diff --git a/cddl/lib/libdtrace/tcp.d b/cddl/lib/libdtrace/tcp.d index afcfdfadaf1e..0ceb4021a16d 100644 --- a/cddl/lib/libdtrace/tcp.d +++ b/cddl/lib/libdtrace/tcp.d @@ -1,418 +1,418 @@ /* * CDDL HEADER START * * The contents of this file are subject to the terms of the * Common Development and Distribution License (the "License"). * You may not use this file except in compliance with the License. * * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE * or http://www.opensolaris.org/os/licensing. * See the License for the specific language governing permissions * and limitations under the License. * * When distributing Covered Code, include this CDDL HEADER in each * file and include the License file at usr/src/OPENSOLARIS.LICENSE. * If applicable, add the following below this CDDL HEADER, with the * fields enclosed by brackets "[]" replaced with your own identifying * information: Portions Copyright [yyyy] [name of copyright owner] * * CDDL HEADER END * * $FreeBSD$ */ /* * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved. * Copyright (c) 2013 Mark Johnston */ #pragma D depends_on library ip.d #pragma D depends_on module kernel #pragma D depends_on provider tcp /* * Convert a TCP state value to a string. */ #pragma D binding "1.6.3" TCPS_CLOSED inline int TCPS_CLOSED = 0; #pragma D binding "1.6.3" TCPS_LISTEN inline int TCPS_LISTEN = 1; #pragma D binding "1.6.3" TCPS_SYN_SENT inline int TCPS_SYN_SENT = 2; #pragma D binding "1.6.3" TCPS_SYN_RECEIVED inline int TCPS_SYN_RECEIVED = 3; #pragma D binding "1.6.3" TCPS_ESTABLISHED inline int TCPS_ESTABLISHED = 4; #pragma D binding "1.6.3" TCPS_CLOSE_WAIT inline int TCPS_CLOSE_WAIT = 5; #pragma D binding "1.6.3" TCPS_FIN_WAIT_1 inline int TCPS_FIN_WAIT_1 = 6; #pragma D binding "1.6.3" TCPS_CLOSING inline int TCPS_CLOSING = 7; #pragma D binding "1.6.3" TCPS_LAST_ACK inline int TCPS_LAST_ACK = 8; #pragma D binding "1.6.3" TCPS_FIN_WAIT_2 inline int TCPS_FIN_WAIT_2 = 9; #pragma D binding "1.6.3" TCPS_TIME_WAIT inline int TCPS_TIME_WAIT = 10; /* * For compatibility also provide the names used by Solaris. */ #pragma D binding "1.13" TCP_STATE_CLOSED inline int TCP_STATE_CLOSED = TCPS_CLOSED; #pragma D binding "1.13" TCP_STATE_LISTEN inline int TCP_STATE_LISTEN = TCPS_LISTEN; #pragma D binding "1.13" TCP_STATE_SYN_SENT inline int TCP_STATE_SYN_SENT = TCPS_SYN_SENT; #pragma D binding "1.13" TCP_STATE_SYN_RECEIVED inline int TCP_STATE_SYN_RECEIVED = TCPS_SYN_RECEIVED; #pragma D binding "1.13" TCP_STATE_ESTABLISHED inline int TCP_STATE_ESTABLISHED = TCPS_ESTABLISHED; #pragma D binding "1.13" TCP_STATE_CLOSE_WAIT inline int TCP_STATE_CLOSE_WAIT = TCPS_CLOSE_WAIT; #pragma D binding "1.13" TCP_STATE_FIN_WAIT_1 inline int TCP_STATE_FIN_WAIT_1 = TCPS_FIN_WAIT_1; #pragma D binding "1.13" TCP_STATE_CLOSING inline int TCP_STATE_CLOSING = TCPS_CLOSING; #pragma D binding "1.13" TCP_STATE_LAST_ACK inline int TCP_STATE_LAST_ACK = TCPS_LAST_ACK; #pragma D binding "1.13" TCP_STATE_FIN_WAIT_2 inline int TCP_STATE_FIN_WAIT_2 = TCPS_FIN_WAIT_2; #pragma D binding "1.13" TCP_STATE_TIME_WAIT inline int TCP_STATE_TIME_WAIT = TCPS_TIME_WAIT; /* TCP segment flags. */ #pragma D binding "1.6.3" TH_FIN inline uint8_t TH_FIN = 0x01; #pragma D binding "1.6.3" TH_SYN inline uint8_t TH_SYN = 0x02; #pragma D binding "1.6.3" TH_RST inline uint8_t TH_RST = 0x04; #pragma D binding "1.6.3" TH_PUSH inline uint8_t TH_PUSH = 0x08; #pragma D binding "1.6.3" TH_ACK inline uint8_t TH_ACK = 0x10; #pragma D binding "1.6.3" TH_URG inline uint8_t TH_URG = 0x20; #pragma D binding "1.6.3" TH_ECE inline uint8_t TH_ECE = 0x40; #pragma D binding "1.6.3" TH_CWR inline uint8_t TH_CWR = 0x80; /* TCP connection state strings. */ #pragma D binding "1.6.3" tcp_state_string inline string tcp_state_string[int32_t state] = state == TCPS_CLOSED ? "state-closed" : state == TCPS_LISTEN ? "state-listen" : state == TCPS_SYN_SENT ? "state-syn-sent" : state == TCPS_SYN_RECEIVED ? "state-syn-received" : state == TCPS_ESTABLISHED ? "state-established" : state == TCPS_CLOSE_WAIT ? "state-close-wait" : state == TCPS_FIN_WAIT_1 ? "state-fin-wait-1" : state == TCPS_CLOSING ? "state-closing" : state == TCPS_LAST_ACK ? "state-last-ack" : state == TCPS_FIN_WAIT_2 ? "state-fin-wait-2" : state == TCPS_TIME_WAIT ? "state-time-wait" : ""; /* * tcpsinfo contains stable TCP details from tcp_t. */ typedef struct tcpsinfo { uintptr_t tcps_addr; int tcps_local; /* is delivered locally, boolean */ int tcps_active; /* active open (from here), boolean */ uint16_t tcps_lport; /* local port */ uint16_t tcps_rport; /* remote port */ string tcps_laddr; /* local address, as a string */ string tcps_raddr; /* remote address, as a string */ int32_t tcps_state; /* TCP state */ uint32_t tcps_iss; /* Initial sequence # sent */ uint32_t tcps_irs; /* Initial sequence # received */ uint32_t tcps_suna; /* sequence # sent but unacked */ uint32_t tcps_smax; /* highest sequence number sent */ uint32_t tcps_snxt; /* next sequence # to send */ uint32_t tcps_rack; /* sequence # we have acked */ uint32_t tcps_rnxt; /* next sequence # expected */ u_long tcps_swnd; /* send window size */ int32_t tcps_snd_ws; /* send window scaling */ uint32_t tcps_swl1; /* window update seg seq number */ uint32_t tcps_swl2; /* window update seg ack number */ uint32_t tcps_rup; /* receive urgent pointer */ uint32_t tcps_radv; /* advertised window */ u_long tcps_rwnd; /* receive window size */ int32_t tcps_rcv_ws; /* receive window scaling */ u_long tcps_cwnd; /* congestion window */ u_long tcps_cwnd_ssthresh; /* threshold for congestion avoidance */ uint32_t tcps_srecover; /* for use in NewReno Fast Recovery */ uint32_t tcps_sack_fack; /* SACK sequence # we have acked */ uint32_t tcps_sack_snxt; /* next SACK seq # for retransmission */ uint32_t tcps_rto; /* round-trip timeout, msec */ uint32_t tcps_mss; /* max segment size */ int tcps_retransmit; /* retransmit send event, boolean */ int tcps_srtt; /* smoothed RTT in units of (TCP_RTT_SCALE*hz) */ int tcps_debug; /* socket has SO_DEBUG set */ int tcps_cookie; /* expose the socket's SO_USER_COOKIE */ int32_t tcps_dupacks; /* consecutive dup acks received */ uint32_t tcps_rtttime; /* RTT measurement start time */ uint32_t tcps_rtseq; /* sequence # being timed */ uint32_t tcps_ts_recent; /* timestamp echo data */ } tcpsinfo_t; /* * tcplsinfo provides the old tcp state for state changes. */ typedef struct tcplsinfo { int32_t tcps_state; /* previous TCP state */ } tcplsinfo_t; /* * tcpinfo is the TCP header fields. */ typedef struct tcpinfo { uint16_t tcp_sport; /* source port */ uint16_t tcp_dport; /* destination port */ uint32_t tcp_seq; /* sequence number */ uint32_t tcp_ack; /* acknowledgment number */ uint8_t tcp_offset; /* data offset, in bytes */ uint8_t tcp_flags; /* flags */ uint16_t tcp_window; /* window size */ uint16_t tcp_checksum; /* checksum */ uint16_t tcp_urgent; /* urgent data pointer */ struct tcphdr *tcp_hdr; /* raw TCP header */ } tcpinfo_t; /* * A clone of tcpinfo_t used to handle the fact that the TCP input path * overwrites some fields of the TCP header with their host-order equivalents. * Unfortunately, DTrace doesn't let us simply typedef a new name for struct * tcpinfo and define a separate translator for it. */ typedef struct tcpinfoh { uint16_t tcp_sport; /* source port */ uint16_t tcp_dport; /* destination port */ uint32_t tcp_seq; /* sequence number */ uint32_t tcp_ack; /* acknowledgment number */ uint8_t tcp_offset; /* data offset, in bytes */ uint8_t tcp_flags; /* flags */ uint16_t tcp_window; /* window size */ uint16_t tcp_checksum; /* checksum */ uint16_t tcp_urgent; /* urgent data pointer */ struct tcphdr *tcp_hdr; /* raw TCP header */ } tcpinfoh_t; #pragma D binding "1.6.3" translator translator csinfo_t < struct tcpcb *p > { cs_addr = NULL; - cs_cid = (uint64_t)(p == NULL ? 0 : p->t_inpcb); + cs_cid = (uint64_t)(p == NULL ? 0 : &p->t_inpcb); cs_pid = 0; cs_zoneid = 0; }; #pragma D binding "1.6.3" translator translator tcpsinfo_t < struct tcpcb *p > { tcps_addr = (uintptr_t)p; tcps_local = -1; /* XXX */ tcps_active = -1; /* XXX */ - tcps_lport = p == NULL ? 0 : ntohs(p->t_inpcb->inp_inc.inc_ie.ie_lport); - tcps_rport = p == NULL ? 0 : ntohs(p->t_inpcb->inp_inc.inc_ie.ie_fport); + tcps_lport = p == NULL ? 0 : ntohs(p->t_inpcb.inp_inc.inc_ie.ie_lport); + tcps_rport = p == NULL ? 0 : ntohs(p->t_inpcb.inp_inc.inc_ie.ie_fport); tcps_laddr = p == NULL ? "" : - p->t_inpcb->inp_vflag == INP_IPV4 ? - inet_ntoa(&p->t_inpcb->inp_inc.inc_ie.ie_dependladdr.id46_addr.ia46_addr4.s_addr) : - inet_ntoa6(&p->t_inpcb->inp_inc.inc_ie.ie_dependladdr.id6_addr); + p->t_inpcb.inp_vflag == INP_IPV4 ? + inet_ntoa(&p->t_inpcb.inp_inc.inc_ie.ie_dependladdr.id46_addr.ia46_addr4.s_addr) : + inet_ntoa6(&p->t_inpcb.inp_inc.inc_ie.ie_dependladdr.id6_addr); tcps_raddr = p == NULL ? "" : - p->t_inpcb->inp_vflag == INP_IPV4 ? - inet_ntoa(&p->t_inpcb->inp_inc.inc_ie.ie_dependfaddr.id46_addr.ia46_addr4.s_addr) : - inet_ntoa6(&p->t_inpcb->inp_inc.inc_ie.ie_dependfaddr.id6_addr); + p->t_inpcb.inp_vflag == INP_IPV4 ? + inet_ntoa(&p->t_inpcb.inp_inc.inc_ie.ie_dependfaddr.id46_addr.ia46_addr4.s_addr) : + inet_ntoa6(&p->t_inpcb.inp_inc.inc_ie.ie_dependfaddr.id6_addr); tcps_state = p == NULL ? -1 : p->t_state; tcps_iss = p == NULL ? 0 : p->iss; tcps_irs = p == NULL ? 0 : p->irs; tcps_suna = p == NULL ? 0 : p->snd_una; tcps_smax = p == NULL ? 0 : p->snd_max; tcps_snxt = p == NULL ? 0 : p->snd_nxt; tcps_rack = p == NULL ? 0 : p->last_ack_sent; tcps_rnxt = p == NULL ? 0 : p->rcv_nxt; tcps_swnd = p == NULL ? -1 : p->snd_wnd; tcps_snd_ws = p == NULL ? -1 : p->snd_scale; tcps_swl1 = p == NULL ? -1 : p->snd_wl1; tcps_swl2 = p == NULL ? -1 : p->snd_wl2; tcps_radv = p == NULL ? -1 : p->rcv_adv; tcps_rwnd = p == NULL ? -1 : p->rcv_wnd; tcps_rup = p == NULL ? -1 : p->rcv_up; tcps_rcv_ws = p == NULL ? -1 : p->rcv_scale; tcps_cwnd = p == NULL ? -1 : p->snd_cwnd; tcps_cwnd_ssthresh = p == NULL ? -1 : p->snd_ssthresh; tcps_srecover = p == NULL ? -1 : p->snd_recover; tcps_sack_fack = p == NULL ? 0 : p->snd_fack; tcps_sack_snxt = p == NULL ? 0 : p->snd_recover; tcps_rto = p == NULL ? -1 : (p->t_rxtcur * 1000) / `hz; tcps_mss = p == NULL ? -1 : p->t_maxseg; tcps_retransmit = p == NULL ? -1 : p->t_rxtshift > 0 ? 1 : 0; tcps_srtt = p == NULL ? -1 : p->t_srtt; /* smoothed RTT in units of (TCP_RTT_SCALE*hz) */ tcps_debug = p == NULL ? 0 : - p->t_inpcb->inp_socket->so_options & 1; + p->t_inpcb.inp_socket->so_options & 1; tcps_cookie = p == NULL ? -1 : - p->t_inpcb->inp_socket->so_user_cookie; + p->t_inpcb.inp_socket->so_user_cookie; tcps_dupacks = p == NULL ? -1 : p->t_dupacks; tcps_rtttime = p == NULL ? -1 : p->t_rtttime; tcps_rtseq = p == NULL ? -1 : p->t_rtseq; tcps_ts_recent = p == NULL ? -1 : p->ts_recent; }; #pragma D binding "1.6.3" translator translator tcpinfo_t < struct tcphdr *p > { tcp_sport = p == NULL ? 0 : ntohs(p->th_sport); tcp_dport = p == NULL ? 0 : ntohs(p->th_dport); tcp_seq = p == NULL ? -1 : ntohl(p->th_seq); tcp_ack = p == NULL ? -1 : ntohl(p->th_ack); tcp_offset = p == NULL ? -1 : (p->th_off >> 2); tcp_flags = p == NULL ? 0 : p->th_flags; tcp_window = p == NULL ? 0 : ntohs(p->th_win); tcp_checksum = p == NULL ? 0 : ntohs(p->th_sum); tcp_urgent = p == NULL ? 0 : ntohs(p->th_urp); tcp_hdr = (struct tcphdr *)p; }; /* * This translator differs from the one for tcpinfo_t in that the sequence * number, acknowledgement number, window size and urgent pointer are already * in host order and thus don't need to be converted. */ #pragma D binding "1.6.3" translator translator tcpinfoh_t < struct tcphdr *p > { tcp_sport = p == NULL ? 0 : ntohs(p->th_sport); tcp_dport = p == NULL ? 0 : ntohs(p->th_dport); tcp_seq = p == NULL ? -1 : p->th_seq; tcp_ack = p == NULL ? -1 : p->th_ack; tcp_offset = p == NULL ? -1 : (p->th_off >> 2); tcp_flags = p == NULL ? 0 : p->th_flags; tcp_window = p == NULL ? 0 : p->th_win; tcp_checksum = p == NULL ? 0 : ntohs(p->th_sum); tcp_urgent = p == NULL ? 0 : p->th_urp; tcp_hdr = (struct tcphdr *)p; }; #pragma D binding "1.6.3" translator translator tcplsinfo_t < int s > { tcps_state = s; }; /* Support for TCP debug */ #pragma D binding "1.12.1" TA_INPUT inline int TA_INPUT = 0; #pragma D binding "1.12.1" TA_OUTPUT inline int TA_OUTPUT = 1; #pragma D binding "1.12.1" TA_USER inline int TA_USER = 2; #pragma D binding "1.12.1" TA_RESPOND inline int TA_RESPOND = 3; #pragma D binding "1.12.1" TA_DROP inline int TA_DROP = 4; /* direction strings. */ #pragma D binding "1.12.1" tcpdebug_dir_string inline string tcpdebug_dir_string[uint8_t direction] = direction == TA_INPUT ? "input" : direction == TA_OUTPUT ? "output" : direction == TA_USER ? "user" : direction == TA_RESPOND ? "respond" : direction == TA_OUTPUT ? "drop" : "unknown" ; #pragma D binding "1.12.1" tcpflag_string inline string tcpflag_string[uint8_t flags] = flags & TH_FIN ? "FIN" : flags & TH_SYN ? "SYN" : flags & TH_RST ? "RST" : flags & TH_PUSH ? "PUSH" : flags & TH_ACK ? "ACK" : flags & TH_URG ? "URG" : flags & TH_ECE ? "ECE" : flags & TH_CWR ? "CWR" : "unknown" ; #pragma D binding "1.12.1" PRU_ATTACH inline int PRU_ATTACH = 0; #pragma D binding "1.12.1" PRU_DETACH inline int PRU_DETACH = 1; #pragma D binding "1.12.1" PRU_BIND inline int PRU_BIND = 2; #pragma D binding "1.12.1" PRU_LISTEN inline int PRU_LISTEN = 3; #pragma D binding "1.12.1" PRU_CONNECT inline int PRU_CONNECT = 4; #pragma D binding "1.12.1" PRU_ACCEPT inline int PRU_ACCEPT = 5 ; #pragma D binding "1.12.1" PRU_DISCONNECT inline int PRU_DISCONNECT = 6; #pragma D binding "1.12.1" PRU_SHUTDOWN inline int PRU_SHUTDOWN = 7; #pragma D binding "1.12.1" PRU_RCVD inline int PRU_RCVD = 8; #pragma D binding "1.12.1" PRU_SEND inline int PRU_SEND = 9; #pragma D binding "1.12.1" PRU_ABORT inline int PRU_ABORT = 10; #pragma D binding "1.12.1" PRU_CONTROL inline int PRU_CONTROL = 11; #pragma D binding "1.12.1" PRU_SENSE inline int PRU_SENSE = 12; #pragma D binding "1.12.1" PRU_RCVOOB inline int PRU_RCVOOB = 13; #pragma D binding "1.12.1" PRU_SENDOOB inline int PRU_SENDOOB = 14; #pragma D binding "1.12.1" PRU_SOCKADDR inline int PRU_SOCKADDR = 15; #pragma D binding "1.12.1" PRU_PEERADDR inline int PRU_PEERADDR = 16; #pragma D binding "1.12.1" PRU_CONNECT2 inline int PRU_CONNECT2 = 17; #pragma D binding "1.12.1" PRU_FASTTIMO inline int PRU_FASTTIMO = 18; #pragma D binding "1.12.1" PRU_SLOWTIMO inline int PRU_SLOWTIMO = 19; #pragma D binding "1.12.1" PRU_PROTORCV inline int PRU_PROTORCV = 20; #pragma D binding "1.12.1" PRU_PROTOSEND inline int PRU_PROTOSEND = 21; #pragma D binding "1.12.1" PRU_SEND_EOF inline int PRU_SEND_EOF = 22; #pragma D binding "1.12.1" PRU_SOSETLABEL inline int PRU_SOSETLABEL = 23; #pragma D binding "1.12.1" PRU_CLOSE inline int PRU_CLOSE = 24; #pragma D binding "1.12.1" PRU_FLUSH inline int PRU_FLUSH = 25; #pragma D binding "1.12.1" prureq_string inline string prureq_string[uint8_t req] = req == PRU_ATTACH ? "ATTACH" : req == PRU_DETACH ? "DETACH" : req == PRU_BIND ? "BIND" : req == PRU_LISTEN ? "LISTEN" : req == PRU_CONNECT ? "CONNECT" : req == PRU_ACCEPT ? "ACCEPT" : req == PRU_DISCONNECT ? "DISCONNECT" : req == PRU_SHUTDOWN ? "SHUTDOWN" : req == PRU_RCVD ? "RCVD" : req == PRU_SEND ? "SEND" : req == PRU_ABORT ? "ABORT" : req == PRU_CONTROL ? "CONTROL" : req == PRU_SENSE ? "SENSE" : req == PRU_RCVOOB ? "RCVOOB" : req == PRU_SENDOOB ? "SENDOOB" : req == PRU_SOCKADDR ? "SOCKADDR" : req == PRU_PEERADDR ? "PEERADDR" : req == PRU_CONNECT2 ? "CONNECT2" : req == PRU_FASTTIMO ? "FASTTIMO" : req == PRU_SLOWTIMO ? "SLOWTIMO" : req == PRU_PROTORCV ? "PROTORCV" : req == PRU_PROTOSEND ? "PROTOSEND" : req == PRU_SEND ? "SEND_EOF" : req == PRU_SOSETLABEL ? "SOSETLABEL" : req == PRU_CLOSE ? "CLOSE" : req == PRU_FLUSH ? "FLUSE" : "unknown" ; diff --git a/sys/conf/options b/sys/conf/options index e2720980f356..00019b1c2e5c 100644 --- a/sys/conf/options +++ b/sys/conf/options @@ -1,1032 +1,1032 @@ # $FreeBSD$ # # On the handling of kernel options # # All kernel options should be listed in NOTES, with suitable # descriptions. Negative options (options that make some code not # compile) should be commented out; LINT (generated from NOTES) should # compile as much code as possible. Try to structure option-using # code so that a single option only switch code on, or only switch # code off, to make it possible to have a full compile-test. If # necessary, you can check for COMPILING_LINT to get maximum code # coverage. # # All new options shall also be listed in either "conf/options" or # "conf/options.". Options that affect a single source-file # .[c|s] should be directed into "opt_.h", while options # that affect multiple files should either go in "opt_global.h" if # this is a kernel-wide option (used just about everywhere), or in # "opt_.h" if it affects only some files. # Note that the effect of listing only an option without a # header-file-name in conf/options (and cousins) is that the last # convention is followed. # # This handling scheme is not yet fully implemented. # # # Format of this file: # Option name filename # # If filename is missing, the default is # opt_.h AAC_DEBUG opt_aac.h AACRAID_DEBUG opt_aacraid.h AHC_ALLOW_MEMIO opt_aic7xxx.h AHC_TMODE_ENABLE opt_aic7xxx.h AHC_DUMP_EEPROM opt_aic7xxx.h AHC_DEBUG opt_aic7xxx.h AHC_DEBUG_OPTS opt_aic7xxx.h AHC_REG_PRETTY_PRINT opt_aic7xxx.h AHD_DEBUG opt_aic79xx.h AHD_DEBUG_OPTS opt_aic79xx.h AHD_TMODE_ENABLE opt_aic79xx.h AHD_REG_PRETTY_PRINT opt_aic79xx.h # Debugging options. ALT_BREAK_TO_DEBUGGER opt_kdb.h BREAK_TO_DEBUGGER opt_kdb.h BUF_TRACKING opt_global.h DDB DDB_BUFR_SIZE opt_ddb.h DDB_CAPTURE_DEFAULTBUFSIZE opt_ddb.h DDB_CAPTURE_MAXBUFSIZE opt_ddb.h DDB_CTF opt_ddb.h DDB_NUMSYM opt_ddb.h EARLY_PRINTF opt_global.h FULL_BUF_TRACKING opt_global.h GDB KDB opt_global.h KDB_TRACE opt_kdb.h KDB_UNATTENDED opt_kdb.h KLD_DEBUG opt_kld.h NUM_CORE_FILES opt_global.h QUEUE_MACRO_DEBUG_TRACE opt_global.h QUEUE_MACRO_DEBUG_TRASH opt_global.h SYSCTL_DEBUG opt_sysctl.h TEXTDUMP_PREFERRED opt_ddb.h TEXTDUMP_VERBOSE opt_ddb.h TSLOG opt_global.h TSLOGSIZE opt_global.h # Miscellaneous options. ALQ ALTERA_SDCARD_FAST_SIM opt_altera_sdcard.h ATSE_CFI_HACK opt_cfi.h AUDIT opt_global.h BOOTHOWTO opt_global.h BOOTVERBOSE opt_global.h CALLOUT_PROFILING CAPABILITIES opt_capsicum.h CAPABILITY_MODE opt_capsicum.h CC_CDG opt_global.h CC_CHD opt_global.h CC_CUBIC opt_global.h CC_DEFAULT opt_cc.h CC_DCTCP opt_global.h CC_HD opt_global.h CC_HTCP opt_global.h CC_NEWRENO opt_global.h CC_VEGAS opt_global.h COMPAT_43 opt_global.h COMPAT_43TTY opt_global.h COMPAT_FREEBSD4 opt_global.h COMPAT_FREEBSD5 opt_global.h COMPAT_FREEBSD6 opt_global.h COMPAT_FREEBSD7 opt_global.h COMPAT_FREEBSD9 opt_global.h COMPAT_FREEBSD10 opt_global.h COMPAT_FREEBSD11 opt_global.h COMPAT_FREEBSD12 opt_global.h COMPAT_FREEBSD13 opt_global.h COMPAT_LINUXKPI opt_dontuse.h _COMPAT_LINUX32 opt_compat.h # XXX: make sure opt_compat.h exists COMPILING_LINT opt_global.h CY_PCI_FASTINTR DEADLKRES opt_watchdog.h EXPERIMENTAL opt_global.h DIRECTIO FILEMON opt_dontuse.h FFCLOCK FULL_PREEMPTION opt_sched.h GZIO opt_gzio.h IMGACT_BINMISC opt_dontuse.h IPI_PREEMPTION opt_sched.h GEOM_BDE opt_geom.h GEOM_CACHE opt_geom.h GEOM_CONCAT opt_geom.h GEOM_ELI opt_geom.h GEOM_GATE opt_geom.h GEOM_JOURNAL opt_geom.h GEOM_LABEL opt_geom.h GEOM_LABEL_GPT opt_geom.h GEOM_LINUX_LVM opt_geom.h GEOM_MAP opt_geom.h GEOM_MIRROR opt_geom.h GEOM_MOUNTVER opt_geom.h GEOM_MULTIPATH opt_geom.h GEOM_NOP opt_geom.h GEOM_PART_APM opt_geom.h GEOM_PART_BSD opt_geom.h GEOM_PART_BSD64 opt_geom.h GEOM_PART_EBR opt_geom.h GEOM_PART_GPT opt_geom.h GEOM_PART_LDM opt_geom.h GEOM_PART_MBR opt_geom.h GEOM_PART_VTOC8 opt_geom.h GEOM_RAID opt_geom.h GEOM_RAID3 opt_geom.h GEOM_SHSEC opt_geom.h GEOM_STRIPE opt_geom.h GEOM_UZIP opt_geom.h GEOM_UZIP_DEBUG opt_geom.h GEOM_VINUM opt_geom.h GEOM_VIRSTOR opt_geom.h GEOM_ZERO opt_geom.h IFLIB opt_iflib.h KDTRACE_HOOKS opt_global.h KDTRACE_FRAME opt_kdtrace.h KN_HASHSIZE opt_kqueue.h KSTACK_MAX_PAGES KSTACK_PAGES KSTACK_USAGE_PROF KTRACE KTRACE_REQUEST_POOL opt_ktrace.h LIBICONV MAC opt_global.h MAC_BIBA opt_dontuse.h MAC_BSDEXTENDED opt_dontuse.h MAC_DDB opt_dontuse.h MAC_IFOFF opt_dontuse.h MAC_LOMAC opt_dontuse.h MAC_MLS opt_dontuse.h MAC_NONE opt_dontuse.h MAC_NTPD opt_dontuse.h MAC_PARTITION opt_dontuse.h MAC_PORTACL opt_dontuse.h MAC_PRIORITY opt_dontuse.h MAC_SEEOTHERUIDS opt_dontuse.h MAC_STATIC opt_mac.h MAC_STUB opt_dontuse.h MAC_TEST opt_dontuse.h MAC_VERIEXEC opt_dontuse.h MAC_VERIEXEC_SHA1 opt_dontuse.h MAC_VERIEXEC_SHA256 opt_dontuse.h MAC_VERIEXEC_SHA384 opt_dontuse.h MAC_VERIEXEC_SHA512 opt_dontuse.h MD_ROOT opt_md.h MD_ROOT_FSTYPE opt_md.h MD_ROOT_READONLY opt_md.h MD_ROOT_SIZE opt_md.h MD_ROOT_MEM opt_md.h MFI_DEBUG opt_mfi.h MFI_DECODE_LOG opt_mfi.h MPROF_BUFFERS opt_mprof.h MPROF_HASH_SIZE opt_mprof.h NEW_PCIB opt_global.h NO_ADAPTIVE_MUTEXES opt_adaptive_mutexes.h NO_ADAPTIVE_RWLOCKS NO_ADAPTIVE_SX NO_OBSOLETE_CODE opt_global.h NO_SYSCTL_DESCR opt_global.h NSWBUF_MIN opt_param.h MBUF_PACKET_ZONE_DISABLE opt_global.h PANIC_REBOOT_WAIT_TIME opt_panic.h PCI_HP opt_pci.h PCI_IOV opt_global.h PPC_DEBUG opt_ppc.h PPC_PROBE_CHIPSET opt_ppc.h PPS_SYNC opt_ntp.h PREEMPTION opt_sched.h QUOTA SCHED_4BSD opt_sched.h SCHED_STATS opt_sched.h SCHED_ULE opt_sched.h SLEEPQUEUE_PROFILING SLHCI_DEBUG opt_slhci.h STACK opt_stack.h SUIDDIR MSGMNB opt_sysvipc.h MSGMNI opt_sysvipc.h MSGSEG opt_sysvipc.h MSGSSZ opt_sysvipc.h MSGTQL opt_sysvipc.h SEMMNI opt_sysvipc.h SEMMNS opt_sysvipc.h SEMMNU opt_sysvipc.h SEMMSL opt_sysvipc.h SEMOPM opt_sysvipc.h SEMUME opt_sysvipc.h SHMALL opt_sysvipc.h SHMMAX opt_sysvipc.h SHMMAXPGS opt_sysvipc.h SHMMIN opt_sysvipc.h SHMMNI opt_sysvipc.h SHMSEG opt_sysvipc.h SYSVMSG opt_sysvipc.h SYSVSEM opt_sysvipc.h SYSVSHM opt_sysvipc.h SW_WATCHDOG opt_watchdog.h TCPHPTS opt_inet.h TURNSTILE_PROFILING UMTX_PROFILING UMTX_CHAINS opt_global.h VERBOSE_SYSINIT ZSTDIO opt_zstdio.h # Sanitizers COVERAGE opt_global.h KASAN opt_global.h KCOV KCSAN opt_global.h KMSAN opt_global.h KUBSAN opt_global.h # POSIX kernel options P1003_1B_MQUEUE opt_posix.h P1003_1B_SEMAPHORES opt_posix.h _KPOSIX_PRIORITY_SCHEDULING opt_posix.h # Do we want the config file compiled into the kernel? INCLUDE_CONFIG_FILE opt_config.h # Options for static filesystems. These should only be used at config # time, since the corresponding lkms cannot work if there are any static # dependencies. Unusability is enforced by hiding the defines for the # options in a never-included header. AUTOFS opt_dontuse.h CD9660 opt_dontuse.h EXT2FS opt_dontuse.h FDESCFS opt_dontuse.h FFS opt_dontuse.h FUSEFS opt_dontuse.h MSDOSFS opt_dontuse.h NULLFS opt_dontuse.h PROCFS opt_dontuse.h PSEUDOFS opt_dontuse.h SMBFS opt_dontuse.h TMPFS opt_dontuse.h UDF opt_dontuse.h UNIONFS opt_dontuse.h ZFS opt_dontuse.h # Pseudofs debugging PSEUDOFS_TRACE opt_pseudofs.h # In-kernel GSS-API KGSSAPI opt_kgssapi.h KGSSAPI_DEBUG opt_kgssapi.h # These static filesystems have one slightly bogus static dependency in # sys/i386/i386/autoconf.c. If any of these filesystems are # statically compiled into the kernel, code for mounting them as root # filesystems will be enabled - but look below. # NFSCL - client # NFSD - server NFSCL opt_nfs.h NFSD opt_nfs.h # filesystems and libiconv bridge CD9660_ICONV opt_dontuse.h MSDOSFS_ICONV opt_dontuse.h UDF_ICONV opt_dontuse.h # If you are following the conditions in the copyright, # you can enable soft-updates which will speed up a lot of thigs # and make the system safer from crashes at the same time. # otherwise a STUB module will be compiled in. SOFTUPDATES opt_ffs.h # On small, embedded systems, it can be useful to turn off support for # snapshots. It saves about 30-40k for a feature that would be lightly # used, if it is used at all. NO_FFS_SNAPSHOT opt_ffs.h # Enabling this option turns on support for Access Control Lists in UFS, # which can be used to support high security configurations. Depends on # UFS_EXTATTR. UFS_ACL opt_ufs.h # Enabling this option turns on support for extended attributes in UFS-based # filesystems, which can be used to support high security configurations # as well as new filesystem features. UFS_EXTATTR opt_ufs.h UFS_EXTATTR_AUTOSTART opt_ufs.h # Enable fast hash lookups for large directories on UFS-based filesystems. UFS_DIRHASH opt_ufs.h # Enable gjournal-based UFS journal. UFS_GJOURNAL opt_ufs.h # The below sentence is not in English, and neither is this one. # We plan to remove the static dependences above, with a # _ROOT option to control if it usable as root. This list # allows these options to be present in config files already (though # they won't make any difference yet). NFS_ROOT opt_nfsroot.h # SMB/CIFS requester NETSMB opt_netsmb.h # Enable debugnet(4) networking support. DEBUGNET opt_global.h # Enable netdump(4) client support. NETDUMP opt_global.h # Enable netgdb(4) support. NETGDB opt_global.h # Options used only in subr_param.c. HZ opt_param.h MAXFILES opt_param.h NBUF opt_param.h NSFBUFS opt_param.h VM_BCACHE_SIZE_MAX opt_param.h VM_SWZONE_SIZE_MAX opt_param.h MAXUSERS DFLDSIZ opt_param.h MAXDSIZ opt_param.h MAXSSIZ opt_param.h # Generic SCSI options. CAM_MAX_HIGHPOWER opt_cam.h CAMDEBUG opt_cam.h CAM_DEBUG_COMPILE opt_cam.h CAM_DEBUG_DELAY opt_cam.h CAM_DEBUG_BUS opt_cam.h CAM_DEBUG_TARGET opt_cam.h CAM_DEBUG_LUN opt_cam.h CAM_DEBUG_FLAGS opt_cam.h CAM_BOOT_DELAY opt_cam.h CAM_IOSCHED_DYNAMIC opt_cam.h CAM_IO_STATS opt_cam.h CAM_TEST_FAILURE opt_cam.h SCSI_DELAY opt_scsi.h SCSI_NO_SENSE_STRINGS opt_scsi.h SCSI_NO_OP_STRINGS opt_scsi.h # Options used only in cam/ata/ata_da.c ATA_STATIC_ID opt_ada.h # Options used only in cam/scsi/scsi_cd.c CHANGER_MIN_BUSY_SECONDS opt_cd.h CHANGER_MAX_BUSY_SECONDS opt_cd.h # Options used only in cam/scsi/scsi_da.c DA_TRACK_REFS opt_da.h # Options used only in cam/scsi/scsi_sa.c. SA_IO_TIMEOUT opt_sa.h SA_SPACE_TIMEOUT opt_sa.h SA_REWIND_TIMEOUT opt_sa.h SA_ERASE_TIMEOUT opt_sa.h SA_1FM_AT_EOD opt_sa.h # Options used only in cam/scsi/scsi_pt.c SCSI_PT_DEFAULT_TIMEOUT opt_pt.h # Options used only in cam/scsi/scsi_ses.c SES_ENABLE_PASSTHROUGH opt_ses.h # Options used in dev/sym/ (Symbios SCSI driver). SYM_SETUP_SCSI_DIFF opt_sym.h #-HVD support for 825a, 875, 885 # disabled:0 (default), enabled:1 SYM_SETUP_PCI_PARITY opt_sym.h #-PCI parity checking # disabled:0, enabled:1 (default) SYM_SETUP_MAX_LUN opt_sym.h #-Number of LUNs supported # default:8, range:[1..64] # Options used only in dev/isp/* ISP_TARGET_MODE opt_isp.h ISP_FW_CRASH_DUMP opt_isp.h ISP_DEFAULT_ROLES opt_isp.h ISP_INTERNAL_TARGET opt_isp.h ISP_FCTAPE_OFF opt_isp.h # Options used only in dev/iscsi ISCSI_INITIATOR_DEBUG opt_iscsi_initiator.h # Net stuff. ACCEPT_FILTER_DATA ACCEPT_FILTER_DNS ACCEPT_FILTER_HTTP ALTQ opt_global.h ALTQ_CBQ opt_altq.h ALTQ_CDNR opt_altq.h ALTQ_CODEL opt_altq.h ALTQ_DEBUG opt_altq.h ALTQ_HFSC opt_altq.h ALTQ_FAIRQ opt_altq.h ALTQ_NOPCC opt_altq.h ALTQ_PRIQ opt_altq.h ALTQ_RED opt_altq.h ALTQ_RIO opt_altq.h BOOTP opt_bootp.h BOOTP_BLOCKSIZE opt_bootp.h BOOTP_COMPAT opt_bootp.h BOOTP_NFSROOT opt_bootp.h BOOTP_NFSV3 opt_bootp.h BOOTP_WIRED_TO opt_bootp.h DEVICE_POLLING DUMMYNET opt_ipdn.h RATELIMIT opt_ratelimit.h RATELIMIT_DEBUG opt_ratelimit.h INET opt_inet.h INET6 opt_inet6.h STATS opt_global.h IPDIVERT IPFILTER opt_ipfilter.h IPFILTER_DEFAULT_BLOCK opt_ipfilter.h IPFILTER_LOG opt_ipfilter.h IPFILTER_LOOKUP opt_ipfilter.h IPFIREWALL opt_ipfw.h IPFIREWALL_DEFAULT_TO_ACCEPT opt_ipfw.h IPFIREWALL_NAT opt_ipfw.h IPFIREWALL_NAT64 opt_ipfw.h IPFIREWALL_NPTV6 opt_ipfw.h IPFIREWALL_VERBOSE opt_ipfw.h IPFIREWALL_VERBOSE_LIMIT opt_ipfw.h IPFIREWALL_PMOD opt_ipfw.h IPSEC opt_ipsec.h IPSEC_DEBUG opt_ipsec.h IPSEC_SUPPORT opt_ipsec.h IPSTEALTH KERN_TLS KRPC LIBALIAS LIBMCHAIN MBUF_PROFILING MBUF_STRESS_TEST MROUTING opt_mrouting.h NFSLOCKD PF_DEFAULT_TO_DROP opt_pf.h ROUTE_MPATH opt_route.h ROUTETABLES opt_route.h FIB_ALGO opt_route.h RSS opt_rss.h SLIP_IFF_OPTS opt_slip.h TCPDEBUG TCPPCAP opt_global.h SIFTR TCP_BLACKBOX opt_global.h -TCP_HHOOK opt_inet.h +TCP_HHOOK opt_global.h TCP_OFFLOAD opt_inet.h # Enable code to dispatch TCP offloading TCP_RFC7413 opt_inet.h TCP_RFC7413_MAX_KEYS opt_inet.h TCP_RFC7413_MAX_PSKS opt_inet.h TCP_SIGNATURE opt_ipsec.h VLAN_ARRAY opt_vlan.h XDR XBONEHACK # # SCTP # SCTP opt_sctp.h SCTP_SUPPORT opt_sctp.h SCTP_DEBUG opt_sctp.h # Enable debug printfs SCTP_LOCK_LOGGING opt_sctp.h # Log to KTR lock activity SCTP_MBUF_LOGGING opt_sctp.h # Log to KTR general mbuf aloc/free SCTP_MBCNT_LOGGING opt_sctp.h # Log to KTR mbcnt activity SCTP_PACKET_LOGGING opt_sctp.h # Log to a packet buffer last N packets SCTP_LTRACE_CHUNKS opt_sctp.h # Log to KTR chunks processed SCTP_LTRACE_ERRORS opt_sctp.h # Log to KTR error returns. SCTP_USE_PERCPU_STAT opt_sctp.h # Use per cpu stats. SCTP_MCORE_INPUT opt_sctp.h # Have multiple input threads for input mbufs SCTP_LOCAL_TRACE_BUF opt_sctp.h # Use tracebuffer exported via sysctl SCTP_DETAILED_STR_STATS opt_sctp.h # Use per PR-SCTP policy stream stats # # # # Netgraph(4). Use option NETGRAPH to enable the base netgraph code. # Each netgraph node type can be either be compiled into the kernel # or loaded dynamically. To get the former, include the corresponding # option below. Each type has its own man page, e.g. ng_async(4). NETGRAPH NETGRAPH_DEBUG opt_netgraph.h NETGRAPH_ASYNC opt_netgraph.h NETGRAPH_ATMLLC opt_netgraph.h NETGRAPH_ATM_ATMPIF opt_netgraph.h NETGRAPH_BLUETOOTH opt_netgraph.h NETGRAPH_BLUETOOTH_BT3C opt_netgraph.h NETGRAPH_BLUETOOTH_H4 opt_netgraph.h NETGRAPH_BLUETOOTH_HCI opt_netgraph.h NETGRAPH_BLUETOOTH_L2CAP opt_netgraph.h NETGRAPH_BLUETOOTH_SOCKET opt_netgraph.h NETGRAPH_BLUETOOTH_UBT opt_netgraph.h NETGRAPH_BLUETOOTH_UBTBCMFW opt_netgraph.h NETGRAPH_BPF opt_netgraph.h NETGRAPH_BRIDGE opt_netgraph.h NETGRAPH_CAR opt_netgraph.h NETGRAPH_CHECKSUM opt_netgraph.h NETGRAPH_CISCO opt_netgraph.h NETGRAPH_DEFLATE opt_netgraph.h NETGRAPH_DEVICE opt_netgraph.h NETGRAPH_ECHO opt_netgraph.h NETGRAPH_EIFACE opt_netgraph.h NETGRAPH_ETHER opt_netgraph.h NETGRAPH_ETHER_ECHO opt_netgraph.h NETGRAPH_FEC opt_netgraph.h NETGRAPH_FRAME_RELAY opt_netgraph.h NETGRAPH_GIF opt_netgraph.h NETGRAPH_GIF_DEMUX opt_netgraph.h NETGRAPH_HOLE opt_netgraph.h NETGRAPH_IFACE opt_netgraph.h NETGRAPH_IP_INPUT opt_netgraph.h NETGRAPH_IPFW opt_netgraph.h NETGRAPH_KSOCKET opt_netgraph.h NETGRAPH_L2TP opt_netgraph.h NETGRAPH_LMI opt_netgraph.h NETGRAPH_MPPC_COMPRESSION opt_netgraph.h NETGRAPH_MPPC_ENCRYPTION opt_netgraph.h NETGRAPH_NAT opt_netgraph.h NETGRAPH_NETFLOW opt_netgraph.h NETGRAPH_ONE2MANY opt_netgraph.h NETGRAPH_PATCH opt_netgraph.h NETGRAPH_PIPE opt_netgraph.h NETGRAPH_PPP opt_netgraph.h NETGRAPH_PPPOE opt_netgraph.h NETGRAPH_PPTPGRE opt_netgraph.h NETGRAPH_PRED1 opt_netgraph.h NETGRAPH_RFC1490 opt_netgraph.h NETGRAPH_SOCKET opt_netgraph.h NETGRAPH_SPLIT opt_netgraph.h NETGRAPH_SPPP opt_netgraph.h NETGRAPH_TAG opt_netgraph.h NETGRAPH_TCPMSS opt_netgraph.h NETGRAPH_TEE opt_netgraph.h NETGRAPH_TTY opt_netgraph.h NETGRAPH_UI opt_netgraph.h NETGRAPH_VJC opt_netgraph.h NETGRAPH_VLAN opt_netgraph.h # NgATM options NGATM_ATM opt_netgraph.h NGATM_ATMBASE opt_netgraph.h NGATM_SSCOP opt_netgraph.h NGATM_SSCFU opt_netgraph.h NGATM_UNI opt_netgraph.h NGATM_CCATM opt_netgraph.h # DRM options DRM_DEBUG opt_drm.h TI_SF_BUF_JUMBO opt_ti.h TI_JUMBO_HDRSPLIT opt_ti.h # Misc debug flags. Most of these should probably be replaced with # 'DEBUG', and then let people recompile just the interesting modules # with 'make CC="cc -DDEBUG"'. DEBUG_1284 opt_ppb_1284.h LPT_DEBUG opt_lpt.h PLIP_DEBUG opt_plip.h LOCKF_DEBUG opt_debug_lockf.h SI_DEBUG opt_debug_si.h IFMEDIA_DEBUG opt_ifmedia.h # Fb options FB_DEBUG opt_fb.h # ppbus related options PERIPH_1284 opt_ppb_1284.h DONTPROBE_1284 opt_ppb_1284.h # smbus related options ENABLE_ALART opt_intpm.h # These cause changes all over the kernel BLKDEV_IOSIZE opt_global.h BURN_BRIDGES opt_global.h DEBUG opt_global.h DEBUG_LOCKS opt_global.h DEBUG_VFS_LOCKS opt_global.h DFLTPHYS opt_global.h DIAGNOSTIC opt_global.h INVARIANT_SUPPORT opt_global.h INVARIANTS opt_global.h KASSERT_PANIC_OPTIONAL opt_global.h MAXCPU opt_global.h MAXMEMDOM opt_global.h MAXPHYS opt_maxphys.h MCLSHIFT opt_global.h MUTEX_NOINLINE opt_global.h LOCK_PROFILING opt_global.h MSIZE opt_global.h REGRESSION opt_global.h RWLOCK_NOINLINE opt_global.h SX_NOINLINE opt_global.h VFS_BIO_DEBUG opt_global.h # These are VM related options VM_KMEM_SIZE opt_vm.h VM_KMEM_SIZE_SCALE opt_vm.h VM_KMEM_SIZE_MAX opt_vm.h VM_NRESERVLEVEL opt_vm.h VM_LEVEL_0_ORDER opt_vm.h NO_SWAPPING opt_vm.h MALLOC_MAKE_FAILURES opt_vm.h MALLOC_PROFILE opt_vm.h MALLOC_DEBUG_MAXZONES opt_vm.h # The MemGuard replacement allocator used for tamper-after-free detection DEBUG_MEMGUARD opt_vm.h # The RedZone malloc(9) protection DEBUG_REDZONE opt_vm.h # Standard SMP options EARLY_AP_STARTUP opt_global.h SMP opt_global.h NUMA opt_global.h # Size of the kernel message buffer MSGBUF_SIZE opt_msgbuf.h # NFS options NFS_MINATTRTIMO opt_nfs.h NFS_MAXATTRTIMO opt_nfs.h NFS_MINDIRATTRTIMO opt_nfs.h NFS_MAXDIRATTRTIMO opt_nfs.h NFS_DEBUG opt_nfs.h # TMPFS options TMPFS_PAGES_MINRESERVED opt_tmpfs.h # Options for uart(4) UART_PPS_ON_CTS opt_uart.h UART_POLL_FREQ opt_uart.h UART_DEV_TOLERANCE_PCT opt_uart.h # options for bus/device framework BUS_DEBUG opt_bus.h # options for USB support USB_DEBUG opt_usb.h USB_HOST_ALIGN opt_usb.h USB_REQ_DEBUG opt_usb.h USB_TEMPLATE opt_usb.h USB_VERBOSE opt_usb.h USB_DMA_SINGLE_ALLOC opt_usb.h USB_EHCI_BIG_ENDIAN_DESC opt_usb.h U3G_DEBUG opt_u3g.h UKBD_DFLT_KEYMAP opt_ukbd.h UPLCOM_INTR_INTERVAL opt_uplcom.h UVSCOM_DEFAULT_OPKTSIZE opt_uvscom.h UVSCOM_INTR_INTERVAL opt_uvscom.h # options for the Realtek rtwn driver RTWN_DEBUG opt_rtwn.h RTWN_WITHOUT_UCODE opt_rtwn.h # Embedded system options INIT_PATH ROOTDEVNAME FDC_DEBUG opt_fdc.h PCFCLOCK_VERBOSE opt_pcfclock.h PCFCLOCK_MAX_RETRIES opt_pcfclock.h KTR opt_global.h KTR_ALQ opt_ktr.h KTR_MASK opt_ktr.h KTR_CPUMASK opt_ktr.h KTR_COMPILE opt_global.h KTR_BOOT_ENTRIES opt_global.h KTR_ENTRIES opt_global.h KTR_VERBOSE opt_ktr.h WITNESS opt_global.h WITNESS_KDB opt_witness.h WITNESS_NO_VNODE opt_witness.h WITNESS_SKIPSPIN opt_witness.h WITNESS_COUNT opt_witness.h OPENSOLARIS_WITNESS opt_global.h EPOCH_TRACE opt_global.h # options for ACPI support ACPI_DEBUG opt_acpi.h ACPI_MAX_TASKS opt_acpi.h ACPI_MAX_THREADS opt_acpi.h DEV_ACPI opt_acpi.h ACPI_EARLY_EPYC_WAR opt_acpi.h # options for IOMMU support IOMMU opt_iommu.h # ISA support DEV_ISA opt_isa.h ISAPNP opt_dontuse.h # various 'device presence' options. DEV_BPF opt_bpf.h DEV_CARP opt_carp.h DEV_NETMAP opt_global.h DEV_PCI opt_pci.h DEV_PF opt_pf.h DEV_PFLOG opt_pf.h DEV_PFSYNC opt_pf.h DEV_SPLASH opt_splash.h DEV_VLAN opt_vlan.h # bce driver BCE_DEBUG opt_bce.h BCE_NVRAM_WRITE_SUPPORT opt_bce.h SOCKBUF_DEBUG opt_global.h # options for hifn driver HIFN_DEBUG opt_hifn.h HIFN_RNDTEST opt_hifn.h # options for safenet driver SAFE_DEBUG opt_safe.h SAFE_NO_RNG opt_safe.h SAFE_RNDTEST opt_safe.h # syscons/vt options MAXCONS opt_syscons.h SC_ALT_MOUSE_IMAGE opt_syscons.h SC_CUT_SPACES2TABS opt_syscons.h SC_CUT_SEPCHARS opt_syscons.h SC_DEBUG_LEVEL opt_syscons.h SC_DFLT_FONT opt_syscons.h SC_DFLT_TERM opt_syscons.h SC_DISABLE_KDBKEY opt_syscons.h SC_DISABLE_REBOOT opt_syscons.h SC_HISTORY_SIZE opt_syscons.h SC_KERNEL_CONS_ATTR opt_syscons.h SC_KERNEL_CONS_ATTRS opt_syscons.h SC_KERNEL_CONS_REV_ATTR opt_syscons.h SC_MOUSE_CHAR opt_syscons.h SC_NO_CUTPASTE opt_syscons.h SC_NO_FONT_LOADING opt_syscons.h SC_NO_HISTORY opt_syscons.h SC_NO_MODE_CHANGE opt_syscons.h SC_NO_SUSPEND_VTYSWITCH opt_syscons.h SC_NO_SYSMOUSE opt_syscons.h SC_NO_TERM_DUMB opt_syscons.h SC_NO_TERM_SC opt_syscons.h SC_NO_TERM_TEKEN opt_syscons.h SC_NORM_ATTR opt_syscons.h SC_NORM_REV_ATTR opt_syscons.h SC_PIXEL_MODE opt_syscons.h SC_RENDER_DEBUG opt_syscons.h SC_TWOBUTTON_MOUSE opt_syscons.h VT_ALT_TO_ESC_HACK opt_syscons.h VT_FB_MAX_WIDTH opt_syscons.h VT_FB_MAX_HEIGHT opt_syscons.h VT_MAXWINDOWS opt_syscons.h VT_TWOBUTTON_MOUSE opt_syscons.h DEV_SC opt_syscons.h DEV_VT opt_syscons.h # teken terminal emulator options TEKEN_CONS25 opt_teken.h TEKEN_UTF8 opt_teken.h TERMINAL_KERN_ATTR opt_teken.h TERMINAL_NORM_ATTR opt_teken.h # options for printf PRINTF_BUFR_SIZE opt_printf.h BOOT_TAG opt_printf.h BOOT_TAG_SZ opt_printf.h # kbd options KBD_DISABLE_KEYMAP_LOAD opt_kbd.h KBD_INSTALL_CDEV opt_kbd.h KBD_MAXRETRY opt_kbd.h KBD_MAXWAIT opt_kbd.h KBD_RESETDELAY opt_kbd.h KBDIO_DEBUG opt_kbd.h KBDMUX_DFLT_KEYMAP opt_kbdmux.h # options for the Atheros driver ATH_DEBUG opt_ath.h ATH_TXBUF opt_ath.h ATH_RXBUF opt_ath.h ATH_DIAGAPI opt_ath.h ATH_TX99_DIAG opt_ath.h ATH_ENABLE_11N opt_ath.h ATH_ENABLE_DFS opt_ath.h ATH_EEPROM_FIRMWARE opt_ath.h ATH_ENABLE_RADIOTAP_VENDOR_EXT opt_ath.h ATH_DEBUG_ALQ opt_ath.h ATH_KTR_INTR_DEBUG opt_ath.h # options for the Atheros hal # XXX For now, this breaks non-AR9130 chipsets, so only use it # XXX when actually targeting AR9130. AH_SUPPORT_AR9130 opt_ah.h # This is required for AR933x SoC support AH_SUPPORT_AR9330 opt_ah.h AH_SUPPORT_AR9340 opt_ah.h AH_SUPPORT_QCA9530 opt_ah.h AH_SUPPORT_QCA9550 opt_ah.h AH_DEBUG opt_ah.h AH_ASSERT opt_ah.h AH_DEBUG_ALQ opt_ah.h AH_REGOPS_FUNC opt_ah.h AH_WRITE_REGDOMAIN opt_ah.h AH_DEBUG_COUNTRY opt_ah.h AH_WRITE_EEPROM opt_ah.h AH_PRIVATE_DIAG opt_ah.h AH_NEED_DESC_SWAP opt_ah.h AH_USE_INIPDGAIN opt_ah.h AH_MAXCHAN opt_ah.h AH_RXCFG_SDMAMW_4BYTES opt_ah.h AH_INTERRUPT_DEBUGGING opt_ah.h # AR5416 and later interrupt mitigation # XXX do not use this for AR9130 AH_AR5416_INTERRUPT_MITIGATION opt_ah.h # options for the Altera mSGDMA driver (altera_msgdma) ALTERA_MSGDMA_DESC_STD opt_altera_msgdma.h ALTERA_MSGDMA_DESC_EXT opt_altera_msgdma.h ALTERA_MSGDMA_DESC_PF_STD opt_altera_msgdma.h ALTERA_MSGDMA_DESC_PF_EXT opt_altera_msgdma.h # options for the Broadcom BCM43xx driver (bwi) BWI_DEBUG opt_bwi.h BWI_DEBUG_VERBOSE opt_bwi.h # options for the Brodacom BCM43xx driver (bwn) BWN_DEBUG opt_bwn.h BWN_GPL_PHY opt_bwn.h BWN_USE_SIBA opt_bwn.h # Options for the SIBA driver SIBA_DEBUG opt_siba.h # options for the Marvell 8335 wireless driver MALO_DEBUG opt_malo.h MALO_TXBUF opt_malo.h MALO_RXBUF opt_malo.h # options for the Marvell wireless driver MWL_DEBUG opt_mwl.h MWL_TXBUF opt_mwl.h MWL_RXBUF opt_mwl.h MWL_DIAGAPI opt_mwl.h MWL_AGGR_SIZE opt_mwl.h MWL_TX_NODROP opt_mwl.h # Options for the Marvell NETA driver MVNETA_MULTIQUEUE opt_mvneta.h MVNETA_KTR opt_mvneta.h # Options for the Intel 802.11ac wireless driver IWM_DEBUG opt_iwm.h # Options for the Intel 802.11n wireless driver IWN_DEBUG opt_iwn.h # Options for the Intel 3945ABG wireless driver WPI_DEBUG opt_wpi.h # dcons options DCONS_BUF_SIZE opt_dcons.h DCONS_POLL_HZ opt_dcons.h DCONS_FORCE_CONSOLE opt_dcons.h DCONS_FORCE_GDB opt_dcons.h # HWPMC options HWPMC_DEBUG opt_global.h HWPMC_HOOKS HWPMC_MIPS_BACKTRACE opt_hwpmc_hooks.h # 802.11 support layer IEEE80211_DEBUG opt_wlan.h IEEE80211_DEBUG_REFCNT opt_wlan.h IEEE80211_SUPPORT_MESH opt_wlan.h IEEE80211_SUPPORT_SUPERG opt_wlan.h IEEE80211_SUPPORT_TDMA opt_wlan.h IEEE80211_ALQ opt_wlan.h IEEE80211_DFS_DEBUG opt_wlan.h # 802.11 TDMA support TDMA_SLOTLEN_DEFAULT opt_tdma.h TDMA_SLOTCNT_DEFAULT opt_tdma.h TDMA_BINTVAL_DEFAULT opt_tdma.h TDMA_TXRATE_11B_DEFAULT opt_tdma.h TDMA_TXRATE_11G_DEFAULT opt_tdma.h TDMA_TXRATE_11A_DEFAULT opt_tdma.h TDMA_TXRATE_TURBO_DEFAULT opt_tdma.h TDMA_TXRATE_HALF_DEFAULT opt_tdma.h TDMA_TXRATE_QUARTER_DEFAULT opt_tdma.h TDMA_TXRATE_11NA_DEFAULT opt_tdma.h TDMA_TXRATE_11NG_DEFAULT opt_tdma.h # VideoMode PICKMODE_DEBUG opt_videomode.h # Network stack virtualization options VIMAGE opt_global.h VNET_DEBUG opt_global.h # Common Flash Interface (CFI) options CFI_SUPPORT_STRATAFLASH opt_cfi.h CFI_ARMEDANDDANGEROUS opt_cfi.h CFI_HARDWAREBYTESWAP opt_cfi.h # Sound options SND_DEBUG opt_snd.h SND_DIAGNOSTIC opt_snd.h SND_FEEDER_MULTIFORMAT opt_snd.h SND_FEEDER_FULL_MULTIFORMAT opt_snd.h SND_FEEDER_RATE_HP opt_snd.h SND_PCM_64 opt_snd.h SND_OLDSTEREO opt_snd.h X86BIOS # Flattened device tree options FDT opt_platform.h FDT_DTB_STATIC opt_platform.h # OFED Infiniband stack OFED opt_ofed.h OFED_DEBUG_INIT opt_ofed.h SDP opt_ofed.h SDP_DEBUG opt_ofed.h IPOIB opt_ofed.h IPOIB_DEBUG opt_ofed.h IPOIB_CM opt_ofed.h # Resource Accounting RACCT opt_global.h RACCT_DEFAULT_TO_DISABLED opt_global.h # Resource Limits RCTL opt_global.h # Random number generator(s) # Alternative RNG algorithm. RANDOM_FENESTRASX opt_global.h # With this, no entropy processor is loaded, but the entropy # harvesting infrastructure is present. This means an entropy # processor may be loaded as a module. RANDOM_LOADABLE opt_global.h # This turns on high-rate and potentially expensive harvesting in # the uma slab allocator. RANDOM_ENABLE_UMA opt_global.h RANDOM_ENABLE_ETHER opt_global.h # This options turns TPM into entropy source. TPM_HARVEST opt_tpm.h # BHND(4) driver BHND_LOGLEVEL opt_global.h # GPIO and child devices GPIO_SPI_DEBUG opt_gpio.h # SPI devices SPIGEN_LEGACY_CDEVNAME opt_spi.h # etherswitch(4) driver RTL8366_SOFT_RESET opt_etherswitch.h # evdev protocol support EVDEV_SUPPORT opt_evdev.h EVDEV_DEBUG opt_evdev.h UINPUT_DEBUG opt_evdev.h # Hyper-V network driver HN_DEBUG opt_hn.h # CAM-based MMC stack MMCCAM # Encrypted kernel crash dumps EKCD opt_ekcd.h # NVME options NVME_USE_NVD opt_nvme.h NVME_2X_RESET opt_nvme.h # amdsbwd options AMDSBWD_DEBUG opt_amdsbwd.h # gcov support GCOV opt_global.h LINDEBUGFS # options for HID support HID_DEBUG opt_hid.h IICHID_DEBUG opt_hid.h IICHID_SAMPLING opt_hid.h HKBD_DFLT_KEYMAP opt_hkbd.h HIDRAW_MAKE_UHID_ALIAS opt_hid.h # kenv options # The early kernel environment (loader environment, config(8)-provided static) # is typically cleared after the dynamic environment comes up to ensure that # we're not inadvertently holding on to 'secret' values in these stale envs. # This option is insecure except in controlled environments where the static # environment's contents are known to be safe. PRESERVE_EARLY_KENV opt_global.h diff --git a/sys/dev/cxgbe/tom/t4_tom_l2t.c b/sys/dev/cxgbe/tom/t4_tom_l2t.c index 97a1e08bda58..e91d5c955758 100644 --- a/sys/dev/cxgbe/tom/t4_tom_l2t.c +++ b/sys/dev/cxgbe/tom/t4_tom_l2t.c @@ -1,456 +1,457 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2012 Chelsio Communications, Inc. * 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. * * 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" #ifdef TCP_OFFLOAD #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include +#include #include #include #include "common/common.h" #include "common/t4_msg.h" #include "tom/t4_tom_l2t.h" #include "tom/t4_tom.h" #define VLAN_NONE 0xfff static inline void l2t_hold(struct l2t_data *d, struct l2t_entry *e) { if (atomic_fetchadd_int(&e->refcnt, 1) == 0) /* 0 -> 1 transition */ atomic_subtract_int(&d->nfree, 1); } static inline u_int l2_hash(struct l2t_data *d, const struct sockaddr *sa, int ifindex) { u_int hash, half = d->l2t_size / 2, start = 0; const void *key; size_t len; KASSERT(sa->sa_family == AF_INET || sa->sa_family == AF_INET6, ("%s: sa %p has unexpected sa_family %d", __func__, sa, sa->sa_family)); if (sa->sa_family == AF_INET) { const struct sockaddr_in *sin = (const void *)sa; key = &sin->sin_addr; len = sizeof(sin->sin_addr); } else { const struct sockaddr_in6 *sin6 = (const void *)sa; key = &sin6->sin6_addr; len = sizeof(sin6->sin6_addr); start = half; } hash = fnv_32_buf(key, len, FNV1_32_INIT); hash = fnv_32_buf(&ifindex, sizeof(ifindex), hash); hash %= half; return (hash + start); } static inline int l2_cmp(const struct sockaddr *sa, struct l2t_entry *e) { KASSERT(sa->sa_family == AF_INET || sa->sa_family == AF_INET6, ("%s: sa %p has unexpected sa_family %d", __func__, sa, sa->sa_family)); if (sa->sa_family == AF_INET) { const struct sockaddr_in *sin = (const void *)sa; return (e->addr[0] != sin->sin_addr.s_addr); } else { const struct sockaddr_in6 *sin6 = (const void *)sa; return (memcmp(&e->addr[0], &sin6->sin6_addr, sizeof(e->addr))); } } static inline void l2_store(const struct sockaddr *sa, struct l2t_entry *e) { KASSERT(sa->sa_family == AF_INET || sa->sa_family == AF_INET6, ("%s: sa %p has unexpected sa_family %d", __func__, sa, sa->sa_family)); if (sa->sa_family == AF_INET) { const struct sockaddr_in *sin = (const void *)sa; e->addr[0] = sin->sin_addr.s_addr; e->ipv6 = 0; } else { const struct sockaddr_in6 *sin6 = (const void *)sa; memcpy(&e->addr[0], &sin6->sin6_addr, sizeof(e->addr)); e->ipv6 = 1; } } /* * Add a WR to an L2T entry's queue of work requests awaiting resolution. * Must be called with the entry's lock held. */ static inline void arpq_enqueue(struct l2t_entry *e, struct wrqe *wr) { mtx_assert(&e->lock, MA_OWNED); STAILQ_INSERT_TAIL(&e->wr_list, wr, link); } static inline void send_pending(struct adapter *sc, struct l2t_entry *e) { struct wrqe *wr; mtx_assert(&e->lock, MA_OWNED); while ((wr = STAILQ_FIRST(&e->wr_list)) != NULL) { STAILQ_REMOVE_HEAD(&e->wr_list, link); t4_wrq_tx(sc, wr); } } static void resolution_failed(struct adapter *sc, struct l2t_entry *e) { struct tom_data *td = sc->tom_softc; mtx_assert(&e->lock, MA_OWNED); mtx_lock(&td->unsent_wr_lock); STAILQ_CONCAT(&td->unsent_wr_list, &e->wr_list); mtx_unlock(&td->unsent_wr_lock); taskqueue_enqueue(taskqueue_thread, &td->reclaim_wr_resources); } static void update_entry(struct adapter *sc, struct l2t_entry *e, uint8_t *lladdr, uint16_t vtag) { mtx_assert(&e->lock, MA_OWNED); /* * The entry may be in active use (e->refcount > 0) or not. We update * it even when it's not as this simplifies the case where we decide to * reuse the entry later. */ if (lladdr == NULL && (e->state == L2T_STATE_RESOLVING || e->state == L2T_STATE_FAILED)) { /* * Never got a valid L2 address for this one. Just mark it as * failed instead of removing it from the hash (for which we'd * need to wlock the table). */ e->state = L2T_STATE_FAILED; resolution_failed(sc, e); return; } else if (lladdr == NULL) { /* Valid or already-stale entry was deleted (or expired) */ KASSERT(e->state == L2T_STATE_VALID || e->state == L2T_STATE_STALE, ("%s: lladdr NULL, state %d", __func__, e->state)); e->state = L2T_STATE_STALE; } else { if (e->state == L2T_STATE_RESOLVING || e->state == L2T_STATE_FAILED || memcmp(e->dmac, lladdr, ETHER_ADDR_LEN)) { /* unresolved -> resolved; or dmac changed */ memcpy(e->dmac, lladdr, ETHER_ADDR_LEN); e->vlan = vtag; t4_write_l2e(e, 1); } e->state = L2T_STATE_VALID; } } static int resolve_entry(struct adapter *sc, struct l2t_entry *e) { struct tom_data *td = sc->tom_softc; struct toedev *tod = &td->tod; struct sockaddr_in sin = {0}; struct sockaddr_in6 sin6 = {0}; struct sockaddr *sa; uint8_t dmac[ETHER_HDR_LEN]; uint16_t vtag; int rc; if (e->ipv6 == 0) { sin.sin_family = AF_INET; sin.sin_len = sizeof(struct sockaddr_in); sin.sin_addr.s_addr = e->addr[0]; sa = (void *)&sin; } else { sin6.sin6_family = AF_INET6; sin6.sin6_len = sizeof(struct sockaddr_in6); memcpy(&sin6.sin6_addr, &e->addr[0], sizeof(e->addr)); sa = (void *)&sin6; } vtag = EVL_MAKETAG(VLAN_NONE, 0, 0); rc = toe_l2_resolve(tod, e->ifp, sa, dmac, &vtag); if (rc == EWOULDBLOCK) return (rc); mtx_lock(&e->lock); update_entry(sc, e, rc == 0 ? dmac : NULL, vtag); mtx_unlock(&e->lock); return (rc); } int t4_l2t_send_slow(struct adapter *sc, struct wrqe *wr, struct l2t_entry *e) { again: switch (e->state) { case L2T_STATE_STALE: /* entry is stale, kick off revalidation */ resolve_entry(sc, e); /* Fall through */ case L2T_STATE_VALID: /* fast-path, send the packet on */ t4_wrq_tx(sc, wr); return (0); case L2T_STATE_RESOLVING: case L2T_STATE_SYNC_WRITE: mtx_lock(&e->lock); if (e->state != L2T_STATE_SYNC_WRITE && e->state != L2T_STATE_RESOLVING) { /* state changed by the time we got here */ mtx_unlock(&e->lock); goto again; } arpq_enqueue(e, wr); mtx_unlock(&e->lock); if (resolve_entry(sc, e) == EWOULDBLOCK) break; mtx_lock(&e->lock); if (e->state == L2T_STATE_VALID && !STAILQ_EMPTY(&e->wr_list)) send_pending(sc, e); if (e->state == L2T_STATE_FAILED) resolution_failed(sc, e); mtx_unlock(&e->lock); break; case L2T_STATE_FAILED: return (EHOSTUNREACH); } return (0); } int do_l2t_write_rpl2(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_l2t_write_rpl *rpl = (const void *)(rss + 1); unsigned int tid = GET_TID(rpl); unsigned int idx = tid % L2T_SIZE; if (__predict_false(rpl->status != CPL_ERR_NONE)) { log(LOG_ERR, "Unexpected L2T_WRITE_RPL (%u) for entry at hw_idx %u\n", rpl->status, idx); return (EINVAL); } if (tid & F_SYNC_WR) { struct l2t_entry *e = &sc->l2t->l2tab[idx - sc->vres.l2t.start]; mtx_lock(&e->lock); if (e->state != L2T_STATE_SWITCHING) { send_pending(sc, e); e->state = L2T_STATE_VALID; } mtx_unlock(&e->lock); } return (0); } /* * The TOE wants an L2 table entry that it can use to reach the next hop over * the specified port. Produce such an entry - create one if needed. * * Note that the ifnet could be a pseudo-device like if_vlan, if_lagg, etc. on * top of the real cxgbe interface. */ struct l2t_entry * t4_l2t_get(struct port_info *pi, struct ifnet *ifp, struct sockaddr *sa) { struct l2t_entry *e; struct adapter *sc = pi->adapter; struct l2t_data *d = sc->l2t; u_int hash, smt_idx = pi->port_id; uint16_t vid, pcp, vtag; KASSERT(sa->sa_family == AF_INET || sa->sa_family == AF_INET6, ("%s: sa %p has unexpected sa_family %d", __func__, sa, sa->sa_family)); vid = VLAN_NONE; 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); hash = l2_hash(d, sa, ifp->if_index); rw_wlock(&d->lock); for (e = d->l2tab[hash].first; e; e = e->next) { if (l2_cmp(sa, e) == 0 && e->ifp == ifp && e->vlan == vtag && e->smt_idx == smt_idx) { l2t_hold(d, e); goto done; } } /* Need to allocate a new entry */ e = t4_alloc_l2e(d); if (e) { mtx_lock(&e->lock); /* avoid race with t4_l2t_free */ e->next = d->l2tab[hash].first; d->l2tab[hash].first = e; e->state = L2T_STATE_RESOLVING; l2_store(sa, e); e->ifp = ifp; e->smt_idx = smt_idx; e->hash = hash; e->lport = pi->lport; e->wrq = &sc->sge.ctrlq[pi->port_id]; e->iqid = sc->sge.ofld_rxq[pi->vi[0].first_ofld_rxq].iq.abs_id; atomic_store_rel_int(&e->refcnt, 1); e->vlan = vtag; mtx_unlock(&e->lock); } done: rw_wunlock(&d->lock); return e; } /* * Called when the host's ARP layer makes a change to some entry that is loaded * into the HW L2 table. */ void t4_l2_update(struct toedev *tod, struct ifnet *ifp, struct sockaddr *sa, uint8_t *lladdr, uint16_t vtag) { struct adapter *sc = tod->tod_softc; struct l2t_entry *e; struct l2t_data *d = sc->l2t; u_int hash; KASSERT(d != NULL, ("%s: no L2 table", __func__)); hash = l2_hash(d, sa, ifp->if_index); rw_rlock(&d->lock); for (e = d->l2tab[hash].first; e; e = e->next) { if (l2_cmp(sa, e) == 0 && e->ifp == ifp) { mtx_lock(&e->lock); if (atomic_load_acq_int(&e->refcnt)) goto found; e->state = L2T_STATE_STALE; mtx_unlock(&e->lock); break; } } rw_runlock(&d->lock); /* * This is of no interest to us. We've never had an offloaded * connection to this destination, and we aren't attempting one right * now. */ return; found: rw_runlock(&d->lock); KASSERT(e->state != L2T_STATE_UNUSED, ("%s: unused entry in the hash.", __func__)); update_entry(sc, e, lladdr, vtag); mtx_unlock(&e->lock); } #endif diff --git a/sys/netinet/cc/cc.h b/sys/netinet/cc/cc.h index 90c954faffa1..8047df0da499 100644 --- a/sys/netinet/cc/cc.h +++ b/sys/netinet/cc/cc.h @@ -1,251 +1,255 @@ /*- * Copyright (c) 2007-2008 * Swinburne University of Technology, Melbourne, Australia. * Copyright (c) 2009-2010 Lawrence Stewart * Copyright (c) 2010 The FreeBSD Foundation * All rights reserved. * * This software was developed at the Centre for Advanced Internet * Architectures, Swinburne University of Technology, by Lawrence Stewart and * James Healy, 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. * * 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. * * $FreeBSD$ */ /* * This software was first released in 2007 by James Healy and Lawrence Stewart * whilst working on the NewTCP research project at Swinburne University of * Technology's Centre for Advanced Internet Architectures, Melbourne, * Australia, which was made possible in part by a grant from the Cisco * University Research Program Fund at Community Foundation Silicon Valley. * More details are available at: * http://caia.swin.edu.au/urp/newtcp/ */ #ifndef _NETINET_CC_CC_H_ #define _NETINET_CC_CC_H_ #ifdef _KERNEL MALLOC_DECLARE(M_CC_MEM); /* Global CC vars. */ extern STAILQ_HEAD(cc_head, cc_algo) cc_list; extern const int tcprexmtthresh; /* Per-netstack bits. */ VNET_DECLARE(struct cc_algo *, default_cc_ptr); #define V_default_cc_ptr VNET(default_cc_ptr) VNET_DECLARE(int, cc_do_abe); #define V_cc_do_abe VNET(cc_do_abe) VNET_DECLARE(int, cc_abe_frlossreduce); #define V_cc_abe_frlossreduce VNET(cc_abe_frlossreduce) /* Define the new net.inet.tcp.cc sysctl tree. */ +#ifdef _SYS_SYSCTL_H_ SYSCTL_DECL(_net_inet_tcp_cc); +#endif /* For CC modules that use hystart++ */ extern uint32_t hystart_lowcwnd; extern uint32_t hystart_minrtt_thresh; extern uint32_t hystart_maxrtt_thresh; extern uint32_t hystart_n_rttsamples; extern uint32_t hystart_css_growth_div; extern uint32_t hystart_css_rounds; extern uint32_t hystart_bblogs; /* CC housekeeping functions. */ int cc_register_algo(struct cc_algo *add_cc); int cc_deregister_algo(struct cc_algo *remove_cc); +#endif /* _KERNEL */ +#if defined(_KERNEL) || defined(_WANT_TCPCB) /* * Wrapper around transport structs that contain same-named congestion * control variables. Allows algos to be shared amongst multiple CC aware * transprots. */ struct cc_var { void *cc_data; /* Per-connection private CC algorithm data. */ int bytes_this_ack; /* # bytes acked by the current ACK. */ tcp_seq curack; /* Most recent ACK. */ uint32_t flags; /* Flags for cc_var (see below) */ int type; /* Indicates which ptr is valid in ccvc. */ union ccv_container { struct tcpcb *tcp; struct sctp_nets *sctp; } ccvc; uint16_t nsegs; /* # segments coalesced into current chain. */ uint8_t labc; /* Dont use system abc use passed in */ }; /* cc_var flags. */ #define CCF_ABC_SENTAWND 0x0001 /* ABC counted cwnd worth of bytes? */ #define CCF_CWND_LIMITED 0x0002 /* Are we currently cwnd limited? */ #define CCF_USE_LOCAL_ABC 0x0004 /* Dont use the system l_abc val */ #define CCF_ACKNOW 0x0008 /* Will this ack be sent now? */ #define CCF_IPHDR_CE 0x0010 /* Does this packet set CE bit? */ #define CCF_TCPHDR_CWR 0x0020 /* Does this packet set CWR bit? */ #define CCF_MAX_CWND 0x0040 /* Have we reached maximum cwnd? */ #define CCF_CHG_MAX_CWND 0x0080 /* CUBIC max_cwnd changed, for K */ #define CCF_USR_IWND 0x0100 /* User specified initial window */ #define CCF_USR_IWND_INIT_NSEG 0x0200 /* Convert segs to bytes on conn init */ #define CCF_HYSTART_ALLOWED 0x0400 /* If the CC supports it Hystart is allowed */ #define CCF_HYSTART_CAN_SH_CWND 0x0800 /* Can hystart when going CSS -> CA slam the cwnd */ #define CCF_HYSTART_CONS_SSTH 0x1000 /* Should hystart use the more conservative ssthresh */ /* ACK types passed to the ack_received() hook. */ #define CC_ACK 0x0001 /* Regular in sequence ACK. */ #define CC_DUPACK 0x0002 /* Duplicate ACK. */ #define CC_PARTIALACK 0x0004 /* Not yet. */ #define CC_SACK 0x0008 /* Not yet. */ -#endif /* _KERNEL */ +#endif /* defined(_KERNEL) || defined(_WANT_TCPCB) */ /* * Congestion signal types passed to the cong_signal() hook. The highest order 8 * bits (0x01000000 - 0x80000000) are reserved for CC algos to declare their own * congestion signal types. */ #define CC_ECN 0x00000001 /* ECN marked packet received. */ #define CC_RTO 0x00000002 /* RTO fired. */ #define CC_RTO_ERR 0x00000004 /* RTO fired in error. */ #define CC_NDUPACK 0x00000008 /* Threshold of dupack's reached. */ #define CC_SIGPRIVMASK 0xFF000000 /* Mask to check if sig is private. */ #ifdef _KERNEL /* * Structure to hold data and function pointers that together represent a * congestion control algorithm. */ struct cc_algo { char name[TCP_CA_NAME_MAX]; /* Init global module state on kldload. */ int (*mod_init)(void); /* Cleanup global module state on kldunload. */ int (*mod_destroy)(void); /* Return the size of the void pointer the CC needs for state */ size_t (*cc_data_sz)(void); /* * Init CC state for a new control block. The CC * module may be passed a NULL ptr indicating that * it must allocate the memory. If it is passed a * non-null pointer it is pre-allocated memory by * the caller and the cb_init is expected to use that memory. * It is not expected to fail if memory is passed in and * all currently defined modules do not. */ int (*cb_init)(struct cc_var *ccv, void *ptr); /* Cleanup CC state for a terminating control block. */ void (*cb_destroy)(struct cc_var *ccv); /* Init variables for a newly established connection. */ void (*conn_init)(struct cc_var *ccv); /* Called on receipt of an ack. */ void (*ack_received)(struct cc_var *ccv, uint16_t type); /* Called on detection of a congestion signal. */ void (*cong_signal)(struct cc_var *ccv, uint32_t type); /* Called after exiting congestion recovery. */ void (*post_recovery)(struct cc_var *ccv); /* Called when data transfer resumes after an idle period. */ void (*after_idle)(struct cc_var *ccv); /* Called for an additional ECN processing apart from RFC3168. */ void (*ecnpkt_handler)(struct cc_var *ccv); /* Called when a new "round" begins, if the transport is tracking rounds. */ void (*newround)(struct cc_var *ccv, uint32_t round_cnt); /* * Called when a RTT sample is made (fas = flight at send, if you dont have it * send the cwnd in). */ void (*rttsample)(struct cc_var *ccv, uint32_t usec_rtt, uint32_t rxtcnt, uint32_t fas); /* Called for {get|set}sockopt() on a TCP socket with TCP_CCALGOOPT. */ int (*ctl_output)(struct cc_var *, struct sockopt *, void *); STAILQ_ENTRY (cc_algo) entries; u_int cc_refcount; uint8_t flags; }; #define CC_MODULE_BEING_REMOVED 0x01 /* The module is