclusion and ordering limitations the struct ip * and ip6_hdr pointers have to be passed as void pointers. */ char * tcp_log_vain(struct in_conninfo *inc, struct tcphdr *th, void *ip4hdr, const void *ip6hdr) { /* Is logging enabled? */ if (V_tcp_log_in_vain == 0) return (NULL); return (tcp_log_addr(inc, th, ip4hdr, ip6hdr)); } char * tcp_log_addrs(struct in_conninfo *inc, struct tcphdr *th, void *ip4hdr, const void *ip6hdr) { /* Is logging enabled? */ if (tcp_log_debug == 0) return (NULL); return (tcp_log_addr(inc, th, ip4hdr, ip6hdr)); } static char * tcp_log_addr(struct in_conninfo *inc, struct tcphdr *th, void *ip4hdr, const void *ip6hdr) { char *s, *sp; size_t size; struct ip *ip; #ifdef INET6 const struct ip6_hdr *ip6; ip6 = (const struct ip6_hdr *)ip6hdr; #endif /* INET6 */ ip = (struct ip *)ip4hdr; /* * The log line looks like this: * "TCP: [1.2.3.4]:50332 to [1.2.3.4]:80 tcpflags 0x2" */ size = sizeof("TCP: []:12345 to []:12345 tcpflags 0x2<>") + sizeof(PRINT_TH_FLAGS) + 1 + #ifdef INET6 2 * INET6_ADDRSTRLEN; #else 2 * INET_ADDRSTRLEN; #endif /* INET6 */ s = malloc(size, M_TCPLOG, M_ZERO|M_NOWAIT); if (s == NULL) return (NULL); strcat(s, "TCP: ["); sp = s + strlen(s); if (inc && ((inc->inc_flags & INC_ISIPV6) == 0)) { inet_ntoa_r(inc->inc_faddr, sp); sp = s + strlen(s); sprintf(sp, "]:%i to [", ntohs(inc->inc_fport)); sp = s + strlen(s); inet_ntoa_r(inc->inc_laddr, sp); sp = s + strlen(s); sprintf(sp, "]:%i", ntohs(inc->inc_lport)); #ifdef INET6 } else if (inc) { ip6_sprintf(sp, &inc->inc6_faddr); sp = s + strlen(s); sprintf(sp, "]:%i to [", ntohs(inc->inc_fport)); sp = s + strlen(s); ip6_sprintf(sp, &inc->inc6_laddr); sp = s + strlen(s); sprintf(sp, "]:%i", ntohs(inc->inc_lport)); } else if (ip6 && th) { ip6_sprintf(sp, &ip6->ip6_src); sp = s + strlen(s); sprintf(sp, "]:%i to [", ntohs(th->th_sport)); sp = s + strlen(s); ip6_sprintf(sp, &ip6->ip6_dst); sp = s + strlen(s); sprintf(sp, "]:%i", ntohs(th->th_dport)); #endif /* INET6 */ #ifdef INET } else if (ip && th) { inet_ntoa_r(ip->ip_src, sp); sp = s + strlen(s); sprintf(sp, "]:%i to [", ntohs(th->th_sport)); sp = s + strlen(s); inet_ntoa_r(ip->ip_dst, sp); sp = s + strlen(s); sprintf(sp, "]:%i", ntohs(th->th_dport)); #endif /* INET */ } else { free(s, M_TCPLOG); return (NULL); } sp = s + strlen(s); if (th) sprintf(sp, " tcpflags 0x%b", th->th_flags, PRINT_TH_FLAGS); if (*(s + size - 1) != '\0') panic("%s: string too long", __func__); return (s); } /* * A subroutine which makes it easy to track TCP state changes with DTrace. * This function shouldn't be called for t_state initializations that don't * correspond to actual TCP state transitions. */ void tcp_state_change(struct tcpcb *tp, int newstate) { #if defined(KDTRACE_HOOKS) int pstate = tp->t_state; #endif TCPSTATES_DEC(tp->t_state); TCPSTATES_INC(newstate); tp->t_state = newstate; TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, pstate); } /* * Create an external-format (``xtcpcb'') structure using the information in * the kernel-format tcpcb structure pointed to by tp. This is done to * reduce the spew of irrelevant information over this interface, to isolate * user code from changes in the kernel structure, and potentially to provide * information-hiding if we decide that some of this information should be * hidden from users. */ void tcp_inptoxtp(const struct inpcb *inp, struct xtcpcb *xt) { struct tcpcb *tp = intotcpcb(inp); + struct tcptw *tw = intotw(inp); sbintime_t now; bzero(xt, sizeof(*xt)); if (inp->inp_flags & INP_TIMEWAIT) { xt->t_state = TCPS_TIME_WAIT; + xt->xt_encaps_port = tw->t_port; } else { xt->t_state = tp->t_state; xt->t_logstate = tp->t_logstate; xt->t_flags = tp->t_flags; xt->t_sndzerowin = tp->t_sndzerowin; xt->t_sndrexmitpack = tp->t_sndrexmitpack; xt->t_rcvoopack = tp->t_rcvoopack; xt->t_rcv_wnd = tp->rcv_wnd; xt->t_snd_wnd = tp->snd_wnd; xt->t_snd_cwnd = tp->snd_cwnd; xt->t_snd_ssthresh = tp->snd_ssthresh; xt->t_maxseg = tp->t_maxseg; xt->xt_ecn = (tp->t_flags2 & TF2_ECN_PERMIT) ? 1 : 0 + (tp->t_flags2 & TF2_ACE_PERMIT) ? 2 : 0; now = getsbinuptime(); #define COPYTIMER(ttt) do { \ if (callout_active(&tp->t_timers->ttt)) \ xt->ttt = (tp->t_timers->ttt.c_time - now) / \ SBT_1MS; \ else \ xt->ttt = 0; \ } while (0) COPYTIMER(tt_delack); COPYTIMER(tt_rexmt); COPYTIMER(tt_persist); COPYTIMER(tt_keep); COPYTIMER(tt_2msl); #undef COPYTIMER xt->t_rcvtime = 1000 * (ticks - tp->t_rcvtime) / hz; + xt->xt_encaps_port = tp->t_port; bcopy(tp->t_fb->tfb_tcp_block_name, xt->xt_stack, TCP_FUNCTION_NAME_LEN_MAX); bcopy(CC_ALGO(tp)->name, xt->xt_cc, TCP_CA_NAME_MAX); #ifdef TCP_BLACKBOX (void)tcp_log_get_id(tp, xt->xt_logid); #endif } xt->xt_len = sizeof(struct xtcpcb); in_pcbtoxinpcb(inp, &xt->xt_inp); if (inp->inp_socket == NULL) xt->xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; } void tcp_log_end_status(struct tcpcb *tp, uint8_t status) { uint32_t bit, i; if ((tp == NULL) || (status > TCP_EI_STATUS_MAX_VALUE) || (status == 0)) { /* Invalid */ return; } if (status > (sizeof(uint32_t) * 8)) { /* Should this be a KASSERT? */ return; } bit = 1U << (status - 1); if (bit & tp->t_end_info_status) { /* already logged */ return; } for (i = 0; i < TCP_END_BYTE_INFO; i++) { if (tp->t_end_info_bytes[i] == TCP_EI_EMPTY_SLOT) { tp->t_end_info_bytes[i] = status; tp->t_end_info_status |= bit; break; } } } diff --git a/sys/netinet/tcp_syncache.c b/sys/netinet/tcp_syncache.c index 4cd8411af8d5..35d9c091ab96 100644 --- a/sys/netinet/tcp_syncache.c +++ b/sys/netinet/tcp_syncache.c @@ -1,2525 +1,2594 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2001 McAfee, Inc. * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG * All rights reserved. * * This software was developed for the FreeBSD Project by Jonathan Lemon * and McAfee Research, the Security Research Division of McAfee, Inc. under * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the * DARPA CHATS research program. [2001 McAfee, Inc.] * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_pcbgroup.h" #include #include #include #include #include #include #include #include #include #include #include #include /* for proc0 declaration */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef INET6 #include #include #include #include #include #endif #include #include #include #include #include #include #include #ifdef INET6 #include #endif #ifdef TCP_OFFLOAD #include #endif +#include +#include #include #include #include VNET_DEFINE_STATIC(int, tcp_syncookies) = 1; #define V_tcp_syncookies VNET(tcp_syncookies) SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_syncookies), 0, "Use TCP SYN cookies if the syncache overflows"); VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0; #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly) SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_syncookiesonly), 0, "Use only TCP SYN cookies"); VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1; #define V_functions_inherit_listen_socket_stack \ VNET(functions_inherit_listen_socket_stack) SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(functions_inherit_listen_socket_stack), 0, "Inherit listen socket's stack"); #ifdef TCP_OFFLOAD #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL) #endif static void syncache_drop(struct syncache *, struct syncache_head *); static void syncache_free(struct syncache *); static void syncache_insert(struct syncache *, struct syncache_head *); static int syncache_respond(struct syncache *, const struct mbuf *, int); static struct socket *syncache_socket(struct syncache *, struct socket *, struct mbuf *m); static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout); static void syncache_timer(void *); static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t, uint8_t *, uintptr_t); static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *); static struct syncache *syncookie_lookup(struct in_conninfo *, struct syncache_head *, struct syncache *, struct tcphdr *, struct tcpopt *, - struct socket *); + struct socket *, uint16_t); static void syncache_pause(struct in_conninfo *); static void syncache_unpause(void *); static void syncookie_reseed(void *); #ifdef INVARIANTS static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, struct syncache *sc, struct tcphdr *th, struct tcpopt *to, - struct socket *lso); + struct socket *lso, uint16_t port); #endif /* * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies. * 3 retransmits corresponds to a timeout with default values of * tcp_rexmit_initial * ( 1 + * tcp_backoff[1] + * tcp_backoff[2] + * tcp_backoff[3]) + 3 * tcp_rexmit_slop, * 1000 ms * (1 + 2 + 4 + 8) + 3 * 200 ms = 15600 ms, * the odds are that the user has given up attempting to connect by then. */ #define SYNCACHE_MAXREXMTS 3 /* Arbitrary values */ #define TCP_SYNCACHE_HASHSIZE 512 #define TCP_SYNCACHE_BUCKETLIMIT 30 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache); #define V_tcp_syncache VNET(tcp_syncache) static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "TCP SYN cache"); SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN, &VNET_NAME(tcp_syncache.bucket_limit), 0, "Per-bucket hash limit for syncache"); SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN, &VNET_NAME(tcp_syncache.cache_limit), 0, "Overall entry limit for syncache"); SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET, &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache"); SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN, &VNET_NAME(tcp_syncache.hashsize), 0, "Size of TCP syncache hashtable"); SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0, "All syncache(4) entries are visible, ignoring UID/GID, jail(2) " "and mac(4) checks"); static int sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS) { int error; u_int new; new = V_tcp_syncache.rexmt_limit; error = sysctl_handle_int(oidp, &new, 0, req); if ((error == 0) && (req->newptr != NULL)) { if (new > TCP_MAXRXTSHIFT) error = EINVAL; else V_tcp_syncache.rexmt_limit = new; } return (error); } SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(tcp_syncache.rexmt_limit), 0, sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI", "Limit on SYN/ACK retransmissions"); VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, "Send reset on socket allocation failure"); static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) /* * Requires the syncache entry to be already removed from the bucket list. */ static void syncache_free(struct syncache *sc) { if (sc->sc_ipopts) (void) m_free(sc->sc_ipopts); if (sc->sc_cred) crfree(sc->sc_cred); #ifdef MAC mac_syncache_destroy(&sc->sc_label); #endif uma_zfree(V_tcp_syncache.zone, sc); } void syncache_init(void) { int i; V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; V_tcp_syncache.hash_secret = arc4random(); TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", &V_tcp_syncache.hashsize); TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", &V_tcp_syncache.bucket_limit); if (!powerof2(V_tcp_syncache.hashsize) || V_tcp_syncache.hashsize == 0) { printf("WARNING: syncache hash size is not a power of 2.\n"); V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; } V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; /* Set limits. */ V_tcp_syncache.cache_limit = V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", &V_tcp_syncache.cache_limit); /* Allocate the hash table. */ V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); #ifdef VIMAGE V_tcp_syncache.vnet = curvnet; #endif /* Initialize the hash buckets. */ for (i = 0; i < V_tcp_syncache.hashsize; i++) { TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", NULL, MTX_DEF); callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, &V_tcp_syncache.hashbase[i].sch_mtx, 0); V_tcp_syncache.hashbase[i].sch_length = 0; V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache; V_tcp_syncache.hashbase[i].sch_last_overflow = -(SYNCOOKIE_LIFETIME + 1); } /* Create the syncache entry zone. */ V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, V_tcp_syncache.cache_limit); /* Start the SYN cookie reseeder callout. */ callout_init(&V_tcp_syncache.secret.reseed, 1); arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, syncookie_reseed, &V_tcp_syncache); /* Initialize the pause machinery. */ mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF); callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx, 0); V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME; V_tcp_syncache.pause_backoff = 0; V_tcp_syncache.paused = false; } #ifdef VIMAGE void syncache_destroy(void) { struct syncache_head *sch; struct syncache *sc, *nsc; int i; /* * Stop the re-seed timer before freeing resources. No need to * possibly schedule it another time. */ callout_drain(&V_tcp_syncache.secret.reseed); /* Stop the SYN cache pause callout. */ mtx_lock(&V_tcp_syncache.pause_mtx); if (callout_stop(&V_tcp_syncache.pause_co) == 0) { mtx_unlock(&V_tcp_syncache.pause_mtx); callout_drain(&V_tcp_syncache.pause_co); } else mtx_unlock(&V_tcp_syncache.pause_mtx); /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ for (i = 0; i < V_tcp_syncache.hashsize; i++) { sch = &V_tcp_syncache.hashbase[i]; callout_drain(&sch->sch_timer); SCH_LOCK(sch); TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) syncache_drop(sc, sch); SCH_UNLOCK(sch); KASSERT(TAILQ_EMPTY(&sch->sch_bucket), ("%s: sch->sch_bucket not empty", __func__)); KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", __func__, sch->sch_length)); mtx_destroy(&sch->sch_mtx); } KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, ("%s: cache_count not 0", __func__)); /* Free the allocated global resources. */ uma_zdestroy(V_tcp_syncache.zone); free(V_tcp_syncache.hashbase, M_SYNCACHE); mtx_destroy(&V_tcp_syncache.pause_mtx); } #endif /* * Inserts a syncache entry into the specified bucket row. * Locks and unlocks the syncache_head autonomously. */ static void syncache_insert(struct syncache *sc, struct syncache_head *sch) { struct syncache *sc2; SCH_LOCK(sch); /* * Make sure that we don't overflow the per-bucket limit. * If the bucket is full, toss the oldest element. */ if (sch->sch_length >= V_tcp_syncache.bucket_limit) { KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), ("sch->sch_length incorrect")); syncache_pause(&sc->sc_inc); sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); sch->sch_last_overflow = time_uptime; syncache_drop(sc2, sch); } /* Put it into the bucket. */ TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); sch->sch_length++; #ifdef TCP_OFFLOAD if (ADDED_BY_TOE(sc)) { struct toedev *tod = sc->sc_tod; tod->tod_syncache_added(tod, sc->sc_todctx); } #endif /* Reinitialize the bucket row's timer. */ if (sch->sch_length == 1) sch->sch_nextc = ticks + INT_MAX; syncache_timeout(sc, sch, 1); SCH_UNLOCK(sch); TCPSTATES_INC(TCPS_SYN_RECEIVED); TCPSTAT_INC(tcps_sc_added); } /* * Remove and free entry from syncache bucket row. * Expects locked syncache head. */ static void syncache_drop(struct syncache *sc, struct syncache_head *sch) { SCH_LOCK_ASSERT(sch); TCPSTATES_DEC(TCPS_SYN_RECEIVED); TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); sch->sch_length--; #ifdef TCP_OFFLOAD if (ADDED_BY_TOE(sc)) { struct toedev *tod = sc->sc_tod; tod->tod_syncache_removed(tod, sc->sc_todctx); } #endif syncache_free(sc); } /* * Engage/reengage time on bucket row. */ static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) { int rexmt; if (sc->sc_rxmits == 0) rexmt = tcp_rexmit_initial; else TCPT_RANGESET(rexmt, tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits], tcp_rexmit_min, TCPTV_REXMTMAX); sc->sc_rxttime = ticks + rexmt; sc->sc_rxmits++; if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { sch->sch_nextc = sc->sc_rxttime; if (docallout) callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, syncache_timer, (void *)sch); } } /* * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. * If we have retransmitted an entry the maximum number of times, expire it. * One separate timer for each bucket row. */ static void syncache_timer(void *xsch) { struct syncache_head *sch = (struct syncache_head *)xsch; struct syncache *sc, *nsc; struct epoch_tracker et; int tick = ticks; char *s; bool paused; CURVNET_SET(sch->sch_sc->vnet); /* NB: syncache_head has already been locked by the callout. */ SCH_LOCK_ASSERT(sch); /* * In the following cycle we may remove some entries and/or * advance some timeouts, so re-initialize the bucket timer. */ sch->sch_nextc = tick + INT_MAX; /* * If we have paused processing, unconditionally remove * all syncache entries. */ mtx_lock(&V_tcp_syncache.pause_mtx); paused = V_tcp_syncache.paused; mtx_unlock(&V_tcp_syncache.pause_mtx); TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { if (paused) { syncache_drop(sc, sch); continue; } /* * We do not check if the listen socket still exists * and accept the case where the listen socket may be * gone by the time we resend the SYN/ACK. We do * not expect this to happens often. If it does, * then the RST will be sent by the time the remote * host does the SYN/ACK->ACK. */ if (TSTMP_GT(sc->sc_rxttime, tick)) { if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) sch->sch_nextc = sc->sc_rxttime; continue; } if (sc->sc_rxmits > V_tcp_ecn_maxretries) { sc->sc_flags &= ~SCF_ECN; } if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " "giving up and removing syncache entry\n", s, __func__); free(s, M_TCPLOG); } syncache_drop(sc, sch); TCPSTAT_INC(tcps_sc_stale); continue; } if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: Response timeout, " "retransmitting (%u) SYN|ACK\n", s, __func__, sc->sc_rxmits); free(s, M_TCPLOG); } NET_EPOCH_ENTER(et); syncache_respond(sc, NULL, TH_SYN|TH_ACK); NET_EPOCH_EXIT(et); TCPSTAT_INC(tcps_sc_retransmitted); syncache_timeout(sc, sch, 0); } if (!TAILQ_EMPTY(&(sch)->sch_bucket)) callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, syncache_timer, (void *)(sch)); CURVNET_RESTORE(); } /* * Returns true if the system is only using cookies at the moment. * This could be due to a sysadmin decision to only use cookies, or it * could be due to the system detecting an attack. */ static inline bool syncache_cookiesonly(void) { return (V_tcp_syncookies && (V_tcp_syncache.paused || V_tcp_syncookiesonly)); } /* * Find the hash bucket for the given connection. */ static struct syncache_head * syncache_hashbucket(struct in_conninfo *inc) { uint32_t hash; /* * The hash is built on foreign port + local port + foreign address. * We rely on the fact that struct in_conninfo starts with 16 bits * of foreign port, then 16 bits of local port then followed by 128 * bits of foreign address. In case of IPv4 address, the first 3 * 32-bit words of the address always are zeroes. */ hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5, V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask; return (&V_tcp_syncache.hashbase[hash]); } /* * Find an entry in the syncache. * Returns always with locked syncache_head plus a matching entry or NULL. */ static struct syncache * syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) { struct syncache *sc; struct syncache_head *sch; *schp = sch = syncache_hashbucket(inc); SCH_LOCK(sch); /* Circle through bucket row to find matching entry. */ TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie, sizeof(struct in_endpoints)) == 0) break; return (sc); /* Always returns with locked sch. */ } /* * This function is called when we get a RST for a * non-existent connection, so that we can see if the * connection is in the syn cache. If it is, zap it. * If required send a challenge ACK. */ void -syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m) +syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m, + uint16_t port) { struct syncache *sc; struct syncache_head *sch; char *s = NULL; if (syncache_cookiesonly()) return; sc = syncache_lookup(inc, &sch); /* returns locked sch */ SCH_LOCK_ASSERT(sch); /* * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. * See RFC 793 page 65, section SEGMENT ARRIVES. */ if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " "FIN flag set, segment ignored\n", s, __func__); TCPSTAT_INC(tcps_badrst); goto done; } /* * No corresponding connection was found in syncache. * If syncookies are enabled and possibly exclusively * used, or we are under memory pressure, a valid RST * may not find a syncache entry. In that case we're * done and no SYN|ACK retransmissions will happen. * Otherwise the RST was misdirected or spoofed. */ if (sc == NULL) { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Spurious RST without matching " "syncache entry (possibly syncookie only), " "segment ignored\n", s, __func__); TCPSTAT_INC(tcps_badrst); goto done; } + /* The remote UDP encaps port does not match. */ + if (sc->sc_port != port) { + if ((s = tcp_log_addrs(inc, th, NULL, NULL))) + log(LOG_DEBUG, "%s; %s: Spurious RST with matching " + "syncache entry but non-matching UDP encaps port, " + "segment ignored\n", s, __func__); + TCPSTAT_INC(tcps_badrst); + goto done; + } + /* * If the RST bit is set, check the sequence number to see * if this is a valid reset segment. * * RFC 793 page 37: * In all states except SYN-SENT, all reset (RST) segments * are validated by checking their SEQ-fields. A reset is * valid if its sequence number is in the window. * * RFC 793 page 69: * There are four cases for the acceptability test for an incoming * segment: * * Segment Receive Test * Length Window * ------- ------- ------------------------------------------- * 0 0 SEG.SEQ = RCV.NXT * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND * >0 0 not acceptable * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND * * Note that when receiving a SYN segment in the LISTEN state, * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as * described in RFC 793, page 66. */ if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) && SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) || (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) { if (V_tcp_insecure_rst || th->th_seq == sc->sc_irs + 1) { syncache_drop(sc, sch); if ((s = tcp_log_addrs(inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Our SYN|ACK was rejected, " "connection attempt aborted by remote " "endpoint\n", s, __func__); TCPSTAT_INC(tcps_sc_reset); } else { TCPSTAT_INC(tcps_badrst); /* Send challenge ACK. */ if ((s = tcp_log_addrs(inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: RST with invalid " " SEQ %u != NXT %u (+WND %u), " "sending challenge ACK\n", s, __func__, th->th_seq, sc->sc_irs + 1, sc->sc_wnd); syncache_respond(sc, m, TH_ACK); } } else { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " "NXT %u (+WND %u), segment ignored\n", s, __func__, th->th_seq, sc->sc_irs + 1, sc->sc_wnd); TCPSTAT_INC(tcps_badrst); } done: if (s != NULL) free(s, M_TCPLOG); SCH_UNLOCK(sch); } void -syncache_badack(struct in_conninfo *inc) +syncache_badack(struct in_conninfo *inc, uint16_t port) { struct syncache *sc; struct syncache_head *sch; if (syncache_cookiesonly()) return; sc = syncache_lookup(inc, &sch); /* returns locked sch */ SCH_LOCK_ASSERT(sch); - if (sc != NULL) { + if ((sc != NULL) && (sc->sc_port == port)) { syncache_drop(sc, sch); TCPSTAT_INC(tcps_sc_badack); } SCH_UNLOCK(sch); } void -syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq) +syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port) { struct syncache *sc; struct syncache_head *sch; if (syncache_cookiesonly()) return; sc = syncache_lookup(inc, &sch); /* returns locked sch */ SCH_LOCK_ASSERT(sch); if (sc == NULL) goto done; + /* If the port != sc_port, then it's a bogus ICMP msg */ + if (port != sc->sc_port) + goto done; + /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ if (ntohl(th_seq) != sc->sc_iss) goto done; /* * If we've rertransmitted 3 times and this is our second error, * we remove the entry. Otherwise, we allow it to continue on. * This prevents us from incorrectly nuking an entry during a * spurious network outage. * * See tcp_notify(). */ if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { sc->sc_flags |= SCF_UNREACH; goto done; } syncache_drop(sc, sch); TCPSTAT_INC(tcps_sc_unreach); done: SCH_UNLOCK(sch); } /* * Build a new TCP socket structure from a syncache entry. * * On success return the newly created socket with its underlying inp locked. */ static struct socket * syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) { struct tcp_function_block *blk; struct inpcb *inp = NULL; struct socket *so; struct tcpcb *tp; int error; char *s; NET_EPOCH_ASSERT(); /* * Ok, create the full blown connection, and set things up * as they would have been set up if we had created the * connection when the SYN arrived. If we can't create * the connection, abort it. */ so = sonewconn(lso, 0); if (so == NULL) { /* * Drop the connection; we will either send a RST or * have the peer retransmit its SYN again after its * RTO and try again. */ TCPSTAT_INC(tcps_listendrop); if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: Socket create failed " "due to limits or memory shortage\n", s, __func__); free(s, M_TCPLOG); } goto abort2; } #ifdef MAC mac_socketpeer_set_from_mbuf(m, so); #endif inp = sotoinpcb(so); inp->inp_inc.inc_fibnum = so->so_fibnum; INP_WLOCK(inp); /* * Exclusive pcbinfo lock is not required in syncache socket case even * if two inpcb locks can be acquired simultaneously: * - the inpcb in LISTEN state, * - the newly created inp. * * In this case, an inp cannot be at same time in LISTEN state and * just created by an accept() call. */ INP_HASH_WLOCK(&V_tcbinfo); /* Insert new socket into PCB hash list. */ inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; #ifdef INET6 if (sc->sc_inc.inc_flags & INC_ISIPV6) { inp->inp_vflag &= ~INP_IPV4; inp->inp_vflag |= INP_IPV6; inp->in6p_laddr = sc->sc_inc.inc6_laddr; } else { inp->inp_vflag &= ~INP_IPV6; inp->inp_vflag |= INP_IPV4; #endif inp->inp_ip_ttl = sc->sc_ip_ttl; inp->inp_ip_tos = sc->sc_ip_tos; inp->inp_laddr = sc->sc_inc.inc_laddr; #ifdef INET6 } #endif /* * If there's an mbuf and it has a flowid, then let's initialise the * inp with that particular flowid. */ if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { inp->inp_flowid = m->m_pkthdr.flowid; inp->inp_flowtype = M_HASHTYPE_GET(m); #ifdef NUMA inp->inp_numa_domain = m->m_pkthdr.numa_domain; #endif } inp->inp_lport = sc->sc_inc.inc_lport; #ifdef INET6 if (inp->inp_vflag & INP_IPV6PROTO) { struct inpcb *oinp = sotoinpcb(lso); /* * Inherit socket options from the listening socket. * Note that in6p_inputopts are not (and should not be) * copied, since it stores previously received options and is * used to detect if each new option is different than the * previous one and hence should be passed to a user. * If we copied in6p_inputopts, a user would not be able to * receive options just after calling the accept system call. */ inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; if (oinp->in6p_outputopts) inp->in6p_outputopts = ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); inp->in6p_hops = oinp->in6p_hops; } if (sc->sc_inc.inc_flags & INC_ISIPV6) { struct in6_addr laddr6; struct sockaddr_in6 sin6; sin6.sin6_family = AF_INET6; sin6.sin6_len = sizeof(sin6); sin6.sin6_addr = sc->sc_inc.inc6_faddr; sin6.sin6_port = sc->sc_inc.inc_fport; sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; laddr6 = inp->in6p_laddr; if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) inp->in6p_laddr = sc->sc_inc.inc6_laddr; if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6, thread0.td_ucred, m, false)) != 0) { inp->in6p_laddr = laddr6; if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed " "with error %i\n", s, __func__, error); free(s, M_TCPLOG); } INP_HASH_WUNLOCK(&V_tcbinfo); goto abort; } /* Override flowlabel from in6_pcbconnect. */ inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; inp->inp_flow |= sc->sc_flowlabel; } #endif /* INET6 */ #if defined(INET) && defined(INET6) else #endif #ifdef INET { struct in_addr laddr; struct sockaddr_in sin; inp->inp_options = (m) ? ip_srcroute(m) : NULL; if (inp->inp_options == NULL) { inp->inp_options = sc->sc_ipopts; sc->sc_ipopts = NULL; } sin.sin_family = AF_INET; sin.sin_len = sizeof(sin); sin.sin_addr = sc->sc_inc.inc_faddr; sin.sin_port = sc->sc_inc.inc_fport; bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); laddr = inp->inp_laddr; if (inp->inp_laddr.s_addr == INADDR_ANY) inp->inp_laddr = sc->sc_inc.inc_laddr; if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin, thread0.td_ucred, m, false)) != 0) { inp->inp_laddr = laddr; if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: in_pcbconnect failed " "with error %i\n", s, __func__, error); free(s, M_TCPLOG); } INP_HASH_WUNLOCK(&V_tcbinfo); goto abort; } } #endif /* INET */ #if defined(IPSEC) || defined(IPSEC_SUPPORT) /* Copy old policy into new socket's. */ if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0) printf("syncache_socket: could not copy policy\n"); #endif INP_HASH_WUNLOCK(&V_tcbinfo); tp = intotcpcb(inp); tcp_state_change(tp, TCPS_SYN_RECEIVED); tp->iss = sc->sc_iss; tp->irs = sc->sc_irs; + tp->t_port = sc->sc_port; tcp_rcvseqinit(tp); tcp_sendseqinit(tp); blk = sototcpcb(lso)->t_fb; if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) { /* * Our parents t_fb was not the default, * we need to release our ref on tp->t_fb and * pickup one on the new entry. */ struct tcp_function_block *rblk; rblk = find_and_ref_tcp_fb(blk); KASSERT(rblk != NULL, ("cannot find blk %p out of syncache?", blk)); if (tp->t_fb->tfb_tcp_fb_fini) (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); refcount_release(&tp->t_fb->tfb_refcnt); tp->t_fb = rblk; /* * XXXrrs this is quite dangerous, it is possible * for the new function to fail to init. We also * are not asking if the handoff_is_ok though at * the very start thats probalbly ok. */ if (tp->t_fb->tfb_tcp_fb_init) { (*tp->t_fb->tfb_tcp_fb_init)(tp); } } tp->snd_wl1 = sc->sc_irs; tp->snd_max = tp->iss + 1; tp->snd_nxt = tp->iss + 1; tp->rcv_up = sc->sc_irs + 1; tp->rcv_wnd = sc->sc_wnd; tp->rcv_adv += tp->rcv_wnd; tp->last_ack_sent = tp->rcv_nxt; tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); if (sc->sc_flags & SCF_NOOPT) tp->t_flags |= TF_NOOPT; else { if (sc->sc_flags & SCF_WINSCALE) { tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; tp->snd_scale = sc->sc_requested_s_scale; tp->request_r_scale = sc->sc_requested_r_scale; } if (sc->sc_flags & SCF_TIMESTAMP) { tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; tp->ts_recent = sc->sc_tsreflect; tp->ts_recent_age = tcp_ts_getticks(); tp->ts_offset = sc->sc_tsoff; } #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (sc->sc_flags & SCF_SIGNATURE) tp->t_flags |= TF_SIGNATURE; #endif if (sc->sc_flags & SCF_SACK) tp->t_flags |= TF_SACK_PERMIT; } if (sc->sc_flags & SCF_ECN) tp->t_flags2 |= TF2_ECN_PERMIT; /* * Set up MSS and get cached values from tcp_hostcache. * This might overwrite some of the defaults we just set. */ tcp_mss(tp, sc->sc_peer_mss); /* * If the SYN,ACK was retransmitted, indicate that CWND to be * limited to one segment in cc_conn_init(). * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. */ if (sc->sc_rxmits > 1) tp->snd_cwnd = 1; #ifdef TCP_OFFLOAD /* * Allow a TOE driver to install its hooks. Note that we hold the * pcbinfo lock too and that prevents tcp_usr_accept from accepting a * new connection before the TOE driver has done its thing. */ if (ADDED_BY_TOE(sc)) { struct toedev *tod = sc->sc_tod; tod->tod_offload_socket(tod, sc->sc_todctx, so); } #endif /* * Copy and activate timers. */ tp->t_keepinit = sototcpcb(lso)->t_keepinit; tp->t_keepidle = sototcpcb(lso)->t_keepidle; tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); TCPSTAT_INC(tcps_accepts); return (so); abort: INP_WUNLOCK(inp); abort2: if (so != NULL) soabort(so); return (NULL); } /* * This function gets called when we receive an ACK for a * socket in the LISTEN state. We look up the connection * in the syncache, and if its there, we pull it out of * the cache and turn it into a full-blown connection in * the SYN-RECEIVED state. * * On syncache_socket() success the newly created socket * has its underlying inp locked. */ int syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, - struct socket **lsop, struct mbuf *m) + struct socket **lsop, struct mbuf *m, uint16_t port) { struct syncache *sc; struct syncache_head *sch; struct syncache scs; char *s; bool locked; NET_EPOCH_ASSERT(); KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, ("%s: can handle only ACK", __func__)); if (syncache_cookiesonly()) { sc = NULL; sch = syncache_hashbucket(inc); locked = false; } else { sc = syncache_lookup(inc, &sch); /* returns locked sch */ locked = true; SCH_LOCK_ASSERT(sch); } #ifdef INVARIANTS /* * Test code for syncookies comparing the syncache stored * values with the reconstructed values from the cookie. */ if (sc != NULL) - syncookie_cmp(inc, sch, sc, th, to, *lsop); + syncookie_cmp(inc, sch, sc, th, to, *lsop, port); #endif if (sc == NULL) { /* * There is no syncache entry, so see if this ACK is * a returning syncookie. To do this, first: * A. Check if syncookies are used in case of syncache * overflows * B. See if this socket has had a syncache entry dropped in * the recent past. We don't want to accept a bogus * syncookie if we've never received a SYN or accept it * twice. * C. check that the syncookie is valid. If it is, then * cobble up a fake syncache entry, and return. */ if (locked && !V_tcp_syncookies) { SCH_UNLOCK(sch); if ((s = tcp_log_addrs(inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Spurious ACK, " "segment rejected (syncookies disabled)\n", s, __func__); goto failed; } if (locked && !V_tcp_syncookiesonly && sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) { SCH_UNLOCK(sch); if ((s = tcp_log_addrs(inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Spurious ACK, " "segment rejected (no syncache entry)\n", s, __func__); goto failed; } bzero(&scs, sizeof(scs)); - sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop); + sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port); if (locked) SCH_UNLOCK(sch); if (sc == NULL) { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Segment failed " "SYNCOOKIE authentication, segment rejected " "(probably spoofed)\n", s, __func__); goto failed; } #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) /* If received ACK has MD5 signature, check it. */ if ((to->to_flags & TOF_SIGNATURE) != 0 && (!TCPMD5_ENABLED() || TCPMD5_INPUT(m, th, to->to_signature) != 0)) { /* Drop the ACK. */ if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: Segment rejected, " "MD5 signature doesn't match.\n", s, __func__); free(s, M_TCPLOG); } TCPSTAT_INC(tcps_sig_err_sigopt); return (-1); /* Do not send RST */ } #endif /* TCP_SIGNATURE */ } else { + if (sc->sc_port != port) { + SCH_UNLOCK(sch); + return (0); + } #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) /* * If listening socket requested TCP digests, check that * received ACK has signature and it is correct. * If not, drop the ACK and leave sc entry in th cache, * because SYN was received with correct signature. */ if (sc->sc_flags & SCF_SIGNATURE) { if ((to->to_flags & TOF_SIGNATURE) == 0) { /* No signature */ TCPSTAT_INC(tcps_sig_err_nosigopt); SCH_UNLOCK(sch); if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: Segment " "rejected, MD5 signature wasn't " "provided.\n", s, __func__); free(s, M_TCPLOG); } return (-1); /* Do not send RST */ } if (!TCPMD5_ENABLED() || TCPMD5_INPUT(m, th, to->to_signature) != 0) { /* Doesn't match or no SA */ SCH_UNLOCK(sch); if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: Segment " "rejected, MD5 signature doesn't " "match.\n", s, __func__); free(s, M_TCPLOG); } return (-1); /* Do not send RST */ } } #endif /* TCP_SIGNATURE */ /* * RFC 7323 PAWS: If we have a timestamp on this segment and * it's less than ts_recent, drop it. * XXXMT: RFC 7323 also requires to send an ACK. * In tcp_input.c this is only done for TCP segments * with user data, so be consistent here and just drop * the segment. */ if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS && TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) { SCH_UNLOCK(sch); if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: SEG.TSval %u < TS.Recent %u, " "segment dropped\n", s, __func__, to->to_tsval, sc->sc_tsreflect); free(s, M_TCPLOG); } return (-1); /* Do not send RST */ } /* * If timestamps were not negotiated during SYN/ACK and a * segment with a timestamp is received, ignore the * timestamp and process the packet normally. * See section 3.2 of RFC 7323. */ if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: Timestamp not " "expected, segment processed normally\n", s, __func__); free(s, M_TCPLOG); s = NULL; } } /* * If timestamps were negotiated during SYN/ACK and a * segment without a timestamp is received, silently drop * the segment, unless the missing timestamps are tolerated. * See section 3.2 of RFC 7323. */ if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) { if (V_tcp_tolerate_missing_ts) { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: Timestamp missing, " "segment processed normally\n", s, __func__); free(s, M_TCPLOG); } } else { SCH_UNLOCK(sch); if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: Timestamp missing, " "segment silently dropped\n", s, __func__); free(s, M_TCPLOG); } return (-1); /* Do not send RST */ } } /* * Pull out the entry to unlock the bucket row. * * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not * tcp_state_change(). The tcpcb is not existent at this * moment. A new one will be allocated via syncache_socket-> * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then * syncache_socket() will change it to TCPS_SYN_RECEIVED. */ TCPSTATES_DEC(TCPS_SYN_RECEIVED); TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); sch->sch_length--; #ifdef TCP_OFFLOAD if (ADDED_BY_TOE(sc)) { struct toedev *tod = sc->sc_tod; tod->tod_syncache_removed(tod, sc->sc_todctx); } #endif SCH_UNLOCK(sch); } /* * Segment validation: * ACK must match our initial sequence number + 1 (the SYN|ACK). */ if (th->th_ack != sc->sc_iss + 1) { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " "rejected\n", s, __func__, th->th_ack, sc->sc_iss); goto failed; } /* * The SEQ must fall in the window starting at the received * initial receive sequence number + 1 (the SYN). */ if (SEQ_LEQ(th->th_seq, sc->sc_irs) || SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " "rejected\n", s, __func__, th->th_seq, sc->sc_irs); goto failed; } *lsop = syncache_socket(sc, *lsop, m); if (*lsop == NULL) TCPSTAT_INC(tcps_sc_aborted); else TCPSTAT_INC(tcps_sc_completed); /* how do we find the inp for the new socket? */ if (sc != &scs) syncache_free(sc); return (1); failed: if (sc != NULL && sc != &scs) syncache_free(sc); if (s != NULL) free(s, M_TCPLOG); *lsop = NULL; return (0); } static struct socket * syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m, uint64_t response_cookie) { struct inpcb *inp; struct tcpcb *tp; unsigned int *pending_counter; struct socket *so; NET_EPOCH_ASSERT(); pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending; so = syncache_socket(sc, lso, m); if (so == NULL) { TCPSTAT_INC(tcps_sc_aborted); atomic_subtract_int(pending_counter, 1); } else { soisconnected(so); inp = sotoinpcb(so); tp = intotcpcb(inp); tp->t_flags |= TF_FASTOPEN; tp->t_tfo_cookie.server = response_cookie; tp->snd_max = tp->iss; tp->snd_nxt = tp->iss; tp->t_tfo_pending = pending_counter; TCPSTAT_INC(tcps_sc_completed); } return (so); } /* * Given a LISTEN socket and an inbound SYN request, add * this to the syn cache, and send back a segment: * * to the source. * * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. * Doing so would require that we hold onto the data and deliver it * to the application. However, if we are the target of a SYN-flood * DoS attack, an attacker could send data which would eventually * consume all available buffer space if it were ACKed. By not ACKing * the data, we avoid this DoS scenario. * * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) * cookie is processed and a new socket is created. In this case, any data * accompanying the SYN will be queued to the socket by tcp_input() and will * be ACKed either when the application sends response data or the delayed * ACK timer expires, whichever comes first. */ struct socket * syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod, - void *todctx, uint8_t iptos) + void *todctx, uint8_t iptos, uint16_t port) { struct tcpcb *tp; struct socket *rv = NULL; struct syncache *sc = NULL; struct syncache_head *sch; struct mbuf *ipopts = NULL; u_int ltflags; int win, ip_ttl, ip_tos; char *s; #ifdef INET6 int autoflowlabel = 0; #endif #ifdef MAC struct label *maclabel; #endif struct syncache scs; struct ucred *cred; uint64_t tfo_response_cookie; unsigned int *tfo_pending = NULL; int tfo_cookie_valid = 0; int tfo_response_cookie_valid = 0; bool locked; INP_RLOCK_ASSERT(inp); /* listen socket */ KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, ("%s: unexpected tcp flags", __func__)); /* * Combine all so/tp operations very early to drop the INP lock as * soon as possible. */ KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so)); tp = sototcpcb(so); cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred); #ifdef INET6 if (inc->inc_flags & INC_ISIPV6) { if (inp->inp_flags & IN6P_AUTOFLOWLABEL) { autoflowlabel = 1; } ip_ttl = in6_selecthlim(inp, NULL); if ((inp->in6p_outputopts == NULL) || (inp->in6p_outputopts->ip6po_tclass == -1)) { ip_tos = 0; } else { ip_tos = inp->in6p_outputopts->ip6po_tclass; } } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { ip_ttl = inp->inp_ip_ttl; ip_tos = inp->inp_ip_tos; } #endif win = so->sol_sbrcv_hiwat; ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) && (tp->t_tfo_pending != NULL) && (to->to_flags & TOF_FASTOPEN)) { /* * Limit the number of pending TFO connections to * approximately half of the queue limit. This prevents TFO * SYN floods from starving the service by filling the * listen queue with bogus TFO connections. */ if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= (so->sol_qlimit / 2)) { int result; result = tcp_fastopen_check_cookie(inc, to->to_tfo_cookie, to->to_tfo_len, &tfo_response_cookie); tfo_cookie_valid = (result > 0); tfo_response_cookie_valid = (result >= 0); } /* * Remember the TFO pending counter as it will have to be * decremented below if we don't make it to syncache_tfo_expand(). */ tfo_pending = tp->t_tfo_pending; } /* By the time we drop the lock these should no longer be used. */ so = NULL; tp = NULL; #ifdef MAC if (mac_syncache_init(&maclabel) != 0) { INP_RUNLOCK(inp); goto done; } else mac_syncache_create(maclabel, inp); #endif if (!tfo_cookie_valid) INP_RUNLOCK(inp); /* * Remember the IP options, if any. */ #ifdef INET6 if (!