being removed */ /* Macro to obtain the CC algo's struct ptr. */ -#define CC_ALGO(tp) ((tp)->cc_algo) +#define CC_ALGO(tp) ((tp)->t_cc) /* Macro to obtain the CC algo's data ptr. */ -#define CC_DATA(tp) ((tp)->ccv->cc_data) +#define CC_DATA(tp) ((tp)->t_ccv.cc_data) /* Macro to obtain the system default CC algo's struct ptr. */ #define CC_DEFAULT_ALGO() V_default_cc_ptr extern struct rwlock cc_list_lock; #define CC_LIST_LOCK_INIT() rw_init(&cc_list_lock, "cc_list") #define CC_LIST_LOCK_DESTROY() rw_destroy(&cc_list_lock) #define CC_LIST_RLOCK() rw_rlock(&cc_list_lock) #define CC_LIST_RUNLOCK() rw_runlock(&cc_list_lock) #define CC_LIST_WLOCK() rw_wlock(&cc_list_lock) #define CC_LIST_WUNLOCK() rw_wunlock(&cc_list_lock) #define CC_LIST_LOCK_ASSERT() rw_assert(&cc_list_lock, RA_LOCKED) #define CC_ALGOOPT_LIMIT 2048 /* * These routines give NewReno behavior to the caller * they require no state and can be used by any other CC * module that wishes to use NewReno type behaviour (along * with anything else they may add on, pre or post call). */ void newreno_cc_post_recovery(struct cc_var *); void newreno_cc_after_idle(struct cc_var *); void newreno_cc_cong_signal(struct cc_var *, uint32_t ); void newreno_cc_ack_received(struct cc_var *, uint16_t); /* Called to temporarily keep an algo from going away during change */ void cc_refer(struct cc_algo *algo); /* Called to release the temporary hold */ void cc_release(struct cc_algo *algo); /* Called to attach a CC algorithm to a tcpcb */ void cc_attach(struct tcpcb *, struct cc_algo *); /* Called to detach a CC algorithm from a tcpcb */ void cc_detach(struct tcpcb *); #endif /* _KERNEL */ #endif /* _NETINET_CC_CC_H_ */ diff --git a/sys/netinet/cc/cc_cdg.c b/sys/netinet/cc/cc_cdg.c index 1bda96d33e65..a60339bf5d23 100644 --- a/sys/netinet/cc/cc_cdg.c +++ b/sys/netinet/cc/cc_cdg.c @@ -1,728 +1,728 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2009-2013 * Swinburne University of Technology, Melbourne, Australia * All rights reserved. * * This software was developed at the Centre for Advanced Internet * Architectures, Swinburne University of Technology, by David Hayes, made * possible in part by a gift from The Cisco University Research Program Fund, * a corporate advised fund of Silicon Valley Community Foundation. Development * and testing were further assisted by a grant 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. * * 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. */ /* * CAIA Delay-Gradient (CDG) congestion control algorithm * * An implemention of the delay-gradient congestion control algorithm proposed * in the following paper: * * D. A. Hayes and G. Armitage, "Revisiting TCP Congestion Control using Delay * Gradients", in IFIP Networking, Valencia, Spain, 9-13 May 2011. * * Developed as part of the NewTCP research project at Swinburne University of * Technology's Centre for Advanced Internet Architectures, Melbourne, * Australia. More details are available at: * http://caia.swin.edu.au/urp/newtcp/ */ #include __FBSDID("$FreeBSD$"); #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 CDG_VERSION "0.1" /* Private delay-gradient induced congestion control signal. */ #define CC_CDG_DELAY 0x01000000 /* NewReno window deflation factor on loss (as a percentage). */ #define RENO_BETA 50 /* Queue states. */ #define CDG_Q_EMPTY 1 #define CDG_Q_RISING 2 #define CDG_Q_FALLING 3 #define CDG_Q_FULL 4 #define CDG_Q_UNKNOWN 9999 /* Number of bit shifts used in probexp lookup table. */ #define EXP_PREC 15 /* Largest gradient represented in probexp lookup table. */ #define MAXGRAD 5 /* * Delay Precision Enhance - number of bit shifts used for qtrend related * integer arithmetic precision. */ #define D_P_E 7 struct qdiff_sample { long qdiff; STAILQ_ENTRY(qdiff_sample) qdiff_lnk; }; struct cdg { long max_qtrend; long min_qtrend; STAILQ_HEAD(minrtts_head, qdiff_sample) qdiffmin_q; STAILQ_HEAD(maxrtts_head, qdiff_sample) qdiffmax_q; long window_incr; /* rttcount for window increase when in congestion avoidance */ long rtt_count; /* maximum measured rtt within an rtt period */ int maxrtt_in_rtt; /* maximum measured rtt within prev rtt period */ int maxrtt_in_prevrtt; /* minimum measured rtt within an rtt period */ int minrtt_in_rtt; /* minimum measured rtt within prev rtt period */ int minrtt_in_prevrtt; /* consecutive congestion episode counter */ uint32_t consec_cong_cnt; /* when tracking a new reno type loss window */ uint32_t shadow_w; /* maximum number of samples in the moving average queue */ int sample_q_size; /* number of samples in the moving average queue */ int num_samples; /* estimate of the queue state of the path */ int queue_state; }; /* * Lookup table for: * (1 - exp(-x)) << EXP_PREC, where x = [0,MAXGRAD] in 2^-7 increments * * Note: probexp[0] is set to 10 (not 0) as a safety for very low increase * gradients. */ static const int probexp[641] = { 10,255,508,759,1008,1255,1501,1744,1985,2225,2463,2698,2932,3165,3395,3624, 3850,4075,4299,4520,4740,4958,5175,5389,5602,5814,6024,6232,6438,6643,6846, 7048,7248,7447,7644,7839,8033,8226,8417,8606,8794,8981,9166,9350,9532,9713, 9892,10070,10247,10422,10596,10769,10940,11110,11278,11445,11611,11776,11939, 12101,12262,12422,12580,12737,12893,13048,13201,13354,13505,13655,13803,13951, 14097,14243,14387,14530,14672,14813,14952,15091,15229,15365,15500,15635,15768, 15900,16032,16162,16291,16419,16547,16673,16798,16922,17046,17168,17289,17410, 17529,17648,17766,17882,17998,18113,18227,18340,18453,18564,18675,18784,18893, 19001,19108,19215,19320,19425,19529,19632,19734,19835,19936,20036,20135,20233, 20331,20427,20523,20619,20713,20807,20900,20993,21084,21175,21265,21355,21444, 21532,21619,21706,21792,21878,21962,22046,22130,22213,22295,22376,22457,22537, 22617,22696,22774,22852,22929,23006,23082,23157,23232,23306,23380,23453,23525, 23597,23669,23739,23810,23879,23949,24017,24085,24153,24220,24286,24352,24418, 24483,24547,24611,24675,24738,24800,24862,24924,24985,25045,25106,25165,25224, 25283,25341,25399,25456,25513,25570,25626,25681,25737,25791,25846,25899,25953, 26006,26059,26111,26163,26214,26265,26316,26366,26416,26465,26514,26563,26611, 26659,26707,26754,26801,26847,26893,26939,26984,27029,27074,27118,27162,27206, 27249,27292,27335,27377,27419,27460,27502,27543,27583,27624,27664,27703,27743, 27782,27821,27859,27897,27935,27973,28010,28047,28084,28121,28157,28193,28228, 28263,28299,28333,28368,28402,28436,28470,28503,28536,28569,28602,28634,28667, 28699,28730,28762,28793,28824,28854,28885,28915,28945,28975,29004,29034,29063, 29092,29120,29149,29177,29205,29232,29260,29287,29314,29341,29368,29394,29421, 29447,29472,29498,29524,29549,29574,29599,29623,29648,29672,29696,29720,29744, 29767,29791,29814,29837,29860,29882,29905,29927,29949,29971,29993,30014,30036, 30057,30078,30099,30120,30141,30161,30181,30201,30221,30241,30261,30280,30300, 30319,30338,30357,30376,30394,30413,30431,30449,30467,30485,30503,30521,30538, 30555,30573,30590,30607,30624,30640,30657,30673,30690,30706,30722,30738,30753, 30769,30785,30800,30815,30831,30846,30861,30876,30890,30905,30919,30934,30948, 30962,30976,30990,31004,31018,31031,31045,31058,31072,31085,31098,31111,31124, 31137,31149,31162,31174,31187,31199,31211,31223,31235,31247,31259,31271,31283, 31294,31306,31317,31328,31339,31351,31362,31373,31383,31394,31405,31416,31426, 31436,31447,31457,31467,31477,31487,31497,31507,31517,31527,31537,31546,31556, 31565,31574,31584,31593,31602,31611,31620,31629,31638,31647,31655,31664,31673, 31681,31690,31698,31706,31715,31723,31731,31739,31747,31755,31763,31771,31778, 31786,31794,31801,31809,31816,31824,31831,31838,31846,31853,31860,31867,31874, 31881,31888,31895,31902,31908,31915,31922,31928,31935,31941,31948,31954,31960, 31967,31973,31979,31985,31991,31997,32003,32009,32015,32021,32027,32033,32038, 32044,32050,32055,32061,32066,32072,32077,32083,32088,32093,32098,32104,32109, 32114,32119,32124,32129,32134,32139,32144,32149,32154,32158,32163,32168,32173, 32177,32182,32186,32191,32195,32200,32204,32209,32213,32217,32222,32226,32230, 32234,32238,32242,32247,32251,32255,32259,32263,32267,32270,32274,32278,32282, 32286,32290,32293,32297,32301,32304,32308,32311,32315,32318,32322,32325,32329, 32332,32336,32339,32342,32346,32349,32352,32356,32359,32362,32365,32368,32371, 32374,32377,32381,32384,32387,32389,32392,32395,32398,32401,32404,32407,32410, 32412,32415,32418,32421,32423,32426,32429,32431,32434,32437,32439,32442,32444, 32447,32449,32452,32454,32457,32459,32461,32464,32466,32469,32471,32473,32476, 32478,32480,32482,32485,32487,32489,32491,32493,32495,32497,32500,32502,32504, 32506,32508,32510,32512,32514,32516,32518,32520,32522,32524,32526,32527,32529, 32531,32533,32535,32537,32538,32540,32542,32544,32545,32547}; static uma_zone_t qdiffsample_zone; static int ertt_id; VNET_DEFINE_STATIC(uint32_t, cdg_alpha_inc); VNET_DEFINE_STATIC(uint32_t, cdg_beta_delay); VNET_DEFINE_STATIC(uint32_t, cdg_beta_loss); VNET_DEFINE_STATIC(uint32_t, cdg_smoothing_factor); VNET_DEFINE_STATIC(uint32_t, cdg_exp_backoff_scale); VNET_DEFINE_STATIC(uint32_t, cdg_consec_cong); VNET_DEFINE_STATIC(uint32_t, cdg_hold_backoff); #define V_cdg_alpha_inc VNET(cdg_alpha_inc) #define V_cdg_beta_delay VNET(cdg_beta_delay) #define V_cdg_beta_loss VNET(cdg_beta_loss) #define V_cdg_smoothing_factor VNET(cdg_smoothing_factor) #define V_cdg_exp_backoff_scale VNET(cdg_exp_backoff_scale) #define V_cdg_consec_cong VNET(cdg_consec_cong) #define V_cdg_hold_backoff VNET(cdg_hold_backoff) /* Function prototypes. */ static int cdg_mod_init(void); static int cdg_mod_destroy(void); static void cdg_conn_init(struct cc_var *ccv); static int cdg_cb_init(struct cc_var *ccv, void *ptr); static void cdg_cb_destroy(struct cc_var *ccv); static void cdg_cong_signal(struct cc_var *ccv, uint32_t signal_type); static void cdg_ack_received(struct cc_var *ccv, uint16_t ack_type); static size_t cdg_data_sz(void); struct cc_algo cdg_cc_algo = { .name = "cdg", .mod_init = cdg_mod_init, .ack_received = cdg_ack_received, .cb_destroy = cdg_cb_destroy, .cb_init = cdg_cb_init, .conn_init = cdg_conn_init, .cong_signal = cdg_cong_signal, .mod_destroy = cdg_mod_destroy, .cc_data_sz = cdg_data_sz, .post_recovery = newreno_cc_post_recovery, .after_idle = newreno_cc_after_idle, }; /* Vnet created and being initialised. */ static void cdg_init_vnet(const void *unused __unused) { V_cdg_alpha_inc = 0; V_cdg_beta_delay = 70; V_cdg_beta_loss = 50; V_cdg_smoothing_factor = 8; V_cdg_exp_backoff_scale = 3; V_cdg_consec_cong = 5; V_cdg_hold_backoff = 5; } static int cdg_mod_init(void) { VNET_ITERATOR_DECL(v); ertt_id = khelp_get_id("ertt"); if (ertt_id <= 0) return (EINVAL); qdiffsample_zone = uma_zcreate("cdg_qdiffsample", sizeof(struct qdiff_sample), NULL, NULL, NULL, NULL, 0, 0); VNET_LIST_RLOCK(); VNET_FOREACH(v) { CURVNET_SET(v); cdg_init_vnet(NULL); CURVNET_RESTORE(); } VNET_LIST_RUNLOCK(); return (0); } static int cdg_mod_destroy(void) { uma_zdestroy(qdiffsample_zone); return (0); } static size_t cdg_data_sz(void) { return (sizeof(struct cdg)); } static int cdg_cb_init(struct cc_var *ccv, void *ptr) { struct cdg *cdg_data; INP_WLOCK_ASSERT(tptoinpcb(ccv->ccvc.tcp)); if (ptr == NULL) { cdg_data = malloc(sizeof(struct cdg), M_CC_MEM, M_NOWAIT); if (cdg_data == NULL) return (ENOMEM); } else { cdg_data = ptr; } cdg_data->shadow_w = 0; cdg_data->max_qtrend = 0; cdg_data->min_qtrend = 0; cdg_data->queue_state = CDG_Q_UNKNOWN; cdg_data->maxrtt_in_rtt = 0; cdg_data->maxrtt_in_prevrtt = 0; cdg_data->minrtt_in_rtt = INT_MAX; cdg_data->minrtt_in_prevrtt = 0; cdg_data->window_incr = 0; cdg_data->rtt_count = 0; cdg_data->consec_cong_cnt = 0; cdg_data->sample_q_size = V_cdg_smoothing_factor; cdg_data->num_samples = 0; STAILQ_INIT(&cdg_data->qdiffmin_q); STAILQ_INIT(&cdg_data->qdiffmax_q); ccv->cc_data = cdg_data; return (0); } static void cdg_conn_init(struct cc_var *ccv) { struct cdg *cdg_data = ccv->cc_data; /* * Initialise the shadow_cwnd in case we are competing with loss based * flows from the start */ cdg_data->shadow_w = CCV(ccv, snd_cwnd); } static void cdg_cb_destroy(struct cc_var *ccv) { struct cdg *cdg_data; struct qdiff_sample *qds, *qds_n; cdg_data = ccv->cc_data; qds = STAILQ_FIRST(&cdg_data->qdiffmin_q); while (qds != NULL) { qds_n = STAILQ_NEXT(qds, qdiff_lnk); uma_zfree(qdiffsample_zone,qds); qds = qds_n; } qds = STAILQ_FIRST(&cdg_data->qdiffmax_q); while (qds != NULL) { qds_n = STAILQ_NEXT(qds, qdiff_lnk); uma_zfree(qdiffsample_zone,qds); qds = qds_n; } free(ccv->cc_data, M_CC_MEM); } static int cdg_beta_handler(SYSCTL_HANDLER_ARGS) { int error; uint32_t new; new = *(uint32_t *)arg1; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr != NULL) { if (new == 0 || new > 100) error = EINVAL; else *(uint32_t *)arg1 = new; } return (error); } static int cdg_exp_backoff_scale_handler(SYSCTL_HANDLER_ARGS) { int error; uint32_t new; new = *(uint32_t *)arg1; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr != NULL) { if (new < 1) error = EINVAL; else *(uint32_t *)arg1 = new; } return (error); } static inline uint32_t cdg_window_decrease(struct cc_var *ccv, unsigned long owin, unsigned int beta) { return ((ulmin(CCV(ccv, snd_wnd), owin) * beta) / 100); } /* * Window increase function * This window increase function is independent of the initial window size * to ensure small window flows are not discriminated against (i.e. fairness). * It increases at 1pkt/rtt like Reno for alpha_inc rtts, and then 2pkts/rtt for * the next alpha_inc rtts, etc. */ static void cdg_window_increase(struct cc_var *ccv, int new_measurement) { struct cdg *cdg_data; int incr, s_w_incr; cdg_data = ccv->cc_data; incr = s_w_incr = 0; if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh)) { /* Slow start. */ incr = CCV(ccv, t_maxseg); s_w_incr = incr; cdg_data->window_incr = cdg_data->rtt_count = 0; } else { /* Congestion avoidance. */ if (new_measurement) { s_w_incr = CCV(ccv, t_maxseg); if (V_cdg_alpha_inc == 0) { incr = CCV(ccv, t_maxseg); } else { if (++cdg_data->rtt_count >= V_cdg_alpha_inc) { cdg_data->window_incr++; cdg_data->rtt_count = 0; } incr = CCV(ccv, t_maxseg) * cdg_data->window_incr; } } } if (cdg_data->shadow_w > 0) cdg_data->shadow_w = ulmin(cdg_data->shadow_w + s_w_incr, TCP_MAXWIN << CCV(ccv, snd_scale)); CCV(ccv, snd_cwnd) = ulmin(CCV(ccv, snd_cwnd) + incr, TCP_MAXWIN << CCV(ccv, snd_scale)); } static void cdg_cong_signal(struct cc_var *ccv, uint32_t signal_type) { struct cdg *cdg_data = ccv->cc_data; switch(signal_type) { case CC_CDG_DELAY: CCV(ccv, snd_ssthresh) = cdg_window_decrease(ccv, CCV(ccv, snd_cwnd), V_cdg_beta_delay); CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); CCV(ccv, snd_recover) = CCV(ccv, snd_max); cdg_data->window_incr = cdg_data->rtt_count = 0; ENTER_CONGRECOVERY(CCV(ccv, t_flags)); break; case CC_NDUPACK: /* * If already responding to congestion OR we have guessed no * queue in the path is full. */ if (IN_CONGRECOVERY(CCV(ccv, t_flags)) || cdg_data->queue_state < CDG_Q_FULL) { CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd); CCV(ccv, snd_recover) = CCV(ccv, snd_max); } else { /* * Loss is likely to be congestion related. We have * inferred a queue full state, so have shadow window * react to loss as NewReno would. */ if (cdg_data->shadow_w > 0) cdg_data->shadow_w = cdg_window_decrease(ccv, cdg_data->shadow_w, RENO_BETA); CCV(ccv, snd_ssthresh) = max(cdg_data->shadow_w, cdg_window_decrease(ccv, CCV(ccv, snd_cwnd), V_cdg_beta_loss)); cdg_data->window_incr = cdg_data->rtt_count = 0; } ENTER_RECOVERY(CCV(ccv, t_flags)); break; default: newreno_cc_cong_signal(ccv, signal_type); break; } } /* * Using a negative exponential probabilistic backoff so that sources with * varying RTTs which share the same link will, on average, have the same * probability of backoff over time. * * Prob_backoff = 1 - exp(-qtrend / V_cdg_exp_backoff_scale), where * V_cdg_exp_backoff_scale is the average qtrend for the exponential backoff. */ static inline int prob_backoff(long qtrend) { int backoff, idx, p; backoff = (qtrend > ((MAXGRAD * V_cdg_exp_backoff_scale) << D_P_E)); if (!backoff) { if (V_cdg_exp_backoff_scale > 1) idx = (qtrend + V_cdg_exp_backoff_scale / 2) / V_cdg_exp_backoff_scale; else idx = qtrend; /* Backoff probability proportional to rate of queue growth. */ p = (INT_MAX / (1 << EXP_PREC)) * probexp[idx]; backoff = (random() < p); } return (backoff); } static inline void calc_moving_average(struct cdg *cdg_data, long qdiff_max, long qdiff_min) { struct qdiff_sample *qds; ++cdg_data->num_samples; if (cdg_data->num_samples > cdg_data->sample_q_size) { /* Minimum RTT. */ qds = STAILQ_FIRST(&cdg_data->qdiffmin_q); cdg_data->min_qtrend = cdg_data->min_qtrend + (qdiff_min - qds->qdiff) / cdg_data->sample_q_size; STAILQ_REMOVE_HEAD(&cdg_data->qdiffmin_q, qdiff_lnk); qds->qdiff = qdiff_min; STAILQ_INSERT_TAIL(&cdg_data->qdiffmin_q, qds, qdiff_lnk); /* Maximum RTT. */ qds = STAILQ_FIRST(&cdg_data->qdiffmax_q); cdg_data->max_qtrend = cdg_data->max_qtrend + (qdiff_max - qds->qdiff) / cdg_data->sample_q_size; STAILQ_REMOVE_HEAD(&cdg_data->qdiffmax_q, qdiff_lnk); qds->qdiff = qdiff_max; STAILQ_INSERT_TAIL(&cdg_data->qdiffmax_q, qds, qdiff_lnk); --cdg_data->num_samples; } else { qds = uma_zalloc(qdiffsample_zone, M_NOWAIT); if (qds != NULL) { cdg_data->min_qtrend = cdg_data->min_qtrend + qdiff_min / cdg_data->sample_q_size; qds->qdiff = qdiff_min; STAILQ_INSERT_TAIL(&cdg_data->qdiffmin_q, qds, qdiff_lnk); } qds = uma_zalloc(qdiffsample_zone, M_NOWAIT); if (qds) { cdg_data->max_qtrend = cdg_data->max_qtrend + qdiff_max / cdg_data->sample_q_size; qds->qdiff = qdiff_max; STAILQ_INSERT_TAIL(&cdg_data->qdiffmax_q, qds, qdiff_lnk); } } } static void cdg_ack_received(struct cc_var *ccv, uint16_t ack_type) { struct cdg *cdg_data; struct ertt *e_t; long qdiff_max, qdiff_min; int congestion, new_measurement, slowstart; cdg_data = ccv->cc_data; - e_t = (struct ertt *)khelp_get_osd(CCV(ccv, osd), ertt_id); + e_t = (struct ertt *)khelp_get_osd(&CCV(ccv, t_osd), ertt_id); new_measurement = e_t->flags & ERTT_NEW_MEASUREMENT; congestion = 0; cdg_data->maxrtt_in_rtt = imax(e_t->rtt, cdg_data->maxrtt_in_rtt); cdg_data->minrtt_in_rtt = imin(e_t->rtt, cdg_data->minrtt_in_rtt); if (new_measurement) { slowstart = (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh)); /* * Update smoothed gradient measurements. Since we are only * using one measurement per RTT, use max or min rtt_in_rtt. * This is also less noisy than a sample RTT measurement. Max * RTT measurements can have trouble due to OS issues. */ if (cdg_data->maxrtt_in_prevrtt) { qdiff_max = ((long)(cdg_data->maxrtt_in_rtt - cdg_data->maxrtt_in_prevrtt) << D_P_E ); qdiff_min = ((long)(cdg_data->minrtt_in_rtt - cdg_data->minrtt_in_prevrtt) << D_P_E ); if (cdg_data->sample_q_size == 0) { cdg_data->max_qtrend = qdiff_max; cdg_data->min_qtrend = qdiff_min; } else calc_moving_average(cdg_data, qdiff_max, qdiff_min); /* Probabilistic backoff with respect to gradient. */ if (slowstart && qdiff_min > 0) congestion = prob_backoff(qdiff_min); else if (cdg_data->min_qtrend > 0) congestion = prob_backoff(cdg_data->min_qtrend); else if (slowstart && qdiff_max > 0) congestion = prob_backoff(qdiff_max); else if (cdg_data->max_qtrend > 0) congestion = prob_backoff(cdg_data->max_qtrend); /* Update estimate of queue state. */ if (cdg_data->min_qtrend > 0 && cdg_data->max_qtrend <= 0) { cdg_data->queue_state = CDG_Q_FULL; } else if (cdg_data->min_qtrend >= 0 && cdg_data->max_qtrend < 0) { cdg_data->queue_state = CDG_Q_EMPTY; cdg_data->shadow_w = 0; } else if (cdg_data->min_qtrend > 0 && cdg_data->max_qtrend > 0) { cdg_data->queue_state = CDG_Q_RISING; } else if (cdg_data->min_qtrend < 0 && cdg_data->max_qtrend < 0) { cdg_data->queue_state = CDG_Q_FALLING; } if (cdg_data->min_qtrend < 0 || cdg_data->max_qtrend < 0) cdg_data->consec_cong_cnt = 0; } cdg_data->minrtt_in_prevrtt = cdg_data->minrtt_in_rtt; cdg_data->minrtt_in_rtt = INT_MAX; cdg_data->maxrtt_in_prevrtt = cdg_data->maxrtt_in_rtt; cdg_data->maxrtt_in_rtt = 0; e_t->flags &= ~ERTT_NEW_MEASUREMENT; } if (congestion) { cdg_data->consec_cong_cnt++; if (!IN_RECOVERY(CCV(ccv, t_flags))) { if (cdg_data->consec_cong_cnt <= V_cdg_consec_cong) cdg_cong_signal(ccv, CC_CDG_DELAY); else /* * We have been backing off but the queue is not * falling. Assume we are competing with * loss-based flows and don't back off for the * next V_cdg_hold_backoff RTT periods. */ if (cdg_data->consec_cong_cnt >= V_cdg_consec_cong + V_cdg_hold_backoff) cdg_data->consec_cong_cnt = 0; /* Won't see effect until 2nd RTT. */ cdg_data->maxrtt_in_prevrtt = 0; /* * Resync shadow window in case we are competing with a * loss based flow */ cdg_data->shadow_w = ulmax(CCV(ccv, snd_cwnd), cdg_data->shadow_w); } } else if (ack_type == CC_ACK) cdg_window_increase(ccv, new_measurement); } /* When a vnet is created and being initialised, init the per-stack CDG vars. */ VNET_SYSINIT(cdg_init_vnet, SI_SUB_PROTO_BEGIN, SI_ORDER_FIRST, cdg_init_vnet, NULL); SYSCTL_DECL(_net_inet_tcp_cc_cdg); SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, cdg, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, "CAIA delay-gradient congestion control related settings"); SYSCTL_STRING(_net_inet_tcp_cc_cdg, OID_AUTO, version, CTLFLAG_RD, CDG_VERSION, sizeof(CDG_VERSION) - 1, "Current algorithm/implementation version number"); SYSCTL_UINT(_net_inet_tcp_cc_cdg, OID_AUTO, alpha_inc, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(cdg_alpha_inc), 0, "Increment the window increase factor alpha by 1 MSS segment every " "alpha_inc RTTs during congestion avoidance mode."); SYSCTL_PROC(_net_inet_tcp_cc_cdg, OID_AUTO, beta_delay, CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(cdg_beta_delay), 70, &cdg_beta_handler, "IU", "Delay-based window decrease factor as a percentage " "(on delay-based backoff, w = w * beta_delay / 100)"); SYSCTL_PROC(_net_inet_tcp_cc_cdg, OID_AUTO, beta_loss, CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(cdg_beta_loss), 50, &cdg_beta_handler, "IU", "Loss-based window decrease factor as a percentage " "(on loss-based backoff, w = w * beta_loss / 100)"); SYSCTL_PROC(_net_inet_tcp_cc_cdg, OID_AUTO, exp_backoff_scale, CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(cdg_exp_backoff_scale), 2, &cdg_exp_backoff_scale_handler, "IU", "Scaling parameter for the probabilistic exponential backoff"); SYSCTL_UINT(_net_inet_tcp_cc_cdg, OID_AUTO, smoothing_factor, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(cdg_smoothing_factor), 8, "Number of samples used for moving average smoothing (0 = no smoothing)"); SYSCTL_UINT(_net_inet_tcp_cc_cdg, OID_AUTO, loss_compete_consec_cong, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(cdg_consec_cong), 5, "Number of consecutive delay-gradient based congestion episodes which will " "trigger loss based CC compatibility"); SYSCTL_UINT(_net_inet_tcp_cc_cdg, OID_AUTO, loss_compete_hold_backoff, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(cdg_hold_backoff), 5, "Number of consecutive delay-gradient based congestion episodes to hold " "the window backoff for loss based CC compatibility"); DECLARE_CC_MODULE(cdg, &cdg_cc_algo); MODULE_VERSION(cdg, 2); MODULE_DEPEND(cdg, ertt, 1, 1, 1); diff --git a/sys/netinet/cc/cc_chd.c b/sys/netinet/cc/cc_chd.c index d8bd764a1cba..091aeb60f698 100644 --- a/sys/netinet/cc/cc_chd.c +++ b/sys/netinet/cc/cc_chd.c @@ -1,508 +1,508 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2009-2010 * Swinburne University of Technology, Melbourne, Australia * Copyright (c) 2010-2011 The FreeBSD Foundation * All rights reserved. * * This software was developed at the Centre for Advanced Internet * Architectures, Swinburne University of Technology, by David Hayes and * Lawrence Stewart, 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. * * 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. */ /* * An implementation of the CAIA-Hamilton delay based congestion control * algorithm, based on "Improved coexistence and loss tolerance for delay based * TCP congestion control" by D. A. Hayes and G. Armitage., in 35th Annual IEEE * Conference on Local Computer Networks (LCN 2010), Denver, Colorado, USA, * 11-14 October 2010. * * Originally released as part of the NewTCP research project at Swinburne * University of Technology's Centre for Advanced Internet Architectures, * Melbourne, Australia, which was made possible in part by a grant from the * Cisco University Research Program Fund at Community Foundation Silicon * Valley. More details are available at: * http://caia.swin.edu.au/urp/newtcp/ */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* * Private signal type for rate based congestion signal. * See for appropriate bit-range to use for private signals. */ #define CC_CHD_DELAY 0x02000000 /* Largest possible number returned by random(). */ #define RANDOM_MAX INT_MAX static void chd_ack_received(struct cc_var *ccv, uint16_t ack_type); static void chd_cb_destroy(struct cc_var *ccv); static int chd_cb_init(struct cc_var *ccv, void *ptr); static void chd_cong_signal(struct cc_var *ccv, uint32_t signal_type); static void chd_conn_init(struct cc_var *ccv); static int chd_mod_init(void); static size_t chd_data_sz(void); struct chd { /* * Shadow window - keeps track of what the NewReno congestion window * would have been if delay-based cwnd backoffs had not been made. This * functionality aids coexistence with loss-based TCP flows which may be * sharing links along the path. */ unsigned long shadow_w; /* * Loss-based TCP compatibility flag - When set, it turns on the shadow * window functionality. */ int loss_compete; /* The maximum round trip time seen within a measured rtt period. */ int maxrtt_in_rtt; /* The previous qdly that caused cwnd to backoff. */ int prev_backoff_qdly; }; static int ertt_id; VNET_DEFINE_STATIC(uint32_t, chd_qmin) = 5; VNET_DEFINE_STATIC(uint32_t, chd_pmax) = 50; VNET_DEFINE_STATIC(uint32_t, chd_loss_fair) = 1; VNET_DEFINE_STATIC(uint32_t, chd_use_max) = 1; VNET_DEFINE_STATIC(uint32_t, chd_qthresh) = 20; #define V_chd_qthresh VNET(chd_qthresh) #define V_chd_qmin VNET(chd_qmin) #define V_chd_pmax VNET(chd_pmax) #define V_chd_loss_fair VNET(chd_loss_fair) #define V_chd_use_max VNET(chd_use_max) struct cc_algo chd_cc_algo = { .name = "chd", .ack_received = chd_ack_received, .cb_destroy = chd_cb_destroy, .cb_init = chd_cb_init, .cong_signal = chd_cong_signal, .conn_init = chd_conn_init, .mod_init = chd_mod_init, .cc_data_sz = chd_data_sz, .after_idle = newreno_cc_after_idle, .post_recovery = newreno_cc_post_recovery, }; static __inline void chd_window_decrease(struct cc_var *ccv) { unsigned long win; win = min(CCV(ccv, snd_wnd), CCV(ccv, snd_cwnd)) / CCV(ccv, t_maxseg); win -= max((win / 2), 1); CCV(ccv, snd_ssthresh) = max(win, 2) * CCV(ccv, t_maxseg); } /* * Probabilistic backoff function. Returns 1 if we should backoff or 0 * otherwise. The calculation of p is similar to the calculation of p in cc_hd. */ static __inline int should_backoff(int qdly, int maxqdly, struct chd *chd_data) { unsigned long p, rand; rand = random(); if (qdly < V_chd_qthresh) { chd_data->loss_compete = 0; p = (((RANDOM_MAX / 100) * V_chd_pmax) / (V_chd_qthresh - V_chd_qmin)) * (qdly - V_chd_qmin); } else { if (qdly > V_chd_qthresh) { p = (((RANDOM_MAX / 100) * V_chd_pmax) / (maxqdly - V_chd_qthresh)) * (maxqdly - qdly); if (V_chd_loss_fair && rand < p) chd_data->loss_compete = 1; } else { p = (RANDOM_MAX / 100) * V_chd_pmax; chd_data->loss_compete = 0; } } return (rand < p); } static __inline void chd_window_increase(struct cc_var *ccv, int new_measurement) { struct chd *chd_data; int incr; chd_data = ccv->cc_data; incr = 0; if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh)) { /* Adapted from NewReno slow start. */ if (V_tcp_do_rfc3465) { /* In slow-start with ABC enabled. */ if (CCV(ccv, snd_nxt) == CCV(ccv, snd_max)) { /* Not due to RTO. */ incr = min(ccv->bytes_this_ack, V_tcp_abc_l_var * CCV(ccv, t_maxseg)); } else { /* Due to RTO. */ incr = min(ccv->bytes_this_ack, CCV(ccv, t_maxseg)); } } else incr = CCV(ccv, t_maxseg); } else { /* Congestion avoidance. */ if (V_tcp_do_rfc3465) { if (ccv->flags & CCF_ABC_SENTAWND) { ccv->flags &= ~CCF_ABC_SENTAWND; incr = CCV(ccv, t_maxseg); } } else if (new_measurement) incr = CCV(ccv, t_maxseg); } if (chd_data->shadow_w > 0) { /* Track NewReno window. */ chd_data->shadow_w = min(chd_data->shadow_w + incr, TCP_MAXWIN << CCV(ccv, snd_scale)); } CCV(ccv,snd_cwnd) = min(CCV(ccv, snd_cwnd) + incr, TCP_MAXWIN << CCV(ccv, snd_scale)); } /* * All ACK signals are used for timing measurements to determine delay-based * congestion. However, window increases are only performed when * ack_type == CC_ACK. */ static void chd_ack_received(struct cc_var *ccv, uint16_t ack_type) { struct chd *chd_data; struct ertt *e_t; int backoff, new_measurement, qdly, rtt; - e_t = khelp_get_osd(CCV(ccv, osd), ertt_id); + e_t = khelp_get_osd(&CCV(ccv, t_osd), ertt_id); chd_data = ccv->cc_data; new_measurement = e_t->flags & ERTT_NEW_MEASUREMENT; backoff = qdly = 0; chd_data->maxrtt_in_rtt = imax(e_t->rtt, chd_data->maxrtt_in_rtt); if (new_measurement) { /* * There is a new per RTT measurement, so check to see if there * is delay based congestion. */ rtt = V_chd_use_max ? chd_data->maxrtt_in_rtt : e_t->rtt; chd_data->maxrtt_in_rtt = 0; if (rtt && e_t->minrtt && !IN_RECOVERY(CCV(ccv, t_flags))) { qdly = rtt - e_t->minrtt; if (qdly > V_chd_qmin) { /* * Probabilistic delay based congestion * indication. */ backoff = should_backoff(qdly, e_t->maxrtt - e_t->minrtt, chd_data); } else chd_data->loss_compete = 0; } /* Reset per RTT measurement flag to start a new measurement. */ e_t->flags &= ~ERTT_NEW_MEASUREMENT; } if (backoff) { /* * Update shadow_w before delay based backoff. */ if (chd_data->loss_compete || qdly > chd_data->prev_backoff_qdly) { /* * Delay is higher than when we backed off previously, * so it is possible that this flow is competing with * loss based flows. */ chd_data->shadow_w = max(CCV(ccv, snd_cwnd), chd_data->shadow_w); } else { /* * Reset shadow_w, as it is probable that this flow is * not competing with loss based flows at the moment. */ chd_data->shadow_w = 0; } chd_data->prev_backoff_qdly = qdly; /* * Send delay-based congestion signal to the congestion signal * handler. */ chd_cong_signal(ccv, CC_CHD_DELAY); } else if (ack_type == CC_ACK) chd_window_increase(ccv, new_measurement); } static void chd_cb_destroy(struct cc_var *ccv) { free(ccv->cc_data, M_CC_MEM); } size_t chd_data_sz(void) { return (sizeof(struct chd)); } static int chd_cb_init(struct cc_var *ccv, void *ptr) { struct chd *chd_data; INP_WLOCK_ASSERT(tptoinpcb(ccv->ccvc.tcp)); if (ptr == NULL) { chd_data = malloc(sizeof(struct chd), M_CC_MEM, M_NOWAIT); if (chd_data == NULL) return (ENOMEM); } else chd_data = ptr; chd_data->shadow_w = 0; ccv->cc_data = chd_data; return (0); } static void chd_cong_signal(struct cc_var *ccv, uint32_t signal_type) { struct ertt *e_t; struct chd *chd_data; int qdly; - e_t = khelp_get_osd(CCV(ccv, osd), ertt_id); + e_t = khelp_get_osd(&CCV(ccv, t_osd), ertt_id); chd_data = ccv->cc_data; qdly = imax(e_t->rtt, chd_data->maxrtt_in_rtt) - e_t->minrtt; switch(signal_type) { case CC_CHD_DELAY: chd_window_decrease(ccv); /* Set new ssthresh. */ CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); CCV(ccv, snd_recover) = CCV(ccv, snd_max); ENTER_CONGRECOVERY(CCV(ccv, t_flags)); break; case CC_NDUPACK: /* Packet loss. */ /* * Only react to loss as a congestion signal if qdly > * V_chd_qthresh. If qdly is less than qthresh, presume that * this is a non congestion related loss. If qdly is greater * than qthresh, assume that we are competing with loss based * tcp flows and restore window from any unnecessary backoffs, * before the decrease. */ if (!IN_RECOVERY(CCV(ccv, t_flags)) && qdly > V_chd_qthresh) { if (chd_data->loss_compete) { CCV(ccv, snd_cwnd) = max(CCV(ccv, snd_cwnd), chd_data->shadow_w); } chd_window_decrease(ccv); } else { /* * This loss isn't congestion related, or already * recovering from congestion. */ CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd); CCV(ccv, snd_recover) = CCV(ccv, snd_max); } if (chd_data->shadow_w > 0) { chd_data->shadow_w = max(chd_data->shadow_w / CCV(ccv, t_maxseg) / 2, 2) * CCV(ccv, t_maxseg); } ENTER_FASTRECOVERY(CCV(ccv, t_flags)); break; default: newreno_cc_cong_signal(ccv, signal_type); } } static void chd_conn_init(struct cc_var *ccv) { struct chd *chd_data; chd_data = ccv->cc_data; chd_data->prev_backoff_qdly = 0; chd_data->maxrtt_in_rtt = 0; chd_data->loss_compete = 0; /* * Initialise the shadow_cwnd to be equal to snd_cwnd in case we are * competing with loss based flows from the start. */ chd_data->shadow_w = CCV(ccv, snd_cwnd); } static int chd_mod_init(void) { ertt_id = khelp_get_id("ertt"); if (ertt_id <= 0) { printf("%s: h_ertt module not found\n", __func__); return (ENOENT); } return (0); } static int chd_loss_fair_handler(SYSCTL_HANDLER_ARGS) { int error; uint32_t new; new = V_chd_loss_fair; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr != NULL) { if (new > 1) error = EINVAL; else V_chd_loss_fair = new; } return (error); } static int chd_pmax_handler(SYSCTL_HANDLER_ARGS) { int error; uint32_t new; new = V_chd_pmax; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr != NULL) { if (new == 0 || new > 100) error = EINVAL; else V_chd_pmax = new; } return (error); } static int chd_qthresh_handler(SYSCTL_HANDLER_ARGS) { int error; uint32_t new; new = V_chd_qthresh; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr != NULL) { if (new <= V_chd_qmin) error = EINVAL; else V_chd_qthresh = new; } return (error); } SYSCTL_DECL(_net_inet_tcp_cc_chd); SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, chd, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, "CAIA Hamilton delay-based congestion control related settings"); SYSCTL_PROC(_net_inet_tcp_cc_chd, OID_AUTO, loss_fair, CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(chd_loss_fair), 1, &chd_loss_fair_handler, "IU", "Flag to enable shadow window functionality."); SYSCTL_PROC(_net_inet_tcp_cc_chd, OID_AUTO, pmax, CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(chd_pmax), 5, &chd_pmax_handler, "IU", "Per RTT maximum backoff probability as a percentage"); SYSCTL_PROC(_net_inet_tcp_cc_chd, OID_AUTO, queue_threshold, CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(chd_qthresh), 20, &chd_qthresh_handler, "IU", "Queueing congestion threshold in ticks"); SYSCTL_UINT(_net_inet_tcp_cc_chd, OID_AUTO, queue_min, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(chd_qmin), 5, "Minimum queueing delay threshold in ticks"); SYSCTL_UINT(_net_inet_tcp_cc_chd, OID_AUTO, use_max, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(chd_use_max), 1, "Use the maximum RTT seen within the measurement period (RTT) " "as the basic delay measurement for the algorithm."); DECLARE_CC_MODULE(chd, &chd_cc_algo); MODULE_VERSION(chd, 2); MODULE_DEPEND(chd, ertt, 1, 1, 1); diff --git a/sys/netinet/cc/cc_hd.c b/sys/netinet/cc/cc_hd.c index 9e910568e27c..1abca5fb29b7 100644 --- a/sys/netinet/cc/cc_hd.c +++ b/sys/netinet/cc/cc_hd.c @@ -1,261 +1,262 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2009-2010 * Swinburne University of Technology, Melbourne, Australia * Copyright (c) 2010 Lawrence Stewart * Copyright (c) 2010-2011 The FreeBSD Foundation * All rights reserved. * * This software was developed at the Centre for Advanced Internet * Architectures, Swinburne University of Technology, by David Hayes and * Lawrence Stewart, 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. * * 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. */ /* * An implementation of the Hamilton Institute's delay-based congestion control * algorithm for FreeBSD, based on "A strategy for fair coexistence of loss and * delay-based congestion control algorithms," by L. Budzisz, R. Stanojevic, R. * Shorten, and F. Baker, IEEE Commun. Lett., vol. 13, no. 7, pp. 555--557, Jul. * 2009. * * Originally released as part of the NewTCP research project at Swinburne * University of Technology's Centre for Advanced Internet Architectures, * Melbourne, Australia, which was made possible in part by a grant from the * Cisco University Research Program Fund at Community Foundation Silicon * Valley. More details are available at: * http://caia.swin.edu.au/urp/newtcp/ */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include -#include +#include +#include #include #include #include #include #include #include /* Largest possible number returned by random(). */ #define RANDOM_MAX INT_MAX static void hd_ack_received(struct cc_var *ccv, uint16_t ack_type); static int hd_mod_init(void); static size_t hd_data_sz(void); static int ertt_id; VNET_DEFINE_STATIC(uint32_t, hd_qthresh) = 20; VNET_DEFINE_STATIC(uint32_t, hd_qmin) = 5; VNET_DEFINE_STATIC(uint32_t, hd_pmax) = 5; #define V_hd_qthresh VNET(hd_qthresh) #define V_hd_qmin VNET(hd_qmin) #define V_hd_pmax VNET(hd_pmax) struct cc_algo hd_cc_algo = { .name = "hd", .ack_received = hd_ack_received, .mod_init = hd_mod_init, .cc_data_sz = hd_data_sz, .after_idle = newreno_cc_after_idle, .cong_signal = newreno_cc_cong_signal, .post_recovery = newreno_cc_post_recovery, }; static size_t hd_data_sz(void) { return (0); } /* * Hamilton backoff function. Returns 1 if we should backoff or 0 otherwise. */ static __inline int should_backoff(int qdly, int maxqdly) { unsigned long p; if (qdly < V_hd_qthresh) { p = (((RANDOM_MAX / 100) * V_hd_pmax) / (V_hd_qthresh - V_hd_qmin)) * (qdly - V_hd_qmin); } else { if (qdly > V_hd_qthresh) p = (((RANDOM_MAX / 100) * V_hd_pmax) / (maxqdly - V_hd_qthresh)) * (maxqdly - qdly); else p = (RANDOM_MAX / 100) * V_hd_pmax; } return (random() < p); } /* * If the ack type is CC_ACK, and the inferred queueing delay is greater than * the Qmin threshold, cwnd is reduced probabilistically. When backing off due * to delay, HD behaves like NewReno when an ECN signal is received. HD behaves * as NewReno in all other circumstances. */ static void hd_ack_received(struct cc_var *ccv, uint16_t ack_type) { struct ertt *e_t; int qdly; if (ack_type == CC_ACK) { - e_t = khelp_get_osd(CCV(ccv, osd), ertt_id); + e_t = khelp_get_osd(&CCV(ccv, t_osd), ertt_id); if (e_t->rtt && e_t->minrtt && V_hd_qthresh > 0) { qdly = e_t->rtt - e_t->minrtt; if (qdly > V_hd_qmin && !IN_RECOVERY(CCV(ccv, t_flags))) { /* Probabilistic backoff of cwnd. */ if (should_backoff(qdly, e_t->maxrtt - e_t->minrtt)) { /* * Update cwnd and ssthresh update to * half cwnd and behave like an ECN (ie * not a packet loss). */ newreno_cc_cong_signal(ccv, CC_ECN); return; } } } } newreno_cc_ack_received(ccv, ack_type); } static int hd_mod_init(void) { ertt_id = khelp_get_id("ertt"); if (ertt_id <= 0) { printf("%s: h_ertt module not found\n", __func__); return (ENOENT); } return (0); } static int hd_pmax_handler(SYSCTL_HANDLER_ARGS) { int error; uint32_t new; new = V_hd_pmax; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr != NULL) { if (new == 0 || new > 100) error = EINVAL; else V_hd_pmax = new; } return (error); } static int hd_qmin_handler(SYSCTL_HANDLER_ARGS) { int error; uint32_t new; new = V_hd_qmin; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr != NULL) { if (new > V_hd_qthresh) error = EINVAL; else V_hd_qmin = new; } return (error); } static int hd_qthresh_handler(SYSCTL_HANDLER_ARGS) { int error; uint32_t new; new = V_hd_qthresh; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr != NULL) { if (new == 0 || new < V_hd_qmin) error = EINVAL; else V_hd_qthresh = new; } return (error); } SYSCTL_DECL(_net_inet_tcp_cc_hd); SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, hd, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, "Hamilton delay-based congestion control related settings"); SYSCTL_PROC(_net_inet_tcp_cc_hd, OID_AUTO, queue_threshold, CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(hd_qthresh), 20, &hd_qthresh_handler, "IU", "queueing congestion threshold (qth) in ticks"); SYSCTL_PROC(_net_inet_tcp_cc_hd, OID_AUTO, pmax, CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(hd_pmax), 5, &hd_pmax_handler, "IU", "per packet maximum backoff probability as a percentage"); SYSCTL_PROC(_net_inet_tcp_cc_hd, OID_AUTO, queue_min, CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(hd_qmin), 5, &hd_qmin_handler, "IU", "minimum queueing delay threshold (qmin) in ticks"); DECLARE_CC_MODULE(hd, &hd_cc_algo); MODULE_VERSION(hd, 2); MODULE_DEPEND(hd, ertt, 1, 1, 1); diff --git a/sys/netinet/cc/cc_vegas.c b/sys/netinet/cc/cc_vegas.c index 2056537a1df0..cbed488c2b96 100644 --- a/sys/netinet/cc/cc_vegas.c +++ b/sys/netinet/cc/cc_vegas.c @@ -1,314 +1,314 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2009-2010 * Swinburne University of Technology, Melbourne, Australia * Copyright (c) 2010 Lawrence Stewart * Copyright (c) 2010-2011 The FreeBSD Foundation * All rights reserved. * * This software was developed at the Centre for Advanced Internet * Architectures, Swinburne University of Technology, by David Hayes and * Lawrence Stewart, 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. * * 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. */ /* * An implementation of the Vegas congestion control algorithm for FreeBSD, * based on L. S. Brakmo and L. L. Peterson, "TCP Vegas: end to end congestion * avoidance on a global internet", IEEE J. Sel. Areas Commun., vol. 13, no. 8, * pp. 1465-1480, Oct. 1995. The original Vegas duplicate ack policy has not * been implemented, since clock ticks are not as coarse as they were (i.e. * 500ms) when Vegas was designed. Also, packets are timed once per RTT as in * the original paper. * * Originally released as part of the NewTCP research project at Swinburne * University of Technology's Centre for Advanced Internet Architectures, * Melbourne, Australia, which was made possible in part by a grant from the * Cisco University Research Program Fund at Community Foundation Silicon * Valley. More details are available at: * http://caia.swin.edu.au/urp/newtcp/ */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* * Private signal type for rate based congestion signal. * See for appropriate bit-range to use for private signals. */ #define CC_VEGAS_RATE 0x01000000 static void vegas_ack_received(struct cc_var *ccv, uint16_t ack_type); static void vegas_cb_destroy(struct cc_var *ccv); static int vegas_cb_init(struct cc_var *ccv, void *ptr); static void vegas_cong_signal(struct cc_var *ccv, uint32_t signal_type); static void vegas_conn_init(struct cc_var *ccv); static int vegas_mod_init(void); static size_t vegas_data_sz(void); struct vegas { int slow_start_toggle; }; static int32_t ertt_id; VNET_DEFINE_STATIC(uint32_t, vegas_alpha) = 1; VNET_DEFINE_STATIC(uint32_t, vegas_beta) = 3; #define V_vegas_alpha VNET(vegas_alpha) #define V_vegas_beta VNET(vegas_beta) struct cc_algo vegas_cc_algo = { .name = "vegas", .ack_received = vegas_ack_received, .cb_destroy = vegas_cb_destroy, .cb_init = vegas_cb_init, .cong_signal = vegas_cong_signal, .conn_init = vegas_conn_init, .mod_init = vegas_mod_init, .cc_data_sz = vegas_data_sz, .after_idle = newreno_cc_after_idle, .post_recovery = newreno_cc_post_recovery, }; /* * The vegas window adjustment is done once every RTT, as indicated by the * ERTT_NEW_MEASUREMENT flag. This flag is reset once the new measurement data * has been used. */ static void vegas_ack_received(struct cc_var *ccv, uint16_t ack_type) { struct ertt *e_t; struct vegas *vegas_data; long actual_tx_rate, expected_tx_rate, ndiff; - e_t = khelp_get_osd(CCV(ccv, osd), ertt_id); + e_t = khelp_get_osd(&CCV(ccv, t_osd), ertt_id); vegas_data = ccv->cc_data; if (e_t->flags & ERTT_NEW_MEASUREMENT) { /* Once per RTT. */ if (e_t->minrtt && e_t->markedpkt_rtt) { expected_tx_rate = e_t->marked_snd_cwnd / e_t->minrtt; actual_tx_rate = e_t->bytes_tx_in_marked_rtt / e_t->markedpkt_rtt; ndiff = (expected_tx_rate - actual_tx_rate) * e_t->minrtt / CCV(ccv, t_maxseg); if (ndiff < V_vegas_alpha) { if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh)) { vegas_data->slow_start_toggle = vegas_data->slow_start_toggle ? 0 : 1; } else { vegas_data->slow_start_toggle = 0; CCV(ccv, snd_cwnd) = min(CCV(ccv, snd_cwnd) + CCV(ccv, t_maxseg), TCP_MAXWIN << CCV(ccv, snd_scale)); } } else if (ndiff > V_vegas_beta) { /* Rate-based congestion. */ vegas_cong_signal(ccv, CC_VEGAS_RATE); vegas_data->slow_start_toggle = 0; } } e_t->flags &= ~ERTT_NEW_MEASUREMENT; } if (vegas_data->slow_start_toggle) newreno_cc_ack_received(ccv, ack_type); } static void vegas_cb_destroy(struct cc_var *ccv) { free(ccv->cc_data, M_CC_MEM); } static size_t vegas_data_sz(void) { return (sizeof(struct vegas)); } static int vegas_cb_init(struct cc_var *ccv, void *ptr) { struct vegas *vegas_data; INP_WLOCK_ASSERT(tptoinpcb(ccv->ccvc.tcp)); if (ptr == NULL) { vegas_data = malloc(sizeof(struct vegas), M_CC_MEM, M_NOWAIT); if (vegas_data == NULL) return (ENOMEM); } else vegas_data = ptr; vegas_data->slow_start_toggle = 1; ccv->cc_data = vegas_data; return (0); } /* * If congestion has been triggered triggered by the Vegas measured rates, it is * handled here, otherwise it falls back to newreno's congestion handling. */ static void vegas_cong_signal(struct cc_var *ccv, uint32_t signal_type) { struct vegas *vegas_data; int presignalrecov; vegas_data = ccv->cc_data; if (IN_RECOVERY(CCV(ccv, t_flags))) presignalrecov = 1; else presignalrecov = 0; switch(signal_type) { case CC_VEGAS_RATE: if (!IN_RECOVERY(CCV(ccv, t_flags))) { CCV(ccv, snd_cwnd) = max(2 * CCV(ccv, t_maxseg), CCV(ccv, snd_cwnd) - CCV(ccv, t_maxseg)); if (CCV(ccv, snd_cwnd) < CCV(ccv, snd_ssthresh)) /* Exit slow start. */ CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd); } break; default: newreno_cc_cong_signal(ccv, signal_type); } if (IN_RECOVERY(CCV(ccv, t_flags)) && !presignalrecov) vegas_data->slow_start_toggle = (CCV(ccv, snd_cwnd) < CCV(ccv, snd_ssthresh)) ? 1 : 0; } static void vegas_conn_init(struct cc_var *ccv) { struct vegas *vegas_data; vegas_data = ccv->cc_data; vegas_data->slow_start_toggle = 1; } static int vegas_mod_init(void) { ertt_id = khelp_get_id("ertt"); if (ertt_id <= 0) { printf("%s: h_ertt module not found\n", __func__); return (ENOENT); } return (0); } static int vegas_alpha_handler(SYSCTL_HANDLER_ARGS) { int error; uint32_t new; new = V_vegas_alpha; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr != NULL) { if (new == 0 || new > V_vegas_beta) error = EINVAL; else V_vegas_alpha = new; } return (error); } static int vegas_beta_handler(SYSCTL_HANDLER_ARGS) { int error; uint32_t new; new = V_vegas_beta; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr != NULL) { if (new == 0 || new < V_vegas_alpha) error = EINVAL; else V_vegas_beta = new; } return (error); } SYSCTL_DECL(_net_inet_tcp_cc_vegas); SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, vegas, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, "Vegas related settings"); SYSCTL_PROC(_net_inet_tcp_cc_vegas, OID_AUTO, alpha, CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(vegas_alpha), 1, &vegas_alpha_handler, "IU", "vegas alpha, specified as number of \"buffers\" (0 < alpha < beta)"); SYSCTL_PROC(_net_inet_tcp_cc_vegas, OID_AUTO, beta, CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(vegas_beta), 3, &vegas_beta_handler, "IU", "vegas beta, specified as number of \"buffers\" (0 < alpha < beta)"); DECLARE_CC_MODULE(vegas, &vegas_cc_algo); MODULE_VERSION(vegas, 2); MODULE_DEPEND(vegas, ertt, 1, 1, 1); diff --git a/sys/netinet/tcp_debug.c b/sys/netinet/tcp_debug.c index 0708d3ebb29d..e00dc095b01f 100644 --- a/sys/netinet/tcp_debug.c +++ b/sys/netinet/tcp_debug.c @@ -1,224 +1,225 @@ /*- * 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_debug.c 8.1 (Berkeley) 6/10/93 */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_tcpdebug.h" #ifdef TCPDEBUG /* load symbolic names */ #define PRUREQUESTS #define TCPSTATES #define TCPTIMERS #define TANAMES #endif #include #include #include #include #include #include #include #include +#include #include #include #ifdef INET6 #include #endif #include #include #include #include #include #include #include #ifdef TCPDEBUG static int tcpconsdebug = 0; #endif /* * Global ring buffer of TCP debugging state. Each entry captures a snapshot * of TCP connection state at any given moment. tcp_debx addresses at the * next available slot. There is no explicit export of this data structure; * it will be read via /dev/kmem by debugging tools. */ static struct tcp_debug tcp_debug[TCP_NDEBUG]; static int tcp_debx; /* * All global state is protected by tcp_debug_mtx; tcp_trace() is split into * two parts, one of which saves connection and other state into the global * array (locked by tcp_debug_mtx). */ struct mtx tcp_debug_mtx; MTX_SYSINIT(tcp_debug_mtx, &tcp_debug_mtx, "tcp_debug_mtx", MTX_DEF); /* * Save TCP state at a given moment; optionally, both tcpcb and TCP packet * header state will be saved. */ void tcp_trace(short act, short ostate, struct tcpcb *tp, void *ipgen, struct tcphdr *th, int req) { #ifdef INET6 int isipv6; #endif /* INET6 */ tcp_seq seq, ack; int len, flags; struct tcp_debug *td; mtx_lock(&tcp_debug_mtx); td = &tcp_debug[tcp_debx++]; if (tcp_debx == TCP_NDEBUG) tcp_debx = 0; bzero(td, sizeof(*td)); #ifdef INET6 isipv6 = (ipgen != NULL && ((struct ip *)ipgen)->ip_v == 6) ? 1 : 0; #endif /* INET6 */ td->td_family = #ifdef INET6 (isipv6 != 0) ? AF_INET6 : #endif AF_INET; #ifdef INET td->td_time = iptime(); #endif td->td_act = act; td->td_ostate = ostate; td->td_tcb = (caddr_t)tp; if (tp != NULL) td->td_cb = *tp; if (ipgen != NULL) { switch (td->td_family) { #ifdef INET case AF_INET: bcopy(ipgen, &td->td_ti.ti_i, sizeof(td->td_ti.ti_i)); break; #endif #ifdef INET6 case AF_INET6: bcopy(ipgen, td->td_ip6buf, sizeof(td->td_ip6buf)); break; #endif } } if (th != NULL) { switch (td->td_family) { #ifdef INET case AF_INET: td->td_ti.ti_t = *th; break; #endif #ifdef INET6 case AF_INET6: td->td_ti6.th = *th; break; #endif } } td->td_req = req; mtx_unlock(&tcp_debug_mtx); #ifdef TCPDEBUG if (tcpconsdebug == 0) return; if (tp != NULL) printf("%p %s:", tp, tcpstates[ostate]); else printf("???????? "); printf("%s ", tanames[act]); switch (act) { case TA_INPUT: case TA_OUTPUT: case TA_DROP: if (ipgen == NULL || th == NULL) break; seq = th->th_seq; ack = th->th_ack; len = #ifdef INET6 isipv6 ? ntohs(((struct ip6_hdr *)ipgen)->ip6_plen) : #endif ntohs(((struct ip *)ipgen)->ip_len); if (act == TA_OUTPUT) { seq = ntohl(seq); ack = ntohl(ack); } if (act == TA_OUTPUT) len -= sizeof (struct tcphdr); if (len) printf("[%x..%x)", seq, seq+len); else printf("%x", seq); printf("@%x, urp=%x", ack, th->th_urp); flags = tcp_get_flags(th); if (flags) { char *cp = "<"; #define pf(f) { \ if (tcp_get_flags(th) & TH_##f) { \ printf("%s%s", cp, #f); \ cp = ","; \ } \ } pf(SYN); pf(ACK); pf(FIN); pf(RST); pf(PUSH); pf(URG); printf(">"); } break; case TA_USER: printf("%s", prurequests[req&0xff]); if ((req & 0xff) == PRU_SLOWTIMO) printf("<%s>", tcptimers[req>>8]); break; } if (tp != NULL) printf(" -> %s", tcpstates[tp->t_state]); /* print out internal state of tp !?! */ printf("\n"); if (tp == NULL) return; printf( "\trcv_(nxt,wnd,up) (%lx,%lx,%lx) snd_(una,nxt,max) (%lx,%lx,%lx)\n", (u_long)tp->rcv_nxt, (u_long)tp->rcv_wnd, (u_long)tp->rcv_up, (u_long)tp->snd_una, (u_long)tp->snd_nxt, (u_long)tp->snd_max); printf("\tsnd_(wl1,wl2,wnd) (%lx,%lx,%lx)\n", (u_long)tp->snd_wl1, (u_long)tp->snd_wl2, (u_long)tp->snd_wnd); #endif /* TCPDEBUG */ } diff --git a/sys/netinet/tcp_hpts.c b/sys/netinet/tcp_hpts.c index 57114706baa3..d3e923406784 100644 --- a/sys/netinet/tcp_hpts.c +++ b/sys/netinet/tcp_hpts.c @@ -1,2043 +1,2046 @@ /*- * Copyright (c) 2016-2018 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_rss.h" #include "opt_tcpdebug.h" /** * Some notes about usage. * * The tcp_hpts system is designed to provide a high precision timer * system for tcp. Its main purpose is to provide a mechanism for * pacing packets out onto the wire. It can be used in two ways * by a given TCP stack (and those two methods can be used simultaneously). * * First, and probably the main thing its used by Rack and BBR, it can * be used to call tcp_output() of a transport stack at some time in the future. * The normal way this is done is that tcp_output() of the stack schedules * itself to be called again by calling tcp_hpts_insert(tcpcb, slot). The * slot is the time from now that the stack wants to be called but it * must be converted to tcp_hpts's notion of slot. This is done with * one of the macros HPTS_MS_TO_SLOTS or HPTS_USEC_TO_SLOTS. So a typical * call from the tcp_output() routine might look like: * * tcp_hpts_insert(tp, HPTS_USEC_TO_SLOTS(550)); * * The above would schedule tcp_ouput() to be called in 550 useconds. * Note that if using this mechanism the stack will want to add near * its top a check to prevent unwanted calls (from user land or the * arrival of incoming ack's). So it would add something like: * * if (tcp_in_hpts(inp)) * return; * * to prevent output processing until the time alotted has gone by. * Of course this is a bare bones example and the stack will probably * have more consideration then just the above. * * In order to run input queued segments from the HPTS context the * tcp stack must define an input function for * tfb_do_queued_segments(). This function understands * how to dequeue a array of packets that were input and * knows how to call the correct processing routine. * * Locking in this is important as well so most likely the * stack will need to define the tfb_do_segment_nounlock() * splitting tfb_do_segment() into two parts. The main processing * part that does not unlock the INP and returns a value of 1 or 0. * It returns 0 if all is well and the lock was not released. It * returns 1 if we had to destroy the TCB (a reset received etc). * The remains of tfb_do_segment() then become just a simple call * to the tfb_do_segment_nounlock() function and check the return * code and possibly unlock. * * The stack must also set the flag on the INP that it supports this * feature i.e. INP_SUPPORTS_MBUFQ. The LRO code recoginizes * this flag as well and will queue packets when it is set. * There are other flags as well INP_MBUF_QUEUE_READY and * INP_DONT_SACK_QUEUE. The first flag tells the LRO code * that we are in the pacer for output so there is no * need to wake up the hpts system to get immediate * input. The second tells the LRO code that its okay * if a SACK arrives you can still defer input and let * the current hpts timer run (this is usually set when * a rack timer is up so we know SACK's are happening * on the connection already and don't want to wakeup yet). * * There is a common functions within the rack_bbr_common code * version i.e. ctf_do_queued_segments(). This function * knows how to take the input queue of packets from * tp->t_in_pkts and process them digging out * all the arguments, calling any bpf tap and * calling into tfb_do_segment_nounlock(). The common * function (ctf_do_queued_segments()) requires that * you have defined the tfb_do_segment_nounlock() as * described above. */ #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 #ifdef RSS #include #include #endif #define TCPSTATES /* for logging */ #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 #ifdef tcpdebug #include #endif /* tcpdebug */ #ifdef tcp_offload #include #endif /* * The hpts uses a 102400 wheel. The wheel * defines the time in 10 usec increments (102400 x 10). * This gives a range of 10usec - 1024ms to place * an entry within. If the user requests more than * 1.024 second, a remaineder is attached and the hpts * when seeing the remainder will re-insert the * inpcb forward in time from where it is until * the remainder is zero. */ #define NUM_OF_HPTSI_SLOTS 102400 /* Each hpts has its own p_mtx which is used for locking */ #define HPTS_MTX_ASSERT(hpts) mtx_assert(&(hpts)->p_mtx, MA_OWNED) #define HPTS_LOCK(hpts) mtx_lock(&(hpts)->p_mtx) #define HPTS_UNLOCK(hpts) mtx_unlock(&(hpts)->p_mtx) struct tcp_hpts_entry { /* Cache line 0x00 */ struct mtx p_mtx; /* Mutex for hpts */ struct timeval p_mysleep; /* Our min sleep time */ uint64_t syscall_cnt; uint64_t sleeping; /* What the actual sleep was (if sleeping) */ uint16_t p_hpts_active; /* Flag that says hpts is awake */ uint8_t p_wheel_complete; /* have we completed the wheel arc walk? */ uint32_t p_curtick; /* Tick in 10 us the hpts is going to */ uint32_t p_runningslot; /* Current tick we are at if we are running */ uint32_t p_prev_slot; /* Previous slot we were on */ uint32_t p_cur_slot; /* Current slot in wheel hpts is draining */ uint32_t p_nxt_slot; /* The next slot outside the current range of * slots that the hpts is running on. */ int32_t p_on_queue_cnt; /* Count on queue in this hpts */ uint32_t p_lasttick; /* Last tick before the current one */ uint8_t p_direct_wake :1, /* boolean */ p_on_min_sleep:1, /* boolean */ p_hpts_wake_scheduled:1, /* boolean */ p_avail:5; uint8_t p_fill[3]; /* Fill to 32 bits */ /* Cache line 0x40 */ struct hptsh { TAILQ_HEAD(, inpcb) head; uint32_t count; uint32_t gencnt; } *p_hptss; /* Hptsi wheel */ uint32_t p_hpts_sleep_time; /* Current sleep interval having a max * of 255ms */ uint32_t overidden_sleep; /* what was overrided by min-sleep for logging */ uint32_t saved_lasttick; /* for logging */ uint32_t saved_curtick; /* for logging */ uint32_t saved_curslot; /* for logging */ uint32_t saved_prev_slot; /* for logging */ uint32_t p_delayed_by; /* How much were we delayed by */ /* Cache line 0x80 */ struct sysctl_ctx_list hpts_ctx; struct sysctl_oid *hpts_root; struct intr_event *ie; void *ie_cookie; uint16_t p_num; /* The hpts number one per cpu */ uint16_t p_cpu; /* The hpts CPU */ /* There is extra space in here */ /* Cache line 0x100 */ struct callout co __aligned(CACHE_LINE_SIZE); } __aligned(CACHE_LINE_SIZE); static struct tcp_hptsi { struct cpu_group **grps; struct tcp_hpts_entry **rp_ent; /* Array of hptss */ uint32_t *cts_last_ran; uint32_t grp_cnt; uint32_t rp_num_hptss; /* Number of hpts threads */ } tcp_pace; MALLOC_DEFINE(M_TCPHPTS, "tcp_hpts", "TCP hpts"); #ifdef RSS static int tcp_bind_threads = 1; #else static int tcp_bind_threads = 2; #endif static int tcp_use_irq_cpu = 0; static uint32_t *cts_last_ran; static int hpts_does_tp_logging = 0; static int32_t tcp_hptsi(struct tcp_hpts_entry *hpts, int from_callout); static void tcp_hpts_thread(void *ctx); static void tcp_init_hptsi(void *st); int32_t tcp_min_hptsi_time = DEFAULT_MIN_SLEEP; static int conn_cnt_thresh = DEFAULT_CONNECTION_THESHOLD; static int32_t dynamic_min_sleep = DYNAMIC_MIN_SLEEP; static int32_t dynamic_max_sleep = DYNAMIC_MAX_SLEEP; SYSCTL_NODE(_net_inet_tcp, OID_AUTO, hpts, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "TCP Hpts controls"); SYSCTL_NODE(_net_inet_tcp_hpts, OID_AUTO, stats, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "TCP Hpts statistics"); #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) static int32_t tcp_hpts_precision = 120; static struct hpts_domain_info { int count; int cpu[MAXCPU]; } hpts_domains[MAXMEMDOM]; enum { IHPTS_NONE = 0, IHPTS_ONQUEUE, IHPTS_MOVING, }; counter_u64_t hpts_hopelessly_behind; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, hopeless, CTLFLAG_RD, &hpts_hopelessly_behind, "Number of times hpts could not catch up and was behind hopelessly"); counter_u64_t hpts_loops; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, loops, CTLFLAG_RD, &hpts_loops, "Number of times hpts had to loop to catch up"); counter_u64_t back_tosleep; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, no_tcbsfound, CTLFLAG_RD, &back_tosleep, "Number of times hpts found no tcbs"); counter_u64_t combined_wheel_wrap; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, comb_wheel_wrap, CTLFLAG_RD, &combined_wheel_wrap, "Number of times the wheel lagged enough to have an insert see wrap"); counter_u64_t wheel_wrap; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, wheel_wrap, CTLFLAG_RD, &wheel_wrap, "Number of times the wheel lagged enough to have an insert see wrap"); counter_u64_t hpts_direct_call; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, direct_call, CTLFLAG_RD, &hpts_direct_call, "Number of times hpts was called by syscall/trap or other entry"); counter_u64_t hpts_wake_timeout; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, timeout_wakeup, CTLFLAG_RD, &hpts_wake_timeout, "Number of times hpts threads woke up via the callout expiring"); counter_u64_t hpts_direct_awakening; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, direct_awakening, CTLFLAG_RD, &hpts_direct_awakening, "Number of times hpts threads woke up via the callout expiring"); counter_u64_t hpts_back_tosleep; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, back_tosleep, CTLFLAG_RD, &hpts_back_tosleep, "Number of times hpts threads woke up via the callout expiring and went back to sleep no work"); counter_u64_t cpu_uses_flowid; counter_u64_t cpu_uses_random; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, cpusel_flowid, CTLFLAG_RD, &cpu_uses_flowid, "Number of times when setting cpuid we used the flowid field"); SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, cpusel_random, CTLFLAG_RD, &cpu_uses_random, "Number of times when setting cpuid we used the a random value"); TUNABLE_INT("net.inet.tcp.bind_hptss", &tcp_bind_threads); TUNABLE_INT("net.inet.tcp.use_irq", &tcp_use_irq_cpu); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, bind_hptss, CTLFLAG_RD, &tcp_bind_threads, 2, "Thread Binding tunable"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, use_irq, CTLFLAG_RD, &tcp_use_irq_cpu, 0, "Use of irq CPU tunable"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, precision, CTLFLAG_RW, &tcp_hpts_precision, 120, "Value for PRE() precision of callout"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, cnt_thresh, CTLFLAG_RW, &conn_cnt_thresh, 0, "How many connections (below) make us use the callout based mechanism"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, logging, CTLFLAG_RW, &hpts_does_tp_logging, 0, "Do we add to any tp that has logging on pacer logs"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, dyn_minsleep, CTLFLAG_RW, &dynamic_min_sleep, 250, "What is the dynamic minsleep value?"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, dyn_maxsleep, CTLFLAG_RW, &dynamic_max_sleep, 5000, "What is the dynamic maxsleep value?"); static int32_t max_pacer_loops = 10; SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, loopmax, CTLFLAG_RW, &max_pacer_loops, 10, "What is the maximum number of times the pacer will loop trying to catch up"); #define HPTS_MAX_SLEEP_ALLOWED (NUM_OF_HPTSI_SLOTS/2) static uint32_t hpts_sleep_max = HPTS_MAX_SLEEP_ALLOWED; static int sysctl_net_inet_tcp_hpts_max_sleep(SYSCTL_HANDLER_ARGS) { int error; uint32_t new; new = hpts_sleep_max; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr) { if ((new < (dynamic_min_sleep/HPTS_TICKS_PER_SLOT)) || (new > HPTS_MAX_SLEEP_ALLOWED)) error = EINVAL; else hpts_sleep_max = new; } return (error); } static int sysctl_net_inet_tcp_hpts_min_sleep(SYSCTL_HANDLER_ARGS) { int error; uint32_t new; new = tcp_min_hptsi_time; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr) { if (new < LOWEST_SLEEP_ALLOWED) error = EINVAL; else tcp_min_hptsi_time = new; } return (error); } SYSCTL_PROC(_net_inet_tcp_hpts, OID_AUTO, maxsleep, CTLTYPE_UINT | CTLFLAG_RW, &hpts_sleep_max, 0, &sysctl_net_inet_tcp_hpts_max_sleep, "IU", "Maximum time hpts will sleep in slots"); SYSCTL_PROC(_net_inet_tcp_hpts, OID_AUTO, minsleep, CTLTYPE_UINT | CTLFLAG_RW, &tcp_min_hptsi_time, 0, &sysctl_net_inet_tcp_hpts_min_sleep, "IU", "The minimum time the hpts must sleep before processing more slots"); static int ticks_indicate_more_sleep = TICKS_INDICATE_MORE_SLEEP; static int ticks_indicate_less_sleep = TICKS_INDICATE_LESS_SLEEP; static int tcp_hpts_no_wake_over_thresh = 1; SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, more_sleep, CTLFLAG_RW, &ticks_indicate_more_sleep, 0, "If we only process this many or less on a timeout, we need longer sleep on the next callout"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, less_sleep, CTLFLAG_RW, &ticks_indicate_less_sleep, 0, "If we process this many or more on a timeout, we need less sleep on the next callout"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, nowake_over_thresh, CTLFLAG_RW, &tcp_hpts_no_wake_over_thresh, 0, "When we are over the threshold on the pacer do we prohibit wakeups?"); static void tcp_hpts_log(struct tcp_hpts_entry *hpts, struct tcpcb *tp, struct timeval *tv, int slots_to_run, int idx, int from_callout) { union tcp_log_stackspecific log; /* * Unused logs are * 64 bit - delRate, rttProp, bw_inuse * 16 bit - cwnd_gain * 8 bit - bbr_state, bbr_substate, inhpts; */ memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = hpts->p_nxt_slot; log.u_bbr.flex2 = hpts->p_cur_slot; log.u_bbr.flex3 = hpts->p_prev_slot; log.u_bbr.flex4 = idx; log.u_bbr.flex5 = hpts->p_curtick; log.u_bbr.flex6 = hpts->p_on_queue_cnt; log.u_bbr.flex7 = hpts->p_cpu; log.u_bbr.flex8 = (uint8_t)from_callout; log.u_bbr.inflight = slots_to_run; log.u_bbr.applimited = hpts->overidden_sleep; log.u_bbr.delivered = hpts->saved_curtick; log.u_bbr.timeStamp = tcp_tv_to_usectick(tv); log.u_bbr.epoch = hpts->saved_curslot; log.u_bbr.lt_epoch = hpts->saved_prev_slot; log.u_bbr.pkts_out = hpts->p_delayed_by; log.u_bbr.lost = hpts->p_hpts_sleep_time; log.u_bbr.pacing_gain = hpts->p_cpu; log.u_bbr.pkt_epoch = hpts->p_runningslot; log.u_bbr.use_lt_bw = 1; TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv, &tptosocket(tp)->so_snd, BBR_LOG_HPTSDIAG, 0, 0, &log, false, tv); } static void tcp_wakehpts(struct tcp_hpts_entry *hpts) { HPTS_MTX_ASSERT(hpts); if (tcp_hpts_no_wake_over_thresh && (hpts->p_on_queue_cnt >= conn_cnt_thresh)) { hpts->p_direct_wake = 0; return; } if (hpts->p_hpts_wake_scheduled == 0) { hpts->p_hpts_wake_scheduled = 1; swi_sched(hpts->ie_cookie, 0); } } static void hpts_timeout_swi(void *arg) { struct tcp_hpts_entry *hpts; hpts = (struct tcp_hpts_entry *)arg; swi_sched(hpts->ie_cookie, 0); } static void inp_hpts_insert(struct inpcb *inp, struct tcp_hpts_entry *hpts) { struct hptsh *hptsh; INP_WLOCK_ASSERT(inp); HPTS_MTX_ASSERT(hpts); MPASS(hpts->p_cpu == inp->inp_hpts_cpu); MPASS(!(inp->inp_flags & INP_DROPPED)); hptsh = &hpts->p_hptss[inp->inp_hptsslot]; if (inp->inp_in_hpts == IHPTS_NONE) { inp->inp_in_hpts = IHPTS_ONQUEUE; in_pcbref(inp); } else if (inp->inp_in_hpts == IHPTS_MOVING) { inp->inp_in_hpts = IHPTS_ONQUEUE; } else MPASS(inp->inp_in_hpts == IHPTS_ONQUEUE); inp->inp_hpts_gencnt = hptsh->gencnt; TAILQ_INSERT_TAIL(&hptsh->head, inp, inp_hpts); hptsh->count++; hpts->p_on_queue_cnt++; } static struct tcp_hpts_entry * tcp_hpts_lock(struct inpcb *inp) { struct tcp_hpts_entry *hpts; INP_LOCK_ASSERT(inp); hpts = tcp_pace.rp_ent[inp->inp_hpts_cpu]; HPTS_LOCK(hpts); return (hpts); } static void inp_hpts_release(struct inpcb *inp) { bool released __diagused; inp->inp_in_hpts = IHPTS_NONE; released = in_pcbrele_wlocked(inp); MPASS(released == false); } /* * Called normally with the INP_LOCKED but it * does not matter, the hpts lock is the key * but the lock order allows us to hold the * INP lock and then get the hpts lock. */ void tcp_hpts_remove(struct inpcb *inp) { struct tcp_hpts_entry *hpts; struct hptsh *hptsh; INP_WLOCK_ASSERT(inp); hpts = tcp_hpts_lock(inp); if (inp->inp_in_hpts == IHPTS_ONQUEUE) { hptsh = &hpts->p_hptss[inp->inp_hptsslot]; inp->inp_hpts_request = 0; if (__predict_true(inp->inp_hpts_gencnt == hptsh->gencnt)) { TAILQ_REMOVE(&hptsh->head, inp, inp_hpts); MPASS(hptsh->count > 0); hptsh->count--; MPASS(hpts->p_on_queue_cnt > 0); hpts->p_on_queue_cnt--; inp_hpts_release(inp); } else { /* * tcp_hptsi() now owns the TAILQ head of this inp. * Can't TAILQ_REMOVE, just mark it. */ #ifdef INVARIANTS struct inpcb *tmp; TAILQ_FOREACH(tmp, &hptsh->head, inp_hpts) MPASS(tmp != inp); #endif inp->inp_in_hpts = IHPTS_MOVING; inp->inp_hptsslot = -1; } } else if (inp->inp_in_hpts == IHPTS_MOVING) { /* * Handle a special race condition: * tcp_hptsi() moves inpcb to detached tailq * tcp_hpts_remove() marks as IHPTS_MOVING, slot = -1 * tcp_hpts_insert() sets slot to a meaningful value * tcp_hpts_remove() again (we are here!), then in_pcbdrop() * tcp_hptsi() finds pcb with meaningful slot and INP_DROPPED */ inp->inp_hptsslot = -1; } HPTS_UNLOCK(hpts); } bool tcp_in_hpts(struct inpcb *inp) { return (inp->inp_in_hpts == IHPTS_ONQUEUE); } static inline int hpts_slot(uint32_t wheel_slot, uint32_t plus) { /* * Given a slot on the wheel, what slot * is that plus ticks out? */ KASSERT(wheel_slot < NUM_OF_HPTSI_SLOTS, ("Invalid tick %u not on wheel", wheel_slot)); return ((wheel_slot + plus) % NUM_OF_HPTSI_SLOTS); } static inline int tick_to_wheel(uint32_t cts_in_wticks) { /* * Given a timestamp in ticks (so by * default to get it to a real time one * would multiply by 10.. i.e the number * of ticks in a slot) map it to our limited * space wheel. */ return (cts_in_wticks % NUM_OF_HPTSI_SLOTS); } static inline int hpts_slots_diff(int prev_slot, int slot_now) { /* * Given two slots that are someplace * on our wheel. How far are they apart? */ if (slot_now > prev_slot) return (slot_now - prev_slot); else if (slot_now == prev_slot) /* * Special case, same means we can go all of our * wheel less one slot. */ return (NUM_OF_HPTSI_SLOTS - 1); else return ((NUM_OF_HPTSI_SLOTS - prev_slot) + slot_now); } /* * Given a slot on the wheel that is the current time * mapped to the wheel (wheel_slot), what is the maximum * distance forward that can be obtained without * wrapping past either prev_slot or running_slot * depending on the htps state? Also if passed * a uint32_t *, fill it with the slot location. * * Note if you do not give this function the current * time (that you think it is) mapped to the wheel slot * then the results will not be what you expect and * could lead to invalid inserts. */ static inline int32_t max_slots_available(struct tcp_hpts_entry *hpts, uint32_t wheel_slot, uint32_t *target_slot) { uint32_t dis_to_travel, end_slot, pacer_to_now, avail_on_wheel; if ((hpts->p_hpts_active == 1) && (hpts->p_wheel_complete == 0)) { end_slot = hpts->p_runningslot; /* Back up one tick */ if (end_slot == 0) end_slot = NUM_OF_HPTSI_SLOTS - 1; else end_slot--; if (target_slot) *target_slot = end_slot; } else { /* * For the case where we are * not active, or we have * completed the pass over * the wheel, we can use the * prev tick and subtract one from it. This puts us * as far out as possible on the wheel. */ end_slot = hpts->p_prev_slot; if (end_slot == 0) end_slot = NUM_OF_HPTSI_SLOTS - 1; else end_slot--; if (target_slot) *target_slot = end_slot; /* * Now we have close to the full wheel left minus the * time it has been since the pacer went to sleep. Note * that wheel_tick, passed in, should be the current time * from the perspective of the caller, mapped to the wheel. */ if (hpts->p_prev_slot != wheel_slot) dis_to_travel = hpts_slots_diff(hpts->p_prev_slot, wheel_slot); else dis_to_travel = 1; /* * dis_to_travel in this case is the space from when the * pacer stopped (p_prev_slot) and where our wheel_slot * is now. To know how many slots we can put it in we * subtract from the wheel size. We would not want * to place something after p_prev_slot or it will * get ran too soon. */ return (NUM_OF_HPTSI_SLOTS - dis_to_travel); } /* * So how many slots are open between p_runningslot -> p_cur_slot * that is what is currently un-available for insertion. Special * case when we are at the last slot, this gets 1, so that * the answer to how many slots are available is all but 1. */ if (hpts->p_runningslot == hpts->p_cur_slot) dis_to_travel = 1; else dis_to_travel = hpts_slots_diff(hpts->p_runningslot, hpts->p_cur_slot); /* * How long has the pacer been running? */ if (hpts->p_cur_slot != wheel_slot) { /* The pacer is a bit late */ pacer_to_now = hpts_slots_diff(hpts->p_cur_slot, wheel_slot); } else { /* The pacer is right on time, now == pacers start time */ pacer_to_now = 0; } /* * To get the number left we can insert into we simply * subract the distance the pacer has to run from how * many slots there are. */ avail_on_wheel = NUM_OF_HPTSI_SLOTS - dis_to_travel; /* * Now how many of those we will eat due to the pacer's * time (p_cur_slot) of start being behind the * real time (wheel_slot)? */ if (avail_on_wheel <= pacer_to_now) { /* * Wheel wrap, we can't fit on the wheel, that * is unusual the system must be way overloaded! * Insert into the assured slot, and return special * "0". */ counter_u64_add(combined_wheel_wrap, 1); *target_slot = hpts->p_nxt_slot; return (0); } else { /* * We know how many slots are open * on the wheel (the reverse of what * is left to run. Take away the time * the pacer started to now (wheel_slot) * and that tells you how many slots are * open that can be inserted into that won't * be touched by the pacer until later. */ return (avail_on_wheel - pacer_to_now); } } #ifdef INVARIANTS static void check_if_slot_would_be_wrong(struct tcp_hpts_entry *hpts, struct inpcb *inp, uint32_t inp_hptsslot, int line) { /* * Sanity checks for the pacer with invariants * on insert. */ KASSERT(inp_hptsslot < NUM_OF_HPTSI_SLOTS, ("hpts:%p inp:%p slot:%d > max", hpts, inp, inp_hptsslot)); if ((hpts->p_hpts_active) && (hpts->p_wheel_complete == 0)) { /* * If the pacer is processing a arc * of the wheel, we need to make * sure we are not inserting within * that arc. */ int distance, yet_to_run; distance = hpts_slots_diff(hpts->p_runningslot, inp_hptsslot); if (hpts->p_runningslot != hpts->p_cur_slot) yet_to_run = hpts_slots_diff(hpts->p_runningslot, hpts->p_cur_slot); else yet_to_run = 0; /* processing last slot */ KASSERT(yet_to_run <= distance, ("hpts:%p inp:%p slot:%d distance:%d yet_to_run:%d rs:%d cs:%d", hpts, inp, inp_hptsslot, distance, yet_to_run, hpts->p_runningslot, hpts->p_cur_slot)); } } #endif uint32_t tcp_hpts_insert_diag(struct inpcb *inp, uint32_t slot, int32_t line, struct hpts_diag *diag) { struct tcp_hpts_entry *hpts; struct timeval tv; uint32_t slot_on, wheel_cts, last_slot, need_new_to = 0; int32_t wheel_slot, maxslots; bool need_wakeup = false; INP_WLOCK_ASSERT(inp); MPASS(!tcp_in_hpts(inp)); MPASS(!(inp->inp_flags & INP_DROPPED)); /* * We now return the next-slot the hpts will be on, beyond its * current run (if up) or where it was when it stopped if it is * sleeping. */ hpts = tcp_hpts_lock(inp); microuptime(&tv); if (diag) { memset(diag, 0, sizeof(struct hpts_diag)); diag->p_hpts_active = hpts->p_hpts_active; diag->p_prev_slot = hpts->p_prev_slot; diag->p_runningslot = hpts->p_runningslot; diag->p_nxt_slot = hpts->p_nxt_slot; diag->p_cur_slot = hpts->p_cur_slot; diag->p_curtick = hpts->p_curtick; diag->p_lasttick = hpts->p_lasttick; diag->slot_req = slot; diag->p_on_min_sleep = hpts->p_on_min_sleep; diag->hpts_sleep_time = hpts->p_hpts_sleep_time; } if (slot == 0) { /* Ok we need to set it on the hpts in the current slot */ inp->inp_hpts_request = 0; if ((hpts->p_hpts_active == 0) || (hpts->p_wheel_complete)) { /* * A sleeping hpts we want in next slot to run * note that in this state p_prev_slot == p_cur_slot */ inp->inp_hptsslot = hpts_slot(hpts->p_prev_slot, 1); if ((hpts->p_on_min_sleep == 0) && (hpts->p_hpts_active == 0)) need_wakeup = true; } else inp->inp_hptsslot = hpts->p_runningslot; if (__predict_true(inp->inp_in_hpts != IHPTS_MOVING)) inp_hpts_insert(inp, hpts); if (need_wakeup) { /* * Activate the hpts if it is sleeping and its * timeout is not 1. */ hpts->p_direct_wake = 1; tcp_wakehpts(hpts); } slot_on = hpts->p_nxt_slot; HPTS_UNLOCK(hpts); return (slot_on); } /* Get the current time relative to the wheel */ wheel_cts = tcp_tv_to_hptstick(&tv); /* Map it onto the wheel */ wheel_slot = tick_to_wheel(wheel_cts); /* Now what's the max we can place it at? */ maxslots = max_slots_available(hpts, wheel_slot, &last_slot); if (diag) { diag->wheel_slot = wheel_slot; diag->maxslots = maxslots; diag->wheel_cts = wheel_cts; } if (maxslots == 0) { /* The pacer is in a wheel wrap behind, yikes! */ if (slot > 1) { /* * Reduce by 1 to prevent a forever loop in * case something else is wrong. Note this * probably does not hurt because the pacer * if its true is so far behind we will be * > 1second late calling anyway. */ slot--; } inp->inp_hptsslot = last_slot; inp->inp_hpts_request = slot; } else if (maxslots >= slot) { /* It all fits on the wheel */ inp->inp_hpts_request = 0; inp->inp_hptsslot = hpts_slot(wheel_slot, slot); } else { /* It does not fit */ inp->inp_hpts_request = slot - maxslots; inp->inp_hptsslot = last_slot; } if (diag) { diag->slot_remaining = inp->inp_hpts_request; diag->inp_hptsslot = inp->inp_hptsslot; } #ifdef INVARIANTS check_if_slot_would_be_wrong(hpts, inp, inp->inp_hptsslot, line); #endif if (__predict_true(inp->inp_in_hpts != IHPTS_MOVING)) inp_hpts_insert(inp, hpts); if ((hpts->p_hpts_active == 0) && (inp->inp_hpts_request == 0) && (hpts->p_on_min_sleep == 0)) { /* * The hpts is sleeping and NOT on a minimum * sleep time, we need to figure out where * it will wake up at and if we need to reschedule * its time-out. */ uint32_t have_slept, yet_to_sleep; /* Now do we need to restart the hpts's timer? */ have_slept = hpts_slots_diff(hpts->p_prev_slot, wheel_slot); if (have_slept < hpts->p_hpts_sleep_time) yet_to_sleep = hpts->p_hpts_sleep_time - have_slept; else { /* We are over-due */ yet_to_sleep = 0; need_wakeup = 1; } if (diag) { diag->have_slept = have_slept; diag->yet_to_sleep = yet_to_sleep; } if (yet_to_sleep && (yet_to_sleep > slot)) { /* * We need to reschedule the hpts's time-out. */ hpts->p_hpts_sleep_time = slot; need_new_to = slot * HPTS_TICKS_PER_SLOT; } } /* * Now how far is the hpts sleeping to? if active is 1, its * up and ticking we do nothing, otherwise we may need to * reschedule its callout if need_new_to is set from above. */ if (need_wakeup) { hpts->p_direct_wake = 1; tcp_wakehpts(hpts); if (diag) { diag->need_new_to = 0; diag->co_ret = 0xffff0000; } } else if (need_new_to) { int32_t co_ret; struct timeval tv; sbintime_t sb; tv.tv_sec = 0; tv.tv_usec = 0; while (need_new_to > HPTS_USEC_IN_SEC) { tv.tv_sec++; need_new_to -= HPTS_USEC_IN_SEC; } tv.tv_usec = need_new_to; sb = tvtosbt(tv); co_ret = callout_reset_sbt_on(&hpts->co, sb, 0, hpts_timeout_swi, hpts, hpts->p_cpu, (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); if (diag) { diag->need_new_to = need_new_to; diag->co_ret = co_ret; } } slot_on = hpts->p_nxt_slot; HPTS_UNLOCK(hpts); return (slot_on); } uint16_t hpts_random_cpu(struct inpcb *inp){ /* * No flow type set distribute the load randomly. */ uint16_t cpuid; uint32_t ran; /* * Shortcut if it is already set. XXXGL: does it happen? */ if (inp->inp_hpts_cpu_set) { return (inp->inp_hpts_cpu); } /* Nothing set use a random number */ ran = arc4random(); cpuid = (((ran & 0xffff) % mp_ncpus) % tcp_pace.rp_num_hptss); return (cpuid); } static uint16_t hpts_cpuid(struct inpcb *inp, int *failed) { u_int cpuid; #ifdef NUMA struct hpts_domain_info *di; #endif *failed = 0; if (inp->inp_hpts_cpu_set) { return (inp->inp_hpts_cpu); } /* * If we are using the irq cpu set by LRO or * the driver then it overrides all other domains. */ if (tcp_use_irq_cpu) { if (inp->inp_irq_cpu_set == 0) { *failed = 1; return(0); } return(inp->inp_irq_cpu); } /* If one is set the other must be the same */ #ifdef RSS cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype); if (cpuid == NETISR_CPUID_NONE) return (hpts_random_cpu(inp)); else return (cpuid); #endif /* * We don't have a flowid -> cpuid mapping, so cheat and just map * unknown cpuids to curcpu. Not the best, but apparently better * than defaulting to swi 0. */ if (inp->inp_flowtype == M_HASHTYPE_NONE) { counter_u64_add(cpu_uses_random, 1); return (hpts_random_cpu(inp)); } /* * Hash to a thread based on the flowid. If we are using numa, * then restrict the hash to the numa domain where the inp lives. */ #ifdef NUMA if ((vm_ndomains == 1) || (inp->inp_numa_domain == M_NODOM)) { #endif cpuid = inp->inp_flowid % mp_ncpus; #ifdef NUMA } else { /* Hash into the cpu's that use that domain */ di = &hpts_domains[inp->inp_numa_domain]; cpuid = di->cpu[inp->inp_flowid % di->count]; } #endif counter_u64_add(cpu_uses_flowid, 1); return (cpuid); } #ifdef not_longer_used_gleb static void tcp_drop_in_pkts(struct tcpcb *tp) { struct mbuf *m, *n; m = tp->t_in_pkt; if (m) n = m->m_nextpkt; else n = NULL; tp->t_in_pkt = NULL; while (m) { m_freem(m); m = n; if (m) n = m->m_nextpkt; } } #endif static void tcp_hpts_set_max_sleep(struct tcp_hpts_entry *hpts, int wrap_loop_cnt) { uint32_t t = 0, i; if ((hpts->p_on_queue_cnt) && (wrap_loop_cnt < 2)) { /* * Find next slot that is occupied and use that to * be the sleep time. */ for (i = 0, t = hpts_slot(hpts->p_cur_slot, 1); i < NUM_OF_HPTSI_SLOTS; i++) { if (TAILQ_EMPTY(&hpts->p_hptss[t].head) == 0) { break; } t = (t + 1) % NUM_OF_HPTSI_SLOTS; } KASSERT((i != NUM_OF_HPTSI_SLOTS), ("Hpts:%p cnt:%d but none found", hpts, hpts->p_on_queue_cnt)); hpts->p_hpts_sleep_time = min((i + 1), hpts_sleep_max); } else { /* No one on the wheel sleep for all but 400 slots or sleep max */ hpts->p_hpts_sleep_time = hpts_sleep_max; } } static int32_t tcp_hptsi(struct tcp_hpts_entry *hpts, int from_callout) { struct tcpcb *tp; struct inpcb *inp; struct timeval tv; int32_t slots_to_run, i, error; int32_t loop_cnt = 0; int32_t did_prefetch = 0; int32_t prefetch_ninp = 0; int32_t prefetch_tp = 0; int32_t wrap_loop_cnt = 0; int32_t slot_pos_of_endpoint = 0; int32_t orig_exit_slot; int8_t completed_measure = 0, seen_endpoint = 0; HPTS_MTX_ASSERT(hpts); NET_EPOCH_ASSERT(); /* record previous info for any logging */ hpts->saved_lasttick = hpts->p_lasttick; hpts->saved_curtick = hpts->p_curtick; hpts->saved_curslot = hpts->p_cur_slot; hpts->saved_prev_slot = hpts->p_prev_slot; hpts->p_lasttick = hpts->p_curtick; hpts->p_curtick = tcp_gethptstick(&tv); cts_last_ran[hpts->p_num] = tcp_tv_to_usectick(&tv); orig_exit_slot = hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick); if ((hpts->p_on_queue_cnt == 0) || (hpts->p_lasttick == hpts->p_curtick)) { /* * No time has yet passed, * or nothing to do. */ hpts->p_prev_slot = hpts->p_cur_slot; hpts->p_lasttick = hpts->p_curtick; goto no_run; } again: hpts->p_wheel_complete = 0; HPTS_MTX_ASSERT(hpts); slots_to_run = hpts_slots_diff(hpts->p_prev_slot, hpts->p_cur_slot); if (((hpts->p_curtick - hpts->p_lasttick) > ((NUM_OF_HPTSI_SLOTS-1) * HPTS_TICKS_PER_SLOT)) && (hpts->p_on_queue_cnt != 0)) { /* * Wheel wrap is occuring, basically we * are behind and the distance between * run's has spread so much it has exceeded * the time on the wheel (1.024 seconds). This * is ugly and should NOT be happening. We * need to run the entire wheel. We last processed * p_prev_slot, so that needs to be the last slot * we run. The next slot after that should be our * reserved first slot for new, and then starts * the running position. Now the problem is the * reserved "not to yet" place does not exist * and there may be inp's in there that need * running. We can merge those into the * first slot at the head. */ wrap_loop_cnt++; hpts->p_nxt_slot = hpts_slot(hpts->p_prev_slot, 1); hpts->p_runningslot = hpts_slot(hpts->p_prev_slot, 2); /* * Adjust p_cur_slot to be where we are starting from * hopefully we will catch up (fat chance if something * is broken this bad :( ) */ hpts->p_cur_slot = hpts->p_prev_slot; /* * The next slot has guys to run too, and that would * be where we would normally start, lets move them into * the next slot (p_prev_slot + 2) so that we will * run them, the extra 10usecs of late (by being * put behind) does not really matter in this situation. */ TAILQ_FOREACH(inp, &hpts->p_hptss[hpts->p_nxt_slot].head, inp_hpts) { MPASS(inp->inp_hptsslot == hpts->p_nxt_slot); MPASS(inp->inp_hpts_gencnt == hpts->p_hptss[hpts->p_nxt_slot].gencnt); MPASS(inp->inp_in_hpts == IHPTS_ONQUEUE); /* * Update gencnt and nextslot accordingly to match * the new location. This is safe since it takes both * the INP lock and the pacer mutex to change the * inp_hptsslot and inp_hpts_gencnt. */ inp->inp_hpts_gencnt = hpts->p_hptss[hpts->p_runningslot].gencnt; inp->inp_hptsslot = hpts->p_runningslot; } TAILQ_CONCAT(&hpts->p_hptss[hpts->p_runningslot].head, &hpts->p_hptss[hpts->p_nxt_slot].head, inp_hpts); hpts->p_hptss[hpts->p_runningslot].count += hpts->p_hptss[hpts->p_nxt_slot].count; hpts->p_hptss[hpts->p_nxt_slot].count = 0; hpts->p_hptss[hpts->p_nxt_slot].gencnt++; slots_to_run = NUM_OF_HPTSI_SLOTS - 1; counter_u64_add(wheel_wrap, 1); } else { /* * Nxt slot is always one after p_runningslot though * its not used usually unless we are doing wheel wrap. */ hpts->p_nxt_slot = hpts->p_prev_slot; hpts->p_runningslot = hpts_slot(hpts->p_prev_slot, 1); } if (hpts->p_on_queue_cnt == 0) { goto no_one; } for (i = 0; i < slots_to_run; i++) { struct inpcb *inp, *ninp; TAILQ_HEAD(, inpcb) head = TAILQ_HEAD_INITIALIZER(head); struct hptsh *hptsh; uint32_t runningslot; /* * Calculate our delay, if there are no extra ticks there * was not any (i.e. if slots_to_run == 1, no delay). */ hpts->p_delayed_by = (slots_to_run - (i + 1)) * HPTS_TICKS_PER_SLOT; runningslot = hpts->p_runningslot; hptsh = &hpts->p_hptss[runningslot]; TAILQ_SWAP(&head, &hptsh->head, inpcb, inp_hpts); hpts->p_on_queue_cnt -= hptsh->count; hptsh->count = 0; hptsh->gencnt++; HPTS_UNLOCK(hpts); TAILQ_FOREACH_SAFE(inp, &head, inp_hpts, ninp) { bool set_cpu; if (ninp != NULL) { /* We prefetch the next inp if possible */ kern_prefetch(ninp, &prefetch_ninp); prefetch_ninp = 1; } /* For debugging */ if (seen_endpoint == 0) { seen_endpoint = 1; orig_exit_slot = slot_pos_of_endpoint = runningslot; } else if (completed_measure == 0) { /* Record the new position */ orig_exit_slot = runningslot; } INP_WLOCK(inp); if (inp->inp_hpts_cpu_set == 0) { set_cpu = true; } else { set_cpu = false; } if (__predict_false(inp->inp_in_hpts == IHPTS_MOVING)) { if (inp->inp_hptsslot == -1) { inp->inp_in_hpts = IHPTS_NONE; if (in_pcbrele_wlocked(inp) == false) INP_WUNLOCK(inp); } else { HPTS_LOCK(hpts); inp_hpts_insert(inp, hpts); HPTS_UNLOCK(hpts); INP_WUNLOCK(inp); } continue; } MPASS(inp->inp_in_hpts == IHPTS_ONQUEUE); MPASS(!(inp->inp_flags & INP_DROPPED)); KASSERT(runningslot == inp->inp_hptsslot, ("Hpts:%p inp:%p slot mis-aligned %u vs %u", hpts, inp, runningslot, inp->inp_hptsslot)); if (inp->inp_hpts_request) { /* * This guy is deferred out further in time * then our wheel had available on it. * Push him back on the wheel or run it * depending. */ uint32_t maxslots, last_slot, remaining_slots; remaining_slots = slots_to_run - (i + 1); if (inp->inp_hpts_request > remaining_slots) { HPTS_LOCK(hpts); /* * How far out can we go? */ maxslots = max_slots_available(hpts, hpts->p_cur_slot, &last_slot); if (maxslots >= inp->inp_hpts_request) { /* We can place it finally to * be processed. */ inp->inp_hptsslot = hpts_slot( hpts->p_runningslot, inp->inp_hpts_request); inp->inp_hpts_request = 0; } else { /* Work off some more time */ inp->inp_hptsslot = last_slot; inp->inp_hpts_request -= maxslots; } inp_hpts_insert(inp, hpts); HPTS_UNLOCK(hpts); INP_WUNLOCK(inp); continue; } inp->inp_hpts_request = 0; /* Fall through we will so do it now */ } inp_hpts_release(inp); tp = intotcpcb(inp); MPASS(tp); if (set_cpu) { /* * Setup so the next time we will move to * the right CPU. This should be a rare * event. It will sometimes happens when we * are the client side (usually not the * server). Somehow tcp_output() gets called * before the tcp_do_segment() sets the * intial state. This means the r_cpu and * r_hpts_cpu is 0. We get on the hpts, and * then tcp_input() gets called setting up * the r_cpu to the correct value. The hpts * goes off and sees the mis-match. We * simply correct it here and the CPU will * switch to the new hpts nextime the tcb * gets added to the hpts (not this one) * :-) */ tcp_set_hpts(inp); } CURVNET_SET(inp->inp_vnet); /* Lets do any logging that we might want to */ if (hpts_does_tp_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) { tcp_hpts_log(hpts, tp, &tv, slots_to_run, i, from_callout); } if (tp->t_fb_ptr != NULL) { kern_prefetch(tp->t_fb_ptr, &did_prefetch); did_prefetch = 1; } if ((inp->inp_flags2 & INP_SUPPORTS_MBUFQ) && tp->t_in_pkt) { error = (*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0); if (error) { /* The input killed the connection */ goto skip_pacing; } } inp->inp_hpts_calls = 1; error = tcp_output(tp); if (error < 0) goto skip_pacing; inp->inp_hpts_calls = 0; - if (ninp && ninp->inp_ppcb) { + if (ninp) { /* * If we have a nxt inp, see if we can - * prefetch its ppcb. Note this may seem + * prefetch it. Note this may seem * "risky" since we have no locks (other * than the previous inp) and there no * assurance that ninp was not pulled while * we were processing inp and freed. If this * occurred it could mean that either: * * a) Its NULL (which is fine we won't go * here) b) Its valid (which is cool we * will prefetch it) c) The inp got * freed back to the slab which was * reallocated. Then the piece of memory was * re-used and something else (not an * address) is in inp_ppcb. If that occurs * we don't crash, but take a TLB shootdown * performance hit (same as if it was NULL * and we tried to pre-fetch it). * * Considering that the likelyhood of is * quite rare we will take a risk on doing * this. If performance drops after testing * we can always take this out. NB: the * kern_prefetch on amd64 actually has * protection against a bad address now via * the DMAP_() tests. This will prevent the * TLB hit, and instead if occurs just * cause us to load cache with a useless * address (to us). + * + * XXXGL: with tcpcb == inpcb, I'm unsure this + * prefetch is still correct and useful. */ - kern_prefetch(ninp->inp_ppcb, &prefetch_tp); + kern_prefetch(ninp, &prefetch_tp); prefetch_tp = 1; } INP_WUNLOCK(inp); skip_pacing: CURVNET_RESTORE(); } if (seen_endpoint) { /* * We now have a accurate distance between * slot_pos_of_endpoint <-> orig_exit_slot * to tell us how late we were, orig_exit_slot * is where we calculated the end of our cycle to * be when we first entered. */ completed_measure = 1; } HPTS_LOCK(hpts); hpts->p_runningslot++; if (hpts->p_runningslot >= NUM_OF_HPTSI_SLOTS) { hpts->p_runningslot = 0; } } no_one: HPTS_MTX_ASSERT(hpts); hpts->p_delayed_by = 0; /* * Check to see if we took an excess amount of time and need to run * more ticks (if we did not hit eno-bufs). */ hpts->p_prev_slot = hpts->p_cur_slot; hpts->p_lasttick = hpts->p_curtick; if ((from_callout == 0) || (loop_cnt > max_pacer_loops)) { /* * Something is serious slow we have * looped through processing the wheel * and by the time we cleared the * needs to run max_pacer_loops time * we still needed to run. That means * the system is hopelessly behind and * can never catch up :( * * We will just lie to this thread * and let it thing p_curtick is * correct. When it next awakens * it will find itself further behind. */ if (from_callout) counter_u64_add(hpts_hopelessly_behind, 1); goto no_run; } hpts->p_curtick = tcp_gethptstick(&tv); hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick); if (seen_endpoint == 0) { /* We saw no endpoint but we may be looping */ orig_exit_slot = hpts->p_cur_slot; } if ((wrap_loop_cnt < 2) && (hpts->p_lasttick != hpts->p_curtick)) { counter_u64_add(hpts_loops, 1); loop_cnt++; goto again; } no_run: cts_last_ran[hpts->p_num] = tcp_tv_to_usectick(&tv); /* * Set flag to tell that we are done for * any slot input that happens during * input. */ hpts->p_wheel_complete = 1; /* * Now did we spend too long running input and need to run more ticks? * Note that if wrap_loop_cnt < 2 then we should have the conditions * in the KASSERT's true. But if the wheel is behind i.e. wrap_loop_cnt * is greater than 2, then the condtion most likely are *not* true. * Also if we are called not from the callout, we don't run the wheel * multiple times so the slots may not align either. */ KASSERT(((hpts->p_prev_slot == hpts->p_cur_slot) || (wrap_loop_cnt >= 2) || (from_callout == 0)), ("H:%p p_prev_slot:%u not equal to p_cur_slot:%u", hpts, hpts->p_prev_slot, hpts->p_cur_slot)); KASSERT(((hpts->p_lasttick == hpts->p_curtick) || (wrap_loop_cnt >= 2) || (from_callout == 0)), ("H:%p p_lasttick:%u not equal to p_curtick:%u", hpts, hpts->p_lasttick, hpts->p_curtick)); if (from_callout && (hpts->p_lasttick != hpts->p_curtick)) { hpts->p_curtick = tcp_gethptstick(&tv); counter_u64_add(hpts_loops, 1); hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick); goto again; } if (from_callout){ tcp_hpts_set_max_sleep(hpts, wrap_loop_cnt); } if (seen_endpoint) return(hpts_slots_diff(slot_pos_of_endpoint, orig_exit_slot)); else return (0); } void __tcp_set_hpts(struct inpcb *inp, int32_t line) { struct tcp_hpts_entry *hpts; int failed; INP_WLOCK_ASSERT(inp); hpts = tcp_hpts_lock(inp); if ((inp->inp_in_hpts == 0) && (inp->inp_hpts_cpu_set == 0)) { inp->inp_hpts_cpu = hpts_cpuid(inp, &failed); if (failed == 0) inp->inp_hpts_cpu_set = 1; } mtx_unlock(&hpts->p_mtx); } static void __tcp_run_hpts(struct tcp_hpts_entry *hpts) { int ticks_ran; if (hpts->p_hpts_active) { /* Already active */ return; } if (mtx_trylock(&hpts->p_mtx) == 0) { /* Someone else got the lock */ return; } if (hpts->p_hpts_active) goto out_with_mtx; hpts->syscall_cnt++; counter_u64_add(hpts_direct_call, 1); hpts->p_hpts_active = 1; ticks_ran = tcp_hptsi(hpts, 0); /* We may want to adjust the sleep values here */ if (hpts->p_on_queue_cnt >= conn_cnt_thresh) { if (ticks_ran > ticks_indicate_less_sleep) { struct timeval tv; sbintime_t sb; hpts->p_mysleep.tv_usec /= 2; if (hpts->p_mysleep.tv_usec < dynamic_min_sleep) hpts->p_mysleep.tv_usec = dynamic_min_sleep; /* Reschedule with new to value */ tcp_hpts_set_max_sleep(hpts, 0); tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_SLOT; /* Validate its in the right ranges */ if (tv.tv_usec < hpts->p_mysleep.tv_usec) { hpts->overidden_sleep = tv.tv_usec; tv.tv_usec = hpts->p_mysleep.tv_usec; } else if (tv.tv_usec > dynamic_max_sleep) { /* Lets not let sleep get above this value */ hpts->overidden_sleep = tv.tv_usec; tv.tv_usec = dynamic_max_sleep; } /* * In this mode the timer is a backstop to * all the userret/lro_flushes so we use * the dynamic value and set the on_min_sleep * flag so we will not be awoken. */ sb = tvtosbt(tv); /* Store off to make visible the actual sleep time */ hpts->sleeping = tv.tv_usec; callout_reset_sbt_on(&hpts->co, sb, 0, hpts_timeout_swi, hpts, hpts->p_cpu, (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); } else if (ticks_ran < ticks_indicate_more_sleep) { /* For the further sleep, don't reschedule hpts */ hpts->p_mysleep.tv_usec *= 2; if (hpts->p_mysleep.tv_usec > dynamic_max_sleep) hpts->p_mysleep.tv_usec = dynamic_max_sleep; } hpts->p_on_min_sleep = 1; } hpts->p_hpts_active = 0; out_with_mtx: HPTS_MTX_ASSERT(hpts); mtx_unlock(&hpts->p_mtx); } static struct tcp_hpts_entry * tcp_choose_hpts_to_run(void) { int i, oldest_idx, start, end; uint32_t cts, time_since_ran, calc; cts = tcp_get_usecs(NULL); time_since_ran = 0; /* Default is all one group */ start = 0; end = tcp_pace.rp_num_hptss; /* * If we have more than one L3 group figure out which one * this CPU is in. */ if (tcp_pace.grp_cnt > 1) { for (i = 0; i < tcp_pace.grp_cnt; i++) { if (CPU_ISSET(curcpu, &tcp_pace.grps[i]->cg_mask)) { start = tcp_pace.grps[i]->cg_first; end = (tcp_pace.grps[i]->cg_last + 1); break; } } } oldest_idx = -1; for (i = start; i < end; i++) { if (TSTMP_GT(cts, cts_last_ran[i])) calc = cts - cts_last_ran[i]; else calc = 0; if (calc > time_since_ran) { oldest_idx = i; time_since_ran = calc; } } if (oldest_idx >= 0) return(tcp_pace.rp_ent[oldest_idx]); else return(tcp_pace.rp_ent[(curcpu % tcp_pace.rp_num_hptss)]); } void tcp_run_hpts(void) { static struct tcp_hpts_entry *hpts; struct epoch_tracker et; NET_EPOCH_ENTER(et); hpts = tcp_choose_hpts_to_run(); __tcp_run_hpts(hpts); NET_EPOCH_EXIT(et); } static void tcp_hpts_thread(void *ctx) { struct tcp_hpts_entry *hpts; struct epoch_tracker et; struct timeval tv; sbintime_t sb; int ticks_ran; hpts = (struct tcp_hpts_entry *)ctx; mtx_lock(&hpts->p_mtx); if (hpts->p_direct_wake) { /* Signaled by input or output with low occupancy count. */ callout_stop(&hpts->co); counter_u64_add(hpts_direct_awakening, 1); } else { /* Timed out, the normal case. */ counter_u64_add(hpts_wake_timeout, 1); if (callout_pending(&hpts->co) || !callout_active(&hpts->co)) { mtx_unlock(&hpts->p_mtx); return; } } callout_deactivate(&hpts->co); hpts->p_hpts_wake_scheduled = 0; NET_EPOCH_ENTER(et); if (hpts->p_hpts_active) { /* * We are active already. This means that a syscall * trap or LRO is running in behalf of hpts. In that case * we need to double our timeout since there seems to be * enough activity in the system that we don't need to * run as often (if we were not directly woken). */ if (hpts->p_direct_wake == 0) { counter_u64_add(hpts_back_tosleep, 1); if (hpts->p_on_queue_cnt >= conn_cnt_thresh) { hpts->p_mysleep.tv_usec *= 2; if (hpts->p_mysleep.tv_usec > dynamic_max_sleep) hpts->p_mysleep.tv_usec = dynamic_max_sleep; tv.tv_usec = hpts->p_mysleep.tv_usec; hpts->p_on_min_sleep = 1; } else { /* * Here we have low count on the wheel, but * somehow we still collided with one of the * connections. Lets go back to sleep for a * min sleep time, but clear the flag so we * can be awoken by insert. */ hpts->p_on_min_sleep = 0; tv.tv_usec = tcp_min_hptsi_time; } } else { /* * Directly woken most likely to reset the * callout time. */ tv.tv_sec = 0; tv.tv_usec = hpts->p_mysleep.tv_usec; } goto back_to_sleep; } hpts->sleeping = 0; hpts->p_hpts_active = 1; ticks_ran = tcp_hptsi(hpts, 1); tv.tv_sec = 0; tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_SLOT; if (hpts->p_on_queue_cnt >= conn_cnt_thresh) { if(hpts->p_direct_wake == 0) { /* * Only adjust sleep time if we were * called from the callout i.e. direct_wake == 0. */ if (ticks_ran < ticks_indicate_more_sleep) { hpts->p_mysleep.tv_usec *= 2; if (hpts->p_mysleep.tv_usec > dynamic_max_sleep) hpts->p_mysleep.tv_usec = dynamic_max_sleep; } else if (ticks_ran > ticks_indicate_less_sleep) { hpts->p_mysleep.tv_usec /= 2; if (hpts->p_mysleep.tv_usec < dynamic_min_sleep) hpts->p_mysleep.tv_usec = dynamic_min_sleep; } } if (tv.tv_usec < hpts->p_mysleep.tv_usec) { hpts->overidden_sleep = tv.tv_usec; tv.tv_usec = hpts->p_mysleep.tv_usec; } else if (tv.tv_usec > dynamic_max_sleep) { /* Lets not let sleep get above this value */ hpts->overidden_sleep = tv.tv_usec; tv.tv_usec = dynamic_max_sleep; } /* * In this mode the timer is a backstop to * all the userret/lro_flushes so we use * the dynamic value and set the on_min_sleep * flag so we will not be awoken. */ hpts->p_on_min_sleep = 1; } else if (hpts->p_on_queue_cnt == 0) { /* * No one on the wheel, please wake us up * if you insert on the wheel. */ hpts->p_on_min_sleep = 0; hpts->overidden_sleep = 0; } else { /* * We hit here when we have a low number of * clients on the wheel (our else clause). * We may need to go on min sleep, if we set * the flag we will not be awoken if someone * is inserted ahead of us. Clearing the flag * means we can be awoken. This is "old mode" * where the timer is what runs hpts mainly. */ if (tv.tv_usec < tcp_min_hptsi_time) { /* * Yes on min sleep, which means * we cannot be awoken. */ hpts->overidden_sleep = tv.tv_usec; tv.tv_usec = tcp_min_hptsi_time; hpts->p_on_min_sleep = 1; } else { /* Clear the min sleep flag */ hpts->overidden_sleep = 0; hpts->p_on_min_sleep = 0; } } HPTS_MTX_ASSERT(hpts); hpts->p_hpts_active = 0; back_to_sleep: hpts->p_direct_wake = 0; sb = tvtosbt(tv); /* Store off to make visible the actual sleep time */ hpts->sleeping = tv.tv_usec; callout_reset_sbt_on(&hpts->co, sb, 0, hpts_timeout_swi, hpts, hpts->p_cpu, (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); NET_EPOCH_EXIT(et); mtx_unlock(&hpts->p_mtx); } #undef timersub static int32_t hpts_count_level(struct cpu_group *cg) { int32_t count_l3, i; count_l3 = 0; if (cg->cg_level == CG_SHARE_L3) count_l3++; /* Walk all the children looking for L3 */ for (i = 0; i < cg->cg_children; i++) { count_l3 += hpts_count_level(&cg->cg_child[i]); } return (count_l3); } static void hpts_gather_grps(struct cpu_group **grps, int32_t *at, int32_t max, struct cpu_group *cg) { int32_t idx, i; idx = *at; if (cg->cg_level == CG_SHARE_L3) { grps[idx] = cg; idx++; if (idx == max) { *at = idx; return; } } *at = idx; /* Walk all the children looking for L3 */ for (i = 0; i < cg->cg_children; i++) { hpts_gather_grps(grps, at, max, &cg->cg_child[i]); } } static void tcp_init_hptsi(void *st) { struct cpu_group *cpu_top; int32_t error __diagused; int32_t i, j, bound = 0, created = 0; size_t sz, asz; struct timeval tv; sbintime_t sb; struct tcp_hpts_entry *hpts; struct pcpu *pc; char unit[16]; uint32_t ncpus = mp_ncpus ? mp_ncpus : MAXCPU; int count, domain; #ifdef SMP cpu_top = smp_topo(); #else cpu_top = NULL; #endif tcp_pace.rp_num_hptss = ncpus; hpts_hopelessly_behind = counter_u64_alloc(M_WAITOK); hpts_loops = counter_u64_alloc(M_WAITOK); back_tosleep = counter_u64_alloc(M_WAITOK); combined_wheel_wrap = counter_u64_alloc(M_WAITOK); wheel_wrap = counter_u64_alloc(M_WAITOK); hpts_wake_timeout = counter_u64_alloc(M_WAITOK); hpts_direct_awakening = counter_u64_alloc(M_WAITOK); hpts_back_tosleep = counter_u64_alloc(M_WAITOK); hpts_direct_call = counter_u64_alloc(M_WAITOK); cpu_uses_flowid = counter_u64_alloc(M_WAITOK); cpu_uses_random = counter_u64_alloc(M_WAITOK); sz = (tcp_pace.rp_num_hptss * sizeof(struct tcp_hpts_entry *)); tcp_pace.rp_ent = malloc(sz, M_TCPHPTS, M_WAITOK | M_ZERO); sz = (sizeof(uint32_t) * tcp_pace.rp_num_hptss); cts_last_ran = malloc(sz, M_TCPHPTS, M_WAITOK); tcp_pace.grp_cnt = 0; if (cpu_top == NULL) { tcp_pace.grp_cnt = 1; } else { /* Find out how many cache level 3 domains we have */ count = 0; tcp_pace.grp_cnt = hpts_count_level(cpu_top); if (tcp_pace.grp_cnt == 0) { tcp_pace.grp_cnt = 1; } sz = (tcp_pace.grp_cnt * sizeof(struct cpu_group *)); tcp_pace.grps = malloc(sz, M_TCPHPTS, M_WAITOK); /* Now populate the groups */ if (tcp_pace.grp_cnt == 1) { /* * All we need is the top level all cpu's are in * the same cache so when we use grp[0]->cg_mask * with the cg_first <-> cg_last it will include * all cpu's in it. The level here is probably * zero which is ok. */ tcp_pace.grps[0] = cpu_top; } else { /* * Here we must find all the level three cache domains * and setup our pointers to them. */ count = 0; hpts_gather_grps(tcp_pace.grps, &count, tcp_pace.grp_cnt, cpu_top); } } asz = sizeof(struct hptsh) * NUM_OF_HPTSI_SLOTS; for (i = 0; i < tcp_pace.rp_num_hptss; i++) { tcp_pace.rp_ent[i] = malloc(sizeof(struct tcp_hpts_entry), M_TCPHPTS, M_WAITOK | M_ZERO); tcp_pace.rp_ent[i]->p_hptss = malloc(asz, M_TCPHPTS, M_WAITOK); hpts = tcp_pace.rp_ent[i]; /* * Init all the hpts structures that are not specifically * zero'd by the allocations. Also lets attach them to the * appropriate sysctl block as well. */ mtx_init(&hpts->p_mtx, "tcp_hpts_lck", "hpts", MTX_DEF | MTX_DUPOK); for (j = 0; j < NUM_OF_HPTSI_SLOTS; j++) { TAILQ_INIT(&hpts->p_hptss[j].head); hpts->p_hptss[j].count = 0; hpts->p_hptss[j].gencnt = 0; } sysctl_ctx_init(&hpts->hpts_ctx); sprintf(unit, "%d", i); hpts->hpts_root = SYSCTL_ADD_NODE(&hpts->hpts_ctx, SYSCTL_STATIC_CHILDREN(_net_inet_tcp_hpts), OID_AUTO, unit, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, ""); SYSCTL_ADD_INT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "out_qcnt", CTLFLAG_RD, &hpts->p_on_queue_cnt, 0, "Count TCB's awaiting output processing"); SYSCTL_ADD_U16(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "active", CTLFLAG_RD, &hpts->p_hpts_active, 0, "Is the hpts active"); SYSCTL_ADD_UINT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "curslot", CTLFLAG_RD, &hpts->p_cur_slot, 0, "What the current running pacers goal"); SYSCTL_ADD_UINT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "runtick", CTLFLAG_RD, &hpts->p_runningslot, 0, "What the running pacers current slot is"); SYSCTL_ADD_UINT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "curtick", CTLFLAG_RD, &hpts->p_curtick, 0, "What the running pacers last tick mapped to the wheel was"); SYSCTL_ADD_UINT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "lastran", CTLFLAG_RD, &cts_last_ran[i], 0, "The last usec tick that this hpts ran"); SYSCTL_ADD_LONG(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "cur_min_sleep", CTLFLAG_RD, &hpts->p_mysleep.tv_usec, "What the running pacers is using for p_mysleep.tv_usec"); SYSCTL_ADD_U64(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "now_sleeping", CTLFLAG_RD, &hpts->sleeping, 0, "What the running pacers is actually sleeping for"); SYSCTL_ADD_U64(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "syscall_cnt", CTLFLAG_RD, &hpts->syscall_cnt, 0, "How many times we had syscalls on this hpts"); hpts->p_hpts_sleep_time = hpts_sleep_max; hpts->p_num = i; hpts->p_curtick = tcp_gethptstick(&tv); cts_last_ran[i] = tcp_tv_to_usectick(&tv); hpts->p_prev_slot = hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick); hpts->p_cpu = 0xffff; hpts->p_nxt_slot = hpts_slot(hpts->p_cur_slot, 1); callout_init(&hpts->co, 1); } /* Don't try to bind to NUMA domains if we don't have any */ if (vm_ndomains == 1 && tcp_bind_threads == 2) tcp_bind_threads = 0; /* * Now lets start ithreads to handle the hptss. */ for (i = 0; i < tcp_pace.rp_num_hptss; i++) { hpts = tcp_pace.rp_ent[i]; hpts->p_cpu = i; error = swi_add(&hpts->ie, "hpts", tcp_hpts_thread, (void *)hpts, SWI_NET, INTR_MPSAFE, &hpts->ie_cookie); KASSERT(error == 0, ("Can't add hpts:%p i:%d err:%d", hpts, i, error)); created++; hpts->p_mysleep.tv_sec = 0; hpts->p_mysleep.tv_usec = tcp_min_hptsi_time; if (tcp_bind_threads == 1) { if (intr_event_bind(hpts->ie, i) == 0) bound++; } else if (tcp_bind_threads == 2) { /* Find the group for this CPU (i) and bind into it */ for (j = 0; j < tcp_pace.grp_cnt; j++) { if (CPU_ISSET(i, &tcp_pace.grps[j]->cg_mask)) { if (intr_event_bind_ithread_cpuset(hpts->ie, &tcp_pace.grps[j]->cg_mask) == 0) { bound++; pc = pcpu_find(i); domain = pc->pc_domain; count = hpts_domains[domain].count; hpts_domains[domain].cpu[count] = i; hpts_domains[domain].count++; break; } } } } tv.tv_sec = 0; tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_SLOT; hpts->sleeping = tv.tv_usec; sb = tvtosbt(tv); callout_reset_sbt_on(&hpts->co, sb, 0, hpts_timeout_swi, hpts, hpts->p_cpu, (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); } /* * If we somehow have an empty domain, fall back to choosing * among all htps threads. */ for (i = 0; i < vm_ndomains; i++) { if (hpts_domains[i].count == 0) { tcp_bind_threads = 0; break; } } printf("TCP Hpts created %d swi interrupt threads and bound %d to %s\n", created, bound, tcp_bind_threads == 2 ? "NUMA domains" : "cpus"); #ifdef INVARIANTS printf("HPTS is in INVARIANT mode!!\n"); #endif } SYSINIT(tcphptsi, SI_SUB_SOFTINTR, SI_ORDER_ANY, tcp_init_hptsi, NULL); MODULE_VERSION(tcphpts, 1); diff --git a/sys/netinet/tcp_input.c b/sys/netinet/tcp_input.c index e4f2663cafdc..96812afd3b37 100644 --- a/sys/netinet/tcp_input.c +++ b/sys/netinet/tcp_input.c @@ -1,4103 +1,4103 @@ /*- * 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 #include #ifdef TCP_OFFLOAD #include #endif #include #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(bool, blackhole_local) = false; #define V_blackhole_local VNET(blackhole_local) SYSCTL_BOOL(_net_inet_tcp, OID_AUTO, blackhole_local, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(blackhole_local), false, "Enforce net.inet.tcp.blackhole for locally originated packets"); 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_lrd) = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_lrd, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_lrd), 1, "Perform Lost Retransmission Detection"); 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"); 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 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"); /* * 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); } /* * Make sure that we only start a SACK loss recovery when * receiving a duplicate ACK with a SACK block, and also * complete SACK loss recovery in case the other end * reneges. */ static bool inline tcp_is_sack_recovery(struct tcpcb *tp, struct tcpopt *to) { return ((tp->t_flags & TF_SACK_PERMIT) && ((to->to_flags & TOF_SACK) || (!TAILQ_EMPTY(&tp->snd_holes)))); } #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); + &tp->t_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(tptoinpcb(tp)); - tp->ccv->nsegs = nsegs; - tp->ccv->bytes_this_ack = BYTES_THIS_ACK(tp, th); + tp->t_ccv.nsegs = nsegs; + tp->t_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; + tp->t_ccv.flags |= CCF_CWND_LIMITED; else - tp->ccv->flags &= ~CCF_CWND_LIMITED; + tp->t_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)); + tp->t_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)SEQ_SUB(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; + tp->t_bytes_acked += tp->t_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; + tp->t_ccv.flags |= CCF_ABC_SENTAWND; } } else { - tp->ccv->flags &= ~CCF_ABC_SENTAWND; + tp->t_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); + tp->t_ccv.curack = th->th_ack; + CC_ALGO(tp)->ack_received(&tp->t_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 = tptoinpcb(tp); u_int maxseg; int rtt; INP_WLOCK_ASSERT(inp); tcp_hc_get(&inp->inp_inc, &metrics); maxseg = tcp_maxseg(tp); if (tp->t_srtt == 0 && (rtt = metrics.rmx_rtt)) { tp->t_srtt = rtt; 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); + CC_ALGO(tp)->conn_init(&tp->t_ccv); } void inline cc_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type) { INP_WLOCK_ASSERT(tptoinpcb(tp)); #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); + tp->t_ccv.curack = th->th_ack; + CC_ALGO(tp)->cong_signal(&tp->t_ccv, type); } } void inline cc_post_recovery(struct tcpcb *tp, struct tcphdr *th) { INP_WLOCK_ASSERT(tptoinpcb(tp)); /* 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); + tp->t_ccv.curack = th->th_ack; + CC_ALGO(tp)->post_recovery(&tp->t_ccv); } /* XXXLAS: EXIT_RECOVERY ? */ tp->t_bytes_acked = 0; tp->sackhint.delivered_data = 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_flags(struct tcpcb *tp, uint16_t flags, uint8_t iptos) { INP_WLOCK_ASSERT(tptoinpcb(tp)); if (CC_ALGO(tp)->ecnpkt_handler != NULL) { switch (iptos & IPTOS_ECN_MASK) { case IPTOS_ECN_CE: - tp->ccv->flags |= CCF_IPHDR_CE; + tp->t_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; + tp->t_ccv.flags &= ~CCF_IPHDR_CE; break; } if (flags & TH_CWR) - tp->ccv->flags |= CCF_TCPHDR_CWR; + tp->t_ccv.flags |= CCF_TCPHDR_CWR; else - tp->ccv->flags &= ~CCF_TCPHDR_CWR; + tp->t_ccv.flags &= ~CCF_TCPHDR_CWR; - CC_ALGO(tp)->ecnpkt_handler(tp->ccv); + CC_ALGO(tp)->ecnpkt_handler(&tp->t_ccv); - if (tp->ccv->flags & CCF_ACKNOW) { + if (tp->t_ccv.flags & CCF_ACKNOW) { tcp_timer_activate(tp, TT_DELACK, tcp_delacktime); tp->t_flags |= TF_ACKNOW; } } } void inline cc_ecnpkt_handler(struct tcpcb *tp, struct tcphdr *th, uint8_t iptos) { cc_ecnpkt_handler_flags(tp, tcp_get_flags(th), iptos); } /* * 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_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_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_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; } 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 = tcp_get_flags(th); /* * 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 we don't need write lock on matching * PCB, be it a listening one or a synchronized one. The packet * shall not modify its state. */ lookupflag = INPLOOKUP_WILDCARD | ((thflags & (TH_ACK|TH_SYN)) == TH_SYN ? INPLOOKUP_RLOCKPCB : 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, 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, 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, lookupflag, m->m_pkthdr.rcvif); } } else inp = in_pcblookup_mbuf(&V_tcbinfo, ip->ip_src, th->th_sport, ip->ip_dst, th->th_dport, 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) { if (rstreason != 0) { /* We came here after second (safety) lookup. */ MPASS((lookupflag & INPLOOKUP_WILDCARD) == 0); goto dropwithreset; } /* * 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) && (V_blackhole_local || ( #ifdef INET6 isipv6 ? !in6_localaddr(&ip6->ip6_src) : #endif #ifdef INET !in_localip(ip->ip_src) #else true #endif ))) goto dropunlock; rstreason = BANDLIM_RST_CLOSEDPORT; goto dropwithreset; } INP_LOCK_ASSERT(inp); if ((inp->inp_flowtype == M_HASHTYPE_NONE) && (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) && !SOLISTENING(inp->inp_socket)) { 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; } tp = intotcpcb(inp); switch (tp->t_state) { case TCPS_TIME_WAIT: /* * 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. */ tcp_dooptions(&to, optp, optlen, (thflags & TH_SYN) ? TO_SYN : 0); /* * tcp_twcheck unlocks the inp always, and frees the m if fails. */ if (tcp_twcheck(inp, &to, th, m, tlen)) goto findpcb; return (IPPROTO_DONE); case TCPS_CLOSED: /* * 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. */ 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 || !SOLISTENING(so), ("%s: so accepting but tp %p not listening", __func__, tp)); if (tp->t_state == TCPS_LISTEN && SOLISTENING(so)) { 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. */ 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. Do our protection against double * ACK. If peer sent us 2 ACKs, then for the * first one syncache_expand() successfully * converted syncache entry into a socket, * while we were waiting on the inpcb lock. We * don't want to sent RST for the second ACK, * so we perform second lookup without wildcard * match, hoping to find the new socket. If * the ACK is stray indeed, rstreason would * hint the above code that the lookup was a * second attempt. * * NB: syncache did its own logging * of the failure cause. */ INP_WUNLOCK(inp); rstreason = BANDLIM_RST_OPENPORT; lookupflag &= ~INPLOOKUP_WILDCARD; goto findpcb; } 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, 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, 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, port)) != NULL) goto tfo_socket_result; /* * Entry added to syncache and mbuf consumed. * Only the listen socket is unlocked by syncache_add(). */ 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. * * XXXGL: in case of a pure SYN arriving on existing connection * TCP stacks won't need to modify the PCB, they would either drop * the segment silently, or send a challenge ACK. However, we try * to upgrade the lock, because calling convention for stacks is * write-lock on PCB. If upgrade fails, drop the SYN. */ if ((lookupflag & INPLOOKUP_RLOCKPCB) && INP_TRY_UPGRADE(inp) == 0) goto dropunlock; 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: 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)); } static void tcp_handle_wakeup(struct tcpcb *tp) { INP_WLOCK_ASSERT(tptoinpcb(tp)); if (tp->t_flags & TF_WAKESOR) { struct socket *so = tptosocket(tp); tp->t_flags &= ~TF_WAKESOR; SOCKBUF_LOCK_ASSERT(&so->so_rcv); sorwakeup_locked(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) { uint16_t thflags; int acked, ourfinisacked, needoutput = 0, sack_changed; int rstreason, todrop, win, incforsyn = 0; uint32_t tiwin; uint16_t nsegs; char *s; struct inpcb *inp = tptoinpcb(tp); struct in_conninfo *inc = &inp->inp_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 = tcp_get_flags(th); tp->sackhint.last_sack_ack = 0; sack_changed = 0; nsegs = max(1, m->m_pkthdr.lro_nsegs); NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(inp); 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; tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 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. */ if (tp->t_idle_reduce && (tp->snd_max == tp->snd_una) && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) cc_after_idle(tp); tp->t_rcvtime = ticks; if (thflags & TH_FIN) tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN); /* * 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 (tcp_ecn_input_segment(tp, thflags, tlen, tcp_packets_this_ack(tp, th->th_ack), iptos)) 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_rxtshift == 1 && tp->t_flags & TF_PREVVALID && tp->t_badrxtwin != 0 && TSTMP_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 */ tcp_ecn_input_parallel_syn(tp, thflags, iptos); 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 ((inp->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 && tp->t_badrxtwin != 0 && TSTMP_LT(ticks, tp->t_badrxtwin)) { 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); /* * Clear t_acktime if remote side has ACKd * all data in the socket buffer. * Otherwise, update t_acktime if we received * a sufficiently large ACK. */ if (sbavail(&so->so_snd) == 0) tp->t_acktime = 0; else if (acked > 1) tp->t_acktime = ticks; 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_RXTCUR(tp)); sowwakeup(so); if (sbavail(&so->so_snd)) (void) 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, NULL)) so->so_rcv.sb_flags &= ~SB_AUTOSIZE; m_adj(m, drop_hdrlen); /* delayed header drop */ sbappendstream_locked(&so->so_rcv, m, 0); } /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); if (DELAY_ACK(tp, tlen)) { tp->t_flags |= TF_DELACK; } else { tp->t_flags |= TF_ACKNOW; 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; tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 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; tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 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); tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 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; tcp_ecn_input_syn_sent(tp, thflags, iptos); /* * Received in SYN_SENT[*] state. * Transitions: * SYN_SENT --> ESTABLISHED * SYN_SENT* --> FIN_WAIT_1 */ tp->t_starttime = ticks; if (tp->t_flags & TF_NEEDFIN) { tp->t_acktime = ticks; 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 | TF_SONOTCONN); tcp_timer_activate(tp, TT_REXMT, 0); tcp_state_change(tp, TCPS_SYN_RECEIVED); } /* * 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; } /* * 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: tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_RST); 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)) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); tp = tcp_drop(tp, ECONNRESET); rstreason = BANDLIM_UNLIMITED; } else { tcp_ecn_input_syn_sent(tp, thflags, iptos); /* 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; tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 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 ((tp->t_flags & TF_CLOSED) && 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); } 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); 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); if (tp->t_flags & TF_SONOTCONN) { /* * Usually SYN_RECEIVED had been created from a LISTEN, * and solisten_enqueue() has already marked the socket * layer as connected. If it didn't, which can happen * only with an accept_filter(9), then the tp is marked * with TF_SONOTCONN. The other reason for this mark * to be set is a simultaneous open, a SYN_RECEIVED * that had been created from SYN_SENT. */ tp->t_flags &= ~TF_SONOTCONN; 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) { tp->t_acktime = ticks; 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); tcp_handle_wakeup(tp); } 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 (tcp_is_sack_recovery(tp, &to)) { if (((sack_changed = tcp_sack_doack(tp, &to, th->th_ack)) != 0) && (tp->t_flags & TF_LRD)) { tcp_sack_lost_retransmission(tp, th); } } 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 || (tcp_is_sack_recovery(tp, &to) && !