(inc->inc_flags & INC_ISIPV6)) #endif #ifdef INET ipopts = (m) ? ip_srcroute(m) : NULL; #else ipopts = NULL; #endif #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) /* * If listening socket requested TCP digests, check that received * SYN has signature and it is correct. If signature doesn't match * or TCP_SIGNATURE support isn't enabled, drop the packet. */ if (ltflags & TF_SIGNATURE) { if ((to->to_flags & TOF_SIGNATURE) == 0) { TCPSTAT_INC(tcps_sig_err_nosigopt); goto done; } if (!TCPMD5_ENABLED() || TCPMD5_INPUT(m, th, to->to_signature) != 0) goto done; } #endif /* TCP_SIGNATURE */ /* * See if we already have an entry for this connection. * If we do, resend the SYN,ACK, and reset the retransmit timer. * * XXX: should the syncache be re-initialized with the contents * of the new SYN here (which may have different options?) * * XXX: We do not check the sequence number to see if this is a * real retransmit or a new connection attempt. The question is * how to handle such a case; either ignore it as spoofed, or * drop the current entry and create a new one? */ if (syncache_cookiesonly()) { sc = NULL; sch = syncache_hashbucket(inc); locked = false; } else { sc = syncache_lookup(inc, &sch); /* returns locked sch */ locked = true; SCH_LOCK_ASSERT(sch); } if (sc != NULL) { if (tfo_cookie_valid) INP_RUNLOCK(inp); TCPSTAT_INC(tcps_sc_dupsyn); if (ipopts) { /* * If we were remembering a previous source route, * forget it and use the new one we've been given. */ if (sc->sc_ipopts) (void) m_free(sc->sc_ipopts); sc->sc_ipopts = ipopts; } /* * Update timestamp if present. */ if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) sc->sc_tsreflect = to->to_tsval; else sc->sc_flags &= ~SCF_TIMESTAMP; /* * Disable ECN if needed. */ if ((sc->sc_flags & SCF_ECN) && ((th->th_flags & (TH_ECE|TH_CWR)) != (TH_ECE|TH_CWR))) { sc->sc_flags &= ~SCF_ECN; } #ifdef MAC /* * Since we have already unconditionally allocated label * storage, free it up. The syncache entry will already * have an initialized label we can use. */ mac_syncache_destroy(&maclabel); #endif TCP_PROBE5(receive, NULL, NULL, m, NULL, th); /* Retransmit SYN|ACK and reset retransmit count. */ if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " "resetting timer and retransmitting SYN|ACK\n", s, __func__); free(s, M_TCPLOG); } if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { sc->sc_rxmits = 0; syncache_timeout(sc, sch, 1); TCPSTAT_INC(tcps_sndacks); TCPSTAT_INC(tcps_sndtotal); } SCH_UNLOCK(sch); goto donenoprobe; } if (tfo_cookie_valid) { bzero(&scs, sizeof(scs)); sc = &scs; goto skip_alloc; } /* * Skip allocating a syncache entry if we are just going to discard * it later. */ if (!locked) { bzero(&scs, sizeof(scs)); sc = &scs; } else sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); if (sc == NULL) { /* * The zone allocator couldn't provide more entries. * Treat this as if the cache was full; drop the oldest * entry and insert the new one. */ TCPSTAT_INC(tcps_sc_zonefail); if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) { sch->sch_last_overflow = time_uptime; syncache_drop(sc, sch); syncache_pause(inc); } sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); if (sc == NULL) { if (V_tcp_syncookies) { bzero(&scs, sizeof(scs)); sc = &scs; } else { KASSERT(locked, ("%s: bucket unexpectedly unlocked", __func__)); SCH_UNLOCK(sch); if (ipopts) (void) m_free(ipopts); goto done; } } } skip_alloc: if (!tfo_cookie_valid && tfo_response_cookie_valid) sc->sc_tfo_cookie = &tfo_response_cookie; /* * Fill in the syncache values. */ #ifdef MAC sc->sc_label = maclabel; #endif sc->sc_cred = cred; + sc->sc_port = port; cred = NULL; sc->sc_ipopts = ipopts; bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); sc->sc_ip_tos = ip_tos; sc->sc_ip_ttl = ip_ttl; #ifdef TCP_OFFLOAD sc->sc_tod = tod; sc->sc_todctx = todctx; #endif sc->sc_irs = th->th_seq; sc->sc_flags = 0; sc->sc_flowlabel = 0; /* * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. * win was derived from socket earlier in the function. */ win = imax(win, 0); win = imin(win, TCP_MAXWIN); sc->sc_wnd = win; if (V_tcp_do_rfc1323 && !(ltflags & TF_NOOPT)) { /* * A timestamp received in a SYN makes * it ok to send timestamp requests and replies. */ if (to->to_flags & TOF_TS) { sc->sc_tsreflect = to->to_tsval; sc->sc_flags |= SCF_TIMESTAMP; sc->sc_tsoff = tcp_new_ts_offset(inc); } if (to->to_flags & TOF_SCALE) { int wscale = 0; /* * Pick the smallest possible scaling factor that * will still allow us to scale up to sb_max, aka * kern.ipc.maxsockbuf. * * We do this because there are broken firewalls that * will corrupt the window scale option, leading to * the other endpoint believing that our advertised * window is unscaled. At scale factors larger than * 5 the unscaled window will drop below 1500 bytes, * leading to serious problems when traversing these * broken firewalls. * * With the default maxsockbuf of 256K, a scale factor * of 3 will be chosen by this algorithm. Those who * choose a larger maxsockbuf should watch out * for the compatibility problems mentioned above. * * RFC1323: The Window field in a SYN (i.e., a * or ) segment itself is never scaled. */ while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) wscale++; sc->sc_requested_r_scale = wscale; sc->sc_requested_s_scale = to->to_wscale; sc->sc_flags |= SCF_WINSCALE; } } #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) /* * If listening socket requested TCP digests, flag this in the * syncache so that syncache_respond() will do the right thing * with the SYN+ACK. */ if (ltflags & TF_SIGNATURE) sc->sc_flags |= SCF_SIGNATURE; #endif /* TCP_SIGNATURE */ if (to->to_flags & TOF_SACKPERM) sc->sc_flags |= SCF_SACK; if (to->to_flags & TOF_MSS) sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ if (ltflags & TF_NOOPT) sc->sc_flags |= SCF_NOOPT; if (((th->th_flags & (TH_ECE|TH_CWR)) == (TH_ECE|TH_CWR)) && V_tcp_do_ecn) sc->sc_flags |= SCF_ECN; if (V_tcp_syncookies) sc->sc_iss = syncookie_generate(sch, sc); else sc->sc_iss = arc4random(); #ifdef INET6 if (autoflowlabel) { if (V_tcp_syncookies) sc->sc_flowlabel = sc->sc_iss; else sc->sc_flowlabel = ip6_randomflowlabel(); sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; } #endif if (locked) SCH_UNLOCK(sch); if (tfo_cookie_valid) { rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie); /* INP_RUNLOCK(inp) will be performed by the caller */ goto tfo_expanded; } TCP_PROBE5(receive, NULL, NULL, m, NULL, th); /* * Do a standard 3-way handshake. */ if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) syncache_free(sc); else if (sc != &scs) syncache_insert(sc, sch); /* locks and unlocks sch */ TCPSTAT_INC(tcps_sndacks); TCPSTAT_INC(tcps_sndtotal); } else { if (sc != &scs) syncache_free(sc); TCPSTAT_INC(tcps_sc_dropped); } goto donenoprobe; done: TCP_PROBE5(receive, NULL, NULL, m, NULL, th); donenoprobe: if (m) m_freem(m); /* * If tfo_pending is not NULL here, then a TFO SYN that did not * result in a new socket was processed and the associated pending * counter has not yet been decremented. All such TFO processing paths * transit this point. */ if (tfo_pending != NULL) tcp_fastopen_decrement_counter(tfo_pending); tfo_expanded: if (cred != NULL) crfree(cred); #ifdef MAC if (sc == &scs) mac_syncache_destroy(&maclabel); #endif return (rv); } /* * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment, * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. */ static int syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags) { struct ip *ip = NULL; struct mbuf *m; struct tcphdr *th = NULL; + struct udphdr *udp = NULL; int optlen, error = 0; /* Make compiler happy */ - u_int16_t hlen, tlen, mssopt; + u_int16_t hlen, tlen, mssopt, ulen; struct tcpopt to; #ifdef INET6 struct ip6_hdr *ip6 = NULL; #endif NET_EPOCH_ASSERT(); hlen = #ifdef INET6 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : #endif sizeof(struct ip); tlen = hlen + sizeof(struct tcphdr); - + if (sc->sc_port) { + tlen += sizeof(struct udphdr); + } /* Determine MSS we advertize to other end of connection. */ - mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss); + mssopt = tcp_mssopt(&sc->sc_inc); + if (sc->sc_port) + mssopt -= V_tcp_udp_tunneling_overhead; + mssopt = max(mssopt, V_tcp_minmss); /* XXX: Assume that the entire packet will fit in a header mbuf. */ KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, ("syncache: mbuf too small")); /* Create the IP+TCP header from scratch. */ m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) return (ENOBUFS); #ifdef MAC mac_syncache_create_mbuf(sc->sc_label, m); #endif m->m_data += max_linkhdr; m->m_len = tlen; m->m_pkthdr.len = tlen; m->m_pkthdr.rcvif = NULL; #ifdef INET6 if (sc->sc_inc.inc_flags & INC_ISIPV6) { ip6 = mtod(m, struct ip6_hdr *); ip6->ip6_vfc = IPV6_VERSION; - ip6->ip6_nxt = IPPROTO_TCP; ip6->ip6_src = sc->sc_inc.inc6_laddr; ip6->ip6_dst = sc->sc_inc.inc6_faddr; ip6->ip6_plen = htons(tlen - hlen); /* ip6_hlim is set after checksum */ /* Zero out traffic class and flow label. */ ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK; ip6->ip6_flow |= sc->sc_flowlabel; + if (sc->sc_port != 0) { + ip6->ip6_nxt = IPPROTO_UDP; + udp = (struct udphdr *)(ip6 + 1); + udp->uh_sport = htons(V_tcp_udp_tunneling_port); + udp->uh_dport = sc->sc_port; + ulen = (tlen - sizeof(struct ip6_hdr)); + th = (struct tcphdr *)(udp + 1); + } else { + ip6->ip6_nxt = IPPROTO_TCP; + th = (struct tcphdr *)(ip6 + 1); + } ip6->ip6_flow |= htonl(sc->sc_ip_tos << 20); - - th = (struct tcphdr *)(ip6 + 1); } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { ip = mtod(m, struct ip *); ip->ip_v = IPVERSION; ip->ip_hl = sizeof(struct ip) >> 2; ip->ip_len = htons(tlen); ip->ip_id = 0; ip->ip_off = 0; ip->ip_sum = 0; - ip->ip_p = IPPROTO_TCP; ip->ip_src = sc->sc_inc.inc_laddr; ip->ip_dst = sc->sc_inc.inc_faddr; ip->ip_ttl = sc->sc_ip_ttl; ip->ip_tos = sc->sc_ip_tos; /* * See if we should do MTU discovery. Route lookups are * expensive, so we will only unset the DF bit if: * * 1) path_mtu_discovery is disabled * 2) the SCF_UNREACH flag has been set */ if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) ip->ip_off |= htons(IP_DF); - - th = (struct tcphdr *)(ip + 1); + if (sc->sc_port == 0) { + ip->ip_p = IPPROTO_TCP; + th = (struct tcphdr *)(ip + 1); + } else { + ip->ip_p = IPPROTO_UDP; + udp = (struct udphdr *)(ip + 1); + udp->uh_sport = htons(V_tcp_udp_tunneling_port); + udp->uh_dport = sc->sc_port; + ulen = (tlen - sizeof(struct ip)); + th = (struct tcphdr *)(udp + 1); + } } #endif /* INET */ th->th_sport = sc->sc_inc.inc_lport; th->th_dport = sc->sc_inc.inc_fport; if (flags & TH_SYN) th->th_seq = htonl(sc->sc_iss); else th->th_seq = htonl(sc->sc_iss + 1); th->th_ack = htonl(sc->sc_irs + 1); th->th_off = sizeof(struct tcphdr) >> 2; th->th_x2 = 0; th->th_flags = flags; th->th_win = htons(sc->sc_wnd); th->th_urp = 0; if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) { th->th_flags |= TH_ECE; TCPSTAT_INC(tcps_ecn_shs); } /* Tack on the TCP options. */ if ((sc->sc_flags & SCF_NOOPT) == 0) { to.to_flags = 0; if (flags & TH_SYN) { to.to_mss = mssopt; to.to_flags = TOF_MSS; if (sc->sc_flags & SCF_WINSCALE) { to.to_wscale = sc->sc_requested_r_scale; to.to_flags |= TOF_SCALE; } if (sc->sc_flags & SCF_SACK) to.to_flags |= TOF_SACKPERM; #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (sc->sc_flags & SCF_SIGNATURE) to.to_flags |= TOF_SIGNATURE; #endif if (sc->sc_tfo_cookie) { to.to_flags |= TOF_FASTOPEN; to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; to.to_tfo_cookie = sc->sc_tfo_cookie; /* don't send cookie again when retransmitting response */ sc->sc_tfo_cookie = NULL; } } if (sc->sc_flags & SCF_TIMESTAMP) { to.to_tsval = sc->sc_tsoff + tcp_ts_getticks(); to.to_tsecr = sc->sc_tsreflect; to.to_flags |= TOF_TS; } optlen = tcp_addoptions(&to, (u_char *)(th + 1)); /* Adjust headers by option size. */ th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; m->m_len += optlen; m->m_pkthdr.len += optlen; #ifdef INET6 if (sc->sc_inc.inc_flags & INC_ISIPV6) ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); else #endif ip->ip_len = htons(ntohs(ip->ip_len) + optlen); #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (sc->sc_flags & SCF_SIGNATURE) { KASSERT(to.to_flags & TOF_SIGNATURE, ("tcp_addoptions() didn't set tcp_signature")); /* NOTE: to.to_signature is inside of mbuf */ if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { m_freem(m); return (EACCES); } } #endif } else optlen = 0; + if (udp) { + ulen += optlen; + udp->uh_ulen = htons(ulen); + } M_SETFIB(m, sc->sc_inc.inc_fibnum); - m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); /* * If we have peer's SYN and it has a flowid, then let's assign it to * our SYN|ACK. ip6_output() and ip_output() will not assign flowid * to SYN|ACK due to lack of inp here. */ if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { m->m_pkthdr.flowid = m0->m_pkthdr.flowid; M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); } #ifdef INET6 if (sc->sc_inc.inc_flags & INC_ISIPV6) { - m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; - th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, - IPPROTO_TCP, 0); + if (sc->sc_port) { + m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; + m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); + udp->uh_sum = in6_cksum_pseudo(ip6, ulen, + IPPROTO_UDP, 0); + th->th_sum = htons(0); + } else { + m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; + m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); + th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, + IPPROTO_TCP, 0); + } ip6->ip6_hlim = sc->sc_ip_ttl; #ifdef TCP_OFFLOAD if (ADDED_BY_TOE(sc)) { struct toedev *tod = sc->sc_tod; error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); return (error); } #endif TCP_PROBE5(send, NULL, NULL, ip6, NULL, th); error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { - m->m_pkthdr.csum_flags = CSUM_TCP; - th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, - htons(tlen + optlen - hlen + IPPROTO_TCP)); + if (sc->sc_port) { + m->m_pkthdr.csum_flags = CSUM_UDP; + m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); + udp->uh_sum = in_pseudo(ip->ip_src.s_addr, + ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); + th->th_sum = htons(0); + } else { + m->m_pkthdr.csum_flags = CSUM_TCP; + m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); + th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, + htons(tlen + optlen - hlen + IPPROTO_TCP)); + } #ifdef TCP_OFFLOAD if (ADDED_BY_TOE(sc)) { struct toedev *tod = sc->sc_tod; error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); return (error); } #endif TCP_PROBE5(send, NULL, NULL, ip, NULL, th); error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); } #endif return (error); } /* * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks * that exceed the capacity of the syncache by avoiding the storage of any * of the SYNs we receive. Syncookies defend against blind SYN flooding * attacks where the attacker does not have access to our responses. * * Syncookies encode and include all necessary information about the * connection setup within the SYN|ACK that we send back. That way we * can avoid keeping any local state until the ACK to our SYN|ACK returns * (if ever). Normally the syncache and syncookies are running in parallel * with the latter taking over when the former is exhausted. When matching * syncache entry is found the syncookie is ignored. * * The only reliable information persisting the 3WHS is our initial sequence * number ISS of 32 bits. Syncookies embed a cryptographically sufficient * strong hash (MAC) value and a few bits of TCP SYN options in the ISS * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK * returns and signifies a legitimate connection if it matches the ACK. * * The available space of 32 bits to store the hash and to encode the SYN * option information is very tight and we should have at least 24 bits for * the MAC to keep the number of guesses by blind spoofing reasonably high. * * SYN option information we have to encode to fully restore a connection: * MSS: is imporant to chose an optimal segment size to avoid IP level * fragmentation along the path. The common MSS values can be encoded * in a 3-bit table. Uncommon values are captured by the next lower value * in the table leading to a slight increase in packetization overhead. * WSCALE: is necessary to allow large windows to be used for high delay- * bandwidth product links. Not scaling the window when it was initially * negotiated is bad for performance as lack of scaling further decreases * the apparent available send window. We only need to encode the WSCALE * we received from the remote end. Our end can be recalculated at any * time. The common WSCALE values can be encoded in a 3-bit table. * Uncommon values are captured by the next lower value in the table * making us under-estimate the available window size halving our * theoretically possible maximum throughput for that connection. * SACK: Greatly assists in packet loss recovery and requires 1 bit. * TIMESTAMP and SIGNATURE is not encoded because they are permanent options * that are included in all segments on a connection. We enable them when * the ACK has them. * * Security of syncookies and attack vectors: * * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) * together with the gloabl secret to make it unique per connection attempt. * Thus any change of any of those parameters results in a different MAC output * in an unpredictable way unless a collision is encountered. 24 bits of the * MAC are embedded into the ISS. * * To prevent replay attacks two rotating global secrets are updated with a * new random value every 15 seconds. The life-time of a syncookie is thus * 15-30 seconds. * * Vector 1: Attacking the secret. This requires finding a weakness in the * MAC itself or the way it is used here. The attacker can do a chosen plain * text attack by varying and testing the all parameters under his control. * The strength depends on the size and randomness of the secret, and the * cryptographic security of the MAC function. Due to the constant updating * of the secret the attacker has at most 29.999 seconds to find the secret * and launch spoofed connections. After that he has to start all over again. * * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC * size an average of 4,823 attempts are required for a 50% chance of success * to spoof a single syncookie (birthday collision paradox). However the * attacker is blind and doesn't know if one of his attempts succeeded unless * he has a side channel to interfere success from. A single connection setup * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. * This many attempts are required for each one blind spoofed connection. For * every additional spoofed connection he has to launch another N attempts. * Thus for a sustained rate 100 spoofed connections per second approximately * 1,800,000 packets per second would have to be sent. * * NB: The MAC function should be fast so that it doesn't become a CPU * exhaustion attack vector itself. * * References: * RFC4987 TCP SYN Flooding Attacks and Common Mitigations * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 * http://cr.yp.to/syncookies.html (overview) * http://cr.yp.to/syncookies/archive (details) * * * Schematic construction of a syncookie enabled Initial Sequence Number: * 0 1 2 3 * 12345678901234567890123456789012 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| * * x 24 MAC (truncated) * W 3 Send Window Scale index * M 3 MSS index * S 1 SACK permitted * P 1 Odd/even secret */ /* * Distribution and probability of certain MSS values. Those in between are * rounded down to the next lower one. * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] * .2% .3% 5% 7% 7% 20% 15% 45% */ static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; /* * Distribution and probability of certain WSCALE values. We have to map the * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 * bits based on prevalence of certain values. Where we don't have an exact * match for are rounded down to the next lower one letting us under-estimate * the true available window. At the moment this would happen only for the * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer * and window size). The absence of the WSCALE option (no scaling in either * direction) is encoded with index zero. * [WSCALE values histograms, Allman, 2012] * X 10 10 35 5 6 14 10% by host * X 11 4 5 5 18 49 3% by connections */ static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; /* * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed * and good cryptographic properties. */ static uint32_t syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, uint8_t *secbits, uintptr_t secmod) { SIPHASH_CTX ctx; uint32_t siphash[2]; SipHash24_Init(&ctx); SipHash_SetKey(&ctx, secbits); switch (inc->inc_flags & INC_ISIPV6) { #ifdef INET case 0: SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); break; #endif #ifdef INET6 case INC_ISIPV6: SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); break; #endif } SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); SipHash_Update(&ctx, &irs, sizeof(irs)); SipHash_Update(&ctx, &flags, sizeof(flags)); SipHash_Update(&ctx, &secmod, sizeof(secmod)); SipHash_Final((u_int8_t *)&siphash, &ctx); return (siphash[0] ^ siphash[1]); } static tcp_seq syncookie_generate(struct syncache_head *sch, struct syncache *sc) { u_int i, secbit, wscale; uint32_t iss, hash; uint8_t *secbits; union syncookie cookie; cookie.cookie = 0; /* Map our computed MSS into the 3-bit index. */ for (i = nitems(tcp_sc_msstab) - 1; tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0; i--) ; cookie.flags.mss_idx = i; /* * Map the send window scale into the 3-bit index but only if * the wscale option was received. */ if (sc->sc_flags & SCF_WINSCALE) { wscale = sc->sc_requested_s_scale; for (i = nitems(tcp_sc_wstab) - 1; tcp_sc_wstab[i] > wscale && i > 0; i--) ; cookie.flags.wscale_idx = i; } /* Can we do SACK? */ if (sc->sc_flags & SCF_SACK) cookie.flags.sack_ok = 1; /* Which of the two secrets to use. */ secbit = V_tcp_syncache.secret.oddeven & 0x1; cookie.flags.odd_even = secbit; secbits = V_tcp_syncache.secret.key[secbit]; hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, (uintptr_t)sch); /* * Put the flags into the hash and XOR them to get better ISS number * variance. This doesn't enhance the cryptographic strength and is * done to prevent the 8 cookie bits from showing up directly on the * wire. */ iss = hash & ~0xff; iss |= cookie.cookie ^ (hash >> 24); TCPSTAT_INC(tcps_sc_sendcookie); return (iss); } static struct syncache * syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, struct syncache *sc, struct tcphdr *th, struct tcpopt *to, - struct socket *lso) + struct socket *lso, uint16_t port) { uint32_t hash; uint8_t *secbits; tcp_seq ack, seq; int wnd, wscale = 0; union syncookie cookie; /* * Pull information out of SYN-ACK/ACK and revert sequence number * advances. */ ack = th->th_ack - 1; seq = th->th_seq - 1; /* * Unpack the flags containing enough information to restore the * connection. */ cookie.cookie = (ack & 0xff) ^ (ack >> 24); /* Which of the two secrets to use. */ secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even]; hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); /* The recomputed hash matches the ACK if this was a genuine cookie. */ if ((ack & ~0xff) != (hash & ~0xff)) return (NULL); /* Fill in the syncache values. */ sc->sc_flags = 0; bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); sc->sc_ipopts = NULL; sc->sc_irs = seq; sc->sc_iss = ack; switch (inc->inc_flags & INC_ISIPV6) { #ifdef INET case 0: sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; break; #endif #ifdef INET6 case INC_ISIPV6: if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) sc->sc_flowlabel = htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK; break; #endif } sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; /* We can simply recompute receive window scale we sent earlier. */ while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) wscale++; /* Only use wscale if it was enabled in the orignal SYN. */ if (cookie.flags.wscale_idx > 0) { sc->sc_requested_r_scale = wscale; sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; sc->sc_flags |= SCF_WINSCALE; } wnd = lso->sol_sbrcv_hiwat; wnd = imax(wnd, 0); wnd = imin(wnd, TCP_MAXWIN); sc->sc_wnd = wnd; if (cookie.flags.sack_ok) sc->sc_flags |= SCF_SACK; if (to->to_flags & TOF_TS) { sc->sc_flags |= SCF_TIMESTAMP; sc->sc_tsreflect = to->to_tsval; sc->sc_tsoff = tcp_new_ts_offset(inc); } if (to->to_flags & TOF_SIGNATURE) sc->sc_flags |= SCF_SIGNATURE; sc->sc_rxmits = 0; + sc->sc_port = port; + TCPSTAT_INC(tcps_sc_recvcookie); return (sc); } #ifdef INVARIANTS static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, struct syncache *sc, struct tcphdr *th, struct tcpopt *to, - struct socket *lso) + struct socket *lso, uint16_t port) { struct syncache scs, *scx; char *s; bzero(&scs, sizeof(scs)); - scx = syncookie_lookup(inc, sch, &scs, th, to, lso); + scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port); if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) return (0); if (scx != NULL) { if (sc->sc_peer_mss != scx->sc_peer_mss) log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", s, __func__, sc->sc_requested_r_scale, scx->sc_requested_r_scale); if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", s, __func__, sc->sc_requested_s_scale, scx->sc_requested_s_scale); if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); } if (s != NULL) free(s, M_TCPLOG); return (0); } #endif /* INVARIANTS */ static void syncookie_reseed(void *arg) { struct tcp_syncache *sc = arg; uint8_t *secbits; int secbit; /* * Reseeding the secret doesn't have to be protected by a lock. * It only must be ensured that the new random values are visible * to all CPUs in a SMP environment. The atomic with release * semantics ensures that. */ secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; secbits = sc->secret.key[secbit]; arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); atomic_add_rel_int(&sc->secret.oddeven, 1); /* Reschedule ourself. */ callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); } /* * We have overflowed a bucket. Let's pause dealing with the syncache. * This function will increment the bucketoverflow statistics appropriately * (once per pause when pausing is enabled; otherwise, once per overflow). */ static void syncache_pause(struct in_conninfo *inc) { time_t delta; const char *s; /* XXX: * 2. Add sysctl read here so we don't get the benefit of this * change without the new sysctl. */ /* * Try an unlocked read. If we already know that another thread * has activated the feature, there is no need to proceed. */ if (V_tcp_syncache.paused) return; /* Are cookied enabled? If not, we can't pause. */ if (!V_tcp_syncookies) { TCPSTAT_INC(tcps_sc_bucketoverflow); return; } /* * We may be the first thread to find an overflow. Get the lock * and evaluate if we need to take action. */ mtx_lock(&V_tcp_syncache.pause_mtx); if (V_tcp_syncache.paused) { mtx_unlock(&V_tcp_syncache.pause_mtx); return; } /* Activate protection. */ V_tcp_syncache.paused = true; TCPSTAT_INC(tcps_sc_bucketoverflow); /* * Determine the last backoff time. If we are seeing a re-newed * attack within that same time after last reactivating the syncache, * consider it an extension of the same attack. */ delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff; if (V_tcp_syncache.pause_until + delta - time_uptime > 0) { if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) { delta <<= 1; V_tcp_syncache.pause_backoff++; } } else { delta = TCP_SYNCACHE_PAUSE_TIME; V_tcp_syncache.pause_backoff = 0; } /* Log a warning, including IP addresses, if able. */ if (inc != NULL) s = tcp_log_addrs(inc, NULL, NULL, NULL); else s = (const char *)NULL; log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for " "the next %lld seconds%s%s%s\n", (long long)delta, (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "", (s != NULL) ? ")" : ""); free(__DECONST(void *, s), M_TCPLOG); /* Use the calculated delta to set a new pause time. */ V_tcp_syncache.pause_until = time_uptime + delta; callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause, &V_tcp_syncache); mtx_unlock(&V_tcp_syncache.pause_mtx); } /* Evaluate whether we need to unpause. */ static void syncache_unpause(void *arg) { struct tcp_syncache *sc; time_t delta; sc = arg; mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED); callout_deactivate(&sc->pause_co); /* * Check to make sure we are not running early. If the pause * time has expired, then deactivate the protection. */ if ((delta = sc->pause_until - time_uptime) > 0) callout_schedule(&sc->pause_co, delta * hz); else sc->paused = false; } /* * Exports the syncache entries to userland so that netstat can display * them alongside the other sockets. This function is intended to be * called only from tcp_pcblist. * * Due to concurrency on an active system, the number of pcbs exported * may have no relation to max_pcbs. max_pcbs merely indicates the * amount of space the caller allocated for this function to use. */ int syncache_pcblist(struct sysctl_req *req) { struct xtcpcb xt; struct syncache *sc; struct syncache_head *sch; int error, i; bzero(&xt, sizeof(xt)); xt.xt_len = sizeof(xt); xt.t_state = TCPS_SYN_RECEIVED; xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket); xt.xt_inp.xi_socket.so_type = SOCK_STREAM; xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING; for (i = 0; i < V_tcp_syncache.hashsize; i++) { sch = &V_tcp_syncache.hashbase[i]; SCH_LOCK(sch); TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { if (sc->sc_cred != NULL && cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) continue; if (sc->sc_inc.inc_flags & INC_ISIPV6) xt.xt_inp.inp_vflag = INP_IPV6; else xt.xt_inp.inp_vflag = INP_IPV4; + xt.xt_encaps_port = sc->sc_port; bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, sizeof (struct in_conninfo)); error = SYSCTL_OUT(req, &xt, sizeof xt); if (error) { SCH_UNLOCK(sch); return (0); } } SCH_UNLOCK(sch); } return (0); } diff --git a/sys/netinet/tcp_syncache.h b/sys/netinet/tcp_syncache.h index 03e34a89c112..a16a80c483d5 100644 --- a/sys/netinet/tcp_syncache.h +++ b/sys/netinet/tcp_syncache.h @@ -1,152 +1,154 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1993, 1994, 1995 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)tcp_var.h 8.4 (Berkeley) 5/24/95 * $FreeBSD$ */ #ifndef _NETINET_TCP_SYNCACHE_H_ #define _NETINET_TCP_SYNCACHE_H_ #ifdef _KERNEL void syncache_init(void); #ifdef VIMAGE void syncache_destroy(void); #endif -void syncache_unreach(struct in_conninfo *, tcp_seq); +void syncache_unreach(struct in_conninfo *, tcp_seq, uint16_t); int syncache_expand(struct in_conninfo *, struct tcpopt *, - struct tcphdr *, struct socket **, struct mbuf *); + struct tcphdr *, struct socket **, struct mbuf *, uint16_t); struct socket * syncache_add(struct in_conninfo *, struct tcpopt *, struct tcphdr *, struct inpcb *, struct socket *, struct mbuf *, - void *, void *, uint8_t); -void syncache_chkrst(struct in_conninfo *, struct tcphdr *, struct mbuf *); -void syncache_badack(struct in_conninfo *); + void *, void *, uint8_t, uint16_t); +void syncache_chkrst(struct in_conninfo *, struct tcphdr *, struct mbuf *, + uint16_t); +void syncache_badack(struct in_conninfo *, uint16_t); int syncache_pcblist(struct sysctl_req *); struct syncache { TAILQ_ENTRY(syncache) sc_hash; struct in_conninfo sc_inc; /* addresses */ int sc_rxttime; /* retransmit time */ u_int16_t sc_rxmits; /* retransmit counter */ + u_int16_t sc_port; /* remote UDP encaps port */ u_int32_t sc_tsreflect; /* timestamp to reflect */ u_int32_t sc_tsoff; /* ts offset w/ syncookies */ u_int32_t sc_flowlabel; /* IPv6 flowlabel */ tcp_seq sc_irs; /* seq from peer */ tcp_seq sc_iss; /* our ISS */ struct mbuf *sc_ipopts; /* source route */ u_int16_t sc_peer_mss; /* peer's MSS */ u_int16_t sc_wnd; /* advertised window */ u_int8_t sc_ip_ttl; /* TTL / Hop Limit */ u_int8_t sc_ip_tos; /* TOS / Traffic Class */ u_int8_t sc_requested_s_scale:4, sc_requested_r_scale:4; u_int16_t sc_flags; #if defined(TCP_OFFLOAD) || !defined(TCP_OFFLOAD_DISABLE) struct toedev *sc_tod; /* entry added by this TOE */ void *sc_todctx; /* TOE driver context */ #endif struct label *sc_label; /* MAC label reference */ struct ucred *sc_cred; /* cred cache for jail checks */ void *sc_tfo_cookie; /* for TCP Fast Open response */ void *sc_pspare; /* TCP_SIGNATURE */ u_int32_t sc_spare[2]; /* UTO */ }; /* * Flags for the sc_flags field. */ #define SCF_NOOPT 0x01 /* no TCP options */ #define SCF_WINSCALE 0x02 /* negotiated window scaling */ #define SCF_TIMESTAMP 0x04 /* negotiated timestamps */ /* MSS is implicit */ #define SCF_UNREACH 0x10 /* icmp unreachable received */ #define SCF_SIGNATURE 0x20 /* send MD5 digests */ #define SCF_SACK 0x80 /* send SACK option */ #define SCF_ECN 0x100 /* send ECN setup packet */ #define SCF_ACE_N 0x200 /* send ACE non-ECT setup */ #define SCF_ACE_0 0x400 /* send ACE ECT0 setup */ #define SCF_ACE_1 0x800 /* send ACE ECT1 setup */ #define SCF_ACE_CE 0x1000 /* send ACE CE setup */ struct syncache_head { struct mtx sch_mtx; TAILQ_HEAD(sch_head, syncache) sch_bucket; struct callout sch_timer; int sch_nextc; u_int sch_length; struct tcp_syncache *sch_sc; time_t sch_last_overflow; }; #define SYNCOOKIE_SECRET_SIZE 16 #define SYNCOOKIE_LIFETIME 15 /* seconds */ struct syncookie_secret { volatile u_int oddeven; uint8_t key[2][SYNCOOKIE_SECRET_SIZE]; struct callout reseed; u_int lifetime; }; #define TCP_SYNCACHE_PAUSE_TIME SYNCOOKIE_LIFETIME #define TCP_SYNCACHE_MAX_BACKOFF 6 /* 16 minutes */ struct tcp_syncache { struct syncache_head *hashbase; uma_zone_t zone; u_int hashsize; u_int hashmask; u_int bucket_limit; u_int cache_limit; u_int rexmt_limit; uint32_t hash_secret; struct vnet *vnet; struct syncookie_secret secret; struct mtx pause_mtx; struct callout pause_co; time_t pause_until; uint8_t pause_backoff; volatile bool paused; bool see_other; }; /* Internal use for the syncookie functions. */ union syncookie { uint8_t cookie; struct { uint8_t odd_even:1, sack_ok:1, wscale_idx:3, mss_idx:3; } flags; }; #endif /* _KERNEL */ #endif /* !_NETINET_TCP_SYNCACHE_H_ */ diff --git a/sys/netinet/tcp_timewait.c b/sys/netinet/tcp_timewait.c index f98927b196fc..b62386ddca05 100644 --- a/sys/netinet/tcp_timewait.c +++ b/sys/netinet/tcp_timewait.c @@ -1,762 +1,822 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95 */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_tcpdebug.h" #include #include #include #include #include #include #include #include #include #include #include #ifndef INVARIANTS #include #endif #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef INET6 #include #include #include #include #include #endif #include #include #include #include #include #ifdef INET6 #include #endif #include #ifdef TCPDEBUG #include #endif #ifdef INET6 #include #endif +#include +#include #include #include VNET_DEFINE_STATIC(uma_zone_t, tcptw_zone); #define V_tcptw_zone VNET(tcptw_zone) static int maxtcptw; /* * The timed wait queue contains references to each of the TCP sessions * currently in the TIME_WAIT state. The queue pointers, including the * queue pointers in each tcptw structure, are protected using the global * timewait lock, which must be held over queue iteration and modification. * * Rules on tcptw usage: * - a inpcb is always freed _after_ its tcptw * - a tcptw relies on its inpcb reference counting for memory stability * - a tcptw is dereferenceable only while its inpcb is locked */ VNET_DEFINE_STATIC(TAILQ_HEAD(, tcptw), twq_2msl); #define V_twq_2msl VNET(twq_2msl) /* Global timewait lock */ VNET_DEFINE_STATIC(struct rwlock, tw_lock); #define V_tw_lock VNET(tw_lock) #define TW_LOCK_INIT(tw, d) rw_init_flags(&(tw), (d), 0) #define TW_LOCK_DESTROY(tw) rw_destroy(&(tw)) #define TW_RLOCK(tw) rw_rlock(&(tw)) #define TW_WLOCK(tw) rw_wlock(&(tw)) #define TW_RUNLOCK(tw) rw_runlock(&(tw)) #define TW_WUNLOCK(tw) rw_wunlock(&(tw)) #define TW_LOCK_ASSERT(tw) rw_assert(&(tw), RA_LOCKED) #define TW_RLOCK_ASSERT(tw) rw_assert(&(tw), RA_RLOCKED) #define TW_WLOCK_ASSERT(tw) rw_assert(&(tw), RA_WLOCKED) #define TW_UNLOCK_ASSERT(tw) rw_assert(&(tw), RA_UNLOCKED) static void tcp_tw_2msl_reset(struct tcptw *, int); static void tcp_tw_2msl_stop(struct tcptw *, int); static int tcp_twrespond(struct tcptw *, int); static int tcptw_auto_size(void) { int halfrange; /* * Max out at half the ephemeral port range so that TIME_WAIT * sockets don't tie up too many ephemeral ports. */ if (V_ipport_lastauto > V_ipport_firstauto) halfrange = (V_ipport_lastauto - V_ipport_firstauto) / 2; else halfrange = (V_ipport_firstauto - V_ipport_lastauto) / 2; /* Protect against goofy port ranges smaller than 32. */ return (imin(imax(halfrange, 32), maxsockets / 5)); } static int sysctl_maxtcptw(SYSCTL_HANDLER_ARGS) { int error, new; if (maxtcptw == 0) new = tcptw_auto_size(); else new = maxtcptw; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr) if (new >= 32) { maxtcptw = new; uma_zone_set_max(V_tcptw_zone, maxtcptw); } return (error); } SYSCTL_PROC(_net_inet_tcp, OID_AUTO, maxtcptw, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &maxtcptw, 0, sysctl_maxtcptw, "IU", "Maximum number of compressed TCP TIME_WAIT entries"); VNET_DEFINE_STATIC(int, nolocaltimewait) = 0; #define V_nolocaltimewait VNET(nolocaltimewait) SYSCTL_INT(_net_inet_tcp, OID_AUTO, nolocaltimewait, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(nolocaltimewait), 0, "Do not create compressed TCP TIME_WAIT entries for local connections"); void tcp_tw_zone_change(void) { if (maxtcptw == 0) uma_zone_set_max(V_tcptw_zone, tcptw_auto_size()); } void tcp_tw_init(void) { V_tcptw_zone = uma_zcreate("tcptw", sizeof(struct tcptw), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); TUNABLE_INT_FETCH("net.inet.tcp.maxtcptw", &maxtcptw); if (maxtcptw == 0) uma_zone_set_max(V_tcptw_zone, tcptw_auto_size()); else uma_zone_set_max(V_tcptw_zone, maxtcptw); TAILQ_INIT(&V_twq_2msl); TW_LOCK_INIT(V_tw_lock, "tcptw"); } #ifdef VIMAGE void tcp_tw_destroy(void) { struct tcptw *tw; struct epoch_tracker et; NET_EPOCH_ENTER(et); while ((tw = TAILQ_FIRST(&V_twq_2msl)) != NULL) tcp_twclose(tw, 0); NET_EPOCH_EXIT(et); TW_LOCK_DESTROY(V_tw_lock); uma_zdestroy(V_tcptw_zone); } #endif /* * Move a TCP connection into TIME_WAIT state. * tcbinfo is locked. * inp is locked, and is unlocked before returning. */ void tcp_twstart(struct tcpcb *tp) { struct tcptw twlocal, *tw; struct inpcb *inp = tp->t_inpcb; struct socket *so; uint32_t recwin; bool acknow, local; #ifdef INET6 bool isipv6 = inp->inp_inc.inc_flags & INC_ISIPV6; #endif NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(inp); /* A dropped inp should never transition to TIME_WAIT state. */ KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("tcp_twstart: " "(inp->inp_flags & INP_DROPPED) != 0")); if (V_nolocaltimewait) { #ifdef INET6 if (isipv6) local = in6_localaddr(&inp->in6p_faddr); else #endif #ifdef INET local = in_localip(inp->inp_faddr); #else local = false; #endif } else local = false; /* * For use only by DTrace. We do not reference the state * after this point so modifying it in place is not a problem. */ tcp_state_change(tp, TCPS_TIME_WAIT); if (local) tw = &twlocal; else tw = uma_zalloc(V_tcptw_zone, M_NOWAIT); if (tw == NULL) { /* * Reached limit on total number of TIMEWAIT connections * allowed. Remove a connection from TIMEWAIT queue in LRU * fashion to make room for this connection. * * XXX: Check if it possible to always have enough room * in advance based on guarantees provided by uma_zalloc(). */ tw = tcp_tw_2msl_scan(1); if (tw == NULL) { tp = tcp_close(tp); if (tp != NULL) INP_WUNLOCK(inp); return; } } /* * For !local case the tcptw will hold a reference on its inpcb * until tcp_twclose is called. */ tw->tw_inpcb = inp; /* * Recover last window size sent. */ so = inp->inp_socket; recwin = lmin(lmax(sbspace(&so->so_rcv), 0), (long)TCP_MAXWIN << tp->rcv_scale); if (recwin < (so->so_rcv.sb_hiwat / 4) && recwin < tp->t_maxseg) recwin = 0; if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && recwin < (tp->rcv_adv - tp->rcv_nxt)) recwin = (tp->rcv_adv - tp->rcv_nxt); tw->last_win = (u_short)(recwin >> tp->rcv_scale); /* * Set t_recent if timestamps are used on the connection. */ if ((tp->t_flags & (TF_REQ_TSTMP|TF_RCVD_TSTMP|TF_NOOPT)) == (TF_REQ_TSTMP|TF_RCVD_TSTMP)) { tw->t_recent = tp->ts_recent; tw->ts_offset = tp->ts_offset; } else { tw->t_recent = 0; tw->ts_offset = 0; } tw->snd_nxt = tp->snd_nxt; + tw->t_port = tp->t_port; tw->rcv_nxt = tp->rcv_nxt; tw->iss = tp->iss; tw->irs = tp->irs; tw->t_starttime = tp->t_starttime; tw->tw_time = 0; /* XXX * If this code will * be used for fin-wait-2 state also, then we may need * a ts_recent from the last segment. */ acknow = tp->t_flags & TF_ACKNOW; /* * First, discard tcpcb state, which includes stopping its timers and * freeing it. tcp_discardcb() used to also release the inpcb, but * that work is now done in the caller. * * Note: soisdisconnected() call used to be made in tcp_discardcb(), * and might not be needed here any longer. */ tcp_discardcb(tp); soisdisconnected(so); tw->tw_so_options = so->so_options; inp->inp_flags |= INP_TIMEWAIT; if (acknow) tcp_twrespond(tw, TH_ACK); if (local) in_pcbdrop(inp); else { in_pcbref(inp); /* Reference from tw */ tw->tw_cred = crhold(so->so_cred); inp->inp_ppcb = tw; TCPSTATES_INC(TCPS_TIME_WAIT); tcp_tw_2msl_reset(tw, 0); } /* * If the inpcb owns the sole reference to the socket, then we can * detach and free the socket as it is not needed in time wait. */ if (inp->inp_flags & INP_SOCKREF) { KASSERT(so->so_state & SS_PROTOREF, ("tcp_twstart: !SS_PROTOREF")); inp->inp_flags &= ~INP_SOCKREF; INP_WUNLOCK(inp); SOCK_LOCK(so); so->so_state &= ~SS_PROTOREF; sofree(so); } else INP_WUNLOCK(inp); } /* * Returns 1 if the TIME_WAIT state was killed and we should start over, * looking for a pcb in the listen state. Returns 0 otherwise. * It be called with to == NULL only for pure SYN-segments. */ int tcp_twcheck(struct inpcb *inp, struct tcpopt *to, struct tcphdr *th, struct mbuf *m, int tlen) { struct tcptw *tw; int thflags; tcp_seq seq; NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(inp); /* * XXXRW: Time wait state for inpcb has been recycled, but inpcb is * still present. This is undesirable, but temporarily necessary * until we work out how to handle inpcb's who's timewait state has * been removed. */ tw = intotw(inp); if (tw == NULL) goto drop; thflags = th->th_flags; KASSERT(to != NULL || (thflags & (TH_SYN | TH_ACK)) == TH_SYN, ("tcp_twcheck: called without options on a non-SYN segment")); /* * NOTE: for FIN_WAIT_2 (to be added later), * must validate sequence number before accepting RST */ /* * If the segment contains RST: * Drop the segment - see Stevens, vol. 2, p. 964 and * RFC 1337. */ if (thflags & TH_RST) goto drop; #if 0 /* PAWS not needed at the moment */ /* * RFC 1323 PAWS: If we have a timestamp reply on this segment * and it's less than ts_recent, drop it. */ if ((to.to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to.to_tsval, tp->ts_recent)) { if ((thflags & TH_ACK) == 0) goto drop; goto ack; } /* * ts_recent is never updated because we never accept new segments. */ #endif /* * If a new connection request is received * while in TIME_WAIT, drop the old connection * and start over if the sequence numbers * are above the previous ones. + * Allow UDP port number changes in this case. */ if ((thflags & TH_SYN) && SEQ_GT(th->th_seq, tw->rcv_nxt)) { tcp_twclose(tw, 0); return (1); } + /* + * Send RST if UDP port numbers don't match + */ + if (tw->t_port != m->m_pkthdr.tcp_tun_port) { + if (th->th_flags & TH_ACK) { + tcp_respond(NULL, mtod(m, void *), th, m, + (tcp_seq)0, th->th_ack, TH_RST); + } else { + if (th->th_flags & TH_SYN) + tlen++; + if (th->th_flags & TH_FIN) + tlen++; + tcp_respond(NULL, mtod(m, void *), th, m, + th->th_seq+tlen, (tcp_seq)0, TH_RST|TH_ACK); + } + INP_WUNLOCK(inp); + return (0); + } + /* * Drop the segment if it does not contain an ACK. */ if ((thflags & TH_ACK) == 0) goto drop; /* * If timestamps were negotiated during SYN/ACK and a * segment without a timestamp is received, silently drop * the segment, unless the missing timestamps are tolerated. * See section 3.2 of RFC 7323. */ if (((to->to_flags & TOF_TS) == 0) && (tw->t_recent != 0) && (V_tcp_tolerate_missing_ts == 0)) { goto drop; } /* * Reset the 2MSL timer if this is a duplicate FIN. */ if (thflags & TH_FIN) { seq = th->th_seq + tlen + (thflags & TH_SYN ? 