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)) { tcp_do_prr_ack(tp, th, &to); } else if (tcp_is_sack_recovery(tp, &to) && 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) 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. */ if (tcp_is_sack_recovery(tp, &to)) { tp->sackhint.prr_delivered = tp->sackhint.sacked_bytes; } else { tp->sackhint.prr_delivered = imin(tp->snd_max - tp->snd_una, imin(INT_MAX / 65536, tp->t_dupacks) * maxseg); } tp->sackhint.recover_fs = max(1, tp->snd_nxt - tp->snd_una); } if (tcp_is_sack_recovery(tp, &to)) { TCPSTAT_INC( tcps_sack_recovery_episode); tp->snd_recover = tp->snd_nxt; tp->snd_cwnd = maxseg; (void) 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) 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 * or we need to send an ACK. */ SOCKBUF_LOCK(&so->so_snd); avail = sbavail(&so->so_snd) - (tp->snd_nxt - tp->snd_una); SOCKBUF_UNLOCK(&so->so_snd); if (avail > 0 || tp->t_flags & TF_ACKNOW) (void) 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 (tcp_is_sack_recovery(tp, &to) && 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(inp); /* * 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 != 0 && to.to_flags & TOF_TS && to.to_tsecr != 0 && TSTMP_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); } SOCKBUF_LOCK(&so->so_snd); /* * Clear t_acktime if remote side has ACKd all data in the * socket buffer and FIN (if applicable). * Otherwise, update t_acktime if we received a sufficiently * large ACK. */ if ((tp->t_state <= TCPS_CLOSE_WAIT && acked == sbavail(&so->so_snd)) || acked > sbavail(&so->so_snd)) tp->t_acktime = 0; else if (acked > 1) tp->t_acktime = ticks; /* * 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_RXTCUR(tp)); /* * If no data (only SYN) was ACK'd, * skip rest of ACK processing. */ if (acked == 0) { SOCKBUF_UNLOCK(&so->so_snd); 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); 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; } /* NB: sowwakeup_locked() does an implicit unlock. */ sowwakeup_locked(so); 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. */ 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(inp); /* * 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(inp); /* * 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 = tcp_get_flags(th) & 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); 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); } } tcp_handle_wakeup(tp); #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) { /* The socket upcall is handled by socantrcvmore. */ socantrcvmore(so); /* * 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) tcp_output(tp); check_delack: INP_WLOCK_ASSERT(inp); if (tp->t_flags & TF_DELACK) { tp->t_flags &= ~TF_DELACK; tcp_timer_activate(tp, TT_DELACK, tcp_delacktime); } INP_WUNLOCK(inp); 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; tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 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) tcp_output(tp); INP_WUNLOCK(inp); m_freem(m); return; dropwithreset: if (tp != NULL) { tcp_dropwithreset(m, th, tp, tlen, rstreason); INP_WUNLOCK(inp); } else tcp_dropwithreset(m, th, NULL, tlen, rstreason); return; drop: /* * Drop space held by incoming segment and return. */ #ifdef TCPDEBUG if (tp == NULL || (inp->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) { INP_WUNLOCK(inp); } 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(tptoinpcb(tp)); } /* Don't bother if destination was broadcast/multicast. */ if ((tcp_get_flags(th) & 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 (tcp_get_flags(th) & TH_ACK) { tcp_respond(tp, mtod(m, void *), th, m, (tcp_seq)0, th->th_ack, TH_RST); } else { if (tcp_get_flags(th) & TH_SYN) tlen++; if (tcp_get_flags(th) & 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(tptoinpcb(tp)); 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(tptoinpcb(tp)); 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; } 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_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 = tptoinpcb(tp); 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 size_t min_protoh = sizeof(struct tcpiphdr); #endif INP_WLOCK_ASSERT(inp); 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 = tptoinpcb(tp); 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; /* * 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, 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, 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(tptoinpcb(tp)); /* * Compute the amount of data that this ACK is indicating * (del_data) and an estimate of how many bytes are in the * network. */ if (tcp_is_sack_recovery(tp, to) || (IN_CONGRECOVERY(tp->t_flags) && !IN_FASTRECOVERY(tp->t_flags))) { 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; } else { if (tp->sackhint.prr_delivered < (tcprexmtthresh * maxseg + tp->snd_recover - tp->snd_una)) del_data = maxseg; pipe = imax(0, tp->snd_max - tp->snd_una - imin(INT_MAX / 65536, tp->t_dupacks) * maxseg); } 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 || (del_data == 0)) 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)) { if (tcp_is_sack_recovery(tp, to)) { tp->snd_cwnd = tp->snd_nxt - tp->snd_recover + tp->sackhint.sack_bytes_rexmit + (snd_cnt * maxseg); } else { tp->snd_cwnd = (tp->snd_max - tp->snd_una) + (snd_cnt * maxseg); } } else if (IN_CONGRECOVERY(tp->t_flags)) tp->snd_cwnd = pipe - del_data + (snd_cnt * maxseg); tp->snd_cwnd = imax(maxseg, tp->snd_cwnd); } /* * 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(tptoinpcb(tp)); 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) 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) { if (tp->t_fb->tfb_compute_pipe == NULL) { return (tp->snd_max - tp->snd_una + tp->sackhint.sack_bytes_rexmit - tp->sackhint.sacked_bytes); } else { return((*tp->t_fb->tfb_compute_pipe)(tp)); } } 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 3ebf36196ac0..aa4bfef9a9ff 100644 --- a/sys/netinet/tcp_output.c +++ b/sys/netinet/tcp_output.c @@ -1,2207 +1,2207 @@ /*- * 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 #include #ifdef TCPPCAP #include #endif #ifdef TCPDEBUG #include #endif #ifdef TCP_OFFLOAD #include #endif #include #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")) #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); + &tp->t_osd); } } #endif /* * CC wrapper hook functions */ void cc_after_idle(struct tcpcb *tp) { INP_WLOCK_ASSERT(tptoinpcb(tp)); if (CC_ALGO(tp)->after_idle != NULL) - CC_ALGO(tp)->after_idle(tp->ccv); + CC_ALGO(tp)->after_idle(&tp->t_ccv); } /* * Tcp output routine: figure out what should be sent and send it. */ int tcp_default_output(struct tcpcb *tp) { struct socket *so = tptosocket(tp); struct inpcb *inp = tptoinpcb(tp); int32_t len; uint32_t recwin, sendwin; uint16_t flags; int off, 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, 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; struct tcp_log_buffer *lgb; 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 = (inp->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(inp); #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, SEQ_SUB(tp->snd_recover, p->rxmit))); } } else { len = ((int32_t)ulmin(cwin, SEQ_SUB(p->end, p->rxmit))); } if (len > 0) { off = SEQ_SUB(p->rxmit, tp->snd_una); KASSERT(off >= 0,("%s: sack block to the left of una : %d", __func__, off)); 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 - imax(0, (int32_t) (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_active(tp, TT_PERSIST)) { 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, inp); #ifdef INET else #endif #endif /* INET6 */ #ifdef INET if (IPSEC_ENABLED(ipv4)) ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp); #endif /* INET */ #endif /* IPSEC */ #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) 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_RXTCUR(tp)); 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(&inp->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(inp, m); #endif #ifdef INET6 if (isipv6) { ip6 = mtod(m, struct ip6_hdr *); if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip6 + 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(inp, tp->t_port, ip6, th); } else #endif /* INET6 */ { ip = mtod(m, struct ip *); #ifdef TCPDEBUG ipov = (struct ipovly *)ip; #endif if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip + 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(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) { flags |= tcp_ecn_output_syn_sent(tp); } /* Also handle parallel SYN for ECN */ if ((TCPS_HAVERCVDSYN(tp->t_state)) && (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) { int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit); if ((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_flags2 & TF2_ECN_SND_ECE)) tp->t_flags2 &= ~TF2_ECN_SND_ECE; #ifdef INET6 if (isipv6) { ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20); ip6->ip6_flow |= htonl(ect << 20); } else #endif { ip->ip_tos &= ~IPTOS_ECN_MASK; ip->ip_tos |= ect; } } /* * 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; /* * Lost Retransmission Detection * trigger resending of a (then * still existing) hole, when * fack acks recoverypoint. */ if ((tp->t_flags & TF_LRD) && SEQ_GEQ(p->rxmit, p->end)) p->rxmit = tp->snd_recover; 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; } tcp_set_flags(th, 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 */ #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. */ 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. */ 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. */ /* XXXMT: We are not honoring verbose logging. */ if (tp->t_logstate != TCP_LOG_STATE_OFF) lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, len, NULL, false, NULL, NULL, 0, NULL); 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 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(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); #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, inp->in6p_outputopts, &inp->inp_route6, ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 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); #ifdef TCPPCAP /* Save packet, if requested. */ tcp_pcap_add(th, m, &(tp->t_outpkts)); #endif error = ip_output(m, inp->inp_options, &inp->inp_route, ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, inp); if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL) mtu = inp->inp_route.ro_nh->nh_mtu; } #endif /* INET */ if (lgb != NULL) { lgb->tlb_errno = error; lgb = NULL; } out: if (error == 0) tcp_account_for_send(tp, len, (tp->snd_nxt != tp->snd_max), 0, hw_tls); /* * 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)) { /* * Update "made progress" indication if we just * added new data to an empty socket buffer. */ if (tp->snd_una == tp->snd_max) tp->t_acktime = ticks; 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_RXTCUR(tp)); } 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) { /* * 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")); KASSERT((flags & TH_FIN) == 0, ("error while FIN with SACK rxmit")); } else { tp->snd_nxt -= len; if (flags & TH_FIN) tp->snd_nxt--; } } 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; int maxunacktime; tp->t_flags &= ~TF_PREVVALID; if (tcp_timer_active(tp, TT_REXMT)) panic("tcp_setpersist: retransmit pending"); /* * If the state is already closed, don't bother. */ if (tp->t_state == TCPS_CLOSED) return; /* * Start/restart persistence timer. */ TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], tcp_persmin, tcp_persmax); if (TP_MAXUNACKTIME(tp) && tp->t_acktime) { maxunacktime = tp->t_acktime + TP_MAXUNACKTIME(tp) - ticks; if (maxunacktime < 1) maxunacktime = 1; if (maxunacktime < tt) tt = maxunacktime; } 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 __diagused; #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), curthread)) so->so_snd.sb_flags &= ~SB_AUTOSIZE; } } } diff --git a/sys/netinet/tcp_pcap.c b/sys/netinet/tcp_pcap.c index 1faa1500ec5a..eb2af0c84a83 100644 --- a/sys/netinet/tcp_pcap.c +++ b/sys/netinet/tcp_pcap.c @@ -1,452 +1,454 @@ /*- * Copyright (c) 2015 * Jonathan Looney. 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. * * 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. * * $FreeBSD$ */ #include #include #include #include #include #include #include #include #include #include +#include +#include #include #include #define M_LEADINGSPACE_NOWRITE(m) \ ((m)->m_data - M_START(m)) int tcp_pcap_aggressive_free = 1; static int tcp_pcap_clusters_referenced_cur = 0; static int tcp_pcap_clusters_referenced_max = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_pcap_aggressive_free, CTLFLAG_RW, &tcp_pcap_aggressive_free, 0, "Free saved packets when the memory system comes under pressure"); SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_pcap_clusters_referenced_cur, CTLFLAG_RD, &tcp_pcap_clusters_referenced_cur, 0, "Number of clusters currently referenced on TCP PCAP queues"); SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_pcap_clusters_referenced_max, CTLFLAG_RW, &tcp_pcap_clusters_referenced_max, 0, "Maximum number of clusters allowed to be referenced on TCP PCAP " "queues"); static int tcp_pcap_alloc_reuse_ext = 0; static int tcp_pcap_alloc_reuse_mbuf = 0; static int tcp_pcap_alloc_new_mbuf = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_pcap_alloc_reuse_ext, CTLFLAG_RD, &tcp_pcap_alloc_reuse_ext, 0, "Number of mbufs with external storage reused for the TCP PCAP " "functionality"); SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_pcap_alloc_reuse_mbuf, CTLFLAG_RD, &tcp_pcap_alloc_reuse_mbuf, 0, "Number of mbufs with internal storage reused for the TCP PCAP " "functionality"); SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_pcap_alloc_new_mbuf, CTLFLAG_RD, &tcp_pcap_alloc_new_mbuf, 0, "Number of new mbufs allocated for the TCP PCAP functionality"); VNET_DEFINE(int, tcp_pcap_packets) = 0; #define V_tcp_pcap_packets VNET(tcp_pcap_packets) SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_pcap_packets, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_pcap_packets), 0, "Default number of packets saved per direction per TCPCB"); /* Initialize the values. */ static void tcp_pcap_max_set(void) { tcp_pcap_clusters_referenced_max = nmbclusters / 4; } void tcp_pcap_init(void) { tcp_pcap_max_set(); EVENTHANDLER_REGISTER(nmbclusters_change, tcp_pcap_max_set, NULL, EVENTHANDLER_PRI_ANY); } /* * If we are below the maximum allowed cluster references, * increment the reference count and return TRUE. Otherwise, * leave the reference count alone and return FALSE. */ static __inline bool tcp_pcap_take_cluster_reference(void) { if (atomic_fetchadd_int(&tcp_pcap_clusters_referenced_cur, 1) >= tcp_pcap_clusters_referenced_max) { atomic_add_int(&tcp_pcap_clusters_referenced_cur, -1); return FALSE; } return TRUE; } /* * For all the external entries in m, apply the given adjustment. * This can be used to adjust the counter when an mbuf chain is * copied or freed. */ static __inline void tcp_pcap_adj_cluster_reference(struct mbuf *m, int adj) { while (m) { if (m->m_flags & M_EXT) atomic_add_int(&tcp_pcap_clusters_referenced_cur, adj); m = m->m_next; } } /* * Free all mbufs in a chain, decrementing the reference count as * necessary. * * Functions in this file should use this instead of m_freem() when * they are freeing mbuf chains that may contain clusters that were * already included in tcp_pcap_clusters_referenced_cur. */ static void tcp_pcap_m_freem(struct mbuf *mb) { while (mb != NULL) { if (mb->m_flags & M_EXT) atomic_subtract_int(&tcp_pcap_clusters_referenced_cur, 1); mb = m_free(mb); } } /* * Copy data from m to n, where n cannot fit all the data we might * want from m. * * Prioritize data like this: * 1. TCP header * 2. IP header * 3. Data */ static void tcp_pcap_copy_bestfit(struct tcphdr *th, struct mbuf *m, struct mbuf *n) { struct mbuf *m_cur = m; int bytes_to_copy=0, trailing_data, skip=0, tcp_off; /* Below, we assume these will be non-NULL. */ KASSERT(th, ("%s: called with th == NULL", __func__)); KASSERT(m, ("%s: called with m == NULL", __func__)); KASSERT(n, ("%s: called with n == NULL", __func__)); /* We assume this initialization occurred elsewhere. */ KASSERT(n->m_len == 0, ("%s: called with n->m_len=%d (expected 0)", __func__, n->m_len)); KASSERT(n->m_data == M_START(n), ("%s: called with n->m_data != M_START(n)", __func__)); /* * Calculate the size of the TCP header. We use this often * enough that it is worth just calculating at the start. */ tcp_off = th->th_off << 2; /* Trim off leading empty mbufs. */ while (m && m->m_len == 0) m = m->m_next; if (m) { m_cur = m; } else { /* * No data? Highly unusual. We would expect to at * least see a TCP header in the mbuf. * As we have a pointer to the TCP header, I guess * we should just copy that. (???) */ fallback: bytes_to_copy = tcp_off; if (bytes_to_copy > M_SIZE(n)) bytes_to_copy = M_SIZE(n); bcopy(th, n->m_data, bytes_to_copy); n->m_len = bytes_to_copy; return; } /* * Find TCP header. Record the total number of bytes up to, * and including, the TCP header. */ while (m_cur) { if ((caddr_t) th >= (caddr_t) m_cur->m_data && (caddr_t) th < (caddr_t) (m_cur->m_data + m_cur->m_len)) break; bytes_to_copy += m_cur->m_len; m_cur = m_cur->m_next; } if (m_cur) bytes_to_copy += (caddr_t) th - (caddr_t) m_cur->m_data; else goto fallback; bytes_to_copy += tcp_off; /* * If we already want to copy more bytes than we can hold * in the destination mbuf, skip leading bytes and copy * what we can. * * Otherwise, consider trailing data. */ if (bytes_to_copy > M_SIZE(n)) { skip = bytes_to_copy - M_SIZE(n); bytes_to_copy = M_SIZE(n); } else { /* * Determine how much trailing data is in the chain. * We start with the length of this mbuf (the one * containing th) and subtract the size of the TCP * header (tcp_off) and the size of the data prior * to th (th - m_cur->m_data). * * This *should not* be negative, as the TCP code * should put the whole TCP header in a single * mbuf. But, it isn't a problem if it is. We will * simple work off our negative balance as we look * at subsequent mbufs. */ trailing_data = m_cur->m_len - tcp_off; trailing_data -= (caddr_t) th - (caddr_t) m_cur->m_data; m_cur = m_cur->m_next; while (m_cur) { trailing_data += m_cur->m_len; m_cur = m_cur->m_next; } if ((bytes_to_copy + trailing_data) > M_SIZE(n)) bytes_to_copy = M_SIZE(n); else bytes_to_copy += trailing_data; } m_copydata(m, skip, bytes_to_copy, n->m_data); n->m_len = bytes_to_copy; } void tcp_pcap_add(struct tcphdr *th, struct mbuf *m, struct mbufq *queue) { struct mbuf *n = NULL, *mhead; KASSERT(th, ("%s: called with th == NULL", __func__)); KASSERT(m, ("%s: called with m == NULL", __func__)); KASSERT(queue, ("%s: called with queue == NULL", __func__)); /* We only care about data packets. */ while (m && m->m_type != MT_DATA) m = m->m_next; /* We only need to do something if we still have an mbuf. */ if (!m) return; /* If we are not saving mbufs, return now. */ if (queue->mq_maxlen == 0) return; /* * Check to see if we will need to recycle mbufs. * * If we need to get rid of mbufs to stay below * our packet count, try to reuse the mbuf. Once * we already have a new mbuf (n), then we can * simply free subsequent mbufs. * * Note that most of the logic in here is to deal * with the reuse. If we are fine with constant * mbuf allocs/deallocs, we could ditch this logic. * But, it only seems to make sense to reuse * mbufs we already have. */ while (mbufq_full(queue)) { mhead = mbufq_dequeue(queue); if (n) { tcp_pcap_m_freem(mhead); } else { /* * If this held an external cluster, try to * detach the cluster. But, if we held the * last reference, go through the normal * free-ing process. */ if (mhead->m_flags & M_EXTPG) { /* Don't mess around with these. */ tcp_pcap_m_freem(mhead); continue; } else if (mhead->m_flags & M_EXT) { switch (mhead->m_ext.ext_type) { case EXT_SFBUF: /* Don't mess around with these. */ tcp_pcap_m_freem(mhead); continue; default: if (atomic_fetchadd_int( mhead->m_ext.ext_cnt, -1) == 1) { /* * We held the last reference * on this cluster. Restore * the reference count and put * it back in the pool. */ *(mhead->m_ext.ext_cnt) = 1; tcp_pcap_m_freem(mhead); continue; } /* * We were able to cleanly free the * reference. */ atomic_subtract_int( &tcp_pcap_clusters_referenced_cur, 1); tcp_pcap_alloc_reuse_ext++; break; } } else { tcp_pcap_alloc_reuse_mbuf++; } n = mhead; tcp_pcap_m_freem(n->m_next); m_init(n, M_NOWAIT, MT_DATA, 0); } } /* Check to see if we need to get a new mbuf. */ if (!n) { if (!(n = m_get(M_NOWAIT, MT_DATA))) return; tcp_pcap_alloc_new_mbuf++; } /* * What are we dealing with? If a cluster, attach it. Otherwise, * try to copy the data from the beginning of the mbuf to the * end of data. (There may be data between the start of the data * area and the current data pointer. We want to get this, because * it may contain header information that is useful.) * In cases where that isn't possible, settle for what we can * get. */ if ((m->m_flags & (M_EXT|M_EXTPG)) && tcp_pcap_take_cluster_reference()) { n->m_data = m->m_data; n->m_len = m->m_len; mb_dupcl(n, m); } else if (((m->m_data + m->m_len) - M_START(m)) <= M_SIZE(n)) { /* * At this point, n is guaranteed to be a normal mbuf * with no cluster and no packet header. Because the * logic in this code block requires this, the assert * is here to catch any instances where someone * changes the logic to invalidate that assumption. */ KASSERT((n->m_flags & (M_EXT | M_PKTHDR)) == 0, ("%s: Unexpected flags (%#x) for mbuf", __func__, n->m_flags)); n->m_data = n->m_dat + M_LEADINGSPACE_NOWRITE(m); n->m_len = m->m_len; if (m->m_flags & M_EXTPG) m_copydata(m, 0, m->m_len, n->m_data); else bcopy(M_START(m), n->m_dat, m->m_len + M_LEADINGSPACE_NOWRITE(m)); } else { /* * This is the case where we need to "settle for what * we can get". The most probable way to this code * path is that we've already taken references to the * maximum number of mbuf clusters we can, and the data * is too long to fit in an mbuf's internal storage. * Try for a "best fit". */ tcp_pcap_copy_bestfit(th, m, n); /* Don't try to get additional data. */ goto add_to_queue; } if (m->m_next) { n->m_next = m_copym(m->m_next, 0, M_COPYALL, M_NOWAIT); tcp_pcap_adj_cluster_reference(n->m_next, 1); } add_to_queue: /* Add the new mbuf to the list. */ if (mbufq_enqueue(queue, n)) { /* This shouldn't happen. If INVARIANTS is defined, panic. */ KASSERT(0, ("%s: mbufq was unexpectedly full!", __func__)); tcp_pcap_m_freem(n); } } void tcp_pcap_drain(struct mbufq *queue) { struct mbuf *m; while ((m = mbufq_dequeue(queue))) tcp_pcap_m_freem(m); } void tcp_pcap_tcpcb_init(struct tcpcb *tp) { mbufq_init(&(tp->t_inpkts), V_tcp_pcap_packets); mbufq_init(&(tp->t_outpkts), V_tcp_pcap_packets); } void tcp_pcap_set_sock_max(struct mbufq *queue, int newval) { queue->mq_maxlen = newval; while (queue->mq_len > queue->mq_maxlen) tcp_pcap_m_freem(mbufq_dequeue(queue)); } int tcp_pcap_get_sock_max(struct mbufq *queue) { return queue->mq_maxlen; } diff --git a/sys/netinet/tcp_stacks/rack.c b/sys/netinet/tcp_stacks/rack.c index 7105ce8bab83..c3aed32e02ea 100644 --- a/sys/netinet/tcp_stacks/rack.c +++ b/sys/netinet/tcp_stacks/rack.c @@ -1,20945 +1,20945 @@ /*- * 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 "opt_kern_tls.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 #ifdef TCP_ACCOUNTING #include #include #endif #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 #include #include #include #ifdef NETFLIX_SHARED_CWND #include #endif #ifdef TCPDEBUG #include #endif /* TCPDEBUG */ #ifdef TCP_OFFLOAD #include #endif #ifdef INET6 #include #endif #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_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 VNET_DECLARE(uint32_t, newreno_beta); VNET_DECLARE(uint32_t, newreno_beta_ecn); #define V_newreno_beta VNET(newreno_beta) #define V_newreno_beta_ecn VNET(newreno_beta_ecn) MALLOC_DEFINE(M_TCPFSB, "tcp_fsb", "TCP fast send block"); MALLOC_DEFINE(M_TCPDO, "tcp_do", "TCP deferred options"); 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_segment * 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 = 60000000; /* 0 - never fade, def 60,000,000 * - 60 seconds */ static uint8_t rack_req_measurements = 1; /* 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_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */ static int32_t rack_hw_pace_extra_slots = 2; /* 2 extra MSS time betweens */ static int32_t rack_hw_rate_caps = 1; /* 1; */ static int32_t rack_hw_rate_min = 0; /* 1500000;*/ static int32_t rack_hw_rate_to_low = 0; /* 1200000; */ static int32_t rack_hw_up_only = 1; static int32_t rack_stats_gets_ms_rtt = 1; static int32_t rack_prr_addbackmax = 2; static int32_t rack_do_hystart = 0; static int32_t rack_apply_rtt_with_reduced_conf = 0; static int32_t rack_pkt_delay = 1000; static int32_t rack_send_a_lot_in_prr = 1; static int32_t rack_min_to = 1000; /* Number of microsecond min timeout */ static int32_t rack_verbose_logging = 0; static int32_t rack_ignore_data_after_close = 1; static int32_t rack_enable_shared_cwnd = 1; static int32_t rack_use_cmp_acks = 1; static int32_t rack_use_fsb = 1; static int32_t rack_use_rfo = 1; static int32_t rack_use_rsm_rfo = 1; static int32_t rack_max_abc_post_recovery = 2; static int32_t rack_client_low_buf = 0; static int32_t rack_dsack_std_based = 0x3; /* bit field bit 1 sets rc_rack_tmr_std_based and bit 2 sets rc_rack_use_dsack */ #ifdef TCP_ACCOUNTING static int32_t rack_tcp_accounting = 0; #endif 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 = 250000; /* 250usec */ static int32_t rack_persist_max = 2000000; /* 2 Second in usec's */ static int32_t rack_sack_not_required = 1; /* 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_autosndbuf_inc = 20; /* In percentage form */ static int32_t rack_enobuf_hw_boost_mult = 2; /* How many times the hw rate we boost slot using time_between */ static int32_t rack_enobuf_hw_max = 12000; /* 12 ms in usecs */ static int32_t rack_enobuf_hw_min = 10000; /* 10 ms in usecs */ static int32_t rack_hw_rwnd_factor = 2; /* How many max_segs the rwnd must be before we hold off sending */ /* * 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 = 10000; /* 10ms */ static int32_t rack_rto_min = 30000; /* 30,000 usec same as main freebsd */ static int32_t rack_rto_max = 4000000; /* 4 seconds in usec's */ 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 = 40000; /* 40ms in usecs */ 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_def_profile = 0; static int32_t rack_lower_cwnd_at_tlp = 0; 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 usecs */ 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 = 40000; /* 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 = 250000; /* Must move at least 250ms (in microseconds) 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 multiplier */ static int32_t rack_gp_increase_per = 2; /* 2% increase in multiplier */ 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; static uint64_t rack_bw_rate_cap = 0; static uint32_t rack_trace_point_config = 0; static uint32_t rack_trace_point_bb_mode = 4; static int32_t rack_trace_point_count = 0; /* Weird delayed ack mode */ static int32_t rack_use_imac_dack = 0; /* Rack specific counters */ counter_u64_t rack_saw_enobuf; counter_u64_t rack_saw_enobuf_hw; counter_u64_t rack_saw_enetunreach; counter_u64_t rack_persists_sends; counter_u64_t rack_persists_acks; counter_u64_t rack_persists_loss; counter_u64_t rack_persists_lost_ends; #ifdef INVARIANTS counter_u64_t rack_adjust_map_bw; #endif /* 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_to_tot; counter_u64_t rack_hot_alloc; 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_multi_single_eq; counter_u64_t rack_proc_non_comp_ack; counter_u64_t rack_fto_send; counter_u64_t rack_fto_rsm_send; counter_u64_t rack_nfto_resend; counter_u64_t rack_non_fto_send; counter_u64_t rack_extended_rfo; 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_input_idle_reduces; counter_u64_t rack_collapsed_win; counter_u64_t rack_collapsed_win_seen; counter_u64_t rack_collapsed_win_rxt; counter_u64_t rack_collapsed_win_rxt_bytes; counter_u64_t rack_try_scwnd; counter_u64_t rack_hw_pace_init_fail; counter_u64_t rack_hw_pace_lost; counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE]; counter_u64_t rack_opts_arry[RACK_OPTS_SIZE]; #define RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2))) #define RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do { \ (tv) = (value) + slop; \ if ((u_long)(tv) < (u_long)(tvmin)) \ (tv) = (tvmin); \ if ((u_long)(tv) > (u_long)(tvmax)) \ (tv) = (tvmax); \ } while (0) 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, uint32_t th_ack, 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, uint32_t type, uint32_t ack, int ); static void rack_counter_destroy(void); static int rack_ctloutput(struct inpcb *inp, struct sockopt *sopt); 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, uint64_t *fill_override); 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_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, uint8_t quality); 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 inpcb *inp, struct sockopt *sopt); static void rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, int line, uint8_t quality); 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, int entered_rec, int dup_ack_struck); static void rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts, struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls); 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, uint32_t th_seq); static void rack_remxt_tmr(struct tcpcb *tp); static int rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt); 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, uint64_t ts, int32_t * lenp, uint16_t add_flag); static void rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag); 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); static int rack_set_profile(struct tcp_rack *rack, int prof); static void rack_apply_deferred_options(struct tcp_rack *rack); int32_t rack_clear_counter=0; static inline void rack_trace_point(struct tcp_rack *rack, int num) { if (((rack_trace_point_config == num) || (rack_trace_point_config = 0xffffffff)) && (rack_trace_point_bb_mode != 0) && (rack_trace_point_count > 0) && (rack->rc_tp->t_logstate == 0)) { int res; res = atomic_fetchadd_int(&rack_trace_point_count, -1); if (res > 0) { rack->rc_tp->t_logstate = rack_trace_point_bb_mode; } else { /* Loss a race assure its zero now */ rack_trace_point_count = 0; } } } static void rack_set_cc_pacing(struct tcp_rack *rack) { struct sockopt sopt; struct cc_newreno_opts opt; struct newreno old, *ptr; struct tcpcb *tp; int error; if (rack->rc_pacing_cc_set) return; tp = rack->rc_tp; - if (tp->cc_algo == NULL) { + if (tp->t_cc == NULL) { /* Tcb is leaving */ return; } rack->rc_pacing_cc_set = 1; - if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) { + if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) { /* Not new-reno we can't play games with beta! */ goto out; } - ptr = ((struct newreno *)tp->ccv->cc_data); + ptr = ((struct newreno *)tp->t_ccv.cc_data); if (CC_ALGO(tp)->ctl_output == NULL) { /* Huh, why does new_reno no longer have a set function? */ goto out; } if (ptr == NULL) { /* Just the default values */ old.beta = V_newreno_beta_ecn; old.beta_ecn = V_newreno_beta_ecn; old.newreno_flags = 0; } else { old.beta = ptr->beta; old.beta_ecn = ptr->beta_ecn; old.newreno_flags = ptr->newreno_flags; } sopt.sopt_valsize = sizeof(struct cc_newreno_opts); sopt.sopt_dir = SOPT_SET; opt.name = CC_NEWRENO_BETA; opt.val = rack->r_ctl.rc_saved_beta.beta; - error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt); + error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt); if (error) { goto out; } /* * Hack alert we need to set in our newreno_flags * so that Abe behavior is also applied. */ - ((struct newreno *)tp->ccv->cc_data)->newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED; + ((struct newreno *)tp->t_ccv.cc_data)->newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED; opt.name = CC_NEWRENO_BETA_ECN; opt.val = rack->r_ctl.rc_saved_beta.beta_ecn; - error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt); + error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt); if (error) { goto out; } /* Save off the original values for restoral */ memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno)); out: if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; struct timeval tv; - ptr = ((struct newreno *)tp->ccv->cc_data); + ptr = ((struct newreno *)tp->t_ccv.cc_data); memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.timeStamp = tcp_get_usecs(&tv); if (ptr) { log.u_bbr.flex1 = ptr->beta; log.u_bbr.flex2 = ptr->beta_ecn; log.u_bbr.flex3 = ptr->newreno_flags; } log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta; log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn; log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags; log.u_bbr.flex7 = rack->gp_ready; log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= rack->use_fixed_rate; log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= rack->rc_pacing_cc_set; log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex8 = 3; tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, error, 0, &log, false, NULL, NULL, 0, &tv); } } static void rack_undo_cc_pacing(struct tcp_rack *rack) { struct newreno old, *ptr; struct tcpcb *tp; if (rack->rc_pacing_cc_set == 0) return; tp = rack->rc_tp; rack->rc_pacing_cc_set = 0; - if (tp->cc_algo == NULL) + if (tp->t_cc == NULL) /* Tcb is leaving */ return; - if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) { + if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) { /* Not new-reno nothing to do! */ return; } - ptr = ((struct newreno *)tp->ccv->cc_data); + ptr = ((struct newreno *)tp->t_ccv.cc_data); if (ptr == NULL) { /* * This happens at rack_fini() if the * cc module gets freed on us. In that * case we loose our "new" settings but * thats ok, since the tcb is going away anyway. */ return; } /* Grab out our set values */ memcpy(&old, ptr, sizeof(struct newreno)); /* Copy back in the original values */ memcpy(ptr, &rack->r_ctl.rc_saved_beta, sizeof(struct newreno)); /* Now save back the values we had set in (for when pacing is restored) */ memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno)); if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; struct timeval tv; - ptr = ((struct newreno *)tp->ccv->cc_data); + ptr = ((struct newreno *)tp->t_ccv.cc_data); memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.flex1 = ptr->beta; log.u_bbr.flex2 = ptr->beta_ecn; log.u_bbr.flex3 = ptr->newreno_flags; log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta; log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn; log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags; log.u_bbr.flex7 = rack->gp_ready; log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= rack->use_fixed_rate; log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= rack->rc_pacing_cc_set; log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex8 = 4; tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0, 0, &log, false, NULL, NULL, 0, &tv); } } #ifdef NETFLIX_PEAKRATE static inline void rack_update_peakrate_thr(struct tcpcb *tp) { /* Keep in mind that t_maxpeakrate is in B/s. */ uint64_t peak; peak = uqmax((tp->t_maxseg * 2), (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC)); tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX); } #endif 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_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_to_tot); counter_u64_zero(rack_saw_enobuf); counter_u64_zero(rack_saw_enobuf_hw); counter_u64_zero(rack_saw_enetunreach); counter_u64_zero(rack_persists_sends); counter_u64_zero(rack_persists_acks); counter_u64_zero(rack_persists_loss); counter_u64_zero(rack_persists_lost_ends); #ifdef INVARIANTS counter_u64_zero(rack_adjust_map_bw); #endif 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_fto_send); counter_u64_zero(rack_fto_rsm_send); counter_u64_zero(rack_extended_rfo); counter_u64_zero(rack_hw_pace_init_fail); counter_u64_zero(rack_hw_pace_lost); counter_u64_zero(rack_non_fto_send); counter_u64_zero(rack_nfto_resend); 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_multi_single_eq); counter_u64_zero(rack_proc_non_comp_ack); 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_try_scwnd); counter_u64_zero(rack_collapsed_win); counter_u64_zero(rack_collapsed_win_rxt); counter_u64_zero(rack_collapsed_win_seen); counter_u64_zero(rack_collapsed_win_rxt_bytes); } 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_features; struct sysctl_oid *rack_measure; struct sysctl_oid *rack_probertt; struct sysctl_oid *rack_hw_pacing; struct sysctl_oid *rack_tracepoint; 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"); rack_tracepoint = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "tp", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Rack tracepoint facility"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tracepoint), OID_AUTO, "number", CTLFLAG_RW, &rack_trace_point_config, 0, "What is the trace point number to activate (0=none, 0xffffffff = all)?"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tracepoint), OID_AUTO, "bbmode", CTLFLAG_RW, &rack_trace_point_bb_mode, 4, "What is BB logging mode that is activated?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tracepoint), OID_AUTO, "count", CTLFLAG_RW, &rack_trace_point_count, 0, "How many connections will have BB logging turned on that hit the tracepoint?"); /* 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_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"); SYSCTL_ADD_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "rate_cap", CTLFLAG_RW, &rack_bw_rate_cap, 0, "If set we apply this value to the absolute rate cap used by pacing"); SYSCTL_ADD_U8(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "req_measure_cnt", CTLFLAG_RW, &rack_req_measurements, 1, "If doing dynamic pacing, how many measurements must be in before we start pacing?"); /* Hardware pacing */ rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "hdwr_pacing", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Pacing related Controls"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "rwnd_factor", CTLFLAG_RW, &rack_hw_rwnd_factor, 2, "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW, &rack_enobuf_hw_boost_mult, 2, "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "pace_enobuf_max", CTLFLAG_RW, &rack_enobuf_hw_max, 2, "What is the max boost the pacing time if we see a ENOBUFS?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "pace_enobuf_min", CTLFLAG_RW, &rack_enobuf_hw_min, 2, "What is the min boost the pacing time if we see a ENOBUFS?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "enable", CTLFLAG_RW, &rack_enable_hw_pacing, 0, "Should RACK attempt to use hw pacing?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "rate_cap", CTLFLAG_RW, &rack_hw_rate_caps, 1, "Does the highest hardware pacing rate cap the rate we will send at??"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "rate_min", CTLFLAG_RW, &rack_hw_rate_min, 0, "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "rate_to_low", CTLFLAG_RW, &rack_hw_rate_to_low, 0, "If we fall below this rate, dis-engage hw pacing?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "up_only", CTLFLAG_RW, &rack_hw_up_only, 1, "Do we allow hw pacing to lower the rate selected?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "extra_mss_precise", CTLFLAG_RW, &rack_hw_pace_extra_slots, 2, "If the rates between software and hardware match precisely how many extra time_betweens do we get?"); 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 multiplier 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 highest 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 highest 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, "post_rec_labc", CTLFLAG_RW, &rack_max_abc_post_recovery, 2, "Since we do early recovery, do we override the l_abc to a value, if so what?"); 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, 10000, "TLP minimum timeout per the specification (in microseconds)"); 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_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, 60000000, "Does reorder detection fade, if so how many microseconds (0 means never)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "pktdelay", CTLFLAG_RW, &rack_pkt_delay, 1000, "Extra RACK time (in microseconds) 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, 250000, "What is the minimum time in microseconds between persists"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "persmax", CTLFLAG_RW, &rack_persist_max, 2000000, "What is the largest delay in microseconds between persists"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "delayed_ack", CTLFLAG_RW, &rack_delayed_ack_time, 40000, "Delayed ack time (40ms in microseconds)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "minrto", CTLFLAG_RW, &rack_rto_min, 30000, "Minimum RTO in microseconds -- 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, 4000000, "Maximum RTO in microseconds -- 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, 1000, "Minimum rack timeout in microseconds"); /* 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"); /* Features */ rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "features", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Feature controls"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_features), OID_AUTO, "cmpack", CTLFLAG_RW, &rack_use_cmp_acks, 1, "Should RACK have LRO send compressed acks"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_features), OID_AUTO, "fsb", CTLFLAG_RW, &rack_use_fsb, 1, "Should RACK use the fast send block?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_features), OID_AUTO, "rfo", CTLFLAG_RW, &rack_use_rfo, 1, "Should RACK use rack_fast_output()?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_features), OID_AUTO, "rsmrfo", CTLFLAG_RW, &rack_use_rsm_rfo, 1, "Should RACK use rack_fast_rsm_output()?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_features), 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_features), OID_AUTO, "hystartplusplus", CTLFLAG_RW, &rack_do_hystart, 0, "Should RACK enable HyStart++ on connections?"); /* 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"); #ifdef TCP_ACCOUNTING SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "tcp_acct", CTLFLAG_RW, &rack_tcp_accounting, 0, "Should we turn on TCP accounting for all rack sessions?"); #endif SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW, &rack_apply_rtt_with_reduced_conf, 0, "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW, &rack_dsack_std_based, 3, "How do we process dsack with respect to rack timers, bit field, 3 is standards based?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "prr_addback_max", CTLFLAG_RW, &rack_prr_addbackmax, 2, "What is the maximum number of MSS we allow to be added back if prr can't send all its data?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "stats_gets_ms", CTLFLAG_RW, &rack_stats_gets_ms_rtt, 1, "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "clientlowbuf", CTLFLAG_RW, &rack_client_low_buf, 0, "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "defprofile", CTLFLAG_RW, &rack_def_profile, 0, "Should RACK use a default profile (0=no, num == profile num)?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "shared_cwnd", CTLFLAG_RW, &rack_enable_shared_cwnd, 1, "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, "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, 1, "Do we allow rack to run on connections not supporting SACK"); 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, "autoscale", CTLFLAG_RW, &rack_autosndbuf_inc, 20, "What percentage should rack scale up its snd buffer by?"); /* 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 positions 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 positions 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_fto_send = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "fto_send", CTLFLAG_RD, &rack_fto_send, "Total number of rack_fast_output sends"); rack_fto_rsm_send = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "fto_rsm_send", CTLFLAG_RD, &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends"); rack_nfto_resend = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "nfto_resend", CTLFLAG_RD, &rack_nfto_resend, "Total number of rack_output retransmissions"); rack_non_fto_send = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "nfto_send", CTLFLAG_RD, &rack_non_fto_send, "Total number of rack_output first sends"); rack_extended_rfo = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "rfo_extended", CTLFLAG_RD, &rack_extended_rfo, "Total number of times we extended rfo"); rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "hwpace_init_fail", CTLFLAG_RD, &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing"); rack_hw_pace_lost = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "hwpace_lost", CTLFLAG_RD, &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing"); 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_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_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 sends returned enobuf for non-hdwr paced connections"); rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD, &rack_saw_enobuf_hw, "Total number of times a send returned enobuf for hdwr paced connections"); 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 send received a enetunreachable"); rack_hot_alloc = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "alloc_hot", CTLFLAG_RD, &rack_hot_alloc, "Total allocations from the top of our list"); 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_persists_sends = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "persist_sends", CTLFLAG_RD, &rack_persists_sends, "Number of times we sent a persist probe"); rack_persists_acks = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "persist_acks", CTLFLAG_RD, &rack_persists_acks, "Number of times a persist probe was acked"); rack_persists_loss = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "persist_loss", CTLFLAG_RD, &rack_persists_loss, "Number of times we detected a lost persist probe (no ack)"); rack_persists_lost_ends = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "persist_loss_ends", CTLFLAG_RD, &rack_persists_lost_ends, "Number of lost persist probe (no ack) that the run ended with a PERSIST abort"); #ifdef INVARIANTS rack_adjust_map_bw = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "map_adjust_req", CTLFLAG_RD, &rack_adjust_map_bw, "Number of times we hit the case where the sb went up and down on a sendmap entry"); #endif rack_multi_single_eq = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD, &rack_multi_single_eq, "Number of compressed acks total represented"); rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "cmp_ack_not", CTLFLAG_RD, &rack_proc_non_comp_ack, "Number of non compresseds acks that we processed"); 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_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_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_seen = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "collapsed_win_seen", CTLFLAG_RD, &rack_collapsed_win_seen, "Total number of collapsed window events seen (where our window shrinks)"); 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 window events where we mark packets"); rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD, &rack_collapsed_win_rxt, "Total number of packets that were retransmitted"); rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD, &rack_collapsed_win_rxt_bytes, "Total number of bytes that were retransmitted"); 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"); 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_FASTRECOVERY(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)rack->rc_tp->t_srtt; bw *= (uint64_t)USECS_IN_SECOND; bw /= srtt; if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap)) bw = rack->r_ctl.bw_rate_cap; return (bw); } else { uint64_t bw; if (rack->r_ctl.num_measurements >= 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_measurements; } #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 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap)) bw = rack->r_ctl.bw_rate_cap; 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_FASTRECOVERY(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 void rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6) { /* * Types of logs (mod value) * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit. * 2 = a dsack round begins, persist is reset to 16. * 3 = a dsack round ends * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh. */ 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 = rack->rc_rack_tmr_std_based; log.u_bbr.flex1 <<= 1; log.u_bbr.flex1 |= rack->rc_rack_use_dsack; log.u_bbr.flex1 <<= 1; log.u_bbr.flex1 |= rack->rc_dsack_round_seen; log.u_bbr.flex2 = rack->r_ctl.dsack_round_end; log.u_bbr.flex3 = rack->r_ctl.num_dsack; log.u_bbr.flex4 = flex4; log.u_bbr.flex5 = flex5; log.u_bbr.flex6 = flex6; log.u_bbr.flex7 = rack->r_ctl.dsack_persist; log.u_bbr.flex8 = mod; log.u_bbr.timeStamp = tcp_get_usecs(&tv); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, RACK_DSACK_HANDLING, 0, 0, &log, false, &tv); } } static void rack_log_hdwr_pacing(struct tcp_rack *rack, uint64_t rate, uint64_t hw_rate, int line, int error, uint16_t mod) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; const struct ifnet *ifp; memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff); log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff); if (rack->r_ctl.crte) { ifp = rack->r_ctl.crte->ptbl->rs_ifp; } else if (rack->rc_inp->inp_route.ro_nh && rack->rc_inp->inp_route.ro_nh->nh_ifp) { ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp; } else ifp = NULL; if (ifp) { 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.flex7 = mod; 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; log.u_bbr.delRate = rack->r_ctl.crte_prev_rate; if (rack->r_ctl.crte) log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate; else log.u_bbr.cur_del_rate = 0; log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req; 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 uint64_t rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped) { /* * We allow rack_per_of_gp_xx to dictate our bw rate we want. */ uint64_t bw_est, high_rate; 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; if (rack->r_rack_hw_rate_caps) { /* Rate caps are in place */ if (rack->r_ctl.crte != NULL) { /* We have a hdwr rate already */ high_rate = tcp_hw_highest_rate(rack->r_ctl.crte); if (bw_est >= high_rate) { /* We are capping bw at the highest rate table entry */ rack_log_hdwr_pacing(rack, bw_est, high_rate, __LINE__, 0, 3); bw_est = high_rate; if (capped) *capped = 1; } } else if ((rack->rack_hdrw_pacing == 0) && (rack->rack_hdw_pace_ena) && (rack->rack_attempt_hdwr_pace == 0) && (rack->rc_inp->inp_route.ro_nh != NULL) && (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) { /* * Special case, we have not yet attempted hardware * pacing, and yet we may, when we do, find out if we are * above the highest rate. We need to know the maxbw for the interface * in question (if it supports ratelimiting). We get back * a 0, if the interface is not found in the RL lists. */ high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp); if (high_rate) { /* Yep, we have a rate is it above this rate? */ if (bw_est > high_rate) { bw_est = high_rate; if (capped) *capped = 1; } } } } 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 = tcp_in_hpts(rack->rc_inp); 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.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; 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 = rack->rc_tp->t_srtt; log.u_bbr.flex2 = to; 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 = tcp_in_hpts(rack->rc_inp); 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.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; log.u_bbr.cwnd_gain = rack->rc_has_collapsed; log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift; log.u_bbr.lost = rack_rto_min; 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 = tcp_in_hpts(rack->rc_inp); 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); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; 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_map_chg(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *prev, struct rack_sendmap *rsm, struct rack_sendmap *next, int flag, uint32_t th_ack, 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.flex8 = flag; log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); log.u_bbr.cur_del_rate = (uint64_t)prev; log.u_bbr.delRate = (uint64_t)rsm; log.u_bbr.rttProp = (uint64_t)next; log.u_bbr.flex7 = 0; if (prev) { log.u_bbr.flex1 = prev->r_start; log.u_bbr.flex2 = prev->r_end; log.u_bbr.flex7 |= 0x4; } if (rsm) { log.u_bbr.flex3 = rsm->r_start; log.u_bbr.flex4 = rsm->r_end; log.u_bbr.flex7 |= 0x2; } if (next) { log.u_bbr.flex5 = next->r_start; log.u_bbr.flex6 = next->r_end; log.u_bbr.flex7 |= 0x1; } log.u_bbr.applimited = line; log.u_bbr.pkts_out = th_ack; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); if (rack->rack_no_prr) log.u_bbr.lost = 0; else log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, TCP_LOG_MAPCHG, 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 = tcp_in_hpts(rack->rc_inp); log.u_bbr.flex1 = t; log.u_bbr.flex2 = len; log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt; log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest; log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest; log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt; log.u_bbr.flex7 = conf; log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot; log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt; 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; /* We loose any upper of the 24 bits */ log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags; } else { /* Its a SYN */ log.u_bbr.pkt_epoch = rack->rc_tp->iss; log.u_bbr.lost = 0; log.u_bbr.cwnd_gain = 0; log.u_bbr.pacing_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; log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); log.u_bbr.bw_inuse <<= 32; if (rsm) log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]); 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; 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.flex6 = rack->rc_tp->t_rxtcur; log.u_bbr.flex7 = 1; 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); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; /* * We capture in delRate the upper 32 bits as * the confidence level we had declared, and the * lower 32 bits as the actual RTT using the arrival * timestamp. */ log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence; log.u_bbr.delRate <<= 32; log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt; /* Lets capture all the things that make up t_rtxcur */ log.u_bbr.applimited = rack_rto_min; log.u_bbr.epoch = rack_rto_max; log.u_bbr.lt_epoch = rack->r_ctl.timer_slop; log.u_bbr.lost = rack_rto_min; log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop); log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp); log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec; log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC; log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec; 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 void rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where) { if (rack_verbose_logging && (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; log.u_bbr.flex2 = send_time; log.u_bbr.flex3 = ack_time; log.u_bbr.flex4 = where; log.u_bbr.flex7 = 2; log.u_bbr.timeStamp = tcp_get_usecs(&tv); 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 = tcp_in_hpts(rack->rc_inp); 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); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; 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 = tcp_in_hpts(rack->rc_inp); 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); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; 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, int nsegs) { 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.flex6 = nsegs; log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs; log.u_bbr.flex7 = rack->rc_ack_can_sendout_data; /* Do we have ack-can-send set */ log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= rack->r_fast_output; /* is fast output primed */ log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= rack->r_wanted_output; /* Do we want output */ log.u_bbr.flex8 = rack->rc_in_persist; log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); 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.use_lt_bw = rack->r_ent_rec_ns; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->r_might_revert; log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; 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_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log, 0, sizeof(log)); 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 = arg1; log.u_bbr.flex5 = arg2; log.u_bbr.flex6 = arg3; 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); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.applimited = rack->r_ctl.rc_sacked; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv, &tptosocket(tp)->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 = tcp_in_hpts(rack->rc_inp); 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); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; log.u_bbr.cwnd_gain = rack->rc_has_collapsed; 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 = tcp_in_hpts(rack->rc_inp); 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); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; 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.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; 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, int line) { 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.flex7 = line; 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); log.u_bbr.use_lt_bw = rack->r_ent_rec_ns; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->r_might_revert; 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_fto_send); counter_u64_free(rack_fto_rsm_send); counter_u64_free(rack_nfto_resend); counter_u64_free(rack_hw_pace_init_fail); counter_u64_free(rack_hw_pace_lost); counter_u64_free(rack_non_fto_send); counter_u64_free(rack_extended_rfo); 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_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_to_tot); counter_u64_free(rack_saw_enobuf); counter_u64_free(rack_saw_enobuf_hw); counter_u64_free(rack_saw_enetunreach); counter_u64_free(rack_hot_alloc); 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_multi_single_eq); counter_u64_free(rack_proc_non_comp_ack); 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_sack_skipped_acked); counter_u64_free(rack_sack_splits); counter_u64_free(rack_input_idle_reduces); counter_u64_free(rack_collapsed_win); counter_u64_free(rack_collapsed_win_rxt); counter_u64_free(rack_collapsed_win_rxt_bytes); counter_u64_free(rack_collapsed_win_seen); counter_u64_free(rack_try_scwnd); counter_u64_free(rack_persists_sends); counter_u64_free(rack_persists_acks); counter_u64_free(rack_persists_loss); counter_u64_free(rack_persists_lost_ends); #ifdef INVARIANTS counter_u64_free(rack_adjust_map_bw); #endif 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; /* * First get the top of the list it in * theory is the "hottest" rsm we have, * possibly just freed by ack processing. */ if (rack->rc_free_cnt > rack_free_cache) { rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext); counter_u64_add(rack_hot_alloc, 1); rack->rc_free_cnt--; return (rsm); } /* * Once we get under our free cache we probably * no longer have a "hot" one available. Lets * get one from UMA. */ 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); } /* * Dig in to our aux rsm's (the last two) since * UMA failed to get us one. */ 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_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; memset(rsm, 0, sizeof(struct rack_sendmap)); TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext); rack->rc_free_cnt++; } static void rack_free_trim(struct tcp_rack *rack) { struct rack_sendmap *rsm; /* * Free up all the tail entries until * we get our list down to the limit. */ while (rack->rc_free_cnt > rack_free_cache) { rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head); TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext); rack->rc_free_cnt--; 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 20MSS). */ 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)tp->t_srtt; 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, uint8_t *quality) { 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 */ *quality = RACK_QUALITY_ALLACKED; 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 && (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) { /* * We are up to the app limited send point * we have to measure irrespective of the time.. */ *quality = RACK_QUALITY_APPLIMITED; 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) { *quality = RACK_QUALITY_HIGH; 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 multiplier * 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_QUALITY_PROBERTT); } 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__, RACK_QUALITY_PROBERTT); } 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; + rack->rc_tp->t_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 multiplier 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 multiplier */ 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 multiplier */ 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 multiplier (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 multiplier */ 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, uint8_t quality) { uint64_t tim, bytes_ps, ltim, stim, utim; uint32_t segsiz, bytes, reqbytes, us_cts; int32_t gput, new_rtt_diff, timely_says; uint64_t resid_bw, subpart = 0, addpart = 0, srtt; int did_add = 0; 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 (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 get a msec time ltim too for the old stuff */ 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, quality); 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, quality); 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, quality); 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, quality); 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, quality); 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 measurement */ if (bytes_ps) { rack->r_ctl.gp_bw = bytes_ps; rack->rc_gp_filled = 1; rack->r_ctl.num_measurements = 1; rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL); } else { rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes, rack->r_ctl.rc_app_limited_cnt, 0, 0, 10, __LINE__, NULL, quality); } if (tcp_in_hpts(rack->rc_inp) && (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); rack->r_ctl.rc_hpts_flags = 0; rack->r_ctl.rc_last_output_to = 0; } did_add = 2; } else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) { /* Still a small number run an average */ rack->r_ctl.gp_bw += bytes_ps; addpart = rack->r_ctl.num_measurements; rack->r_ctl.num_measurements++; if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) { /* We have collected enough to move forward */ rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements; } did_add = 3; } 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. */ if (rack->r_ctl.num_measurements < 0xff) { rack->r_ctl.num_measurements++; } srtt = (uint64_t)tp->t_srtt; 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; 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; did_add = 1; } 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 measurement 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. */ did_add = 1; resid_bw = rack->r_ctl.gp_bw - subpart; rack->r_ctl.gp_bw = resid_bw + addpart; } } } if ((rack->gp_ready == 0) && (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) { /* We have enough measurements now */ rack->gp_ready = 1; rack_set_cc_pacing(rack); if (rack->defer_options) rack_apply_deferred_options(rack); } rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim, rack_get_bw(rack), 22, did_add, NULL, quality); /* 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, quality); /* 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_end - 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; 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_end - 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_end - th_ack); if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) { /* * We don't have enough to make a measurement. */ tp->t_flags &= ~TF_GPUTINPROG; rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq, 0, 0, 0, 6, __LINE__, NULL, quality); return; } } if (tp->t_state >= TCPS_FIN_WAIT_1) { /* * We will get no more data into the SB * this means we need to have the data available * before we start a measurement. */ if (sbavail(&tptosocket(tp)->so_snd) < (tp->gput_ack - tp->gput_seq)) { /* Nope not enough data. */ return; } } tp->t_flags |= TF_GPUTINPROG; /* * 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 = (uint32_t)rsm->r_ack_arrival; else rack->app_limited_needs_set = 1; rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; } else { /* * If we don't find the rsm due to some * send-limit set the current time, which * basically disables the send-limit. */ struct timeval tv; microuptime(&tv); rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv); } 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, quality); } } /* * CC wrapper hook functions */ static void rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery) { uint32_t prior_cwnd, acked; struct tcp_log_buffer *lgb = NULL; uint8_t labc_to_use, quality; INP_WLOCK_ASSERT(tptoinpcb(tp)); - tp->ccv->nsegs = nsegs; - acked = tp->ccv->bytes_this_ack = (th_ack - tp->snd_una); + tp->t_ccv.nsegs = nsegs; + acked = tp->t_ccv.bytes_this_ack = (th_ack - tp->snd_una); 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 (tp->t_ccv.bytes_this_ack > max) { + tp->t_ccv.bytes_this_ack = max; } } #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 quality = RACK_QUALITY_NONE; if ((tp->t_flags & TF_GPUTINPROG) && rack_enough_for_measurement(tp, rack, th_ack, &quality)) { /* Measure the Goodput */ rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality); #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. */ rack_update_peakrate_thr(tp); } #endif } /* Which way our we limited, if not cwnd limited no advance in CA */ if (tp->snd_cwnd <= tp->snd_wnd) - tp->ccv->flags |= CCF_CWND_LIMITED; + tp->t_ccv.flags |= CCF_CWND_LIMITED; else - tp->ccv->flags &= ~CCF_CWND_LIMITED; + tp->t_ccv.flags &= ~CCF_CWND_LIMITED; if (tp->snd_cwnd > tp->snd_ssthresh) { - tp->t_bytes_acked += min(tp->ccv->bytes_this_ack, + tp->t_bytes_acked += min(tp->t_ccv.bytes_this_ack, nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp)); /* For the setting of a window past use the actual scwnd we are using */ 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; + tp->t_ccv.flags |= CCF_ABC_SENTAWND; } } else { - tp->ccv->flags &= ~CCF_ABC_SENTAWND; + tp->t_ccv.flags &= ~CCF_ABC_SENTAWND; tp->t_bytes_acked = 0; } prior_cwnd = tp->snd_cwnd; if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec || (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf))) labc_to_use = rack->rc_labc; else labc_to_use = rack_max_abc_post_recovery; if (rack_verbose_logging && (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.timeStamp = tcp_get_usecs(&tv); log.u_bbr.flex1 = th_ack; - log.u_bbr.flex2 = tp->ccv->flags; - log.u_bbr.flex3 = tp->ccv->bytes_this_ack; - log.u_bbr.flex4 = tp->ccv->nsegs; + log.u_bbr.flex2 = tp->t_ccv.flags; + log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack; + log.u_bbr.flex4 = tp->t_ccv.nsegs; log.u_bbr.flex5 = labc_to_use; log.u_bbr.flex6 = prior_cwnd; log.u_bbr.flex7 = V_tcp_do_newsack; log.u_bbr.flex8 = 1; lgb = tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0, 0, &log, false, NULL, NULL, 0, &tv); } if (CC_ALGO(tp)->ack_received != NULL) { /* XXXLAS: Find a way to live without this */ - tp->ccv->curack = th_ack; - tp->ccv->labc = labc_to_use; - tp->ccv->flags |= CCF_USE_LOCAL_ABC; - CC_ALGO(tp)->ack_received(tp->ccv, type); + tp->t_ccv.curack = th_ack; + tp->t_ccv.labc = labc_to_use; + tp->t_ccv.flags |= CCF_USE_LOCAL_ABC; + CC_ALGO(tp)->ack_received(&tp->t_ccv, type); } if (lgb) { lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd; } if (rack->r_must_retran) { if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) { /* * We now are beyond the rxt point so lets disable * the flag. */ rack->r_ctl.rc_out_at_rto = 0; rack->r_must_retran = 0; } else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) { /* * Only decrement the rc_out_at_rto if the cwnd advances * at least a whole segment. Otherwise next time the peer * acks, we won't be able to send this generaly happens * when we are in Congestion Avoidance. */ if (acked <= rack->r_ctl.rc_out_at_rto){ rack->r_ctl.rc_out_at_rto -= acked; } else { rack->r_ctl.rc_out_at_rto = 0; } } } #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 tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; INP_WLOCK_ASSERT(tptoinpcb(tp)); /* * 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, uint32_t th_ack) { struct tcp_rack *rack; uint32_t orig_cwnd; orig_cwnd = tp->snd_cwnd; INP_WLOCK_ASSERT(tptoinpcb(tp)); rack = (struct tcp_rack *)tp->t_fb_ptr; /* only alert CC if we alerted when we entered */ if (CC_ALGO(tp)->post_recovery != NULL) { - tp->ccv->curack = th_ack; - CC_ALGO(tp)->post_recovery(tp->ccv); + tp->t_ccv.curack = th_ack; + CC_ALGO(tp)->post_recovery(&tp->t_ccv); if (tp->snd_cwnd < tp->snd_ssthresh) { /* * Rack has burst control and pacing * so lets not set this any lower than * snd_ssthresh per RFC-6582 (option 2). */ tp->snd_cwnd = tp->snd_ssthresh; } } if (rack_verbose_logging && (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.timeStamp = tcp_get_usecs(&tv); log.u_bbr.flex1 = th_ack; - log.u_bbr.flex2 = tp->ccv->flags; - log.u_bbr.flex3 = tp->ccv->bytes_this_ack; - log.u_bbr.flex4 = tp->ccv->nsegs; + log.u_bbr.flex2 = tp->t_ccv.flags; + log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack; + log.u_bbr.flex4 = tp->t_ccv.nsegs; log.u_bbr.flex5 = V_tcp_abc_l_var; log.u_bbr.flex6 = orig_cwnd; log.u_bbr.flex7 = V_tcp_do_newsack; log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex8 = 2; tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0, 0, &log, false, NULL, NULL, 0, &tv); } if ((rack->rack_no_prr == 0) && (rack->no_prr_addback == 0) && (rack->r_ctl.rc_prr_sndcnt > 0)) { /* * Suck the next prr cnt back into cwnd, but * only do that if we are not application limited. */ if (ctf_outstanding(tp) <= sbavail(&tptosocket(tp)->so_snd)) { /* * We are allowed to add back to the cwnd the amount we did * not get out if: * a) no_prr_addback is off. * b) we are not app limited * c) we are doing prr * * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none). */ tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax), rack->r_ctl.rc_prr_sndcnt); } rack->r_ctl.rc_prr_sndcnt = 0; rack_log_to_prr(rack, 1, 0, __LINE__); } rack_log_to_prr(rack, 14, orig_cwnd, __LINE__); tp->snd_recover = tp->snd_una; if (rack->r_ctl.dsack_persist) { rack->r_ctl.dsack_persist--; if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) { rack->r_ctl.num_dsack = 0; } rack_log_dsack_event(rack, 1, __LINE__, 0, 0); } EXIT_RECOVERY(tp->t_flags); } static void rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line) { struct tcp_rack *rack; uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd; INP_WLOCK_ASSERT(tptoinpcb(tp)); #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type); #endif if (IN_RECOVERY(tp->t_flags) == 0) { in_rec_at_entry = 0; ssthresh_enter = tp->snd_ssthresh; cwnd_enter = tp->snd_cwnd; } else in_rec_at_entry = 1; 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, in_rec_at_entry, line); } 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(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); orig_cwnd = tp->snd_cwnd; tp->snd_cwnd = ctf_fixed_maxseg(tp); rack_log_to_prr(rack, 16, orig_cwnd, line); 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 ((CC_ALGO(tp)->cong_signal != NULL) && (type != CC_RTO)){ - tp->ccv->curack = ack; - CC_ALGO(tp)->cong_signal(tp->ccv, type); + tp->t_ccv.curack = ack; + CC_ALGO(tp)->cong_signal(&tp->t_ccv, type); } if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) { rack_log_to_prr(rack, 15, cwnd_enter, line); rack->r_ctl.dsack_byte_cnt = 0; rack->r_ctl.retran_during_recovery = 0; rack->r_ctl.rc_cwnd_at_erec = cwnd_enter; rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter; rack->r_ent_rec_ns = 1; } } static inline void rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp) { uint32_t i_cwnd; INP_WLOCK_ASSERT(tptoinpcb(tp)); #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); + CC_ALGO(tp)->after_idle(&tp->t_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 different 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). */ 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; } if (rack->rc_rack_tmr_std_based == 0) { thresh = srtt + rack->r_ctl.rc_pkt_delay; } else { /* Standards based pkt-delay is 1/4 srtt */ thresh = srtt + (srtt >> 2); } if (lro && (rack->rc_rack_tmr_std_based == 0)) { /* 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); } if (rack->rc_rack_use_dsack && lro && (rack->r_ctl.num_dsack > 0)) { /* * We only increase the reordering window if we * have seen reordering we have a DSACK count. */ thresh += rack->r_ctl.num_dsack * (srtt >> 2); rack_log_dsack_event(rack, 4, __LINE__, srtt, thresh); } /* SRTT * 2 is the ceiling */ if (thresh > (srtt * 2)) { thresh = srtt * 2; } /* And we don't want it above the RTO max either */ if (thresh > rack_rto_max) { thresh = rack_rto_max; } rack_log_dsack_event(rack, 6, __LINE__, srtt, thresh); 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); 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. */ 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; idx = rsm->r_rtr_cnt - 1; nidx = prsm->r_rtr_cnt - 1; if (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; 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. */ alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; if (alt_thresh > thresh) thresh = alt_thresh; } } /* Not above an RTO */ if (thresh > tp->t_rxtcur) { thresh = 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 (tp->t_srtt); } 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, ((uint32_t)rsm->r_tim_lastsent[idx]))) { return (NULL); } if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) { return (NULL); } /* Ok if we reach here we are over-due and this guy can be sent */ rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__); 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 = (tp->t_srtt + (tp->t_rttvar << 2)); RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop); 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 or no rack */ activate_rxt: time_since_sent = 0; rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rsm) { /* * Should we discount the RTX timer any? * * We want to discount it the smallest amount. * If a timer (Rack/TLP or RXT) has gone off more * recently thats the discount we want to use (now - timer time). * If the retransmit of the oldest packet was more recent then * we want to use that (now - oldest-packet-last_transmit_time). * */ idx = rsm->r_rtr_cnt - 1; if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx]))) tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time; else tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx]; if (TSTMP_GT(cts, tstmp_touse)) time_since_sent = cts - tstmp_touse; } if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&tptosocket(tp)->so_snd)) { rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT; to = 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; /* Special case for KEEPINIT */ if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) && (TP_KEEPINIT(tp) != 0) && rsm) { /* * We have to put a ceiling on the rxt timer * of the keep-init timeout. */ uint32_t max_time, red; max_time = TICKS_2_USEC(TP_KEEPINIT(tp)); if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) { red = (cts - (uint32_t)rsm->r_tim_lastsent[0]); if (red < max_time) max_time -= red; else max_time = 1; } /* Reduce timeout to the keep value if needed */ if (max_time < to) to = max_time; } 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_flags & RACK_RWND_COLLAPSED) || (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_FASTRECOVERY(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 = ((uint32_t)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(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time)) tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx]; else tstmp_touse = (uint32_t)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) { if ((rack->rc_srtt_measure_made == 0) && (tp->t_srtt == 1)) { /* * If another stack as run and set srtt to 1, * then the srtt was 0, so lets use the initial. */ srtt = RACK_INITIAL_RTO; } else { srtt_cur = tp->t_srtt; srtt = 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 && tp->t_srtt && (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 */ (uint32_t)rsm->r_tim_lastsent[idx], srtt, idx, 99); } if (to < rack_tlp_min) { to = rack_tlp_min; } if (to > TICKS_2_USEC(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__, RACK_QUALITY_PERSIST); } #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__); rack->r_ctl.persist_lost_ends = 0; rack->probe_not_answered = 0; rack->forced_ack = 0; tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); rack->rc_in_persist = 1; } } static void rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { if (tcp_in_hpts(rack->rc_inp)) { tcp_hpts_remove(rack->rc_inp); 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_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); 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_runningslot; log.u_bbr.bw_inuse = diag->wheel_slot; log.u_bbr.rttProp = diag->wheel_cts; log.u_bbr.timeStamp = cts; log.u_bbr.delRate = diag->maxslots; 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_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type) { if (rack_verbose_logging && 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 = sb->sb_flags; log.u_bbr.flex2 = len; log.u_bbr.flex3 = sb->sb_state; log.u_bbr.flex8 = type; log.u_bbr.timeStamp = tcp_get_usecs(&tv); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, TCP_LOG_SB_WAKE, 0, len, &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 = tptoinpcb(tp); struct timeval tv; uint32_t delayed_ack = 0; uint32_t hpts_timeout; uint32_t entry_slot = slot; uint8_t stopped; uint32_t left = 0; uint32_t us_cts; if ((tp->t_state == TCPS_CLOSED) || (tp->t_state == TCPS_LISTEN)) { return; } if (tcp_in_hpts(inp)) { /* 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 */ rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0); if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) { /* * We have a early carry over set, * we can always add more time so we * can always make this compensation. * * Note if ack's are allowed to wake us do not * penalize the next timer for being awoke * by an ack aka the rc_agg_early (non-paced mode). */ 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_SLOT) { /* We gain delay */ rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot); slot = HPTS_TICKS_PER_SLOT; } else { /* We take off some */ rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT); slot = HPTS_TICKS_PER_SLOT; } } 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_SLOT) { rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot; slot = HPTS_TICKS_PER_SLOT; } 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_USEC(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 = TICKS_2_USEC(TP_KEEPIDLE(tp)); } else { /* * Get the initial setup keep-alive time, * note that this is probably not going to * happen, since rack will be running a rxt timer * if a SYN of some sort is outstanding. It is * actually handled in rack_timeout_rxt(). */ hpts_timeout = TICKS_2_USEC(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; } } } 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; } rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0); if ((rack->gp_ready == 0) && (rack->use_fixed_rate == 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; } /** * Turn off all the flags for queuing by default. The * flags have important meanings to what happens when * LRO interacts with the transport. Most likely (by default now) * mbuf_queueing and ack compression are on. So the transport * has a couple of flags that control what happens (if those * are not on then these flags won't have any effect since it * won't go through the queuing LRO path). * * INP_MBUF_QUEUE_READY - This flags says that I am busy * pacing output, so don't disturb. But * it also means LRO can wake me if there * is a SACK arrival. * * INP_DONT_SACK_QUEUE - This flag is used in conjunction * with the above flag (QUEUE_READY) and * when present it says don't even wake me * if a SACK arrives. * * The idea behind these flags is that if we are pacing we * set the MBUF_QUEUE_READY and only get woken up if * a SACK arrives (which could change things) or if * our pacing timer expires. If, however, we have a rack * timer running, then we don't even want a sack to wake * us since the rack timer has to expire before we can send. * * Other cases should usually have none of the flags set * so LRO can call into us. */ inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY); if (slot) { rack->r_ctl.rc_last_output_to = us_cts + slot; /* * A pacing timer (slot) is being set, in * such a case we cannot send (we are blocked by * the timer). So lets tell LRO that it should not * wake us unless there is a SACK. Note this only * will be effective if mbuf queueing is on or * compressed acks are being processed. */ inp->inp_flags2 |= INP_MBUF_QUEUE_READY; /* * But wait if we have a Rack timer running * even a SACK should not disturb us (with * the exception of r_rr_config 3). */ if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && (rack->r_rr_config != 3)) inp->inp_flags2 |= INP_DONT_SACK_QUEUE; if (rack->rc_ack_can_sendout_data) { /* * Ahh but wait, this is that special case * where the pacing timer can be disturbed * backout the changes (used for non-paced * burst limiting). */ inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY); } if ((rack->use_rack_rr) && (rack->r_rr_config < 2) && ((hpts_timeout) && (hpts_timeout < 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(inp, HPTS_USEC_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(inp, 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) { /* * With respect to inp_flags2 here, lets let any new acks wake * us up here. Since we are not pacing (no pacing timer), output * can happen so we should let it. If its a Rack timer, then any inbound * packet probably won't change the sending (we will be blocked) * but it may change the prr stats so letting it in (the set defaults * at the start of this block) are good enough. */ (void)tcp_hpts_insert_diag(inp, HPTS_USEC_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; - if (tp->t_timers->tt_flags & TT_STOPPED) { + if (tp->tt_flags & TT_STOPPED) { return (1); } 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) { rack->r_ctl.rc_resend = rsm; rack->r_timer_override = 1; 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_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT; } } 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 void rack_adjust_orig_mlen(struct rack_sendmap *rsm) { if (rsm->m->m_len > rsm->orig_m_len) { /* * Mbuf grew, caused by sbcompress, our offset does * not change. */ rsm->orig_m_len = rsm->m->m_len; } else if (rsm->m->m_len < rsm->orig_m_len) { /* * Mbuf shrank, trimmed off the top by an ack, our * offset changes. */ rsm->soff -= (rsm->orig_m_len - rsm->m->m_len); rsm->orig_m_len = rsm->m->m_len; } } static void rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm) { struct mbuf *m; uint32_t soff; if (src_rsm->m && (src_rsm->orig_m_len != src_rsm->m->m_len)) { /* Fix up the orig_m_len and possibly the mbuf offset */ rack_adjust_orig_mlen(src_rsm); } m = src_rsm->m; soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start); while (soff >= m->m_len) { /* Move out past this mbuf */ soff -= m->m_len; m = m->m_next; KASSERT((m != NULL), ("rsm:%p nrsm:%p hit at soff:%u null m", src_rsm, rsm, soff)); } rsm->m = m; rsm->soff = soff; rsm->orig_m_len = m->m_len; } 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->r_no_rtt_allowed = rsm->r_no_rtt_allowed; nrsm->r_rtr_bytes = 0; nrsm->r_fas = rsm->r_fas; 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]; } /* Now if we have SYN flag we keep it on the left edge */ if (nrsm->r_flags & RACK_HAS_SYN) nrsm->r_flags &= ~RACK_HAS_SYN; /* Now if we have a FIN flag we keep it on the right edge */ if (rsm->r_flags & RACK_HAS_FIN) rsm->r_flags &= ~RACK_HAS_FIN; /* Push bit must go to the right edge as well */ if (rsm->r_flags & RACK_HAD_PUSH) rsm->r_flags &= ~RACK_HAD_PUSH; /* Clone over the state of the hw_tls flag */ nrsm->r_hw_tls = rsm->r_hw_tls; /* * Now we need to find nrsm's new location in the mbuf chain * we basically calculate a new offset, which is soff + * how much is left in original rsm. Then we walk out the mbuf * chain to find the righ position, it may be the same mbuf * or maybe not. */ KASSERT(((rsm->m != NULL) || (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))), ("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack)); if (rsm->m) rack_setup_offset_for_rsm(rsm, nrsm); } 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). */ #ifdef INVARIANTS struct rack_sendmap *rm; #endif rack_log_map_chg(rack->rc_tp, rack, NULL, l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__); 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; } #ifndef INVARIANTS (void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm); #else rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm); 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, uint8_t *doing_tlp) { /* * Tail Loss Probe. */ struct rack_sendmap *rsm = NULL; #ifdef INVARIANTS struct rack_sendmap *insret; #endif struct socket *so = tptosocket(tp); uint32_t amm; uint32_t out, avail; int collapsed_win = 0; - if (tp->t_timers->tt_flags & TT_STOPPED) { + if (tp->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__); return (-ETIMEDOUT); /* tcp_drop() */ } /* * 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); rack->r_ctl.retran_during_recovery = 0; rack->r_ctl.dsack_byte_cnt = 0; counter_u64_add(rack_tlp_tot, 1); if (rack->r_state && (rack->r_state != tp->t_state)) rack_set_state(tp, rack); avail = sbavail(&so->so_snd); out = tp->snd_max - tp->snd_una; if ((out > tp->snd_wnd) || rack->rc_has_collapsed) { /* special case, we need a retransmission */ collapsed_win = 1; goto need_retran; } if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) { rack->r_ctl.dsack_persist--; if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) { rack->r_ctl.num_dsack = 0; } rack_log_dsack_event(rack, 1, __LINE__, 0, 0); } if ((tp->t_flags & TF_GPUTINPROG) && (rack->r_ctl.rc_tlp_cnt_out == 1)) { /* * If this is the second in a row * TLP and we are doing a measurement * its time to abandon the measurement. * Something is likely broken on * the clients network and measuring a * broken network does us no good. */ tp->t_flags &= ~TF_GPUTINPROG; rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/, rack->r_ctl.rc_gp_srtt /*flex1*/, tp->gput_seq, 0, 0, 18, __LINE__, NULL, 0); } /* * 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_FASTRECOVERY(tp->t_flags)) { /* Unlikely */ if (rack->rack_no_prr == 0) { if (out + amm <= tp->snd_wnd) { rack->r_ctl.rc_prr_sndcnt = amm; rack->r_ctl.rc_tlp_new_data = amm; rack_log_to_prr(rack, 4, 0, __LINE__); } } 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) { #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) { #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. */ goto out; } rack_clone_rsm(rack, nrsm, rsm, (rsm->r_end - ctf_fixed_maxseg(tp))); rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__); #ifndef INVARIANTS (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #else insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); 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 = nrsm; } rack->r_ctl.rc_tlpsend = rsm; send: /* Make sure output path knows we are doing a TLP */ *doing_tlp = 1; 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) { + if (tp->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; int32_t retval = 1; - if (tp->t_timers->tt_flags & TT_STOPPED) { + if (tp->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__); counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends); return (-ETIMEDOUT); /* tcp_drop() */ } /* * 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_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) { KMOD_TCPSTAT_INC(tcps_persistdrop); tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends); retval = -ETIMEDOUT; /* tcp_drop() */ 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) { KMOD_TCPSTAT_INC(tcps_persistdrop); tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends); retval = -ETIMEDOUT; /* tcp_drop() */ 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); } else { rack->probe_not_answered = 1; counter_u64_add(rack_persists_loss, 1); rack->r_ctl.persist_lost_ends++; } counter_u64_add(rack_persists_sends, 1); 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 = tptoinpcb(tp); - if (tp->t_timers->tt_flags & TT_STOPPED) { + if (tp->tt_flags & TT_STOPPED) { return (1); } rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP; 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); } else { rack->probe_not_answered = 1; } 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); return (-ETIMEDOUT); /* tcp_drop() */ } /* * 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; rack = (struct tcp_rack *)tp->t_fb_ptr; rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __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. * * Also we really need to stick them back in sequence * order. This way we send in the proper order and any * sacks that come floating in will "re-ack" the data. * To do this we zap the tmap with an INIT and then * walk through and place every rsm in the RB tree * back in its seq ordered place. */ TAILQ_INIT(&rack->r_ctl.rc_tmap); RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) { rsm->r_dupack = 0; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); /* 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 | RACK_RWND_COLLAPSED); rsm->r_flags |= RACK_MUST_RXT; } /* Clear the count (we just un-acked them) */ rack->r_ctl.rc_last_timeout_snduna = tp->snd_una; rack->r_ctl.rc_sacked = 0; rack->r_ctl.rc_sacklast = NULL; 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); if (rack->r_ctl.rc_resend != NULL) rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT; rack->r_ctl.rc_prr_sndcnt = 0; rack_log_to_prr(rack, 6, 0, __LINE__); rack->r_timer_override = 1; if ((((tp->t_flags & TF_SACK_PERMIT) == 0) #ifdef NETFLIX_EXP_DETECTION || (rack->sack_attack_disable != 0) #endif ) && ((tp->t_flags & TF_SENTFIN) == 0)) { /* * For non-sack customers new data * needs to go out as retransmits until * we retransmit up to snd_max. */ rack->r_must_retran = 1; rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); } rack->r_ctl.rc_snd_max_at_rto = tp->snd_max; } static void rack_convert_rtts(struct tcpcb *tp) { if (tp->t_srtt > 1) { uint32_t val, frac; val = tp->t_srtt >> TCP_RTT_SHIFT; frac = tp->t_srtt & 0x1f; tp->t_srtt = TICKS_2_USEC(val); /* * frac is the fractional part of the srtt (if any) * but its in ticks and every bit represents * 1/32nd of a hz. */ if (frac) { if (hz == 1000) { frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE); } else { frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE)); } tp->t_srtt += frac; } } if (tp->t_rttvar) { uint32_t val, frac; val = tp->t_rttvar >> TCP_RTTVAR_SHIFT; frac = tp->t_rttvar & 0x1f; tp->t_rttvar = TICKS_2_USEC(val); /* * frac is the fractional part of the srtt (if any) * but its in ticks and every bit represents * 1/32nd of a hz. */ if (frac) { if (hz == 1000) { frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE); } else { frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE)); } tp->t_rttvar += frac; } } tp->t_rxtcur = RACK_REXMTVAL(tp); if (TCPS_HAVEESTABLISHED(tp->t_state)) { tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop); } if (tp->t_rxtcur > rack_rto_max) { tp->t_rxtcur = rack_rto_max; } } static void rack_cc_conn_init(struct tcpcb *tp) { struct tcp_rack *rack; uint32_t srtt; rack = (struct tcp_rack *)tp->t_fb_ptr; srtt = tp->t_srtt; cc_conn_init(tp); /* * Now convert to rack's internal format, * if required. */ if ((srtt == 0) && (tp->t_srtt != 0)) rack_convert_rtts(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) { struct inpcb *inp = tptoinpcb(tp); int32_t rexmt; int32_t retval = 0; bool isipv6; - if (tp->t_timers->tt_flags & TT_STOPPED) { + if (tp->tt_flags & TT_STOPPED) { return (1); } if ((tp->t_flags & TF_GPUTINPROG) && (tp->t_rxtshift)) { /* * We have had a second timeout * measurements on successive rxt's are not profitable. * It is unlikely to be of any use (the network is * broken or the client went away). */ tp->t_flags &= ~TF_GPUTINPROG; rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/, rack->r_ctl.rc_gp_srtt /*flex1*/, tp->gput_seq, 0, 0, 18, __LINE__, NULL, 0); } 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__); return (-ETIMEDOUT); /* tcp_drop() */ } rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; rack->r_ctl.retran_during_recovery = 0; rack->rc_ack_required = 1; rack->r_ctl.dsack_byte_cnt = 0; 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; if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_una == tp->snd_max)) { /* Nothing outstanding .. nothing to do */ return (0); } if (rack->r_ctl.dsack_persist) { rack->r_ctl.dsack_persist--; if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) { rack->r_ctl.num_dsack = 0; } rack_log_dsack_event(rack, 1, __LINE__, 0, 0); } /* * Rack can only run one timer at a time, so we cannot * run a KEEPINIT (gating SYN sending) and a retransmit * timer for the SYN. So if we are in a front state and * have a KEEPINIT timer we need to check the first transmit * against now to see if we have exceeded the KEEPINIT time * (if one is set). */ if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) && (TP_KEEPINIT(tp) != 0)) { struct rack_sendmap *rsm; rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm) { /* Ok we have something outstanding to test keepinit with */ if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) && ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) { /* We have exceeded the KEEPINIT time */ tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX); goto drop_it; } } } /* * 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) { tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN); drop_it: tp->t_rxtshift = TCP_MAXRXTSHIFT; KMOD_TCPSTAT_INC(tcps_timeoutdrop); /* XXXGL: previously t_softerror was casted to uint16_t */ MPASS(tp->t_softerror >= 0); retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT; goto out; /* tcp_drop() */ } 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; tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2); tp->t_flags |= TF_PREVVALID; } else if ((tp->t_flags & TF_RCVD_TSTMP) == 0) tp->t_flags &= ~TF_PREVVALID; KMOD_TCPSTAT_INC(tcps_rexmttimeo); if ((tp->t_state == TCPS_SYN_SENT) || (tp->t_state == TCPS_SYN_RECEIVED)) rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift]; else rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift]; RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt, max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop); /* * 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 = (inp->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); } } } /* * 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 ((inp->inp_vflag & INP_IPV6) != 0) in6_losing(inp); else #endif in_losing(inp); tp->t_rttvar += tp->t_srtt; 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, CC_RTO, tp->snd_una, __LINE__); out: return (retval); } static int rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp) { int32_t ret = 0; int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK); if ((tp->t_state >= TCPS_FIN_WAIT_1) && (tp->t_flags & TF_GPUTINPROG)) { /* * We have a goodput in progress * and we have entered a late state. * Do we have enough data in the sb * to handle the GPUT request? */ uint32_t bytes; bytes = tp->gput_ack - tp->gput_seq; if (SEQ_GT(tp->gput_seq, tp->snd_una)) bytes += tp->gput_seq - tp->snd_una; if (bytes > sbavail(&tptosocket(tp)->so_snd)) { /* * There are not enough bytes in the socket * buffer that have been sent to cover this * measurement. Cancel it. */ rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/, rack->r_ctl.rc_gp_srtt /*flex1*/, tp->gput_seq, 0, 0, 18, __LINE__, NULL, 0); tp->t_flags &= ~TF_GPUTINPROG; } } 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(tptoinpcb(tp), 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; rack->r_fast_output = 0; 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, doing_tlp); } else if (timers & PACE_TMR_RXT) { rack->r_ctl.rc_tlp_rxt_last_time = cts; rack->r_fast_output = 0; 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_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 (tcp_in_hpts(rack->rc_inp) && ((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_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, uint64_t ts, uint16_t add_flag) { 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; /* * Here we don't add in the len of send, since its already * in snduna <->snd_max. */ rsm->r_fas = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); 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; /* Take off the must retransmit flag, if its on */ if (rsm->r_flags & RACK_MUST_RXT) { if (rack->r_must_retran) rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start); if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) { /* * We have retransmitted all we need. Clear * any must retransmit flags. */ rack->r_must_retran = 0; rack->r_ctl.rc_out_at_rto = 0; } rsm->r_flags &= ~RACK_MUST_RXT; } 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, uint64_t ts, int32_t *lenp, uint16_t add_flag) { /* * We (re-)transmitted starting at rsm->r_start for some length * (possibly less than r_end. */ struct rack_sendmap *nrsm; #ifdef INVARIANTS struct rack_sendmap *insret; #endif 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, add_flag); 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); #ifndef INVARIANTS (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #else insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); 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, add_flag); /* Log a split of rsm into rsm and nrsm */ rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__); *lenp = 0; return (0); } static void rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts, struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff, int hw_tls) { struct tcp_rack *rack; struct rack_sendmap *rsm, *nrsm, fe; #ifdef INVARIANTS struct rack_sendmap *insret; #endif 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(tptoinpcb(tp)); 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; 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 (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; } if (len == 0) { /* We don't log zero window probes */ return; } if (IN_FASTRECOVERY(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|add_flag; } else { rsm->r_flags = add_flag; } if (hw_tls) rsm->r_hw_tls = 1; rsm->r_tim_lastsent[0] = cts; rsm->r_rtr_cnt = 1; rsm->r_rtr_bytes = 0; if (th_flags & TH_SYN) { /* The data space is one beyond snd_una */ rsm->r_flags |= RACK_HAS_SYN; } rsm->r_start = seq_out; rsm->r_end = rsm->r_start + len; rsm->r_dupack = 0; /* * save off the mbuf location that * sndmbuf_noadv returned (which is * where we started copying from).. */ rsm->m = s_mb; rsm->soff = s_moff; /* * Here we do add in the len of send, since its not yet * reflected in in snduna <->snd_max */ rsm->r_fas = (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) + (rsm->r_end - rsm->r_start)); /* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */ if (rsm->m) { if (rsm->m->m_len <= rsm->soff) { /* * XXXrrs Question, will this happen? * * If sbsndptr is set at the correct place * then s_moff should always be somewhere * within rsm->m. But if the sbsndptr was * off then that won't be true. If it occurs * we need to walkout to the correct location. */ struct mbuf *lm; lm = rsm->m; while (lm->m_len <= rsm->soff) { rsm->soff -= lm->m_len; lm = lm->m_next; KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u", __func__, rack, s_moff, s_mb, rsm->soff)); } rsm->m = lm; } rsm->orig_m_len = rsm->m->m_len; } else rsm->orig_m_len = 0; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); /* Log a new rsm */ rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__); #ifndef INVARIANTS (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #else insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); 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_FASTRECOVERY(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, cts, &len, add_flag); 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, cts, &len, add_flag); 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, cts, add_flag); 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); rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__); #ifndef INVARIANTS (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #else insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); 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, cts, &len, add_flag); 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; 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; } 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 = (struct tcp_rack *)tp->t_fb_ptr; if (tp->t_srtt != 0) { /* * We keep a simple srtt in microseconds, like our rtt * measurement. We don't need to do any tricks with shifting * etc. Instead we just add in 1/8th of the new measurement * and subtract out 1/8 of the old srtt. We do the same with * the variance after finding the absolute value of the * difference between this sample and the current srtt. */ delta = tp->t_srtt - rtt; /* Take off 1/8th of the current sRTT */ tp->t_srtt -= (tp->t_srtt >> 3); /* Add in 1/8th of the new RTT just measured */ tp->t_srtt += (rtt >> 3); if (tp->t_srtt <= 0) tp->t_srtt = 1; /* Now lets make the absolute value of the variance */ if (delta < 0) delta = -delta; /* Subtract out 1/8th */ tp->t_rttvar -= (tp->t_rttvar >> 3); /* Add in 1/8th of the new variance we just saw */ tp->t_rttvar += (delta >> 3); if (tp->t_rttvar <= 0) tp->t_rttvar = 1; } 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; tp->t_rttvar = rtt >> 1; } rack->rc_srtt_measure_made = 1; KMOD_TCPSTAT_INC(tcps_rttupdated); tp->t_rttupdated++; #ifdef STATS if (rack_stats_gets_ms_rtt == 0) { /* Send in the microsecond rtt used for rxt timeout purposes */ stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt)); } else if (rack_stats_gets_ms_rtt == 1) { /* Send in the millisecond rtt used for rxt timeout purposes */ int32_t ms_rtt; /* Round up */ ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC; stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt)); } else if (rack_stats_gets_ms_rtt == 2) { /* Send in the millisecond rtt has close to the path RTT as we can get */ int32_t ms_rtt; /* Round up */ ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC; stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt)); } else { /* Send in the microsecond rtt has close to the path RTT as we can get */ stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt)); } #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). */ tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop); rack_log_rtt_sample(rack, rtt); tp->t_softerror = 0; } 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 (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) { uint32_t us_rtt; int32_t i, all; uint32_t t, len_acked; if ((rsm->r_flags & RACK_ACKED) || (rsm->r_flags & RACK_WAS_ACKED)) /* Already done */ return (0); if (rsm->r_no_rtt_allowed) { /* Not allowed */ return (0); } if (ack_type == CUM_ACKED) { if (SEQ_GT(th_ack, rsm->r_end)) { len_acked = rsm->r_end - rsm->r_start; all = 1; } else { len_acked = th_ack - rsm->r_start; all = 0; } } else { len_acked = rsm->r_end - rsm->r_start; all = 0; } if (rsm->r_rtr_cnt == 1) { t = cts - (uint32_t)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; } } if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; else us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; if (us_rtt == 0) us_rtt = 1; if (CC_ALGO(tp)->rttsample != NULL) { /* Kick the RTT to the CC */ - CC_ALGO(tp)->rttsample(tp->ccv, us_rtt, 1, rsm->r_fas); + CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas); } rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time)); if (ack_type == SACKED) { rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1); tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt); } else { /* * We need to setup what our confidence * is in this ack. * * If the rsm was app limited and it is * less than a mss in length (the end * of the send) then we have a gap. If we * were app limited but say we were sending * multiple MSS's then we are more confident * int it. * * When we are not app-limited then we see if * the rsm is being included in the current * measurement, we tell this by the app_limited_needs_set * flag. * * Note that being cwnd blocked is not applimited * as well as the pacing delay between packets which * are sending only 1 or 2 MSS's also will show up * in the RTT. We probably need to examine this algorithm * a bit more and enhance it to account for the delay * between rsm's. We could do that by saving off the * pacing delay of each rsm (in an rsm) and then * factoring that in somehow though for now I am * not sure how :) */ int calc_conf = 0; if (rsm->r_flags & RACK_APP_LIMITED) { if (all && (len_acked <= ctf_fixed_maxseg(tp))) calc_conf = 0; else calc_conf = 1; } else if (rack->app_limited_needs_set == 0) { calc_conf = 1; } else { calc_conf = 0; } rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2); tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, calc_conf, rsm, rsm->r_rtr_cnt); } if ((rsm->r_flags & RACK_TLP) && (!IN_FASTRECOVERY(tp->t_flags))) { /* Segment was a TLP and our retrans matched */ if (rack->r_ctl.rc_tlp_cwnd_reduce) { rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__); } } if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { /* New more recent rack_tmit_time */ rack->r_ctl.rc_rack_tmit_time = (uint32_t)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; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); tp->t_softerror = 0; if (to && (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 (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) { t = cts - (uint32_t)rsm->r_tim_lastsent[i]; if ((int)t <= 0) t = 1; if (CC_ALGO(tp)->rttsample != NULL) { /* * Kick the RTT to the CC, here * we lie a bit in that we know the * retransmission is correct even though * we retransmitted. This is because * we match the timestamps. */ if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i])) us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i]; else us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i]; - CC_ALGO(tp)->rttsample(tp->ccv, us_rtt, 1, rsm->r_fas); + CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas); } if ((i + 1) < rsm->r_rtr_cnt) { /* * The peer ack'd from our previous * transmission. We have a spurious * retransmission and thus we dont * want to update our rack_rtt. * * Hmm should there be a CC revert here? * */ return (0); } 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, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { /* New more recent rack_tmit_time */ rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; rack->rc_rack_rtt = t; } rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3); tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 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 - (uint32_t)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 * likely did an improper retransmit as outlined in * 6.2 Step 2 point 2 in the rack-draft so we * don't want to update our rack_rtt. We in * theory (in future) might want to think about reverting our * cwnd state but we won't for now. */ return (0); } 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, (uint32_t)rsm->r_tim_lastsent[i])) { /* New more recent rack_tmit_time */ rack->r_ctl.rc_rack_tmit_time = (uint32_t)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_RWND_COLLAPSED) { /* * If the peer dropped the rwnd on * these then we don't worry about them. */ 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->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; } 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->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; } 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->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; } 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, 0); if (rack->rc_gp_filled && ((tp->gput_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp))))) { uint32_t ideal_amount; ideal_amount = rack_get_measure_window(tp, rack); if (ideal_amount > sbavail(&tptosocket(tp)->so_snd)) { /* * 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. */ tp->t_flags &= ~TF_GPUTINPROG; rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq, 0, 0, 0, 6, __LINE__, NULL, 0); } else { /* * Reset the window further out. */ tp->gput_ack = tp->gput_seq + ideal_amount; } } } } static inline int