1 : 0); if (seq + 1 == tw->rcv_nxt) tcp_tw_2msl_reset(tw, 1); } /* * Acknowledge the segment if it has data or is not a duplicate ACK. */ if (thflags != TH_ACK || tlen != 0 || th->th_seq != tw->rcv_nxt || th->th_ack != tw->snd_nxt) { TCP_PROBE5(receive, NULL, NULL, m, NULL, th); tcp_twrespond(tw, TH_ACK); goto dropnoprobe; } drop: TCP_PROBE5(receive, NULL, NULL, m, NULL, th); dropnoprobe: INP_WUNLOCK(inp); m_freem(m); return (0); } void tcp_twclose(struct tcptw *tw, int reuse) { struct socket *so; struct inpcb *inp; /* * At this point, we are in one of two situations: * * (1) We have no socket, just an inpcb<->twtcp pair. We can free * all state. * * (2) We have a socket -- if we own a reference, release it and * notify the socket layer. */ inp = tw->tw_inpcb; KASSERT((inp->inp_flags & INP_TIMEWAIT), ("tcp_twclose: !timewait")); KASSERT(intotw(inp) == tw, ("tcp_twclose: inp_ppcb != tw")); NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(inp); tcp_tw_2msl_stop(tw, reuse); inp->inp_ppcb = NULL; in_pcbdrop(inp); so = inp->inp_socket; if (so != NULL) { /* * If there's a socket, handle two cases: first, we own a * strong reference, which we will now release, or we don't * in which case another reference exists (XXXRW: think * about this more), and we don't need to take action. */ if (inp->inp_flags & INP_SOCKREF) { inp->inp_flags &= ~INP_SOCKREF; INP_WUNLOCK(inp); SOCK_LOCK(so); KASSERT(so->so_state & SS_PROTOREF, ("tcp_twclose: INP_SOCKREF && !SS_PROTOREF")); so->so_state &= ~SS_PROTOREF; sofree(so); } else { /* * If we don't own the only reference, the socket and * inpcb need to be left around to be handled by * tcp_usr_detach() later. */ INP_WUNLOCK(inp); } } else { /* * The socket has been already cleaned-up for us, only free the * inpcb. */ in_pcbfree(inp); } TCPSTAT_INC(tcps_closed); } static int tcp_twrespond(struct tcptw *tw, int flags) { struct inpcb *inp = tw->tw_inpcb; #if defined(INET6) || defined(INET) struct tcphdr *th = NULL; #endif struct mbuf *m; #ifdef INET struct ip *ip = NULL; #endif - u_int hdrlen, optlen; + u_int hdrlen, optlen, ulen; int error = 0; /* Keep compiler happy */ struct tcpopt to; #ifdef INET6 struct ip6_hdr *ip6 = NULL; int isipv6 = inp->inp_inc.inc_flags & INC_ISIPV6; #endif + struct udphdr *udp = NULL; hdrlen = 0; /* Keep compiler happy */ INP_WLOCK_ASSERT(inp); m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) return (ENOBUFS); m->m_data += max_linkhdr; #ifdef MAC mac_inpcb_create_mbuf(inp, m); #endif #ifdef INET6 if (isipv6) { hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); ip6 = mtod(m, struct ip6_hdr *); - th = (struct tcphdr *)(ip6 + 1); - tcpip_fillheaders(inp, ip6, th); + if (tw->t_port) { + udp = (struct udphdr *)(ip6 + 1); + hdrlen += sizeof(struct udphdr); + udp->uh_sport = htons(V_tcp_udp_tunneling_port); + udp->uh_dport = tw->t_port; + ulen = (hdrlen - sizeof(struct ip6_hdr)); + th = (struct tcphdr *)(udp + 1); + } else + th = (struct tcphdr *)(ip6 + 1); + tcpip_fillheaders(inp, tw->t_port, ip6, th); } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { hdrlen = sizeof(struct tcpiphdr); ip = mtod(m, struct ip *); - th = (struct tcphdr *)(ip + 1); - tcpip_fillheaders(inp, ip, th); + if (tw->t_port) { + udp = (struct udphdr *)(ip + 1); + hdrlen += sizeof(struct udphdr); + udp->uh_sport = htons(V_tcp_udp_tunneling_port); + udp->uh_dport = tw->t_port; + ulen = (hdrlen - sizeof(struct ip)); + th = (struct tcphdr *)(udp + 1); + } else + th = (struct tcphdr *)(ip + 1); + tcpip_fillheaders(inp, tw->t_port, ip, th); } #endif to.to_flags = 0; /* * Send a timestamp and echo-reply if both our side and our peer * have sent timestamps in our SYN's and this is not a RST. */ if (tw->t_recent && flags == TH_ACK) { to.to_flags |= TOF_TS; to.to_tsval = tcp_ts_getticks() + tw->ts_offset; to.to_tsecr = tw->t_recent; } optlen = tcp_addoptions(&to, (u_char *)(th + 1)); + if (udp) { + ulen += optlen; + udp->uh_ulen = htons(ulen); + } m->m_len = hdrlen + optlen; m->m_pkthdr.len = m->m_len; KASSERT(max_linkhdr + m->m_len <= MHLEN, ("tcptw: mbuf too small")); th->th_seq = htonl(tw->snd_nxt); th->th_ack = htonl(tw->rcv_nxt); th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; th->th_flags = flags; th->th_win = htons(tw->last_win); - m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); #ifdef INET6 if (isipv6) { - m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; - th->th_sum = in6_cksum_pseudo(ip6, - sizeof(struct tcphdr) + optlen, IPPROTO_TCP, 0); + if (tw->t_port) { + m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; + m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); + udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); + th->th_sum = htons(0); + } else { + m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; + m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); + th->th_sum = in6_cksum_pseudo(ip6, + sizeof(struct tcphdr) + optlen, IPPROTO_TCP, 0); + } ip6->ip6_hlim = in6_selecthlim(inp, NULL); TCP_PROBE5(send, NULL, NULL, ip6, NULL, th); error = ip6_output(m, inp->in6p_outputopts, NULL, (tw->tw_so_options & SO_DONTROUTE), NULL, NULL, inp); } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { - m->m_pkthdr.csum_flags = CSUM_TCP; - th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, - htons(sizeof(struct tcphdr) + optlen + IPPROTO_TCP)); + if (tw->t_port) { + m->m_pkthdr.csum_flags = CSUM_UDP; + m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); + udp->uh_sum = in_pseudo(ip->ip_src.s_addr, + ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); + th->th_sum = htons(0); + } else { + m->m_pkthdr.csum_flags = CSUM_TCP; + m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); + th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, + htons(sizeof(struct tcphdr) + optlen + IPPROTO_TCP)); + } ip->ip_len = htons(m->m_pkthdr.len); if (V_path_mtu_discovery) ip->ip_off |= htons(IP_DF); TCP_PROBE5(send, NULL, NULL, ip, NULL, th); error = ip_output(m, inp->inp_options, NULL, ((tw->tw_so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), NULL, inp); } #endif if (flags & TH_ACK) TCPSTAT_INC(tcps_sndacks); else TCPSTAT_INC(tcps_sndctrl); TCPSTAT_INC(tcps_sndtotal); return (error); } static void tcp_tw_2msl_reset(struct tcptw *tw, int rearm) { NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(tw->tw_inpcb); TW_WLOCK(V_tw_lock); if (rearm) TAILQ_REMOVE(&V_twq_2msl, tw, tw_2msl); tw->tw_time = ticks + 2 * tcp_msl; TAILQ_INSERT_TAIL(&V_twq_2msl, tw, tw_2msl); TW_WUNLOCK(V_tw_lock); } static void tcp_tw_2msl_stop(struct tcptw *tw, int reuse) { struct ucred *cred; struct inpcb *inp; int released __unused; NET_EPOCH_ASSERT(); TW_WLOCK(V_tw_lock); inp = tw->tw_inpcb; tw->tw_inpcb = NULL; TAILQ_REMOVE(&V_twq_2msl, tw, tw_2msl); cred = tw->tw_cred; tw->tw_cred = NULL; TW_WUNLOCK(V_tw_lock); if (cred != NULL) crfree(cred); released = in_pcbrele_wlocked(inp); KASSERT(!released, ("%s: inp should not be released here", __func__)); if (!reuse) uma_zfree(V_tcptw_zone, tw); TCPSTATES_DEC(TCPS_TIME_WAIT); } struct tcptw * tcp_tw_2msl_scan(int reuse) { struct tcptw *tw; struct inpcb *inp; NET_EPOCH_ASSERT(); for (;;) { TW_RLOCK(V_tw_lock); tw = TAILQ_FIRST(&V_twq_2msl); if (tw == NULL || (!reuse && (tw->tw_time - ticks) > 0)) { TW_RUNLOCK(V_tw_lock); break; } KASSERT(tw->tw_inpcb != NULL, ("%s: tw->tw_inpcb == NULL", __func__)); inp = tw->tw_inpcb; in_pcbref(inp); TW_RUNLOCK(V_tw_lock); INP_WLOCK(inp); tw = intotw(inp); if (in_pcbrele_wlocked(inp)) { if (__predict_true(tw == NULL)) { continue; } else { /* This should not happen as in TIMEWAIT * state the inp should not be destroyed * before its tcptw. If INVARIANTS is * defined panic. */ #ifdef INVARIANTS panic("%s: Panic before an infinite " "loop: INP_TIMEWAIT && (INP_FREED " "|| inp last reference) && tw != " "NULL", __func__); #else log(LOG_ERR, "%s: Avoid an infinite " "loop: INP_TIMEWAIT && (INP_FREED " "|| inp last reference) && tw != " "NULL", __func__); #endif break; } } if (tw == NULL) { /* tcp_twclose() has already been called */ INP_WUNLOCK(inp); continue; } tcp_twclose(tw, reuse); if (reuse) return tw; } return NULL; } diff --git a/sys/netinet/tcp_usrreq.c b/sys/netinet/tcp_usrreq.c index 31b580bfafcc..c4cfb5ea199f 100644 --- a/sys/netinet/tcp_usrreq.c +++ b/sys/netinet/tcp_usrreq.c @@ -1,2966 +1,2996 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1988, 1993 * The Regents of the University of California. * Copyright (c) 2006-2007 Robert N. M. Watson * Copyright (c) 2010-2011 Juniper Networks, Inc. * All rights reserved. * * Portions of this software were developed by Robert N. M. Watson under * contract to Juniper Networks, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * From: @(#)tcp_usrreq.c 8.2 (Berkeley) 1/3/94 */ #include __FBSDID("$FreeBSD$"); #include "opt_ddb.h" #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_kern_tls.h" #include "opt_tcpdebug.h" #include #include #include #include #include #include #include #include #include #include #include #ifdef INET6 #include #endif /* INET6 */ #include #include #include #include #include #include #include #ifdef DDB #include #endif #include #include #include #include #include #include #include #include #include #include #ifdef INET6 #include #include #include #include #endif #include #include #include #include #include #include #include #include #include #include #ifdef TCPPCAP #include #endif #ifdef TCPDEBUG #include #endif #ifdef TCP_OFFLOAD #include #endif #include #include #include #include #include #include #include /* * TCP protocol interface to socket abstraction. */ #ifdef INET static int tcp_connect(struct tcpcb *, struct sockaddr *, struct thread *td); #endif /* INET */ #ifdef INET6 static int tcp6_connect(struct tcpcb *, struct sockaddr *, struct thread *td); #endif /* INET6 */ static void tcp_disconnect(struct tcpcb *); static void tcp_usrclosed(struct tcpcb *); static void tcp_fill_info(struct tcpcb *, struct tcp_info *); static int tcp_pru_options_support(struct tcpcb *tp, int flags); #ifdef TCPDEBUG #define TCPDEBUG0 int ostate = 0 #define TCPDEBUG1() ostate = tp ? tp->t_state : 0 #define TCPDEBUG2(req) if (tp && (so->so_options & SO_DEBUG)) \ tcp_trace(TA_USER, ostate, tp, 0, 0, req) #else #define TCPDEBUG0 #define TCPDEBUG1() #define TCPDEBUG2(req) #endif /* * tcp_require_unique port requires a globally-unique source port for each * outgoing connection. The default is to require the 4-tuple to be unique. */ VNET_DEFINE(int, tcp_require_unique_port) = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, require_unique_port, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_require_unique_port), 0, "Require globally-unique ephemeral port for outgoing connections"); #define V_tcp_require_unique_port VNET(tcp_require_unique_port) /* * TCP attaches to socket via pru_attach(), reserving space, * and an internet control block. */ static int tcp_usr_attach(struct socket *so, int proto, struct thread *td) { struct inpcb *inp; struct tcpcb *tp = NULL; int error; TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp == NULL, ("tcp_usr_attach: inp != NULL")); TCPDEBUG1(); if (so->so_snd.sb_hiwat == 0 || so->so_rcv.sb_hiwat == 0) { error = soreserve(so, V_tcp_sendspace, V_tcp_recvspace); if (error) goto out; } so->so_rcv.sb_flags |= SB_AUTOSIZE; so->so_snd.sb_flags |= SB_AUTOSIZE; error = in_pcballoc(so, &V_tcbinfo); if (error) goto out; inp = sotoinpcb(so); #ifdef INET6 if (inp->inp_vflag & INP_IPV6PROTO) { inp->inp_vflag |= INP_IPV6; if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0) inp->inp_vflag |= INP_IPV4; inp->in6p_hops = -1; /* use kernel default */ } else #endif inp->inp_vflag |= INP_IPV4; tp = tcp_newtcpcb(inp); if (tp == NULL) { error = ENOBUFS; in_pcbdetach(inp); in_pcbfree(inp); goto out; } tp->t_state = TCPS_CLOSED; INP_WUNLOCK(inp); TCPSTATES_INC(TCPS_CLOSED); out: TCPDEBUG2(PRU_ATTACH); TCP_PROBE2(debug__user, tp, PRU_ATTACH); return (error); } /* * tcp_usr_detach is called when the socket layer loses its final reference * to the socket, be it a file descriptor reference, a reference from TCP, * etc. At this point, there is only one case in which we will keep around * inpcb state: time wait. */ static void tcp_usr_detach(struct socket *so) { struct inpcb *inp; struct tcpcb *tp; inp = sotoinpcb(so); KASSERT(inp != NULL, ("%s: inp == NULL", __func__)); INP_WLOCK(inp); KASSERT(so->so_pcb == inp && inp->inp_socket == so, ("%s: socket %p inp %p mismatch", __func__, so, inp)); tp = intotcpcb(inp); if (inp->inp_flags & INP_TIMEWAIT) { /* * There are two cases to handle: one in which the time wait * state is being discarded (INP_DROPPED), and one in which * this connection will remain in timewait. In the former, * it is time to discard all state (except tcptw, which has * already been discarded by the timewait close code, which * should be further up the call stack somewhere). In the * latter case, we detach from the socket, but leave the pcb * present until timewait ends. * * XXXRW: Would it be cleaner to free the tcptw here? * * Astute question indeed, from twtcp perspective there are * four cases to consider: * * #1 tcp_usr_detach is called at tcptw creation time by * tcp_twstart, then do not discard the newly created tcptw * and leave inpcb present until timewait ends * #2 tcp_usr_detach is called at tcptw creation time by * tcp_twstart, but connection is local and tw will be * discarded immediately * #3 tcp_usr_detach is called at timewait end (or reuse) by * tcp_twclose, then the tcptw has already been discarded * (or reused) and inpcb is freed here * #4 tcp_usr_detach is called() after timewait ends (or reuse) * (e.g. by soclose), then tcptw has already been discarded * (or reused) and inpcb is freed here * * In all three cases the tcptw should not be freed here. */ if (inp->inp_flags & INP_DROPPED) { in_pcbdetach(inp); if (__predict_true(tp == NULL)) { in_pcbfree(inp); } else { /* * This case should not happen as in TIMEWAIT * state the inp should not be destroyed before * its tcptw. If INVARIANTS is defined, panic. */ #ifdef INVARIANTS panic("%s: Panic before an inp double-free: " "INP_TIMEWAIT && INP_DROPPED && tp != NULL" , __func__); #else log(LOG_ERR, "%s: Avoid an inp double-free: " "INP_TIMEWAIT && INP_DROPPED && tp != NULL" , __func__); #endif INP_WUNLOCK(inp); } } else { in_pcbdetach(inp); INP_WUNLOCK(inp); } } else { /* * If the connection is not in timewait, we consider two * two conditions: one in which no further processing is * necessary (dropped || embryonic), and one in which TCP is * not yet done, but no longer requires the socket, so the * pcb will persist for the time being. * * XXXRW: Does the second case still occur? */ if (inp->inp_flags & INP_DROPPED || tp->t_state < TCPS_SYN_SENT) { tcp_discardcb(tp); in_pcbdetach(inp); in_pcbfree(inp); } else { in_pcbdetach(inp); INP_WUNLOCK(inp); } } } #ifdef INET /* * Give the socket an address. */ static int tcp_usr_bind(struct socket *so, struct sockaddr *nam, struct thread *td) { int error = 0; struct inpcb *inp; struct tcpcb *tp = NULL; struct sockaddr_in *sinp; sinp = (struct sockaddr_in *)nam; if (nam->sa_len != sizeof (*sinp)) return (EINVAL); /* * Must check for multicast addresses and disallow binding * to them. */ if (sinp->sin_family == AF_INET && IN_MULTICAST(ntohl(sinp->sin_addr.s_addr))) return (EAFNOSUPPORT); TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_bind: inp == NULL")); INP_WLOCK(inp); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { error = EINVAL; goto out; } tp = intotcpcb(inp); TCPDEBUG1(); INP_HASH_WLOCK(&V_tcbinfo); error = in_pcbbind(inp, nam, td->td_ucred); INP_HASH_WUNLOCK(&V_tcbinfo); out: TCPDEBUG2(PRU_BIND); TCP_PROBE2(debug__user, tp, PRU_BIND); INP_WUNLOCK(inp); return (error); } #endif /* INET */ #ifdef INET6 static int tcp6_usr_bind(struct socket *so, struct sockaddr *nam, struct thread *td) { int error = 0; struct inpcb *inp; struct tcpcb *tp = NULL; struct sockaddr_in6 *sin6; u_char vflagsav; sin6 = (struct sockaddr_in6 *)nam; if (nam->sa_len != sizeof (*sin6)) return (EINVAL); /* * Must check for multicast addresses and disallow binding * to them. */ if (sin6->sin6_family == AF_INET6 && IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) return (EAFNOSUPPORT); TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp6_usr_bind: inp == NULL")); INP_WLOCK(inp); vflagsav = inp->inp_vflag; if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { error = EINVAL; goto out; } tp = intotcpcb(inp); TCPDEBUG1(); INP_HASH_WLOCK(&V_tcbinfo); inp->inp_vflag &= ~INP_IPV4; inp->inp_vflag |= INP_IPV6; #ifdef INET if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0) { if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) inp->inp_vflag |= INP_IPV4; else if (IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) { struct sockaddr_in sin; in6_sin6_2_sin(&sin, sin6); if (IN_MULTICAST(ntohl(sin.sin_addr.s_addr))) { error = EAFNOSUPPORT; INP_HASH_WUNLOCK(&V_tcbinfo); goto out; } inp->inp_vflag |= INP_IPV4; inp->inp_vflag &= ~INP_IPV6; error = in_pcbbind(inp, (struct sockaddr *)&sin, td->td_ucred); INP_HASH_WUNLOCK(&V_tcbinfo); goto out; } } #endif error = in6_pcbbind(inp, nam, td->td_ucred); INP_HASH_WUNLOCK(&V_tcbinfo); out: if (error != 0) inp->inp_vflag = vflagsav; TCPDEBUG2(PRU_BIND); TCP_PROBE2(debug__user, tp, PRU_BIND); INP_WUNLOCK(inp); return (error); } #endif /* INET6 */ #ifdef INET /* * Prepare to accept connections. */ static int tcp_usr_listen(struct socket *so, int backlog, struct thread *td) { int error = 0; struct inpcb *inp; struct tcpcb *tp = NULL; TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_listen: inp == NULL")); INP_WLOCK(inp); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { error = EINVAL; goto out; } tp = intotcpcb(inp); TCPDEBUG1(); SOCK_LOCK(so); error = solisten_proto_check(so); INP_HASH_WLOCK(&V_tcbinfo); if (error == 0 && inp->inp_lport == 0) error = in_pcbbind(inp, (struct sockaddr *)0, td->td_ucred); INP_HASH_WUNLOCK(&V_tcbinfo); if (error == 0) { tcp_state_change(tp, TCPS_LISTEN); solisten_proto(so, backlog); #ifdef TCP_OFFLOAD if ((so->so_options & SO_NO_OFFLOAD) == 0) tcp_offload_listen_start(tp); #endif } SOCK_UNLOCK(so); if (IS_FASTOPEN(tp->t_flags)) tp->t_tfo_pending = tcp_fastopen_alloc_counter(); out: TCPDEBUG2(PRU_LISTEN); TCP_PROBE2(debug__user, tp, PRU_LISTEN); INP_WUNLOCK(inp); return (error); } #endif /* INET */ #ifdef INET6 static int tcp6_usr_listen(struct socket *so, int backlog, struct thread *td) { int error = 0; struct inpcb *inp; struct tcpcb *tp = NULL; u_char vflagsav; TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp6_usr_listen: inp == NULL")); INP_WLOCK(inp); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { error = EINVAL; goto out; } vflagsav = inp->inp_vflag; tp = intotcpcb(inp); TCPDEBUG1(); SOCK_LOCK(so); error = solisten_proto_check(so); INP_HASH_WLOCK(&V_tcbinfo); if (error == 0 && inp->inp_lport == 0) { inp->inp_vflag &= ~INP_IPV4; if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0) inp->inp_vflag |= INP_IPV4; error = in6_pcbbind(inp, (struct sockaddr *)0, td->td_ucred); } INP_HASH_WUNLOCK(&V_tcbinfo); if (error == 0) { tcp_state_change(tp, TCPS_LISTEN); solisten_proto(so, backlog); #ifdef TCP_OFFLOAD if ((so->so_options & SO_NO_OFFLOAD) == 0) tcp_offload_listen_start(tp); #endif } SOCK_UNLOCK(so); if (IS_FASTOPEN(tp->t_flags)) tp->t_tfo_pending = tcp_fastopen_alloc_counter(); if (error != 0) inp->inp_vflag = vflagsav; out: TCPDEBUG2(PRU_LISTEN); TCP_PROBE2(debug__user, tp, PRU_LISTEN); INP_WUNLOCK(inp); return (error); } #endif /* INET6 */ #ifdef INET /* * Initiate connection to peer. * Create a template for use in transmissions on this connection. * Enter SYN_SENT state, and mark socket as connecting. * Start keep-alive timer, and seed output sequence space. * Send initial segment on connection. */ static int tcp_usr_connect(struct socket *so, struct sockaddr *nam, struct thread *td) { struct epoch_tracker et; int error = 0; struct inpcb *inp; struct tcpcb *tp = NULL; struct sockaddr_in *sinp; sinp = (struct sockaddr_in *)nam; if (nam->sa_len != sizeof (*sinp)) return (EINVAL); /* * Must disallow TCP ``connections'' to multicast addresses. */ if (sinp->sin_family == AF_INET && IN_MULTICAST(ntohl(sinp->sin_addr.s_addr))) return (EAFNOSUPPORT); if ((sinp->sin_family == AF_INET) && (ntohl(sinp->sin_addr.s_addr) == INADDR_BROADCAST)) return (EACCES); if ((error = prison_remote_ip4(td->td_ucred, &sinp->sin_addr)) != 0) return (error); TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_connect: inp == NULL")); INP_WLOCK(inp); if (inp->inp_flags & INP_TIMEWAIT) { error = EADDRINUSE; goto out; } if (inp->inp_flags & INP_DROPPED) { error = ECONNREFUSED; goto out; } tp = intotcpcb(inp); TCPDEBUG1(); NET_EPOCH_ENTER(et); if ((error = tcp_connect(tp, nam, td)) != 0) goto out_in_epoch; #ifdef TCP_OFFLOAD if (registered_toedevs > 0 && (so->so_options & SO_NO_OFFLOAD) == 0 && (error = tcp_offload_connect(so, nam)) == 0) goto out_in_epoch; #endif tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); error = tp->t_fb->tfb_tcp_output(tp); out_in_epoch: NET_EPOCH_EXIT(et); out: TCPDEBUG2(PRU_CONNECT); TCP_PROBE2(debug__user, tp, PRU_CONNECT); INP_WUNLOCK(inp); return (error); } #endif /* INET */ #ifdef INET6 static int tcp6_usr_connect(struct socket *so, struct sockaddr *nam, struct thread *td) { struct epoch_tracker et; int error = 0; struct inpcb *inp; struct tcpcb *tp = NULL; struct sockaddr_in6 *sin6; u_int8_t incflagsav; u_char vflagsav; TCPDEBUG0; sin6 = (struct sockaddr_in6 *)nam; if (nam->sa_len != sizeof (*sin6)) return (EINVAL); /* * Must disallow TCP ``connections'' to multicast addresses. */ if (sin6->sin6_family == AF_INET6 && IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) return (EAFNOSUPPORT); inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp6_usr_connect: inp == NULL")); INP_WLOCK(inp); vflagsav = inp->inp_vflag; incflagsav = inp->inp_inc.inc_flags; if (inp->inp_flags & INP_TIMEWAIT) { error = EADDRINUSE; goto out; } if (inp->inp_flags & INP_DROPPED) { error = ECONNREFUSED; goto out; } tp = intotcpcb(inp); TCPDEBUG1(); #ifdef INET /* * XXXRW: Some confusion: V4/V6 flags relate to binding, and * therefore probably require the hash lock, which isn't held here. * Is this a significant problem? */ if (IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) { struct sockaddr_in sin; if ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0) { error = EINVAL; goto out; } if ((inp->inp_vflag & INP_IPV4) == 0) { error = EAFNOSUPPORT; goto out; } in6_sin6_2_sin(&sin, sin6); if (IN_MULTICAST(ntohl(sin.sin_addr.s_addr))) { error = EAFNOSUPPORT; goto out; } if (ntohl(sin.sin_addr.s_addr) == INADDR_BROADCAST) { error = EACCES; goto out; } if ((error = prison_remote_ip4(td->td_ucred, &sin.sin_addr)) != 0) goto out; inp->inp_vflag |= INP_IPV4; inp->inp_vflag &= ~INP_IPV6; NET_EPOCH_ENTER(et); if ((error = tcp_connect(tp, (struct sockaddr *)&sin, td)) != 0) goto out_in_epoch; #ifdef TCP_OFFLOAD if (registered_toedevs > 0 && (so->so_options & SO_NO_OFFLOAD) == 0 && (error = tcp_offload_connect(so, nam)) == 0) goto out_in_epoch; #endif error = tp->t_fb->tfb_tcp_output(tp); goto out_in_epoch; } else { if ((inp->inp_vflag & INP_IPV6) == 0) { error = EAFNOSUPPORT; goto out; } } #endif if ((error = prison_remote_ip6(td->td_ucred, &sin6->sin6_addr)) != 0) goto out; inp->inp_vflag &= ~INP_IPV4; inp->inp_vflag |= INP_IPV6; inp->inp_inc.inc_flags |= INC_ISIPV6; if ((error = tcp6_connect(tp, nam, td)) != 0) goto out; #ifdef TCP_OFFLOAD if (registered_toedevs > 0 && (so->so_options & SO_NO_OFFLOAD) == 0 && (error = tcp_offload_connect(so, nam)) == 0) goto out; #endif tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); NET_EPOCH_ENTER(et); error = tp->t_fb->tfb_tcp_output(tp); #ifdef INET out_in_epoch: #endif NET_EPOCH_EXIT(et); out: /* * If the implicit bind in the connect call fails, restore * the flags we modified. */ if (error != 0 && inp->inp_lport == 0) { inp->inp_vflag = vflagsav; inp->inp_inc.inc_flags = incflagsav; } TCPDEBUG2(PRU_CONNECT); TCP_PROBE2(debug__user, tp, PRU_CONNECT); INP_WUNLOCK(inp); return (error); } #endif /* INET6 */ /* * Initiate disconnect from peer. * If connection never passed embryonic stage, just drop; * else if don't need to let data drain, then can just drop anyways, * else have to begin TCP shutdown process: mark socket disconnecting, * drain unread data, state switch to reflect user close, and * send segment (e.g. FIN) to peer. Socket will be really disconnected * when peer sends FIN and acks ours. * * SHOULD IMPLEMENT LATER PRU_CONNECT VIA REALLOC TCPCB. */ static int tcp_usr_disconnect(struct socket *so) { struct inpcb *inp; struct tcpcb *tp = NULL; struct epoch_tracker et; int error = 0; TCPDEBUG0; NET_EPOCH_ENTER(et); inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_disconnect: inp == NULL")); INP_WLOCK(inp); if (inp->inp_flags & INP_TIMEWAIT) goto out; if (inp->inp_flags & INP_DROPPED) { error = ECONNRESET; goto out; } tp = intotcpcb(inp); TCPDEBUG1(); tcp_disconnect(tp); out: TCPDEBUG2(PRU_DISCONNECT); TCP_PROBE2(debug__user, tp, PRU_DISCONNECT); INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); return (error); } #ifdef INET /* * Accept a connection. Essentially all the work is done at higher levels; * just return the address of the peer, storing through addr. */ static int tcp_usr_accept(struct socket *so, struct sockaddr **nam) { int error = 0; struct inpcb *inp = NULL; struct tcpcb *tp = NULL; struct in_addr addr; in_port_t port = 0; TCPDEBUG0; if (so->so_state & SS_ISDISCONNECTED) return (ECONNABORTED); inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_accept: inp == NULL")); INP_WLOCK(inp); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { error = ECONNABORTED; goto out; } tp = intotcpcb(inp); TCPDEBUG1(); /* * We inline in_getpeeraddr and COMMON_END here, so that we can * copy the data of interest and defer the malloc until after we * release the lock. */ port = inp->inp_fport; addr = inp->inp_faddr; out: TCPDEBUG2(PRU_ACCEPT); TCP_PROBE2(debug__user, tp, PRU_ACCEPT); INP_WUNLOCK(inp); if (error == 0) *nam = in_sockaddr(port, &addr); return error; } #endif /* INET */ #ifdef INET6 static int tcp6_usr_accept(struct socket *so, struct sockaddr **nam) { struct inpcb *inp = NULL; int error = 0; struct tcpcb *tp = NULL; struct in_addr addr; struct in6_addr addr6; struct epoch_tracker et; in_port_t port = 0; int v4 = 0; TCPDEBUG0; if (so->so_state & SS_ISDISCONNECTED) return (ECONNABORTED); inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp6_usr_accept: inp == NULL")); NET_EPOCH_ENTER(et); INP_WLOCK(inp); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { error = ECONNABORTED; goto out; } tp = intotcpcb(inp); TCPDEBUG1(); /* * We inline in6_mapped_peeraddr and COMMON_END here, so that we can * copy the data of interest and defer the malloc until after we * release the lock. */ if (inp->inp_vflag & INP_IPV4) { v4 = 1; port = inp->inp_fport; addr = inp->inp_faddr; } else { port = inp->inp_fport; addr6 = inp->in6p_faddr; } out: TCPDEBUG2(PRU_ACCEPT); TCP_PROBE2(debug__user, tp, PRU_ACCEPT); INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); if (error == 0) { if (v4) *nam = in6_v4mapsin6_sockaddr(port, &addr); else *nam = in6_sockaddr(port, &addr6); } return error; } #endif /* INET6 */ /* * Mark the connection as being incapable of further output. */ static int tcp_usr_shutdown(struct socket *so) { int error = 0; struct inpcb *inp; struct tcpcb *tp = NULL; struct epoch_tracker et; TCPDEBUG0; NET_EPOCH_ENTER(et); inp = sotoinpcb(so); KASSERT(inp != NULL, ("inp == NULL")); INP_WLOCK(inp); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { error = ECONNRESET; goto out; } tp = intotcpcb(inp); TCPDEBUG1(); socantsendmore(so); tcp_usrclosed(tp); if (!(inp->inp_flags & INP_DROPPED)) error = tp->t_fb->tfb_tcp_output(tp); out: TCPDEBUG2(PRU_SHUTDOWN); TCP_PROBE2(debug__user, tp, PRU_SHUTDOWN); INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); return (error); } /* * After a receive, possibly send window update to peer. */ static int tcp_usr_rcvd(struct socket *so, int flags) { struct epoch_tracker et; struct inpcb *inp; struct tcpcb *tp = NULL; int error = 0; TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_rcvd: inp == NULL")); INP_WLOCK(inp); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { error = ECONNRESET; goto out; } tp = intotcpcb(inp); TCPDEBUG1(); /* * For passively-created TFO connections, don't attempt a window * update while still in SYN_RECEIVED as this may trigger an early * SYN|ACK. It is preferable to have the SYN|ACK be sent along with * application response data, or failing that, when the DELACK timer * expires. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_state == TCPS_SYN_RECEIVED)) goto out; NET_EPOCH_ENTER(et); #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) tcp_offload_rcvd(tp); else #endif tp->t_fb->tfb_tcp_output(tp); NET_EPOCH_EXIT(et); out: TCPDEBUG2(PRU_RCVD); TCP_PROBE2(debug__user, tp, PRU_RCVD); INP_WUNLOCK(inp); return (error); } /* * Do a send by putting data in output queue and updating urgent * marker if URG set. Possibly send more data. Unlike the other * pru_*() routines, the mbuf chains are our responsibility. We * must either enqueue them or free them. The other pru_* routines * generally are caller-frees. */ static int tcp_usr_send(struct socket *so, int flags, struct mbuf *m, struct sockaddr *nam, struct mbuf *control, struct thread *td) { struct epoch_tracker et; int error = 0; struct inpcb *inp; struct tcpcb *tp = NULL; #ifdef INET #ifdef INET6 struct sockaddr_in sin; #endif struct sockaddr_in *sinp; #endif #ifdef INET6 int isipv6; #endif u_int8_t incflagsav; u_char vflagsav; bool restoreflags; TCPDEBUG0; /* * We require the pcbinfo "read lock" if we will close the socket * as part of this call. */ NET_EPOCH_ENTER(et); inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_send: inp == NULL")); INP_WLOCK(inp); vflagsav = inp->inp_vflag; incflagsav = inp->inp_inc.inc_flags; restoreflags = false; if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { if (control) m_freem(control); /* * In case of PRUS_NOTREADY, tcp_usr_ready() is responsible * for freeing memory. */ if (m && (flags & PRUS_NOTREADY) == 0) m_freem(m); error = ECONNRESET; goto out; } tp = intotcpcb(inp); if (flags & PRUS_OOB) { if ((error = tcp_pru_options_support(tp, PRUS_OOB)) != 0) { if (control) m_freem(control); if (m && (flags & PRUS_NOTREADY) == 0) m_freem(m); goto out; } } TCPDEBUG1(); if (nam != NULL && tp->t_state < TCPS_SYN_SENT) { switch (nam->sa_family) { #ifdef INET case AF_INET: sinp = (struct sockaddr_in *)nam; if (sinp->sin_len != sizeof(struct sockaddr_in)) { if (m) m_freem(m); error = EINVAL; goto out; } if ((inp->inp_vflag & INP_IPV6) != 0) { if (m) m_freem(m); error = EAFNOSUPPORT; goto out; } if (IN_MULTICAST(ntohl(sinp->sin_addr.s_addr))) { if (m) m_freem(m); error = EAFNOSUPPORT; goto out; } if (ntohl(sinp->sin_addr.s_addr) == INADDR_BROADCAST) { if (m) m_freem(m); error = EACCES; goto out; } if ((error = prison_remote_ip4(td->td_ucred, &sinp->sin_addr))) { if (m) m_freem(m); goto out; } #ifdef INET6 isipv6 = 0; #endif break; #endif /* INET */ #ifdef INET6 case AF_INET6: { struct sockaddr_in6 *sin6; sin6 = (struct sockaddr_in6 *)nam; if (sin6->sin6_len != sizeof(*sin6)) { if (m) m_freem(m); error = EINVAL; goto out; } if ((inp->inp_vflag & INP_IPV6PROTO) == 0) { if (m != NULL) m_freem(m); error = EAFNOSUPPORT; goto out; } if (IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) { if (m) m_freem(m); error = EAFNOSUPPORT; goto out; } if (IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) { #ifdef INET if ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0) { error = EINVAL; if (m) m_freem(m); goto out; } if ((inp->inp_vflag & INP_IPV4) == 0) { error = EAFNOSUPPORT; if (m) m_freem(m); goto out; } restoreflags = true; inp->inp_vflag &= ~INP_IPV6; sinp = &sin; in6_sin6_2_sin(sinp, sin6); if (IN_MULTICAST( ntohl(sinp->sin_addr.s_addr))) { error = EAFNOSUPPORT; if (m) m_freem(m); goto out; } if ((error = prison_remote_ip4(td->td_ucred, &sinp->sin_addr))) { if (m) m_freem(m); goto out; } isipv6 = 0; #else /* !INET */ error = EAFNOSUPPORT; if (m) m_freem(m); goto out; #endif /* INET */ } else { if ((inp->inp_vflag & INP_IPV6) == 0) { if (m) m_freem(m); error = EAFNOSUPPORT; goto out; } restoreflags = true; inp->inp_vflag &= ~INP_IPV4; inp->inp_inc.inc_flags |= INC_ISIPV6; if ((error = prison_remote_ip6(td->td_ucred, &sin6->sin6_addr))) { if (m) m_freem(m); goto out; } isipv6 = 1; } break; } #endif /* INET6 */ default: if (m) m_freem(m); error = EAFNOSUPPORT; goto out; } } if (control) { /* TCP doesn't do control messages (rights, creds, etc) */ if (control->m_len) { m_freem(control); if (m) m_freem(m); error = EINVAL; goto out; } m_freem(control); /* empty control, just free it */ } if (!(flags & PRUS_OOB)) { sbappendstream(&so->so_snd, m, flags); if (nam && tp->t_state < TCPS_SYN_SENT) { /* * Do implied connect if not yet connected, * initialize window to default value, and * initialize maxseg using peer's cached MSS. */ #ifdef INET6 if (isipv6) error = tcp6_connect(tp, nam, td); #endif /* INET6 */ #if defined(INET6) && defined(INET) else #endif #ifdef INET error = tcp_connect(tp, (struct sockaddr *)sinp, td); #endif /* * The bind operation in tcp_connect succeeded. We * no longer want to restore the flags if later * operations fail. */ if (error == 0 || inp->inp_lport != 0) restoreflags = false; if (error) goto out; if (IS_FASTOPEN(tp->t_flags)) tcp_fastopen_connect(tp); else { tp->snd_wnd = TTCP_CLIENT_SND_WND; tcp_mss(tp, -1); } } if (flags & PRUS_EOF) { /* * Close the send side of the connection after * the data is sent. */ socantsendmore(so); tcp_usrclosed(tp); } if (TCPS_HAVEESTABLISHED(tp->t_state) && ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && (tp->t_fbyte_out == 0) && (so->so_snd.sb_ccc > 0)) { tp->t_fbyte_out = ticks; if (tp->t_fbyte_out == 0) tp->t_fbyte_out = 1; if (tp->t_fbyte_out && tp->t_fbyte_in) tp->t_flags2 |= TF2_FBYTES_COMPLETE; } if (!(inp->inp_flags & INP_DROPPED) && !(flags & PRUS_NOTREADY)) { if (flags & PRUS_MORETOCOME) tp->t_flags |= TF_MORETOCOME; error = tp->t_fb->tfb_tcp_output(tp); if (flags & PRUS_MORETOCOME) tp->t_flags &= ~TF_MORETOCOME; } } else { /* * XXXRW: PRUS_EOF not implemented with PRUS_OOB? */ SOCKBUF_LOCK(&so->so_snd); if (sbspace(&so->so_snd) < -512) { SOCKBUF_UNLOCK(&so->so_snd); m_freem(m); error = ENOBUFS; goto out; } /* * According to RFC961 (Assigned Protocols), * the urgent pointer points to the last octet * of urgent data. We continue, however, * to consider it to indicate the first octet * of data past the urgent section. * Otherwise, snd_up should be one lower. */ sbappendstream_locked(&so->so_snd, m, flags); SOCKBUF_UNLOCK(&so->so_snd); if (nam && tp->t_state < TCPS_SYN_SENT) { /* * Do implied connect if not yet connected, * initialize window to default value, and * initialize maxseg using peer's cached MSS. */ /* * Not going to contemplate SYN|URG */ if (IS_FASTOPEN(tp->t_flags)) tp->t_flags &= ~TF_FASTOPEN; #ifdef INET6 if (isipv6) error = tcp6_connect(tp, nam, td); #endif /* INET6 */ #if defined(INET6) && defined(INET) else #endif #ifdef INET error = tcp_connect(tp, (struct sockaddr *)sinp, td); #endif /* * The bind operation in tcp_connect succeeded. We * no longer want to restore the flags if later * operations fail. */ if (error == 0 || inp->inp_lport != 0) restoreflags = false; if (error) goto out; tp->snd_wnd = TTCP_CLIENT_SND_WND; tcp_mss(tp, -1); } tp->snd_up = tp->snd_una + sbavail(&so->so_snd); if (!(flags & PRUS_NOTREADY)) { tp->t_flags |= TF_FORCEDATA; error = tp->t_fb->tfb_tcp_output(tp); tp->t_flags &= ~TF_FORCEDATA; } } TCP_LOG_EVENT(tp, NULL, &inp->inp_socket->so_rcv, &inp->inp_socket->so_snd, TCP_LOG_USERSEND, error, 0, NULL, false); out: /* * If the request was unsuccessful and we changed flags, * restore the original flags. */ if (error != 0 && restoreflags) { inp->inp_vflag = vflagsav; inp->inp_inc.inc_flags = incflagsav; } TCPDEBUG2((flags & PRUS_OOB) ? PRU_SENDOOB : ((flags & PRUS_EOF) ? PRU_SEND_EOF : PRU_SEND)); TCP_PROBE2(debug__user, tp, (flags & PRUS_OOB) ? PRU_SENDOOB : ((flags & PRUS_EOF) ? PRU_SEND_EOF : PRU_SEND)); INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); return (error); } static int tcp_usr_ready(struct socket *so, struct mbuf *m, int count) { struct epoch_tracker et; struct inpcb *inp; struct tcpcb *tp; int error; inp = sotoinpcb(so); INP_WLOCK(inp); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { INP_WUNLOCK(inp); mb_free_notready(m, count); return (ECONNRESET); } tp = intotcpcb(inp); SOCKBUF_LOCK(&so->so_snd); error = sbready(&so->so_snd, m, count); SOCKBUF_UNLOCK(&so->so_snd); if (error == 0) { NET_EPOCH_ENTER(et); error = tp->t_fb->tfb_tcp_output(tp); NET_EPOCH_EXIT(et); } INP_WUNLOCK(inp); return (error); } /* * Abort the TCP. Drop the connection abruptly. */ static void tcp_usr_abort(struct socket *so) { struct inpcb *inp; struct tcpcb *tp = NULL; struct epoch_tracker et; TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_abort: inp == NULL")); NET_EPOCH_ENTER(et); INP_WLOCK(inp); KASSERT(inp->inp_socket != NULL, ("tcp_usr_abort: inp_socket == NULL")); /* * If we still have full TCP state, and we're not dropped, drop. */ if (!(inp->inp_flags & INP_TIMEWAIT) && !(inp->inp_flags & INP_DROPPED)) { tp = intotcpcb(inp); TCPDEBUG1(); tp = tcp_drop(tp, ECONNABORTED); if (tp == NULL) goto dropped; TCPDEBUG2(PRU_ABORT); TCP_PROBE2(debug__user, tp, PRU_ABORT); } if (!(inp->inp_flags & INP_DROPPED)) { SOCK_LOCK(so); so->so_state |= SS_PROTOREF; SOCK_UNLOCK(so); inp->inp_flags |= INP_SOCKREF; } INP_WUNLOCK(inp); dropped: NET_EPOCH_EXIT(et); } /* * TCP socket is closed. Start friendly disconnect. */ static void tcp_usr_close(struct socket *so) { struct inpcb *inp; struct tcpcb *tp = NULL; struct epoch_tracker et; TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_close: inp == NULL")); NET_EPOCH_ENTER(et); INP_WLOCK(inp); KASSERT(inp->inp_socket != NULL, ("tcp_usr_close: inp_socket == NULL")); /* * If we still have full TCP state, and we're not dropped, initiate * a disconnect. */ if (!(inp->inp_flags & INP_TIMEWAIT) && !(inp->inp_flags & INP_DROPPED)) { tp = intotcpcb(inp); TCPDEBUG1(); tcp_disconnect(tp); TCPDEBUG2(PRU_CLOSE); TCP_PROBE2(debug__user, tp, PRU_CLOSE); } if (!(inp->inp_flags & INP_DROPPED)) { SOCK_LOCK(so); so->so_state |= SS_PROTOREF; SOCK_UNLOCK(so); inp->inp_flags |= INP_SOCKREF; } INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); } static int tcp_pru_options_support(struct tcpcb *tp, int flags) { /* * If the specific TCP stack has a pru_options * specified then it does not always support * all the PRU_XX options and we must ask it. * If the function is not specified then all * of the PRU_XX options are supported. */ int ret = 0; if (tp->t_fb->tfb_pru_options) { ret = (*tp->t_fb->tfb_pru_options)(tp, flags); } return (ret); } /* * Receive out-of-band data. */ static int tcp_usr_rcvoob(struct socket *so, struct mbuf *m, int flags) { int error = 0; struct inpcb *inp; struct tcpcb *tp = NULL; TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_rcvoob: inp == NULL")); INP_WLOCK(inp); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { error = ECONNRESET; goto out; } tp = intotcpcb(inp); error = tcp_pru_options_support(tp, PRUS_OOB); if (error) { goto out; } TCPDEBUG1(); if ((so->so_oobmark == 0 && (so->so_rcv.sb_state & SBS_RCVATMARK) == 0) || so->so_options & SO_OOBINLINE || tp->t_oobflags & TCPOOB_HADDATA) { error = EINVAL; goto out; } if ((tp->t_oobflags & TCPOOB_HAVEDATA) == 0) { error = EWOULDBLOCK; goto out; } m->m_len = 1; *mtod(m, caddr_t) = tp->t_iobc; if ((flags & MSG_PEEK) == 0) tp->t_oobflags ^= (TCPOOB_HAVEDATA | TCPOOB_HADDATA); out: TCPDEBUG2(PRU_RCVOOB); TCP_PROBE2(debug__user, tp, PRU_RCVOOB); INP_WUNLOCK(inp); return (error); } #ifdef INET struct pr_usrreqs tcp_usrreqs = { .pru_abort = tcp_usr_abort, .pru_accept = tcp_usr_accept, .pru_attach = tcp_usr_attach, .pru_bind = tcp_usr_bind, .pru_connect = tcp_usr_connect, .pru_control = in_control, .pru_detach = tcp_usr_detach, .pru_disconnect = tcp_usr_disconnect, .pru_listen = tcp_usr_listen, .pru_peeraddr = in_getpeeraddr, .pru_rcvd = tcp_usr_rcvd, .pru_rcvoob = tcp_usr_rcvoob, .pru_send = tcp_usr_send, .pru_ready = tcp_usr_ready, .pru_shutdown = tcp_usr_shutdown, .pru_sockaddr = in_getsockaddr, .pru_sosetlabel = in_pcbsosetlabel, .pru_close = tcp_usr_close, }; #endif /* INET */ #ifdef INET6 struct pr_usrreqs tcp6_usrreqs = { .pru_abort = tcp_usr_abort, .pru_accept = tcp6_usr_accept, .pru_attach = tcp_usr_attach, .pru_bind = tcp6_usr_bind, .pru_connect = tcp6_usr_connect, .pru_control = in6_control, .pru_detach = tcp_usr_detach, .pru_disconnect = tcp_usr_disconnect, .pru_listen = tcp6_usr_listen, .pru_peeraddr = in6_mapped_peeraddr, .pru_rcvd = tcp_usr_rcvd, .pru_rcvoob = tcp_usr_rcvoob, .pru_send = tcp_usr_send, .pru_ready = tcp_usr_ready, .pru_shutdown = tcp_usr_shutdown, .pru_sockaddr = in6_mapped_sockaddr, .pru_sosetlabel = in_pcbsosetlabel, .pru_close = tcp_usr_close, }; #endif /* INET6 */ #ifdef INET /* * Common subroutine to open a TCP connection to remote host specified * by struct sockaddr_in in mbuf *nam. Call in_pcbbind to assign a local * port number if needed. Call in_pcbconnect_setup to do the routing and * to choose a local host address (interface). If there is an existing * incarnation of the same connection in TIME-WAIT state and if the remote * host was sending CC options and if the connection duration was < MSL, then * truncate the previous TIME-WAIT state and proceed. * Initialize connection parameters and enter SYN-SENT state. */ static int tcp_connect(struct tcpcb *tp, struct sockaddr *nam, struct thread *td) { struct inpcb *inp = tp->t_inpcb, *oinp; struct socket *so = inp->inp_socket; struct in_addr laddr; u_short lport; int error; NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK(&V_tcbinfo); if (V_tcp_require_unique_port && inp->inp_lport == 0) { error = in_pcbbind(inp, (struct sockaddr *)0, td->td_ucred); if (error) goto out; } /* * Cannot simply call in_pcbconnect, because there might be an * earlier incarnation of this same connection still in * TIME_WAIT state, creating an ADDRINUSE error. */ laddr = inp->inp_laddr; lport = inp->inp_lport; error = in_pcbconnect_setup(inp, nam, &laddr.s_addr, &lport, &inp->inp_faddr.s_addr, &inp->inp_fport, &oinp, td->td_ucred); if (error && oinp == NULL) goto out; if (oinp) { error = EADDRINUSE; goto out; } /* Handle initial bind if it hadn't been done in advance. */ if (inp->inp_lport == 0) { inp->inp_lport = lport; if (in_pcbinshash(inp) != 0) { inp->inp_lport = 0; error = EAGAIN; goto out; } } inp->inp_laddr = laddr; in_pcbrehash(inp); INP_HASH_WUNLOCK(&V_tcbinfo); /* * Compute window scaling to request: * Scale to fit into sweet spot. See tcp_syncache.c. * XXX: This should move to tcp_output(). */ while (tp->request_r_scale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << tp->request_r_scale) < sb_max) tp->request_r_scale++; soisconnecting(so); TCPSTAT_INC(tcps_connattempt); tcp_state_change(tp, TCPS_SYN_SENT); tp->iss = tcp_new_isn(&inp->inp_inc); if (tp->t_flags & TF_REQ_TSTMP) tp->ts_offset = tcp_new_ts_offset(&inp->inp_inc); tcp_sendseqinit(tp); return 0; out: INP_HASH_WUNLOCK(&V_tcbinfo); return (error); } #endif /* INET */ #ifdef INET6 static int tcp6_connect(struct tcpcb *tp, struct sockaddr *nam, struct thread *td) { struct inpcb *inp = tp->t_inpcb; int error; INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK(&V_tcbinfo); if (V_tcp_require_unique_port && inp->inp_lport == 0) { error = in6_pcbbind(inp, (struct sockaddr *)0, td->td_ucred); if (error) goto out; } error = in6_pcbconnect(inp, nam, td->td_ucred); if (error != 0) goto out; INP_HASH_WUNLOCK(&V_tcbinfo); /* Compute window scaling to request. */ while (tp->request_r_scale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << tp->request_r_scale) < sb_max) tp->request_r_scale++; soisconnecting(inp->inp_socket); TCPSTAT_INC(tcps_connattempt); tcp_state_change(tp, TCPS_SYN_SENT); tp->iss = tcp_new_isn(&inp->inp_inc); if (tp->t_flags & TF_REQ_TSTMP) tp->ts_offset = tcp_new_ts_offset(&inp->inp_inc); tcp_sendseqinit(tp); return 0; out: INP_HASH_WUNLOCK(&V_tcbinfo); return error; } #endif /* INET6 */ /* * Export TCP internal state information via a struct tcp_info, based on the * Linux 2.6 API. Not ABI compatible as our constants are mapped differently * (TCP state machine, etc). We export all information using FreeBSD-native * constants -- for example, the numeric values for tcpi_state will differ * from Linux. */ static void tcp_fill_info(struct tcpcb *tp, struct tcp_info *ti) { INP_WLOCK_ASSERT(tp->t_inpcb); bzero(ti, sizeof(*ti)); ti->tcpi_state = tp->t_state; if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP)) ti->tcpi_options |= TCPI_OPT_TIMESTAMPS; if (tp->t_flags & TF_SACK_PERMIT) ti->tcpi_options |= TCPI_OPT_SACK; if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) { ti->tcpi_options |= TCPI_OPT_WSCALE; ti->tcpi_snd_wscale = tp->snd_scale; ti->tcpi_rcv_wscale = tp->rcv_scale; } if (tp->t_flags2 & TF2_ECN_PERMIT) ti->tcpi_options |= TCPI_OPT_ECN; ti->tcpi_rto = tp->t_rxtcur * tick; ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick; ti->tcpi_rtt = ((u_int64_t)tp->t_srtt * tick) >> TCP_RTT_SHIFT; ti->tcpi_rttvar = ((u_int64_t)tp->t_rttvar * tick) >> TCP_RTTVAR_SHIFT; ti->tcpi_snd_ssthresh = tp->snd_ssthresh; ti->tcpi_snd_cwnd = tp->snd_cwnd; /* * FreeBSD-specific extension fields for tcp_info. */ ti->tcpi_rcv_space = tp->rcv_wnd; ti->tcpi_rcv_nxt = tp->rcv_nxt; ti->tcpi_snd_wnd = tp->snd_wnd; ti->tcpi_snd_bwnd = 0; /* Unused, kept for compat. */ ti->tcpi_snd_nxt = tp->snd_nxt; ti->tcpi_snd_mss = tp->t_maxseg; ti->tcpi_rcv_mss = tp->t_maxseg; ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack; ti->tcpi_rcv_ooopack = tp->t_rcvoopack; ti->tcpi_snd_zerowin = tp->t_sndzerowin; #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) { ti->tcpi_options |= TCPI_OPT_TOE; tcp_offload_tcp_info(tp, ti); } #endif } /* * tcp_ctloutput() must drop the inpcb lock before performing copyin on * socket option arguments. When it re-acquires the lock after the copy, it * has to revalidate that the connection is still valid for the socket * option. */ #define INP_WLOCK_RECHECK_CLEANUP(inp, cleanup) do { \ INP_WLOCK(inp); \ if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { \ INP_WUNLOCK(inp); \ cleanup; \ return (ECONNRESET); \ } \ tp = intotcpcb(inp); \ } while(0) #define INP_WLOCK_RECHECK(inp) INP_WLOCK_RECHECK_CLEANUP((inp), /* noop */) int tcp_ctloutput(struct socket *so, struct sockopt *sopt) { int error; struct inpcb *inp; struct tcpcb *tp; struct tcp_function_block *blk; struct tcp_function_set fsn; error = 0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_ctloutput: inp == NULL")); if (sopt->sopt_level != IPPROTO_TCP) { #ifdef INET6 if (inp->inp_vflag & INP_IPV6PROTO) { error = ip6_ctloutput(so, sopt); /* * In case of the IPV6_USE_MIN_MTU socket option, * the INC_IPV6MINMTU flag to announce a corresponding * MSS during the initial handshake. * If the TCP connection is not in the front states, * just reduce the MSS being used. * This avoids the sending of TCP segments which will * be fragmented at the IPv6 layer. */ if ((error == 0) && (sopt->sopt_dir == SOPT_SET) && (sopt->sopt_level == IPPROTO_IPV6) && (sopt->sopt_name == IPV6_USE_MIN_MTU)) { INP_WLOCK(inp); if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED))) { INP_WUNLOCK(inp); return (ECONNRESET); } inp->inp_inc.inc_flags |= INC_IPV6MINMTU; tp = intotcpcb(inp); if ((tp->t_state >= TCPS_SYN_SENT) && (inp->inp_inc.inc_flags & INC_ISIPV6)) { struct ip6_pktopts *opt; opt = inp->in6p_outputopts; if ((opt != NULL) && (opt->ip6po_minmtu == IP6PO_MINMTU_ALL)) { if (tp->t_maxseg > TCP6_MSS) { tp->t_maxseg = TCP6_MSS; } } } INP_WUNLOCK(inp); } } #endif /* INET6 */ #if defined(INET6) && defined(INET) else #endif #ifdef INET { error = ip_ctloutput(so, sopt); } #endif return (error); } INP_WLOCK(inp); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { INP_WUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); /* * Protect the TCP option TCP_FUNCTION_BLK so * that a sub-function can *never* overwrite this. */ if ((sopt->sopt_dir == SOPT_SET) && (sopt->sopt_name == TCP_FUNCTION_BLK)) { INP_WUNLOCK(inp); error = sooptcopyin(sopt, &fsn, sizeof fsn, sizeof fsn); if (error) return (error); INP_WLOCK_RECHECK(inp); blk = find_and_ref_tcp_functions(&fsn); if (blk == NULL) { INP_WUNLOCK(inp); return (ENOENT); } if (tp->t_fb == blk) { /* You already have this */ refcount_release(&blk->tfb_refcnt); INP_WUNLOCK(inp); return (0); } if (tp->t_state != TCPS_CLOSED) { /* * The user has advanced the state * past the initial point, we may not * be able to switch. */ if (blk->tfb_tcp_handoff_ok != NULL) { /* * Does the stack provide a * query mechanism, if so it may * still be possible? */ error = (*blk->tfb_tcp_handoff_ok)(tp); } else error = EINVAL; if (error) { refcount_release(&blk->tfb_refcnt); INP_WUNLOCK(inp); return(error); } } if (blk->tfb_flags & TCP_FUNC_BEING_REMOVED) { refcount_release(&blk->tfb_refcnt); INP_WUNLOCK(inp); return (ENOENT); } /* * Release the old refcnt, the * lookup acquired a ref on the * new one already. */ if (tp->t_fb->tfb_tcp_fb_fini) { /* * Tell the stack to cleanup with 0 i.e. * the tcb is not going away. */ (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); } #ifdef TCPHPTS /* Assure that we are not on any hpts */ tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_ALL); #endif if (blk->tfb_tcp_fb_init) { error = (*blk->tfb_tcp_fb_init)(tp); if (error) { refcount_release(&blk->tfb_refcnt); if (tp->t_fb->tfb_tcp_fb_init) { if((*tp->t_fb->tfb_tcp_fb_init)(tp) != 0) { /* Fall back failed, drop the connection */ INP_WUNLOCK(inp); soabort(so); return(error); } } goto err_out; } } refcount_release(&tp->t_fb->tfb_refcnt); tp->t_fb = blk; #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) { tcp_offload_ctloutput(tp, sopt->sopt_dir, sopt->sopt_name); } #endif err_out: INP_WUNLOCK(inp); return (error); } else if ((sopt->sopt_dir == SOPT_GET) && (sopt->sopt_name == TCP_FUNCTION_BLK)) { strncpy(fsn.function_set_name, tp->t_fb->tfb_tcp_block_name, TCP_FUNCTION_NAME_LEN_MAX); fsn.function_set_name[TCP_FUNCTION_NAME_LEN_MAX - 1] = '\0'; fsn.pcbcnt = tp->t_fb->tfb_refcnt; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &fsn, sizeof fsn); return (error); } /* Pass in the INP locked, called must unlock it */ return (tp->t_fb->tfb_tcp_ctloutput(so, sopt, inp, tp)); } /* * If this assert becomes untrue, we need to change the size of the buf * variable in tcp_default_ctloutput(). */ #ifdef CTASSERT CTASSERT(TCP_CA_NAME_MAX <= TCP_LOG_ID_LEN); CTASSERT(TCP_LOG_REASON_LEN <= TCP_LOG_ID_LEN); #endif #ifdef KERN_TLS static int copyin_tls_enable(struct sockopt *sopt, struct tls_enable *tls) { struct tls_enable_v0 tls_v0; int error; if (sopt->sopt_valsize == sizeof(tls_v0)) { error = sooptcopyin(sopt, &tls_v0, sizeof(tls_v0), sizeof(tls_v0)); if (error) return (error); memset(tls, 0, sizeof(*tls)); tls->cipher_key = tls_v0.cipher_key; tls->iv = tls_v0.iv; tls->auth_key = tls_v0.auth_key; tls->cipher_algorithm = tls_v0.cipher_algorithm; tls->cipher_key_len = tls_v0.cipher_key_len; tls->iv_len = tls_v0.iv_len; tls->auth_algorithm = tls_v0.auth_algorithm; tls->auth_key_len = tls_v0.auth_key_len; tls->flags = tls_v0.flags; tls->tls_vmajor = tls_v0.tls_vmajor; tls->tls_vminor = tls_v0.tls_vminor; return (0); } return (sooptcopyin(sopt, tls, sizeof(*tls), sizeof(*tls))); } #endif int tcp_default_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp) { int error, opt, optval; u_int ui; struct tcp_info ti; #ifdef KERN_TLS struct tls_enable tls; #endif struct cc_algo *algo; char *pbuf, buf[TCP_LOG_ID_LEN]; #ifdef STATS struct statsblob *sbp; #endif size_t len; /* * For TCP_CCALGOOPT forward the control to CC module, for both * SOPT_SET and SOPT_GET. */ switch (sopt->sopt_name) { case TCP_CCALGOOPT: INP_WUNLOCK(inp); if (sopt->sopt_valsize > CC_ALGOOPT_LIMIT) return (EINVAL); pbuf = malloc(sopt->sopt_valsize, M_TEMP, M_WAITOK | M_ZERO); error = sooptcopyin(sopt, pbuf, sopt->sopt_valsize, sopt->sopt_valsize); if (error) { free(pbuf, M_TEMP); return (error); } INP_WLOCK_RECHECK_CLEANUP(inp, free(pbuf, M_TEMP)); if (CC_ALGO(tp)->ctl_output != NULL) error = CC_ALGO(tp)->ctl_output(tp->ccv, sopt, pbuf); else error = ENOENT; INP_WUNLOCK(inp); if (error == 0 && sopt->sopt_dir == SOPT_GET) error = sooptcopyout(sopt, pbuf, sopt->sopt_valsize); free(pbuf, M_TEMP); return (error); } switch (sopt->sopt_dir) { case SOPT_SET: switch (sopt->sopt_name) { #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) case TCP_MD5SIG: if (!TCPMD5_ENABLED()) { INP_WUNLOCK(inp); return (ENOPROTOOPT); } error = TCPMD5_PCBCTL(inp, sopt); if (error) return (error); goto unlock_and_done; #endif /* IPSEC */ case TCP_NODELAY: case TCP_NOOPT: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &optval, sizeof optval, sizeof optval); if (error) return (error); INP_WLOCK_RECHECK(inp); switch (sopt->sopt_name) { case TCP_NODELAY: opt = TF_NODELAY; break; case TCP_NOOPT: opt = TF_NOOPT; break; default: opt = 0; /* dead code to fool gcc */ break; } if (optval) tp->t_flags |= opt; else tp->t_flags &= ~opt; unlock_and_done: #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) { tcp_offload_ctloutput(tp, sopt->sopt_dir, sopt->sopt_name); } #endif INP_WUNLOCK(inp); break; case TCP_NOPUSH: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &optval, sizeof optval, sizeof optval); if (error) return (error); INP_WLOCK_RECHECK(inp); if (optval) tp->t_flags |= TF_NOPUSH; else if (tp->t_flags & TF_NOPUSH) { tp->t_flags &= ~TF_NOPUSH; if (TCPS_HAVEESTABLISHED(tp->t_state)) { struct epoch_tracker et; NET_EPOCH_ENTER(et); error = tp->t_fb->tfb_tcp_output(tp); NET_EPOCH_EXIT(et); } } goto unlock_and_done; + case TCP_REMOTE_UDP_ENCAPS_PORT: + INP_WUNLOCK(inp); + error = sooptcopyin(sopt, &optval, sizeof optval, + sizeof optval); + if (error) + return (error); + if ((optval < TCP_TUNNELING_PORT_MIN) || + (optval > TCP_TUNNELING_PORT_MAX)) { + /* Its got to be in range */ + return (EINVAL); + } + if ((V_tcp_udp_tunneling_port == 0) && (optval != 0)) { + /* You have to have enabled a UDP tunneling port first */ + return (EINVAL); + } + INP_WLOCK_RECHECK(inp); + if (tp->t_state != TCPS_CLOSED) { + /* You can't change after you are connected */ + error = EINVAL; + } else { + /* Ok we are all good set the port */ + tp->t_port = htons(optval); + } + goto unlock_and_done; + case TCP_MAXSEG: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &optval, sizeof optval, sizeof optval); if (error) return (error); INP_WLOCK_RECHECK(inp); if (optval > 0 && optval <= tp->t_maxseg && optval + 40 >= V_tcp_minmss) tp->t_maxseg = optval; else error = EINVAL; goto unlock_and_done; case TCP_INFO: INP_WUNLOCK(inp); error = EINVAL; break; case TCP_STATS: INP_WUNLOCK(inp); #ifdef STATS error = sooptcopyin(sopt, &optval, sizeof optval, sizeof optval); if (error) return (error); if (optval > 0) sbp = stats_blob_alloc( V_tcp_perconn_stats_dflt_tpl, 0); else sbp = NULL; INP_WLOCK_RECHECK(inp); if ((tp->t_stats != NULL && sbp == NULL) || (tp->t_stats == NULL && sbp != NULL)) { struct statsblob *t = tp->t_stats; tp->t_stats = sbp; sbp = t; } INP_WUNLOCK(inp); stats_blob_destroy(sbp); #else return (EOPNOTSUPP); #endif /* !STATS */ break; case TCP_CONGESTION: INP_WUNLOCK(inp); error = sooptcopyin(sopt, buf, TCP_CA_NAME_MAX - 1, 1); if (error) break; buf[sopt->sopt_valsize] = '\0'; INP_WLOCK_RECHECK(inp); CC_LIST_RLOCK(); STAILQ_FOREACH(algo, &cc_list, entries) if (strncmp(buf, algo->name, TCP_CA_NAME_MAX) == 0) break; CC_LIST_RUNLOCK(); if (algo == NULL) { INP_WUNLOCK(inp); error = EINVAL; break; } /* * We hold a write lock over the tcb so it's safe to * do these things without ordering concerns. */ if (CC_ALGO(tp)->cb_destroy != NULL) CC_ALGO(tp)->cb_destroy(tp->ccv); CC_DATA(tp) = NULL; CC_ALGO(tp) = algo; /* * If something goes pear shaped initialising the new * algo, fall back to newreno (which does not * require initialisation). */ if (algo->cb_init != NULL && algo->cb_init(tp->ccv) != 0) { CC_ALGO(tp) = &newreno_cc_algo; /* * The only reason init should fail is * because of malloc. */ error = ENOMEM; } INP_WUNLOCK(inp); break; case TCP_REUSPORT_LB_NUMA: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); INP_WLOCK_RECHECK(inp); if (!error) error = in_pcblbgroup_numa(inp, optval); INP_WUNLOCK(inp); break; #ifdef KERN_TLS case TCP_TXTLS_ENABLE: INP_WUNLOCK(inp); error = copyin_tls_enable(sopt, &tls); if (error) break; error = ktls_enable_tx(so, &tls); break; case TCP_TXTLS_MODE: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &ui, sizeof(ui), sizeof(ui)); if (error) return (error); INP_WLOCK_RECHECK(inp); error = ktls_set_tx_mode(so, ui); INP_WUNLOCK(inp); break; case TCP_RXTLS_ENABLE: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &tls, sizeof(tls), sizeof(tls)); if (error) break; error = ktls_enable_rx(so, &tls); break; #endif case TCP_KEEPIDLE: case TCP_KEEPINTVL: case TCP_KEEPINIT: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &ui, sizeof(ui), sizeof(ui)); if (error) return (error); if (ui > (UINT_MAX / hz)) { error = EINVAL; break; } ui *= hz; INP_WLOCK_RECHECK(inp); switch (sopt->sopt_name) { case TCP_KEEPIDLE: tp->t_keepidle = ui; /* * XXX: better check current remaining * timeout and "merge" it with new value. */ if ((tp->t_state > TCPS_LISTEN) && (tp->t_state <= TCPS_CLOSING)) tcp_timer_activate(tp, TT_KEEP, TP_KEEPIDLE(tp)); break; case TCP_KEEPINTVL: tp->t_keepintvl = ui; if ((tp->t_state == TCPS_FIN_WAIT_2) && (TP_MAXIDLE(tp) > 0)) tcp_timer_activate(tp, TT_2MSL, TP_MAXIDLE(tp)); break; case TCP_KEEPINIT: tp->t_keepinit = ui; if (tp->t_state == TCPS_SYN_RECEIVED || tp->t_state == TCPS_SYN_SENT) tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); break; } goto unlock_and_done; case TCP_KEEPCNT: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &ui, sizeof(ui), sizeof(ui)); if (error) return (error); INP_WLOCK_RECHECK(inp); tp->t_keepcnt = ui; if ((tp->t_state == TCPS_FIN_WAIT_2) && (TP_MAXIDLE(tp) > 0)) tcp_timer_activate(tp, TT_2MSL, TP_MAXIDLE(tp)); goto unlock_and_done; #ifdef TCPPCAP case TCP_PCAP_OUT: case TCP_PCAP_IN: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &optval, sizeof optval, sizeof optval); if (error) return (error); INP_WLOCK_RECHECK(inp); if (optval >= 0) tcp_pcap_set_sock_max(TCP_PCAP_OUT ? &(tp->t_outpkts) : &(tp->t_inpkts), optval); else error = EINVAL; goto unlock_and_done; #endif case TCP_FASTOPEN: { struct tcp_fastopen tfo_optval; INP_WUNLOCK(inp); if (!V_tcp_fastopen_client_enable && !V_tcp_fastopen_server_enable) return (EPERM); error = sooptcopyin(sopt, &tfo_optval, sizeof(tfo_optval), sizeof(int)); if (error) return (error); INP_WLOCK_RECHECK(inp); if ((tp->t_state != TCPS_CLOSED) && (tp->t_state != TCPS_LISTEN)) { error = EINVAL; goto unlock_and_done; } if (tfo_optval.enable) { if (tp->t_state == TCPS_LISTEN) { if (!V_tcp_fastopen_server_enable) { error = EPERM; goto unlock_and_done; } if (tp->t_tfo_pending == NULL) tp->t_tfo_pending = tcp_fastopen_alloc_counter(); } else { /* * If a pre-shared key was provided, * stash it in the client cookie * field of the tcpcb for use during * connect. */ if (sopt->sopt_valsize == sizeof(tfo_optval)) { memcpy(tp->t_tfo_cookie.client, tfo_optval.psk, TCP_FASTOPEN_PSK_LEN); tp->t_tfo_client_cookie_len = TCP_FASTOPEN_PSK_LEN; } } tp->t_flags |= TF_FASTOPEN; } else tp->t_flags &= ~TF_FASTOPEN; goto unlock_and_done; } #ifdef TCP_BLACKBOX case TCP_LOG: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &optval, sizeof optval, sizeof optval); if (error) return (error); INP_WLOCK_RECHECK(inp); error = tcp_log_state_change(tp, optval); goto unlock_and_done; case TCP_LOGBUF: INP_WUNLOCK(inp); error = EINVAL; break; case TCP_LOGID: INP_WUNLOCK(inp); error = sooptcopyin(sopt, buf, TCP_LOG_ID_LEN - 1, 0); if (error) break; buf[sopt->sopt_valsize] = '\0'; INP_WLOCK_RECHECK(inp); error = tcp_log_set_id(tp, buf); /* tcp_log_set_id() unlocks the INP. */ break; case TCP_LOGDUMP: case TCP_LOGDUMPID: INP_WUNLOCK(inp); error = sooptcopyin(sopt, buf, TCP_LOG_REASON_LEN - 1, 0); if (error) break; buf[sopt->sopt_valsize] = '\0'; INP_WLOCK_RECHECK(inp); if (sopt->sopt_name == TCP_LOGDUMP) { error = tcp_log_dump_tp_logbuf(tp, buf, M_WAITOK, true); INP_WUNLOCK(inp); } else { tcp_log_dump_tp_bucket_logbufs(tp, buf); /* * tcp_log_dump_tp_bucket_logbufs() drops the * INP lock. */ } break; #endif default: INP_WUNLOCK(inp); error = ENOPROTOOPT; break; } break; case SOPT_GET: tp = intotcpcb(inp); switch (sopt->sopt_name) { #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) case TCP_MD5SIG: if (!TCPMD5_ENABLED()) { INP_WUNLOCK(inp); return (ENOPROTOOPT); } error = TCPMD5_PCBCTL(inp, sopt); break; #endif case TCP_NODELAY: optval = tp->t_flags & TF_NODELAY; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof optval); break; case TCP_MAXSEG: optval = tp->t_maxseg; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof optval); break; + case TCP_REMOTE_UDP_ENCAPS_PORT: + optval = ntohs(tp->t_port); + INP_WUNLOCK(inp); + error = sooptcopyout(sopt, &optval, sizeof optval); + break; case TCP_NOOPT: optval = tp->t_flags & TF_NOOPT; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof optval); break; case TCP_NOPUSH: optval = tp->t_flags & TF_NOPUSH; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof optval); break; case TCP_INFO: tcp_fill_info(tp, &ti); INP_WUNLOCK(inp); error = sooptcopyout(sopt, &ti, sizeof ti); break; case TCP_STATS: { #ifdef STATS int nheld; TYPEOF_MEMBER(struct statsblob, flags) sbflags = 0; error = 0; socklen_t outsbsz = sopt->sopt_valsize; if (tp->t_stats == NULL) error = ENOENT; else if (outsbsz >= tp->t_stats->cursz) outsbsz = tp->t_stats->cursz; else if (outsbsz >= sizeof(struct statsblob)) outsbsz = sizeof(struct statsblob); else error = EINVAL; INP_WUNLOCK(inp); if (error) break; sbp = sopt->sopt_val; nheld = atop(round_page(((vm_offset_t)sbp) + (vm_size_t)outsbsz) - trunc_page((vm_offset_t)sbp)); vm_page_t ma[nheld]; if (vm_fault_quick_hold_pages( &curproc->p_vmspace->vm_map, (vm_offset_t)sbp, outsbsz, VM_PROT_READ | VM_PROT_WRITE, ma, nheld) < 0) { error = EFAULT; break; } if ((error = copyin_nofault(&(sbp->flags), &sbflags, SIZEOF_MEMBER(struct statsblob, flags)))) goto unhold; INP_WLOCK_RECHECK(inp); error = stats_blob_snapshot(&sbp, outsbsz, tp->t_stats, sbflags | SB_CLONE_USRDSTNOFAULT); INP_WUNLOCK(inp); sopt->sopt_valsize = outsbsz; unhold: vm_page_unhold_pages(ma, nheld); #else INP_WUNLOCK(inp); error = EOPNOTSUPP; #endif /* !STATS */ break; } case TCP_CONGESTION: len = strlcpy(buf, CC_ALGO(tp)->name, TCP_CA_NAME_MAX); INP_WUNLOCK(inp); error = sooptcopyout(sopt, buf, len + 1); break; case TCP_KEEPIDLE: case TCP_KEEPINTVL: case TCP_KEEPINIT: case TCP_KEEPCNT: switch (sopt->sopt_name) { case TCP_KEEPIDLE: ui = TP_KEEPIDLE(tp) / hz; break; case TCP_KEEPINTVL: ui = TP_KEEPINTVL(tp) / hz; break; case TCP_KEEPINIT: ui = TP_KEEPINIT(tp) / hz; break; case TCP_KEEPCNT: ui = TP_KEEPCNT(tp); break; } INP_WUNLOCK(inp); error = sooptcopyout(sopt, &ui, sizeof(ui)); break; #ifdef TCPPCAP case TCP_PCAP_OUT: case TCP_PCAP_IN: optval = tcp_pcap_get_sock_max(TCP_PCAP_OUT ? &(tp->t_outpkts) : &(tp->t_inpkts)); INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof optval); break; #endif case TCP_FASTOPEN: optval = tp->t_flags & TF_FASTOPEN; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof optval); break; #ifdef TCP_BLACKBOX case TCP_LOG: optval = tp->t_logstate; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof(optval)); break; case TCP_LOGBUF: /* tcp_log_getlogbuf() does INP_WUNLOCK(inp) */ error = tcp_log_getlogbuf(sopt, tp); break; case TCP_LOGID: len = tcp_log_get_id(tp, buf); INP_WUNLOCK(inp); error = sooptcopyout(sopt, buf, len + 1); break; case TCP_LOGDUMP: case TCP_LOGDUMPID: INP_WUNLOCK(inp); error = EINVAL; break; #endif #ifdef KERN_TLS case TCP_TXTLS_MODE: optval = ktls_get_tx_mode(so); INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof(optval)); break; case TCP_RXTLS_MODE: optval = ktls_get_rx_mode(so); INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof(optval)); break; #endif default: INP_WUNLOCK(inp); error = ENOPROTOOPT; break; } break; } return (error); } #undef INP_WLOCK_RECHECK #undef INP_WLOCK_RECHECK_CLEANUP /* * Initiate (or continue) disconnect. * If embryonic state, just send reset (once). * If in ``let data drain'' option and linger null, just drop. * Otherwise (hard), mark socket disconnecting and drop * current input data; switch states based on user close, and * send segment to peer (with FIN). */ static void tcp_disconnect(struct tcpcb *tp) { struct inpcb *inp = tp->t_inpcb; struct socket *so = inp->inp_socket; NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(inp); /* * Neither tcp_close() nor tcp_drop() should return NULL, as the * socket is still open. */ if (tp->t_state < TCPS_ESTABLISHED && !(tp->t_state > TCPS_LISTEN && IS_FASTOPEN(tp->t_flags))) { tp = tcp_close(tp); KASSERT(tp != NULL, ("tcp_disconnect: tcp_close() returned NULL")); } else if ((so->so_options & SO_LINGER) && so->so_linger == 0) { tp = tcp_drop(tp, 0); KASSERT(tp != NULL, ("tcp_disconnect: tcp_drop() returned NULL")); } else { soisdisconnecting(so); sbflush(&so->so_rcv); tcp_usrclosed(tp); if (!(inp->inp_flags & INP_DROPPED)) tp->t_fb->tfb_tcp_output(tp); } } /* * User issued close, and wish to trail through shutdown states: * if never received SYN, just forget it. If got a SYN from peer, * but haven't sent FIN, then go to FIN_WAIT_1 state to send peer a FIN. * If already got a FIN from peer, then almost done; go to LAST_ACK * state. In all other cases, have already sent FIN to peer (e.g. * after PRU_SHUTDOWN), and just have to play tedious game waiting * for peer to send FIN or not respond to keep-alives, etc. * We can let the user exit from the close as soon as the FIN is acked. */ static void tcp_usrclosed(struct tcpcb *tp) { NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(tp->t_inpcb); switch (tp->t_state) { case TCPS_LISTEN: #ifdef TCP_OFFLOAD tcp_offload_listen_stop(tp); #endif tcp_state_change(tp, TCPS_CLOSED); /* FALLTHROUGH */ case TCPS_CLOSED: tp = tcp_close(tp); /* * tcp_close() should never return NULL here as the socket is * still open. */ KASSERT(tp != NULL, ("tcp_usrclosed: tcp_close() returned NULL")); break; case TCPS_SYN_SENT: case TCPS_SYN_RECEIVED: tp->t_flags |= TF_NEEDFIN; break; case TCPS_ESTABLISHED: tcp_state_change(tp, TCPS_FIN_WAIT_1); break; case TCPS_CLOSE_WAIT: tcp_state_change(tp, TCPS_LAST_ACK); break; } if (tp->t_state >= TCPS_FIN_WAIT_2) { soisdisconnected(tp->t_inpcb->inp_socket); /* Prevent the connection hanging in FIN_WAIT_2 forever. */ if (tp->t_state == TCPS_FIN_WAIT_2) { int timeout; timeout = (tcp_fast_finwait2_recycle) ? tcp_finwait2_timeout : TP_MAXIDLE(tp); tcp_timer_activate(tp, TT_2MSL, timeout); } } } #ifdef DDB static void db_print_indent(int indent) { int i; for (i = 0; i < indent; i++) db_printf(" "); } static void db_print_tstate(int t_state) { switch (t_state) { case TCPS_CLOSED: db_printf("TCPS_CLOSED"); return; case TCPS_LISTEN: db_printf("TCPS_LISTEN"); return; case TCPS_SYN_SENT: db_printf("TCPS_SYN_SENT"); return; case TCPS_SYN_RECEIVED: db_printf("TCPS_SYN_RECEIVED"); return; case TCPS_ESTABLISHED: db_printf("TCPS_ESTABLISHED"); return; case TCPS_CLOSE_WAIT: db_printf("TCPS_CLOSE_WAIT"); return; case TCPS_FIN_WAIT_1: db_printf("TCPS_FIN_WAIT_1"); return; case TCPS_CLOSING: db_printf("TCPS_CLOSING"); return; case TCPS_LAST_ACK: db_printf("TCPS_LAST_ACK"); return; case TCPS_FIN_WAIT_2: db_printf("TCPS_FIN_WAIT_2"); return; case TCPS_TIME_WAIT: db_printf("TCPS_TIME_WAIT"); return; default: db_printf("unknown"); return; } } static void db_print_tflags(u_int t_flags) { int comma; comma = 0; if (t_flags & TF_ACKNOW) { db_printf("%sTF_ACKNOW", comma ? ", " : ""); comma = 1; } if (t_flags & TF_DELACK) { db_printf("%sTF_DELACK", comma ? ", " : ""); comma = 1; } if (t_flags & TF_NODELAY) { db_printf("%sTF_NODELAY", comma ? ", " : ""); comma = 1; } if (t_flags & TF_NOOPT) { db_printf("%sTF_NOOPT", comma ? ", " : ""); comma = 1; } if (t_flags & TF_SENTFIN) { db_printf("%sTF_SENTFIN", comma ? ", " : ""); comma = 1; } if (t_flags & TF_REQ_SCALE) { db_printf("%sTF_REQ_SCALE", comma ? ", " : ""); comma = 1; } if (t_flags & TF_RCVD_SCALE) { db_printf("%sTF_RECVD_SCALE", comma ? ", " : ""); comma = 1; } if (t_flags & TF_REQ_TSTMP) { db_printf("%sTF_REQ_TSTMP", comma ? ", " : ""); comma = 1; } if (t_flags & TF_RCVD_TSTMP) { db_printf("%sTF_RCVD_TSTMP", comma ? ", " : ""); comma = 1; } if (t_flags & TF_SACK_PERMIT) { db_printf("%sTF_SACK_PERMIT", comma ? ", " : ""); comma = 1; } if (t_flags & TF_NEEDSYN) { db_printf("%sTF_NEEDSYN", comma ? ", " : ""); comma = 1; } if (t_flags & TF_NEEDFIN) { db_printf("%sTF_NEEDFIN", comma ? ", " : ""); comma = 1; } if (t_flags & TF_NOPUSH) { db_printf("%sTF_NOPUSH", comma ? ", " : ""); comma = 1; } if (t_flags & TF_MORETOCOME) { db_printf("%sTF_MORETOCOME", comma ? ", " : ""); comma = 1; } if (t_flags & TF_LQ_OVERFLOW) { db_printf("%sTF_LQ_OVERFLOW", comma ? ", " : ""); comma = 1; } if (t_flags & TF_LASTIDLE) { db_printf("%sTF_LASTIDLE", comma ? ", " : ""); comma = 1; } if (t_flags & TF_RXWIN0SENT) { db_printf("%sTF_RXWIN0SENT", comma ? ", " : ""); comma = 1; } if (t_flags & TF_FASTRECOVERY) { db_printf("%sTF_FASTRECOVERY", comma ? ", " : ""); comma = 1; } if (t_flags & TF_CONGRECOVERY) { db_printf("%sTF_CONGRECOVERY", comma ? ", " : ""); comma = 1; } if (t_flags & TF_WASFRECOVERY) { db_printf("%sTF_WASFRECOVERY", comma ? ", " : ""); comma = 1; } if (t_flags & TF_SIGNATURE) { db_printf("%sTF_SIGNATURE", comma ? ", " : ""); comma = 1; } if (t_flags & TF_FORCEDATA) { db_printf("%sTF_FORCEDATA", comma ? ", " : ""); comma = 1; } if (t_flags & TF_TSO) { db_printf("%sTF_TSO", comma ? ", " : ""); comma = 1; } if (t_flags & TF_FASTOPEN) { db_printf("%sTF_FASTOPEN", comma ? ", " : ""); comma = 1; } } static void db_print_tflags2(u_int t_flags2) { int comma; comma = 0; if (t_flags2 & TF2_ECN_PERMIT) { db_printf("%sTF2_ECN_PERMIT", comma ? ", " : ""); comma = 1; } } static void db_print_toobflags(char t_oobflags) { int comma; comma = 0; if (t_oobflags & TCPOOB_HAVEDATA) { db_printf("%sTCPOOB_HAVEDATA", comma ? ", " : ""); comma = 1; } if (t_oobflags & TCPOOB_HADDATA) { db_printf("%sTCPOOB_HADDATA", comma ? ", " : ""); comma = 1; } } static void db_print_tcpcb(struct tcpcb *tp, const char *name, int indent) { db_print_indent(indent); db_printf("%s at %p\n", name, tp); indent += 2; db_print_indent(indent); db_printf("t_segq first: %p t_segqlen: %d t_dupacks: %d\n", TAILQ_FIRST(&tp->t_segq), tp->t_segqlen, tp->t_dupacks); db_print_indent(indent); db_printf("tt_rexmt: %p tt_persist: %p tt_keep: %p\n", &tp->t_timers->tt_rexmt, &tp->t_timers->tt_persist, &tp->t_timers->tt_keep); db_print_indent(indent); db_printf("tt_2msl: %p tt_delack: %p t_inpcb: %p\n", &tp->t_timers->tt_2msl, &tp->t_timers->tt_delack, tp->t_inpcb); db_print_indent(indent); db_printf("t_state: %d (", tp->t_state); db_print_tstate(tp->t_state); db_printf(")\n"); db_print_indent(indent); db_printf("t_flags: 0x%x (", tp->t_flags); db_print_tflags(tp->t_flags); db_printf(")\n"); db_print_indent(indent); db_printf("t_flags2: 0x%x (", tp->t_flags2); db_print_tflags2(tp->t_flags2); db_printf(")\n"); db_print_indent(indent); db_printf("snd_una: 0x%08x snd_max: 0x%08x snd_nxt: x0%08x\n", tp->snd_una, tp->snd_max, tp->snd_nxt); db_print_indent(indent); db_printf("snd_up: 0x%08x snd_wl1: 0x%08x snd_wl2: 0x%08x\n", tp->snd_up, tp->snd_wl1, tp->snd_wl2); db_print_indent(indent); db_printf("iss: 0x%08x irs: 0x%08x rcv_nxt: 0x%08x\n", tp->iss, tp->irs, tp->rcv_nxt); db_print_indent(indent); db_printf("rcv_adv: 0x%08x rcv_wnd: %u rcv_up: 0x%08x\n", tp->rcv_adv, tp->rcv_wnd, tp->rcv_up); db_print_indent(indent); db_printf("snd_wnd: %u snd_cwnd: %u\n", tp->snd_wnd, tp->snd_cwnd); db_print_indent(indent); db_printf("snd_ssthresh: %u snd_recover: " "0x%08x\n", tp->snd_ssthresh, tp->snd_recover); db_print_indent(indent); db_printf("t_rcvtime: %u t_startime: %u\n", tp->t_rcvtime, tp->t_starttime); db_print_indent(indent); db_printf("t_rttime: %u t_rtsq: 0x%08x\n", tp->t_rtttime, tp->t_rtseq); db_print_indent(indent); db_printf("t_rxtcur: %d t_maxseg: %u t_srtt: %d\n", tp->t_rxtcur, tp->t_maxseg, tp->t_srtt); db_print_indent(indent); db_printf("t_rttvar: %d t_rxtshift: %d t_rttmin: %u " "t_rttbest: %u\n", tp->t_rttvar, tp->t_rxtshift, tp->t_rttmin, tp->t_rttbest); db_print_indent(indent); db_printf("t_rttupdated: %lu max_sndwnd: %u t_softerror: %d\n", tp->t_rttupdated, tp->max_sndwnd, tp->t_softerror); db_print_indent(indent); db_printf("t_oobflags: 0x%x (", tp->t_oobflags); db_print_toobflags(tp->t_oobflags); db_printf(") t_iobc: 0x%02x\n", tp->t_iobc); db_print_indent(indent); db_printf("snd_scale: %u rcv_scale: %u request_r_scale: %u\n", tp->snd_scale, tp->rcv_scale, tp->request_r_scale); db_print_indent(indent); db_printf("ts_recent: %u ts_recent_age: %u\n", tp->ts_recent, tp->ts_recent_age); db_print_indent(indent); db_printf("ts_offset: %u last_ack_sent: 0x%08x snd_cwnd_prev: " "%u\n", tp->ts_offset, tp->last_ack_sent, tp->snd_cwnd_prev); db_print_indent(indent); db_printf("snd_ssthresh_prev: %u snd_recover_prev: 0x%08x " "t_badrxtwin: %u\n", tp->snd_ssthresh_prev, tp->snd_recover_prev, tp->t_badrxtwin); db_print_indent(indent); db_printf("snd_numholes: %d snd_holes first: %p\n", tp->snd_numholes, TAILQ_FIRST(&tp->snd_holes)); db_print_indent(indent); db_printf("snd_fack: 0x%08x rcv_numsacks: %d\n", tp->snd_fack, tp->rcv_numsacks); /* Skip sackblks, sackhint. */ db_print_indent(indent); db_printf("t_rttlow: %d rfbuf_ts: %u rfbuf_cnt: %d\n", tp->t_rttlow, tp->rfbuf_ts, tp->rfbuf_cnt); } DB_SHOW_COMMAND(tcpcb, db_show_tcpcb) { struct tcpcb *tp; if (!have_addr) { db_printf("usage: show tcpcb \n"); return; } tp = (struct tcpcb *)addr; db_print_tcpcb(tp, "tcpcb", 0); } #endif diff --git a/sys/netinet/tcp_var.h b/sys/netinet/tcp_var.h index 4d28cab80d89..dfd2f239d007 100644 --- a/sys/netinet/tcp_var.h +++ b/sys/netinet/tcp_var.h @@ -1,1105 +1,1126 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1993, 1994, 1995 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)tcp_var.h 8.4 (Berkeley) 5/24/95 * $FreeBSD$ */ #ifndef _NETINET_TCP_VAR_H_ #define _NETINET_TCP_VAR_H_ #include #include #ifdef _KERNEL #include #include #endif #define TCP_END_BYTE_INFO 8 /* Bytes that makeup the "end information array" */ /* Types of ending byte info */ #define TCP_EI_EMPTY_SLOT 0 #define TCP_EI_STATUS_CLIENT_FIN 0x1 #define TCP_EI_STATUS_CLIENT_RST 0x2 #define TCP_EI_STATUS_SERVER_FIN 0x3 #define TCP_EI_STATUS_SERVER_RST 0x4 #define TCP_EI_STATUS_RETRAN 0x5 #define TCP_EI_STATUS_PROGRESS 0x6 #define TCP_EI_STATUS_PERSIST_MAX 0x7 #define TCP_EI_STATUS_KEEP_MAX 0x8 #define TCP_EI_STATUS_DATA_A_CLOSE 0x9 #define TCP_EI_STATUS_RST_IN_FRONT 0xa #define TCP_EI_STATUS_2MSL 0xb #define TCP_EI_STATUS_MAX_VALUE 0xb /************************************************/ /* Status bits we track to assure no duplicates, * the bits here are not used by the code but * for human representation. To check a bit we * take and shift over by 1 minus the value (1-8). */ /************************************************/ #define TCP_EI_BITS_CLIENT_FIN 0x001 #define TCP_EI_BITS_CLIENT_RST 0x002 #define TCP_EI_BITS_SERVER_FIN 0x004 #define TCP_EI_BITS_SERVER_RST 0x008 #define TCP_EI_BITS_RETRAN 0x010 #define TCP_EI_BITS_PROGRESS 0x020 #define TCP_EI_BITS_PRESIST_MAX 0x040 #define TCP_EI_BITS_KEEP_MAX 0x080 #define TCP_EI_BITS_DATA_A_CLO 0x100 #define TCP_EI_BITS_RST_IN_FR 0x200 /* a front state reset */ #define TCP_EI_BITS_2MS_TIMER 0x400 /* 2 MSL timer expired */ #if defined(_KERNEL) || defined(_WANT_TCPCB) /* TCP segment queue entry */ struct tseg_qent { TAILQ_ENTRY(tseg_qent) tqe_q; struct mbuf *tqe_m; /* mbuf contains packet */ struct mbuf *tqe_last; /* last mbuf in chain */ tcp_seq tqe_start; /* TCP Sequence number start */ int tqe_len; /* TCP segment data length */ uint32_t tqe_flags; /* The flags from the th->th_flags */ uint32_t tqe_mbuf_cnt; /* Count of mbuf overhead */ }; TAILQ_HEAD(tsegqe_head, tseg_qent); struct sackblk { tcp_seq start; /* start seq no. of sack block */ tcp_seq end; /* end seq no. */ }; struct sackhole { tcp_seq start; /* start seq no. of hole */ tcp_seq end; /* end seq no. */ tcp_seq rxmit; /* next seq. no in hole to be retransmitted */ TAILQ_ENTRY(sackhole) scblink; /* scoreboard linkage */ }; struct sackhint { struct sackhole *nexthole; int32_t sack_bytes_rexmit; tcp_seq last_sack_ack; /* Most recent/largest sacked ack */ int32_t delivered_data; /* Newly acked data from last SACK */ int32_t sacked_bytes; /* Total sacked bytes reported by the * receiver via sack option */ uint32_t recover_fs; /* Flight Size at the start of Loss recovery */ uint32_t prr_delivered; /* Total bytes delivered using PRR */ uint32_t prr_out; /* Bytes sent during IN_RECOVERY */ }; #define SEGQ_EMPTY(tp) TAILQ_EMPTY(&(tp)->t_segq) STAILQ_HEAD(tcp_log_stailq, tcp_log_mem); /* * Tcp control block, one per tcp; fields: * Organized for 64 byte cacheline efficiency based * on common tcp_input/tcp_output processing. */ struct tcpcb { /* Cache line 1 */ struct inpcb *t_inpcb; /* back pointer to internet pcb */ struct tcp_function_block *t_fb;/* TCP function call block */ void *t_fb_ptr; /* Pointer to t_fb specific data */ uint32_t t_maxseg:24, /* maximum segment size */ t_logstate:8; /* State of "black box" logging */ uint32_t t_port:16, /* Tunneling (over udp) port */ t_state:4, /* state of this connection */ t_idle_reduce : 1, t_delayed_ack: 7, /* Delayed ack variable */ t_fin_is_rst: 1, /* Are fin's treated as resets */ t_log_state_set: 1, bits_spare : 2; u_int t_flags; tcp_seq snd_una; /* sent but unacknowledged */ tcp_seq snd_max; /* highest sequence number sent; * used to recognize retransmits */ tcp_seq snd_nxt; /* send next */ tcp_seq snd_up; /* send urgent pointer */ uint32_t snd_wnd; /* send window */ uint32_t snd_cwnd; /* congestion-controlled window */ uint32_t t_peakrate_thr; /* pre-calculated peak rate threshold */ /* Cache line 2 */ u_int32_t ts_offset; /* our timestamp offset */ u_int32_t rfbuf_ts; /* recv buffer autoscaling timestamp */ int rcv_numsacks; /* # distinct sack blks present */ u_int t_tsomax; /* TSO total burst length limit in bytes */ u_int t_tsomaxsegcount; /* TSO maximum segment count */ u_int t_tsomaxsegsize; /* TSO maximum segment size in bytes */ tcp_seq rcv_nxt; /* receive next */ tcp_seq rcv_adv; /* advertised window */ uint32_t rcv_wnd; /* receive window */ u_int t_flags2; /* More tcpcb flags storage */ int t_srtt; /* smoothed round-trip time */ int t_rttvar; /* variance in round-trip time */ u_int32_t ts_recent; /* timestamp echo data */ u_char snd_scale; /* window scaling for send window */ u_char rcv_scale; /* window scaling for recv window */ u_char snd_limited; /* segments limited transmitted */ u_char request_r_scale; /* pending window scaling */ tcp_seq last_ack_sent; u_int t_rcvtime; /* inactivity time */ /* Cache line 3 */ tcp_seq rcv_up; /* receive urgent pointer */ int t_segqlen; /* segment reassembly queue length */ uint32_t t_segqmbuflen; /* Count of bytes mbufs on all entries */ struct tsegqe_head t_segq; /* segment reassembly queue */ struct mbuf *t_in_pkt; struct mbuf *t_tail_pkt; struct tcp_timer *t_timers; /* All the TCP timers in one struct */ struct vnet *t_vnet; /* back pointer to parent vnet */ uint32_t snd_ssthresh; /* snd_cwnd size threshold for * for slow start exponential to * linear switch */ tcp_seq snd_wl1; /* window update seg seq number */ /* Cache line 4 */ tcp_seq snd_wl2; /* window update seg ack number */ tcp_seq irs; /* initial receive sequence number */ tcp_seq iss; /* initial send sequence number */ u_int t_acktime; /* RACK and BBR incoming new data was acked */ u_int t_sndtime; /* time last data was sent */ u_int ts_recent_age; /* when last updated */ tcp_seq snd_recover; /* for use in NewReno Fast Recovery */ uint16_t cl4_spare; /* Spare to adjust CL 4 */ char t_oobflags; /* have some */ char t_iobc; /* input character */ int t_rxtcur; /* current retransmit value (ticks) */ int t_rxtshift; /* log(2) of rexmt exp. backoff */ u_int t_rtttime; /* RTT measurement start time */ tcp_seq t_rtseq; /* sequence number being timed */ u_int t_starttime; /* time connection was established */ u_int t_fbyte_in; /* ticks time when first byte queued in */ u_int t_fbyte_out; /* ticks time when first byte queued out */ u_int t_pmtud_saved_maxseg; /* pre-blackhole MSS */ int t_blackhole_enter; /* when to enter blackhole detection */ int t_blackhole_exit; /* when to exit blackhole detection */ u_int t_rttmin; /* minimum rtt allowed */ u_int t_rttbest; /* best rtt we've seen */ int t_softerror; /* possible error not yet reported */ uint32_t max_sndwnd; /* largest window peer has offered */ /* Cache line 5 */ uint32_t snd_cwnd_prev; /* cwnd prior to retransmit */ uint32_t snd_ssthresh_prev; /* ssthresh prior to retransmit */ tcp_seq snd_recover_prev; /* snd_recover prior to retransmit */ int t_sndzerowin; /* zero-window updates sent */ u_long t_rttupdated; /* number of times rtt sampled */ int snd_numholes; /* number of holes seen by sender */ u_int t_badrxtwin; /* window for retransmit recovery */ TAILQ_HEAD(sackhole_head, sackhole) snd_holes; /* SACK scoreboard (sorted) */ tcp_seq snd_fack; /* last seq number(+1) sack'd by rcv'r*/ struct sackblk sackblks[MAX_SACK_BLKS]; /* seq nos. of sack blocks */ struct sackhint sackhint; /* SACK scoreboard hint */ int t_rttlow; /* smallest observerved RTT */ int rfbuf_cnt; /* recv buffer autoscaling byte count */ struct toedev *tod; /* toedev handling this connection */ int t_sndrexmitpack; /* retransmit packets sent */ int t_rcvoopack; /* out-of-order packets received */ void *t_toe; /* TOE pcb pointer */ struct cc_algo *cc_algo; /* congestion control algorithm */ struct cc_var *ccv; /* congestion control specific vars */ struct osd *osd; /* storage for Khelp module data */ int t_bytes_acked; /* # bytes acked during current RTT */ u_int t_maxunacktime; u_int t_keepinit; /* time to establish connection */ u_int t_keepidle; /* time before keepalive probes begin */ u_int t_keepintvl; /* interval between keepalives */ u_int t_keepcnt; /* number of keepalives before close */ int t_dupacks; /* consecutive dup acks recd */ int t_lognum; /* Number of log entries */ int t_loglimit; /* Maximum number of log entries */ int64_t t_pacing_rate; /* bytes / sec, -1 => unlimited */ struct tcp_log_stailq t_logs; /* Log buffer */ struct tcp_log_id_node *t_lin; struct tcp_log_id_bucket *t_lib; const char *t_output_caller; /* Function that called tcp_output */ struct statsblob *t_stats; /* Per-connection stats */ uint32_t t_logsn; /* Log "serial number" */ uint32_t gput_ts; /* Time goodput measurement started */ tcp_seq gput_seq; /* Outbound measurement seq */ tcp_seq gput_ack; /* Inbound measurement ack */ int32_t t_stats_gput_prev; /* XXXLAS: Prev gput measurement */ uint8_t t_tfo_client_cookie_len; /* TCP Fast Open client cookie length */ uint32_t t_end_info_status; /* Status flag of end info */ unsigned int *t_tfo_pending; /* TCP Fast Open server pending counter */ union { uint8_t client[TCP_FASTOPEN_MAX_COOKIE_LEN]; uint64_t server; } t_tfo_cookie; /* TCP Fast Open cookie to send */ union { uint8_t t_end_info_bytes[TCP_END_BYTE_INFO]; uint64_t t_end_info; }; #ifdef TCPPCAP struct mbufq t_inpkts; /* List of saved input packets. */ struct mbufq t_outpkts; /* List of saved output packets. */ #endif }; #endif /* _KERNEL || _WANT_TCPCB */ #ifdef _KERNEL struct tcptemp { u_char tt_ipgen[40]; /* the size must be of max ip header, now IPv6 */ struct tcphdr tt_t; }; +/* Enable TCP/UDP tunneling port */ +#define TCP_TUNNELING_PORT_MIN 0 +#define TCP_TUNNELING_PORT_MAX 65535 +#define TCP_TUNNELING_PORT_DEFAULT 0 + +/* Enable TCP/UDP tunneling port */ +#define TCP_TUNNELING_OVERHEAD_MIN sizeof(struct udphdr) +#define TCP_TUNNELING_OVERHEAD_MAX 1024 +#define TCP_TUNNELING_OVERHEAD_DEFAULT TCP_TUNNELING_OVERHEAD_MIN + /* Minimum map entries limit value, if set */ #define TCP_MIN_MAP_ENTRIES_LIMIT 128 /* * TODO: We yet need to brave plowing in * to tcp_input() and the pru_usrreq() block. * Right now