is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm) { if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) { /* Behind our TLP definition or right at */ return (0); } if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) { /* The start is beyond or right at our end of TLP definition */ return (0); } /* It has to be a sub-part of the original TLP recorded */ return (1); } 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, *prev, *next; #ifdef INVARIANTS struct rack_sendmap *insret; #endif 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) { /* * Before any splitting or hookery is * done is it a TLP of interest i.e. rxt? */ if ((rsm->r_flags & RACK_TLP) && (rsm->r_rtr_cnt > 1)) { /* * We are splitting a rxt TLP, check * if we need to save off the start/end */ if (rack->rc_last_tlp_acked_set && (is_rsm_inside_declared_tlp_block(rack, rsm))) { /* * We already turned this on since we are inside * the previous one was a partially sack now we * are getting another one (maybe all of it). * */ rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end); /* * Lets make sure we have all of it though. */ if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) { rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) { rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } } else { rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack->rc_last_tlp_past_cumack = 0; rack->rc_last_tlp_acked_set = 1; rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end); } } /** * 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; /* Now we must adjust back where next->m is */ rack_setup_offset_for_rsm(rsm, next); /* We don't need to adjust rsm, it did not change */ /* 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) { 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; } rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__); 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; #ifndef INVARIANTS (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #else insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); 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; } rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__); 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 * reposition 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) { /* * Is it a TLP of interest? */ if ((rsm->r_flags & RACK_TLP) && (rsm->r_rtr_cnt > 1)) { /* * We are splitting a rxt TLP, check * if we need to save off the start/end */ if (rack->rc_last_tlp_acked_set && (is_rsm_inside_declared_tlp_block(rack, rsm))) { /* * We already turned this on since we are inside * the previous one was a partially sack now we * are getting another one (maybe all of it). */ rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end); /* * Lets make sure we have all of it though. */ if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) { rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) { rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } } else { rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack->rc_last_tlp_past_cumack = 0; rack->rc_last_tlp_acked_set = 1; rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end); } } 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; 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 = rack_to_usec_ts(&rack->r_ctl.act_rcv_time); rsm->r_flags |= RACK_ACKED; if (rsm->r_in_tmap) { TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__); } 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) { /* * Is it a TLP of interest? */ if ((rsm->r_flags & RACK_TLP) && (rsm->r_rtr_cnt > 1)) { /* * We are splitting a rxt TLP, check * if we need to save off the start/end */ if (rack->rc_last_tlp_acked_set && (is_rsm_inside_declared_tlp_block(rack, rsm))) { /* * We already turned this on since we are inside * the previous one was a partially sack now we * are getting another one (maybe all of it). */ rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end); /* * Lets make sure we have all of it though. */ if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) { rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) { rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } } else { rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack->rc_last_tlp_past_cumack = 0; rack->rc_last_tlp_acked_set = 1; rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end); } } 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) * * Note if either prev/rsm is a TLP we don't * do this. */ 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 that the rsm has had its start moved forward * lets go ahead and get its new place in the world. */ rack_setup_offset_for_rsm(prev, rsm); /* * 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) { rack->r_ctl.rc_reorder_ts = cts; } rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__); 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; } if ((rsm->r_flags & RACK_TLP) && (rsm->r_rtr_cnt > 1)) { /* * We are splitting a rxt TLP, check * if we need to save off the start/end */ if (rack->rc_last_tlp_acked_set && (is_rsm_inside_declared_tlp_block(rack, rsm))) { /* * We already turned this on since this block is inside * the previous one was a partially sack now we * are getting another one (maybe all of it). */ rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end); /* * Lets make sure we have all of it though. */ if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) { rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) { rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } } else { rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack->rc_last_tlp_acked_set = 1; rack->rc_last_tlp_past_cumack = 0; rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end); } } /** * 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; #ifndef INVARIANTS (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #else insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); 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; 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 = rack_to_usec_ts(&rack->r_ctl.act_rcv_time); rsm->r_flags |= RACK_ACKED; rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__); 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_TLP) == 0) && (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_TLP) break; 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_TLP) break; 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. */ #ifdef NETFLIX_EXP_DETECTION uint32_t pkt_delta; pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt; #endif /* 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_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to) { struct rack_sendmap *rsm; #ifdef INVARIANTS struct rack_sendmap *rm; #endif /* * 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; /* Tend any TLP that has been marked for 1/2 the seq space (its old) */ if ((rack->rc_last_tlp_acked_set == 1)&& (rack->rc_last_tlp_past_cumack == 1) && (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) { /* * We have reached the point where our last rack * tlp retransmit sequence is ahead of the cum-ack. * This can only happen when the cum-ack moves all * the way around (its been a full 2^^31+1 bytes * or more since we sent a retransmitted TLP). Lets * turn off the valid flag since its not really valid. * * Note since sack's also turn on this event we have * a complication, we have to wait to age it out until * the cum-ack is by the TLP before checking which is * what the next else clause does. */ rack_log_dsack_event(rack, 9, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); rack->rc_last_tlp_acked_set = 0; rack->rc_last_tlp_past_cumack = 0; } else if ((rack->rc_last_tlp_acked_set == 1) && (rack->rc_last_tlp_past_cumack == 0) && (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) { /* * It is safe to start aging TLP's out. */ rack->rc_last_tlp_past_cumack = 1; } /* We do the same for the tlp send seq as well */ if ((rack->rc_last_sent_tlp_seq_valid == 1) && (rack->rc_last_sent_tlp_past_cumack == 1) && (SEQ_GT(rack->r_ctl.last_sent_tlp_seq, th_ack))) { rack_log_dsack_event(rack, 9, __LINE__, rack->r_ctl.last_sent_tlp_seq, (rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len)); rack->rc_last_sent_tlp_seq_valid = 0; rack->rc_last_sent_tlp_past_cumack = 0; } else if ((rack->rc_last_sent_tlp_seq_valid == 1) && (rack->rc_last_sent_tlp_past_cumack == 0) && (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) { /* * It is safe to start aging TLP's send. */ rack->rc_last_sent_tlp_past_cumack = 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. */ return; } if (tp->t_flags & TF_SENTFIN) { /* if we sent a FIN we often will not have map */ return; } #ifdef INVARIANTS panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n", tp, tp->t_state, th_ack, rack, tp->snd_una, tp->snd_max, tp->snd_nxt); #endif return; } 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 return; } rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack); /* Now was it a retransmitted TLP? */ if ((rsm->r_flags & RACK_TLP) && (rsm->r_rtr_cnt > 1)) { /* * Yes, this rsm was a TLP and retransmitted, remember that * since if a DSACK comes back on this we don't want * to think of it as a reordered segment. This may * get updated again with possibly even other TLPs * in flight, but thats ok. Only when we don't send * a retransmitted TLP for 1/2 the sequences space * will it get turned off (above). */ if (rack->rc_last_tlp_acked_set && (is_rsm_inside_declared_tlp_block(rack, rsm))) { /* * We already turned this on since the end matches, * the previous one was a partially ack now we * are getting another one (maybe all of it). */ rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end); /* * Lets make sure we have all of it though. */ if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) { rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) { rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } } else { rack->rc_last_tlp_past_cumack = 1; rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack->rc_last_tlp_acked_set = 1; rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end); } } /* 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_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__); 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->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; #ifndef INVARIANTS (void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #else rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); 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; rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time); rsm->r_flags |= RACK_ACKED; rack->r_ctl.rc_reorder_ts = cts; if (rack->r_ent_rec_ns) { /* * We have sent no more, and we saw an sack * then ack arrive. */ rack->r_might_revert = 1; } } if ((rsm->r_flags & RACK_TO_REXT) && (tp->t_flags & TF_RCVD_TSTMP) && (to->to_flags & TOF_TS) && (to->to_tsecr != 0) && (tp->t_flags & TF_PREVVALID)) { /* * We can use the timestamp to see * if this retransmission was from the * first transmit. If so we made a mistake. */ tp->t_flags &= ~TF_PREVVALID; if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) { /* The first transmit is what this ack is for */ rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__); } } 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; } /* 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_ack); } return; } 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->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__); /* Now we need to move our offset forward too */ if (rsm->m && (rsm->orig_m_len != rsm->m->m_len)) { /* Fix up the orig_m_len and possibly the mbuf offset */ rack_adjust_orig_mlen(rsm); } rsm->soff += (th_ack - rsm->r_start); rsm->r_start = th_ack; /* Now do we need to move the mbuf fwd too? */ if (rsm->m) { while (rsm->soff >= rsm->m->m_len) { rsm->soff -= rsm->m->m_len; rsm->m = rsm->m->m_next; KASSERT((rsm->m != NULL), (" nrsm:%p hit at soff:%u null m", rsm, rsm->soff)); } rsm->orig_m_len = rsm->m->m_len; } if (rack->app_limited_needs_set) rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG); } static void rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack) { struct rack_sendmap *rsm; int sack_pass_fnd = 0; if (rack->r_might_revert) { /* * Ok we have reordering, have not sent anything, we * might want to revert the congestion state if nothing * further has SACK_PASSED on it. Lets check. * * We also get here when we have DSACKs come in for * all the data that we FR'd. Note that a rxt or tlp * timer clears this from happening. */ TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) { if (rsm->r_flags & RACK_SACK_PASSED) { sack_pass_fnd = 1; break; } } if (sack_pass_fnd == 0) { /* * We went into recovery * incorrectly due to reordering! */ int orig_cwnd; rack->r_ent_rec_ns = 0; orig_cwnd = tp->snd_cwnd; tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec; tp->snd_recover = tp->snd_una; rack_log_to_prr(rack, 14, orig_cwnd, __LINE__); EXIT_RECOVERY(tp->t_flags); } rack->r_might_revert = 0; } } #ifdef NETFLIX_EXP_DETECTION static void rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack, uint32_t bytes_this_ack, uint32_t segsiz) { 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 / segsiz)); 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 static int rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end) { uint32_t am, l_end; int was_tlp = 0; if (SEQ_GT(end, start)) am = end - start; else am = 0; if ((rack->rc_last_tlp_acked_set ) && (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) && (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) { /* * The DSACK is because of a TLP which we don't * do anything with the reordering window over since * it was not reordering that caused the DSACK but * our previous retransmit TLP. */ rack_log_dsack_event(rack, 7, __LINE__, start, end); was_tlp = 1; goto skip_dsack_round; } if (rack->rc_last_sent_tlp_seq_valid) { l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len; if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) && (SEQ_LEQ(end, l_end))) { /* * This dsack is from the last sent TLP, ignore it * for reordering purposes. */ rack_log_dsack_event(rack, 7, __LINE__, start, end); was_tlp = 1; goto skip_dsack_round; } } if (rack->rc_dsack_round_seen == 0) { rack->rc_dsack_round_seen = 1; rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max; rack->r_ctl.num_dsack++; rack->r_ctl.dsack_persist = 16; /* 16 is from the standard */ rack_log_dsack_event(rack, 2, __LINE__, 0, 0); } skip_dsack_round: /* * We keep track of how many DSACK blocks we get * after a recovery incident. */ rack->r_ctl.dsack_byte_cnt += am; if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) && rack->r_ctl.retran_during_recovery && (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) { /* * False recovery most likely culprit is reordering. If * nothing else is missing we need to revert. */ rack->r_might_revert = 1; rack_handle_might_revert(rack->rc_tp, rack); rack->r_might_revert = 0; rack->r_ctl.retran_during_recovery = 0; rack->r_ctl.dsack_byte_cnt = 0; } return (was_tlp); } static uint32_t do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una) { return (((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt); } static int32_t rack_compute_pipe(struct tcpcb *tp) { return ((int32_t)do_rack_compute_pipe(tp, (struct tcp_rack *)tp->t_fb_ptr, tp->snd_una)); } static void rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack) { /* Deal with changed and PRR here (in recovery only) */ uint32_t pipe, snd_una; rack->r_ctl.rc_prr_delivered += changed; if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) { /* * It is all outstanding, we are application limited * and thus we don't need more room to send anything. * Note we use tp->snd_una here and not th_ack because * the data as yet not been cut from the sb. */ rack->r_ctl.rc_prr_sndcnt = 0; return; } /* Compute prr_sndcnt */ if (SEQ_GT(tp->snd_una, th_ack)) { snd_una = tp->snd_una; } else { snd_una = th_ack; } pipe = do_rack_compute_pipe(tp, rack, snd_una); 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, __LINE__); 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, __LINE__); } 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, __LINE__); } else { rack->r_ctl.rc_prr_sndcnt = min(0, limit); rack_log_to_prr(rack, 12, 0, __LINE__); } } } static void rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck) { uint32_t changed; struct tcp_rack *rack; struct rack_sendmap *rsm; 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; int loop_start = 0, moved_two = 0; uint32_t tsused; INP_WLOCK_ASSERT(tptoinpcb(tp)); if (tcp_get_flags(th) & TH_RST) { /* We don't log resets */ return; } rack = (struct tcp_rack *)tp->t_fb_ptr; cts = tcp_get_usecs(NULL); 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) { rack_process_to_cumack(tp, rack, th_ack, cts, to); } if ((to->to_flags & TOF_SACK) == 0) { /* We are done nothing left and no sack. */ rack_handle_might_revert(tp, rack); /* * For cases where we struck a dup-ack * with no SACK, add to the changes so * PRR will work right. */ if (dup_ack_struck && (changed == 0)) { changed += ctf_fixed_maxseg(rack->rc_tp); } 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++; } else if (SEQ_LEQ(sack.start, th_ack) && SEQ_LEQ(sack.end, th_ack)) { int was_tlp; was_tlp = rack_note_dsack(rack, sack.start, sack.end); /* * Its a D-SACK block. */ tcp_record_dsack(tp, sack.start, sack.end, was_tlp); } } if (rack->rc_dsack_round_seen) { /* Is the dsack roound over? */ if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) { /* Yes it is */ rack->rc_dsack_round_seen = 0; rack_log_dsack_event(rack, 3, __LINE__, 0, 0); } } /* * 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; } 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; } 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 rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp)); #endif if (changed) { /* Something changed cancel the rack timer */ rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); } tsused = tcp_get_usecs(NULL); rsm = tcp_rack_output(tp, rack, tsused); if ((!IN_FASTRECOVERY(tp->t_flags)) && rsm && ((rsm->r_flags & RACK_MUST_RXT) == 0)) { /* Enter recovery */ entered_recovery = 1; rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__); /* * 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, __LINE__); } rack->r_timer_override = 1; rack->r_early = 0; rack->r_ctl.rc_agg_early = 0; } else if (IN_FASTRECOVERY(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_FASTRECOVERY(tp->t_flags) && (rack->rack_no_prr == 0) && (entered_recovery == 0)) { rack_update_prr(tp, rack, changed, th_ack); if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) && ((tcp_in_hpts(rack->rc_inp) == 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); while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) { rsm = TAILQ_NEXT(rsm, r_tnext); if (rsm->r_flags & RACK_MUST_RXT) { /* Sendmap entries that are marked to * be retransmitted do not need dupack's * struck. We get these marks for a number * of reasons (rxt timeout with no sack, * mtu change, or rwnd collapses). When * these events occur, we know we must retransmit * them and mark the sendmap entries. Dupack counting * is not needed since we are already set to retransmit * it as soon as we can. */ continue; } } if (rsm && (rsm->r_dupack < 0xff)) { rsm->r_dupack++; if (rsm->r_dupack >= DUP_ACK_THRESHOLD) { struct timeval tv; uint32_t cts; /* * Here we see if we need to retransmit. For * a SACK type connection if enough time has passed * we will get a return of the rsm. For a non-sack * connection we will get the rsm returned if the * dupack value is 3 or more. */ cts = tcp_get_usecs(&tv); rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts); if (rack->r_ctl.rc_resend != NULL) { if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) { rack_cong_signal(rack->rc_tp, CC_NDUPACK, rack->rc_tp->snd_una, __LINE__); } 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 multiplier. */ 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; if (rtt == 0) { /* no us sample is there a ms one? */ if (rack->r_ctl.rack_rs.rs_rtt_lowest) { rtt = rack->r_ctl.rack_rs.rs_rtt_lowest; } else { goto no_measurement; } } 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; if (rack->r_ctl.num_measurements < RACK_REQ_AVG) rack->r_ctl.num_measurements = RACK_REQ_AVG; rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL); } else if (calc_bw > rack->r_ctl.gp_bw) { rack->r_ctl.rc_rtt_diff = 0; if (rack->r_ctl.num_measurements < RACK_REQ_AVG) rack->r_ctl.num_measurements = RACK_REQ_AVG; rack->r_ctl.gp_bw = calc_bw; rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL); } else rack_increase_bw_mul(rack, -1, 0, 0, 1); if ((rack->gp_ready == 0) && (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) { /* We have enough measurements now */ rack->gp_ready = 1; rack_set_cc_pacing(rack); if (rack->defer_options) rack_apply_deferred_options(rack); } /* * 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 { /* * zero rtt possibly?, settle for just an old increase. */ no_measurement: rack_increase_bw_mul(rack, -1, 0, 0, 1); } } else if ((IN_FASTRECOVERY(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); } } static void rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount) { /* * The fast output path is enabled and we * have moved the cumack forward. Lets see if * we can expand forward the fast path length by * that amount. What we would ideally like to * do is increase the number of bytes in the * fast path block (left_to_send) by the * acked amount. However we have to gate that * by two factors: * 1) The amount outstanding and the rwnd of the peer * (i.e. we don't want to exceed the rwnd of the peer). * * 2) The amount of data left in the socket buffer (i.e. * we can't send beyond what is in the buffer). * * Note that this does not take into account any increase * in the cwnd. We will only extend the fast path by * what was acked. */ uint32_t new_total, gating_val; new_total = acked_amount + rack->r_ctl.fsb.left_to_send; gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)), (tp->snd_wnd - (tp->snd_max - tp->snd_una))); if (new_total <= gating_val) { /* We can increase left_to_send by the acked amount */ counter_u64_add(rack_extended_rfo, 1); rack->r_ctl.fsb.left_to_send = new_total; KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))), ("rack:%p left_to_send:%u sbavail:%u out:%u", rack, rack->r_ctl.fsb.left_to_send, sbavail(&rack->rc_inp->inp_socket->so_snd), (tp->snd_max - tp->snd_una))); } } static void rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una) { /* * Here any sendmap entry that points to the * beginning mbuf must be adjusted to the correct * offset. This must be called with: * 1) The socket buffer locked * 2) snd_una adjusted to its new postion. * * Note that (2) implies rack_ack_received has also * been called. * * We grab the first mbuf in the socket buffer and * then go through the front of the sendmap, recalculating * the stored offset for any sendmap entry that has * that mbuf. We must use the sb functions to do this * since its possible an add was done has well as * the subtraction we may have just completed. This should * not be a penalty though, since we just referenced the sb * to go in and trim off the mbufs that we freed (of course * there will be a penalty for the sendmap references though). */ struct mbuf *m; struct rack_sendmap *rsm; SOCKBUF_LOCK_ASSERT(sb); m = sb->sb_mb; rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if ((rsm == NULL) || (m == NULL)) { /* Nothing outstanding */ return; } while (rsm->m && (rsm->m == m)) { /* one to adjust */ #ifdef INVARIANTS struct mbuf *tm; uint32_t soff; tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff); if (rsm->orig_m_len != m->m_len) { rack_adjust_orig_mlen(rsm); } if (rsm->soff != soff) { /* * This is not a fatal error, we anticipate it * might happen (the else code), so we count it here * so that under invariant we can see that it really * does happen. */ counter_u64_add(rack_adjust_map_bw, 1); } rsm->m = tm; rsm->soff = soff; if (tm) rsm->orig_m_len = rsm->m->m_len; else rsm->orig_m_len = 0; #else rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff); if (rsm->m) rsm->orig_m_len = rsm->m->m_len; else rsm->orig_m_len = 0; #endif rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); if (rsm == NULL) break; } } /* * 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; INP_WLOCK_ASSERT(tptoinpcb(tp)); 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_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt); rack->r_wanted_output = 1; return (1); } if (rack->gp_ready && (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { under_pacing = 1; } if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { int in_rec, dup_ack_struck = 0; in_rec = IN_FASTRECOVERY(tp->t_flags); if (rack->rc_in_persist) { tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); } if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd) && ((to->to_flags & TOF_SACK) == 0)) { rack_strike_dupack(rack); dup_ack_struck = 1; } rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck); } if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { /* * Old ack, behind (or duplicate to) the last one rcv'd * Note: We mark reordering is occuring if its * less than and we have not closed our window. */ if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) { rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); } 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); acked = BYTES_THIS_ACK(tp, th); if (acked) { /* * Any time we move the cum-ack forward clear * keep-alive tied probe-not-answered. The * persists clears its own on entry. */ rack->probe_not_answered = 0; } 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_RCVD_TSTMP) == 0)) { tp->t_flags &= ~TF_PREVVALID; if (tp->t_rxtshift == 1 && (int)(ticks - tp->t_badrxtwin) < 0) rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__); } if (acked) { /* assure we are not backed off */ tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); 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 (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); } else { rack_post_recovery(tp, th->th_ack); 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->th_ack, 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; } tp->snd_una = th->th_ack; if (acked_amount && sbavail(&so->so_snd)) rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una); rack_log_wakeup(tp,rack, &so->so_snd, acked, 2); /* NB: sowwakeup_locked() does an implicit unlock. */ sowwakeup_locked(so); m_freem(mfree); 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 */ tp->t_flags &= ~TF_PREVVALID; rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL); rack->r_ctl.retran_during_recovery = 0; rack->r_ctl.dsack_byte_cnt = 0; 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(&tptosocket(tp)->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 (now or in the past), 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_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line, int dir, uint32_t flags, 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 = cnt; log.u_bbr.flex2 = split; log.u_bbr.flex3 = out; log.u_bbr.flex4 = line; log.u_bbr.flex5 = rack->r_must_retran; log.u_bbr.flex6 = flags; log.u_bbr.flex7 = rack->rc_has_collapsed; log.u_bbr.flex8 = dir; /* * 1 is collapsed, 0 is uncollapsed, * 2 is log of a rsm being marked, 3 is a split. */ if (rsm == NULL) log.u_bbr.rttProp = 0; else log.u_bbr.rttProp = (uint64_t)rsm; 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_RACK_LOG_COLLAPSE, 0, 0, &log, false, &tv); } } static void rack_collapsed_window(struct tcp_rack *rack, uint32_t out, int line) { /* * Here all we do is mark the collapsed point and set the flag. * This may happen again and again, but there is no * sense splitting our map until we know where the * peer finally lands in the collapse. */ rack_trace_point(rack, RACK_TP_COLLAPSED_WND); if ((rack->rc_has_collapsed == 0) || (rack->r_ctl.last_collapse_point != (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd))) counter_u64_add(rack_collapsed_win_seen, 1); rack->r_ctl.last_collapse_point = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd; rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max; rack->rc_has_collapsed = 1; rack->r_collapse_point_valid = 1; rack_log_collapse(rack, 0, 0, rack->r_ctl.last_collapse_point, line, 1, 0, NULL); } static void rack_un_collapse_window(struct tcp_rack *rack, int line) { struct rack_sendmap *nrsm, *rsm, fe; int cnt = 0, split = 0; #ifdef INVARIANTS struct rack_sendmap *insret; #endif memset(&fe, 0, sizeof(fe)); rack->rc_has_collapsed = 0; fe.r_start = rack->r_ctl.last_collapse_point; rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); if (rsm == NULL) { /* Nothing to do maybe the peer ack'ed it all */ rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL); return; } /* Now do we need to split this one? */ if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) { rack_log_collapse(rack, rsm->r_start, rsm->r_end, rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm); 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 */ split = 1; rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point); #ifndef INVARIANTS (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #else insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT, rack->r_ctl.last_collapse_point, __LINE__); 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: RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) { nrsm->r_flags |= RACK_RWND_COLLAPSED; rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm); cnt++; } if (cnt) { counter_u64_add(rack_collapsed_win, 1); } rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL); } 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; } } } static void rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack) { /* * If fast output is in progress, lets validate that * the new window did not shrink on us and make it * so fast output should end. */ if (rack->r_fast_output) { uint32_t out; /* * Calculate what we will send if left as is * and compare that to our send window. */ out = ctf_outstanding(tp); if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) { /* ok we have an issue */ if (out >= tp->snd_wnd) { /* Turn off fast output the window is met or collapsed */ rack->r_fast_output = 0; } else { /* we have some room left */ rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out; if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) { /* If not at least 1 full segment never mind */ rack->r_fast_output = 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; INP_WLOCK_ASSERT(tptoinpcb(tp)); rack = (struct tcp_rack *)tp->t_fb_ptr; 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; rack_validate_fo_sendwin_up(tp, rack); 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; rack_validate_fo_sendwin_up(tp, rack); 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, ctf_outstanding(tp), __LINE__); else if (rack->rc_has_collapsed) rack_un_collapse_window(rack, __LINE__); if ((rack->r_collapse_point_valid) && (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point))) rack->r_collapse_point_valid = 0; /* 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) || rack->rc_has_collapsed) && sbavail(&tptosocket(tp)->so_snd) && (sbavail(&tptosocket(tp)->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. */ 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; /* * 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 = tcp_get_flags(th) & 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); rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1); /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); #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_WAKESOR) { tp->t_flags &= ~TF_WAKESOR; /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); } } 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) { /* The socket upcall is handled by socantrcvmore. */ socantrcvmore(so); /* * 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; } 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, 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); } rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1); /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); #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->gp_ready && (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, 0, 0); /* Did the window get updated? */ if (tiwin != tp->snd_wnd) { tp->snd_wnd = tiwin; rack_validate_fo_sendwin_up(tp, rack); 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) || rack->rc_has_collapsed) && sbavail(&tptosocket(tp)->so_snd) && (sbavail(&tptosocket(tp)->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. */ 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_RCVD_TSTMP) == 0)) { tp->t_flags &= ~TF_PREVVALID; if (tp->t_rxtshift == 1 && (int)(ticks - tp->t_badrxtwin) < 0) rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__); } /* * 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); if (acked) { struct mbuf *mfree; rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0); SOCKBUF_LOCK(&so->so_snd); mfree = sbcut_locked(&so->so_snd, acked); tp->snd_una = th->th_ack; /* Note we want to hold the sb lock through the sendmap adjust */ rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una); /* Wake up the socket if we have room to write more */ rack_log_wakeup(tp,rack, &so->so_snd, acked, 2); sowwakeup_locked(so); m_freem(mfree); tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); 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. */ if (tp->snd_wnd < ctf_outstanding(tp)) { /* The peer collapsed the window */ rack_collapsed_window(rack, ctf_outstanding(tp), __LINE__); } else if (rack->rc_has_collapsed) rack_un_collapse_window(rack, __LINE__); if ((rack->r_collapse_point_valid) && (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point))) rack->r_collapse_point_valid = 0; /* * 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) { tp->t_flags &= ~TF_PREVVALID; rack->r_ctl.retran_during_recovery = 0; rack->r_ctl.dsack_byte_cnt = 0; 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(&tptosocket(tp)->so_snd) == 0) tp->t_acktime = 0; rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); } if (acked && rack->r_fast_output) rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked); 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; INP_WLOCK_ASSERT(tptoinpcb(tp)); 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; } tcp_ecn_input_syn_sent(tp, thflags, iptos); 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 | TF_SONOTCONN); tcp_state_change(tp, TCPS_SYN_RECEIVED); } /* * 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, mcts; mcts = tcp_ts_getticks(); t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4); tcp_rack_xmit_timer(rack, t + 1, 1, t, 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, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)); /* * 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, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { 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; rack_validate_fo_sendwin_up(tp, rack); /* * 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); if (tp->t_flags & TF_SONOTCONN) { tp->t_flags &= ~TF_SONOTCONN; 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); if (tp->t_flags & TF_WAKESOR) { tp->t_flags &= ~TF_WAKESOR; /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); } } 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, mcts; mcts = tcp_ts_getticks(); t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5); tcp_rack_xmit_timer(rack, t + 1, 1, t, 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, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, iptos, &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, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { 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__); ctf_do_dropwithreset_conn(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; struct tcp_rack *rack; rack = (struct tcp_rack *)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, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, iptos, &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, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { 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__); ctf_do_dropwithreset_conn(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; struct tcp_rack *rack; rack = (struct tcp_rack *)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, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, iptos, &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, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((tp->t_flags & TF_CLOSED) && tlen && 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__); ctf_do_dropwithreset_conn(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; struct tcp_rack *rack; rack = (struct tcp_rack *)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, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, iptos, &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, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((tp->t_flags & TF_CLOSED) && tlen && 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__); ctf_do_dropwithreset_conn(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; struct tcp_rack *rack; rack = (struct tcp_rack *)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, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, iptos, &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, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((tp->t_flags & TF_CLOSED) && tlen && 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__); ctf_do_dropwithreset_conn(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; struct tcp_rack *rack; rack = (struct tcp_rack *)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, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, iptos, &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, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((tp->t_flags & TF_CLOSED) && tlen && 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__); ctf_do_dropwithreset_conn(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 *fill_override) { uint64_t bw_est, rate_wanted; int chged = 0; uint32_t user_max, orig_min, orig_max; orig_min = rack->r_ctl.rc_pace_min_segs; orig_max = rack->r_ctl.rc_pace_max_segs; 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; if (fill_override) rate_wanted = *fill_override; else rate_wanted = rack_get_output_bw(rack, bw_est, NULL, 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_pacing_sizes(tp, rack, orig_min, orig_max, line, 2); } static void rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack) { #ifdef INET6 struct ip6_hdr *ip6 = NULL; #endif #ifdef INET struct ip *ip = NULL; #endif struct udphdr *udp = NULL; /* Ok lets fill in the fast block, it can only be used with no IP options! */ #ifdef INET6 if (rack->r_is_v6) { rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr; if (tp->t_port) { rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr); udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; rack->r_ctl.fsb.udp = udp; rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1); } else { rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1); rack->r_ctl.fsb.udp = NULL; } tcpip_fillheaders(rack->rc_inp, tp->t_port, ip6, rack->r_ctl.fsb.th); } else #endif /* INET6 */ { rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr); ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr; if (tp->t_port) { rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr); udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; rack->r_ctl.fsb.udp = udp; rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1); } else { rack->r_ctl.fsb.udp = NULL; rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1); } tcpip_fillheaders(rack->rc_inp, tp->t_port, ip, rack->r_ctl.fsb.th); } rack->r_fsb_inited = 1; } static int rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack) { /* * Allocate the larger of spaces V6 if available else just * V4 and include udphdr (overbook) */ #ifdef INET6 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr); #else rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr); #endif rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len, M_TCPFSB, M_NOWAIT|M_ZERO); if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) { return (ENOMEM); } rack->r_fsb_inited = 0; return (0); } static int rack_init(struct tcpcb *tp) { struct inpcb *inp = tptoinpcb(tp); struct tcp_rack *rack = NULL; #ifdef INVARIANTS struct rack_sendmap *insret; #endif uint32_t iwin, snt, us_cts; int err; 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 recursive (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; rack->rc_inp = inp; /* Set the flag */ rack->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0; /* Probably not needed but lets be sure */ rack_clear_rate_sample(rack); /* * Save off the default values, socket options will poke * at these if pacing is not on or we have not yet * reached where pacing is on (gp_ready/fixed enabled). * When they get set into the CC module (when gp_ready * is enabled or we enable fixed) then we will set these * values into the CC and place in here the old values * so we have a restoral. Then we will set the flag * rc_pacing_cc_set. That way whenever we turn off pacing * or switch off this stack, we will know to go restore * the saved values. */ rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn; rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn; /* We want abe like behavior as well */ rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED; 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; rack->r_ctl.roundends = tp->snd_max; 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; #ifdef TCP_ACCOUNTING if (rack_tcp_accounting) { tp->t_flags2 |= TF2_TCP_ACCOUNTING; } #endif 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; TAILQ_INIT(&rack->r_ctl.opt_list); rack->r_ctl.rc_reorder_shift = rack_reorder_thresh; rack->r_ctl.rc_pkt_delay = rack_pkt_delay; rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp; rack->r_ctl.rc_lowest_us_rtt = 0xffffffff; rack->r_ctl.rc_highest_us_rtt = 0; rack->r_ctl.bw_rate_cap = rack_bw_rate_cap; rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop); if (rack_use_cmp_acks) rack->r_use_cmp_ack = 1; if (rack_disable_prr) rack->rack_no_prr = 1; if (rack_gp_no_rec_chg) rack->rc_gp_no_rec_chg = 1; if (rack_pace_every_seg && tcp_can_enable_pacing()) { rack->rc_always_pace = 1; if (rack->use_fixed_rate || rack->gp_ready) rack_set_cc_pacing(rack); } else rack->rc_always_pace = 0; if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) rack->r_mbuf_queue = 1; else rack->r_mbuf_queue = 0; if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack) inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; else inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; rack_set_pace_segments(tp, rack, __LINE__, NULL); if (rack_limits_scwnd) rack->r_limit_scw = 1; else rack->r_limit_scw = 0; rack->rc_labc = V_tcp_abc_l_var; 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->rc_init_win = rack_default_init_window; rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss; if (rack_hw_up_only) rack->r_up_only = 1; 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.challenge_ack_ts = tcp_ts_getticks(); rack->r_ctl.rc_time_probertt_starts = 0; if (rack_dsack_std_based & 0x1) { /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */ rack->rc_rack_tmr_std_based = 1; } if (rack_dsack_std_based & 0x2) { /* Basically this means rack timers are extended based on dsack by up to (2 * srtt) */ rack->rc_rack_use_dsack = 1; } /* We require at least one measurement, even if the sysctl is 0 */ if (rack_req_measurements) rack->r_ctl.req_measurements = rack_req_measurements; else rack->r_ctl.req_measurements = 1; if (rack_enable_hw_pacing) rack->rack_hdw_pace_ena = 1; if (rack_hw_rate_caps) rack->r_rack_hw_rate_caps = 1; /* 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; err = rack_init_fsb(tp, rack); if (err) { uma_zfree(rack_pcb_zone, tp->t_fb_ptr); tp->t_fb_ptr = NULL; return (err); } 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_no_rtt_allowed = 1; rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time); rsm->r_rtr_cnt = 1; rsm->r_rtr_bytes = 0; if (tp->t_flags & TF_SENTFIN) rsm->r_flags |= RACK_HAS_FIN; if ((tp->snd_una == tp->iss) && !TCPS_HAVEESTABLISHED(tp->t_state)) rsm->r_flags |= RACK_HAS_SYN; rsm->r_start = tp->snd_una; rsm->r_end = tp->snd_max; rsm->r_dupack = 0; if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) { rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff); if (rsm->m) rsm->orig_m_len = rsm->m->m_len; else rsm->orig_m_len = 0; } else { /* * This can happen if we have a stand-alone FIN or * SYN. */ rsm->m = NULL; rsm->orig_m_len = 0; rsm->soff = 0; } #ifndef INVARIANTS (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #else insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); 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; } /* * Timers in Rack are kept in microseconds so lets * convert any initial incoming variables * from ticks into usecs. Note that we * also change the values of t_srtt and t_rttvar, if * they are non-zero. They are kept with a 5 * bit decimal so we have to carefully convert * these to get the full precision. */ rack_convert_rtts(tp); tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow); if (rack_do_hystart) { - tp->ccv->flags |= CCF_HYSTART_ALLOWED; + tp->t_ccv.flags |= CCF_HYSTART_ALLOWED; if (rack_do_hystart > 1) - tp->ccv->flags |= CCF_HYSTART_CAN_SH_CWND; + tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND; if (rack_do_hystart > 2) - tp->ccv->flags |= CCF_HYSTART_CONS_SSTH; + tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH; } if (rack_def_profile) rack_set_profile(rack, rack_def_profile); /* 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); /* Lets setup the fsb block */ rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0); rack_log_rtt_shrinks(rack, us_cts, tp->t_rxtcur, __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) { struct inpcb *inp = tptoinpcb(tp); if (tp->t_fb_ptr) { struct tcp_rack *rack; struct rack_sendmap *rsm, *nrsm; #ifdef INVARIANTS struct rack_sendmap *rm; #endif rack = (struct tcp_rack *)tp->t_fb_ptr; if (tp->t_in_pkt) { /* * It is unsafe to process the packets since a * reset may be lurking in them (its rare but it * can occur). If we were to find a RST, then we * would end up dropping the connection and the * INP lock, so when we return the caller (tcp_usrreq) * will blow up when it trys to unlock the inp. */ struct mbuf *save, *m; m = tp->t_in_pkt; tp->t_in_pkt = NULL; tp->t_tail_pkt = NULL; while (m) { save = m->m_nextpkt; m->m_nextpkt = NULL; m_freem(m); m = save; } } tp->t_flags &= ~TF_FORCEDATA; #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 if (rack->r_ctl.fsb.tcp_ip_hdr) { free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB); rack->r_ctl.fsb.tcp_ip_hdr = NULL; rack->r_ctl.fsb.th = NULL; } /* Convert back to ticks, with */ if (tp->t_srtt > 1) { uint32_t val, frac; val = USEC_2_TICKS(tp->t_srtt); frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz); tp->t_srtt = val << TCP_RTT_SHIFT; /* * frac is the fractional part here is left * over from converting to hz and shifting. * We need to convert this to the 5 bit * remainder. */ if (frac) { if (hz == 1000) { frac = (((uint64_t)frac * (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC); } else { frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC); } tp->t_srtt += frac; } } if (tp->t_rttvar) { uint32_t val, frac; val = USEC_2_TICKS(tp->t_rttvar); frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz); tp->t_rttvar = val << TCP_RTTVAR_SHIFT; /* * frac is the fractional part here is left * over from converting to hz and shifting. * We need to convert this to the 5 bit * remainder. */ if (frac) { if (hz == 1000) { frac = (((uint64_t)frac * (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC); } else { frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC); } tp->t_rttvar += frac; } } tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur); tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow); if (rack->rc_always_pace) { tcp_decrement_paced_conn(); rack_undo_cc_pacing(rack); rack->rc_always_pace = 0; } /* Clean up any options if they were not applied */ while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) { struct deferred_opt_list *dol; dol = TAILQ_FIRST(&rack->r_ctl.opt_list); TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next); free(dol, M_TCPDO); } /* rack does not use force data but other stacks may clear it */ if (rack->r_ctl.crte != NULL) { tcp_rel_pacing_rate(rack->r_ctl.crte, tp); rack->rack_hdrw_pacing = 0; rack->r_ctl.crte = NULL; } #ifdef TCP_BLACKBOX tcp_log_flowend(tp); #endif RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) { #ifndef INVARIANTS (void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #else rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); 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; } inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY; inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; inp->inp_flags2 &= ~INP_MBUF_ACKCMP; /* Cancel the GP measurement in progress */ tp->t_flags &= ~TF_GPUTINPROG; inp->inp_flags2 &= ~INP_MBUF_L_ACKS; /* 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) { if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) { rack->r_is_v6 = (tptoinpcb(tp)->inp_vflag & INP_IPV6) != 0; } 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__, NULL); rack->r_state = TCPS_ESTABLISHED; rack->r_substate = rack_do_established; break; case TCPS_CLOSE_WAIT: rack_set_pace_segments(tp, rack, __LINE__, NULL); rack->r_state = TCPS_CLOSE_WAIT; rack->r_substate = rack_do_close_wait; break; case TCPS_FIN_WAIT_1: rack_set_pace_segments(tp, rack, __LINE__, NULL); rack->r_state = TCPS_FIN_WAIT_1; rack->r_substate = rack_do_fin_wait_1; break; case TCPS_CLOSING: rack_set_pace_segments(tp, rack, __LINE__, NULL); rack->r_state = TCPS_CLOSING; rack->r_substate = rack_do_closing; break; case TCPS_LAST_ACK: rack_set_pace_segments(tp, rack, __LINE__, NULL); rack->r_state = TCPS_LAST_ACK; rack->r_substate = rack_do_lastack; break; case TCPS_FIN_WAIT_2: rack_set_pace_segments(tp, rack, __LINE__, NULL); 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; }; if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state)) rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP; } 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(&tptosocket(tp)->so_snd) && (tmr_up == PACE_TMR_RXT)) { /* * if we hit enobufs then we would expect the possibility * 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 (tcp_in_hpts(rack->rc_inp)) { 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(rack->rc_inp); } rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0); } static void rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts, uint32_t high_seq) { if ((SEQ_LT(tp->snd_wl1, seq) || (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) || (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) { /* keep track of pure window updates */ if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd)) KMOD_TCPSTAT_INC(tcps_rcvwinupd); tp->snd_wnd = tiwin; rack_validate_fo_sendwin_up(tp, rack); tp->snd_wl1 = seq; tp->snd_wl2 = ack; if (tp->snd_wnd > tp->max_sndwnd) tp->max_sndwnd = tp->snd_wnd; rack->r_wanted_output = 1; } else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) { tp->snd_wnd = tiwin; rack_validate_fo_sendwin_up(tp, rack); tp->snd_wl1 = seq; tp->snd_wl2 = ack; } else { /* Not a valid win update */ return; } /* 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) || rack->rc_has_collapsed) && sbavail(&tptosocket(tp)->so_snd) && (sbavail(&tptosocket(tp)->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. */ rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime); } } static void rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { struct inpcb *inp = tptoinpcb(tp); union tcp_log_stackspecific log; struct timeval ltv; char tcp_hdr_buf[60]; struct tcphdr *th; struct timespec ts; uint32_t orig_snd_una; uint8_t xx = 0; #ifdef NETFLIX_HTTP_LOGGING struct http_sendfile_track *http_req; if (SEQ_GT(ae->ack, tp->snd_una)) { http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1)); } else { http_req = tcp_http_find_req_for_seq(tp, ae->ack); } #endif memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); 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.use_lt_bw = rack->r_ent_rec_ns; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->r_might_revert; log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced; log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = tp->t_maxseg; log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; log.u_bbr.flex7 = 1; log.u_bbr.lost = ae->flags; log.u_bbr.cwnd_gain = ackval; log.u_bbr.pacing_gain = 0x2; if (ae->flags & TSTMP_HDWR) { /* Record the hardware timestamp if present */ log.u_bbr.flex3 = M_TSTMP; ts.tv_sec = ae->timestamp / 1000000000; ts.tv_nsec = ae->timestamp % 1000000000; 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 (ae->flags & TSTMP_LRO) { /* Record the LRO the arrival timestamp */ log.u_bbr.flex3 = M_TSTMP_LRO; ts.tv_sec = ae->timestamp / 1000000000; ts.tv_nsec = ae->timestamp % 1000000000; 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 = ae->timestamp; #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 memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf)); th = (struct tcphdr *)tcp_hdr_buf; th->th_seq = ae->seq; th->th_ack = ae->ack; th->th_win = ae->win; /* Now fill in the ports */ th->th_sport = inp->inp_fport; th->th_dport = inp->inp_lport; tcp_set_flags(th, ae->flags); /* Now do we have a timestamp option? */ if (ae->flags & HAS_TSTMP) { u_char *cp; uint32_t val; th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2); cp = (u_char *)(th + 1); *cp = TCPOPT_NOP; cp++; *cp = TCPOPT_NOP; cp++; *cp = TCPOPT_TIMESTAMP; cp++; *cp = TCPOLEN_TIMESTAMP; cp++; val = htonl(ae->ts_value); bcopy((char *)&val, (char *)cp, sizeof(uint32_t)); val = htonl(ae->ts_echo); bcopy((char *)&val, (char *)(cp + 4), sizeof(uint32_t)); } else th->th_off = (sizeof(struct tcphdr) >> 2); /* * For sane logging we need to play a little trick. * If the ack were fully processed we would have moved * snd_una to high_seq, but since compressed acks are * processed in two phases, at this point (logging) snd_una * won't be advanced. So we would see multiple acks showing * the advancement. We can prevent that by "pretending" that * snd_una was advanced and then un-advancing it so that the * logging code has the right value for tlb_snd_una. */ if (tp->snd_una != high_seq) { orig_snd_una = tp->snd_una; tp->snd_una = high_seq; xx = 1; } else xx = 0; TCP_LOG_EVENTP(tp, th, &tptosocket(tp)->so_rcv, &tptosocket(tp)->so_snd, TCP_LOG_IN, 0, 0, &log, true, <v); if (xx) { tp->snd_una = orig_snd_una; } } } static void rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts) { uint32_t us_rtt; /* * A persist or keep-alive was forced out, update our * min rtt time. Note now worry about lost responses. * When a subsequent keep-alive or persist times out * and forced_ack is still on, then the last probe * was not responded to. In such cases we have a * sysctl that controls the behavior. Either we apply * the rtt but with reduced confidence (0). Or we just * plain don't apply the rtt estimate. Having data flow * will clear the probe_not_answered flag i.e. cum-ack * move forward exiting and reentering persists. */ rack->forced_ack = 0; rack->rc_tp->t_rxtshift = 0; if ((rack->rc_in_persist && (tiwin == rack->rc_tp->snd_wnd)) || (rack->rc_in_persist == 0)) { /* * In persists only apply the RTT update if this is * a response to our window probe. And that * means the rwnd sent must match the current * snd_wnd. If it does not, then we got a * window update ack instead. For keepalive * we allow the answer no matter what the window. * * Note that if the probe_not_answered is set then * the forced_ack_ts is the oldest one i.e. the first * probe sent that might have been lost. This assures * us that if we do calculate an RTT it is longer not * some short thing. */ if (rack->rc_in_persist) counter_u64_add(rack_persists_acks, 1); us_rtt = us_cts - rack->r_ctl.forced_ack_ts; if (us_rtt == 0) us_rtt = 1; if (rack->probe_not_answered == 0) { rack_apply_updated_usrtt(rack, us_rtt, us_cts); tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1); } else { /* We have a retransmitted probe here too */ if (rack_apply_rtt_with_reduced_conf) { rack_apply_updated_usrtt(rack, us_rtt, us_cts); tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1); } } } } static int rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv) { /* * Handle a "special" compressed ack mbuf. Each incoming * ack has only four possible dispositions: * * A) It moves the cum-ack forward * B) It is behind the cum-ack. * C) It is a window-update ack. * D) It is a dup-ack. * * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES * in the incoming mbuf. We also need to still pay attention * to nxt_pkt since there may be another packet after this * one. */ #ifdef TCP_ACCOUNTING uint64_t ts_val; uint64_t rdstc; #endif int segsiz; struct timespec ts; struct tcp_rack *rack; struct tcp_ackent *ae; uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack; int cnt, i, did_out, ourfinisacked = 0; struct tcpopt to_holder, *to = NULL; #ifdef TCP_ACCOUNTING int win_up_req = 0; #endif int nsegs = 0; int under_pacing = 1; int recovery = 0; #ifdef TCP_ACCOUNTING sched_pin(); #endif rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack->gp_ready && (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) under_pacing = 0; else under_pacing = 1; if (rack->r_state != tp->t_state) rack_set_state(tp, rack); if ((tp->t_state >= TCPS_FIN_WAIT_1) && (tp->t_flags & TF_GPUTINPROG)) { /* * We have a goodput in progress * and we have entered a late state. * Do we have enough data in the sb * to handle the GPUT request? */ uint32_t bytes; bytes = tp->gput_ack - tp->gput_seq; if (SEQ_GT(tp->gput_seq, tp->snd_una)) bytes += tp->gput_seq - tp->snd_una; if (bytes > sbavail(&tptosocket(tp)->so_snd)) { /* * There are not enough bytes in the socket * buffer that have been sent to cover this * measurement. Cancel it. */ rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/, rack->r_ctl.rc_gp_srtt /*flex1*/, tp->gput_seq, 0, 0, 18, __LINE__, NULL, 0); tp->t_flags &= ~TF_GPUTINPROG; } } to = &to_holder; to->to_flags = 0; KASSERT((m->m_len >= sizeof(struct tcp_ackent)), ("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len)); cnt = m->m_len / sizeof(struct tcp_ackent); counter_u64_add(rack_multi_single_eq, cnt); high_seq = tp->snd_una; the_win = tp->snd_wnd; win_seq = tp->snd_wl1; win_upd_ack = tp->snd_wl2; cts = tcp_tv_to_usectick(tv); ms_cts = tcp_tv_to_mssectick(tv); rack->r_ctl.rc_rcvtime = cts; segsiz = ctf_fixed_maxseg(tp); if ((rack->rc_gp_dyn_mul) && (rack->use_fixed_rate == 0) && (rack->rc_always_pace)) { /* Check in on probertt */ rack_check_probe_rtt(rack, cts); } for (i = 0; i < cnt; i++) { #ifdef TCP_ACCOUNTING ts_val = get_cyclecount(); #endif rack_clear_rate_sample(rack); ae = ((mtod(m, struct tcp_ackent *)) + i); /* Setup the window */ tiwin = ae->win << tp->snd_scale; if (tiwin > rack->r_ctl.rc_high_rwnd) rack->r_ctl.rc_high_rwnd = tiwin; /* figure out the type of ack */ if (SEQ_LT(ae->ack, high_seq)) { /* Case B*/ ae->ack_val_set = ACK_BEHIND; } else if (SEQ_GT(ae->ack, high_seq)) { /* Case A */ ae->ack_val_set = ACK_CUMACK; } else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){ /* Case D */ ae->ack_val_set = ACK_DUPACK; } else { /* Case C */ ae->ack_val_set = ACK_RWND; } rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq); /* Validate timestamp */ if (ae->flags & HAS_TSTMP) { /* Setup for a timestamp */ to->to_flags = TOF_TS; ae->ts_echo -= tp->ts_offset; to->to_tsecr = ae->ts_echo; to->to_tsval = ae->ts_value; /* * 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 (TSTMP_GT(ae->ts_echo, ms_cts)) to->to_tsecr = 0; if (tp->ts_recent && TSTMP_LT(ae->ts_value, tp->ts_recent)) { if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) { #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val); } } #endif continue; } } if (SEQ_LEQ(ae->seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, ae->seq)) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = ae->ts_value; } } else { /* Setup for a no options */ to->to_flags = 0; } /* Update the rcv time and perform idle reduction possibly */ if (tp->t_idle_reduce && (tp->snd_max == tp->snd_una) && (TICKS_2_USEC(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; /* Now what about ECN of a chain of pure ACKs? */ if (tcp_ecn_input_segment(tp, ae->flags, 0, tcp_packets_this_ack(tp, ae->ack), ae->codepoint)) rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__); #ifdef TCP_ACCOUNTING /* Count for the specific type of ack in */ counter_u64_add(tcp_cnt_counters[ae->ack_val_set], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[ae->ack_val_set]++; } #endif /* * Note how we could move up these in the determination * above, but we don't so that way the timestamp checks (and ECN) * is done first before we do any processing on the ACK. * The non-compressed path through the code has this * weakness (noted by @jtl) that it actually does some * processing before verifying the timestamp information. * We don't take that path here which is why we set * the ack_val_set first, do the timestamp and ecn * processing, and then look at what we have setup. */ if (ae->ack_val_set == ACK_BEHIND) { /* * Case B flag reordering, if window is not closed * or it could be a keep-alive or persists */ if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) { rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); } } else if (ae->ack_val_set == ACK_DUPACK) { /* Case D */ rack_strike_dupack(rack); } else if (ae->ack_val_set == ACK_RWND) { /* Case C */ if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) { ts.tv_sec = ae->timestamp / 1000000000; ts.tv_nsec = ae->timestamp % 1000000000; 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; } if (rack->forced_ack) { rack_handle_probe_response(rack, tiwin, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time)); } #ifdef TCP_ACCOUNTING win_up_req = 1; #endif win_upd_ack = ae->ack; win_seq = ae->seq; the_win = tiwin; rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq); } else { /* Case A */ if (SEQ_GT(ae->ack, tp->snd_max)) { /* * We just send an ack since the incoming * ack is beyond the largest seq we sent. */ if ((tp->t_flags & TF_ACKNOW) == 0) { ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt); if (tp->t_flags && TF_ACKNOW) rack->r_wanted_output = 1; } } else { nsegs++; /* If the window changed setup to update */ if (tiwin != tp->snd_wnd) { win_upd_ack = ae->ack; win_seq = ae->seq; the_win = tiwin; rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq); } #ifdef TCP_ACCOUNTING /* Account for the acks */ if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz); } counter_u64_add(tcp_cnt_counters[CNT_OF_ACKS_IN], (((ae->ack - high_seq) + segsiz - 1) / segsiz)); #endif high_seq = ae->ack; if (rack_verbose_logging && (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.timeStamp = tcp_get_usecs(&tv); log.u_bbr.flex1 = high_seq; log.u_bbr.flex2 = rack->r_ctl.roundends; log.u_bbr.flex3 = rack->r_ctl.current_round; log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround; log.u_bbr.flex8 = 8; tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0, 0, &log, false, NULL, NULL, 0, &tv); } /* * The draft (v3) calls for us to use SEQ_GEQ, but that * causes issues when we are just going app limited. Lets * instead use SEQ_GT where its equal but more data * is outstanding. */ if ((SEQ_GT(high_seq, rack->r_ctl.roundends)) || ((high_seq == rack->r_ctl.roundends) && SEQ_GT(tp->snd_max, tp->snd_una))) { rack->r_ctl.current_round++; rack->r_ctl.roundends = tp->snd_max; if (CC_ALGO(tp)->newround != NULL) { - CC_ALGO(tp)->newround(tp->ccv, rack->r_ctl.current_round); + CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round); } } /* Setup our act_rcv_time */ if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) { ts.tv_sec = ae->timestamp / 1000000000; ts.tv_nsec = ae->timestamp % 1000000000; 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; } rack_process_to_cumack(tp, rack, ae->ack, cts, to); if (rack->rc_dsack_round_seen) { /* Is the dsack round over? */ if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) { /* Yes it is */ rack->rc_dsack_round_seen = 0; rack_log_dsack_event(rack, 3, __LINE__, 0, 0); } } } } /* And lets be sure to commit the rtt measurements for this ack */ tcp_rack_xmit_timer_commit(rack, tp); #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val); if (ae->ack_val_set == ACK_CUMACK) tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val); } } #endif } #ifdef TCP_ACCOUNTING ts_val = get_cyclecount(); #endif /* Tend to any collapsed window */ if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) { /* The peer collapsed the window */ rack_collapsed_window(rack, (tp->snd_max - high_seq), __LINE__); } else if (rack->rc_has_collapsed) rack_un_collapse_window(rack, __LINE__); if ((rack->r_collapse_point_valid) && (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point))) rack->r_collapse_point_valid = 0; acked_amount = acked = (high_seq - tp->snd_una); if (acked) { /* * Clear the probe not answered flag * since cum-ack moved forward. */ rack->probe_not_answered = 0; if (rack->sack_attack_disable == 0) rack_do_decay(rack); if (acked >= segsiz) { /* * You only get credit for * MSS and greater (and you get extra * credit for larger cum-ack moves). */ int ac; ac = acked / segsiz; 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 (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++; acked_amount = acked = (high_seq - tp->snd_una); } if (acked > sbavail(&so->so_snd)) acked_amount = sbavail(&so->so_snd); #ifdef NETFLIX_EXP_DETECTION /* * We only care on a cum-ack move if we are in a sack-disabled * state. We have already added in to the ack_count, and we never * would disable on a cum-ack move, so we only care to do the * detection if it may "undo" it, i.e. we were in disabled already. */ if (rack->sack_attack_disable) rack_do_detection(tp, rack, acked_amount, segsiz); #endif if (IN_FASTRECOVERY(tp->t_flags) && (rack->rack_no_prr == 0)) rack_update_prr(tp, rack, acked_amount, high_seq); if (IN_RECOVERY(tp->t_flags)) { if (SEQ_LT(high_seq, tp->snd_recover) && (SEQ_LT(high_seq, tp->snd_max))) { tcp_rack_partialack(tp); } else { rack_post_recovery(tp, high_seq); recovery = 1; } } /* Handle the rack-log-ack part (sendmap) */ 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; /* * Lets make sure snd_una is updated * since most likely acked_amount = 0 (it * should be). */ tp->snd_una = high_seq; } /* Did we make a RTO error? */ if ((tp->t_flags & TF_PREVVALID) && ((tp->t_flags & TF_RCVD_TSTMP) == 0)) { tp->t_flags &= ~TF_PREVVALID; if (tp->t_rxtshift == 1 && (int)(ticks - tp->t_badrxtwin) < 0) rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__); } /* Handle the data in the socket buffer */ KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1); KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); if (acked_amount > 0) { struct mbuf *mfree; rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery); SOCKBUF_LOCK(&so->so_snd); mfree = sbcut_locked(&so->so_snd, acked_amount); tp->snd_una = high_seq; /* Note we want to hold the sb lock through the sendmap adjust */ rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una); /* Wake up the socket if we have room to write more */ rack_log_wakeup(tp,rack, &so->so_snd, acked, 2); sowwakeup_locked(so); m_freem(mfree); } /* update progress */ tp->t_acktime = ticks; rack_log_progress_event(rack, tp, tp->t_acktime, PROGRESS_UPDATE, __LINE__); /* Clear out shifts and such */ tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); rack->rc_tlp_in_progress = 0; rack->r_ctl.rc_tlp_cnt_out = 0; /* Send recover and snd_nxt must be dragged along */ 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 the RXT timer is running we want to * stop it, so we can restart a TLP (or new RXT). */ 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, high_seq); #endif tp->snd_wl2 = high_seq; tp->t_dupacks = 0; 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) { tp->t_flags &= ~TF_PREVVALID; rack->r_ctl.retran_during_recovery = 0; rack->r_ctl.dsack_byte_cnt = 0; 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(&tptosocket(tp)->so_snd) == 0) tp->t_acktime = 0; /* Set so we might enter persists... */ rack->r_wanted_output = 1; rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); 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 (not now in the past), time to * reset him. */ rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); /* tcp_close will kill the inp pre-log the Reset */ tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val); tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val); } } #endif m_freem(m); tp = tcp_close(tp); if (tp == NULL) { #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (1); } /* * We would normally do drop-with-reset which would * send back a reset. We can't since we don't have * all the needed bits. Instead lets arrange for * a call to tcp_output(). That way since we * are in the closed state we will generate a reset. * * Note if tcp_accounting is on we don't unpin since * we do that after the goto label. */ goto send_out_a_rst; } if ((sbused(&so->so_snd) == 0) && (tp->t_state >= TCPS_FIN_WAIT_1) && (tp->t_flags & TF_SENTFIN)) { /* * 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. * */ 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))); } if (ourfinisacked == 0) { /* * We don't change to fin-wait-2 if we have our fin acked * which means we are probably in TCPS_CLOSING. */ tcp_state_change(tp, TCPS_FIN_WAIT_2); } } } /* Wake up the socket if we have room to write more */ if (sbavail(&so->so_snd)) { rack->r_wanted_output = 1; if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr, tp, tick, PROGRESS_DROP, __LINE__); /* * We cheat here and don't send a RST, we should send one * when the pacer drops the connection. */ #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val); tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val); } } sched_unpin(); #endif (void)tcp_drop(tp, ETIMEDOUT); m_freem(m); return (1); } } if (ourfinisacked) { switch(tp->t_state) { case TCPS_CLOSING: #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val); tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val); } } sched_unpin(); #endif tcp_twstart(tp); m_freem(m); return (1); break; case TCPS_LAST_ACK: #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val); tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val); } } sched_unpin(); #endif tp = tcp_close(tp); ctf_do_drop(m, tp); return (1); break; case TCPS_FIN_WAIT_1: #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val); tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val); } } #endif 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; default: break; } } if (rack->r_fast_output) { /* * We re doing fast output.. can we expand that? */ rack_gain_for_fastoutput(rack, tp, so, acked_amount); } #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val); tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val); } } } else if (win_up_req) { rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_RWND] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val); } } #endif } /* Now is there a next packet, if so we are done */ m_freem(m); did_out = 0; if (nxt_pkt) { #ifdef TCP_ACCOUNTING sched_unpin(); #endif rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs); return (0); } rack_handle_might_revert(tp, rack); ctf_calc_rwin(so, tp); if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) { send_out_a_rst: if (tcp_output(tp) < 0) { #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (1); } did_out = 1; } rack_free_trim(rack); #ifdef TCP_ACCOUNTING sched_unpin(); #endif rack_timer_audit(tp, rack, &so->so_snd); rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs); return (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) { struct inpcb *inp = tptoinpcb(tp); #ifdef TCP_ACCOUNTING uint64_t ts_val; #endif int32_t thflags, retval, did_out = 0; int32_t way_out = 0; /* * cts - is the current time from tv (caller gets ts) in microseconds. * ms_cts - is the current time from tv in milliseconds. * us_cts - is the time that LRO or hardware actually got the packet in microseconds. */ uint32_t cts, us_cts, ms_cts; uint32_t tiwin, high_seq; struct timespec ts; struct tcpopt to; struct tcp_rack *rack; struct rack_sendmap *rsm; int32_t prev_state = 0; #ifdef TCP_ACCOUNTING int ack_val_set = 0xf; #endif int nsegs; NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(inp); /* * tv passed from common code is from either M_TSTMP_LRO or * tcp_get_usecs() if no LRO m_pkthdr timestamp is present. */ rack = (struct tcp_rack *)tp->t_fb_ptr; if (m->m_flags & M_ACKCMP) { /* * All compressed ack's are ack's by definition so * remove any ack required flag and then do the processing. */ rack->rc_ack_required = 0; return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv)); } if (m->m_flags & M_ACKCMP) { panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp); } cts = tcp_tv_to_usectick(tv); ms_cts = tcp_tv_to_mssectick(tv); nsegs = m->m_pkthdr.lro_nsegs; counter_u64_add(rack_proc_non_comp_ack, 1); thflags = tcp_get_flags(th); #ifdef TCP_ACCOUNTING sched_pin(); if (thflags & TH_ACK) ts_val = get_cyclecount(); #endif 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; /* * 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 TCP_ACCOUNTING if (thflags & TH_ACK) { /* * We have a tradeoff here. We can either do what we are * doing i.e. pinning to this CPU and then doing the accounting * we could do a critical enter, setup the rdtsc and cpu * as in below, and then validate we are on the same CPU on * exit. I have choosen to not do the critical enter since * that often will gain you a context switch, and instead lock * us (line above this if) to the same CPU with sched_pin(). This * means we may be context switched out for a higher priority * interupt but we won't be moved to another CPU. * * If this occurs (which it won't very often since we most likely * are running this code in interupt context and only a higher * priority will bump us ... clock?) we will falsely add in * to the time the interupt processing time plus the ack processing * time. This is ok since its a rare event. */ ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin, ctf_fixed_maxseg(tp)); } #endif /* * Parse options on any incoming segment. */ memset(&to, 0, sizeof(to)); tcp_dooptions(&to, (u_char *)(th + 1), (th->th_off << 2) - sizeof(struct tcphdr), (thflags & TH_SYN) ? TO_SYN : 0); 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_state >= TCPS_FIN_WAIT_1) && (tp->t_flags & TF_GPUTINPROG)) { /* * We have a goodput in progress * and we have entered a late state. * Do we have enough data in the sb * to handle the GPUT request? */ uint32_t bytes; bytes = tp->gput_ack - tp->gput_seq; if (SEQ_GT(tp->gput_seq, tp->snd_una)) bytes += tp->gput_seq - tp->snd_una; if (bytes > sbavail(&tptosocket(tp)->so_snd)) { /* * There are not enough bytes in the socket * buffer that have been sent to cover this * measurement. Cancel it. */ rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/, rack->r_ctl.rc_gp_srtt /*flex1*/, tp->gput_seq, 0, 0, 18, __LINE__, NULL, 0); tp->t_flags &= ~TF_GPUTINPROG; } } high_seq = th->th_ack; 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 = tcp_in_hpts(rack->rc_inp); 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.use_lt_bw = rack->r_ent_rec_ns; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->r_might_revert; 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; log.u_bbr.lost = thflags; log.u_bbr.pacing_gain = 0x1; #ifdef TCP_ACCOUNTING log.u_bbr.cwnd_gain = ack_val_set; #endif log.u_bbr.flex7 = 2; 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); } /* Remove ack required flag if set, we have one */ if (thflags & TH_ACK) rack->rc_ack_required = 0; if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) { way_out = 4; 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(m, tp, th, BANDLIM_RST_OPENPORT, tlen); #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (1); } /* * 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; m_freem(m); 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_2_USEC(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; #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 (tcp_ecn_input_segment(tp, thflags, tlen, tcp_packets_this_ack(tp, th->th_ack), iptos)) rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__); /* * 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, ms_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 = inp; } /* * 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 */ tcp_ecn_input_parallel_syn(tp, thflags, iptos); 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; rack_validate_fo_sendwin_up(tp, rack); 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 ((inp->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); #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (1); } tcp_set_hpts(inp); 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); } rack_clear_rate_sample(rack); if ((rack->forced_ack) && ((tcp_get_flags(th) & TH_RST) == 0)) { rack_handle_probe_response(rack, tiwin, 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; retval = (*rack->r_substate) (m, th, so, tp, &to, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt, iptos); if (retval == 0) { /* * If retval is 1 the tcb is unlocked and most likely the tp * is gone. */ INP_WLOCK_ASSERT(inp); 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__, NULL); } } tcp_rack_xmit_timer_commit(rack, tp); #ifdef TCP_ACCOUNTING /* * If we set the ack_val_se to what ack processing we are doing * we also want to track how many cycles we burned. Note * the bits after tcp_output we let be "free". This is because * we are also tracking the tcp_output times as well. Note the * use of 0xf here since we only have 11 counter (0 - 0xa) and * 0xf cannot be returned and is what we initialize it too to * indicate we are not doing the tabulations. */ if (ack_val_set != 0xf) { uint64_t crtsc; crtsc = get_cyclecount(); counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val); } } #endif if (nxt_pkt == 0) { if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) { do_output_now: if (tcp_output(tp) < 0) return (1); did_out = 1; } rack_start_hpts_timer(rack, tp, cts, 0, 0, 0); rack_free_trim(rack); } /* Update any rounds needed */ if (rack_verbose_logging && (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.timeStamp = tcp_get_usecs(&tv); log.u_bbr.flex1 = high_seq; log.u_bbr.flex2 = rack->r_ctl.roundends; log.u_bbr.flex3 = rack->r_ctl.current_round; log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround; log.u_bbr.flex8 = 9; tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0, 0, &log, false, NULL, NULL, 0, &tv); } /* * The draft (v3) calls for us to use SEQ_GEQ, but that * causes issues when we are just going app limited. Lets * instead use SEQ_GT where its equal but more data * is outstanding. */ if ((SEQ_GT(tp->snd_una, rack->r_ctl.roundends)) || ((tp->snd_una == rack->r_ctl.roundends) && SEQ_GT(tp->snd_max, tp->snd_una))) { rack->r_ctl.current_round++; rack->r_ctl.roundends = tp->snd_max; if (CC_ALGO(tp)->newround != NULL) { - CC_ALGO(tp)->newround(tp->ccv, rack->r_ctl.current_round); + CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round); } } 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) && (tcp_in_hpts(rack->rc_inp)) && (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { /* keep alive not needed if we are hptsi output yet */ ; } else { int late = 0; if (tcp_in_hpts(inp)) { 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(inp); } 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_get_usecs(NULL), 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, max(1, nsegs)); if (did_out) rack->r_wanted_output = 0; #ifdef TCP_ACCOUNTING } else { /* * Track the time (see above). */ if (ack_val_set != 0xf) { uint64_t crtsc; crtsc = get_cyclecount(); counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val)); /* * Note we *DO NOT* increment the per-tcb counters since * in the else the TP may be gone!! */ } #endif } #ifdef TCP_ACCOUNTING sched_unpin(); #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) { mbuf_tstmp2timeval(m, &tv); } 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) { INP_WUNLOCK(tptoinpcb(tp)); } } 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 (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) { return (rsm); } 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 (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) { /* * No sack so we automatically do the 3 strikes and * retransmit (no rack timer would be started). */ return (rsm); } 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 = (uint32_t)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); rack->r_fast_output = 0; 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, uint8_t quality) { 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->rc_ack_can_sendout_data; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->r_late; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->r_early; 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->measure_saw_probe_rtt; 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->gp_ready; log.u_bbr.pkt_epoch = line; log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed; log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early; 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; log.u_bbr.bbr_substate = quality; 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 int32_t pace_to_fill_cwnd(struct tcp_rack *rack, int32_t slot, uint32_t len, uint32_t segsiz, int *capped, uint64_t *rate_wanted, uint8_t non_paced) { uint64_t lentim, fill_bw; /* Lets first see if we are full, if so continue with normal rate */ rack->r_via_fill_cw = 0; 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; if (rack->r_fill_less_agg) { /* * Now take away the inflight (this will reduce our * aggressiveness and yeah, if we get that much out in 1RTT * we will have had acks come back and still be behind). */ fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); } /* Now lets make it into a b/w */ 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 (non_paced) *rate_wanted = fill_bw; if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted)) return (slot); if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap)) fill_bw = rack->r_ctl.bw_rate_cap; rack->r_via_fill_cw = 1; if (rack->r_rack_hw_rate_caps && (rack->r_ctl.crte != NULL)) { uint64_t high_rate; high_rate = tcp_hw_highest_rate(rack->r_ctl.crte); if (fill_bw > high_rate) { /* We are capping bw at the highest rate table entry */ if (*rate_wanted > high_rate) { /* The original rate was also capped */ rack->r_via_fill_cw = 0; } rack_log_hdwr_pacing(rack, fill_bw, high_rate, __LINE__, 0, 3); fill_bw = high_rate; if (capped) *capped = 1; } } else if ((rack->r_ctl.crte == NULL) && (rack->rack_hdrw_pacing == 0) && (rack->rack_hdw_pace_ena) && rack->r_rack_hw_rate_caps && (rack->rack_attempt_hdwr_pace == 0) && (rack->rc_inp->inp_route.ro_nh != NULL) && (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) { /* * Ok we may have a first attempt that is greater than our top rate * lets check. */ uint64_t high_rate; high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp); if (high_rate) { if (fill_bw > high_rate) { fill_bw = high_rate; if (capped) *capped = 1; } } } /* * 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; *rate_wanted = fill_bw; if (non_paced || (lentim < slot)) { rack_log_pacing_delay_calc(rack, len, slot, fill_bw, 0, lentim, 12, __LINE__, NULL, 0); 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) { uint64_t srtt; int32_t slot = 0; int can_start_hw_pacing = 1; 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. */ uint64_t 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. */ if (rack->r_ctl.rc_rack_min_rtt) srtt = rack->r_ctl.rc_rack_min_rtt; else srtt = max(tp->t_srtt, 1); if (rack->r_ctl.rc_rack_largest_cwnd) cwnd = rack->r_ctl.rc_rack_largest_cwnd; else cwnd = rack->r_ctl.cwnd_to_use; /* Inflate cwnd by 1000 so srtt of usecs is in ms */ tr_perms = (cwnd * 1000) / 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 (rack->rc_pace_to_cwnd) { uint64_t rate_wanted = 0; slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1); rack->rc_ack_can_sendout_data = 1; rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0); } else rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0); } else { uint64_t bw_est, res, lentim, rate_wanted; uint32_t orig_val, segs, oh; int capped = 0; int prev_fill; if ((rack->r_rr_config == 1) && rsm) { return (rack->r_ctl.rc_min_to); } 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, &capped); } if ((bw_est == 0) || (rate_wanted == 0) || ((rack->gp_ready == 0) && (rack->use_fixed_rate == 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 (rack_hw_rate_min && (bw_est < rack_hw_rate_min)) can_start_hw_pacing = 0; 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; if (rack->r_ctl.crte == NULL) { /* * Only do this if we are not hardware pacing * since if we are doing hw-pacing below we will * set make a call after setting up or changing * the rate. */ rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL); } else if (rack->rc_inp->inp_snd_tag == NULL) { /* * We lost our rate somehow, this can happen * if the interface changed underneath us. */ tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp); rack->r_ctl.crte = NULL; /* Lets re-allow attempting to setup pacing */ rack->rack_hdrw_pacing = 0; rack->rack_attempt_hdwr_pace = 0; rack_log_hdwr_pacing(rack, rate_wanted, bw_est, __LINE__, 0, 6); } /* 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, 0); prev_fill = rack->r_via_fill_cw; if ((rack->rc_pace_to_cwnd) && (capped == 0) && (rack->use_fixed_rate == 0) && (rack->in_probe_rtt == 0) && (IN_FASTRECOVERY(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, &capped, &rate_wanted, 0); } 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) && (can_start_hw_pacing > 0) && (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, &rack->r_ctl.crte_prev_rate); if (rack->r_ctl.crte) { rack->rack_hdrw_pacing = 1; rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz, 0, rack->r_ctl.crte, NULL); rack_log_hdwr_pacing(rack, rate_wanted, rack->r_ctl.crte->rate, __LINE__, err, 0); rack->r_ctl.last_hw_bw_req = rate_wanted; } else { counter_u64_add(rack_hw_pace_init_fail, 1); } } else if (rack->rack_hdrw_pacing && (rack->r_ctl.last_hw_bw_req != rate_wanted)) { /* Do we need to adjust our rate? */ const struct tcp_hwrate_limit_table *nrte; if (rack->r_up_only && (rate_wanted < rack->r_ctl.crte->rate)) { /** * We have four possible states here * having to do with the previous time * and this time. * previous | this-time * A) 0 | 0 -- fill_cw not in the picture * B) 1 | 0 -- we were doing a fill-cw but now are not * C) 1 | 1 -- all rates from fill_cw * D) 0 | 1 -- we were doing non-fill and now we are filling * * For case A, C and D we don't allow a drop. But for * case B where we now our on our steady rate we do * allow a drop. * */ if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0))) goto done_w_hdwr; } if ((rate_wanted > rack->r_ctl.crte->rate) || (rate_wanted <= rack->r_ctl.crte_prev_rate)) { if (rack_hw_rate_to_low && (bw_est < rack_hw_rate_to_low)) { /* * The pacing rate is too low for hardware, but * do allow hardware pacing to be restarted. */ rack_log_hdwr_pacing(rack, bw_est, rack->r_ctl.crte->rate, __LINE__, 0, 5); tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp); rack->r_ctl.crte = NULL; rack->rack_attempt_hdwr_pace = 0; rack->rack_hdrw_pacing = 0; rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted); goto done_w_hdwr; } 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, &rack->r_ctl.crte_prev_rate); if (nrte == NULL) { /* Lost the rate */ rack->rack_hdrw_pacing = 0; rack->r_ctl.crte = NULL; rack_log_hdwr_pacing(rack, rate_wanted, 0, __LINE__, err, 1); rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted); counter_u64_add(rack_hw_pace_lost, 1); } else if (nrte != rack->r_ctl.crte) { rack->r_ctl.crte = nrte; rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz, 0, rack->r_ctl.crte, NULL); rack_log_hdwr_pacing(rack, rate_wanted, rack->r_ctl.crte->rate, __LINE__, err, 2); rack->r_ctl.last_hw_bw_req = rate_wanted; } } else { /* We just need to adjust the segment size */ rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted); rack_log_hdwr_pacing(rack, rate_wanted, rack->r_ctl.crte->rate, __LINE__, 0, 4); rack->r_ctl.last_hw_bw_req = rate_wanted; } } } if ((rack->r_ctl.crte != NULL) && (rack->r_ctl.crte->rate == rate_wanted)) { /* * We need to add a extra if the rates * are exactly matched. The idea is * we want the software to make sure the * queue is empty before adding more, this * gives us N MSS extra pace times where * N is our sysctl */ slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots); } done_w_hdwr: 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 = rack->rc_tp->t_srtt; else srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC; /* its in ms convert */ if (srtt < (uint64_t)slot) { rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0); slot = srtt; } } rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0); } if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) { /* * If this rate is seeing enobufs when it * goes to send then either the nic is out * of gas or we are mis-estimating the time * somehow and not letting the queue empty * completely. Lets add to the pacing time. */ int hw_boost_delay; hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult; if (hw_boost_delay > rack_enobuf_hw_max) hw_boost_delay = rack_enobuf_hw_max; else if (hw_boost_delay < rack_enobuf_hw_min) hw_boost_delay = rack_enobuf_hw_min; slot += hw_boost_delay; } 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; } if (tp->t_state >= TCPS_FIN_WAIT_1) { /* * We will get no more data into the SB * this means we need to have the data available * before we start a measurement. */ if (sbavail(&tptosocket(tp)->so_snd) < max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)))) { /* Nope not enough data */ 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 */ struct timeval tv; tp->gput_ts = tcp_get_usecs(&tv); rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv); } /* * 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, 0); 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 = (uint32_t)my_rsm->r_ack_arrival; rack->app_limited_needs_set = 0; } rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)]; 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, 0); 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->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)]; if (my_rsm->r_flags & RACK_ACKED) { /* * Unlikely since its probably what was * just transmitted (but I am paranoid). */ tp->gput_ts = (uint32_t)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). */ struct timeval tv; microuptime(&tv); rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv); } 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, 0); } 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 void rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags, unsigned ipoptlen, int32_t orig_len, int32_t len, int error, int rsm_is_null, int optlen, int line, uint16_t mode) { 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 = tcp_in_hpts(rack->rc_inp); log.u_bbr.flex1 = error; log.u_bbr.flex2 = flags; log.u_bbr.flex3 = rsm_is_null; log.u_bbr.flex4 = ipoptlen; log.u_bbr.flex5 = tp->rcv_numsacks; log.u_bbr.flex6 = rack->r_ctl.rc_agg_early; log.u_bbr.flex7 = optlen; log.u_bbr.flex8 = rack->r_fsb_inited; log.u_bbr.applimited = rack->r_fast_output; log.u_bbr.bw_inuse = rack_get_bw(rack); log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL); log.u_bbr.cwnd_gain = mode; log.u_bbr.pkts_out = orig_len; log.u_bbr.lt_epoch = len; log.u_bbr.delivered = line; 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_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0, len, &log, false, NULL, NULL, 0, &tv); } } static struct mbuf * rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen, struct rack_fast_send_blk *fsb, int32_t seglimit, int32_t segsize, int hw_tls) { #ifdef KERN_TLS struct ktls_session *tls, *ntls; #ifdef INVARIANTS struct mbuf *start; #endif #endif struct mbuf *m, *n, **np, *smb; struct mbuf *top; int32_t off, soff; int32_t len = *plen; int32_t fragsize; int32_t len_cp = 0; uint32_t mlen, frags; soff = off = the_off; smb = m = the_m; np = ⊤ top = NULL; #ifdef KERN_TLS if (hw_tls && (m->m_flags & M_EXTPG)) tls = m->m_epg_tls; else tls = NULL; #ifdef INVARIANTS start = m; #endif #endif while (len > 0) { if (m == NULL) { *plen = 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; 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; 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; } frags += howmany(mlen, fragsize); if (frags == 0) frags++; seglimit -= frags; KASSERT(seglimit > 0, ("%s: seglimit went too low", __func__)); } n = m_get(M_NOWAIT, m->m_type); *np = n; if (n == NULL) goto nospace; n->m_len = mlen; soff += 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); } len -= n->m_len; off = 0; m = m->m_next; np = &n->m_next; if (len || (soff == smb->m_len)) { /* * We have more so we move forward or * we have consumed the entire mbuf and * len has fell to 0. */ soff = 0; smb = m; } } if (fsb != NULL) { fsb->m = smb; fsb->off = soff; if (smb) { /* * Save off the size of the mbuf. We do * this so that we can recognize when it * has been trimmed by sbcut() as acks * come in. */ fsb->o_m_len = smb->m_len; } else { /* * This is the case where the next mbuf went to NULL. This * means with this copy we have sent everything in the sb. * In theory we could clear the fast_output flag, but lets * not since its possible that we could get more added * and acks that call the extend function which would let * us send more. */ fsb->o_m_len = 0; } } return (top); nospace: if (top) m_freem(top); return (NULL); } /* * This is a copy of m_copym(), taking the TSO segment size/limit * constraints into account, and advancing the sndptr as it goes. */ static struct mbuf * rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen, int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff) { struct mbuf *m, *n; int32_t soff; soff = rack->r_ctl.fsb.off; m = rack->r_ctl.fsb.m; if (rack->r_ctl.fsb.o_m_len > m->m_len) { /* * The mbuf had the front of it chopped off by an ack * we need to adjust the soff/off by that difference. */ uint32_t delta; delta = rack->r_ctl.fsb.o_m_len - m->m_len; soff -= delta; } else if (rack->r_ctl.fsb.o_m_len < m->m_len) { /* * The mbuf was expanded probably by * a m_compress. Just update o_m_len. */ rack->r_ctl.fsb.o_m_len = m->m_len; } KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff)); KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen)); KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?", __FUNCTION__, rack, *plen, m, m->m_len)); /* Save off the right location before we copy and advance */ *s_soff = soff; *s_mb = rack->r_ctl.fsb.m; n = rack_fo_base_copym(m, soff, plen, &rack->r_ctl.fsb, seglimit, segsize, rack->r_ctl.fsb.hw_tls); return (n); } static int rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp) { /* * Enter the fast retransmit path. We are given that a sched_pin is * in place (if accounting is compliled in) and the cycle count taken * at the entry is in the ts_val. The concept her is that the rsm * now holds the mbuf offsets and such so we can directly transmit * without a lot of overhead, the len field is already set for * us to prohibit us from sending too much (usually its 1MSS). */ struct ip *ip = NULL; struct udphdr *udp = NULL; struct tcphdr *th = NULL; struct mbuf *m = NULL; struct inpcb *inp; uint8_t *cpto; struct tcp_log_buffer *lgb; #ifdef TCP_ACCOUNTING uint64_t crtsc; int cnt_thru = 1; #endif struct tcpopt to; u_char opt[TCP_MAXOLEN]; uint32_t hdrlen, optlen; int32_t slot, segsiz, max_val, tso = 0, error, ulen = 0; uint16_t flags; uint32_t if_hw_tsomaxsegcount = 0, startseq; uint32_t if_hw_tsomaxsegsize; #ifdef INET6 struct ip6_hdr *ip6 = NULL; if (rack->r_is_v6) { ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr; hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); } else #endif /* INET6 */ { ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr; hdrlen = sizeof(struct tcpiphdr); } if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) { goto failed; } if (doing_tlp) { /* Its a TLP add the flag, it may already be there but be sure */ rsm->r_flags |= RACK_TLP; } else { /* If it was a TLP it is not not on this retransmit */ rsm->r_flags &= ~RACK_TLP; } startseq = rsm->r_start; segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); inp = rack->rc_inp; to.to_flags = 0; flags = tcp_outflags[tp->t_state]; if (flags & (TH_SYN|TH_RST)) { goto failed; } if (rsm->r_flags & RACK_HAS_FIN) { /* We can't send a FIN here */ goto failed; } if (flags & TH_FIN) { /* We never send a FIN */ flags &= ~TH_FIN; } if (tp->t_flags & TF_RCVD_TSTMP) { to.to_tsval = ms_cts + tp->ts_offset; to.to_tsecr = tp->ts_recent; to.to_flags = TOF_TS; } optlen = tcp_addoptions(&to, opt); hdrlen += optlen; udp = rack->r_ctl.fsb.udp; if (udp) hdrlen += sizeof(struct udphdr); 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; if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && (len > segsiz) && (tp->t_port == 0)) tso = 1; #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) goto failed; m->m_data += max_linkhdr; m->m_len = hdrlen; th = rack->r_ctl.fsb.th; /* Establish the len to send */ if (len > max_val) len = max_val; if ((tso) && (len + optlen > tp->t_maxseg)) { uint32_t if_hw_tsomax; 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; /* * 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) { goto failed; } else if (len > max_len) { len = max_len; } } if (len <= segsiz) { /* * In case there are too many small fragments don't * use TSO: */ tso = 0; } } else { tso = 0; } if ((tso == 0) && (len > segsiz)) len = segsiz; if ((len == 0) || (len <= MHLEN - hdrlen - max_linkhdr)) { goto failed; } th->th_seq = htonl(rsm->r_start); th->th_ack = htonl(tp->rcv_nxt); /* * The PUSH bit should only be applied * if the full retransmission is made. If * we are sending less than this is the * left hand edge and should not have * the PUSH bit. */ if ((rsm->r_flags & RACK_HAD_PUSH) && (len == (rsm->r_end - rsm->r_start))) flags |= TH_PUSH; th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale)); if (th->th_win == 0) { tp->t_sndzerowin++; tp->t_flags |= TF_RXWIN0SENT; } else tp->t_flags &= ~TF_RXWIN0SENT; if (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 if (rsm->m == NULL) goto failed; if (rsm->orig_m_len != rsm->m->m_len) { /* Fix up the orig_m_len and possibly the mbuf offset */ rack_adjust_orig_mlen(rsm); } m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls); if (len <= segsiz) { /* * 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) || (len <= 0)){ goto failed; } if (udp) { if (rack->r_is_v6) ulen = hdrlen + len - sizeof(struct ip6_hdr); else ulen = hdrlen + len - sizeof(struct ip); udp->uh_ulen = htons(ulen); } m->m_pkthdr.rcvif = (struct ifnet *)0; if (TCPS_HAVERCVDSYN(tp->t_state) && (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) { int ect = tcp_ecn_output_established(tp, &flags, len, true); if ((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_flags2 & TF2_ECN_SND_ECE)) tp->t_flags2 &= ~TF2_ECN_SND_ECE; #ifdef INET6 if (rack->r_is_v6) { ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20); ip6->ip6_flow |= htonl(ect << 20); } else #endif { ip->ip_tos &= ~IPTOS_ECN_MASK; ip->ip_tos |= ect; } } tcp_set_flags(th, flags); m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ #ifdef INET6 if (rack->r_is_v6) { 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 if (tso) { KASSERT(len > tp->t_maxseg - optlen, ("%s: len <= tso_segsz tp:%p", __func__, tp)); m->m_pkthdr.csum_flags |= CSUM_TSO; m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; } #ifdef INET6 if (rack->r_is_v6) { ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit; 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; } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { ip->ip_len = htons(m->m_pkthdr.len); ip->ip_ttl = rack->r_ctl.fsb.hoplimit; 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; } } #endif /* Time to copy in our header */ cpto = mtod(m, uint8_t *); memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len); th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr)); if (optlen) { bcopy(opt, th + 1, optlen); th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; } else { th->th_off = sizeof(struct tcphdr) >> 2; } if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; if (rsm->r_flags & RACK_RWND_COLLAPSED) { rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm); counter_u64_add(rack_collapsed_win_rxt, 1); counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start)); } memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); 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 = max_val; 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 (doing_tlp == 0) log.u_bbr.flex8 = 1; else log.u_bbr.flex8 = 2; log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL); log.u_bbr.flex7 = 55; log.u_bbr.pkts_out = tp->t_maxseg; log.u_bbr.timeStamp = cts; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use; log.u_bbr.delivered = 0; lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK, len, &log, false, NULL, NULL, 0, tv); } else lgb = NULL; #ifdef INET6 if (rack->r_is_v6) { error = ip6_output(m, NULL, &inp->inp_route6, 0, NULL, NULL, inp); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { error = ip_output(m, NULL, &inp->inp_route, 0, 0, inp); } #endif m = NULL; if (lgb) { lgb->tlb_errno = error; lgb = NULL; } if (error) { goto failed; } rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv), rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls); if (doing_tlp && (rack->fast_rsm_hack == 0)) { rack->rc_tlp_in_progress = 1; rack->r_ctl.rc_tlp_cnt_out++; } if (error == 0) { tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls); if (doing_tlp) { rack->rc_last_sent_tlp_past_cumack = 0; rack->rc_last_sent_tlp_seq_valid = 1; rack->r_ctl.last_sent_tlp_seq = rsm->r_start; rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start; } } tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); rack->forced_ack = 0; /* If we send something zap the FA flag */ if (IN_FASTRECOVERY(tp->t_flags) && rsm) rack->r_ctl.retran_during_recovery += len; { 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 (tp->t_rtttime == 0) { tp->t_rtttime = ticks; tp->t_rtseq = startseq; KMOD_TCPSTAT_INC(tcps_segstimed); } counter_u64_add(rack_fto_rsm_send, 1); if (error && (error == ENOBUFS)) { if (rack->r_ctl.crte != NULL) { rack_trace_point(rack, RACK_TP_HWENOBUF); } else rack_trace_point(rack, RACK_TP_ENOBUF); slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC); if (rack->rc_enobuf < 0x7f) rack->rc_enobuf++; if (slot < (10 * HPTS_USEC_IN_MSEC)) slot = 10 * HPTS_USEC_IN_MSEC; } else slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz); 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 overridden to use the old 1ms pacing. */ slot = rack->r_ctl.rc_min_to; } rack_start_hpts_timer(rack, tp, cts, slot, len, 0); #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru; } counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz); } counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz)); sched_unpin(); #endif return (0); failed: if (m) m_free(m); return (-1); } static void rack_sndbuf_autoscale(struct tcp_rack *rack) { /* * 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. * * Note that the rack version moves things much faster since * we want to avoid hitting cache lines in the rack_fast_output() * path so this is called much less often and thus moves * the SB forward by a percentage. */ struct socket *so; struct tcpcb *tp; uint32_t sendwin, scaleup; tp = rack->rc_tp; so = rack->rc_inp->inp_socket; sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd); if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat && sbused(&so->so_snd) >= (so->so_snd.sb_hiwat / 8 * 7) && sbused(&so->so_snd) < V_tcp_autosndbuf_max && sendwin >= (sbused(&so->so_snd) - (tp->snd_nxt - tp->snd_una))) { if (rack_autosndbuf_inc) scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100; else scaleup = V_tcp_autosndbuf_inc; if (scaleup < V_tcp_autosndbuf_inc) scaleup = V_tcp_autosndbuf_inc; scaleup += so->so_snd.sb_hiwat; if (scaleup > V_tcp_autosndbuf_max) scaleup = V_tcp_autosndbuf_max; if (!sbreserve_locked(so, SO_SND, scaleup, curthread)) so->so_snd.sb_flags &= ~SB_AUTOSIZE; } } } static int rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err) { /* * Enter to do fast output. We are given that the sched_pin is * in place (if accounting is compiled in) and the cycle count taken * at entry is in place in ts_val. The idea here is that * we know how many more bytes needs to be sent (presumably either * during pacing or to fill the cwnd and that was greater than * the max-burst). We have how much to send and all the info we * need to just send. */ struct ip *ip = NULL; struct udphdr *udp = NULL; struct tcphdr *th = NULL; struct mbuf *m, *s_mb; struct inpcb *inp; uint8_t *cpto; struct tcp_log_buffer *lgb; #ifdef TCP_ACCOUNTING uint64_t crtsc; #endif struct tcpopt to; u_char opt[TCP_MAXOLEN]; uint32_t hdrlen, optlen; #ifdef TCP_ACCOUNTING int cnt_thru = 1; #endif int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0; uint16_t flags; uint32_t s_soff; uint32_t if_hw_tsomaxsegcount = 0, startseq; uint32_t if_hw_tsomaxsegsize; uint16_t add_flag = RACK_SENT_FP; #ifdef INET6 struct ip6_hdr *ip6 = NULL; if (rack->r_is_v6) { ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr; hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); } else #endif /* INET6 */ { ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr; hdrlen = sizeof(struct tcpiphdr); } if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) { m = NULL; goto failed; } startseq = tp->snd_max; segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); inp = rack->rc_inp; len = rack->r_ctl.fsb.left_to_send; to.to_flags = 0; flags = rack->r_ctl.fsb.tcp_flags; if (tp->t_flags & TF_RCVD_TSTMP) { to.to_tsval = ms_cts + tp->ts_offset; to.to_tsecr = tp->ts_recent; to.to_flags = TOF_TS; } optlen = tcp_addoptions(&to, opt); hdrlen += optlen; udp = rack->r_ctl.fsb.udp; if (udp) hdrlen += sizeof(struct udphdr); 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; if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && (len > segsiz) && (tp->t_port == 0)) tso = 1; again: #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) goto failed; m->m_data += max_linkhdr; m->m_len = hdrlen; th = rack->r_ctl.fsb.th; /* Establish the len to send */ if (len > max_val) len = max_val; if ((tso) && (len + optlen > tp->t_maxseg)) { uint32_t if_hw_tsomax; 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; /* * 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) { goto failed; } else if (len > max_len) { len = max_len; } } if (len <= segsiz) { /* * In case there are too many small fragments don't * use TSO: */ tso = 0; } } else { tso = 0; } if ((tso == 0) && (len > segsiz)) len = segsiz; if ((len == 0) || (len <= MHLEN - hdrlen - max_linkhdr)) { goto failed; } sb_offset = tp->snd_max - tp->snd_una; th->th_seq = htonl(tp->snd_max); th->th_ack = htonl(tp->rcv_nxt); th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale)); 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 */ 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 (rack->r_ctl.fsb.m == NULL) goto failed; /* s_mb and s_soff are saved for rack_log_output */ m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, &s_mb, &s_soff); if (len <= segsiz) { /* * 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 (rack->r_ctl.fsb.rfo_apply_push && (len == rack->r_ctl.fsb.left_to_send)) { flags |= TH_PUSH; add_flag |= RACK_HAD_PUSH; } if ((m->m_next == NULL) || (len <= 0)){ goto failed; } if (udp) { if (rack->r_is_v6) ulen = hdrlen + len - sizeof(struct ip6_hdr); else ulen = hdrlen + len - sizeof(struct ip); udp->uh_ulen = htons(ulen); } m->m_pkthdr.rcvif = (struct ifnet *)0; if (TCPS_HAVERCVDSYN(tp->t_state) && (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) { int ect = tcp_ecn_output_established(tp, &flags, len, false); if ((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_flags2 & TF2_ECN_SND_ECE)) tp->t_flags2 &= ~TF2_ECN_SND_ECE; #ifdef INET6 if (rack->r_is_v6) { ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20); ip6->ip6_flow |= htonl(ect << 20); } else #endif { ip->ip_tos &= ~IPTOS_ECN_MASK; ip->ip_tos |= ect; } } tcp_set_flags(th, flags); m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ #ifdef INET6 if (rack->r_is_v6) { 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 if (tso) { KASSERT(len > tp->t_maxseg - optlen, ("%s: len <= tso_segsz tp:%p", __func__, tp)); m->m_pkthdr.csum_flags |= CSUM_TSO; m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; } #ifdef INET6 if (rack->r_is_v6) { ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit; 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; } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { ip->ip_len = htons(m->m_pkthdr.len); ip->ip_ttl = rack->r_ctl.fsb.hoplimit; 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; } } #endif /* Time to copy in our header */ cpto = mtod(m, uint8_t *); memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len); th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr)); if (optlen) { bcopy(opt, th + 1, optlen); th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; } else { th->th_off = sizeof(struct tcphdr) >> 2; } if (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 = tcp_in_hpts(rack->rc_inp); 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 = max_val; 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); log.u_bbr.flex8 = 0; log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL); log.u_bbr.flex7 = 44; log.u_bbr.pkts_out = tp->t_maxseg; log.u_bbr.timeStamp = cts; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use; log.u_bbr.delivered = 0; lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK, len, &log, false, NULL, NULL, 0, tv); } else lgb = NULL; #ifdef INET6 if (rack->r_is_v6) { error = ip6_output(m, NULL, &inp->inp_route6, 0, NULL, NULL, inp); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { error = ip_output(m, NULL, &inp->inp_route, 0, 0, inp); } #endif if (lgb) { lgb->tlb_errno = error; lgb = NULL; } if (error) { *send_err = error; m = NULL; goto failed; } rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv), NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls); m = NULL; if (tp->snd_una == tp->snd_max) { rack->r_ctl.rc_tlp_rxt_last_time = cts; rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__); tp->t_acktime = ticks; } if (error == 0) tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls); rack->forced_ack = 0; /* If we send something zap the FA flag */ tot_len += len; if ((tp->t_flags & TF_GPUTINPROG) == 0) rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset); tp->snd_max += len; tp->snd_nxt = tp->snd_max; { 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 (len <= rack->r_ctl.fsb.left_to_send) rack->r_ctl.fsb.left_to_send -= len; else rack->r_ctl.fsb.left_to_send = 0; if (rack->r_ctl.fsb.left_to_send < segsiz) { rack->r_fast_output = 0; rack->r_ctl.fsb.left_to_send = 0; /* At the end of fast_output scale up the sb */ SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd); rack_sndbuf_autoscale(rack); SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd); } if (tp->t_rtttime == 0) { tp->t_rtttime = ticks; tp->t_rtseq = startseq; KMOD_TCPSTAT_INC(tcps_segstimed); } if ((rack->r_ctl.fsb.left_to_send >= segsiz) && (max_val > len) && (tso == 0)) { max_val -= len; len = segsiz; th = rack->r_ctl.fsb.th; #ifdef TCP_ACCOUNTING cnt_thru++; #endif goto again; } tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); counter_u64_add(rack_fto_send, 1); slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz); rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0); #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru; } counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz); } counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz)); sched_unpin(); #endif return (0); failed: if (m) m_free(m); rack->r_fast_output = 0; return (-1); } static struct rack_sendmap * rack_check_collapsed(struct tcp_rack *rack, uint32_t cts) { struct rack_sendmap *rsm = NULL; struct rack_sendmap fe; int thresh; restart: fe.r_start = rack->r_ctl.last_collapse_point; rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) { /* Nothing, strange turn off validity */ rack->r_collapse_point_valid = 0; return (NULL); } /* Can we send it yet? */ if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) { /* * Receiver window has not grown enough for * the segment to be put on the wire. */ return (NULL); } if (rsm->r_flags & RACK_ACKED) { /* * It has been sacked, lets move to the * next one if possible. */ rack->r_ctl.last_collapse_point = rsm->r_end; /* Are we done? */ if (SEQ_GEQ(rack->r_ctl.last_collapse_point, rack->r_ctl.high_collapse_point)) { rack->r_collapse_point_valid = 0; return (NULL); } goto restart; } /* Now has it been long enough ? */ thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts); if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) { rack_log_collapse(rack, rsm->r_start, (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])), thresh, __LINE__, 6, rsm->r_flags, rsm); return (rsm); } /* Not enough time */ rack_log_collapse(rack, rsm->r_start, (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])), thresh, __LINE__, 7, rsm->r_flags, rsm); return (NULL); } static int rack_output(struct tcpcb *tp) { struct socket *so; uint32_t recwin; uint32_t sb_offset, s_moff = 0; int32_t len, error = 0; uint16_t flags; struct mbuf *m, *s_mb = NULL; struct mbuf *mb; uint32_t if_hw_tsomaxsegcount = 0; uint32_t if_hw_tsomaxsegsize; int32_t segsiz, minseg; long tot_len_this_send = 0; #ifdef INET struct ip *ip = 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, ms_cts, delayed, early; uint16_t add_flag = RACK_SENT_SP; /* The doing_tlp flag will be set by the actual rack_timeout_tlp() */ uint8_t hpts_calling, doing_tlp = 0; uint32_t cwnd_to_use, pace_max_seg; int32_t do_a_prefetch = 0; int32_t prefetch_rsm = 0; int32_t orig_len = 0; struct timeval tv; int32_t