these go to the old standards which * are somewhat ok, but in the long term may * need to be changed. If we do tackle tcp_input() * then we need to get rid of the tcp_do_segment() * function below. */ /* Flags for tcp functions */ #define TCP_FUNC_BEING_REMOVED 0x01 /* Can no longer be referenced */ /* * If defining the optional tcp_timers, in the * tfb_tcp_timer_stop call you must use the * callout_async_drain() function with the * tcp_timer_discard callback. You should check * the return of callout_async_drain() and if 0 * increment tt_draincnt. Since the timer sub-system * does not know your callbacks you must provide a * stop_all function that loops through and calls * tcp_timer_stop() with each of your defined timers. * Adding a tfb_tcp_handoff_ok function allows the socket * option to change stacks to query you even if the * connection is in a later stage. You return 0 to * say you can take over and run your stack, you return * non-zero (an error number) to say no you can't. * If the function is undefined you can only change * in the early states (before connect or listen). * tfb_tcp_fb_fini is changed to add a flag to tell * the old stack if the tcb is being destroyed or * not. A one in the flag means the TCB is being * destroyed, a zero indicates its transitioning to * another stack (via socket option). */ struct tcp_function_block { char tfb_tcp_block_name[TCP_FUNCTION_NAME_LEN_MAX]; int (*tfb_tcp_output)(struct tcpcb *); int (*tfb_tcp_output_wtime)(struct tcpcb *, const struct timeval *); void (*tfb_tcp_do_segment)(struct mbuf *, struct tcphdr *, struct socket *, struct tcpcb *, int, int, uint8_t); int (*tfb_do_queued_segments)(struct socket *, struct tcpcb *, int); int (*tfb_do_segment_nounlock)(struct mbuf *, struct tcphdr *, struct socket *, struct tcpcb *, int, int, uint8_t, int, struct timeval *); void (*tfb_tcp_hpts_do_segment)(struct mbuf *, struct tcphdr *, struct socket *, struct tcpcb *, int, int, uint8_t, int, struct timeval *); int (*tfb_tcp_ctloutput)(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp); /* Optional memory allocation/free routine */ int (*tfb_tcp_fb_init)(struct tcpcb *); void (*tfb_tcp_fb_fini)(struct tcpcb *, int); /* Optional timers, must define all if you define one */ int (*tfb_tcp_timer_stop_all)(struct tcpcb *); void (*tfb_tcp_timer_activate)(struct tcpcb *, uint32_t, u_int); int (*tfb_tcp_timer_active)(struct tcpcb *, uint32_t); void (*tfb_tcp_timer_stop)(struct tcpcb *, uint32_t); void (*tfb_tcp_rexmit_tmr)(struct tcpcb *); int (*tfb_tcp_handoff_ok)(struct tcpcb *); void (*tfb_tcp_mtu_chg)(struct tcpcb *); int (*tfb_pru_options)(struct tcpcb *, int); volatile uint32_t tfb_refcnt; uint32_t tfb_flags; uint8_t tfb_id; }; struct tcp_function { TAILQ_ENTRY(tcp_function) tf_next; char tf_name[TCP_FUNCTION_NAME_LEN_MAX]; struct tcp_function_block *tf_fb; }; TAILQ_HEAD(tcp_funchead, tcp_function); #endif /* _KERNEL */ /* * Flags and utility macros for the t_flags field. */ #define TF_ACKNOW 0x00000001 /* ack peer immediately */ #define TF_DELACK 0x00000002 /* ack, but try to delay it */ #define TF_NODELAY 0x00000004 /* don't delay packets to coalesce */ #define TF_NOOPT 0x00000008 /* don't use tcp options */ #define TF_SENTFIN 0x00000010 /* have sent FIN */ #define TF_REQ_SCALE 0x00000020 /* have/will request window scaling */ #define TF_RCVD_SCALE 0x00000040 /* other side has requested scaling */ #define TF_REQ_TSTMP 0x00000080 /* have/will request timestamps */ #define TF_RCVD_TSTMP 0x00000100 /* a timestamp was received in SYN */ #define TF_SACK_PERMIT 0x00000200 /* other side said I could SACK */ #define TF_NEEDSYN 0x00000400 /* send SYN (implicit state) */ #define TF_NEEDFIN 0x00000800 /* send FIN (implicit state) */ #define TF_NOPUSH 0x00001000 /* don't push */ #define TF_PREVVALID 0x00002000 /* saved values for bad rxmit valid */ #define TF_WAKESOR 0x00004000 /* wake up receive socket */ #define TF_GPUTINPROG 0x00008000 /* Goodput measurement in progress */ #define TF_MORETOCOME 0x00010000 /* More data to be appended to sock */ #define TF_LQ_OVERFLOW 0x00020000 /* listen queue overflow */ #define TF_LASTIDLE 0x00040000 /* connection was previously idle */ #define TF_RXWIN0SENT 0x00080000 /* sent a receiver win 0 in response */ #define TF_FASTRECOVERY 0x00100000 /* in NewReno Fast Recovery */ #define TF_WASFRECOVERY 0x00200000 /* was in NewReno Fast Recovery */ #define TF_SIGNATURE 0x00400000 /* require MD5 digests (RFC2385) */ #define TF_FORCEDATA 0x00800000 /* force out a byte */ #define TF_TSO 0x01000000 /* TSO enabled on this connection */ #define TF_TOE 0x02000000 /* this connection is offloaded */ #define TF_WAKESOW 0x04000000 /* wake up send socket */ #define TF_UNUSED1 0x08000000 /* unused */ #define TF_UNUSED2 0x10000000 /* unused */ #define TF_CONGRECOVERY 0x20000000 /* congestion recovery mode */ #define TF_WASCRECOVERY 0x40000000 /* was in congestion recovery */ #define TF_FASTOPEN 0x80000000 /* TCP Fast Open indication */ #define IN_FASTRECOVERY(t_flags) (t_flags & TF_FASTRECOVERY) #define ENTER_FASTRECOVERY(t_flags) t_flags |= TF_FASTRECOVERY #define EXIT_FASTRECOVERY(t_flags) t_flags &= ~TF_FASTRECOVERY #define IN_CONGRECOVERY(t_flags) (t_flags & TF_CONGRECOVERY) #define ENTER_CONGRECOVERY(t_flags) t_flags |= TF_CONGRECOVERY #define EXIT_CONGRECOVERY(t_flags) t_flags &= ~TF_CONGRECOVERY #define IN_RECOVERY(t_flags) (t_flags & (TF_CONGRECOVERY | TF_FASTRECOVERY)) #define ENTER_RECOVERY(t_flags) t_flags |= (TF_CONGRECOVERY | TF_FASTRECOVERY) #define EXIT_RECOVERY(t_flags) t_flags &= ~(TF_CONGRECOVERY | TF_FASTRECOVERY) #if defined(_KERNEL) && !defined(TCP_RFC7413) #define IS_FASTOPEN(t_flags) (false) #else #define IS_FASTOPEN(t_flags) (t_flags & TF_FASTOPEN) #endif #define BYTES_THIS_ACK(tp, th) (th->th_ack - tp->snd_una) /* * Flags for the t_oobflags field. */ #define TCPOOB_HAVEDATA 0x01 #define TCPOOB_HADDATA 0x02 /* * Flags for the extended TCP flags field, t_flags2 */ #define TF2_PLPMTU_BLACKHOLE 0x00000001 /* Possible PLPMTUD Black Hole. */ #define TF2_PLPMTU_PMTUD 0x00000002 /* Allowed to attempt PLPMTUD. */ #define TF2_PLPMTU_MAXSEGSNT 0x00000004 /* Last seg sent was full seg. */ #define TF2_LOG_AUTO 0x00000008 /* Session is auto-logging. */ #define TF2_DROP_AF_DATA 0x00000010 /* Drop after all data ack'd */ #define TF2_ECN_PERMIT 0x00000020 /* connection ECN-ready */ #define TF2_ECN_SND_CWR 0x00000040 /* ECN CWR in queue */ #define TF2_ECN_SND_ECE 0x00000080 /* ECN ECE in queue */ #define TF2_ACE_PERMIT 0x00000100 /* Accurate ECN mode */ #define TF2_FBYTES_COMPLETE 0x00000400 /* We have first bytes in and out */ /* * Structure to hold TCP options that are only used during segment * processing (in tcp_input), but not held in the tcpcb. * It's basically used to reduce the number of parameters * to tcp_dooptions and tcp_addoptions. * The binary order of the to_flags is relevant for packing of the * options in tcp_addoptions. */ struct tcpopt { u_int32_t to_flags; /* which options are present */ #define TOF_MSS 0x0001 /* maximum segment size */ #define TOF_SCALE 0x0002 /* window scaling */ #define TOF_SACKPERM 0x0004 /* SACK permitted */ #define TOF_TS 0x0010 /* timestamp */ #define TOF_SIGNATURE 0x0040 /* TCP-MD5 signature option (RFC2385) */ #define TOF_SACK 0x0080 /* Peer sent SACK option */ #define TOF_FASTOPEN 0x0100 /* TCP Fast Open (TFO) cookie */ #define TOF_MAXOPT 0x0200 u_int32_t to_tsval; /* new timestamp */ u_int32_t to_tsecr; /* reflected timestamp */ u_char *to_sacks; /* pointer to the first SACK blocks */ u_char *to_signature; /* pointer to the TCP-MD5 signature */ u_int8_t *to_tfo_cookie; /* pointer to the TFO cookie */ u_int16_t to_mss; /* maximum segment size */ u_int8_t to_wscale; /* window scaling */ u_int8_t to_nsacks; /* number of SACK blocks */ u_int8_t to_tfo_len; /* TFO cookie length */ u_int32_t to_spare; /* UTO */ }; /* * Flags for tcp_dooptions. */ #define TO_SYN 0x01 /* parse SYN-only options */ struct hc_metrics_lite { /* must stay in sync with hc_metrics */ uint32_t rmx_mtu; /* MTU for this path */ uint32_t rmx_ssthresh; /* outbound gateway buffer limit */ uint32_t rmx_rtt; /* estimated round trip time */ uint32_t rmx_rttvar; /* estimated rtt variance */ uint32_t rmx_cwnd; /* congestion window */ uint32_t rmx_sendpipe; /* outbound delay-bandwidth product */ uint32_t rmx_recvpipe; /* inbound delay-bandwidth product */ }; /* * Used by tcp_maxmtu() to communicate interface specific features * and limits at the time of connection setup. */ struct tcp_ifcap { int ifcap; u_int tsomax; u_int tsomaxsegcount; u_int tsomaxsegsize; }; #ifndef _NETINET_IN_PCB_H_ struct in_conninfo; #endif /* _NETINET_IN_PCB_H_ */ struct tcptw { struct inpcb *tw_inpcb; /* XXX back pointer to internet pcb */ + uint32_t t_port:16, /* UDP port number if TCPoUDP */ + t_unused:16; tcp_seq snd_nxt; tcp_seq rcv_nxt; tcp_seq iss; tcp_seq irs; u_short last_win; /* cached window value */ short tw_so_options; /* copy of so_options */ struct ucred *tw_cred; /* user credentials */ u_int32_t t_recent; u_int32_t ts_offset; /* our timestamp offset */ u_int t_starttime; int tw_time; TAILQ_ENTRY(tcptw) tw_2msl; void *tw_pspare; /* TCP_SIGNATURE */ u_int *tw_spare; /* TCP_SIGNATURE */ }; #define intotcpcb(ip) ((struct tcpcb *)(ip)->inp_ppcb) #define intotw(ip) ((struct tcptw *)(ip)->inp_ppcb) #define sototcpcb(so) (intotcpcb(sotoinpcb(so))) /* * The smoothed round-trip time and estimated variance * are stored as fixed point numbers scaled by the values below. * For convenience, these scales are also used in smoothing the average * (smoothed = (1/scale)sample + ((scale-1)/scale)smoothed). * With these scales, srtt has 3 bits to the right of the binary point, * and thus an "ALPHA" of 0.875. rttvar has 2 bits to the right of the * binary point, and is smoothed with an ALPHA of 0.75. */ #define TCP_RTT_SCALE 32 /* multiplier for srtt; 3 bits frac. */ #define TCP_RTT_SHIFT 5 /* shift for srtt; 3 bits frac. */ #define TCP_RTTVAR_SCALE 16 /* multiplier for rttvar; 2 bits */ #define TCP_RTTVAR_SHIFT 4 /* shift for rttvar; 2 bits */ #define TCP_DELTA_SHIFT 2 /* see tcp_input.c */ /* * The initial retransmission should happen at rtt + 4 * rttvar. * Because of the way we do the smoothing, srtt and rttvar * will each average +1/2 tick of bias. When we compute * the retransmit timer, we want 1/2 tick of rounding and * 1 extra tick because of +-1/2 tick uncertainty in the * firing of the timer. The bias will give us exactly the * 1.5 tick we need. But, because the bias is * statistical, we have to test that we don't drop below * the minimum feasible timer (which is 2 ticks). * This version of the macro adapted from a paper by Lawrence * Brakmo and Larry Peterson which outlines a problem caused * by insufficient precision in the original implementation, * which results in inappropriately large RTO values for very * fast networks. */ #define TCP_REXMTVAL(tp) \ max((tp)->t_rttmin, (((tp)->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)) \ + (tp)->t_rttvar) >> TCP_DELTA_SHIFT) /* * TCP statistics. * Many of these should be kept per connection, * but that's inconvenient at the moment. */ struct tcpstat { uint64_t tcps_connattempt; /* connections initiated */ uint64_t tcps_accepts; /* connections accepted */ uint64_t tcps_connects; /* connections established */ uint64_t tcps_drops; /* connections dropped */ uint64_t tcps_conndrops; /* embryonic connections dropped */ uint64_t tcps_minmssdrops; /* average minmss too low drops */ uint64_t tcps_closed; /* conn. closed (includes drops) */ uint64_t tcps_segstimed; /* segs where we tried to get rtt */ uint64_t tcps_rttupdated; /* times we succeeded */ uint64_t tcps_delack; /* delayed acks sent */ uint64_t tcps_timeoutdrop; /* conn. dropped in rxmt timeout */ uint64_t tcps_rexmttimeo; /* retransmit timeouts */ uint64_t tcps_persisttimeo; /* persist timeouts */ uint64_t tcps_keeptimeo; /* keepalive timeouts */ uint64_t tcps_keepprobe; /* keepalive probes sent */ uint64_t tcps_keepdrops; /* connections dropped in keepalive */ uint64_t tcps_sndtotal; /* total packets sent */ uint64_t tcps_sndpack; /* data packets sent */ uint64_t tcps_sndbyte; /* data bytes sent */ uint64_t tcps_sndrexmitpack; /* data packets retransmitted */ uint64_t tcps_sndrexmitbyte; /* data bytes retransmitted */ uint64_t tcps_sndrexmitbad; /* unnecessary packet retransmissions */ uint64_t tcps_sndacks; /* ack-only packets sent */ uint64_t tcps_sndprobe; /* window probes sent */ uint64_t tcps_sndurg; /* packets sent with URG only */ uint64_t tcps_sndwinup; /* window update-only packets sent */ uint64_t tcps_sndctrl; /* control (SYN|FIN|RST) packets sent */ uint64_t tcps_rcvtotal; /* total packets received */ uint64_t tcps_rcvpack; /* packets received in sequence */ uint64_t tcps_rcvbyte; /* bytes received in sequence */ uint64_t tcps_rcvbadsum; /* packets received with ccksum errs */ uint64_t tcps_rcvbadoff; /* packets received with bad offset */ uint64_t tcps_rcvreassfull; /* packets dropped for no reass space */ uint64_t tcps_rcvshort; /* packets received too short */ uint64_t tcps_rcvduppack; /* duplicate-only packets received */ uint64_t tcps_rcvdupbyte; /* duplicate-only bytes received */ uint64_t tcps_rcvpartduppack; /* packets with some duplicate data */ uint64_t tcps_rcvpartdupbyte; /* dup. bytes in part-dup. packets */ uint64_t tcps_rcvoopack; /* out-of-order packets received */ uint64_t tcps_rcvoobyte; /* out-of-order bytes received */ uint64_t tcps_rcvpackafterwin; /* packets with data after window */ uint64_t tcps_rcvbyteafterwin; /* bytes rcvd after window */ uint64_t tcps_rcvafterclose; /* packets rcvd after "close" */ uint64_t tcps_rcvwinprobe; /* rcvd window probe packets */ uint64_t tcps_rcvdupack; /* rcvd duplicate acks */ uint64_t tcps_rcvacktoomuch; /* rcvd acks for unsent data */ uint64_t tcps_rcvackpack; /* rcvd ack packets */ uint64_t tcps_rcvackbyte; /* bytes acked by rcvd acks */ uint64_t tcps_rcvwinupd; /* rcvd window update packets */ uint64_t tcps_pawsdrop; /* segments dropped due to PAWS */ uint64_t tcps_predack; /* times hdr predict ok for acks */ uint64_t tcps_preddat; /* times hdr predict ok for data pkts */ uint64_t tcps_pcbcachemiss; uint64_t tcps_cachedrtt; /* times cached RTT in route updated */ uint64_t tcps_cachedrttvar; /* times cached rttvar updated */ uint64_t tcps_cachedssthresh; /* times cached ssthresh updated */ uint64_t tcps_usedrtt; /* times RTT initialized from route */ uint64_t tcps_usedrttvar; /* times RTTVAR initialized from rt */ uint64_t tcps_usedssthresh; /* times ssthresh initialized from rt*/ uint64_t tcps_persistdrop; /* timeout in persist state */ uint64_t tcps_badsyn; /* bogus SYN, e.g. premature ACK */ uint64_t tcps_mturesent; /* resends due to MTU discovery */ uint64_t tcps_listendrop; /* listen queue overflows */ uint64_t tcps_badrst; /* ignored RSTs in the window */ uint64_t tcps_sc_added; /* entry added to syncache */ uint64_t tcps_sc_retransmitted; /* syncache entry was retransmitted */ uint64_t tcps_sc_dupsyn; /* duplicate SYN packet */ uint64_t tcps_sc_dropped; /* could not reply to packet */ uint64_t tcps_sc_completed; /* successful extraction of entry */ uint64_t tcps_sc_bucketoverflow;/* syncache per-bucket limit hit */ uint64_t tcps_sc_cacheoverflow; /* syncache cache limit hit */ uint64_t tcps_sc_reset; /* RST removed entry from syncache */ uint64_t tcps_sc_stale; /* timed out or listen socket gone */ uint64_t tcps_sc_aborted; /* syncache entry aborted */ uint64_t tcps_sc_badack; /* removed due to bad ACK */ uint64_t tcps_sc_unreach; /* ICMP unreachable received */ uint64_t tcps_sc_zonefail; /* zalloc() failed */ uint64_t tcps_sc_sendcookie; /* SYN cookie sent */ uint64_t tcps_sc_recvcookie; /* SYN cookie received */ uint64_t tcps_hc_added; /* entry added to hostcache */ uint64_t tcps_hc_bucketoverflow;/* hostcache per bucket limit hit */ uint64_t tcps_finwait2_drops; /* Drop FIN_WAIT_2 connection after time limit */ /* SACK related stats */ uint64_t tcps_sack_recovery_episode; /* SACK recovery episodes */ uint64_t tcps_sack_rexmits; /* SACK rexmit segments */ uint64_t tcps_sack_rexmit_bytes; /* SACK rexmit bytes */ uint64_t tcps_sack_rcv_blocks; /* SACK blocks (options) received */ uint64_t tcps_sack_send_blocks; /* SACK blocks (options) sent */ uint64_t tcps_sack_sboverflow; /* times scoreboard overflowed */ /* ECN related stats */ uint64_t tcps_ecn_ce; /* ECN Congestion Experienced */ uint64_t tcps_ecn_ect0; /* ECN Capable Transport */ uint64_t tcps_ecn_ect1; /* ECN Capable Transport */ uint64_t tcps_ecn_shs; /* ECN successful handshakes */ uint64_t tcps_ecn_rcwnd; /* # times ECN reduced the cwnd */ /* TCP_SIGNATURE related stats */ uint64_t tcps_sig_rcvgoodsig; /* Total matching signature received */ uint64_t tcps_sig_rcvbadsig; /* Total bad signature received */ uint64_t tcps_sig_err_buildsig; /* Failed to make signature */ uint64_t tcps_sig_err_sigopt; /* No signature expected by socket */ uint64_t tcps_sig_err_nosigopt; /* No signature provided by segment */ /* Path MTU Discovery Black Hole Detection related stats */ uint64_t tcps_pmtud_blackhole_activated; /* Black Hole Count */ uint64_t tcps_pmtud_blackhole_activated_min_mss; /* BH at min MSS Count */ uint64_t tcps_pmtud_blackhole_failed; /* Black Hole Failure Count */ - uint64_t _pad[12]; /* 6 UTO, 6 TBD */ + uint64_t tcps_tunneled_pkts; /* Packets encap's in UDP received */ + uint64_t tcps_tunneled_errs; /* Packets that had errors that were UDP encaped */ + + uint64_t _pad[10]; /* 6 UTO, 6 TBD */ }; #define tcps_rcvmemdrop tcps_rcvreassfull /* compat */ #ifdef _KERNEL #define TI_UNLOCKED 1 #define TI_RLOCKED 2 #include VNET_PCPUSTAT_DECLARE(struct tcpstat, tcpstat); /* tcp statistics */ /* * In-kernel consumers can use these accessor macros directly to update * stats. */ #define TCPSTAT_ADD(name, val) \ VNET_PCPUSTAT_ADD(struct tcpstat, tcpstat, name, (val)) #define TCPSTAT_INC(name) TCPSTAT_ADD(name, 1) /* * Kernel module consumers must use this accessor macro. */ void kmod_tcpstat_add(int statnum, int val); #define KMOD_TCPSTAT_ADD(name, val) \ kmod_tcpstat_add(offsetof(struct tcpstat, name) / sizeof(uint64_t), val) #define KMOD_TCPSTAT_INC(name) KMOD_TCPSTAT_ADD(name, 1) /* * Running TCP connection count by state. */ VNET_DECLARE(counter_u64_t, tcps_states[TCP_NSTATES]); #define V_tcps_states VNET(tcps_states) #define TCPSTATES_INC(state) counter_u64_add(V_tcps_states[state], 1) #define TCPSTATES_DEC(state) counter_u64_add(V_tcps_states[state], -1) /* * TCP specific helper hook point identifiers. */ #define HHOOK_TCP_EST_IN 0 #define HHOOK_TCP_EST_OUT 1 #define HHOOK_TCP_LAST HHOOK_TCP_EST_OUT struct tcp_hhook_data { struct tcpcb *tp; struct tcphdr *th; struct tcpopt *to; uint32_t len; int tso; tcp_seq curack; }; #ifdef TCP_HHOOK void hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th, struct tcpopt *to, uint32_t len, int tso); #endif #endif /* * TCB structure exported to user-land via sysctl(3). * * Fields prefixed with "xt_" are unique to the export structure, and fields * with "t_" or other prefixes match corresponding fields of 'struct tcpcb'. * * Legend: * (s) - used by userland utilities in src * (p) - used by utilities in ports * (3) - is known to be used by third party software not in ports * (n) - no known usage * * Evil hack: declare only if in_pcb.h and sys/socketvar.h have been * included. Not all of our clients do. */ #if defined(_NETINET_IN_PCB_H_) && defined(_SYS_SOCKETVAR_H_) struct xtcpcb { ksize_t xt_len; /* length of this structure */ struct xinpcb xt_inp; char xt_stack[TCP_FUNCTION_NAME_LEN_MAX]; /* (s) */ char xt_logid[TCP_LOG_ID_LEN]; /* (s) */ char xt_cc[TCP_CA_NAME_MAX]; /* (s) */ int64_t spare64[6]; int32_t t_state; /* (s,p) */ uint32_t t_flags; /* (s,p) */ int32_t t_sndzerowin; /* (s) */ int32_t t_sndrexmitpack; /* (s) */ int32_t t_rcvoopack; /* (s) */ int32_t t_rcvtime; /* (s) */ int32_t tt_rexmt; /* (s) */ int32_t tt_persist; /* (s) */ int32_t tt_keep; /* (s) */ int32_t tt_2msl; /* (s) */ int32_t tt_delack; /* (s) */ int32_t t_logstate; /* (3) */ uint32_t t_snd_cwnd; /* (s) */ uint32_t t_snd_ssthresh; /* (s) */ uint32_t t_maxseg; /* (s) */ uint32_t t_rcv_wnd; /* (s) */ uint32_t t_snd_wnd; /* (s) */ uint32_t xt_ecn; /* (s) */ - int32_t spare32[26]; + uint16_t xt_encaps_port; /* (s) */ + int16_t spare16; + int32_t spare32[25]; } __aligned(8); #ifdef _KERNEL void tcp_inptoxtp(const struct inpcb *, struct xtcpcb *); #endif #endif /* * TCP function information (name-to-id mapping, aliases, and refcnt) * exported to user-land via sysctl(3). */ struct tcp_function_info { uint32_t tfi_refcnt; uint8_t tfi_id; char tfi_name[TCP_FUNCTION_NAME_LEN_MAX]; char tfi_alias[TCP_FUNCTION_NAME_LEN_MAX]; }; /* * Identifiers for TCP sysctl nodes */ #define TCPCTL_DO_RFC1323 1 /* use RFC-1323 extensions */ #define TCPCTL_MSSDFLT 3 /* MSS default */ #define TCPCTL_STATS 4 /* statistics */ #define TCPCTL_RTTDFLT 5 /* default RTT estimate */ #define TCPCTL_KEEPIDLE 6 /* keepalive idle timer */ #define TCPCTL_KEEPINTVL 7 /* interval to send keepalives */ #define TCPCTL_SENDSPACE 8 /* send buffer space */ #define TCPCTL_RECVSPACE 9 /* receive buffer space */ #define TCPCTL_KEEPINIT 10 /* timeout for establishing syn */ #define TCPCTL_PCBLIST 11 /* list of all outstanding PCBs */ #define TCPCTL_DELACKTIME 12 /* time before sending delayed ACK */ #define TCPCTL_V6MSSDFLT 13 /* MSS default for IPv6 */ #define TCPCTL_SACK 14 /* Selective Acknowledgement,rfc 2018 */ #define TCPCTL_DROP 15 /* drop tcp connection */ #define TCPCTL_STATES 16 /* connection counts by TCP state */ #ifdef _KERNEL #ifdef SYSCTL_DECL SYSCTL_DECL(_net_inet_tcp); SYSCTL_DECL(_net_inet_tcp_sack); MALLOC_DECLARE(M_TCPLOG); #endif VNET_DECLARE(int, tcp_log_in_vain); #define V_tcp_log_in_vain VNET(tcp_log_in_vain) /* * Global TCP tunables shared between different stacks. * Please keep the list sorted. */ VNET_DECLARE(int, drop_synfin); VNET_DECLARE(int, path_mtu_discovery); VNET_DECLARE(int, tcp_abc_l_var); VNET_DECLARE(int, tcp_autorcvbuf_max); VNET_DECLARE(int, tcp_autosndbuf_inc); VNET_DECLARE(int, tcp_autosndbuf_max); VNET_DECLARE(int, tcp_delack_enabled); VNET_DECLARE(int, tcp_do_autorcvbuf); VNET_DECLARE(int, tcp_do_autosndbuf); VNET_DECLARE(int, tcp_do_ecn); VNET_DECLARE(int, tcp_do_prr); VNET_DECLARE(int, tcp_do_prr_conservative); VNET_DECLARE(int, tcp_do_newcwv); VNET_DECLARE(int, tcp_do_rfc1323); VNET_DECLARE(int, tcp_tolerate_missing_ts); VNET_DECLARE(int, tcp_do_rfc3042); VNET_DECLARE(int, tcp_do_rfc3390); VNET_DECLARE(int, tcp_do_rfc3465); VNET_DECLARE(int, tcp_do_newsack); VNET_DECLARE(int, tcp_do_sack); VNET_DECLARE(int, tcp_do_tso); VNET_DECLARE(int, tcp_ecn_maxretries); VNET_DECLARE(int, tcp_initcwnd_segments); VNET_DECLARE(int, tcp_insecure_rst); VNET_DECLARE(int, tcp_insecure_syn); VNET_DECLARE(uint32_t, tcp_map_entries_limit); VNET_DECLARE(uint32_t, tcp_map_split_limit); VNET_DECLARE(int, tcp_minmss); VNET_DECLARE(int, tcp_mssdflt); #ifdef STATS VNET_DECLARE(int, tcp_perconn_stats_dflt_tpl); VNET_DECLARE(int, tcp_perconn_stats_enable); #endif /* STATS */ VNET_DECLARE(int, tcp_recvspace); VNET_DECLARE(int, tcp_sack_globalholes); VNET_DECLARE(int, tcp_sack_globalmaxholes); VNET_DECLARE(int, tcp_sack_maxholes); VNET_DECLARE(int, tcp_sc_rst_sock_fail); VNET_DECLARE(int, tcp_sendspace); +VNET_DECLARE(int, tcp_udp_tunneling_overhead); +VNET_DECLARE(int, tcp_udp_tunneling_port); VNET_DECLARE(struct inpcbhead, tcb); VNET_DECLARE(struct inpcbinfo, tcbinfo); #define V_tcp_do_prr VNET(tcp_do_prr) #define V_tcp_do_prr_conservative VNET(tcp_do_prr_conservative) #define V_tcp_do_newcwv VNET(tcp_do_newcwv) #define V_drop_synfin VNET(drop_synfin) #define V_path_mtu_discovery VNET(path_mtu_discovery) #define V_tcb VNET(tcb) #define V_tcbinfo VNET(tcbinfo) #define V_tcp_abc_l_var VNET(tcp_abc_l_var) #define V_tcp_autorcvbuf_max VNET(tcp_autorcvbuf_max) #define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc) #define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max) #define V_tcp_delack_enabled VNET(tcp_delack_enabled) #define V_tcp_do_autorcvbuf VNET(tcp_do_autorcvbuf) #define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf) #define V_tcp_do_ecn VNET(tcp_do_ecn) #define V_tcp_do_rfc1323 VNET(tcp_do_rfc1323) #define V_tcp_tolerate_missing_ts VNET(tcp_tolerate_missing_ts) #define V_tcp_ts_offset_per_conn VNET(tcp_ts_offset_per_conn) #define V_tcp_do_rfc3042 VNET(tcp_do_rfc3042) #define V_tcp_do_rfc3390 VNET(tcp_do_rfc3390) #define V_tcp_do_rfc3465 VNET(tcp_do_rfc3465) #define V_tcp_do_newsack VNET(tcp_do_newsack) #define V_tcp_do_sack VNET(tcp_do_sack) #define V_tcp_do_tso VNET(tcp_do_tso) #define V_tcp_ecn_maxretries VNET(tcp_ecn_maxretries) #define V_tcp_initcwnd_segments VNET(tcp_initcwnd_segments) #define V_tcp_insecure_rst VNET(tcp_insecure_rst) #define V_tcp_insecure_syn VNET(tcp_insecure_syn) #define V_tcp_map_entries_limit VNET(tcp_map_entries_limit) #define V_tcp_map_split_limit VNET(tcp_map_split_limit) #define V_tcp_minmss VNET(tcp_minmss) #define V_tcp_mssdflt VNET(tcp_mssdflt) #ifdef STATS #define V_tcp_perconn_stats_dflt_tpl VNET(tcp_perconn_stats_dflt_tpl) #define V_tcp_perconn_stats_enable VNET(tcp_perconn_stats_enable) #endif /* STATS */ #define V_tcp_recvspace VNET(tcp_recvspace) #define V_tcp_sack_globalholes VNET(tcp_sack_globalholes) #define V_tcp_sack_globalmaxholes VNET(tcp_sack_globalmaxholes) #define V_tcp_sack_maxholes VNET(tcp_sack_maxholes) #define V_tcp_sc_rst_sock_fail VNET(tcp_sc_rst_sock_fail) #define V_tcp_sendspace VNET(tcp_sendspace) #define V_tcp_udp_tunneling_overhead VNET(tcp_udp_tunneling_overhead) #define V_tcp_udp_tunneling_port VNET(tcp_udp_tunneling_port) #ifdef TCP_HHOOK VNET_DECLARE(struct hhook_head *, tcp_hhh[HHOOK_TCP_LAST + 1]); #define V_tcp_hhh VNET(tcp_hhh) #endif int tcp_addoptions(struct tcpopt *, u_char *); int tcp_ccalgounload(struct cc_algo *unload_algo); struct tcpcb * tcp_close(struct tcpcb *); void tcp_discardcb(struct tcpcb *); void tcp_twstart(struct tcpcb *); void tcp_twclose(struct tcptw *, int); void tcp_ctlinput(int, struct sockaddr *, void *); int tcp_ctloutput(struct socket *, struct sockopt *); +void tcp_ctlinput_viaudp(int, struct sockaddr *, void *, void *); struct tcpcb * tcp_drop(struct tcpcb *, int); void tcp_drain(void); void tcp_init(void); void tcp_fini(void *); char *tcp_log_addrs(struct in_conninfo *, struct tcphdr *, void *, const void *); char *tcp_log_vain(struct in_conninfo *, struct tcphdr *, void *, const void *); int tcp_reass(struct tcpcb *, struct tcphdr *, tcp_seq *, int *, struct mbuf *); void tcp_reass_global_init(void); void tcp_reass_flush(struct tcpcb *); void tcp_dooptions(struct tcpopt *, u_char *, int, int); void tcp_dropwithreset(struct mbuf *, struct tcphdr *, struct tcpcb *, int, int); void tcp_pulloutofband(struct socket *, struct tcphdr *, struct mbuf *, int); void tcp_xmit_timer(struct tcpcb *, int); void tcp_newreno_partial_ack(struct tcpcb *, struct tcphdr *); void cc_ack_received(struct tcpcb *tp, struct tcphdr *th, uint16_t nsegs, uint16_t type); void cc_conn_init(struct tcpcb *tp); void cc_post_recovery(struct tcpcb *tp, struct tcphdr *th); void cc_ecnpkt_handler(struct tcpcb *tp, struct tcphdr *th, uint8_t iptos); void cc_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type); #ifdef TCP_HHOOK void hhook_run_tcp_est_in(struct tcpcb *tp, struct tcphdr *th, struct tcpopt *to); #endif int tcp_input(struct mbuf **, int *, int); int tcp_autorcvbuf(struct mbuf *, struct tcphdr *, struct socket *, struct tcpcb *, int); +int tcp_input_with_port(struct mbuf **, int *, int, uint16_t); void tcp_handle_wakeup(struct tcpcb *, struct socket *); void tcp_do_segment(struct mbuf *, struct tcphdr *, struct socket *, struct tcpcb *, int, int, uint8_t); int register_tcp_functions(struct tcp_function_block *blk, int wait); int register_tcp_functions_as_names(struct tcp_function_block *blk, int wait, const char *names[], int *num_names); int register_tcp_functions_as_name(struct tcp_function_block *blk, const char *name, int wait); int deregister_tcp_functions(struct tcp_function_block *blk, bool quiesce, bool force); struct tcp_function_block *find_and_ref_tcp_functions(struct tcp_function_set *fs); void tcp_switch_back_to_default(struct tcpcb *tp); struct tcp_function_block * find_and_ref_tcp_fb(struct tcp_function_block *fs); int tcp_default_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp); extern counter_u64_t tcp_inp_lro_direct_queue; extern counter_u64_t tcp_inp_lro_wokeup_queue; extern counter_u64_t tcp_inp_lro_compressed; extern counter_u64_t tcp_inp_lro_single_push; extern counter_u64_t tcp_inp_lro_locks_taken; extern counter_u64_t tcp_inp_lro_sack_wake; extern counter_u64_t tcp_extra_mbuf; extern counter_u64_t tcp_would_have_but; extern counter_u64_t tcp_comp_total; extern counter_u64_t tcp_uncomp_total; extern counter_u64_t tcp_csum_hardware; extern counter_u64_t tcp_csum_hardware_w_ph; extern counter_u64_t tcp_csum_software; #ifdef NETFLIX_EXP_DETECTION /* Various SACK attack thresholds */ extern int32_t tcp_force_detection; extern int32_t tcp_sack_to_ack_thresh; extern int32_t tcp_sack_to_move_thresh; extern int32_t tcp_restoral_thresh; extern int32_t tcp_sad_decay_val; extern int32_t tcp_sad_pacing_interval; extern int32_t tcp_sad_low_pps; extern int32_t tcp_map_minimum; extern int32_t tcp_attack_on_turns_on_logging; #endif uint32_t tcp_maxmtu(struct in_conninfo *, struct tcp_ifcap *); uint32_t tcp_maxmtu6(struct in_conninfo *, struct tcp_ifcap *); u_int tcp_maxseg(const struct tcpcb *); void tcp_mss_update(struct tcpcb *, int, int, struct hc_metrics_lite *, struct tcp_ifcap *); void tcp_mss(struct tcpcb *, int); int tcp_mssopt(struct in_conninfo *); struct inpcb * tcp_drop_syn_sent(struct inpcb *, int); struct tcpcb * tcp_newtcpcb(struct inpcb *); int tcp_output(struct tcpcb *); void tcp_state_change(struct tcpcb *, int); void tcp_respond(struct tcpcb *, void *, struct tcphdr *, struct mbuf *, tcp_seq, tcp_seq, int); void tcp_tw_init(void); #ifdef VIMAGE void tcp_tw_destroy(void); #endif void tcp_tw_zone_change(void); int tcp_twcheck(struct inpcb *, struct tcpopt *, struct tcphdr *, struct mbuf *, int); void tcp_setpersist(struct tcpcb *); void tcp_slowtimo(void); struct tcptemp * tcpip_maketemplate(struct inpcb *); -void tcpip_fillheaders(struct inpcb *, void *, void *); +void tcpip_fillheaders(struct inpcb *, uint16_t, void *, void *); void tcp_timer_activate(struct tcpcb *, uint32_t, u_int); int tcp_timer_suspend(struct tcpcb *, uint32_t); void tcp_timers_unsuspend(struct tcpcb *, uint32_t); int tcp_timer_active(struct tcpcb *, uint32_t); void tcp_timer_stop(struct tcpcb *, uint32_t); void tcp_trace(short, short, struct tcpcb *, void *, struct tcphdr *, int); int inp_to_cpuid(struct inpcb *inp); /* * All tcp_hc_* functions are IPv4 and IPv6 (via in_conninfo) */ void tcp_hc_init(void); #ifdef VIMAGE void tcp_hc_destroy(void); #endif void tcp_hc_get(struct in_conninfo *, struct hc_metrics_lite *); uint32_t tcp_hc_getmtu(struct in_conninfo *); void tcp_hc_updatemtu(struct in_conninfo *, uint32_t); void tcp_hc_update(struct in_conninfo *, struct hc_metrics_lite *); extern struct pr_usrreqs tcp_usrreqs; uint32_t tcp_new_ts_offset(struct in_conninfo *); tcp_seq tcp_new_isn(struct in_conninfo *); int tcp_sack_doack(struct tcpcb *, struct tcpopt *, tcp_seq); void tcp_update_dsack_list(struct tcpcb *, tcp_seq, tcp_seq); void tcp_update_sack_list(struct tcpcb *tp, tcp_seq rcv_laststart, tcp_seq rcv_lastend); void tcp_clean_dsack_blocks(struct tcpcb *tp); void tcp_clean_sackreport(struct tcpcb *tp); void tcp_sack_adjust(struct tcpcb *tp); struct sackhole *tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt); void tcp_do_prr_ack(struct tcpcb *, struct tcphdr *, struct tcpopt *); void tcp_sack_partialack(struct tcpcb *, struct tcphdr *); void tcp_free_sackholes(struct tcpcb *tp); int tcp_newreno(struct tcpcb *, struct tcphdr *); int tcp_compute_pipe(struct tcpcb *); uint32_t tcp_compute_initwnd(uint32_t); void tcp_sndbuf_autoscale(struct tcpcb *, struct socket *, uint32_t); int tcp_stats_sample_rollthedice(struct tcpcb *tp, void *seed_bytes, size_t seed_len); struct mbuf * tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen, int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls); int tcp_stats_init(void); void tcp_log_end_status(struct tcpcb *tp, uint8_t status); static inline void tcp_fields_to_host(struct tcphdr *th) { th->th_seq = ntohl(th->th_seq); th->th_ack = ntohl(th->th_ack); th->th_win = ntohs(th->th_win); th->th_urp = ntohs(th->th_urp); } static inline void tcp_fields_to_net(struct tcphdr *th) { th->th_seq = htonl(th->th_seq); th->th_ack = htonl(th->th_ack); th->th_win = htons(th->th_win); th->th_urp = htons(th->th_urp); } #endif /* _KERNEL */ #endif /* _NETINET_TCP_VAR_H_ */ diff --git a/sys/netinet/toecore.c b/sys/netinet/toecore.c index 480aa64c1bf7..d8d499a6fde3 100644 --- a/sys/netinet/toecore.c +++ b/sys/netinet/toecore.c @@ -1,592 +1,592 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2012 Chelsio Communications, Inc. * All rights reserved. * Written by: Navdeep Parhar * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define TCPSTATES #include #include #include #include #include #include #include static struct mtx toedev_lock; static TAILQ_HEAD(, toedev) toedev_list; static eventhandler_tag listen_start_eh; static eventhandler_tag listen_stop_eh; static eventhandler_tag lle_event_eh; static int toedev_connect(struct toedev *tod __unused, struct socket *so __unused, struct nhop_object *nh __unused, struct sockaddr *nam __unused) { return (ENOTSUP); } static int toedev_listen_start(struct toedev *tod __unused, struct tcpcb *tp __unused) { return (ENOTSUP); } static int toedev_listen_stop(struct toedev *tod __unused, struct tcpcb *tp __unused) { return (ENOTSUP); } static void toedev_input(struct toedev *tod __unused, struct tcpcb *tp __unused, struct mbuf *m) { m_freem(m); return; } static void toedev_rcvd(struct toedev *tod __unused, struct tcpcb *tp __unused) { return; } static int toedev_output(struct toedev *tod __unused, struct tcpcb *tp __unused) { return (ENOTSUP); } static void toedev_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp __unused) { return; } static void toedev_l2_update(struct toedev *tod __unused, struct ifnet *ifp __unused, struct sockaddr *sa __unused, uint8_t *lladdr __unused, uint16_t vtag __unused) { return; } static void toedev_route_redirect(struct toedev *tod __unused, struct ifnet *ifp __unused, struct nhop_object *nh0 __unused, struct nhop_object *nh1 __unused) { return; } static void toedev_syncache_added(struct toedev *tod __unused, void *ctx __unused) { return; } static void toedev_syncache_removed(struct toedev *tod __unused, void *ctx __unused) { return; } static int toedev_syncache_respond(struct toedev *tod __unused, void *ctx __unused, struct mbuf *m) { m_freem(m); return (0); } static void toedev_offload_socket(struct toedev *tod __unused, void *ctx __unused, struct socket *so __unused) { return; } static void toedev_ctloutput(struct toedev *tod __unused, struct tcpcb *tp __unused, int sopt_dir __unused, int sopt_name __unused) { return; } static void toedev_tcp_info(struct toedev *tod __unused, struct tcpcb *tp __unused, struct tcp_info *ti __unused) { return; } static int toedev_alloc_tls_session(struct toedev *tod __unused, struct tcpcb *tp __unused, struct ktls_session *tls __unused, int direction __unused) { return (EINVAL); } /* * Inform one or more TOE devices about a listening socket. */ static void toe_listen_start(struct inpcb *inp, void *arg) { struct toedev *t, *tod; struct tcpcb *tp; INP_WLOCK_ASSERT(inp); KASSERT(inp->inp_pcbinfo == &V_tcbinfo, ("%s: inp is not a TCP inp", __func__)); if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) return; tp = intotcpcb(inp); if (tp->t_state != TCPS_LISTEN) return; t = arg; mtx_lock(&toedev_lock); TAILQ_FOREACH(tod, &toedev_list, link) { if (t == NULL || t == tod) tod->tod_listen_start(tod, tp); } mtx_unlock(&toedev_lock); } static void toe_listen_start_event(void *arg __unused, struct tcpcb *tp) { struct inpcb *inp = tp->t_inpcb; INP_WLOCK_ASSERT(inp); KASSERT(tp->t_state == TCPS_LISTEN, ("%s: t_state %s", __func__, tcpstates[tp->t_state])); toe_listen_start(inp, NULL); } static void toe_listen_stop_event(void *arg __unused, struct tcpcb *tp) { struct toedev *tod; #ifdef INVARIANTS struct inpcb *inp = tp->t_inpcb; #endif INP_WLOCK_ASSERT(inp); KASSERT(tp->t_state == TCPS_LISTEN, ("%s: t_state %s", __func__, tcpstates[tp->t_state])); mtx_lock(&toedev_lock); TAILQ_FOREACH(tod, &toedev_list, link) tod->tod_listen_stop(tod, tp); mtx_unlock(&toedev_lock); } /* * Fill up a freshly allocated toedev struct with reasonable defaults. */ void init_toedev(struct toedev *tod) { tod->tod_softc = NULL; /* * Provide no-op defaults so that the kernel can call any toedev * function without having to check whether the TOE driver supplied one * or not. */ tod->tod_connect = toedev_connect; tod->tod_listen_start = toedev_listen_start; tod->tod_listen_stop = toedev_listen_stop; tod->tod_input = toedev_input; tod->tod_rcvd = toedev_rcvd; tod->tod_output = toedev_output; tod->tod_send_rst = toedev_output; tod->tod_send_fin = toedev_output; tod->tod_pcb_detach = toedev_pcb_detach; tod->tod_l2_update = toedev_l2_update; tod->tod_route_redirect = toedev_route_redirect; tod->tod_syncache_added = toedev_syncache_added; tod->tod_syncache_removed = toedev_syncache_removed; tod->tod_syncache_respond = toedev_syncache_respond; tod->tod_offload_socket = toedev_offload_socket; tod->tod_ctloutput = toedev_ctloutput; tod->tod_tcp_info = toedev_tcp_info; tod->tod_alloc_tls_session = toedev_alloc_tls_session; } /* * Register an active TOE device with the system. This allows it to receive * notifications from the kernel. */ int register_toedev(struct toedev *tod) { struct toedev *t; mtx_lock(&toedev_lock); TAILQ_FOREACH(t, &toedev_list, link) { if (t == tod) { mtx_unlock(&toedev_lock); return (EEXIST); } } TAILQ_INSERT_TAIL(&toedev_list, tod, link); registered_toedevs++; mtx_unlock(&toedev_lock); inp_apply_all(toe_listen_start, tod); return (0); } /* * Remove the TOE device from the global list of active TOE devices. It is the * caller's responsibility to ensure that the TOE device is quiesced prior to * this call. */ int unregister_toedev(struct toedev *tod) { struct toedev *t, *t2; int rc = ENODEV; mtx_lock(&toedev_lock); TAILQ_FOREACH_SAFE(t, &toedev_list, link, t2) { if (t == tod) { TAILQ_REMOVE(&toedev_list, tod, link); registered_toedevs--; rc = 0; break; } } KASSERT(registered_toedevs >= 0, ("%s: registered_toedevs (%d) < 0", __func__, registered_toedevs)); mtx_unlock(&toedev_lock); return (rc); } void toe_syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, struct inpcb *inp, void *tod, void *todctx, uint8_t iptos) { INP_RLOCK_ASSERT(inp); (void )syncache_add(inc, to, th, inp, inp->inp_socket, NULL, tod, - todctx, iptos); + todctx, iptos, htons(0)); } int toe_syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, struct socket **lsop) { NET_EPOCH_ASSERT(); - return (syncache_expand(inc, to, th, lsop, NULL)); + return (syncache_expand(inc, to, th, lsop, NULL, htons(0))); } /* * General purpose check to see if a 4-tuple is in use by the kernel. If a TCP * header (presumably for an incoming SYN) is also provided, an existing 4-tuple * in TIME_WAIT may be assassinated freeing it up for re-use. * * Note that the TCP header must have been run through tcp_fields_to_host() or * equivalent. */ int toe_4tuple_check(struct in_conninfo *inc, struct tcphdr *th, struct ifnet *ifp) { struct inpcb *inp; if (inc->inc_flags & INC_ISIPV6) { inp = in6_pcblookup(&V_tcbinfo, &inc->inc6_faddr, inc->inc_fport, &inc->inc6_laddr, inc->inc_lport, INPLOOKUP_WLOCKPCB, ifp); } else { inp = in_pcblookup(&V_tcbinfo, inc->inc_faddr, inc->inc_fport, inc->inc_laddr, inc->inc_lport, INPLOOKUP_WLOCKPCB, ifp); } if (inp != NULL) { INP_WLOCK_ASSERT(inp); if ((inp->inp_flags & INP_TIMEWAIT) && th != NULL) { if (!tcp_twcheck(inp, NULL, th, NULL, 0)) return (EADDRINUSE); } else { INP_WUNLOCK(inp); return (EADDRINUSE); } } return (0); } static void toe_lle_event(void *arg __unused, struct llentry *lle, int evt) { struct toedev *tod; struct ifnet *ifp; struct sockaddr *sa; uint8_t *lladdr; uint16_t vid, pcp; int family; struct sockaddr_in6 sin6; LLE_WLOCK_ASSERT(lle); ifp = lltable_get_ifp(lle->lle_tbl); family = lltable_get_af(lle->lle_tbl); if (family != AF_INET && family != AF_INET6) return; /* * Not interested if the interface's TOE capability is not enabled. */ if ((family == AF_INET && !(ifp->if_capenable & IFCAP_TOE4)) || (family == AF_INET6 && !(ifp->if_capenable & IFCAP_TOE6))) return; tod = TOEDEV(ifp); if (tod == NULL) return; sa = (struct sockaddr *)&sin6; lltable_fill_sa_entry(lle, sa); vid = 0xfff; pcp = 0; if (evt != LLENTRY_RESOLVED) { /* * LLENTRY_TIMEDOUT, LLENTRY_DELETED, LLENTRY_EXPIRED all mean * this entry is going to be deleted. */ lladdr = NULL; } else { KASSERT(lle->la_flags & LLE_VALID, ("%s: %p resolved but not valid?", __func__, lle)); lladdr = (uint8_t *)lle->ll_addr; VLAN_TAG(ifp, &vid); VLAN_PCP(ifp, &pcp); } tod->tod_l2_update(tod, ifp, sa, lladdr, EVL_MAKETAG(vid, pcp, 0)); } /* * Returns 0 or EWOULDBLOCK on success (any other value is an error). 0 means * lladdr and vtag are valid on return, EWOULDBLOCK means the TOE driver's * tod_l2_update will be called later, when the entry is resolved or times out. */ int toe_l2_resolve(struct toedev *tod, struct ifnet *ifp, struct sockaddr *sa, uint8_t *lladdr, uint16_t *vtag) { int rc; uint16_t vid, pcp; switch (sa->sa_family) { #ifdef INET case AF_INET: rc = arpresolve(ifp, 0, NULL, sa, lladdr, NULL, NULL); break; #endif #ifdef INET6 case AF_INET6: rc = nd6_resolve(ifp, 0, NULL, sa, lladdr, NULL, NULL); break; #endif default: return (EPROTONOSUPPORT); } if (rc == 0) { vid = 0xfff; pcp = 0; if (ifp->if_type == IFT_L2VLAN) { VLAN_TAG(ifp, &vid); VLAN_PCP(ifp, &pcp); } else if (ifp->if_pcp != IFNET_PCP_NONE) { vid = 0; pcp = ifp->if_pcp; } *vtag = EVL_MAKETAG(vid, pcp, 0); } return (rc); } void toe_connect_failed(struct toedev *tod, struct inpcb *inp, int err) { NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(inp); if (!(inp->inp_flags & INP_DROPPED)) { struct tcpcb *tp = intotcpcb(inp); KASSERT(tp->t_flags & TF_TOE, ("%s: tp %p not offloaded.", __func__, tp)); if (err == EAGAIN) { /* * Temporary failure during offload, take this PCB back. * Detach from the TOE driver and do the rest of what * TCP's pru_connect would have done if the connection * wasn't offloaded. */ tod->tod_pcb_detach(tod, tp); KASSERT(!(tp->t_flags & TF_TOE), ("%s: tp %p still offloaded.", __func__, tp)); tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); (void) tp->t_fb->tfb_tcp_output(tp); } else { tp = tcp_drop(tp, err); if (tp == NULL) INP_WLOCK(inp); /* re-acquire */ } } INP_WLOCK_ASSERT(inp); } static int toecore_load(void) { mtx_init(&toedev_lock, "toedev lock", NULL, MTX_DEF); TAILQ_INIT(&toedev_list); listen_start_eh = EVENTHANDLER_REGISTER(tcp_offload_listen_start, toe_listen_start_event, NULL, EVENTHANDLER_PRI_ANY); listen_stop_eh = EVENTHANDLER_REGISTER(tcp_offload_listen_stop, toe_listen_stop_event, NULL, EVENTHANDLER_PRI_ANY); lle_event_eh = EVENTHANDLER_REGISTER(lle_event, toe_lle_event, NULL, EVENTHANDLER_PRI_ANY); return (0); } static int toecore_unload(void) { mtx_lock(&toedev_lock); if (!TAILQ_EMPTY(&toedev_list)) { mtx_unlock(&toedev_lock); return (EBUSY); } EVENTHANDLER_DEREGISTER(tcp_offload_listen_start, listen_start_eh); EVENTHANDLER_DEREGISTER(tcp_offload_listen_stop, listen_stop_eh); EVENTHANDLER_DEREGISTER(lle_event, lle_event_eh); mtx_unlock(&toedev_lock); mtx_destroy(&toedev_lock); return (0); } static int toecore_mod_handler(module_t mod, int cmd, void *arg) { if (cmd == MOD_LOAD) return (toecore_load()); if (cmd == MOD_UNLOAD) return (toecore_unload()); return (EOPNOTSUPP); } static moduledata_t mod_data= { "toecore", toecore_mod_handler, 0 }; MODULE_VERSION(toecore, 1); DECLARE_MODULE(toecore, mod_data, SI_SUB_EXEC, SI_ORDER_ANY); diff --git a/sys/netinet6/tcp6_var.h b/sys/netinet6/tcp6_var.h index 7c758fbd3479..2e411963676e 100644 --- a/sys/netinet6/tcp6_var.h +++ b/sys/netinet6/tcp6_var.h @@ -1,84 +1,86 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the project nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ /*- * Copyright (c) 1982, 1986, 1993, 1994, 1995 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)tcp_var.h 8.4 (Berkeley) 5/24/95 * $FreeBSD$ */ #ifndef _NETINET_TCP6_VAR_H_ #define _NETINET_TCP6_VAR_H_ #ifdef _KERNEL #ifdef SYSCTL_DECL SYSCTL_DECL(_net_inet6_tcp6); VNET_DECLARE(int, tcp_v6mssdflt); /* XXX */ #define V_tcp_v6mssdflt VNET(tcp_v6mssdflt) #endif struct ip6_hdr; void tcp6_ctlinput(int, struct sockaddr *, void *); +void tcp6_ctlinput_viaudp(int, struct sockaddr *, void *, void *); void tcp6_init(void); int tcp6_input(struct mbuf **, int *, int); +int tcp6_input_with_port(struct mbuf **, int *, int, uint16_t); extern struct pr_usrreqs tcp6_usrreqs; #endif /* _KERNEL */ #endif /* _NETINET_TCP6_VAR_H_ */ diff --git a/sys/sys/mbuf.h b/sys/sys/mbuf.h index 371ae8feae46..dd92103cb1fd 100644 --- a/sys/sys/mbuf.h +++ b/sys/sys/mbuf.h @@ -1,1630 +1,1631 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1988, 1993 * The Regents of the University of California. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)mbuf.h 8.5 (Berkeley) 2/19/95 * $FreeBSD$ */ #ifndef _SYS_MBUF_H_ #define _SYS_MBUF_H_ /* XXX: These includes suck. Sorry! */ #include #ifdef _KERNEL #include #include #include #ifdef WITNESS #include #endif #endif #ifdef _KERNEL #include #define MBUF_PROBE1(probe, arg0) \ SDT_PROBE1(sdt, , , probe, arg0) #define MBUF_PROBE2(probe, arg0, arg1) \ SDT_PROBE2(sdt, , , probe, arg0, arg1) #define MBUF_PROBE3(probe, arg0, arg1, arg2) \ SDT_PROBE3(sdt, , , probe, arg0, arg1, arg2) #define MBUF_PROBE4(probe, arg0, arg1, arg2, arg3) \ SDT_PROBE4(sdt, , , probe, arg0, arg1, arg2, arg3) #define MBUF_PROBE5(probe, arg0, arg1, arg2, arg3, arg4) \ SDT_PROBE5(sdt, , , probe, arg0, arg1, arg2, arg3, arg4) SDT_PROBE_DECLARE(sdt, , , m__init); SDT_PROBE_DECLARE(sdt, , , m__gethdr); SDT_PROBE_DECLARE(sdt, , , m__get); SDT_PROBE_DECLARE(sdt, , , m__getcl); SDT_PROBE_DECLARE(sdt, , , m__getjcl); SDT_PROBE_DECLARE(sdt, , , m__clget); SDT_PROBE_DECLARE(sdt, , , m__cljget); SDT_PROBE_DECLARE(sdt, , , m__cljset); SDT_PROBE_DECLARE(sdt, , , m__free); SDT_PROBE_DECLARE(sdt, , , m__freem); #endif /* _KERNEL */ /* * Mbufs are of a single size, MSIZE (sys/param.h), which includes overhead. * An mbuf may add a single "mbuf cluster" of size MCLBYTES (also in * sys/param.h), which has no additional overhead and is used instead of the * internal data area; this is done when at least MINCLSIZE of data must be * stored. Additionally, it is possible to allocate a separate buffer * externally and attach it to the mbuf in a way similar to that of mbuf * clusters. * * NB: These calculation do not take actual compiler-induced alignment and * padding inside the complete struct mbuf into account. Appropriate * attention is required when changing members of struct mbuf. * * MLEN is data length in a normal mbuf. * MHLEN is data length in an mbuf with pktheader. * MINCLSIZE is a smallest amount of data that should be put into cluster. * * Compile-time assertions in uipc_mbuf.c test these values to ensure that * they are sensible. */ struct mbuf; #define MHSIZE offsetof(struct mbuf, m_dat) #define MPKTHSIZE offsetof(struct mbuf, m_pktdat) #define MLEN ((int)(MSIZE - MHSIZE)) #define MHLEN ((int)(MSIZE - MPKTHSIZE)) #define MINCLSIZE (MHLEN + 1) #define M_NODOM 255 #ifdef _KERNEL /*- * Macro for type conversion: convert mbuf pointer to data pointer of correct * type: * * mtod(m, t) -- Convert mbuf pointer to data pointer of correct type. * mtodo(m, o) -- Same as above but with offset 'o' into data. */ #define mtod(m, t) ((t)((m)->m_data)) #define mtodo(m, o) ((void *)(((m)->m_data) + (o))) /* * Argument structure passed to UMA routines during mbuf and packet * allocations. */ struct mb_args { int flags; /* Flags for mbuf being allocated */ short type; /* Type of mbuf being allocated */ }; #endif /* _KERNEL */ /* * Packet tag structure (see below for details). */ struct m_tag { SLIST_ENTRY(m_tag) m_tag_link; /* List of packet tags */ u_int16_t m_tag_id; /* Tag ID */ u_int16_t m_tag_len; /* Length of data */ u_int32_t m_tag_cookie; /* ABI/Module ID */ void (*m_tag_free)(struct m_tag *); }; /* * Static network interface owned tag. * Allocated through ifp->if_snd_tag_alloc(). */ struct m_snd_tag { struct ifnet *ifp; /* network interface tag belongs to */ volatile u_int refcount; u_int type; /* One of IF_SND_TAG_TYPE_*. */ }; /* * Record/packet header in first mbuf of chain; valid only if M_PKTHDR is set. * Size ILP32: 48 * LP64: 56 * Compile-time assertions in uipc_mbuf.c test these values to ensure that * they are correct. */ struct pkthdr { union { struct m_snd_tag *snd_tag; /* send tag, if any */ struct ifnet *rcvif; /* rcv interface */ }; SLIST_HEAD(packet_tags, m_tag) tags; /* list of packet tags */ int32_t len; /* total packet length */ /* Layer crossing persistent information. */ uint32_t flowid; /* packet's 4-tuple system */ uint32_t csum_flags; /* checksum and offload features */ uint16_t fibnum; /* this packet should use this fib */ uint8_t numa_domain; /* NUMA domain of recvd pkt */ uint8_t rsstype; /* hash type */ union { uint64_t rcv_tstmp; /* timestamp in ns */ struct { uint8_t l2hlen; /* layer 2 hdr len */ uint8_t l3hlen; /* layer 3 hdr len */ uint8_t l4hlen; /* layer 4 hdr len */ uint8_t l5hlen; /* layer 5 hdr len */ uint8_t inner_l2hlen; uint8_t inner_l3hlen; uint8_t inner_l4hlen; uint8_t inner_l5hlen; }; }; union { uint8_t eight[8]; uint16_t sixteen[4]; uint32_t thirtytwo[2]; uint64_t sixtyfour[1]; uintptr_t unintptr[1]; void *ptr; } PH_per; /* Layer specific non-persistent local storage for reassembly, etc. */ union { uint8_t eight[8]; uint16_t sixteen[4]; uint32_t thirtytwo[2]; uint64_t sixtyfour[1]; uintptr_t unintptr[1]; void *ptr; } PH_loc; }; #define ether_vtag PH_per.sixteen[0] +#define tcp_tun_port PH_per.sixteen[0] /* outbound */ #define PH_vt PH_per #define vt_nrecs sixteen[0] /* mld and v6-ND */ #define tso_segsz PH_per.sixteen[1] /* inbound after LRO */ #define lro_nsegs tso_segsz /* inbound after LRO */ #define csum_data PH_per.thirtytwo[1] /* inbound from hardware up */ #define lro_len PH_loc.sixteen[0] /* inbound during LRO (no reassembly) */ #define lro_csum PH_loc.sixteen[1] /* inbound during LRO (no reassembly) */ #define lro_etype PH_loc.sixteen[2] /* inbound during LRO (no reassembly) */ /* Note PH_loc is used during IP reassembly (all 8 bytes as a ptr) */ /* * TLS records for TLS 1.0-1.2 can have the following header lengths: * - 5 (AES-CBC with implicit IV) * - 21 (AES-CBC with explicit IV) * - 13 (AES-GCM with 8 byte explicit IV) */ #define MBUF_PEXT_HDR_LEN 23 /* * TLS records for TLS 1.0-1.2 can have the following maximum trailer * lengths: * - 16 (AES-GCM) * - 36 (AES-CBC with SHA1 and up to 16 bytes of padding) * - 48 (AES-CBC with SHA2-256 and up to 16 bytes of padding) * - 64 (AES-CBC with SHA2-384 and up to 16 bytes of padding) */ #define MBUF_PEXT_TRAIL_LEN 64 #if defined(__LP64__) #define MBUF_PEXT_MAX_PGS (40 / sizeof(vm_paddr_t)) #else #define MBUF_PEXT_MAX_PGS (72 / sizeof(vm_paddr_t)) #endif #define MBUF_PEXT_MAX_BYTES \ (MBUF_PEXT_MAX_PGS * PAGE_SIZE + MBUF_PEXT_HDR_LEN + MBUF_PEXT_TRAIL_LEN) struct ktls_session; struct socket; /* * Description of external storage mapped into mbuf; valid only if M_EXT is * set. * Size ILP32: 28 * LP64: 48 * Compile-time assertions in uipc_mbuf.c test these values to ensure that * they are correct. */ typedef void m_ext_free_t(struct mbuf *); struct m_ext { union { /* * If EXT_FLAG_EMBREF is set, then we use refcount in the * mbuf, the 'ext_count' member. Otherwise, we have a * shadow copy and we use pointer 'ext_cnt'. The original * mbuf is responsible to carry the pointer to free routine * and its arguments. They aren't copied into shadows in * mb_dupcl() to avoid dereferencing next cachelines. */ volatile u_int ext_count; volatile u_int *ext_cnt; }; uint32_t ext_size; /* size of buffer, for ext_free */ uint32_t ext_type:8, /* type of external storage */ ext_flags:24; /* external storage mbuf flags */ union { struct { /* * Regular M_EXT mbuf: * o ext_buf always points to the external buffer. * o ext_free (below) and two optional arguments * ext_arg1 and ext_arg2 store the free context for * the external storage. They are set only in the * refcount carrying mbuf, the one with * EXT_FLAG_EMBREF flag, with exclusion for * EXT_EXTREF type, where the free context is copied * into all mbufs that use same external storage. */ char *ext_buf; /* start of buffer */ #define m_ext_copylen offsetof(struct m_ext, ext_arg2) void *ext_arg2; }; struct { /* * Multi-page M_EXTPG mbuf: * o extpg_pa - page vector. * o extpg_trail and extpg_hdr - TLS trailer and * header. * Uses ext_free and may also use ext_arg1. */ vm_paddr_t extpg_pa[MBUF_PEXT_MAX_PGS]; char extpg_trail[MBUF_PEXT_TRAIL_LEN]; char extpg_hdr[MBUF_PEXT_HDR_LEN]; /* Pretend these 3 fields are part of mbuf itself. */ #define m_epg_pa m_ext.extpg_pa #define m_epg_trail m_ext.extpg_trail #define m_epg_hdr m_ext.extpg_hdr #define m_epg_ext_copylen offsetof(struct m_ext, ext_free) }; }; /* * Free method and optional argument pointer, both * used by M_EXT and M_EXTPG. */ m_ext_free_t *ext_free; void *ext_arg1; }; /* * The core of the mbuf object along with some shortcut defines for practical * purposes. */ struct mbuf { /* * Header present at the beginning of every mbuf. * Size ILP32: 24 * LP64: 32 * Compile-time assertions in uipc_mbuf.c test these values to ensure * that they are correct. */ union { /* next buffer in chain */ struct mbuf *m_next; SLIST_ENTRY(mbuf) m_slist; STAILQ_ENTRY(mbuf) m_stailq; }; union { /* next chain in queue/record */ struct mbuf *m_nextpkt; SLIST_ENTRY(mbuf) m_slistpkt; STAILQ_ENTRY(mbuf) m_stailqpkt; }; caddr_t m_data; /* location of data */ int32_t m_len; /* amount of data in this mbuf */ uint32_t m_type:8, /* type of data in this mbuf */ m_flags:24; /* flags; see below */ #if !defined(__LP64__) uint32_t m_pad; /* pad for 64bit alignment */ #endif /* * A set of optional headers (packet header, external storage header) * and internal data storage. Historically, these arrays were sized * to MHLEN (space left after a packet header) and MLEN (space left * after only a regular mbuf header); they are now variable size in * order to support future work on variable-size mbufs. */ union { struct { union { /* M_PKTHDR set. */ struct pkthdr m_pkthdr; /* M_EXTPG set. * Multi-page M_EXTPG mbuf has its meta data * split between the below anonymous structure * and m_ext. It carries vector of pages, * optional header and trailer char vectors * and pointers to socket/TLS data. */ #define m_epg_startcopy m_epg_npgs #define m_epg_endcopy m_epg_stailq struct { /* Overall count of pages and count of * pages with I/O pending. */ uint8_t m_epg_npgs; uint8_t m_epg_nrdy; /* TLS header and trailer lengths. * The data itself resides in m_ext. */ uint8_t m_epg_hdrlen; uint8_t m_epg_trllen; /* Offset into 1st page and length of * data in the last page. */ uint16_t m_epg_1st_off; uint16_t m_epg_last_len; uint8_t m_epg_flags; #define EPG_FLAG_ANON 0x1 /* Data can be encrypted in place. */ #define EPG_FLAG_2FREE 0x2 /* Scheduled for free. */ uint8_t m_epg_record_type; uint8_t __spare[2]; int m_epg_enc_cnt; struct ktls_session *m_epg_tls; struct socket *m_epg_so; uint64_t m_epg_seqno; STAILQ_ENTRY(mbuf) m_epg_stailq; }; }; union { /* M_EXT or M_EXTPG set. */ struct m_ext m_ext; /* M_PKTHDR set, neither M_EXT nor M_EXTPG. */ char m_pktdat[0]; }; }; char m_dat[0]; /* !M_PKTHDR, !M_EXT */ }; }; #ifdef _KERNEL static inline int m_epg_pagelen(const struct mbuf *m, int pidx, int pgoff) { KASSERT(pgoff == 0 || pidx == 0, ("page %d with non-zero offset %d in %p", pidx, pgoff, m)); if (pidx == m->m_epg_npgs - 1) { return (m->m_epg_last_len); } else { return (PAGE_SIZE - pgoff); } } #ifdef INVARIANTS #define MCHECK(ex, msg) KASSERT((ex), \ ("Multi page mbuf %p with " #msg " at %s:%d", \ m, __FILE__, __LINE__)) /* * NB: This expects a non-empty buffer (npgs > 0 and * last_pg_len > 0). */ #define MBUF_EXT_PGS_ASSERT_SANITY(m) do { \ MCHECK(m->m_epg_npgs > 0, "no valid pages"); \ MCHECK(m->m_epg_npgs <= nitems(m->m_epg_pa), \ "too many pages"); \ MCHECK(m->m_epg_nrdy <= m->m_epg_npgs, \ "too many ready pages"); \ MCHECK(m->m_epg_1st_off < PAGE_SIZE, \ "too large page offset"); \ MCHECK(m->m_epg_last_len > 0, "zero last page length"); \ MCHECK(m->m_epg_last_len <= PAGE_SIZE, \ "too large last page length"); \ if (m->m_epg_npgs == 1) \ MCHECK(m->m_epg_1st_off + \ m->m_epg_last_len <= PAGE_SIZE, \ "single page too large"); \ MCHECK(m->m_epg_hdrlen <= sizeof(m->m_epg_hdr), \ "too large header length"); \ MCHECK(m->m_epg_trllen <= sizeof(m->m_epg_trail), \ "too large header length"); \ } while (0) #else #define MBUF_EXT_PGS_ASSERT_SANITY(m) do {} while (0); #endif #endif /* * mbuf flags of global significance and layer crossing. * Those of only protocol/layer specific significance are to be mapped * to M_PROTO[1-11] and cleared at layer handoff boundaries. * NB: Limited to the lower 24 bits. */ #define M_EXT 0x00000001 /* has associated external storage */ #define M_PKTHDR 0x00000002 /* start of record */ #define M_EOR 0x00000004 /* end of record */ #define M_RDONLY 0x00000008 /* associated data is marked read-only */ #define M_BCAST 0x00000010 /* send/received as link-level broadcast */ #define M_MCAST 0x00000020 /* send/received as link-level multicast */ #define M_PROMISC 0x00000040 /* packet was not for us */ #define M_VLANTAG 0x00000080 /* ether_vtag is valid */ #define M_EXTPG 0x00000100 /* has array of unmapped pages and TLS */ #define M_NOFREE 0x00000200 /* do not free mbuf, embedded in cluster */ #define M_TSTMP 0x00000400 /* rcv_tstmp field is valid */ #define M_TSTMP_HPREC 0x00000800 /* rcv_tstmp is high-prec, typically hw-stamped on port (useful for IEEE 1588 and 802.1AS) */ #define M_TSTMP_LRO 0x00001000 /* Time LRO pushed in pkt is valid in (PH_loc) */ #define M_PROTO1 0x00002000 /* protocol-specific */ #define M_PROTO2 0x00004000 /* protocol-specific */ #define M_PROTO3 0x00008000 /* protocol-specific */ #define M_PROTO4 0x00010000 /* protocol-specific */ #define M_PROTO5 0x00020000 /* protocol-specific */ #define M_PROTO6 0x00040000 /* protocol-specific */ #define M_PROTO7 0x00080000 /* protocol-specific */ #define M_PROTO8 0x00100000 /* protocol-specific */ #define M_PROTO9 0x00200000 /* protocol-specific */ #define M_PROTO10 0x00400000 /* protocol-specific */ #define M_PROTO11 0x00800000 /* protocol-specific */ #define MB_DTOR_SKIP 0x1 /* don't pollute the cache by touching a freed mbuf */ /* * Flags to purge when crossing layers. */ #define M_PROTOFLAGS \ (M_PROTO1|M_PROTO2|M_PROTO3|M_PROTO4|M_PROTO5|M_PROTO6|M_PROTO7|M_PROTO8|\ M_PROTO9|M_PROTO10|M_PROTO11) /* * Flags preserved when copying m_pkthdr. */ #define M_COPYFLAGS \ (M_PKTHDR|M_EOR|M_RDONLY|M_BCAST|M_MCAST|M_PROMISC|M_VLANTAG|M_TSTMP| \ M_TSTMP_HPREC|M_TSTMP_LRO|M_PROTOFLAGS) /* * Mbuf flag description for use with printf(9) %b identifier. */ #define M_FLAG_BITS \ "\20\1M_EXT\2M_PKTHDR\3M_EOR\4M_RDONLY\5M_BCAST\6M_MCAST" \ "\7M_PROMISC\10M_VLANTAG\11M_EXTPG\12M_NOFREE\13M_TSTMP\14M_TSTMP_HPREC\15M_TSTMP_LRO" #define M_FLAG_PROTOBITS \ "\16M_PROTO1\17M_PROTO2\20M_PROTO3\21M_PROTO4" \ "\22M_PROTO5\23M_PROTO6\24M_PROTO7\25M_PROTO8\26M_PROTO9" \ "\27M_PROTO10\28M_PROTO11" #define M_FLAG_PRINTF (M_FLAG_BITS M_FLAG_PROTOBITS) /* * Network interface cards are able to hash protocol fields (such as IPv4 * addresses and TCP port numbers) classify packets into flows. These flows * can then be used to maintain ordering while delivering packets to the OS * via parallel input queues, as well as to provide a stateless affinity * model. NIC drivers can pass up the hash via m->m_pkthdr.flowid, and set * m_flag fields to indicate how the hash should be interpreted by the * network stack. * * Most NICs support RSS, which provides ordering and explicit affinity, and * use the hash m_flag bits to indicate what header fields were covered by * the hash. M_HASHTYPE_OPAQUE and M_HASHTYPE_OPAQUE_HASH can be set by non- * RSS cards or configurations that provide an opaque flow identifier, allowing * for ordering and distribution without explicit affinity. Additionally, * M_HASHTYPE_OPAQUE_HASH indicates that the flow identifier has hash * properties. * * The meaning of the IPV6_EX suffix: * "o Home address from the home address option in the IPv6 destination * options header. If the extension header is not present, use the Source * IPv6 Address. * o IPv6 address that is contained in the Routing-Header-Type-2 from the * associated extension header. If the extension header is not present, * use the Destination IPv6 Address." * Quoted from: * https://docs.microsoft.com/en-us/windows-hardware/drivers/network/rss-hashing-types#ndishashipv6ex */ #define M_HASHTYPE_HASHPROP 0x80 /* has hash properties */ #define M_HASHTYPE_INNER 0x40 /* calculated from inner headers */ #define M_HASHTYPE_HASH(t) (M_HASHTYPE_HASHPROP | (t)) /* Microsoft RSS standard hash types */ #define M_HASHTYPE_NONE 0 #define M_HASHTYPE_RSS_IPV4 M_HASHTYPE_HASH(1) /* IPv4 2-tuple */ #define M_HASHTYPE_RSS_TCP_IPV4 M_HASHTYPE_HASH(2) /* TCPv4 4-tuple */ #define M_HASHTYPE_RSS_IPV6 M_HASHTYPE_HASH(3) /* IPv6 2-tuple */ #define M_HASHTYPE_RSS_TCP_IPV6 M_HASHTYPE_HASH(4) /* TCPv6 4-tuple */ #define M_HASHTYPE_RSS_IPV6_EX M_HASHTYPE_HASH(5) /* IPv6 2-tuple + * ext hdrs */ #define M_HASHTYPE_RSS_TCP_IPV6_EX M_HASHTYPE_HASH(6) /* TCPv6 4-tuple + * ext hdrs */ #define M_HASHTYPE_RSS_UDP_IPV4 M_HASHTYPE_HASH(7) /* IPv4 UDP 4-tuple*/ #define M_HASHTYPE_RSS_UDP_IPV6 M_HASHTYPE_HASH(9) /* IPv6 UDP 4-tuple*/ #define M_HASHTYPE_RSS_UDP_IPV6_EX M_HASHTYPE_HASH(10)/* IPv6 UDP 4-tuple + * ext hdrs */ #define M_HASHTYPE_OPAQUE 0x3f /* ordering, not affinity */ #define M_HASHTYPE_OPAQUE_HASH M_HASHTYPE_HASH(M_HASHTYPE_OPAQUE) /* ordering+hash, not affinity*/ #define M_HASHTYPE_CLEAR(m) ((m)->m_pkthdr.rsstype = 0) #define M_HASHTYPE_GET(m) ((m)->m_pkthdr.rsstype & ~M_HASHTYPE_INNER) #define M_HASHTYPE_SET(m, v) ((m)->m_pkthdr.rsstype = (v)) #define M_HASHTYPE_TEST(m, v) (M_HASHTYPE_GET(m) == (v)) #define M_HASHTYPE_ISHASH(m) \ (((m)->m_pkthdr.rsstype & M_HASHTYPE_HASHPROP) != 0) #define M_HASHTYPE_SETINNER(m) do { \ (m)->m_pkthdr.rsstype |= M_HASHTYPE_INNER; \ } while (0) /* * External mbuf storage buffer types. */ #define EXT_CLUSTER 1 /* mbuf cluster */ #define EXT_SFBUF 2 /* sendfile(2)'s sf_buf */ #define EXT_JUMBOP 3 /* jumbo cluster page sized */ #define EXT_JUMBO9 4 /* jumbo cluster 9216 bytes */ #define EXT_JUMBO16 5 /* jumbo cluster 16184 bytes */ #define EXT_PACKET 6 /* mbuf+cluster from packet zone */ #define EXT_MBUF 7 /* external mbuf reference */ #define EXT_RXRING 8 /* data in NIC receive ring */ #define EXT_VENDOR1 224 /* for vendor-internal use */ #define EXT_VENDOR2 225 /* for vendor-internal use */ #define EXT_VENDOR3 226 /* for vendor-internal use */ #define EXT_VENDOR4 227 /* for vendor-internal use */ #define EXT_EXP1 244 /* for experimental use */ #define EXT_EXP2 245 /* for experimental use */ #define EXT_EXP3 246 /* for experimental use */ #define EXT_EXP4 247 /* for experimental use */ #define EXT_NET_DRV 252 /* custom ext_buf provided by net driver(s) */ #define EXT_MOD_TYPE 253 /* custom module's ext_buf type */ #define EXT_DISPOSABLE 254 /* can throw this buffer away w/page flipping */ #define EXT_EXTREF 255 /* has externally maintained ext_cnt ptr */ /* * Flags for external mbuf buffer types. * NB: limited to the lower 24 bits. */ #define EXT_FLAG_EMBREF 0x000001 /* embedded ext_count */ #define EXT_FLAG_EXTREF 0x000002 /* external ext_cnt, notyet */ #define EXT_FLAG_NOFREE 0x000010 /* don't free mbuf to pool, notyet */ #define EXT_FLAG_VENDOR1 0x010000 /* These flags are vendor */ #define EXT_FLAG_VENDOR2 0x020000 /* or submodule specific, */ #define EXT_FLAG_VENDOR3 0x040000 /* not used by mbuf code. */ #define EXT_FLAG_VENDOR4 0x080000 /* Set/read by submodule. */ #define EXT_FLAG_EXP1 0x100000 /* for experimental use */ #define EXT_FLAG_EXP2 0x200000 /* for experimental use */ #define EXT_FLAG_EXP3 0x400000 /* for experimental use */ #define EXT_FLAG_EXP4 0x800000 /* for experimental use */ /* * EXT flag description for use with printf(9) %b identifier. */ #define EXT_FLAG_BITS \ "\20\1EXT_FLAG_EMBREF\2EXT_FLAG_EXTREF\5EXT_FLAG_NOFREE" \ "\21EXT_FLAG_VENDOR1\22EXT_FLAG_VENDOR2\23EXT_FLAG_VENDOR3" \ "\24EXT_FLAG_VENDOR4\25EXT_FLAG_EXP1\26EXT_FLAG_EXP2\27EXT_FLAG_EXP3" \ "\30EXT_FLAG_EXP4" /* * Flags indicating checksum, segmentation and other offload work to be * done, or already done, by hardware or lower layers. It is split into * separate inbound and outbound flags. * * Outbound flags that are set by upper protocol layers requesting lower * layers, or ideally the hardware, to perform these offloading tasks. * For outbound packets this field and its flags can be directly tested * against ifnet if_hwassist. Note that the outbound and the inbound flags do * not collide right now but they could be allowed to (as long as the flags are * scrubbed appropriately when the direction of an mbuf changes). CSUM_BITS * would also have to split into CSUM_BITS_TX and CSUM_BITS_RX. * * CSUM_INNER_ is the same as CSUM_ but it applies to the inner frame. * The CSUM_ENCAP_ bits identify the outer encapsulation. */ #define CSUM_IP 0x00000001 /* IP header checksum offload */ #define CSUM_IP_UDP 0x00000002 /* UDP checksum offload */ #define CSUM_IP_TCP 0x00000004 /* TCP checksum offload */ #define CSUM_IP_SCTP 0x00000008 /* SCTP checksum offload */ #define CSUM_IP_TSO 0x00000010 /* TCP segmentation offload */ #define CSUM_IP_ISCSI 0x00000020 /* iSCSI checksum offload */ #define CSUM_INNER_IP6_UDP 0x00000040 #define CSUM_INNER_IP6_TCP 0x00000080 #define CSUM_INNER_IP6_TSO 0x00000100 #define CSUM_IP6_UDP 0x00000200 /* UDP checksum offload */ #define CSUM_IP6_TCP 0x00000400 /* TCP checksum offload */ #define CSUM_IP6_SCTP 0x00000800 /* SCTP checksum offload */ #define CSUM_IP6_TSO 0x00001000 /* TCP segmentation offload */ #define CSUM_IP6_ISCSI 0x00002000 /* iSCSI checksum offload */ #define CSUM_INNER_IP 0x00004000 #define CSUM_INNER_IP_UDP 0x00008000 #define CSUM_INNER_IP_TCP 0x00010000 #define CSUM_INNER_IP_TSO 0x00020000 #define CSUM_ENCAP_VXLAN 0x00040000 /* VXLAN outer encapsulation */ #define CSUM_ENCAP_RSVD1 0x00080000 /* Inbound checksum support where the checksum was verified by hardware. */ #define CSUM_INNER_L3_CALC 0x00100000 #define CSUM_INNER_L3_VALID 0x00200000 #define CSUM_INNER_L4_CALC 0x00400000 #define CSUM_INNER_L4_VALID 0x00800000 #define CSUM_L3_CALC 0x01000000 /* calculated layer 3 csum */ #define CSUM_L3_VALID 0x02000000 /* checksum is correct */ #define CSUM_L4_CALC 0x04000000 /* calculated layer 4 csum */ #define CSUM_L4_VALID 0x08000000 /* checksum is correct */ #define CSUM_L5_CALC 0x10000000 /* calculated layer 5 csum */ #define CSUM_L5_VALID 0x20000000 /* checksum is correct */ #define CSUM_COALESCED 0x40000000 /* contains merged segments */ #define CSUM_SND_TAG 0x80000000 /* Packet header has send tag */ #define CSUM_FLAGS_TX (CSUM_IP | CSUM_IP_UDP | CSUM_IP_TCP | CSUM_IP_SCTP | \ CSUM_IP_TSO | CSUM_IP_ISCSI | CSUM_INNER_IP6_UDP | CSUM_INNER_IP6_TCP | \ CSUM_INNER_IP6_TSO | CSUM_IP6_UDP | CSUM_IP6_TCP | CSUM_IP6_SCTP | \ CSUM_IP6_TSO | CSUM_IP6_ISCSI | CSUM_INNER_IP | CSUM_INNER_IP_UDP | \ CSUM_INNER_IP_TCP | CSUM_INNER_IP_TSO | CSUM_ENCAP_VXLAN | \ CSUM_ENCAP_RSVD1 | CSUM_SND_TAG) #define CSUM_FLAGS_RX (CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID | \ CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID | CSUM_L3_CALC | CSUM_L3_VALID | \ CSUM_L4_CALC | CSUM_L4_VALID | CSUM_L5_CALC | CSUM_L5_VALID | \ CSUM_COALESCED) /* * CSUM flag description for use with printf(9) %b identifier. */ #define CSUM_BITS \ "\20\1CSUM_IP\2CSUM_IP_UDP\3CSUM_IP_TCP\4CSUM_IP_SCTP\5CSUM_IP_TSO" \ "\6CSUM_IP_ISCSI\7CSUM_INNER_IP6_UDP\10CSUM_INNER_IP6_TCP" \ "\11CSUM_INNER_IP6_TSO\12CSUM_IP6_UDP\13CSUM_IP6_TCP\14CSUM_IP6_SCTP" \ "\15CSUM_IP6_TSO\16CSUM_IP6_ISCSI\17CSUM_INNER_IP\20CSUM_INNER_IP_UDP" \ "\21CSUM_INNER_IP_TCP\22CSUM_INNER_IP_TSO\23CSUM_ENCAP_VXLAN" \ "\24CSUM_ENCAP_RSVD1\25CSUM_INNER_L3_CALC\26CSUM_INNER_L3_VALID" \ "\27CSUM_INNER_L4_CALC\30CSUM_INNER_L4_VALID\31CSUM_L3_CALC" \ "\32CSUM_L3_VALID\33CSUM_L4_CALC\34CSUM_L4_VALID\35CSUM_L5_CALC" \ "\36CSUM_L5_VALID\37CSUM_COALESCED\40CSUM_SND_TAG" /* CSUM flags compatibility mappings. */ #define CSUM_IP_CHECKED CSUM_L3_CALC #define CSUM_IP_VALID CSUM_L3_VALID #define CSUM_DATA_VALID CSUM_L4_VALID #define CSUM_PSEUDO_HDR CSUM_L4_CALC #define CSUM_SCTP_VALID CSUM_L4_VALID #define CSUM_DELAY_DATA (CSUM_TCP|CSUM_UDP) #define CSUM_DELAY_IP CSUM_IP /* Only v4, no v6 IP hdr csum */ #define CSUM_DELAY_DATA_IPV6 (CSUM_TCP_IPV6|CSUM_UDP_IPV6) #define CSUM_DATA_VALID_IPV6 CSUM_DATA_VALID #define CSUM_TCP CSUM_IP_TCP #define CSUM_UDP CSUM_IP_UDP #define CSUM_SCTP CSUM_IP_SCTP #define CSUM_TSO (CSUM_IP_TSO|CSUM_IP6_TSO) #define CSUM_INNER_TSO (CSUM_INNER_IP_TSO|CSUM_INNER_IP6_TSO) #define CSUM_UDP_IPV6 CSUM_IP6_UDP #define CSUM_TCP_IPV6 CSUM_IP6_TCP #define CSUM_SCTP_IPV6 CSUM_IP6_SCTP /* * mbuf types describing the content of the mbuf (including external storage). */ #define MT_NOTMBUF 0 /* USED INTERNALLY ONLY! Object is not mbuf */ #define MT_DATA 1 /* dynamic (data) allocation */ #define MT_HEADER MT_DATA /* packet header, use M_PKTHDR instead */ #define MT_VENDOR1 4 /* for vendor-internal use */ #define MT_VENDOR2 5 /* for vendor-internal use */ #define MT_VENDOR3 6 /* for vendor-internal use */ #define MT_VENDOR4 7 /* for vendor-internal use */ #define MT_SONAME 8 /* socket name */ #define MT_EXP1 9 /* for experimental use */ #define MT_EXP2 10 /* for experimental use */ #define MT_EXP3 11 /* for experimental use */ #define MT_EXP4 12 /* for experimental use */ #define MT_CONTROL 14 /* extra-data protocol message */ #define MT_EXTCONTROL 15 /* control message with externalized contents */ #define MT_OOBDATA 16 /* expedited data */ #define MT_NOINIT 255 /* Not a type but a flag to allocate a non-initialized mbuf */ /* * String names of mbuf-related UMA(9) and malloc(9) types. Exposed to * !_KERNEL so that monitoring tools can look up the zones with * libmemstat(3). */ #define MBUF_MEM_NAME "mbuf" #define MBUF_CLUSTER_MEM_NAME "mbuf_cluster" #define MBUF_PACKET_MEM_NAME "mbuf_packet" #define MBUF_JUMBOP_MEM_NAME "mbuf_jumbo_page" #define MBUF_JUMBO9_MEM_NAME "mbuf_jumbo_9k" #define MBUF_JUMBO16_MEM_NAME "mbuf_jumbo_16k" #define MBUF_TAG_MEM_NAME "mbuf_tag" #define MBUF_EXTREFCNT_MEM_NAME "mbuf_ext_refcnt" #define MBUF_EXTPGS_MEM_NAME "mbuf_extpgs" #ifdef _KERNEL union if_snd_tag_alloc_params; #ifdef WITNESS #define MBUF_CHECKSLEEP(how) do { \ if (how == M_WAITOK) \ WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, \ "Sleeping in \"%s\"", __func__); \ } while (0) #else #define MBUF_CHECKSLEEP(how) #endif /* * Network buffer allocation API * * The rest of it is defined in kern/kern_mbuf.c */ extern uma_zone_t zone_mbuf; extern uma_zone_t zone_clust; extern uma_zone_t zone_pack; extern uma_zone_t zone_jumbop; extern uma_zone_t zone_jumbo9; extern uma_zone_t zone_jumbo16; extern uma_zone_t zone_extpgs; void mb_dupcl(struct mbuf *, struct mbuf *); void mb_free_ext(struct mbuf *); void mb_free_extpg(struct mbuf *); void mb_free_mext_pgs(struct mbuf *); struct mbuf *mb_alloc_ext_pgs(int, m_ext_free_t); struct mbuf *mb_alloc_ext_plus_pages(int, int); struct mbuf *mb_mapped_to_unmapped(struct mbuf *, int, int, int, struct mbuf **); int mb_unmapped_compress(struct mbuf *m); struct mbuf *mb_unmapped_to_ext(struct mbuf *m); void mb_free_notready(struct mbuf *m, int count); void m_adj(struct mbuf *, int); void m_adj_decap(struct mbuf *, int); int m_apply(struct mbuf *, int, int, int (*)(void *, void *, u_int), void *); int m_append(struct mbuf *, int, c_caddr_t); void m_cat(struct mbuf *, struct mbuf *); void m_catpkt(struct mbuf *, struct mbuf *); int m_clget(struct mbuf *m, int how); void *m_cljget(struct mbuf *m, int how, int size); struct mbuf *m_collapse(struct mbuf *, int, int); void m_copyback(struct mbuf *, int, int, c_caddr_t); void m_copydata(const struct mbuf *, int, int, caddr_t); struct mbuf *m_copym(struct mbuf *, int, int, int); struct mbuf *m_copypacket(struct mbuf *, int); void m_copy_pkthdr(struct mbuf *, struct mbuf *); struct mbuf *m_copyup(struct mbuf *, int, int); struct mbuf *m_defrag(struct mbuf *, int); void m_demote_pkthdr(struct mbuf *); void m_demote(struct mbuf *, int, int); struct mbuf *m_devget(char *, int, int, struct ifnet *, void (*)(char *, caddr_t, u_int)); void m_dispose_extcontrolm(struct mbuf *m); struct mbuf *m_dup(const struct mbuf *, int); int m_dup_pkthdr(struct mbuf *, const struct mbuf *, int); void m_extadd(struct mbuf *, char *, u_int, m_ext_free_t, void *, void *, int, int); u_int m_fixhdr(struct mbuf *); struct mbuf *m_fragment(struct mbuf *, int, int); void m_freem(struct mbuf *); struct mbuf *m_get2(int, int, short, int); struct mbuf *m_getjcl(int, short, int, int); struct mbuf *m_getm2(struct mbuf *, int, int, short, int); struct mbuf *m_getptr(struct mbuf *, int, int *); u_int m_length(struct mbuf *, struct mbuf **); int m_mbuftouio(struct uio *, const struct mbuf *, int); int m_unmappedtouio(const struct mbuf *, int, struct uio *, int); void m_move_pkthdr(struct mbuf *, struct mbuf *); int m_pkthdr_init(struct mbuf *, int); struct mbuf *m_prepend(struct mbuf *, int, int); void m_print(const struct mbuf *, int); struct mbuf *m_pulldown(struct mbuf *, int, int, int *); struct mbuf *m_pullup(struct mbuf *, int); int m_sanity(struct mbuf *, int); struct mbuf *m_split(struct mbuf *, int, int); struct mbuf *m_uiotombuf(struct uio *, int, int, int, int); struct mbuf *m_unshare(struct mbuf *, int); int m_snd_tag_alloc(struct ifnet *, union if_snd_tag_alloc_params *, struct m_snd_tag **); void m_snd_tag_init(struct m_snd_tag *, struct ifnet *, u_int); void m_snd_tag_destroy(struct m_snd_tag *); static __inline int m_gettype(int size) { int type; switch (size) { case MSIZE: type = EXT_MBUF; break; case MCLBYTES: type = EXT_CLUSTER; break; #if MJUMPAGESIZE != MCLBYTES case MJUMPAGESIZE: type = EXT_JUMBOP; break; #endif case MJUM9BYTES: type = EXT_JUMBO9; break; case MJUM16BYTES: type = EXT_JUMBO16; break; default: panic("%s: invalid cluster size %d", __func__, size); } return (type); } /* * Associated an external reference counted buffer with an mbuf. */ static __inline void m_extaddref(struct mbuf *m, char *buf, u_int size, u_int *ref_cnt, m_ext_free_t freef, void *arg1, void *arg2) { KASSERT(ref_cnt != NULL, ("%s: ref_cnt not provided", __func__)); atomic_add_int(ref_cnt, 1); m->m_flags |= M_EXT; m->m_ext.ext_buf = buf; m->m_ext.ext_cnt = ref_cnt; m->m_data = m->m_ext.ext_buf; m->m_ext.ext_size = size; m->m_ext.ext_free = freef; m->m_ext.ext_arg1 = arg1; m->m_ext.ext_arg2 = arg2; m->m_ext.ext_type = EXT_EXTREF; m->m_ext.ext_flags = 0; } static __inline uma_zone_t m_getzone(int size) { uma_zone_t zone; switch (size) { case MCLBYTES: zone = zone_clust; break; #if MJUMPAGESIZE != MCLBYTES case MJUMPAGESIZE: zone = zone_jumbop; break; #endif case MJUM9BYTES: zone = zone_jumbo9; break; case MJUM16BYTES: zone = zone_jumbo16; break; default: panic("%s: invalid cluster size %d", __func__, size); } return (zone); } /* * Initialize an mbuf with linear storage. * * Inline because the consumer text overhead will be roughly the same to * initialize or call a function with this many parameters and M_PKTHDR * should go away with constant propagation for !MGETHDR. */ static __inline int m_init(struct mbuf *m, int how, short type, int flags) { int error; m->m_next = NULL; m->m_nextpkt = NULL; m->m_data = m->m_dat; m->m_len = 0; m->m_flags = flags; m->m_type = type; if (flags & M_PKTHDR) error = m_pkthdr_init(m, how); else error = 0; MBUF_PROBE5(m__init, m, how, type, flags, error); return (error); } static __inline struct mbuf * m_get(int how, short type) { struct mbuf *m; struct mb_args args; args.flags = 0; args.type = type; m = uma_zalloc_arg(zone_mbuf, &args, how); MBUF_PROBE3(m__get, how, type, m); return (m); } static __inline struct mbuf * m_gethdr(int how, short type) { struct mbuf *m; struct mb_args args; args.flags = M_PKTHDR; args.type = type; m = uma_zalloc_arg(zone_mbuf, &args, how); MBUF_PROBE3(m__gethdr, how, type, m); return (m); } static __inline struct mbuf * m_getcl(int how, short type, int flags) { struct mbuf *m; struct mb_args args; args.flags = flags; args.type = type; m = uma_zalloc_arg(zone_pack, &args, how); MBUF_PROBE4(m__getcl, how, type, flags, m); return (m); } /* * XXX: m_cljset() is a dangerous API. One must attach only a new, * unreferenced cluster to an mbuf(9). It is not possible to assert * that, so care can be taken only by users of the API. */ static __inline void m_cljset(struct mbuf *m, void *cl, int type) { int size; switch (type) { case EXT_CLUSTER: size = MCLBYTES; break; #if MJUMPAGESIZE != MCLBYTES case EXT_JUMBOP: size = MJUMPAGESIZE; break; #endif case EXT_JUMBO9: size = MJUM9BYTES; break; case EXT_JUMBO16: size = MJUM16BYTES; break; default: panic("%s: unknown cluster type %d", __func__, type); break; } m->m_data = m->m_ext.ext_buf = cl; m->m_ext.ext_free = m->m_ext.ext_arg1 = m->m_ext.ext_arg2 = NULL; m->m_ext.ext_size = size; m->m_ext.ext_type = type; m->m_ext.ext_flags = EXT_FLAG_EMBREF; m->m_ext.ext_count = 1; m->m_flags |= M_EXT; MBUF_PROBE3(m__cljset, m, cl, type); } static __inline void m_chtype(struct mbuf *m, short new_type) { m->m_type = new_type; } static __inline void m_clrprotoflags(struct mbuf *m) { while (m) { m->m_flags &= ~M_PROTOFLAGS; m = m->m_next; } } static __inline struct mbuf * m_last(struct mbuf *m) { while (m->m_next) m = m->m_next; return (m); } static inline u_int m_extrefcnt(struct mbuf *m) { KASSERT(m->m_flags & M_EXT, ("%s: M_EXT missing", __func__)); return ((m->m_ext.ext_flags & EXT_FLAG_EMBREF) ? m->m_ext.ext_count : *m->m_ext.ext_cnt); } /* * mbuf, cluster, and external object allocation macros (for compatibility * purposes). */ #define M_MOVE_PKTHDR(to, from) m_move_pkthdr((to), (from)) #define MGET(m, how, type) ((m) = m_get((how), (type))) #define MGETHDR(m, how, type) ((m) = m_gethdr((how), (type))) #define MCLGET(m, how) m_clget((m), (how)) #define MEXTADD(m, buf, size, free, arg1, arg2, flags, type) \ m_extadd((m), (char *)(buf), (size), (free), (arg1), (arg2), \ (flags), (type)) #define m_getm(m, len, how, type) \ m_getm2((m), (len), (how), (type), M_PKTHDR) /* * Evaluate TRUE if it's safe to write to the mbuf m's data region (this can * be both the local data payload, or an external buffer area, depending on * whether M_EXT is set). */ #define M_WRITABLE(m) (((m)->m_flags & (M_RDONLY | M_EXTPG)) == 0 && \ (!(((m)->m_flags & M_EXT)) || \ (m_extrefcnt(m) == 1))) /* Check if the supplied mbuf has a packet header, or else panic. */ #define M_ASSERTPKTHDR(m) \ KASSERT((m) != NULL && (m)->m_flags & M_PKTHDR, \ ("%s: no mbuf packet header!", __func__)) /* Check if mbuf is multipage. */ #define M_ASSERTEXTPG(m) \ KASSERT(((m)->m_flags & (M_EXTPG|M_PKTHDR)) == M_EXTPG, \ ("%s: m %p is not multipage!", __func__, m)) /* * Ensure that the supplied mbuf is a valid, non-free mbuf. * * XXX: Broken at the moment. Need some UMA magic to make it work again. */ #define M_ASSERTVALID(m) \ KASSERT((((struct mbuf *)m)->m_flags & 0) == 0, \ ("%s: attempted use of a free mbuf!", __func__)) /* * Return the address of the start of the buffer associated with an mbuf, * handling external storage, packet-header mbufs, and regular data mbufs. */ #define M_START(m) \ (((m)->m_flags & M_EXTPG) ? NULL : \ ((m)->m_flags & M_EXT) ? (m)->m_ext.ext_buf : \ ((m)->m_flags & M_PKTHDR) ? &(m)->m_pktdat[0] : \ &(m)->m_dat[0]) /* * Return the size of the buffer associated with an mbuf, handling external * storage, packet-header mbufs, and regular data mbufs. */ #define M_SIZE(m) \ (((m)->m_flags & M_EXT) ? (m)->m_ext.ext_size : \ ((m)->m_flags & M_PKTHDR) ? MHLEN : \ MLEN) /* * Set the m_data pointer of a newly allocated mbuf to place an object of the * specified size at the end of the mbuf, longword aligned. * * NB: Historically, we had M_ALIGN(), MH_ALIGN(), and MEXT_ALIGN() as * separate macros, each asserting that it was called at the proper moment. * This required callers to themselves test the storage type and call the * right one. Rather than require callers to be aware of those layout * decisions, we centralize here. */ static __inline void m_align(struct mbuf *m, int len) { #ifdef INVARIANTS const char *msg = "%s: not a virgin mbuf"; #endif int adjust; KASSERT(m->m_data == M_START(m), (msg, __func__)); adjust = M_SIZE(m) - len; m->m_data += adjust &~ (sizeof(long)-1); } #define M_ALIGN(m, len) m_align(m, len) #define MH_ALIGN(m, len) m_align(m, len) #define MEXT_ALIGN(m, len) m_align(m, len) /* * Compute the amount of space available before the current start of data in * an mbuf. * * The M_WRITABLE() is a temporary, conservative safety measure: the burden * of checking writability of the mbuf data area rests solely with the caller. * * NB: In previous versions, M_LEADINGSPACE() would only check M_WRITABLE() * for mbufs with external storage. We now allow mbuf-embedded data to be * read-only as well. */ #define M_LEADINGSPACE(m) \ (M_WRITABLE(m) ? ((m)->m_data - M_START(m)) : 0) /* * Compute the amount of space available after the end of data in an mbuf. * * The M_WRITABLE() is a temporary, conservative safety measure: the burden * of checking writability of the mbuf data area rests solely with the caller. * * NB: In previous versions, M_TRAILINGSPACE() would only check M_WRITABLE() * for mbufs with external storage. We now allow mbuf-embedded data to be * read-only as well. */ #define M_TRAILINGSPACE(m) \ (M_WRITABLE(m) ? \ ((M_START(m) + M_SIZE(m)) - ((m)->m_data + (m)->m_len)) : 0) /* * Arrange to prepend space of size plen to mbuf m. If a new mbuf must be * allocated, how specifies whether to wait. If the allocation fails, the * original mbuf chain is freed and m is set to NULL. */ #define M_PREPEND(m, plen, how) do { \ struct mbuf **_mmp = &(m); \ struct mbuf *_mm = *_mmp; \ int _mplen = (plen); \ int __mhow = (how); \ \ MBUF_CHECKSLEEP(how); \ if (M_LEADINGSPACE(_mm) >= _mplen) { \ _mm->m_data -= _mplen; \ _mm->m_len += _mplen; \ } else \ _mm = m_prepend(_mm, _mplen, __mhow); \ if (_mm != NULL && _mm->m_flags & M_PKTHDR) \ _mm->m_pkthdr.len += _mplen; \ *_mmp = _mm; \ } while (0) /* * Change mbuf to new type. This is a relatively expensive operation and * should be avoided. */ #define MCHTYPE(m, t) m_chtype((m), (t)) /* Return the rcvif of a packet header. */ static __inline struct ifnet * m_rcvif(struct mbuf *m) { M_ASSERTPKTHDR(m); if (m->m_pkthdr.csum_flags & CSUM_SND_TAG) return (NULL); return (m->m_pkthdr.rcvif); } /* Length to m_copy to copy all. */ #define M_COPYALL 1000000000 extern int max_datalen; /* MHLEN - max_hdr */ extern int max_hdr; /* Largest link + protocol header */ extern int max_linkhdr; /* Largest link-level header */ extern int max_protohdr; /* Largest protocol header */ extern int nmbclusters; /* Maximum number of clusters */ extern bool mb_use_ext_pgs; /* Use ext_pgs for sendfile */ /*- * Network packets may have annotations attached by affixing a list of * "packet tags" to the pkthdr structure. Packet tags are dynamically * allocated semi-opaque data structures that have a fixed header * (struct m_tag) that specifies the size of the memory block and a * pair that identifies it. The cookie is a 32-bit unique * unsigned value used to identify a module or ABI. By convention this value * is chosen as the date+time that the module is created, expressed as the * number of seconds since the epoch (e.g., using date -u +'%s'). The type * value is an ABI/module-specific value that identifies a particular * annotation and is private to the module. For compatibility with systems * like OpenBSD that define packet tags w/o an ABI/module cookie, the value * PACKET_ABI_COMPAT is used to implement m_tag_get and m_tag_find * compatibility shim functions and several tag types are defined below. * Users that do not require compatibility should use a private cookie value * so that packet tag-related definitions can be maintained privately. * * Note that the packet tag returned by m_tag_alloc has the default memory * alignment implemented by malloc. To reference private data one can use a * construct like: * * struct m_tag *mtag = m_tag_alloc(...); * struct foo *p = (struct foo *)(mtag+1); * * if the alignment of struct m_tag is sufficient for referencing members of * struct foo. Otherwise it is necessary to embed struct m_tag within the * private data structure to insure proper alignment; e.g., * * struct foo { * struct m_tag tag; * ... * }; * struct foo *p = (struct foo *) m_tag_alloc(...); * struct m_tag *mtag = &p->tag; */ /* * Persistent tags stay with an mbuf until the mbuf is reclaimed. Otherwise * tags are expected to ``vanish'' when they pass through a network * interface. For most interfaces this happens normally as the tags are * reclaimed when the mbuf is free'd. However in some special cases * reclaiming must be done manually. An example is packets that pass through * the loopback interface. Also, one must be careful to do this when * ``turning around'' packets (e.g., icmp_reflect). * * To mark a tag persistent bit-or this flag in when defining the tag id. * The tag will then be treated as described above. */ #define MTAG_PERSISTENT 0x800 #define PACKET_TAG_NONE 0 /* Nadda */ /* Packet tags for use with PACKET_ABI_COMPAT. */ #define PACKET_TAG_IPSEC_IN_DONE 1 /* IPsec applied, in */ #define PACKET_TAG_IPSEC_OUT_DONE 2 /* IPsec applied, out */ #define PACKET_TAG_IPSEC_IN_CRYPTO_DONE 3 /* NIC IPsec crypto done */ #define PACKET_TAG_IPSEC_OUT_CRYPTO_NEEDED 4 /* NIC IPsec crypto req'ed */ #define PACKET_TAG_IPSEC_IN_COULD_DO_CRYPTO 5 /* NIC notifies IPsec */ #define PACKET_TAG_IPSEC_PENDING_TDB 6 /* Reminder to do IPsec */ #define PACKET_TAG_BRIDGE 7 /* Bridge processing done */ #define PACKET_TAG_GIF 8 /* GIF processing done */ #define PACKET_TAG_GRE 9 /* GRE processing done */ #define PACKET_TAG_IN_PACKET_CHECKSUM 10 /* NIC checksumming done */ #define PACKET_TAG_ENCAP 11 /* Encap. processing */ #define PACKET_TAG_IPSEC_SOCKET 12 /* IPSEC socket ref */ #define PACKET_TAG_IPSEC_HISTORY 13 /* IPSEC history */ #define PACKET_TAG_IPV6_INPUT 14 /* IPV6 input processing */ #define PACKET_TAG_DUMMYNET 15 /* dummynet info */ #define PACKET_TAG_DIVERT 17 /* divert info */ #define PACKET_TAG_IPFORWARD 18 /* ipforward info */ #define PACKET_TAG_MACLABEL (19 | MTAG_PERSISTENT) /* MAC label */ #define PACKET_TAG_PF (21 | MTAG_PERSISTENT) /* PF/ALTQ information */ #define PACKET_TAG_RTSOCKFAM 25 /* rtsock sa family */ #define PACKET_TAG_IPOPTIONS 27 /* Saved IP options */ #define PACKET_TAG_CARP 28 /* CARP info */ #define PACKET_TAG_IPSEC_NAT_T_PORTS 29 /* two uint16_t */ #define PACKET_TAG_ND_OUTGOING 30 /* ND outgoing */ /* Specific cookies and tags. */ /* Packet tag routines. */ struct m_tag *m_tag_alloc(u_int32_t, int, int, int); void m_tag_delete(struct mbuf *, struct m_tag *); void m_tag_delete_chain(struct mbuf *, struct m_tag *); void m_tag_free_default(struct m_tag *); struct m_tag *m_tag_locate(struct mbuf *, u_int32_t, int, struct m_tag *); struct m_tag *m_tag_copy(struct m_tag *, int); int m_tag_copy_chain(struct mbuf *, const struct mbuf *, int); void m_tag_delete_nonpersistent(struct mbuf *); /* * Initialize the list of tags associated with an mbuf. */ static __inline void m_tag_init(struct mbuf *m) { SLIST_INIT(&m->m_pkthdr.tags); } /* * Set up the contents of a tag. Note that this does not fill in the free * method; the caller is expected to do that. * * XXX probably should be called m_tag_init, but that was already taken. */ static __inline void m_tag_setup(struct m_tag *t, u_int32_t cookie, int type, int len) { t->m_tag_id = type; t->m_tag_len = len; t->m_tag_cookie = cookie; } /* * Reclaim resources associated with a tag. */ static __inline void m_tag_free(struct m_tag *t) { (*t->m_tag_free)(t); } /* * Return the first tag associated with an mbuf. */ static __inline struct m_tag * m_tag_first(struct mbuf *m) { return (SLIST_FIRST(&m->m_pkthdr.tags)); } /* * Return the next tag in the list of tags associated with an mbuf. */ static __inline struct m_tag * m_tag_next(struct mbuf *m __unused, struct m_tag *t) { return (SLIST_NEXT(t, m_tag_link)); } /* * Prepend a tag to the list of tags associated with an mbuf. */ static __inline void m_tag_prepend(struct mbuf *m, struct m_tag *t) { SLIST_INSERT_HEAD(&m->m_pkthdr.tags, t, m_tag_link); } /* * Unlink a tag from the list of tags associated with an mbuf. */ static __inline void m_tag_unlink(struct mbuf *m, struct m_tag *t) { SLIST_REMOVE(&m->m_pkthdr.tags, t, m_tag, m_tag_link); } /* These are for OpenBSD compatibility. */ #define MTAG_ABI_COMPAT 0 /* compatibility ABI */ static __inline struct m_tag * m_tag_get(int type, int length, int wait) { return (m_tag_alloc(MTAG_ABI_COMPAT, type, length, wait)); } static __inline struct m_tag * m_tag_find(struct mbuf *m, int type, struct m_tag *start) { return (SLIST_EMPTY(&m->m_pkthdr.tags) ? (struct m_tag *)NULL : m_tag_locate(m, MTAG_ABI_COMPAT, type, start)); } static inline struct m_snd_tag * m_snd_tag_ref(struct m_snd_tag *mst) { refcount_acquire(&mst->refcount); return (mst); } static inline void m_snd_tag_rele(struct m_snd_tag *mst) { if (refcount_release(&mst->refcount)) m_snd_tag_destroy(mst); } static __inline struct mbuf * m_free(struct mbuf *m) { struct mbuf *n = m->m_next; MBUF_PROBE1(m__free, m); if ((m->m_flags & (M_PKTHDR|M_NOFREE)) == (M_PKTHDR|M_NOFREE)) m_tag_delete_chain(m, NULL); if (m->m_flags & M_PKTHDR && m->m_pkthdr.csum_flags & CSUM_SND_TAG) m_snd_tag_rele(m->m_pkthdr.snd_tag); if (m->m_flags & M_EXTPG) mb_free_extpg(m); else if (m->m_flags & M_EXT) mb_free_ext(m); else if ((m->m_flags & M_NOFREE) == 0) uma_zfree(zone_mbuf, m); return (n); } static __inline int rt_m_getfib(struct mbuf *m) { KASSERT(m->m_flags & M_PKTHDR , ("Attempt to get FIB from non header mbuf.")); return (m->m_pkthdr.fibnum); } #define M_GETFIB(_m) rt_m_getfib(_m) #define M_SETFIB(_m, _fib) do { \ KASSERT((_m)->m_flags & M_PKTHDR, ("Attempt to set FIB on non header mbuf.")); \ ((_m)->m_pkthdr.fibnum) = (_fib); \ } while (0) /* flags passed as first argument for "m_xxx_tcpip_hash()" */ #define MBUF_HASHFLAG_L2 (1 << 2) #define MBUF_HASHFLAG_L3 (1 << 3) #define MBUF_HASHFLAG_L4 (1 << 4) /* mbuf hashing helper routines */ uint32_t m_ether_tcpip_hash_init(void); uint32_t m_ether_tcpip_hash(const uint32_t, const struct mbuf *, uint32_t); uint32_t m_infiniband_tcpip_hash_init(void); uint32_t m_infiniband_tcpip_hash(const uint32_t, const struct mbuf *, uint32_t); #ifdef MBUF_PROFILING void m_profile(struct mbuf *m); #define M_PROFILE(m) m_profile(m) #else #define M_PROFILE(m) #endif struct mbufq { STAILQ_HEAD(, mbuf) mq_head; int mq_len; int mq_maxlen; }; static inline void mbufq_init(struct mbufq *mq, int maxlen) { STAILQ_INIT(&mq->mq_head); mq->mq_maxlen = maxlen; mq->mq_len = 0; } static inline struct mbuf * mbufq_flush(struct mbufq *mq) { struct mbuf *m; m = STAILQ_FIRST(&mq->mq_head); STAILQ_INIT(&mq->mq_head); mq->mq_len = 0; return (m); } static inline void mbufq_drain(struct mbufq *mq) { struct mbuf *m, *n; n = mbufq_flush(mq); while ((m = n) != NULL) { n = STAILQ_NEXT(m, m_stailqpkt); m_freem(m); } } static inline struct mbuf * mbufq_first(const struct mbufq *mq) { return (STAILQ_FIRST(&mq->mq_head)); } static inline struct mbuf * mbufq_last(const struct mbufq *mq) { return (STAILQ_LAST(&mq->mq_head, mbuf, m_stailqpkt)); } static inline int mbufq_full(const struct mbufq *mq) { return (mq->mq_maxlen > 0 && mq->mq_len >= mq->mq_maxlen); } static inline int mbufq_len(const struct mbufq *mq) { return (mq->mq_len); } static inline int mbufq_enqueue(struct mbufq *mq, struct mbuf *m) { if (mbufq_full(mq)) return (ENOBUFS); STAILQ_INSERT_TAIL(&mq->mq_head, m, m_stailqpkt); mq->mq_len++; return (0); } static inline struct mbuf * mbufq_dequeue(struct mbufq *mq) { struct mbuf *m; m = STAILQ_FIRST(&mq->mq_head); if (m) { STAILQ_REMOVE_HEAD(&mq->mq_head, m_stailqpkt); m->m_nextpkt = NULL; mq->mq_len--; } return (m); } static inline void mbufq_prepend(struct mbufq *mq, struct mbuf *m) { STAILQ_INSERT_HEAD(&mq->mq_head, m, m_stailqpkt); mq->mq_len++; } /* * Note: this doesn't enforce the maximum list size for dst. */ static inline void mbufq_concat(struct mbufq *mq_dst, struct mbufq *mq_src) { mq_dst->mq_len += mq_src->mq_len; STAILQ_CONCAT(&mq_dst->mq_head, &mq_src->mq_head); mq_src->mq_len = 0; } #ifdef _SYS_TIMESPEC_H_ static inline void mbuf_tstmp2timespec(struct mbuf *m, struct timespec *ts) { KASSERT((m->m_flags & M_PKTHDR) != 0, ("mbuf %p no M_PKTHDR", m)); KASSERT((m->m_flags & (M_TSTMP|M_TSTMP_LRO)) != 0, ("mbuf %p no M_TSTMP or M_TSTMP_LRO", m)); ts->tv_sec = m->m_pkthdr.rcv_tstmp / 1000000000; ts->tv_nsec = m->m_pkthdr.rcv_tstmp % 1000000000; } #endif #ifdef DEBUGNET /* Invoked from the debugnet client code. */ void debugnet_mbuf_drain(void); void debugnet_mbuf_start(void); void debugnet_mbuf_finish(void); void debugnet_mbuf_reinit(int nmbuf, int nclust, int clsize); #endif static inline bool mbuf_has_tls_session(struct mbuf *m) { if (m->m_flags & M_EXTPG) { if (m->m_epg_tls != NULL) { return (true); } } return (false); } #endif /* _KERNEL */ #endif /* !_SYS_MBUF_H_ */ diff --git a/usr.bin/netstat/inet.c b/usr.bin/netstat/inet.c index 95b0d8931b26..49478c4a9247 100644 --- a/usr.bin/netstat/inet.c +++ b/usr.bin/netstat/inet.c @@ -1,1503 +1,1507 @@ /*- * Copyright (c) 1983, 1988, 1993, 1995 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #if 0 #ifndef lint static char sccsid[] = "@(#)inet.c 8.5 (Berkeley) 5/24/95"; #endif /* not lint */ #endif #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #define _WANT_SOCKET #include #include #include #include #include #include #include #include #ifdef INET6 #include #endif /* INET6 */ #include #include #include #include #include #include #include #include #include #define TCPSTATES #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "netstat.h" #include "nl_defs.h" #define max(a, b) (((a) > (b)) ? (a) : (b)) #ifdef INET static void inetprint(const char *, struct in_addr *, int, const char *, int, const int); #endif #ifdef INET6 static int udp_done, tcp_done, sdp_done; #endif /* INET6 */ static int pcblist_sysctl(int proto, const char *name, char **bufp) { const char *mibvar; char *buf; size_t len; switch (proto) { case IPPROTO_TCP: mibvar = "net.inet.tcp.pcblist"; break; case IPPROTO_UDP: mibvar = "net.inet.udp.pcblist"; break; case IPPROTO_DIVERT: mibvar = "net.inet.divert.pcblist"; break; default: mibvar = "net.inet.raw.pcblist"; break; } if (strncmp(name, "sdp", 3) == 0) mibvar = "net.inet.sdp.pcblist"; len = 0; if (sysctlbyname(mibvar, 0, &len, 0, 0) < 0) { if (errno != ENOENT) xo_warn("sysctl: %s", mibvar); return (0); } if ((buf = malloc(len)) == NULL) { xo_warnx("malloc %lu bytes", (u_long)len); return (0); } if (sysctlbyname(mibvar, buf, &len, 0, 0) < 0) { xo_warn("sysctl: %s", mibvar); free(buf); return (0); } *bufp = buf; return (1); } /* * Copied directly from uipc_socket2.c. We leave out some fields that are in * nested structures that aren't used to avoid extra work. */ static void sbtoxsockbuf(struct sockbuf *sb, struct xsockbuf *xsb) { xsb->sb_cc = sb->sb_ccc; xsb->sb_hiwat = sb->sb_hiwat; xsb->sb_mbcnt = sb->sb_mbcnt; xsb->sb_mcnt = sb->sb_mcnt; xsb->sb_ccnt = sb->sb_ccnt; xsb->sb_mbmax = sb->sb_mbmax; xsb->sb_lowat = sb->sb_lowat; xsb->sb_flags = sb->sb_flags; xsb->sb_timeo = sb->sb_timeo; } int sotoxsocket(struct socket *so, struct xsocket *xso) { struct protosw proto; struct domain domain; bzero(xso, sizeof *xso); xso->xso_len = sizeof *xso; xso->xso_so = (uintptr_t)so; xso->so_type = so->so_type; xso->so_options = so->so_options; xso->so_linger = so->so_linger; xso->so_state = so->so_state; xso->so_pcb = (uintptr_t)so->so_pcb; if (kread((uintptr_t)so->so_proto, &proto, sizeof(proto)) != 0) return (-1); xso->xso_protocol = proto.pr_protocol; if (kread((uintptr_t)proto.pr_domain, &domain, sizeof(domain)) != 0) return (-1); xso->xso_family = domain.dom_family; xso->so_timeo = so->so_timeo; xso->so_error = so->so_error; if ((so->so_options & SO_ACCEPTCONN) != 0) { xso->so_qlen = so->sol_qlen; xso->so_incqlen = so->sol_incqlen; xso->so_qlimit = so->sol_qlimit; } else { sbtoxsockbuf(&so->so_snd, &xso->so_snd); sbtoxsockbuf(&so->so_rcv, &xso->so_rcv); xso->so_oobmark = so->so_oobmark; } return (0); } /* * Print a summary of connections related to an Internet * protocol. For TCP, also give state of connection. * Listening processes (aflag) are suppressed unless the * -a (all) flag is specified. */ void protopr(u_long off, const char *name, int af1, int proto) { static int first = 1; int istcp; char *buf; const char *vchar; struct xtcpcb *tp; struct xinpcb *inp; struct xinpgen *xig, *oxig; struct xsocket *so; int fnamelen, cnamelen; istcp = 0; switch (proto) { case IPPROTO_TCP: #ifdef INET6 if (strncmp(name, "sdp", 3) != 0) { if (tcp_done != 0) return; else tcp_done = 1; } else { if (sdp_done != 0) return; else sdp_done = 1; } #endif istcp = 1; break; case IPPROTO_UDP: #ifdef INET6 if (udp_done != 0) return; else udp_done = 1; #endif break; } if (!pcblist_sysctl(proto, name, &buf)) return; if (cflag || Cflag) { fnamelen = strlen("Stack"); cnamelen = strlen("CC"); oxig = xig = (struct xinpgen *)buf; for (xig = (struct xinpgen*)((char *)xig + xig->xig_len); xig->xig_len > sizeof(struct xinpgen); xig = (struct xinpgen *)((char *)xig + xig->xig_len)) { if (istcp) { tp = (struct xtcpcb *)xig; inp = &tp->xt_inp; } else { continue; } if (so->xso_protocol != proto) continue; if (inp->inp_gencnt > oxig->xig_gen) continue; fnamelen = max(fnamelen, (int)strlen(tp->xt_stack)); cnamelen = max(cnamelen, (int)strlen(tp->xt_cc)); } } oxig = xig = (struct xinpgen *)buf; for (xig = (struct xinpgen *)((char *)xig + xig->xig_len); xig->xig_len > sizeof(struct xinpgen); xig = (struct xinpgen *)((char *)xig + xig->xig_len)) { if (istcp) { tp = (struct xtcpcb *)xig; inp = &tp->xt_inp; } else { inp = (struct xinpcb *)xig; } so = &inp->xi_socket; /* Ignore sockets for protocols other than the desired one. */ if (so->xso_protocol != proto) continue; /* Ignore PCBs which were freed during copyout. */ if (inp->inp_gencnt > oxig->xig_gen) continue; if ((af1 == AF_INET && (inp->inp_vflag & INP_IPV4) == 0) #ifdef INET6 || (af1 == AF_INET6 && (inp->inp_vflag & INP_IPV6) == 0) #endif /* INET6 */ || (af1 == AF_UNSPEC && ((inp->inp_vflag & INP_IPV4) == 0 #ifdef INET6 && (inp->inp_vflag & INP_IPV6) == 0 #endif /* INET6 */ )) ) continue; if (!aflag && ( (istcp && tp->t_state == TCPS_LISTEN) || (af1 == AF_INET && inet_lnaof(inp->inp_laddr) == INADDR_ANY) #ifdef INET6 || (af1 == AF_INET6 && IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) #endif /* INET6 */ || (af1 == AF_UNSPEC && (((inp->inp_vflag & INP_IPV4) != 0 && inet_lnaof(inp->inp_laddr) == INADDR_ANY) #ifdef INET6 || ((inp->inp_vflag & INP_IPV6) != 0 && IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) #endif )) )) continue; if (first) { if (!Lflag) { xo_emit("Active Internet connections"); if (aflag) xo_emit(" (including servers)"); } else xo_emit( "Current listen queue sizes (qlen/incqlen/maxqlen)"); xo_emit("\n"); if (Aflag) xo_emit("{T:/%-*s} ", 2 * (int)sizeof(void *), "Tcpcb"); if (Lflag) xo_emit((Aflag && !Wflag) ? "{T:/%-5.5s} {T:/%-32.32s} {T:/%-18.18s}" : ((!Wflag || af1 == AF_INET) ? "{T:/%-5.5s} {T:/%-32.32s} {T:/%-22.22s}" : "{T:/%-5.5s} {T:/%-32.32s} {T:/%-45.45s}"), "Proto", "Listen", "Local Address"); else if (Tflag) xo_emit((Aflag && !Wflag) ? "{T:/%-5.5s} {T:/%-6.6s} {T:/%-6.6s} {T:/%-6.6s} {T:/%-18.18s} {T:/%s}" : ((!Wflag || af1 == AF_INET) ? "{T:/%-5.5s} {T:/%-6.6s} {T:/%-6.6s} {T:/%-6.6s} {T:/%-22.22s} {T:/%s}" : "{T:/%-5.5s} {T:/%-6.6s} {T:/%-6.6s} {T:/%-6.6s} {T:/%-45.45s} {T:/%s}"), "Proto", "Rexmit", "OOORcv", "0-win", "Local Address", "Foreign Address"); else { xo_emit((Aflag && !Wflag) ? "{T:/%-5.5s} {T:/%-6.6s} {T:/%-6.6s} {T:/%-18.18s} {T:/%-18.18s}" : ((!Wflag || af1 == AF_INET) ? "{T:/%-5.5s} {T:/%-6.6s} {T:/%-6.6s} {T:/%-22.22s} {T:/%-22.22s}" : "{T:/%-5.5s} {T:/%-6.6s} {T:/%-6.6s} {T:/%-45.45s} {T:/%-45.45s}"), "Proto", "Recv-Q", "Send-Q", "Local Address", "Foreign Address"); if (!xflag && !Rflag) xo_emit(" {T:/%-11.11s}", "(state)"); } if (xflag) { xo_emit(" {T:/%-6.6s} {T:/%-6.6s} {T:/%-6.6s} " "{T:/%-6.6s} {T:/%-6.6s} {T:/%-6.6s} " "{T:/%-6.6s} {T:/%-6.6s} {T:/%-6.6s} " "{T:/%-6.6s} {T:/%-6.6s} {T:/%-6.6s}", "R-MBUF", "S-MBUF", "R-CLUS", "S-CLUS", "R-HIWA", "S-HIWA", "R-LOWA", "S-LOWA", "R-BCNT", "S-BCNT", "R-BMAX", "S-BMAX"); xo_emit(" {T:/%7.7s} {T:/%7.7s} {T:/%7.7s} " "{T:/%7.7s} {T:/%7.7s} {T:/%7.7s}", "rexmt", "persist", "keep", "2msl", "delack", "rcvtime"); } else if (Rflag) { xo_emit(" {T:/%8.8s} {T:/%5.5s}", "flowid", "ftype"); } if (cflag) { xo_emit(" {T:/%-*.*s}", fnamelen, fnamelen, "Stack"); } if (Cflag) xo_emit(" {T:/%-*.*s} {T:/%10.10s}" " {T:/%10.10s} {T:/%5.5s}" " {T:/%3.3s}", cnamelen, cnamelen, "CC", "cwin", "ssthresh", "MSS", "ECN"); if (Pflag) xo_emit(" {T:/%s}", "Log ID"); xo_emit("\n"); first = 0; } if (Lflag && so->so_qlimit == 0) continue; xo_open_instance("socket"); if (Aflag) { if (istcp) xo_emit("{q:address/%*lx} ", 2 * (int)sizeof(void *), (u_long)inp->inp_ppcb); else xo_emit("{q:address/%*lx} ", 2 * (int)sizeof(void *), (u_long)so->so_pcb); } #ifdef INET6 if ((inp->inp_vflag & INP_IPV6) != 0) vchar = ((inp->inp_vflag & INP_IPV4) != 0) ? "46" : "6"; else #endif vchar = ((inp->inp_vflag & INP_IPV4) != 0) ? "4" : ""; if (istcp && (tp->t_flags & TF_TOE) != 0) xo_emit("{:protocol/%-3.3s%-2.2s/%s%s} ", "toe", vchar); else xo_emit("{:protocol/%-3.3s%-2.2s/%s%s} ", name, vchar); if (Lflag) { char buf1[33]; snprintf(buf1, sizeof buf1, "%u/%u/%u", so->so_qlen, so->so_incqlen, so->so_qlimit); xo_emit("{:listen-queue-sizes/%-32.32s} ", buf1); } else if (Tflag) { if (istcp) xo_emit("{:sent-retransmit-packets/%6u} " "{:received-out-of-order-packets/%6u} " "{:sent-zero-window/%6u} ", tp->t_sndrexmitpack, tp->t_rcvoopack, tp->t_sndzerowin); else xo_emit("{P:/%21s}", ""); } else { xo_emit("{:receive-bytes-waiting/%6u} " "{:send-bytes-waiting/%6u} ", so->so_rcv.sb_cc, so->so_snd.sb_cc); } if (numeric_port) { #ifdef INET if (inp->inp_vflag & INP_IPV4) { inetprint("local", &inp->inp_laddr, (int)inp->inp_lport, name, 1, af1); if (!Lflag) inetprint("remote", &inp->inp_faddr, (int)inp->inp_fport, name, 1, af1); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET6 if (inp->inp_vflag & INP_IPV6) { inet6print("local", &inp->in6p_laddr, (int)inp->inp_lport, name, 1); if (!Lflag) inet6print("remote", &inp->in6p_faddr, (int)inp->inp_fport, name, 1); } /* else nothing printed now */ #endif /* INET6 */ } else if (inp->inp_flags & INP_ANONPORT) { #ifdef INET if (inp->inp_vflag & INP_IPV4) { inetprint("local", &inp->inp_laddr, (int)inp->inp_lport, name, 1, af1); if (!Lflag) inetprint("remote", &inp->inp_faddr, (int)inp->inp_fport, name, 0, af1); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET6 if (inp->inp_vflag & INP_IPV6) { inet6print("local", &inp->in6p_laddr, (int)inp->inp_lport, name, 1); if (!Lflag) inet6print("remote", &inp->in6p_faddr, (int)inp->inp_fport, name, 0); } /* else nothing printed now */ #endif /* INET6 */ } else { #ifdef INET if (inp->inp_vflag & INP_IPV4) { inetprint("local", &inp->inp_laddr, (int)inp->inp_lport, name, 0, af1); if (!Lflag) inetprint("remote", &inp->inp_faddr, (int)inp->inp_fport, name, inp->inp_lport != inp->inp_fport, af1); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET6 if (inp->inp_vflag & INP_IPV6) { inet6print("local", &inp->in6p_laddr, (int)inp->inp_lport, name, 0); if (!Lflag) inet6print("remote", &inp->in6p_faddr, (int)inp->inp_fport, name, inp->inp_lport != inp->inp_fport); } /* else nothing printed now */ #endif /* INET6 */ } if (xflag) { xo_emit("{:receive-mbufs/%6u} {:send-mbufs/%6u} " "{:receive-clusters/%6u} {:send-clusters/%6u} " "{:receive-high-water/%6u} {:send-high-water/%6u} " "{:receive-low-water/%6u} {:send-low-water/%6u} " "{:receive-mbuf-bytes/%6u} {:send-mbuf-bytes/%6u} " "{:receive-mbuf-bytes-max/%6u} " "{:send-mbuf-bytes-max/%6u}", so->so_rcv.sb_mcnt, so->so_snd.sb_mcnt, so->so_rcv.sb_ccnt, so->so_snd.sb_ccnt, so->so_rcv.sb_hiwat, so->so_snd.sb_hiwat, so->so_rcv.sb_lowat, so->so_snd.sb_lowat, so->so_rcv.sb_mbcnt, so->so_snd.sb_mbcnt, so->so_rcv.sb_mbmax, so->so_snd.sb_mbmax); if (istcp) xo_emit(" {:retransmit-timer/%4d.%02d} " "{:persist-timer/%4d.%02d} " "{:keepalive-timer/%4d.%02d} " "{:msl2-timer/%4d.%02d} " "{:delay-ack-timer/%4d.%02d} " "{:inactivity-timer/%4d.%02d}", tp->tt_rexmt / 1000, (tp->tt_rexmt % 1000) / 10, tp->tt_persist / 1000, (tp->tt_persist % 1000) / 10, tp->tt_keep / 1000, (tp->tt_keep % 1000) / 10, tp->tt_2msl / 1000, (tp->tt_2msl % 1000) / 10, tp->tt_delack / 1000, (tp->tt_delack % 1000) / 10, tp->t_rcvtime / 1000, (tp->t_rcvtime % 1000) / 10); } if (istcp && !Lflag && !xflag && !Tflag && !Rflag) { if (tp->t_state < 0 || tp->t_state >= TCP_NSTATES) xo_emit("{:tcp-state/%-11d}", tp->t_state); else { xo_emit("{:tcp-state/%-11s}", tcpstates[tp->t_state]); #if defined(TF_NEEDSYN) && defined(TF_NEEDFIN) /* Show T/TCP `hidden state' */ if (tp->t_flags & (TF_NEEDSYN|TF_NEEDFIN)) xo_emit("{:need-syn-or-fin/*}"); #endif /* defined(TF_NEEDSYN) && defined(TF_NEEDFIN) */ } } if (Rflag) { /* XXX: is this right Alfred */ xo_emit(" {:flow-id/%08x} {:flow-type/%5d}", inp->inp_flowid, inp->inp_flowtype); } if (istcp) { if (cflag) xo_emit(" {:stack/%-*.*s}", fnamelen, fnamelen, tp->xt_stack); if (Cflag) xo_emit(" {:cc/%-*.*s}" " {:snd-cwnd/%10lu}" " {:snd-ssthresh/%10lu}" " {:t-maxseg/%5u} {:ecn/%3s}", cnamelen, cnamelen, tp->xt_cc, tp->t_snd_cwnd, tp->t_snd_ssthresh, tp->t_maxseg, (tp->t_state >= TCPS_ESTABLISHED ? (tp->xt_ecn > 0 ? (tp->xt_ecn == 1 ? "ecn" : "ace") : "off") : "n/a")); if (Pflag) xo_emit(" {:log-id/%s}", tp->xt_logid[0] == '\0' ? "-" : tp->xt_logid); } xo_emit("\n"); xo_close_instance("socket"); } if (xig != oxig && xig->xig_gen != oxig->xig_gen) { if (oxig->xig_count > xig->xig_count) { xo_emit("Some {d:lost/%s} sockets may have been " "deleted.\n", name); } else if (oxig->xig_count < xig->xig_count) { xo_emit("Some {d:created/%s} sockets may have been " "created.\n", name); } else { xo_emit("Some {d:changed/%s} sockets may have been " "created or deleted.\n", name); } } free(buf); } /* * Dump TCP statistics structure. */ void tcp_stats(u_long off, const char *name, int af1 __unused, int proto __unused) { struct tcpstat tcpstat; uint64_t tcps_states[TCP_NSTATES]; #ifdef INET6 if (tcp_done != 0) return; else tcp_done = 1; #endif if (fetch_stats("net.inet.tcp.stats", off, &tcpstat, sizeof(tcpstat), kread_counters) != 0) return; if (fetch_stats_ro("net.inet.tcp.states", nl[N_TCPS_STATES].n_value, &tcps_states, sizeof(tcps_states), kread_counters) != 0) return; xo_open_container("tcp"); xo_emit("{T:/%s}:\n", name); #define p(f, m) if (tcpstat.f || sflag <= 1) \ xo_emit(m, (uintmax_t )tcpstat.f, plural(tcpstat.f)) #define p1a(f, m) if (tcpstat.f || sflag <= 1) \ xo_emit(m, (uintmax_t )tcpstat.f) #define p2(f1, f2, m) if (tcpstat.f1 || tcpstat.f2 || sflag <= 1) \ xo_emit(m, (uintmax_t )tcpstat.f1, plural(tcpstat.f1), \ (uintmax_t )tcpstat.f2, plural(tcpstat.f2)) #define p2a(f1, f2, m) if (tcpstat.f1 || tcpstat.f2 || sflag <= 1) \ xo_emit(m, (uintmax_t )tcpstat.f1, plural(tcpstat.f1), \ (uintmax_t )tcpstat.f2) #define p3(f, m) if (tcpstat.f || sflag <= 1) \ xo_emit(m, (uintmax_t )tcpstat.f, pluralies(tcpstat.f)) p(tcps_sndtotal, "\t{:sent-packets/%ju} {N:/packet%s sent}\n"); p2(tcps_sndpack,tcps_sndbyte, "\t\t{:sent-data-packets/%ju} " "{N:/data packet%s} ({:sent-data-bytes/%ju} {N:/byte%s})\n"); p2(tcps_sndrexmitpack, tcps_sndrexmitbyte, "\t\t" "{:sent-retransmitted-packets/%ju} {N:/data packet%s} " "({:sent-retransmitted-bytes/%ju} {N:/byte%s}) " "{N:retransmitted}\n"); p(tcps_sndrexmitbad, "\t\t" "{:sent-unnecessary-retransmitted-packets/%ju} " "{N:/data packet%s unnecessarily retransmitted}\n"); p(tcps_mturesent, "\t\t{:sent-resends-by-mtu-discovery/%ju} " "{N:/resend%s initiated by MTU discovery}\n"); p2a(tcps_sndacks, tcps_delack, "\t\t{:sent-ack-only-packets/%ju} " "{N:/ack-only packet%s/} ({:sent-packets-delayed/%ju} " "{N:delayed})\n"); p(tcps_sndurg, "\t\t{:sent-urg-only-packets/%ju} " "{N:/URG only packet%s}\n"); p(tcps_sndprobe, "\t\t{:sent-window-probe-packets/%ju} " "{N:/window probe packet%s}\n"); p(tcps_sndwinup, "\t\t{:sent-window-update-packets/%ju} " "{N:/window update packet%s}\n"); p(tcps_sndctrl, "\t\t{:sent-control-packets/%ju} " "{N:/control packet%s}\n"); p(tcps_rcvtotal, "\t{:received-packets/%ju} " "{N:/packet%s received}\n"); p2(tcps_rcvackpack, tcps_rcvackbyte, "\t\t" "{:received-ack-packets/%ju} {N:/ack%s} " "{N:(for} {:received-ack-bytes/%ju} {N:/byte%s})\n"); p(tcps_rcvdupack, "\t\t{:received-duplicate-acks/%ju} " "{N:/duplicate ack%s}\n"); + p(tcps_tunneled_pkts, "\t\t{:received-udp-tunneled-pkts/%ju} " + "{N:/UDP tunneled pkt%s}\n"); + p(tcps_tunneled_errs, "\t\t{:received-bad-udp-tunneled-pkts/%ju} " + "{N:/UDP tunneled pkt cnt with error%s}\n"); p(tcps_rcvacktoomuch, "\t\t{:received-acks-for-unsent-data/%ju} " "{N:/ack%s for unsent data}\n"); p2(tcps_rcvpack, tcps_rcvbyte, "\t\t" "{:received-in-sequence-packets/%ju} {N:/packet%s} " "({:received-in-sequence-bytes/%ju} {N:/byte%s}) " "{N:received in-sequence}\n"); p2(tcps_rcvduppack, tcps_rcvdupbyte, "\t\t" "{:received-completely-duplicate-packets/%ju} " "{N:/completely duplicate packet%s} " "({:received-completely-duplicate-bytes/%ju} {N:/byte%s})\n"); p(tcps_pawsdrop, "\t\t{:received-old-duplicate-packets/%ju} " "{N:/old duplicate packet%s}\n"); p2(tcps_rcvpartduppack, tcps_rcvpartdupbyte, "\t\t" "{:received-some-duplicate-packets/%ju} " "{N:/packet%s with some dup. data} " "({:received-some-duplicate-bytes/%ju} {N:/byte%s duped/})\n"); p2(tcps_rcvoopack, tcps_rcvoobyte, "\t\t{:received-out-of-order/%ju} " "{N:/out-of-order packet%s} " "({:received-out-of-order-bytes/%ju} {N:/byte%s})\n"); p2(tcps_rcvpackafterwin, tcps_rcvbyteafterwin, "\t\t" "{:received-after-window-packets/%ju} {N:/packet%s} " "({:received-after-window-bytes/%ju} {N:/byte%s}) " "{N:of data after window}\n"); p(tcps_rcvwinprobe, "\t\t{:received-window-probes/%ju} " "{N:/window probe%s}\n"); p(tcps_rcvwinupd, "\t\t{:receive-window-update-packets/%ju} " "{N:/window update packet%s}\n"); p(tcps_rcvafterclose, "\t\t{:received-after-close-packets/%ju} " "{N:/packet%s received after close}\n"); p(tcps_rcvbadsum, "\t\t{:discard-bad-checksum/%ju} " "{N:/discarded for bad checksum%s}\n"); p(tcps_rcvbadoff, "\t\t{:discard-bad-header-offset/%ju} " "{N:/discarded for bad header offset field%s}\n"); p1a(tcps_rcvshort, "\t\t{:discard-too-short/%ju} " "{N:discarded because packet too short}\n"); p1a(tcps_rcvreassfull, "\t\t{:discard-reassembly-queue-full/%ju} " "{N:discarded due to full reassembly queue}\n"); p(tcps_connattempt, "\t{:connection-requests/%ju} " "{N:/connection request%s}\n"); p(tcps_accepts, "\t{:connections-accepts/%ju} " "{N:/connection accept%s}\n"); p(tcps_badsyn, "\t{:bad-connection-attempts/%ju} " "{N:/bad connection attempt%s}\n"); p(tcps_listendrop, "\t{:listen-queue-overflows/%ju} " "{N:/listen queue overflow%s}\n"); p(tcps_badrst, "\t{:ignored-in-window-resets/%ju} " "{N:/ignored RSTs in the window%s}\n"); p(tcps_connects, "\t{:connections-established/%ju} " "{N:/connection%s established (including accepts)}\n"); p(tcps_usedrtt, "\t\t{:connections-hostcache-rtt/%ju} " "{N:/time%s used RTT from hostcache}\n"); p(tcps_usedrttvar, "\t\t{:connections-hostcache-rttvar/%ju} " "{N:/time%s used RTT variance from hostcache}\n"); p(tcps_usedssthresh, "\t\t{:connections-hostcache-ssthresh/%ju} " "{N:/time%s used slow-start threshold from hostcache}\n"); p2(tcps_closed, tcps_drops, "\t{:connections-closed/%ju} " "{N:/connection%s closed (including} " "{:connection-drops/%ju} {N:/drop%s})\n"); p(tcps_cachedrtt, "\t\t{:connections-updated-rtt-on-close/%ju} " "{N:/connection%s updated cached RTT on close}\n"); p(tcps_cachedrttvar, "\t\t" "{:connections-updated-variance-on-close/%ju} " "{N:/connection%s updated cached RTT variance on close}\n"); p(tcps_cachedssthresh, "\t\t" "{:connections-updated-ssthresh-on-close/%ju} " "{N:/connection%s updated cached ssthresh on close}\n"); p(tcps_conndrops, "\t{:embryonic-connections-dropped/%ju} " "{N:/embryonic connection%s dropped}\n"); p2(tcps_rttupdated, tcps_segstimed, "\t{:segments-updated-rtt/%ju} " "{N:/segment%s updated rtt (of} " "{:segment-update-attempts/%ju} {N:/attempt%s})\n"); p(tcps_rexmttimeo, "\t{:retransmit-timeouts/%ju} " "{N:/retransmit timeout%s}\n"); p(tcps_timeoutdrop, "\t\t" "{:connections-dropped-by-retransmit-timeout/%ju} " "{N:/connection%s dropped by rexmit timeout}\n"); p(tcps_persisttimeo, "\t{:persist-timeout/%ju} " "{N:/persist timeout%s}\n"); p(tcps_persistdrop, "\t\t" "{:connections-dropped-by-persist-timeout/%ju} " "{N:/connection%s dropped by persist timeout}\n"); p(tcps_finwait2_drops, "\t" "{:connections-dropped-by-finwait2-timeout/%ju} " "{N:/Connection%s (fin_wait_2) dropped because of timeout}\n"); p(tcps_keeptimeo, "\t{:keepalive-timeout/%ju} " "{N:/keepalive timeout%s}\n"); p(tcps_keepprobe, "\t\t{:keepalive-probes/%ju} " "{N:/keepalive probe%s sent}\n"); p(tcps_keepdrops, "\t\t{:connections-dropped-by-keepalives/%ju} " "{N:/connection%s dropped by keepalive}\n"); p(tcps_predack, "\t{:ack-header-predictions/%ju} " "{N:/correct ACK header prediction%s}\n"); p(tcps_preddat, "\t{:data-packet-header-predictions/%ju} " "{N:/correct data packet header prediction%s}\n"); xo_open_container("syncache"); p3(tcps_sc_added, "\t{:entries-added/%ju} " "{N:/syncache entr%s added}\n"); p1a(tcps_sc_retransmitted, "\t\t{:retransmitted/%ju} " "{N:/retransmitted}\n"); p1a(tcps_sc_dupsyn, "\t\t{:duplicates/%ju} {N:/dupsyn}\n"); p1a(tcps_sc_dropped, "\t\t{:dropped/%ju} {N:/dropped}\n"); p1a(tcps_sc_completed, "\t\t{:completed/%ju} {N:/completed}\n"); p1a(tcps_sc_bucketoverflow, "\t\t{:bucket-overflow/%ju} " "{N:/bucket overflow}\n"); p1a(tcps_sc_cacheoverflow, "\t\t{:cache-overflow/%ju} " "{N:/cache overflow}\n"); p1a(tcps_sc_reset, "\t\t{:reset/%ju} {N:/reset}\n"); p1a(tcps_sc_stale, "\t\t{:stale/%ju} {N:/stale}\n"); p1a(tcps_sc_aborted, "\t\t{:aborted/%ju} {N:/aborted}\n"); p1a(tcps_sc_badack, "\t\t{:bad-ack/%ju} {N:/badack}\n"); p1a(tcps_sc_unreach, "\t\t{:unreachable/%ju} {N:/unreach}\n"); p(tcps_sc_zonefail, "\t\t{:zone-failures/%ju} {N:/zone failure%s}\n"); p(tcps_sc_sendcookie, "\t{:sent-cookies/%ju} {N:/cookie%s sent}\n"); p(tcps_sc_recvcookie, "\t{:receivd-cookies/%ju} " "{N:/cookie%s received}\n"); xo_close_container("syncache"); xo_open_container("hostcache"); p3(tcps_hc_added, "\t{:entries-added/%ju} " "{N:/hostcache entr%s added}\n"); p1a(tcps_hc_bucketoverflow, "\t\t{:buffer-overflows/%ju} " "{N:/bucket overflow}\n"); xo_close_container("hostcache"); xo_open_container("sack"); p(tcps_sack_recovery_episode, "\t{:recovery-episodes/%ju} " "{N:/SACK recovery episode%s}\n"); p(tcps_sack_rexmits, "\t{:segment-retransmits/%ju} " "{N:/segment rexmit%s in SACK recovery episodes}\n"); p(tcps_sack_rexmit_bytes, "\t{:byte-retransmits/%ju} " "{N:/byte rexmit%s in SACK recovery episodes}\n"); p(tcps_sack_rcv_blocks, "\t{:received-blocks/%ju} " "{N:/SACK option%s (SACK blocks) received}\n"); p(tcps_sack_send_blocks, "\t{:sent-option-blocks/%ju} " "{N:/SACK option%s (SACK blocks) sent}\n"); p1a(tcps_sack_sboverflow, "\t{:scoreboard-overflows/%ju} " "{N:/SACK scoreboard overflow}\n"); xo_close_container("sack"); xo_open_container("ecn"); p(tcps_ecn_ce, "\t{:ce-packets/%ju} " "{N:/packet%s with ECN CE bit set}\n"); p(tcps_ecn_ect0, "\t{:ect0-packets/%ju} " "{N:/packet%s with ECN ECT(0) bit set}\n"); p(tcps_ecn_ect1, "\t{:ect1-packets/%ju} " "{N:/packet%s with ECN ECT(1) bit set}\n"); p(tcps_ecn_shs, "\t{:handshakes/%ju} " "{N:/successful ECN handshake%s}\n"); p(tcps_ecn_rcwnd, "\t{:congestion-reductions/%ju} " "{N:/time%s ECN reduced the congestion window}\n"); xo_close_container("ecn"); xo_open_container("tcp-signature"); p(tcps_sig_rcvgoodsig, "\t{:received-good-signature/%ju} " "{N:/packet%s with matching signature received}\n"); p(tcps_sig_rcvbadsig, "\t{:received-bad-signature/%ju} " "{N:/packet%s with bad signature received}\n"); p(tcps_sig_err_buildsig, "\t{:failed-make-signature/%ju} " "{N:/time%s failed to make signature due to no SA}\n"); p(tcps_sig_err_sigopt, "\t{:no-signature-expected/%ju} " "{N:/time%s unexpected signature received}\n"); p(tcps_sig_err_nosigopt, "\t{:no-signature-provided/%ju} " "{N:/time%s no signature provided by segment}\n"); xo_close_container("tcp-signature"); xo_open_container("pmtud"); p(tcps_pmtud_blackhole_activated, "\t{:pmtud-activated/%ju} " "{N:/Path MTU discovery black hole detection activation%s}\n"); p(tcps_pmtud_blackhole_activated_min_mss, "\t{:pmtud-activated-min-mss/%ju} " "{N:/Path MTU discovery black hole detection min MSS activation%s}\n"); p(tcps_pmtud_blackhole_failed, "\t{:pmtud-failed/%ju} " "{N:/Path MTU discovery black hole detection failure%s}\n"); #undef p #undef p1a #undef p2 #undef p2a #undef p3 xo_close_container("pmtud"); xo_open_container("TCP connection count by state"); xo_emit("{T:/TCP connection count by state}:\n"); for (int i = 0; i < TCP_NSTATES; i++) { /* * XXXGL: is there a way in libxo to use %s * in the "content string" of a format * string? I failed to do that, that's why * a temporary buffer is used to construct * format string for xo_emit(). */ char fmtbuf[80]; if (sflag > 1 && tcps_states[i] == 0) continue; snprintf(fmtbuf, sizeof(fmtbuf), "\t{:%s/%%ju} " "{Np:/connection ,connections} in %s state\n", tcpstates[i], tcpstates[i]); xo_emit(fmtbuf, (uintmax_t )tcps_states[i]); } xo_close_container("TCP connection count by state"); xo_close_container("tcp"); } /* * Dump UDP statistics structure. */ void udp_stats(u_long off, const char *name, int af1 __unused, int proto __unused) { struct udpstat udpstat; uint64_t delivered; #ifdef INET6 if (udp_done != 0) return; else udp_done = 1; #endif if (fetch_stats("net.inet.udp.stats", off, &udpstat, sizeof(udpstat), kread_counters) != 0) return; xo_open_container("udp"); xo_emit("{T:/%s}:\n", name); #define p(f, m) if (udpstat.f || sflag <= 1) \ xo_emit("\t" m, (uintmax_t)udpstat.f, plural(udpstat.f)) #define p1a(f, m) if (udpstat.f || sflag <= 1) \ xo_emit("\t" m, (uintmax_t)udpstat.f) p(udps_ipackets, "{:received-datagrams/%ju} " "{N:/datagram%s received}\n"); p1a(udps_hdrops, "{:dropped-incomplete-headers/%ju} " "{N:/with incomplete header}\n"); p1a(udps_badlen, "{:dropped-bad-data-length/%ju} " "{N:/with bad data length field}\n"); p1a(udps_badsum, "{:dropped-bad-checksum/%ju} " "{N:/with bad checksum}\n"); p1a(udps_nosum, "{:dropped-no-checksum/%ju} " "{N:/with no checksum}\n"); p1a(udps_noport, "{:dropped-no-socket/%ju} " "{N:/dropped due to no socket}\n"); p(udps_noportbcast, "{:dropped-broadcast-multicast/%ju} " "{N:/broadcast\\/multicast datagram%s undelivered}\n"); p1a(udps_fullsock, "{:dropped-full-socket-buffer/%ju} " "{N:/dropped due to full socket buffers}\n"); p1a(udpps_pcbhashmiss, "{:not-for-hashed-pcb/%ju} " "{N:/not for hashed pcb}\n"); delivered = udpstat.udps_ipackets - udpstat.udps_hdrops - udpstat.udps_badlen - udpstat.udps_badsum - udpstat.udps_noport - udpstat.udps_noportbcast - udpstat.udps_fullsock; if (delivered || sflag <= 1) xo_emit("\t{:delivered-packets/%ju} {N:/delivered}\n", (uint64_t)delivered); p(udps_opackets, "{:output-packets/%ju} {N:/datagram%s output}\n"); /* the next statistic is cumulative in udps_noportbcast */ p(udps_filtermcast, "{:multicast-source-filter-matches/%ju} " "{N:/time%s multicast source filter matched}\n"); #undef p #undef p1a xo_close_container("udp"); } /* * Dump CARP statistics structure. */ void carp_stats(u_long off, const char *name, int af1 __unused, int proto __unused) { struct carpstats carpstat; if (fetch_stats("net.inet.carp.stats", off, &carpstat, sizeof(carpstat), kread_counters) != 0) return; xo_open_container(name); xo_emit("{T:/%s}:\n", name); #define p(f, m) if (carpstat.f || sflag <= 1) \ xo_emit(m, (uintmax_t)carpstat.f, plural(carpstat.f)) #define p2(f, m) if (carpstat.f || sflag <= 1) \ xo_emit(m, (uintmax_t)carpstat.f) p(carps_ipackets, "\t{:received-inet-packets/%ju} " "{N:/packet%s received (IPv4)}\n"); p(carps_ipackets6, "\t{:received-inet6-packets/%ju} " "{N:/packet%s received (IPv6)}\n"); p(carps_badttl, "\t\t{:dropped-wrong-ttl/%ju} " "{N:/packet%s discarded for wrong TTL}\n"); p(carps_hdrops, "\t\t{:dropped-short-header/%ju} " "{N:/packet%s shorter than header}\n"); p(carps_badsum, "\t\t{:dropped-bad-checksum/%ju} " "{N:/discarded for bad checksum%s}\n"); p(carps_badver, "\t\t{:dropped-bad-version/%ju} " "{N:/discarded packet%s with a bad version}\n"); p2(carps_badlen, "\t\t{:dropped-short-packet/%ju} " "{N:/discarded because packet too short}\n"); p2(carps_badauth, "\t\t{:dropped-bad-authentication/%ju} " "{N:/discarded for bad authentication}\n"); p2(carps_badvhid, "\t\t{:dropped-bad-vhid/%ju} " "{N:/discarded for bad vhid}\n"); p2(carps_badaddrs, "\t\t{:dropped-bad-address-list/%ju} " "{N:/discarded because of a bad address list}\n"); p(carps_opackets, "\t{:sent-inet-packets/%ju} " "{N:/packet%s sent (IPv4)}\n"); p(carps_opackets6, "\t{:sent-inet6-packets/%ju} " "{N:/packet%s sent (IPv6)}\n"); p2(carps_onomem, "\t\t{:send-failed-memory-error/%ju} " "{N:/send failed due to mbuf memory error}\n"); #if notyet p(carps_ostates, "\t\t{:send-state-updates/%s} " "{N:/state update%s sent}\n"); #endif #undef p #undef p2 xo_close_container(name); } /* * Dump IP statistics structure. */ void ip_stats(u_long off, const char *name, int af1 __unused, int proto __unused) { struct ipstat ipstat; if (fetch_stats("net.inet.ip.stats", off, &ipstat, sizeof(ipstat), kread_counters) != 0) return; xo_open_container(name); xo_emit("{T:/%s}:\n", name); #define p(f, m) if (ipstat.f || sflag <= 1) \ xo_emit(m, (uintmax_t )ipstat.f, plural(ipstat.f)) #define p1a(f, m) if (ipstat.f || sflag <= 1) \ xo_emit(m, (uintmax_t )ipstat.f) p(ips_total, "\t{:received-packets/%ju} " "{N:/total packet%s received}\n"); p(ips_badsum, "\t{:dropped-bad-checksum/%ju} " "{N:/bad header checksum%s}\n"); p1a(ips_toosmall, "\t{:dropped-below-minimum-size/%ju} " "{N:/with size smaller than minimum}\n"); p1a(ips_tooshort, "\t{:dropped-short-packets/%ju} " "{N:/with data size < data length}\n"); p1a(ips_toolong, "\t{:dropped-too-long/%ju} " "{N:/with ip length > max ip packet size}\n"); p1a(ips_badhlen, "\t{:dropped-short-header-length/%ju} " "{N:/with header length < data size}\n"); p1a(ips_badlen, "\t{:dropped-short-data/%ju} " "{N:/with data length < header length}\n"); p1a(ips_badoptions, "\t{:dropped-bad-options/%ju} " "{N:/with bad options}\n"); p1a(ips_badvers, "\t{:dropped-bad-version/%ju} " "{N:/with incorrect version number}\n"); p(ips_fragments, "\t{:received-fragments/%ju} " "{N:/fragment%s received}\n"); p(ips_fragdropped, "\t{:dropped-fragments/%ju} " "{N:/fragment%s dropped (dup or out of space)}\n"); p(ips_fragtimeout, "\t{:dropped-fragments-after-timeout/%ju} " "{N:/fragment%s dropped after timeout}\n"); p(ips_reassembled, "\t{:reassembled-packets/%ju} " "{N:/packet%s reassembled ok}\n"); p(ips_delivered, "\t{:received-local-packets/%ju} " "{N:/packet%s for this host}\n"); p(ips_noproto, "\t{:dropped-unknown-protocol/%ju} " "{N:/packet%s for unknown\\/unsupported protocol}\n"); p(ips_forward, "\t{:forwarded-packets/%ju} " "{N:/packet%s forwarded}"); p(ips_fastforward, " ({:fast-forwarded-packets/%ju} " "{N:/packet%s fast forwarded})"); if (ipstat.ips_forward || sflag <= 1) xo_emit("\n"); p(ips_cantforward, "\t{:packets-cannot-forward/%ju} " "{N:/packet%s not forwardable}\n"); p(ips_notmember, "\t{:received-unknown-multicast-group/%ju} " "{N:/packet%s received for unknown multicast group}\n"); p(ips_redirectsent, "\t{:redirects-sent/%ju} " "{N:/redirect%s sent}\n"); p(ips_localout, "\t{:sent-packets/%ju} " "{N:/packet%s sent from this host}\n"); p(ips_rawout, "\t{:send-packets-fabricated-header/%ju} " "{N:/packet%s sent with fabricated ip header}\n"); p(ips_odropped, "\t{:discard-no-mbufs/%ju} " "{N:/output packet%s dropped due to no bufs, etc.}\n"); p(ips_noroute, "\t{:discard-no-route/%ju} " "{N:/output packet%s discarded due to no route}\n"); p(ips_fragmented, "\t{:sent-fragments/%ju} " "{N:/output datagram%s fragmented}\n"); p(ips_ofragments, "\t{:fragments-created/%ju} " "{N:/fragment%s created}\n"); p(ips_cantfrag, "\t{:discard-cannot-fragment/%ju} " "{N:/datagram%s that can't be fragmented}\n"); p(ips_nogif, "\t{:discard-tunnel-no-gif/%ju} " "{N:/tunneling packet%s that can't find gif}\n"); p(ips_badaddr, "\t{:discard-bad-address/%ju} " "{N:/datagram%s with bad address in header}\n"); #undef p #undef p1a xo_close_container(name); } /* * Dump ARP statistics structure. */ void arp_stats(u_long off, const char *name, int af1 __unused, int proto __unused) { struct arpstat arpstat; if (fetch_stats("net.link.ether.arp.stats", off, &arpstat, sizeof(arpstat), kread_counters) != 0) return; xo_open_container(name); xo_emit("{T:/%s}:\n", name); #define p(f, m) if (arpstat.f || sflag <= 1) \ xo_emit("\t" m, (uintmax_t)arpstat.f, plural(arpstat.f)) #define p2(f, m) if (arpstat.f || sflag <= 1) \ xo_emit("\t" m, (uintmax_t)arpstat.f, pluralies(arpstat.f)) p(txrequests, "{:sent-requests/%ju} {N:/ARP request%s sent}\n"); p(txerrors, "{:sent-failures/%ju} {N:/ARP request%s failed to sent}\n"); p2(txreplies, "{:sent-replies/%ju} {N:/ARP repl%s sent}\n"); p(rxrequests, "{:received-requests/%ju} " "{N:/ARP request%s received}\n"); p2(rxreplies, "{:received-replies/%ju} " "{N:/ARP repl%s received}\n"); p(received, "{:received-packers/%ju} " "{N:/ARP packet%s received}\n"); p(dropped, "{:dropped-no-entry/%ju} " "{N:/total packet%s dropped due to no ARP entry}\n"); p(timeouts, "{:entries-timeout/%ju} " "{N:/ARP entry%s timed out}\n"); p(dupips, "{:dropped-duplicate-address/%ju} " "{N:/Duplicate IP%s seen}\n"); #undef p #undef p2 xo_close_container(name); } static const char *icmpnames[ICMP_MAXTYPE + 1] = { "echo reply", /* RFC 792 */ "#1", "#2", "destination unreachable", /* RFC 792 */ "source quench", /* RFC 792 */ "routing redirect", /* RFC 792 */ "#6", "#7", "echo", /* RFC 792 */ "router advertisement", /* RFC 1256 */ "router solicitation", /* RFC 1256 */ "time exceeded", /* RFC 792 */ "parameter problem", /* RFC 792 */ "time stamp", /* RFC 792 */ "time stamp reply", /* RFC 792 */ "information request", /* RFC 792 */ "information request reply", /* RFC 792 */ "address mask request", /* RFC 950 */ "address mask reply", /* RFC 950 */ "#19", "#20", "#21", "#22", "#23", "#24", "#25", "#26", "#27", "#28", "#29", "icmp traceroute", /* RFC 1393 */ "datagram conversion error", /* RFC 1475 */ "mobile host redirect", "IPv6 where-are-you", "IPv6 i-am-here", "mobile registration req", "mobile registration reply", "domain name request", /* RFC 1788 */ "domain name reply", /* RFC 1788 */ "icmp SKIP", "icmp photuris", /* RFC 2521 */ }; /* * Dump ICMP statistics. */ void icmp_stats(u_long off, const char *name, int af1 __unused, int proto __unused) { struct icmpstat icmpstat; size_t len; int i, first; if (fetch_stats("net.inet.icmp.stats", off, &icmpstat, sizeof(icmpstat), kread_counters) != 0) return; xo_open_container(name); xo_emit("{T:/%s}:\n", name); #define p(f, m) if (icmpstat.f || sflag <= 1) \ xo_emit(m, icmpstat.f, plural(icmpstat.f)) #define p1a(f, m) if (icmpstat.f || sflag <= 1) \ xo_emit(m, icmpstat.f) #define p2(f, m) if (icmpstat.f || sflag <= 1) \ xo_emit(m, icmpstat.f, plurales(icmpstat.f)) p(icps_error, "\t{:icmp-calls/%lu} " "{N:/call%s to icmp_error}\n"); p(icps_oldicmp, "\t{:errors-not-from-message/%lu} " "{N:/error%s not generated in response to an icmp message}\n"); for (first = 1, i = 0; i < ICMP_MAXTYPE + 1; i++) { if (icmpstat.icps_outhist[i] != 0) { if (first) { xo_open_list("output-histogram"); xo_emit("\tOutput histogram:\n"); first = 0; } xo_open_instance("output-histogram"); if (icmpnames[i] != NULL) xo_emit("\t\t{k:name/%s}: {:count/%lu}\n", icmpnames[i], icmpstat.icps_outhist[i]); else xo_emit("\t\tunknown ICMP #{k:name/%d}: " "{:count/%lu}\n", i, icmpstat.icps_outhist[i]); xo_close_instance("output-histogram"); } } if (!first) xo_close_list("output-histogram"); p(icps_badcode, "\t{:dropped-bad-code/%lu} " "{N:/message%s with bad code fields}\n"); p(icps_tooshort, "\t{:dropped-too-short/%lu} " "{N:/message%s less than the minimum length}\n"); p(icps_checksum, "\t{:dropped-bad-checksum/%lu} " "{N:/message%s with bad checksum}\n"); p(icps_badlen, "\t{:dropped-bad-length/%lu} " "{N:/message%s with bad length}\n"); p1a(icps_bmcastecho, "\t{:dropped-multicast-echo/%lu} " "{N:/multicast echo requests ignored}\n"); p1a(icps_bmcasttstamp, "\t{:dropped-multicast-timestamp/%lu} " "{N:/multicast timestamp requests ignored}\n"); for (first = 1, i = 0; i < ICMP_MAXTYPE + 1; i++) { if (icmpstat.icps_inhist[i] != 0) { if (first) { xo_open_list("input-histogram"); xo_emit("\tInput histogram:\n"); first = 0; } xo_open_instance("input-histogram"); if (icmpnames[i] != NULL) xo_emit("\t\t{k:name/%s}: {:count/%lu}\n", icmpnames[i], icmpstat.icps_inhist[i]); else xo_emit( "\t\tunknown ICMP #{k:name/%d}: {:count/%lu}\n", i, icmpstat.icps_inhist[i]); xo_close_instance("input-histogram"); } } if (!first) xo_close_list("input-histogram"); p(icps_reflect, "\t{:sent-packets/%lu} " "{N:/message response%s generated}\n"); p2(icps_badaddr, "\t{:discard-invalid-return-address/%lu} " "{N:/invalid return address%s}\n"); p(icps_noroute, "\t{:discard-no-route/%lu} " "{N:/no return route%s}\n"); #undef p #undef p1a #undef p2 if (live) { len = sizeof i; if (sysctlbyname("net.inet.icmp.maskrepl", &i, &len, NULL, 0) < 0) return; xo_emit("\tICMP address mask responses are " "{q:icmp-address-responses/%sabled}\n", i ? "en" : "dis"); } xo_close_container(name); } /* * Dump IGMP statistics structure. */ void igmp_stats(u_long off, const char *name, int af1 __unused, int proto __unused) { struct igmpstat igmpstat; int error, zflag0; if (fetch_stats("net.inet.igmp.stats", 0, &igmpstat, sizeof(igmpstat), kread) != 0) return; /* * Reread net.inet.igmp.stats when zflag == 1. * This is because this MIB contains version number and * length of the structure which are not set when clearing * the counters. */ zflag0 = zflag; if (zflag) { zflag = 0; error = fetch_stats("net.inet.igmp.stats", 0, &igmpstat, sizeof(igmpstat), kread); zflag = zflag0; if (error) return; } if (igmpstat.igps_version != IGPS_VERSION_3) { xo_warnx("%s: version mismatch (%d != %d)", __func__, igmpstat.igps_version, IGPS_VERSION_3); return; } if (igmpstat.igps_len != IGPS_VERSION3_LEN) { xo_warnx("%s: size mismatch (%d != %d)", __func__, igmpstat.igps_len, IGPS_VERSION3_LEN); return; } xo_open_container(name); xo_emit("{T:/%s}:\n", name); #define p64(f, m) if (igmpstat.f || sflag <= 1) \ xo_emit(m, (uintmax_t) igmpstat.f, plural(igmpstat.f)) #define py64(f, m) if (igmpstat.f || sflag <= 1) \ xo_emit(m, (uintmax_t) igmpstat.f, pluralies(igmpstat.f)) p64(igps_rcv_total, "\t{:received-messages/%ju} " "{N:/message%s received}\n"); p64(igps_rcv_tooshort, "\t{:dropped-too-short/%ju} " "{N:/message%s received with too few bytes}\n"); p64(igps_rcv_badttl, "\t{:dropped-wrong-ttl/%ju} " "{N:/message%s received with wrong TTL}\n"); p64(igps_rcv_badsum, "\t{:dropped-bad-checksum/%ju} " "{N:/message%s received with bad checksum}\n"); py64(igps_rcv_v1v2_queries, "\t{:received-membership-queries/%ju} " "{N:/V1\\/V2 membership quer%s received}\n"); py64(igps_rcv_v3_queries, "\t{:received-v3-membership-queries/%ju} " "{N:/V3 membership quer%s received}\n"); py64(igps_rcv_badqueries, "\t{:dropped-membership-queries/%ju} " "{N:/membership quer%s received with invalid field(s)}\n"); py64(igps_rcv_gen_queries, "\t{:received-general-queries/%ju} " "{N:/general quer%s received}\n"); py64(igps_rcv_group_queries, "\t{:received-group-queries/%ju} " "{N:/group quer%s received}\n"); py64(igps_rcv_gsr_queries, "\t{:received-group-source-queries/%ju} " "{N:/group-source quer%s received}\n"); py64(igps_drop_gsr_queries, "\t{:dropped-group-source-queries/%ju} " "{N:/group-source quer%s dropped}\n"); p64(igps_rcv_reports, "\t{:received-membership-requests/%ju} " "{N:/membership report%s received}\n"); p64(igps_rcv_badreports, "\t{:dropped-membership-reports/%ju} " "{N:/membership report%s received with invalid field(s)}\n"); p64(igps_rcv_ourreports, "\t" "{:received-membership-reports-matching/%ju} " "{N:/membership report%s received for groups to which we belong}" "\n"); p64(igps_rcv_nora, "\t{:received-v3-reports-no-router-alert/%ju} " "{N:/V3 report%s received without Router Alert}\n"); p64(igps_snd_reports, "\t{:sent-membership-reports/%ju} " "{N:/membership report%s sent}\n"); #undef p64 #undef py64 xo_close_container(name); } /* * Dump PIM statistics structure. */ void pim_stats(u_long off __unused, const char *name, int af1 __unused, int proto __unused) { struct pimstat pimstat; if (fetch_stats("net.inet.pim.stats", off, &pimstat, sizeof(pimstat), kread_counters) != 0) return; xo_open_container(name); xo_emit("{T:/%s}:\n", name); #define p(f, m) if (pimstat.f || sflag <= 1) \ xo_emit(m, (uintmax_t)pimstat.f, plural(pimstat.f)) #define py(f, m) if (pimstat.f || sflag <= 1) \ xo_emit(m, (uintmax_t)pimstat.f, pimstat.f != 1 ? "ies" : "y") p(pims_rcv_total_msgs, "\t{:received-messages/%ju} " "{N:/message%s received}\n"); p(pims_rcv_total_bytes, "\t{:received-bytes/%ju} " "{N:/byte%s received}\n"); p(pims_rcv_tooshort, "\t{:dropped-too-short/%ju} " "{N:/message%s received with too few bytes}\n"); p(pims_rcv_badsum, "\t{:dropped-bad-checksum/%ju} " "{N:/message%s received with bad checksum}\n"); p(pims_rcv_badversion, "\t{:dropped-bad-version/%ju} " "{N:/message%s received with bad version}\n"); p(pims_rcv_registers_msgs, "\t{:received-data-register-messages/%ju} " "{N:/data register message%s received}\n"); p(pims_rcv_registers_bytes, "\t{:received-data-register-bytes/%ju} " "{N:/data register byte%s received}\n"); p(pims_rcv_registers_wrongiif, "\t" "{:received-data-register-wrong-interface/%ju} " "{N:/data register message%s received on wrong iif}\n"); p(pims_rcv_badregisters, "\t{:received-bad-registers/%ju} " "{N:/bad register%s received}\n"); p(pims_snd_registers_msgs, "\t{:sent-data-register-messages/%ju} " "{N:/data register message%s sent}\n"); p(pims_snd_registers_bytes, "\t{:sent-data-register-bytes/%ju} " "{N:/data register byte%s sent}\n"); #undef p #undef py xo_close_container(name); } #ifdef INET /* * Pretty print an Internet address (net address + port). */ static void inetprint(const char *container, struct in_addr *in, int port, const char *proto, int num_port, const int af1) { struct servent *sp = 0; char line[80], *cp; int width; size_t alen, plen; if (container) xo_open_container(container); if (Wflag) snprintf(line, sizeof(line), "%s.", inetname(in)); else snprintf(line, sizeof(line), "%.*s.", (Aflag && !num_port) ? 12 : 16, inetname(in)); alen = strlen(line); cp = line + alen; if (!num_port && port) sp = getservbyport((int)port, proto); if (sp || port == 0) snprintf(cp, sizeof(line) - alen, "%.15s ", sp ? sp->s_name : "*"); else snprintf(cp, sizeof(line) - alen, "%d ", ntohs((u_short)port)); width = (Aflag && !Wflag) ? 18 : ((!Wflag || af1 == AF_INET) ? 22 : 45); if (Wflag) xo_emit("{d:target/%-*s} ", width, line); else xo_emit("{d:target/%-*.*s} ", width, width, line); plen = strlen(cp) - 1; alen--; xo_emit("{e:address/%*.*s}{e:port/%*.*s}", alen, alen, line, plen, plen, cp); if (container) xo_close_container(container); } /* * Construct an Internet address representation. * If numeric_addr has been supplied, give * numeric value, otherwise try for symbolic name. */ char * inetname(struct in_addr *inp) { char *cp; static char line[MAXHOSTNAMELEN]; struct hostent *hp; struct netent *np; cp = 0; if (!numeric_addr && inp->s_addr != INADDR_ANY) { int net = inet_netof(*inp); int lna = inet_lnaof(*inp); if (lna == INADDR_ANY) { np = getnetbyaddr(net, AF_INET); if (np) cp = np->n_name; } if (cp == NULL) { hp = gethostbyaddr((char *)inp, sizeof (*inp), AF_INET); if (hp) { cp = hp->h_name; trimdomain(cp, strlen(cp)); } } } if (inp->s_addr == INADDR_ANY) strcpy(line, "*"); else if (cp) { strlcpy(line, cp, sizeof(line)); } else { inp->s_addr = ntohl(inp->s_addr); #define C(x) ((u_int)((x) & 0xff)) snprintf(line, sizeof(line), "%u.%u.%u.%u", C(inp->s_addr >> 24), C(inp->s_addr >> 16), C(inp->s_addr >> 8), C(inp->s_addr)); } return (line); } #endif diff --git a/usr.bin/sockstat/sockstat.1 b/usr.bin/sockstat/sockstat.1 index 8521c50348c9..f602ad467f9f 100644 --- a/usr.bin/sockstat/sockstat.1 +++ b/usr.bin/sockstat/sockstat.1 @@ -1,222 +1,222 @@ .\"- .\" Copyright (c) 1999 Dag-Erling Coïdan Smørgrav .\" All rights reserved. .\" .\" Redistribution and use in source and binary forms, with or without .\" modification, are permitted provided that the following conditions .\" are met: .\" 1. Redistributions of source code must retain the above copyright .\" notice, this list of conditions and the following disclaimer .\" in this position and unchanged. .\" 2. Redistributions in binary form must reproduce the above copyright .\" notice, this list of conditions and the following disclaimer in the .\" documentation and/or other materials provided with the distribution. .\" 3. The name of the author may not be used to endorse or promote products .\" derived from this software without specific prior written permission. .\" .\" THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR .\" IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES .\" OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. .\" IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, .\" INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT .\" NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, .\" DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY .\" THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT .\" (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF .\" THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. .\" .\" $FreeBSD$ .\" -.Dd December 30, 2020 +.Dd March 28, 2021 .Dt SOCKSTAT 1 .Os .Sh NAME .Nm sockstat .Nd list open sockets .Sh SYNOPSIS .Nm .Op Fl 46CcLlnSsUuvw .Op Fl j Ar jid .Op Fl p Ar ports .Op Fl P Ar protocols .Sh DESCRIPTION The .Nm command lists open Internet or .Ux domain sockets. .Pp The following options are available: .Bl -tag -width Fl .It Fl 4 Show .Dv AF_INET (IPv4) sockets. .It Fl 6 Show .Dv AF_INET6 (IPv6) sockets. .It Fl C Display the congestion control module, if applicable. This is currently only implemented for TCP. .It Fl c Show connected sockets. .It Fl j Ar jail Show only sockets belonging to the specified jail ID or name. .It Fl L Only show Internet sockets if the local and foreign addresses are not in the loopback network prefix .Li 127.0.0.0/8 , or do not contain the IPv6 loopback address .Li ::1 . .It Fl l Show listening sockets. .It Fl n Do not resolve numeric UIDs to user names. .It Fl p Ar ports Only show Internet sockets if the local or foreign port number is on the specified list. The .Ar ports argument is a comma-separated list of port numbers and ranges specified as first and last port separated by a dash. .It Fl P Ar protocols Only show sockets of the specified .Ar protocols . The .Ar protocols argument is a comma-separated list of protocol names, as they are defined in .Xr protocols 5 . .It Fl q Quiet mode, do not print the header line. .It Fl S Display the protocol stack, if applicable. This is currently only implemented for TCP. .It Fl s Display the protocol state, if applicable. This is currently only implemented for SCTP and TCP. .It Fl U Display the remote UDP encapsulation port number, if applicable. -This is currently only implemented for SCTP. +This is currently only implemented for SCTP and TCP. .It Fl u Show .Dv AF_LOCAL .Pq Ux sockets. .It Fl v Verbose mode. .It Fl w Use wider field size for displaying addresses. .El .Pp If neither .Fl 4 , 6 or .Fl u is specified, .Nm will list sockets in all three domains. .Pp If neither .Fl c or .Fl l is specified, .Nm will list both listening and connected sockets. .Pp The information listed for each socket is: .Bl -tag -width "FOREIGN ADDRESS" .It Li USER The user who owns the socket. .It Li COMMAND The command which holds the socket. .It Li PID The process ID of the command which holds the socket. .It Li FD The file descriptor number of the socket. .It Li PROTO The transport protocol associated with the socket for Internet sockets, or the type of socket .Pq stream, datagram, or seqpacket for .Ux sockets. .It Li LOCAL ADDRESS For Internet sockets, this is the address the local end of the socket is bound to (see .Xr getsockname 2 ) . For bound .Ux sockets, it is the socket's filename. For other .Ux sockets, it is a right arrow followed by the endpoint's filename, or .Dq Li ?? if the endpoint could not be determined. .It Li FOREIGN ADDRESS (Internet sockets only) The address the foreign end of the socket is bound to (see .Xr getpeername 2 ) . .It Li ENCAPS The remote UDP encapsulation port number if .Fl U -is specified (only for SCTP). +is specified (only for SCTP or TCP). .It Li PATH STATE The path state if .Fl s is specified (only for SCTP). .It Li CONN STATE The connection state if .Fl s is specified (only for SCTP or TCP). .It Li STACK The protocol stack if .Fl S is specified (only for TCP). .It Li CC The congestion control if .Fl C is specified (only for TCP). .El .Pp If a socket is associated with more than one file descriptor, it is shown multiple times. If a socket is not associated with any file descriptor, the first four columns have no meaning. .Sh EXAMPLES Show information for IPv4 sockets listening on port 22 using protocol TCP: .Bd -literal -offset indent $ sockstat -4 -l -P tcp -p 22 .Ed .Pp Show information for sockets using either TCP or UDP, if neither, the local nor the foreign addresses are in the loopback network: .Bd -literal -offset indent $ sockstat -L -P tcp,udp .Ed .Pp Show TCP IPv6 sockets which are listening and connected (default): .Bd -literal -offset indent $ sockstat -6 -P tcp .Ed .Sh SEE ALSO .Xr fstat 1 , .Xr netstat 1 , .Xr procstat 1 , .Xr inet 4 , .Xr inet6 4 , .Xr protocols 5 .Sh HISTORY The .Nm command appeared in .Fx 3.1 . .Sh AUTHORS The .Nm command and this manual page were written by .An Dag-Erling Sm\(/orgrav Aq Mt des@FreeBSD.org . diff --git a/usr.bin/sockstat/sockstat.c b/usr.bin/sockstat/sockstat.c index 26f31d96b8e0..109b254b7438 100644 --- a/usr.bin/sockstat/sockstat.c +++ b/usr.bin/sockstat/sockstat.c @@ -1,1413 +1,1418 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2002 Dag-Erling Coïdan Smørgrav * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer * in this position and unchanged. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. The name of the author may not be used to endorse or promote products * derived from this software without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #define _WANT_UNPCB #include #include #include #include #include #include #define TCPSTATES /* load state names */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define sstosin(ss) ((struct sockaddr_in *)(ss)) #define sstosin6(ss) ((struct sockaddr_in6 *)(ss)) #define sstosun(ss) ((struct sockaddr_un *)(ss)) #define sstosa(ss) ((struct sockaddr *)(ss)) static int opt_4; /* Show IPv4 sockets */ static int opt_6; /* Show IPv6 sockets */ static int opt_C; /* Show congestion control */ static int opt_c; /* Show connected sockets */ static int opt_j; /* Show specified jail */ static int opt_L; /* Don't show IPv4 or IPv6 loopback sockets */ static int opt_l; /* Show listening sockets */ static int opt_n; /* Don't resolve UIDs to user names */ static int opt_q; /* Don't show header */ static int opt_S; /* Show protocol stack if applicable */ static int opt_s; /* Show protocol state if applicable */ static int opt_U; /* Show remote UDP encapsulation port number */ static int opt_u; /* Show Unix domain sockets */ static int opt_v; /* Verbose mode */ static int opt_w; /* Wide print area for addresses */ /* * Default protocols to use if no -P was defined. */ static const char *default_protos[] = {"sctp", "tcp", "udp", "divert" }; static size_t default_numprotos = nitems(default_protos); static int *protos; /* protocols to use */ static size_t numprotos; /* allocated size of protos[] */ static int *ports; #define INT_BIT (sizeof(int)*CHAR_BIT) #define SET_PORT(p) do { ports[p / INT_BIT] |= 1 << (p % INT_BIT); } while (0) #define CHK_PORT(p) (ports[p / INT_BIT] & (1 << (p % INT_BIT))) struct addr { struct sockaddr_storage address; unsigned int encaps_port; int state; struct addr *next; }; struct sock { kvaddr_t socket; kvaddr_t pcb; int shown; int vflag; int family; int proto; int state; const char *protoname; char stack[TCP_FUNCTION_NAME_LEN_MAX]; char cc[TCP_CA_NAME_MAX]; struct addr *laddr; struct addr *faddr; struct sock *next; }; #define HASHSIZE 1009 static struct sock *sockhash[HASHSIZE]; static struct xfile *xfiles; static int nxfiles; static int xprintf(const char *fmt, ...) { va_list ap; int len; va_start(ap, fmt); len = vprintf(fmt, ap); va_end(ap); if (len < 0) err(1, "printf()"); return (len); } static int get_proto_type(const char *proto) { struct protoent *pent; if (strlen(proto) == 0) return (0); pent = getprotobyname(proto); if (pent == NULL) { warn("getprotobyname"); return (-1); } return (pent->p_proto); } static void init_protos(int num) { int proto_count = 0; if (num > 0) { proto_count = num; } else { /* Find the maximum number of possible protocols. */ while (getprotoent() != NULL) proto_count++; endprotoent(); } if ((protos = malloc(sizeof(int) * proto_count)) == NULL) err(1, "malloc"); numprotos = proto_count; } static int parse_protos(char *protospec) { char *prot; int proto_type, proto_index; if (protospec == NULL) return (-1); init_protos(0); proto_index = 0; while ((prot = strsep(&protospec, ",")) != NULL) { if (strlen(prot) == 0) continue; proto_type = get_proto_type(prot); if (proto_type != -1) protos[proto_index++] = proto_type; } numprotos = proto_index; return (proto_index); } static void parse_ports(const char *portspec) { const char *p, *q; int port, end; if (ports == NULL) if ((ports = calloc(65536 / INT_BIT, sizeof(int))) == NULL) err(1, "calloc()"); p = portspec; while (*p != '\0') { if (!isdigit(*p)) errx(1, "syntax error in port range"); for (q = p; *q != '\0' && isdigit(*q); ++q) /* nothing */ ; for (port = 0; p < q; ++p) port = port * 10 + digittoint(*p); if (port < 0 || port > 65535) errx(1, "invalid port number"); SET_PORT(port); switch (*p) { case '-': ++p; break; case ',': ++p; /* fall through */ case '\0': default: continue; } for (q = p; *q != '\0' && isdigit(*q); ++q) /* nothing */ ; for (end = 0; p < q; ++p) end = end * 10 + digittoint(*p); if (end < port || end > 65535) errx(1, "invalid port number"); while (port++ < end) SET_PORT(port); if (*p == ',') ++p; } } static void sockaddr(struct sockaddr_storage *ss, int af, void *addr, int port) { struct sockaddr_in *sin4; struct sockaddr_in6 *sin6; bzero(ss, sizeof(*ss)); switch (af) { case AF_INET: sin4 = sstosin(ss); sin4->sin_len = sizeof(*sin4); sin4->sin_family = af; sin4->sin_port = port; sin4->sin_addr = *(struct in_addr *)addr; break; case AF_INET6: sin6 = sstosin6(ss); sin6->sin6_len = sizeof(*sin6); sin6->sin6_family = af; sin6->sin6_port = port; sin6->sin6_addr = *(struct in6_addr *)addr; #define s6_addr16 __u6_addr.__u6_addr16 if (IN6_IS_ADDR_LINKLOCAL(&sin6->sin6_addr)) { sin6->sin6_scope_id = ntohs(sin6->sin6_addr.s6_addr16[1]); sin6->sin6_addr.s6_addr16[1] = 0; } break; default: abort(); } } static void free_socket(struct sock *sock) { struct addr *cur, *next; cur = sock->laddr; while (cur != NULL) { next = cur->next; free(cur); cur = next; } cur = sock->faddr; while (cur != NULL) { next = cur->next; free(cur); cur = next; } free(sock); } static void gather_sctp(void) { struct sock *sock; struct addr *laddr, *prev_laddr, *faddr, *prev_faddr; struct xsctp_inpcb *xinpcb; struct xsctp_tcb *xstcb; struct xsctp_raddr *xraddr; struct xsctp_laddr *xladdr; const char *varname; size_t len, offset; char *buf; int hash, vflag; int no_stcb, local_all_loopback, foreign_all_loopback; vflag = 0; if (opt_4) vflag |= INP_IPV4; if (opt_6) vflag |= INP_IPV6; varname = "net.inet.sctp.assoclist"; if (sysctlbyname(varname, 0, &len, 0, 0) < 0) { if (errno != ENOENT) err(1, "sysctlbyname()"); return; } if ((buf = (char *)malloc(len)) == NULL) { err(1, "malloc()"); return; } if (sysctlbyname(varname, buf, &len, 0, 0) < 0) { err(1, "sysctlbyname()"); free(buf); return; } xinpcb = (struct xsctp_inpcb *)(void *)buf; offset = sizeof(struct xsctp_inpcb); while ((offset < len) && (xinpcb->last == 0)) { if ((sock = calloc(1, sizeof *sock)) == NULL) err(1, "malloc()"); sock->socket = xinpcb->socket; sock->proto = IPPROTO_SCTP; sock->protoname = "sctp"; if (xinpcb->maxqlen == 0) sock->state = SCTP_CLOSED; else sock->state = SCTP_LISTEN; if (xinpcb->flags & SCTP_PCB_FLAGS_BOUND_V6) { sock->family = AF_INET6; /* * Currently there is no way to distinguish between * IPv6 only sockets or dual family sockets. * So mark it as dual socket. */ sock->vflag = INP_IPV6 | INP_IPV4; } else { sock->family = AF_INET; sock->vflag = INP_IPV4; } prev_laddr = NULL; local_all_loopback = 1; while (offset < len) { xladdr = (struct xsctp_laddr *)(void *)(buf + offset); offset += sizeof(struct xsctp_laddr); if (xladdr->last == 1) break; if ((laddr = calloc(1, sizeof(struct addr))) == NULL) err(1, "malloc()"); switch (xladdr->address.sa.sa_family) { case AF_INET: #define __IN_IS_ADDR_LOOPBACK(pina) \ ((ntohl((pina)->s_addr) >> IN_CLASSA_NSHIFT) == IN_LOOPBACKNET) if (!__IN_IS_ADDR_LOOPBACK( &xladdr->address.sin.sin_addr)) local_all_loopback = 0; #undef __IN_IS_ADDR_LOOPBACK sockaddr(&laddr->address, AF_INET, &xladdr->address.sin.sin_addr, htons(xinpcb->local_port)); break; case AF_INET6: if (!IN6_IS_ADDR_LOOPBACK( &xladdr->address.sin6.sin6_addr)) local_all_loopback = 0; sockaddr(&laddr->address, AF_INET6, &xladdr->address.sin6.sin6_addr, htons(xinpcb->local_port)); break; default: errx(1, "address family %d not supported", xladdr->address.sa.sa_family); } laddr->next = NULL; if (prev_laddr == NULL) sock->laddr = laddr; else prev_laddr->next = laddr; prev_laddr = laddr; } if (sock->laddr == NULL) { if ((sock->laddr = calloc(1, sizeof(struct addr))) == NULL) err(1, "malloc()"); sock->laddr->address.ss_family = sock->family; if (sock->family == AF_INET) sock->laddr->address.ss_len = sizeof(struct sockaddr_in); else sock->laddr->address.ss_len = sizeof(struct sockaddr_in6); local_all_loopback = 0; } if ((sock->faddr = calloc(1, sizeof(struct addr))) == NULL) err(1, "malloc()"); sock->faddr->address.ss_family = sock->family; if (sock->family == AF_INET) sock->faddr->address.ss_len = sizeof(struct sockaddr_in); else sock->faddr->address.ss_len = sizeof(struct sockaddr_in6); no_stcb = 1; while (offset < len) { xstcb = (struct xsctp_tcb *)(void *)(buf + offset); offset += sizeof(struct xsctp_tcb); if (no_stcb) { if (opt_l && (sock->vflag & vflag) && (!opt_L || !local_all_loopback) && ((xinpcb->flags & SCTP_PCB_FLAGS_UDPTYPE) || (xstcb->last == 1))) { hash = (int)((uintptr_t)sock->socket % HASHSIZE); sock->next = sockhash[hash]; sockhash[hash] = sock; } else { free_socket(sock); } } if (xstcb->last == 1) break; no_stcb = 0; if (opt_c) { if ((sock = calloc(1, sizeof *sock)) == NULL) err(1, "malloc()"); sock->socket = xinpcb->socket; sock->proto = IPPROTO_SCTP; sock->protoname = "sctp"; sock->state = (int)xstcb->state; if (xinpcb->flags & SCTP_PCB_FLAGS_BOUND_V6) { sock->family = AF_INET6; /* * Currently there is no way to distinguish * between IPv6 only sockets or dual family * sockets. So mark it as dual socket. */ sock->vflag = INP_IPV6 | INP_IPV4; } else { sock->family = AF_INET; sock->vflag = INP_IPV4; } } prev_laddr = NULL; local_all_loopback = 1; while (offset < len) { xladdr = (struct xsctp_laddr *)(void *)(buf + offset); offset += sizeof(struct xsctp_laddr); if (xladdr->last == 1) break; if (!opt_c) continue; laddr = calloc(1, sizeof(struct addr)); if (laddr == NULL) err(1, "malloc()"); switch (xladdr->address.sa.sa_family) { case AF_INET: #define __IN_IS_ADDR_LOOPBACK(pina) \ ((ntohl((pina)->s_addr) >> IN_CLASSA_NSHIFT) == IN_LOOPBACKNET) if (!__IN_IS_ADDR_LOOPBACK( &xladdr->address.sin.sin_addr)) local_all_loopback = 0; #undef __IN_IS_ADDR_LOOPBACK sockaddr(&laddr->address, AF_INET, &xladdr->address.sin.sin_addr, htons(xstcb->local_port)); break; case AF_INET6: if (!IN6_IS_ADDR_LOOPBACK( &xladdr->address.sin6.sin6_addr)) local_all_loopback = 0; sockaddr(&laddr->address, AF_INET6, &xladdr->address.sin6.sin6_addr, htons(xstcb->local_port)); break; default: errx(1, "address family %d not supported", xladdr->address.sa.sa_family); } laddr->next = NULL; if (prev_laddr == NULL) sock->laddr = laddr; else prev_laddr->next = laddr; prev_laddr = laddr; } prev_faddr = NULL; foreign_all_loopback = 1; while (offset < len) { xraddr = (struct xsctp_raddr *)(void *)(buf + offset); offset += sizeof(struct xsctp_raddr); if (xraddr->last == 1) break; if (!opt_c) continue; faddr = calloc(1, sizeof(struct addr)); if (faddr == NULL) err(1, "malloc()"); switch (xraddr->address.sa.sa_family) { case AF_INET: #define __IN_IS_ADDR_LOOPBACK(pina) \ ((ntohl((pina)->s_addr) >> IN_CLASSA_NSHIFT) == IN_LOOPBACKNET) if (!__IN_IS_ADDR_LOOPBACK( &xraddr->address.sin.sin_addr)) foreign_all_loopback = 0; #undef __IN_IS_ADDR_LOOPBACK sockaddr(&faddr->address, AF_INET, &xraddr->address.sin.sin_addr, htons(xstcb->remote_port)); break; case AF_INET6: if (!IN6_IS_ADDR_LOOPBACK( &xraddr->address.sin6.sin6_addr)) foreign_all_loopback = 0; sockaddr(&faddr->address, AF_INET6, &xraddr->address.sin6.sin6_addr, htons(xstcb->remote_port)); break; default: errx(1, "address family %d not supported", xraddr->address.sa.sa_family); } faddr->encaps_port = xraddr->encaps_port; faddr->state = xraddr->state; faddr->next = NULL; if (prev_faddr == NULL) sock->faddr = faddr; else prev_faddr->next = faddr; prev_faddr = faddr; } if (opt_c) { if ((sock->vflag & vflag) && (!opt_L || !(local_all_loopback || foreign_all_loopback))) { hash = (int)((uintptr_t)sock->socket % HASHSIZE); sock->next = sockhash[hash]; sockhash[hash] = sock; } else { free_socket(sock); } } } xinpcb = (struct xsctp_inpcb *)(void *)(buf + offset); offset += sizeof(struct xsctp_inpcb); } free(buf); } static void gather_inet(int proto) { struct xinpgen *xig, *exig; struct xinpcb *xip; struct xtcpcb *xtp = NULL; struct xsocket *so; struct sock *sock; struct addr *laddr, *faddr; const char *varname, *protoname; size_t len, bufsize; void *buf; int hash, retry, vflag; vflag = 0; if (opt_4) vflag |= INP_IPV4; if (opt_6) vflag |= INP_IPV6; switch (proto) { case IPPROTO_TCP: varname = "net.inet.tcp.pcblist"; protoname = "tcp"; break; case IPPROTO_UDP: varname = "net.inet.udp.pcblist"; protoname = "udp"; break; case IPPROTO_DIVERT: varname = "net.inet.divert.pcblist"; protoname = "div"; break; default: errx(1, "protocol %d not supported", proto); } buf = NULL; bufsize = 8192; retry = 5; do { for (;;) { if ((buf = realloc(buf, bufsize)) == NULL) err(1, "realloc()"); len = bufsize; if (sysctlbyname(varname, buf, &len, NULL, 0) == 0) break; if (errno == ENOENT) goto out; if (errno != ENOMEM || len != bufsize) err(1, "sysctlbyname()"); bufsize *= 2; } xig = (struct xinpgen *)buf; exig = (struct xinpgen *)(void *) ((char *)buf + len - sizeof *exig); if (xig->xig_len != sizeof *xig || exig->xig_len != sizeof *exig) errx(1, "struct xinpgen size mismatch"); } while (xig->xig_gen != exig->xig_gen && retry--); if (xig->xig_gen != exig->xig_gen && opt_v) warnx("warning: data may be inconsistent"); for (;;) { xig = (struct xinpgen *)(void *)((char *)xig + xig->xig_len); if (xig >= exig) break; switch (proto) { case IPPROTO_TCP: xtp = (struct xtcpcb *)xig; xip = &xtp->xt_inp; if (xtp->xt_len != sizeof(*xtp)) { warnx("struct xtcpcb size mismatch"); goto out; } protoname = xtp->t_flags & TF_TOE ? "toe" : "tcp"; break; case IPPROTO_UDP: case IPPROTO_DIVERT: xip = (struct xinpcb *)xig; if (xip->xi_len != sizeof(*xip)) { warnx("struct xinpcb size mismatch"); goto out; } break; default: errx(1, "protocol %d not supported", proto); } so = &xip->xi_socket; if ((xip->inp_vflag & vflag) == 0) continue; if (xip->inp_vflag & INP_IPV4) { if ((xip->inp_fport == 0 && !opt_l) || (xip->inp_fport != 0 && !opt_c)) continue; #define __IN_IS_ADDR_LOOPBACK(pina) \ ((ntohl((pina)->s_addr) >> IN_CLASSA_NSHIFT) == IN_LOOPBACKNET) if (opt_L && (__IN_IS_ADDR_LOOPBACK(&xip->inp_faddr) || __IN_IS_ADDR_LOOPBACK(&xip->inp_laddr))) continue; #undef __IN_IS_ADDR_LOOPBACK } else if (xip->inp_vflag & INP_IPV6) { if ((xip->inp_fport == 0 && !opt_l) || (xip->inp_fport != 0 && !opt_c)) continue; if (opt_L && (IN6_IS_ADDR_LOOPBACK(&xip->in6p_faddr) || IN6_IS_ADDR_LOOPBACK(&xip->in6p_laddr))) continue; } else { if (opt_v) warnx("invalid vflag 0x%x", xip->inp_vflag); continue; } if ((sock = calloc(1, sizeof(*sock))) == NULL) err(1, "malloc()"); if ((laddr = calloc(1, sizeof *laddr)) == NULL) err(1, "malloc()"); if ((faddr = calloc(1, sizeof *faddr)) == NULL) err(1, "malloc()"); sock->socket = so->xso_so; sock->proto = proto; if (xip->inp_vflag & INP_IPV4) { sock->family = AF_INET; sockaddr(&laddr->address, sock->family, &xip->inp_laddr, xip->inp_lport); sockaddr(&faddr->address, sock->family, &xip->inp_faddr, xip->inp_fport); } else if (xip->inp_vflag & INP_IPV6) { sock->family = AF_INET6; sockaddr(&laddr->address, sock->family, &xip->in6p_laddr, xip->inp_lport); sockaddr(&faddr->address, sock->family, &xip->in6p_faddr, xip->inp_fport); } + if (proto == IPPROTO_TCP) + faddr->encaps_port = xtp->xt_encaps_port; laddr->next = NULL; faddr->next = NULL; sock->laddr = laddr; sock->faddr = faddr; sock->vflag = xip->inp_vflag; if (proto == IPPROTO_TCP) { sock->state = xtp->t_state; memcpy(sock->stack, xtp->xt_stack, TCP_FUNCTION_NAME_LEN_MAX); memcpy(sock->cc, xtp->xt_cc, TCP_CA_NAME_MAX); } sock->protoname = protoname; hash = (int)((uintptr_t)sock->socket % HASHSIZE); sock->next = sockhash[hash]; sockhash[hash] = sock; } out: free(buf); } static void gather_unix(int proto) { struct xunpgen *xug, *exug; struct xunpcb *xup; struct sock *sock; struct addr *laddr, *faddr; const char *varname, *protoname; size_t len, bufsize; void *buf; int hash, retry; switch (proto) { case SOCK_STREAM: varname = "net.local.stream.pcblist"; protoname = "stream"; break; case SOCK_DGRAM: varname = "net.local.dgram.pcblist"; protoname = "dgram"; break; case SOCK_SEQPACKET: varname = "net.local.seqpacket.pcblist"; protoname = "seqpac"; break; default: abort(); } buf = NULL; bufsize = 8192; retry = 5; do { for (;;) { if ((buf = realloc(buf, bufsize)) == NULL) err(1, "realloc()"); len = bufsize; if (sysctlbyname(varname, buf, &len, NULL, 0) == 0) break; if (errno != ENOMEM || len != bufsize) err(1, "sysctlbyname()"); bufsize *= 2; } xug = (struct xunpgen *)buf; exug = (struct xunpgen *)(void *) ((char *)buf + len - sizeof(*exug)); if (xug->xug_len != sizeof(*xug) || exug->xug_len != sizeof(*exug)) { warnx("struct xinpgen size mismatch"); goto out; } } while (xug->xug_gen != exug->xug_gen && retry--); if (xug->xug_gen != exug->xug_gen && opt_v) warnx("warning: data may be inconsistent"); for (;;) { xug = (struct xunpgen *)(void *)((char *)xug + xug->xug_len); if (xug >= exug) break; xup = (struct xunpcb *)xug; if (xup->xu_len != sizeof(*xup)) { warnx("struct xunpcb size mismatch"); goto out; } if ((xup->unp_conn == 0 && !opt_l) || (xup->unp_conn != 0 && !opt_c)) continue; if ((sock = calloc(1, sizeof(*sock))) == NULL) err(1, "malloc()"); if ((laddr = calloc(1, sizeof *laddr)) == NULL) err(1, "malloc()"); if ((faddr = calloc(1, sizeof *faddr)) == NULL) err(1, "malloc()"); sock->socket = xup->xu_socket.xso_so; sock->pcb = xup->xu_unpp; sock->proto = proto; sock->family = AF_UNIX; sock->protoname = protoname; if (xup->xu_addr.sun_family == AF_UNIX) laddr->address = *(struct sockaddr_storage *)(void *)&xup->xu_addr; else if (xup->unp_conn != 0) *(kvaddr_t*)&(faddr->address) = xup->unp_conn; laddr->next = NULL; faddr->next = NULL; sock->laddr = laddr; sock->faddr = faddr; hash = (int)((uintptr_t)sock->socket % HASHSIZE); sock->next = sockhash[hash]; sockhash[hash] = sock; } out: free(buf); } static void getfiles(void) { size_t len, olen; olen = len = sizeof(*xfiles); if ((xfiles = malloc(len)) == NULL) err(1, "malloc()"); while (sysctlbyname("kern.file", xfiles, &len, 0, 0) == -1) { if (errno != ENOMEM || len != olen) err(1, "sysctlbyname()"); olen = len *= 2; if ((xfiles = realloc(xfiles, len)) == NULL) err(1, "realloc()"); } if (len > 0 && xfiles->xf_size != sizeof(*xfiles)) errx(1, "struct xfile size mismatch"); nxfiles = len / sizeof(*xfiles); } static int printaddr(struct sockaddr_storage *ss) { struct sockaddr_un *sun; char addrstr[NI_MAXHOST] = { '\0', '\0' }; int error, off, port = 0; switch (ss->ss_family) { case AF_INET: if (inet_lnaof(sstosin(ss)->sin_addr) == INADDR_ANY) addrstr[0] = '*'; port = ntohs(sstosin(ss)->sin_port); break; case AF_INET6: if (IN6_IS_ADDR_UNSPECIFIED(&sstosin6(ss)->sin6_addr)) addrstr[0] = '*'; port = ntohs(sstosin6(ss)->sin6_port); break; case AF_UNIX: sun = sstosun(ss); off = (int)((char *)&sun->sun_path - (char *)sun); return (xprintf("%.*s", sun->sun_len - off, sun->sun_path)); } if (addrstr[0] == '\0') { error = getnameinfo(sstosa(ss), ss->ss_len, addrstr, sizeof(addrstr), NULL, 0, NI_NUMERICHOST); if (error) errx(1, "getnameinfo()"); } if (port == 0) return xprintf("%s:*", addrstr); else return xprintf("%s:%d", addrstr, port); } static const char * getprocname(pid_t pid) { static struct kinfo_proc proc; size_t len; int mib[4]; mib[0] = CTL_KERN; mib[1] = KERN_PROC; mib[2] = KERN_PROC_PID; mib[3] = (int)pid; len = sizeof(proc); if (sysctl(mib, nitems(mib), &proc, &len, NULL, 0) == -1) { /* Do not warn if the process exits before we get its name. */ if (errno != ESRCH) warn("sysctl()"); return ("??"); } return (proc.ki_comm); } static int getprocjid(pid_t pid) { static struct kinfo_proc proc; size_t len; int mib[4]; mib[0] = CTL_KERN; mib[1] = KERN_PROC; mib[2] = KERN_PROC_PID; mib[3] = (int)pid; len = sizeof(proc); if (sysctl(mib, nitems(mib), &proc, &len, NULL, 0) == -1) { /* Do not warn if the process exits before we get its jid. */ if (errno != ESRCH) warn("sysctl()"); return (-1); } return (proc.ki_jid); } static int check_ports(struct sock *s) { int port; struct addr *addr; if (ports == NULL) return (1); if ((s->family != AF_INET) && (s->family != AF_INET6)) return (1); for (addr = s->laddr; addr != NULL; addr = addr->next) { if (s->family == AF_INET) port = ntohs(sstosin(&addr->address)->sin_port); else port = ntohs(sstosin6(&addr->address)->sin6_port); if (CHK_PORT(port)) return (1); } for (addr = s->faddr; addr != NULL; addr = addr->next) { if (s->family == AF_INET) port = ntohs(sstosin(&addr->address)->sin_port); else port = ntohs(sstosin6(&addr->address)->sin6_port); if (CHK_PORT(port)) return (1); } return (0); } static const char * sctp_conn_state(int state) { switch (state) { case SCTP_CLOSED: return "CLOSED"; break; case SCTP_BOUND: return "BOUND"; break; case SCTP_LISTEN: return "LISTEN"; break; case SCTP_COOKIE_WAIT: return "COOKIE_WAIT"; break; case SCTP_COOKIE_ECHOED: return "COOKIE_ECHOED"; break; case SCTP_ESTABLISHED: return "ESTABLISHED"; break; case SCTP_SHUTDOWN_SENT: return "SHUTDOWN_SENT"; break; case SCTP_SHUTDOWN_RECEIVED: return "SHUTDOWN_RECEIVED"; break; case SCTP_SHUTDOWN_ACK_SENT: return "SHUTDOWN_ACK_SENT"; break; case SCTP_SHUTDOWN_PENDING: return "SHUTDOWN_PENDING"; break; default: return "UNKNOWN"; break; } } static const char * sctp_path_state(int state) { switch (state) { case SCTP_UNCONFIRMED: return "UNCONFIRMED"; break; case SCTP_ACTIVE: return "ACTIVE"; break; case SCTP_INACTIVE: return "INACTIVE"; break; default: return "UNKNOWN"; break; } } static void displaysock(struct sock *s, int pos) { kvaddr_t p; int hash, first, offset; struct addr *laddr, *faddr; struct sock *s_tmp; while (pos < 29) pos += xprintf(" "); pos += xprintf("%s", s->protoname); if (s->vflag & INP_IPV4) pos += xprintf("4"); if (s->vflag & INP_IPV6) pos += xprintf("6"); if (s->vflag & (INP_IPV4 | INP_IPV6)) pos += xprintf(" "); laddr = s->laddr; faddr = s->faddr; first = 1; while (laddr != NULL || faddr != NULL) { offset = 36; while (pos < offset) pos += xprintf(" "); switch (s->family) { case AF_INET: case AF_INET6: if (laddr != NULL) { pos += printaddr(&laddr->address); if (s->family == AF_INET6 && pos >= 58) pos += xprintf(" "); } offset += opt_w ? 46 : 22; while (pos < offset) pos += xprintf(" "); if (faddr != NULL) pos += printaddr(&faddr->address); offset += opt_w ? 46 : 22; break; case AF_UNIX: if ((laddr == NULL) || (faddr == NULL)) errx(1, "laddr = %p or faddr = %p is NULL", (void *)laddr, (void *)faddr); /* server */ if (laddr->address.ss_len > 0) { pos += printaddr(&laddr->address); break; } /* client */ p = *(kvaddr_t*)&(faddr->address); if (p == 0) { pos += xprintf("(not connected)"); offset += opt_w ? 92 : 44; break; } pos += xprintf("-> "); for (hash = 0; hash < HASHSIZE; ++hash) { for (s_tmp = sockhash[hash]; s_tmp != NULL; s_tmp = s_tmp->next) if (s_tmp->pcb == p) break; if (s_tmp != NULL) break; } if (s_tmp == NULL || s_tmp->laddr == NULL || s_tmp->laddr->address.ss_len == 0) pos += xprintf("??"); else pos += printaddr(&s_tmp->laddr->address); offset += opt_w ? 92 : 44; break; default: abort(); } if (opt_U) { if (faddr != NULL && - s->proto == IPPROTO_SCTP && - s->state != SCTP_CLOSED && - s->state != SCTP_BOUND && - s->state != SCTP_LISTEN) { + ((s->proto == IPPROTO_SCTP && + s->state != SCTP_CLOSED && + s->state != SCTP_BOUND && + s->state != SCTP_LISTEN) || + (s->proto == IPPROTO_TCP && + s->state != TCPS_CLOSED && + s->state != TCPS_LISTEN))) { while (pos < offset) pos += xprintf(" "); pos += xprintf("%u", ntohs(faddr->encaps_port)); } offset += 7; } if (opt_s) { if (faddr != NULL && s->proto == IPPROTO_SCTP && s->state != SCTP_CLOSED && s->state != SCTP_BOUND && s->state != SCTP_LISTEN) { while (pos < offset) pos += xprintf(" "); pos += xprintf("%s", sctp_path_state(faddr->state)); } offset += 13; } if (first) { if (opt_s) { if (s->proto == IPPROTO_SCTP || s->proto == IPPROTO_TCP) { while (pos < offset) pos += xprintf(" "); switch (s->proto) { case IPPROTO_SCTP: pos += xprintf("%s", sctp_conn_state(s->state)); break; case IPPROTO_TCP: if (s->state >= 0 && s->state < TCP_NSTATES) pos += xprintf("%s", tcpstates[s->state]); else pos += xprintf("?"); break; } } offset += 13; } if (opt_S) { if (s->proto == IPPROTO_TCP) { while (pos < offset) pos += xprintf(" "); pos += xprintf("%.*s", TCP_FUNCTION_NAME_LEN_MAX, s->stack); } offset += TCP_FUNCTION_NAME_LEN_MAX + 1; } if (opt_C) { if (s->proto == IPPROTO_TCP) { while (pos < offset) pos += xprintf(" "); xprintf("%.*s", TCP_CA_NAME_MAX, s->cc); } offset += TCP_CA_NAME_MAX + 1; } } if (laddr != NULL) laddr = laddr->next; if (faddr != NULL) faddr = faddr->next; if ((laddr != NULL) || (faddr != NULL)) { xprintf("\n"); pos = 0; } first = 0; } xprintf("\n"); } static void display(void) { struct passwd *pwd; struct xfile *xf; struct sock *s; int hash, n, pos; if (opt_q != 1) { printf("%-8s %-10s %-5s %-2s %-6s %-*s %-*s", "USER", "COMMAND", "PID", "FD", "PROTO", opt_w ? 45 : 21, "LOCAL ADDRESS", opt_w ? 45 : 21, "FOREIGN ADDRESS"); if (opt_U) printf(" %-6s", "ENCAPS"); if (opt_s) { printf(" %-12s", "PATH STATE"); printf(" %-12s", "CONN STATE"); } if (opt_S) printf(" %-*.*s", TCP_FUNCTION_NAME_LEN_MAX, TCP_FUNCTION_NAME_LEN_MAX, "STACK"); if (opt_C) printf(" %-.*s", TCP_CA_NAME_MAX, "CC"); printf("\n"); } setpassent(1); for (xf = xfiles, n = 0; n < nxfiles; ++n, ++xf) { if (xf->xf_data == 0) continue; if (opt_j >= 0 && opt_j != getprocjid(xf->xf_pid)) continue; hash = (int)((uintptr_t)xf->xf_data % HASHSIZE); for (s = sockhash[hash]; s != NULL; s = s->next) { if (s->socket != xf->xf_data) continue; if (!check_ports(s)) continue; s->shown = 1; pos = 0; if (opt_n || (pwd = getpwuid(xf->xf_uid)) == NULL) pos += xprintf("%lu ", (u_long)xf->xf_uid); else pos += xprintf("%s ", pwd->pw_name); while (pos < 9) pos += xprintf(" "); pos += xprintf("%.10s", getprocname(xf->xf_pid)); while (pos < 20) pos += xprintf(" "); pos += xprintf("%lu ", (u_long)xf->xf_pid); while (pos < 26) pos += xprintf(" "); pos += xprintf("%d ", xf->xf_fd); displaysock(s, pos); } } if (opt_j >= 0) return; for (hash = 0; hash < HASHSIZE; hash++) { for (s = sockhash[hash]; s != NULL; s = s->next) { if (s->shown) continue; if (!check_ports(s)) continue; pos = 0; pos += xprintf("%-8s %-10s %-5s %-2s ", "?", "?", "?", "?"); displaysock(s, pos); } } } static int set_default_protos(void) { struct protoent *prot; const char *pname; size_t pindex; init_protos(default_numprotos); for (pindex = 0; pindex < default_numprotos; pindex++) { pname = default_protos[pindex]; prot = getprotobyname(pname); if (prot == NULL) err(1, "getprotobyname: %s", pname); protos[pindex] = prot->p_proto; } numprotos = pindex; return (pindex); } /* * Return the vnet property of the jail, or -1 on error. */ static int jail_getvnet(int jid) { struct iovec jiov[6]; int vnet; vnet = -1; jiov[0].iov_base = __DECONST(char *, "jid"); jiov[0].iov_len = sizeof("jid"); jiov[1].iov_base = &jid; jiov[1].iov_len = sizeof(jid); jiov[2].iov_base = __DECONST(char *, "vnet"); jiov[2].iov_len = sizeof("vnet"); jiov[3].iov_base = &vnet; jiov[3].iov_len = sizeof(vnet); jiov[4].iov_base = __DECONST(char *, "errmsg"); jiov[4].iov_len = sizeof("errmsg"); jiov[5].iov_base = jail_errmsg; jiov[5].iov_len = JAIL_ERRMSGLEN; jail_errmsg[0] = '\0'; if (jail_get(jiov, nitems(jiov), 0) < 0) { if (!jail_errmsg[0]) snprintf(jail_errmsg, JAIL_ERRMSGLEN, "jail_get: %s", strerror(errno)); return (-1); } return (vnet); } static void usage(void) { fprintf(stderr, "usage: sockstat [-46cLlSsUuvw] [-j jid] [-p ports] [-P protocols]\n"); exit(1); } int main(int argc, char *argv[]) { int protos_defined = -1; int o, i; opt_j = -1; while ((o = getopt(argc, argv, "46Ccj:Llnp:P:qSsUuvw")) != -1) switch (o) { case '4': opt_4 = 1; break; case '6': opt_6 = 1; break; case 'C': opt_C = 1; break; case 'c': opt_c = 1; break; case 'j': opt_j = jail_getid(optarg); if (opt_j < 0) errx(1, "%s", jail_errmsg); break; case 'L': opt_L = 1; break; case 'l': opt_l = 1; break; case 'n': opt_n = 1; break; case 'p': parse_ports(optarg); break; case 'P': protos_defined = parse_protos(optarg); break; case 'q': opt_q = 1; break; case 'S': opt_S = 1; break; case 's': opt_s = 1; break; case 'U': opt_U = 1; break; case 'u': opt_u = 1; break; case 'v': ++opt_v; break; case 'w': opt_w = 1; break; default: usage(); } argc -= optind; argv += optind; if (argc > 0) usage(); if (opt_j > 0) { switch (jail_getvnet(opt_j)) { case -1: errx(2, "%s", jail_errmsg); case JAIL_SYS_NEW: if (jail_attach(opt_j) < 0) err(3, "jail_attach()"); /* Set back to -1 for normal output in vnet jail. */ opt_j = -1; break; default: break; } } if ((!opt_4 && !opt_6) && protos_defined != -1) opt_4 = opt_6 = 1; if (!opt_4 && !opt_6 && !opt_u) opt_4 = opt_6 = opt_u = 1; if ((opt_4 || opt_6) && protos_defined == -1) protos_defined = set_default_protos(); if (!opt_c && !opt_l) opt_c = opt_l = 1; if (opt_4 || opt_6) { for (i = 0; i < protos_defined; i++) if (protos[i] == IPPROTO_SCTP) gather_sctp(); else gather_inet(protos[i]); } if (opt_u || (protos_defined == -1 && !opt_4 && !opt_6)) { gather_unix(SOCK_STREAM); gather_unix(SOCK_DGRAM); gather_unix(SOCK_SEQPACKET); } getfiles(); display(); exit(0); }