prefetch_so_done = 0; struct tcp_log_buffer *lgb; struct inpcb *inp = tptoinpcb(tp); struct sockbuf *sb; uint64_t ts_val = 0; #ifdef TCP_ACCOUNTING uint64_t crtsc; #endif #ifdef INET6 struct ip6_hdr *ip6 = NULL; int32_t isipv6; #endif bool hw_tls = false; NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(inp); /* setup and take the cache hits here */ rack = (struct tcp_rack *)tp->t_fb_ptr; #ifdef TCP_ACCOUNTING sched_pin(); ts_val = get_cyclecount(); #endif hpts_calling = inp->inp_hpts_calls; #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) { #ifdef TCP_ACCOUNTING sched_unpin(); #endif 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 */ #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (0); } #ifdef INET6 if (rack->r_state) { /* Use the cache line loaded if possible */ isipv6 = rack->r_is_v6; } else { isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0; } #endif early = 0; cts = tcp_get_usecs(&tv); ms_cts = tcp_tv_to_mssectick(&tv); if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && tcp_in_hpts(rack->rc_inp)) { /* * 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(cts, rack->r_ctl.rc_last_output_to)) { /* We are delayed */ delayed = 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) { int retval; retval = rack_process_timers(tp, rack, cts, hpts_calling, &doing_tlp); if (retval != 0) { counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1); #ifdef TCP_ACCOUNTING sched_unpin(); #endif /* * If timers want tcp_drop(), then pass error out, * otherwise suppress it. */ return (retval < 0 ? retval : 0); } } if (rack->rc_in_persist) { if (tcp_in_hpts(rack->rc_inp) == 0) { /* Timer is not running */ rack_start_hpts_timer(rack, tp, cts, 0, 0, 0); } #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (0); } if ((rack->rc_ack_required == 1) && (rack->r_timer_override == 0)){ /* A timeout occurred and no ack has arrived */ if (tcp_in_hpts(rack->rc_inp) == 0) { /* Timer is not running */ rack_start_hpts_timer(rack, tp, cts, 0, 0, 0); } #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (0); } if ((rack->r_timer_override) || (rack->rc_ack_can_sendout_data) || (delayed) || (tp->t_state < TCPS_ESTABLISHED)) { rack->rc_ack_can_sendout_data = 0; if (tcp_in_hpts(rack->rc_inp)) tcp_hpts_remove(rack->rc_inp); } else if (tcp_in_hpts(rack->rc_inp)) { /* * On the hpts you can't pass even if ACKNOW is on, we will * when the hpts fires. */ #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_BLOCKED]++; } counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1); sched_unpin(); #endif counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1); return (0); } rack->rc_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, cts)) { early = rack->r_ctl.rc_last_output_to - 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; if ((tp->t_state != rack->r_state) && TCPS_HAVEESTABLISHED(tp->t_state)) { rack_set_state(tp, rack); } if ((rack->r_fast_output) && (doing_tlp == 0) && (tp->rcv_numsacks == 0)) { int ret; error = 0; ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error); if (ret >= 0) return(ret); else if (error) { inp = rack->rc_inp; so = inp->inp_socket; sb = &so->so_snd; goto nomore; } } inp = rack->rc_inp; /* * 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; so = inp->inp_socket; sb = &so->so_snd; 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_2_USEC(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(cts, rack->r_ctl.rc_went_idle_time)) { idle = 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 = 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, cts); } } idle = 0; } if (rack_use_fsb && (rack->r_fsb_inited == 0) && (rack->r_state != TCPS_CLOSED)) rack_init_fsb_block(tp, rack); 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; cts = tcp_get_usecs(&tv); ms_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; if (rack->r_ctl.rc_pace_max_segs == 0) pace_max_seg = rack->rc_user_set_max_segs * segsiz; else pace_max_seg = rack->r_ctl.rc_pace_max_segs; sb_offset = tp->snd_max - tp->snd_una; cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd; 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); so = inp->inp_socket; sb = &so->so_snd; 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(&inp->inp_socket->so_snd); so = inp->inp_socket; sb = &so->so_snd; goto send; } if (rack->r_ctl.rc_resend) { /* Retransmit timer */ rsm = rack->r_ctl.rc_resend; rack->r_ctl.rc_resend = NULL; 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->r_collapse_point_valid && ((rsm = rack_check_collapsed(rack, cts)) != NULL)) { /* * If an RSM is returned then enough time has passed * for us to retransmit it. Move up the collapse point, * since this rsm has its chance to retransmit now. */ rack_trace_point(rack, RACK_TP_COLLAPSED_RXT); rack->r_ctl.last_collapse_point = rsm->r_end; /* Are we done? */ if (SEQ_GEQ(rack->r_ctl.last_collapse_point, rack->r_ctl.high_collapse_point)) rack->r_collapse_point_valid = 0; sack_rxmit = 1; /* We are not doing a TLP */ doing_tlp = 0; len = rsm->r_end - rsm->r_start; sb_offset = rsm->r_start - tp->snd_una; sendalot = 0; if ((rack->full_size_rxt == 0) && (rack->shape_rxt_to_pacing_min == 0) && (len >= segsiz)) len = segsiz; } else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) { /* We have a retransmit that takes precedence */ if ((!IN_FASTRECOVERY(tp->t_flags)) && ((rsm->r_flags & RACK_MUST_RXT) == 0) && ((tp->t_flags & TF_WASFRECOVERY) == 0)) { /* Enter recovery if not induced by a time-out */ rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__); } #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; sendalot = 0; if (len >= segsiz) len = segsiz; if (len > 0) { sack_rxmit = 1; KMOD_TCPSTAT_INC(tcps_sack_rexmits); KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes, min(len, segsiz)); } } else if (rack->r_ctl.rc_tlpsend) { /* Tail loss probe */ long cwin; long tlen; /* * 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; /* We are doing a TLP make sure the flag is preent */ 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; } if (rack->r_must_retran && (doing_tlp == 0) && (SEQ_GT(tp->snd_max, tp->snd_una)) && (rsm == NULL)) { /* * There are two different ways that we * can get into this block: * a) This is a non-sack connection, we had a time-out * and thus r_must_retran was set and everything * left outstanding as been marked for retransmit. * b) The MTU of the path shrank, so that everything * was marked to be retransmitted with the smaller * mtu and r_must_retran was set. * * This means that we expect the sendmap (outstanding) * to all be marked must. We can use the tmap to * look at them. * */ int sendwin, flight; sendwin = min(tp->snd_wnd, tp->snd_cwnd); flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto); if (flight >= sendwin) { /* * We can't send yet. */ so = inp->inp_socket; sb = &so->so_snd; goto just_return_nolock; } /* * This is the case a/b mentioned above. All * outstanding/not-acked should be marked. * We can use the tmap to find them. */ rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rsm == NULL) { /* TSNH */ rack->r_must_retran = 0; rack->r_ctl.rc_out_at_rto = 0; so = inp->inp_socket; sb = &so->so_snd; goto just_return_nolock; } if ((rsm->r_flags & RACK_MUST_RXT) == 0) { /* * The first one does not have the flag, did we collapse * further up in our list? */ rack->r_must_retran = 0; rack->r_ctl.rc_out_at_rto = 0; rsm = NULL; sack_rxmit = 0; } else { sack_rxmit = 1; len = rsm->r_end - rsm->r_start; sb_offset = rsm->r_start - tp->snd_una; sendalot = 0; if ((rack->full_size_rxt == 0) && (rack->shape_rxt_to_pacing_min == 0) && (len >= segsiz)) len = segsiz; /* * Delay removing the flag RACK_MUST_RXT so * that the fastpath for retransmit will * work with this rsm. */ } } /* * 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); } so = inp->inp_socket; sb = &so->so_snd; 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; } } if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo && ((rsm->r_flags & RACK_HAS_FIN) == 0)) { int ret; ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp); if (ret == 0) return (0); } so = inp->inp_socket; sb = &so->so_snd; if (do_a_prefetch == 0) { kern_prefetch(sb, &do_a_prefetch); do_a_prefetch = 1; } #ifdef NETFLIX_SHARED_CWND if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) && rack->rack_enable_scwnd) { /* We are doing cwnd sharing */ if (rack->gp_ready && (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) && sbavail(&so->so_snd)) { /* 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 /* * 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 + sb_offset) > tp->snd_wnd) { if (tp->snd_wnd > sb_offset) len = tp->snd_wnd - sb_offset; else len = 0; } else { len = rack->r_ctl.rc_tlp_new_data; } rack->r_ctl.rc_tlp_new_data = 0; } else { len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset); } if ((rack->r_ctl.crte == NULL) && 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 Fast recovery episode, this * is caused by a SACK or 3 dup acks. At this point * we have sent all the retransmissions and we rely * on PRR to dictate what we will send in the form of * new data. */ 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; } } 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; } if ((len > segsiz) && (tcp_dsack_block_exists(tp))) { /* We only send 1 MSS if we have a DSACK block */ add_flag |= RACK_SENT_W_DSACK; len = segsiz; } 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))) { rack_enter_persist(tp, rack, cts); } } else if ((rsm == NULL) && (doing_tlp == 0) && (len < pace_max_seg)) { /* * 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((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); } } 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; } else if ((rack->r_ctl.crte != NULL) && (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) && (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) && (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) && (len < (int)(sbavail(sb) - sb_offset))) { /* * Here we are doing hardware pacing, this is not a TLP, * we are not sending a pace max segment size, there is rwnd * room to send at least N pace_max_seg, the cwnd is greater * than or equal to a full pacing segments plus 4 mss and we have 2 or * more segments in flight and its not the tail of the socket buffer. * * We don't want to send instead we need to get more ack's in to * allow us to send a full pacing segment. Normally, if we are pacing * about the right speed, we should have finished our pacing * send as most of the acks have come back if we are at the * right rate. This is a bit fuzzy since return path delay * can delay the acks, which is why we want to make sure we * have cwnd space to have a bit more than a max pace segments in flight. * * If we have not gotten our acks back we are pacing at too high a * rate delaying will not hurt and will bring our GP estimate down by * injecting the delay. If we don't do this we will send * 2 MSS out in response to the acks being clocked in which * defeats the point of hw-pacing (i.e. to help us get * larger TSO's out). */ len = 0; } } /* len will be >= 0 after this point. */ KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); rack_sndbuf_autoscale(rack); /* * 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. */ 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 #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 __unused; 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 */ rack->r_ctl.fsb.recwin = recwin; slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz); if ((error == 0) && rack_use_rfo && ((flags & (TH_SYN|TH_FIN)) == 0) && (ipoptlen == 0) && (tp->snd_nxt == tp->snd_max) && (tp->rcv_numsacks == 0) && rack->r_fsb_inited && TCPS_HAVEESTABLISHED(tp->t_state) && (rack->r_must_retran == 0) && ((tp->t_flags & TF_NEEDFIN) == 0) && (len > 0) && (orig_len > 0) && (orig_len > len) && ((orig_len - len) >= segsiz) && ((optlen == 0) || ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) { /* We can send at least one more MSS using our fsb */ rack->r_fast_output = 1; rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off); rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len; rack->r_ctl.fsb.tcp_flags = flags; rack->r_ctl.fsb.left_to_send = orig_len - len; if (hw_tls) rack->r_ctl.fsb.hw_tls = 1; else rack->r_ctl.fsb.hw_tls = 0; KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))), ("rack:%p left_to_send:%u sbavail:%u out:%u", rack, rack->r_ctl.fsb.left_to_send, sbavail(sb), (tp->snd_max - tp->snd_una))); if (rack->r_ctl.fsb.left_to_send < segsiz) rack->r_fast_output = 0; else { if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una))) rack->r_ctl.fsb.rfo_apply_push = 1; else rack->r_ctl.fsb.rfo_apply_push = 0; } } else rack->r_fast_output = 0; rack_log_fsb(rack, tp, so, flags, ipoptlen, orig_len, len, 0, 1, optlen, __LINE__, 1); if (SEQ_GT(tp->snd_max, tp->snd_nxt)) tp->snd_nxt = tp->snd_max; } 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; if (IN_FASTRECOVERY(tp->t_flags)) rack->r_ctl.rc_prr_sndcnt = 0; } else if (ctf_outstanding(tp) >= sbavail(sb)) { /* We are limited by whats available -- app limited */ app_limited = CTF_JR_APP_LIMITED; if (IN_FASTRECOVERY(tp->t_flags)) rack->r_ctl.rc_prr_sndcnt = 0; } 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 (IN_FASTRECOVERY(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 (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; /* Adjust the Gput measurement */ if ((tp->t_flags & TF_GPUTINPROG) && SEQ_GT(tp->gput_ack, tp->snd_max)) { tp->gput_ack = tp->snd_max; if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) { /* * There is not enough to measure. */ tp->t_flags &= ~TF_GPUTINPROG; rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/, rack->r_ctl.rc_gp_srtt /*flex1*/, tp->gput_seq, 0, 0, 18, __LINE__, NULL, 0); } else log = 1; } /* Mark the last packet has app limited */ 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, 0); } } /* Check if we need to go into persists or not */ if ((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 #ifdef TCP_ACCOUNTING if (tot_len_this_send > 0) { crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_DATA]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz); } counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz)); } else { crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_LIMITED]++; } counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val)); } sched_unpin(); #endif return (0); send: if (rsm || sack_rxmit) counter_u64_add(rack_nfto_resend, 1); else counter_u64_add(rack_non_fto_send, 1); 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); 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. */ 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 = ms_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 (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). */ 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); #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_FAIL]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val)); sched_unpin(); #endif return (EHOSTUNREACH); } hdrlen += sizeof(struct udphdr); } #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) 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. */ hw_tls = (sb->sb_flags & SB_TLS_IFNET) != 0; 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); s_mb = mb; s_moff = 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 , &s_mb, &s_moff #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; add_flag |= RACK_HAD_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 if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) { #ifdef INET6 if (isipv6) ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr; else #endif /* INET6 */ ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr; th = rack->r_ctl.fsb.th; udp = rack->r_ctl.fsb.udp; if (udp) { #ifdef INET6 if (isipv6) ulen = hdrlen + len - sizeof(struct ip6_hdr); else #endif /* INET6 */ ulen = hdrlen + len - sizeof(struct ip); udp->uh_ulen = htons(ulen); } } else { #ifdef INET6 if (isipv6) { ip6 = mtod(m, struct ip6_hdr *); if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip6 + 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(inp, tp->t_port, ip6, th); } else #endif /* INET6 */ { ip = mtod(m, struct ip *); if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip + 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(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) { flags |= tcp_ecn_output_syn_sent(tp); } /* Also handle parallel SYN for ECN */ if (TCPS_HAVERCVDSYN(tp->t_state) && (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) { int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit); if ((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_flags2 & TF2_ECN_SND_ECE)) tp->t_flags2 &= ~TF2_ECN_SND_ECE; #ifdef INET6 if (isipv6) { ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20); ip6->ip6_flow |= htonl(ect << 20); } else #endif { ip->ip_tos &= ~IPTOS_ECN_MASK; ip->ip_tos |= ect; } } /* * 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))) { th->th_seq = htonl(tp->snd_nxt); rack_seq = tp->snd_nxt; } 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); tcp_set_flags(th, 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 */ /* Now are we using fsb?, if so copy the template data to the mbuf */ if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) { uint8_t *cpto; cpto = mtod(m, uint8_t *); memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len); /* * We have just copied in: * IP/IP6 * * tcphdr (no options) * * We need to grab the correct pointers into the mbuf * for both the tcp header, and possibly the udp header (if tunneling). * We do this by using the offset in the copy buffer and adding it * to the mbuf base pointer (cpto). */ #ifdef INET6 if (isipv6) ip6 = mtod(m, struct ip6_hdr *); else #endif /* INET6 */ ip = mtod(m, struct ip *); th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr)); /* If we have a udp header lets set it into the mbuf as well */ if (udp) udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr)); } #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 if (optlen) { bcopy(opt, th + 1, optlen); th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; } /* * 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 /* We're getting ready to send; log now. */ if (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 = tcp_in_hpts(rack->rc_inp); 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; /* 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); log.u_bbr.flex8 = 0; if (rsm) { if (rsm->r_flags & RACK_RWND_COLLAPSED) { rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm); counter_u64_add(rack_collapsed_win_rxt, 1); counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start)); } if (doing_tlp) log.u_bbr.flex8 = 2; else log.u_bbr.flex8 = 1; } else { if (doing_tlp) log.u_bbr.flex8 = 3; 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.flex7 <<= 8; log.u_bbr.flex7 |= pass; log.u_bbr.pkts_out = tp->t_maxseg; log.u_bbr.timeStamp = cts; 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. */ rack->r_ctl.fsb.hoplimit = 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, #if defined(IPSEC) || defined(IPSEC_SUPPORT) inp->in6p_outputopts, #else NULL, #endif &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 */ rack->r_ctl.fsb.hoplimit = ip->ip_ttl; /* * 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, #if defined(IPSEC) || defined(IPSEC_SUPPORT) inp->inp_options, #else NULL, #endif &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) { tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls); if (rsm && doing_tlp) { rack->rc_last_sent_tlp_past_cumack = 0; rack->rc_last_sent_tlp_seq_valid = 1; rack->r_ctl.last_sent_tlp_seq = rsm->r_start; rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start; } 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) && 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; if (doing_tlp) { /* Make sure the TLP is added */ add_flag |= RACK_TLP; } else if (rsm) { /* If its a resend without TLP then it must not have the flag */ rsm->r_flags &= ~RACK_TLP; } rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error, rack_to_usec_ts(&tv), rsm, add_flag, s_mb, s_moff, hw_tls); if ((error == 0) && (len > 0) && (tp->snd_una == tp->snd_max)) rack->r_ctl.rc_tlp_rxt_last_time = cts; { 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); } /* * If we are doing FO we need to update the mbuf position and subtract * this happens when the peer sends us duplicate information and * we thus want to send a DSACK. * * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO * turned off? If not then we are going to echo multiple DSACK blocks * out (with the TSO), which we should not be doing. */ if (rack->r_fast_output && len) { if (rack->r_ctl.fsb.left_to_send > len) rack->r_ctl.fsb.left_to_send -= len; else rack->r_ctl.fsb.left_to_send = 0; if (rack->r_ctl.fsb.left_to_send < segsiz) rack->r_fast_output = 0; if (rack->r_fast_output) { rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off); rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_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; #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_FAIL]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val)); sched_unpin(); #endif return (error); case ENOBUFS: /* * Pace us right away to retry in a some * time */ if (rack->r_ctl.crte != NULL) { rack_trace_point(rack, RACK_TP_HWENOBUF); } else rack_trace_point(rack, RACK_TP_ENOBUF); slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC); if (rack->rc_enobuf < 0x7f) rack->rc_enobuf++; if (slot < (10 * HPTS_USEC_IN_MSEC)) slot = 10 * HPTS_USEC_IN_MSEC; if (rack->r_ctl.crte != NULL) { counter_u64_add(rack_saw_enobuf_hw, 1); tcp_rl_log_enobuf(rack->r_ctl.crte); } counter_u64_add(rack_saw_enobuf, 1); 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); #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_FAIL]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val)); sched_unpin(); #endif 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); #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_FAIL]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val)); sched_unpin(); #endif return (error); } } else { rack->rc_enobuf = 0; if (IN_FASTRECOVERY(tp->t_flags) && rsm) rack->r_ctl.retran_during_recovery += len; } 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: 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 && (rsm->r_flags & RACK_HAS_SYN) == 0 && 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 overridden to use the old 1ms pacing. */ slot = rack->r_ctl.rc_min_to; } } /* We have sent clear the flag */ rack->r_ent_rec_ns = 0; if (rack->r_must_retran) { if (rsm) { rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start); if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) { /* * We have retransmitted all. */ rack->r_must_retran = 0; rack->r_ctl.rc_out_at_rto = 0; } } else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) { /* * Sending new data will also kill * the loop. */ rack->r_must_retran = 0; rack->r_ctl.rc_out_at_rto = 0; } } rack->r_ctl.fsb.recwin = recwin; if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) && SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) { /* * We hit an RTO and now have past snd_max at the RTO * clear all the WAS flags. */ tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY); } if (slot) { /* set the rack tcb into the slot N */ if ((error == 0) && rack_use_rfo && ((flags & (TH_SYN|TH_FIN)) == 0) && (rsm == NULL) && (tp->snd_nxt == tp->snd_max) && (ipoptlen == 0) && (tp->rcv_numsacks == 0) && rack->r_fsb_inited && TCPS_HAVEESTABLISHED(tp->t_state) && (rack->r_must_retran == 0) && ((tp->t_flags & TF_NEEDFIN) == 0) && (len > 0) && (orig_len > 0) && (orig_len > len) && ((orig_len - len) >= segsiz) && ((optlen == 0) || ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) { /* We can send at least one more MSS using our fsb */ rack->r_fast_output = 1; rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off); rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len; rack->r_ctl.fsb.tcp_flags = flags; rack->r_ctl.fsb.left_to_send = orig_len - len; if (hw_tls) rack->r_ctl.fsb.hw_tls = 1; else rack->r_ctl.fsb.hw_tls = 0; KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))), ("rack:%p left_to_send:%u sbavail:%u out:%u", rack, rack->r_ctl.fsb.left_to_send, sbavail(sb), (tp->snd_max - tp->snd_una))); if (rack->r_ctl.fsb.left_to_send < segsiz) rack->r_fast_output = 0; else { if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una))) rack->r_ctl.fsb.rfo_apply_push = 1; else rack->r_ctl.fsb.rfo_apply_push = 0; } } else rack->r_fast_output = 0; rack_log_fsb(rack, tp, so, flags, ipoptlen, orig_len, len, error, (rsm == NULL), optlen, __LINE__, 2); } else if (sendalot) { int ret; sack_rxmit = 0; if ((error == 0) && rack_use_rfo && ((flags & (TH_SYN|TH_FIN)) == 0) && (rsm == NULL) && (ipoptlen == 0) && (tp->rcv_numsacks == 0) && (tp->snd_nxt == tp->snd_max) && (rack->r_must_retran == 0) && rack->r_fsb_inited && TCPS_HAVEESTABLISHED(tp->t_state) && ((tp->t_flags & TF_NEEDFIN) == 0) && (len > 0) && (orig_len > 0) && (orig_len > len) && ((orig_len - len) >= segsiz) && ((optlen == 0) || ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) { /* we can use fast_output for more */ rack->r_fast_output = 1; rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off); rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len; rack->r_ctl.fsb.tcp_flags = flags; rack->r_ctl.fsb.left_to_send = orig_len - len; if (hw_tls) rack->r_ctl.fsb.hw_tls = 1; else rack->r_ctl.fsb.hw_tls = 0; KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))), ("rack:%p left_to_send:%u sbavail:%u out:%u", rack, rack->r_ctl.fsb.left_to_send, sbavail(sb), (tp->snd_max - tp->snd_una))); if (rack->r_ctl.fsb.left_to_send < segsiz) { rack->r_fast_output = 0; } if (rack->r_fast_output) { if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una))) rack->r_ctl.fsb.rfo_apply_push = 1; else rack->r_ctl.fsb.rfo_apply_push = 0; rack_log_fsb(rack, tp, so, flags, ipoptlen, orig_len, len, error, (rsm == NULL), optlen, __LINE__, 3); error = 0; ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error); if (ret >= 0) return (ret); else if (error) goto nomore; } } goto again; } /* 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; rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0); #ifdef TCP_ACCOUNTING crtsc = get_cyclecount() - ts_val; if (tot_len_this_send) { if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_DATA]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_DATA] += crtsc; } counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz); } counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz)); } else { if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_ACK]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_ACK] += crtsc; } counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc); } sched_unpin(); #endif if (error == ENOBUFS) error = 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__, NULL); 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, 0); } static void rack_mtu_change(struct tcpcb *tp) { /* * The MSS may have changed */ struct tcp_rack *rack; struct rack_sendmap *rsm; rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) { /* * The MTU has changed we need to resend everything * since all we have sent is lost. We first fix * up the mtu though. */ rack_set_pace_segments(tp, rack, __LINE__, NULL); /* We treat this like a full retransmit timeout without the cwnd adjustment */ rack_remxt_tmr(tp); rack->r_fast_output = 0; rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp, rack->r_ctl.rc_sacked); rack->r_ctl.rc_snd_max_at_rto = tp->snd_max; rack->r_must_retran = 1; /* Mark all inflight to needing to be rxt'd */ TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) { rsm->r_flags |= RACK_MUST_RXT; } } sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); /* We don't use snd_nxt to retransmit */ tp->snd_nxt = tp->snd_max; } static int rack_set_profile(struct tcp_rack *rack, int prof) { int err = EINVAL; if (prof == 1) { /* pace_always=1 */ if (rack->rc_always_pace == 0) { if (tcp_can_enable_pacing() == 0) return (EBUSY); } rack->rc_always_pace = 1; if (rack->use_fixed_rate || rack->gp_ready) rack_set_cc_pacing(rack); rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; rack->rack_attempt_hdwr_pace = 0; /* cmpack=1 */ if (rack_use_cmp_acks) rack->r_use_cmp_ack = 1; if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) && rack->r_use_cmp_ack) rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP; /* scwnd=1 */ rack->rack_enable_scwnd = 1; /* dynamic=100 */ rack->rc_gp_dyn_mul = 1; /* gp_inc_ca */ 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; /* gp_inc_rec */ rack->r_ctl.rack_per_of_gp_rec = 90; err = 0; } else if (prof == 3) { /* Same as profile one execept fill_cw becomes 2 (less aggressive set) */ /* pace_always=1 */ if (rack->rc_always_pace == 0) { if (tcp_can_enable_pacing() == 0) return (EBUSY); } rack->rc_always_pace = 1; if (rack->use_fixed_rate || rack->gp_ready) rack_set_cc_pacing(rack); rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; rack->rack_attempt_hdwr_pace = 0; /* cmpack=1 */ if (rack_use_cmp_acks) rack->r_use_cmp_ack = 1; if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) && rack->r_use_cmp_ack) rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP; /* scwnd=1 */ rack->rack_enable_scwnd = 1; /* dynamic=100 */ rack->rc_gp_dyn_mul = 1; /* gp_inc_ca */ 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=2 */ rack->rc_pace_to_cwnd = 1; rack->r_fill_less_agg = 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; /* gp_inc_rec */ rack->r_ctl.rack_per_of_gp_rec = 90; err = 0; } else if (prof == 2) { /* cmpack=1 */ if (rack->rc_always_pace == 0) { if (tcp_can_enable_pacing() == 0) return (EBUSY); } rack->rc_always_pace = 1; if (rack->use_fixed_rate || rack->gp_ready) rack_set_cc_pacing(rack); rack->r_use_cmp_ack = 1; if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state)) rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP; /* pace_always=1 */ rack->rc_inp->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; err = 0; } else if (prof == 0) { /* This changes things back to the default settings */ err = 0; if (rack->rc_always_pace) { tcp_decrement_paced_conn(); rack_undo_cc_pacing(rack); rack->rc_always_pace = 0; } if (rack_pace_every_seg && tcp_can_enable_pacing()) { rack->rc_always_pace = 1; if (rack->use_fixed_rate || rack->gp_ready) rack_set_cc_pacing(rack); } else rack->rc_always_pace = 0; if (rack_dsack_std_based & 0x1) { /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */ rack->rc_rack_tmr_std_based = 1; } if (rack_dsack_std_based & 0x2) { /* Basically this means rack timers are extended based on dsack by up to (2 * srtt) */ rack->rc_rack_use_dsack = 1; } if (rack_use_cmp_acks) rack->r_use_cmp_ack = 1; else rack->r_use_cmp_ack = 0; if (rack_disable_prr) rack->rack_no_prr = 1; else rack->rack_no_prr = 0; if (rack_gp_no_rec_chg) rack->rc_gp_no_rec_chg = 1; else rack->rc_gp_no_rec_chg = 0; if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) { rack->r_mbuf_queue = 1; if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state)) rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP; rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; } else { rack->r_mbuf_queue = 0; rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; } if (rack_enable_shared_cwnd) rack->rack_enable_scwnd = 1; else rack->rack_enable_scwnd = 0; 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->rc_gp_dyn_mul = 0; } rack->r_rr_config = 0; rack->r_ctl.rc_no_push_at_mrtt = 0; rack->rc_pace_to_cwnd = 0; rack->rc_pace_fill_if_rttin_range = 0; rack->rtt_limit_mul = 0; if (rack_enable_hw_pacing) rack->rack_hdw_pace_ena = 1; else rack->rack_hdw_pace_ena = 0; if (rack_disable_prr) rack->rack_no_prr = 1; else rack->rack_no_prr = 0; if (rack_limits_scwnd) rack->r_limit_scw = 1; else rack->r_limit_scw = 0; err = 0; } return (err); } static int rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval) { struct deferred_opt_list *dol; dol = malloc(sizeof(struct deferred_opt_list), M_TCPFSB, M_NOWAIT|M_ZERO); if (dol == NULL) { /* * No space yikes -- fail out.. */ return (0); } dol->optname = sopt_name; dol->optval = loptval; TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next); return (1); } static int rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name, uint32_t optval, uint64_t loptval) { struct epoch_tracker et; struct sockopt sopt; struct cc_newreno_opts opt; struct inpcb *inp = tptoinpcb(tp); uint64_t val; int error = 0; uint16_t ca, ss; switch (sopt_name) { case TCP_RACK_DSACK_OPT: RACK_OPTS_INC(tcp_rack_dsack_opt); if (optval & 0x1) { rack->rc_rack_tmr_std_based = 1; } else { rack->rc_rack_tmr_std_based = 0; } if (optval & 0x2) { rack->rc_rack_use_dsack = 1; } else { rack->rc_rack_use_dsack = 0; } rack_log_dsack_event(rack, 5, __LINE__, 0, 0); break; case TCP_RACK_PACING_BETA: RACK_OPTS_INC(tcp_rack_beta); - if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) { + if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) { /* This only works for newreno. */ error = EINVAL; break; } if (rack->rc_pacing_cc_set) { /* * Set them into the real CC module * whats in the rack pcb is the old values * to be used on restoral/ */ sopt.sopt_dir = SOPT_SET; opt.name = CC_NEWRENO_BETA; opt.val = optval; if (CC_ALGO(tp)->ctl_output != NULL) - error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt); + error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt); else { error = ENOENT; break; } } else { /* * Not pacing yet so set it into our local * rack pcb storage. */ rack->r_ctl.rc_saved_beta.beta = optval; } break; case TCP_RACK_TIMER_SLOP: RACK_OPTS_INC(tcp_rack_timer_slop); rack->r_ctl.timer_slop = optval; if (rack->rc_tp->t_srtt) { /* * If we have an SRTT lets update t_rxtcur * to have the new slop. */ RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); } break; case TCP_RACK_PACING_BETA_ECN: RACK_OPTS_INC(tcp_rack_beta_ecn); - if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) { + if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) { /* This only works for newreno. */ error = EINVAL; break; } if (rack->rc_pacing_cc_set) { /* * Set them into the real CC module * whats in the rack pcb is the old values * to be used on restoral/ */ sopt.sopt_dir = SOPT_SET; opt.name = CC_NEWRENO_BETA_ECN; opt.val = optval; if (CC_ALGO(tp)->ctl_output != NULL) - error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt); + error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt); else error = ENOENT; } else { /* * Not pacing yet so set it into our local * rack pcb storage. */ rack->r_ctl.rc_saved_beta.beta_ecn = optval; rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED; } break; case TCP_DEFER_OPTIONS: RACK_OPTS_INC(tcp_defer_opt); if (optval) { if (rack->gp_ready) { /* Too late */ error = EINVAL; break; } rack->defer_options = 1; } else rack->defer_options = 0; break; case TCP_RACK_MEASURE_CNT: RACK_OPTS_INC(tcp_rack_measure_cnt); if (optval && (optval <= 0xff)) { rack->r_ctl.req_measurements = optval; } else error = EINVAL; break; case TCP_REC_ABC_VAL: RACK_OPTS_INC(tcp_rec_abc_val); if (optval > 0) rack->r_use_labc_for_rec = 1; else rack->r_use_labc_for_rec = 0; break; case TCP_RACK_ABC_VAL: RACK_OPTS_INC(tcp_rack_abc_val); if ((optval > 0) && (optval < 255)) rack->rc_labc = optval; else error = EINVAL; break; case TCP_HDWR_UP_ONLY: RACK_OPTS_INC(tcp_pacing_up_only); if (optval) rack->r_up_only = 1; else rack->r_up_only = 0; break; case TCP_PACING_RATE_CAP: RACK_OPTS_INC(tcp_pacing_rate_cap); rack->r_ctl.bw_rate_cap = loptval; break; case TCP_RACK_PROFILE: RACK_OPTS_INC(tcp_profile); error = rack_set_profile(rack, optval); break; case TCP_USE_CMP_ACKS: RACK_OPTS_INC(tcp_use_cmp_acks); if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) { /* You can't turn it off once its on! */ error = EINVAL; } else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) { rack->r_use_cmp_ack = 1; rack->r_mbuf_queue = 1; inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; } if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state)) inp->inp_flags2 |= INP_MBUF_ACKCMP; 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 > 1) rack->r_fill_less_agg = 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_use_cmp_ack) rack->r_mbuf_queue = 1; else rack->r_mbuf_queue = 0; if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack) inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; else inp->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 if (optval == 1) rack->rack_no_prr = 1; else if (optval == 2) rack->no_prr_addback = 1; else error = EINVAL; 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_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_TLP_REDUCE: /* RACK TLP cwnd reduction (bool) */ RACK_OPTS_INC(tcp_rack_tlp_reduce); rack->r_ctl.rc_tlp_cwnd_reduce = 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) { if (rack->rc_always_pace) { error = EALREADY; break; } else if (tcp_can_enable_pacing()) { rack->rc_always_pace = 1; if (rack->use_fixed_rate || rack->gp_ready) rack_set_cc_pacing(rack); } else { error = ENOSPC; break; } } else { if (rack->rc_always_pace) { tcp_decrement_paced_conn(); rack->rc_always_pace = 0; rack_undo_cc_pacing(rack); } } if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack) inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; else inp->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__, NULL); 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; if (rack->rc_always_pace) rack_set_cc_pacing(rack); 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,0); 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; if (rack->rc_always_pace) rack_set_cc_pacing(rack); 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, 0); 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; if (rack->rc_always_pace) rack_set_cc_pacing(rack); 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, 0); 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, 0); 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, 0); 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, 0); 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_HDWR_RATE_CAP: RACK_OPTS_INC(tcp_hdwr_rate_cap); if (optval) { if (rack->r_rack_hw_rate_caps == 0) rack->r_rack_hw_rate_caps = 1; else error = EALREADY; } else { rack->r_rack_hw_rate_caps = 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->r_ctl.crte != NULL) { rack->rack_hdrw_pacing = 0; rack->rack_attempt_hdwr_pace = 0; tcp_rel_pacing_rate(rack->r_ctl.crte, tp); rack->r_ctl.crte = NULL; } #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_ENABLE_HYSTART: { if (optval) { - tp->ccv->flags |= CCF_HYSTART_ALLOWED; + tp->t_ccv.flags |= CCF_HYSTART_ALLOWED; if (rack_do_hystart > RACK_HYSTART_ON) - tp->ccv->flags |= CCF_HYSTART_CAN_SH_CWND; + tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND; if (rack_do_hystart > RACK_HYSTART_ON_W_SC) - tp->ccv->flags |= CCF_HYSTART_CONS_SSTH; + tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH; } else { - tp->ccv->flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH); + tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH); } } 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_FAST_RSM_HACK: RACK_OPTS_INC(tcp_rack_fastrsm_hack); if (optval) rack->fast_rsm_hack = 1; else rack->fast_rsm_hack = 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_DELACK: RACK_OPTS_INC(tcp_rack_delayed_ack); 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: RACK_OPTS_INC(tcp_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: RACK_OPTS_INC(tcp_data_after_close); if (optval) rack->rc_allow_data_af_clo = 1; else rack->rc_allow_data_af_clo = 0; break; default: break; } #ifdef NETFLIX_STATS tcp_log_socket_option(tp, sopt_name, optval, error); #endif return (error); } static void rack_apply_deferred_options(struct tcp_rack *rack) { struct deferred_opt_list *dol, *sdol; uint32_t s_optval; TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) { TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next); /* Disadvantage of deferal is you loose the error return */ s_optval = (uint32_t)dol->optval; (void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval); free(dol, M_TCPDO); } } static void rack_hw_tls_change(struct tcpcb *tp, int chg) { /* * HW tls state has changed.. fix all * rsm's in flight. */ struct tcp_rack *rack; struct rack_sendmap *rsm; rack = (struct tcp_rack *)tp->t_fb_ptr; RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) { if (chg) rsm->r_hw_tls = 1; else rsm->r_hw_tls = 0; } if (chg) rack->r_ctl.fsb.hw_tls = 1; else rack->r_ctl.fsb.hw_tls = 0; } 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_tcp_mtu_chg = rack_mtu_change, .tfb_pru_options = rack_pru_options, .tfb_hwtls_change = rack_hw_tls_change, .tfb_compute_pipe = rack_compute_pipe, .tfb_flags = TCP_FUNC_OUTPUT_CANDROP, }; /* * 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 inpcb *inp, struct sockopt *sopt) { #ifdef INET6 struct ip6_hdr *ip6; #endif #ifdef INET struct ip *ip; #endif struct tcpcb *tp; struct tcp_rack *rack; uint64_t loptval; int32_t error = 0, optval; tp = intotcpcb(inp); rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack == NULL) { INP_WUNLOCK(inp); return (EINVAL); } #ifdef INET6 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr; #endif #ifdef INET ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr; #endif switch (sopt->sopt_level) { #ifdef INET6 case IPPROTO_IPV6: MPASS(inp->inp_vflag & INP_IPV6PROTO); switch (sopt->sopt_name) { case IPV6_USE_MIN_MTU: tcp6_use_min_mtu(tp); break; case IPV6_TCLASS: /* * The DSCP codepoint has changed, update the fsb. */ ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | (rack->rc_inp->inp_flow & IPV6_FLOWINFO_MASK); break; } INP_WUNLOCK(inp); return (0); #endif #ifdef INET case IPPROTO_IP: switch (sopt->sopt_name) { case IP_TOS: /* * The DSCP codepoint has changed, update the fsb. */ ip->ip_tos = rack->rc_inp->inp_ip_tos; break; case IP_TTL: /* * The TTL has changed, update the fsb. */ ip->ip_ttl = rack->rc_inp->inp_ip_ttl; break; } INP_WUNLOCK(inp); return (0); #endif } switch (sopt->sopt_name) { case TCP_RACK_TLP_REDUCE: /* URL:tlp_reduce */ /* 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 */ case TCP_HDWR_RATE_CAP: /* URL:hdwrcap boolean */ case TCP_PACING_RATE_CAP: /* URL:cap -- used by side-channel */ case TCP_HDWR_UP_ONLY: /* URL:uponly -- hardware pacing boolean */ /* End pacing related */ case TCP_FAST_RSM_HACK: /* URL:frsm_hack */ case TCP_DELACK: /* URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */ 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_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: /* no URL */ 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 */ case TCP_USE_CMP_ACKS: /* URL:cmpack */ case TCP_RACK_ABC_VAL: /* URL:labc */ case TCP_REC_ABC_VAL: /* URL:reclabc */ case TCP_RACK_MEASURE_CNT: /* URL:measurecnt */ case TCP_DEFER_OPTIONS: /* URL:defer */ case TCP_RACK_DSACK_OPT: /* URL:dsack */ case TCP_RACK_PACING_BETA: /* URL:pacing_beta */ case TCP_RACK_PACING_BETA_ECN: /* URL:pacing_beta_ecn */ case TCP_RACK_TIMER_SLOP: /* URL:timer_slop */ case TCP_RACK_ENABLE_HYSTART: /* URL:hystart */ break; default: /* Filter off all unknown options to the base stack */ return (tcp_default_ctloutput(inp, sopt)); break; } INP_WUNLOCK(inp); if (sopt->sopt_name == TCP_PACING_RATE_CAP) { error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval)); /* * We truncate it down to 32 bits for the socket-option trace this * means rates > 34Gbps won't show right, but thats probably ok. */ optval = (uint32_t)loptval; } else { error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); /* Save it in 64 bit form too */ loptval = optval; } if (error) return (error); INP_WLOCK(inp); if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } if (tp->t_fb != &__tcp_rack) { INP_WUNLOCK(inp); return (ENOPROTOOPT); } if (rack->defer_options && (rack->gp_ready == 0) && (sopt->sopt_name != TCP_DEFER_OPTIONS) && (sopt->sopt_name != TCP_RACK_PACING_BETA) && (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) && (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) { /* Options are beind deferred */ if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) { INP_WUNLOCK(inp); return (0); } else { /* No memory to defer, fail */ INP_WUNLOCK(inp); return (ENOMEM); } } error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval); INP_WUNLOCK(inp); return (error); } static void rack_fill_info(struct tcpcb *tp, struct tcp_info *ti) { INP_WLOCK_ASSERT(tptoinpcb(tp)); 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 | TF2_ACE_PERMIT)) ti->tcpi_options |= TCPI_OPT_ECN; if (tp->t_flags & TF_FASTOPEN) ti->tcpi_options |= TCPI_OPT_TFO; /* still kept in ticks is t_rcvtime */ ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick; /* Since we hold everything in precise useconds this is easy */ ti->tcpi_rtt = tp->t_srtt; ti->tcpi_rttvar = tp->t_rttvar; ti->tcpi_rto = tp->t_rxtcur; 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 NETFLIX_STATS ti->tcpi_total_tlp = tp->t_sndtlppack; ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte; memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo)); #endif #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) { ti->tcpi_options |= TCPI_OPT_TOE; tcp_offload_tcp_info(tp, ti); } #endif } static int rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt) { struct tcpcb *tp; struct tcp_rack *rack; int32_t error, optval; uint64_t val, loptval; struct tcp_info ti; /* * 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; tp = intotcpcb(inp); rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack == NULL) { INP_WUNLOCK(inp); return (EINVAL); } switch (sopt->sopt_name) { case TCP_INFO: /* First get the info filled */ rack_fill_info(tp, &ti); /* Fix up the rtt related fields if needed */ INP_WUNLOCK(inp); error = sooptcopyout(sopt, &ti, sizeof ti); return (error); /* * Beta is the congestion control value for NewReno that influences how * much of a backoff happens when loss is detected. It is normally set * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value * when you exit recovery. */ case TCP_RACK_PACING_BETA: - if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) + if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) error = EINVAL; else if (rack->rc_pacing_cc_set == 0) optval = rack->r_ctl.rc_saved_beta.beta; else { /* * Reach out into the CC data and report back what * I have previously set. Yeah it looks hackish but * we don't want to report the saved values. */ - if (tp->ccv->cc_data) - optval = ((struct newreno *)tp->ccv->cc_data)->beta; + if (tp->t_ccv.cc_data) + optval = ((struct newreno *)tp->t_ccv.cc_data)->beta; else error = EINVAL; } break; /* * Beta_ecn is the congestion control value for NewReno that influences how * much of a backoff happens when a ECN mark is detected. It is normally set * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when * you exit recovery. Note that classic ECN has a beta of 50, it is only * ABE Ecn that uses this "less" value, but we do too with pacing :) */ case TCP_RACK_PACING_BETA_ECN: - if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) + if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) error = EINVAL; else if (rack->rc_pacing_cc_set == 0) optval = rack->r_ctl.rc_saved_beta.beta_ecn; else { /* * Reach out into the CC data and report back what * I have previously set. Yeah it looks hackish but * we don't want to report the saved values. */ - if (tp->ccv->cc_data) - optval = ((struct newreno *)tp->ccv->cc_data)->beta_ecn; + if (tp->t_ccv.cc_data) + optval = ((struct newreno *)tp->t_ccv.cc_data)->beta_ecn; else error = EINVAL; } break; case TCP_RACK_DSACK_OPT: optval = 0; if (rack->rc_rack_tmr_std_based) { optval |= 1; } if (rack->rc_rack_use_dsack) { optval |= 2; } break; case TCP_RACK_ENABLE_HYSTART: { - if (tp->ccv->flags & CCF_HYSTART_ALLOWED) { + if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) { optval = RACK_HYSTART_ON; - if (tp->ccv->flags & CCF_HYSTART_CAN_SH_CWND) + if (tp->t_ccv.flags & CCF_HYSTART_CAN_SH_CWND) optval = RACK_HYSTART_ON_W_SC; - if (tp->ccv->flags & CCF_HYSTART_CONS_SSTH) + if (tp->t_ccv.flags & CCF_HYSTART_CONS_SSTH) optval = RACK_HYSTART_ON_W_SC_C; } else { optval = RACK_HYSTART_OFF; } } break; case TCP_FAST_RSM_HACK: optval = rack->fast_rsm_hack; break; case TCP_DEFER_OPTIONS: optval = rack->defer_options; break; case TCP_RACK_MEASURE_CNT: optval = rack->r_ctl.req_measurements; break; case TCP_REC_ABC_VAL: optval = rack->r_use_labc_for_rec; break; case TCP_RACK_ABC_VAL: optval = rack->rc_labc; break; case TCP_HDWR_UP_ONLY: optval= rack->r_up_only; break; case TCP_PACING_RATE_CAP: loptval = rack->r_ctl.bw_rate_cap; break; case TCP_RACK_PROFILE: /* You cannot retrieve a profile, its write only */ error = EINVAL; break; case TCP_USE_CMP_ACKS: optval = rack->r_use_cmp_ack; break; case TCP_RACK_PACE_TO_FILL: optval = rack->rc_pace_to_cwnd; if (optval && rack->r_fill_less_agg) optval++; 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: if (rack->rack_no_prr == 1) optval = 1; else if (rack->no_prr_addback == 1) optval = 2; else optval = 0; 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_TLP_REDUCE: /* RACK TLP cwnd reduction (bool) */ optval = rack->r_ctl.rc_tlp_cwnd_reduce; 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_HDWR_RATE_CAP: optval = rack->r_rack_hw_rate_caps; 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_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; case TCP_RACK_TIMER_SLOP: optval = rack->r_ctl.timer_slop; break; default: return (tcp_default_ctloutput(inp, sopt)); break; } INP_WUNLOCK(inp); if (error == 0) { if (TCP_PACING_RATE_CAP) error = sooptcopyout(sopt, &loptval, sizeof loptval); else error = sooptcopyout(sopt, &optval, sizeof optval); } return (error); } static int rack_ctloutput(struct inpcb *inp, struct sockopt *sopt) { if (sopt->sopt_dir == SOPT_SET) { return (rack_set_sockopt(inp, sopt)); } else if (sopt->sopt_dir == SOPT_GET) { return (rack_get_sockopt(inp, sopt)); } else { panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir); } } 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_stacks/sack_filter.c b/sys/netinet/tcp_stacks/sack_filter.c index 96728f360463..19b32acb2c7f 100644 --- a/sys/netinet/tcp_stacks/sack_filter.c +++ b/sys/netinet/tcp_stacks/sack_filter.c @@ -1,809 +1,811 @@ /*- * Copyright (c) 2017-9 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$"); #ifndef _KERNEL #define _WANT_TCPCB 1 #endif #include #include #include #ifdef _KERNEL #include #include #endif +#include +#include #include #include #include #ifndef _KERNEL #include #include #include #include #include #include #include #endif #include "sack_filter.h" /* * Sack filter is used to filter out sacks * that have already been processed. The idea * is pretty simple really, consider two sacks * * SACK 1 * cum-ack A * sack B - C * SACK 2 * cum-ack A * sack D - E * sack B - C * * The previous sack information (B-C) is repeated * in SACK 2. If the receiver gets SACK 1 and then * SACK 2 then any work associated with B-C as already * been completed. This only effects where we may have * (as in bbr or rack) cases where we walk a linked list. * * Now the utility trys to keep everything in a single * cache line. This means that its not perfect and * it could be that so big of sack's come that a * "remembered" processed sack falls off the list and * so gets re-processed. Thats ok, it just means we * did some extra work. We could of course take more * cache line hits by expanding the size of this * structure, but then that would cost more. */ #ifndef _KERNEL int detailed_dump = 0; uint64_t cnt_skipped_oldsack = 0; uint64_t cnt_used_oldsack = 0; int highest_used=0; int over_written=0; int empty_avail=0; int no_collapse = 0; FILE *out = NULL; FILE *in = NULL; #endif #define sack_blk_used(sf, i) ((1 << i) & sf->sf_bits) #define sack_blk_set(sf, i) ((1 << i) | sf->sf_bits) #define sack_blk_clr(sf, i) (~(1 << i) & sf->sf_bits) #ifndef _KERNEL static #endif void sack_filter_clear(struct sack_filter *sf, tcp_seq seq) { sf->sf_ack = seq; sf->sf_bits = 0; sf->sf_cur = 0; sf->sf_used = 0; } /* * Given a previous sack filter block, filter out * any entries where the cum-ack moves over them * fully or partially. */ static void sack_filter_prune(struct sack_filter *sf, tcp_seq th_ack) { int32_t i; /* start with the oldest */ for (i = 0; i < SACK_FILTER_BLOCKS; i++) { if (sack_blk_used(sf, i)) { if (SEQ_GT(th_ack, sf->sf_blks[i].end)) { /* This block is consumed */ sf->sf_bits = sack_blk_clr(sf, i); sf->sf_used--; } else if (SEQ_GT(th_ack, sf->sf_blks[i].start)) { /* Some of it is acked */ sf->sf_blks[i].start = th_ack; /* We could in theory break here, but * there are some broken implementations * that send multiple blocks. We want * to catch them all with similar seq's. */ } } } sf->sf_ack = th_ack; } /* * Return true if you find that * the sackblock b is on the score * board. Update it along the way * if part of it is on the board. */ static int32_t is_sack_on_board(struct sack_filter *sf, struct sackblk *b) { int32_t i, cnt; for (i = sf->sf_cur, cnt=0; cnt < SACK_FILTER_BLOCKS; cnt++) { if (sack_blk_used(sf, i)) { if (SEQ_LT(b->start, sf->sf_ack)) { /* Behind cum-ack update */ b->start = sf->sf_ack; } if (SEQ_LT(b->end, sf->sf_ack)) { /* End back behind too */ b->end = sf->sf_ack; } if (b->start == b->end) { return(1); } /* Jonathans Rule 1 */ if (SEQ_LEQ(sf->sf_blks[i].start, b->start) && SEQ_GEQ(sf->sf_blks[i].end, b->end)) { /** * Our board has this entirely in * whole or in part: * * board |-------------| * sack |-------------| * * board |-------------| * sack |----| * */ return(1); } /* Jonathans Rule 2 */ if(SEQ_LT(sf->sf_blks[i].end, b->start)) { /** * Not near each other: * * board |---| * sack |---| */ goto nxt_blk; } /* Jonathans Rule 3 */ if (SEQ_GT(sf->sf_blks[i].start, b->end)) { /** * Not near each other: * * board |---| * sack |---| */ goto nxt_blk; } if (SEQ_LEQ(sf->sf_blks[i].start, b->start)) { /** * The board block partial meets: * * board |--------| * sack |----------| * * board |--------| * sack |--------------| * * up with this one (we have part of it). * 1) Update the board block to the new end * and * 2) Update the start of this block to my end. */ b->start = sf->sf_blks[i].end; sf->sf_blks[i].end = b->end; goto nxt_blk; } if (SEQ_GEQ(sf->sf_blks[i].end, b->end)) { /** * The board block partial meets: * * board |--------| * sack |----------| * * board |----| * sack |----------| * 1) Update the board block to the new start * and * 2) Update the start of this block to my end. */ b->end = sf->sf_blks[i].start; sf->sf_blks[i].start = b->start; goto nxt_blk; } } nxt_blk: i++; i %= SACK_FILTER_BLOCKS; } /* Did we totally consume it in pieces? */ if (b->start != b->end) return(0); else return(1); } static int32_t sack_filter_old(struct sack_filter *sf, struct sackblk *in, int numblks) { int32_t num, i; struct sackblk blkboard[TCP_MAX_SACK]; /* * An old sack has arrived. It may contain data * we do not have. We might not have it since * we could have had a lost ack we might have the * entire thing on our current board. We want to prune * off anything we have. With this function though we * won't add to the board. */ for( i = 0, num = 0; isf_blks[i], &sf->sf_blks[idx], sizeof(struct sackblk)); sf->sf_bits = sack_blk_clr(sf, idx); sf->sf_bits = sack_blk_set(sf, i); return; } i++; i %= SACK_FILTER_BLOCKS; } } static int32_t sack_filter_new(struct sack_filter *sf, struct sackblk *in, int numblks, tcp_seq th_ack) { struct sackblk blkboard[TCP_MAX_SACK]; int32_t num, i; /* * First lets trim the old and possibly * throw any away we have. */ for(i=0, num=0; i=0; i--) { if (is_sack_on_board(sf, &blkboard[i])) { continue; } /* Add this guy its not listed */ sf->sf_cur++; sf->sf_cur %= SACK_FILTER_BLOCKS; if ((sack_blk_used(sf, sf->sf_cur)) && (sf->sf_used < SACK_FILTER_BLOCKS)) { sack_move_to_empty(sf, sf->sf_cur); } #ifndef _KERNEL if (sack_blk_used(sf, sf->sf_cur)) { over_written++; if (sf->sf_used < SACK_FILTER_BLOCKS) empty_avail++; } #endif memcpy(&sf->sf_blks[sf->sf_cur], &in[i], sizeof(struct sackblk)); if (sack_blk_used(sf, sf->sf_cur) == 0) { sf->sf_used++; #ifndef _KERNEL if (sf->sf_used > highest_used) highest_used = sf->sf_used; #endif sf->sf_bits = sack_blk_set(sf, sf->sf_cur); } } return(numblks); } /* * Given a sack block on the board (the skip index) see if * any other used entries overlap or meet, if so return the index. */ static int32_t sack_blocks_overlap_or_meet(struct sack_filter *sf, struct sackblk *sb, uint32_t skip) { int32_t i; for(i=0; isf_blks[i].end, sb->start) && SEQ_LEQ(sf->sf_blks[i].end, sb->end) && SEQ_LEQ(sf->sf_blks[i].start, sb->start)) { /** * The two board blocks meet: * * board1 |--------| * board2 |----------| * * board1 |--------| * board2 |--------------| * * board1 |--------| * board2 |--------| */ return(i); } if (SEQ_LEQ(sf->sf_blks[i].start, sb->end) && SEQ_GEQ(sf->sf_blks[i].start, sb->start) && SEQ_GEQ(sf->sf_blks[i].end, sb->end)) { /** * The board block partial meets: * * board |--------| * sack |----------| * * board |----| * sack |----------| * 1) Update the board block to the new start * and * 2) Update the start of this block to my end. */ return(i); } } return (-1); } /* * Collapse entry src into entry into * and free up the src entry afterwards. */ static void sack_collapse(struct sack_filter *sf, int32_t src, int32_t into) { if (SEQ_LT(sf->sf_blks[src].start, sf->sf_blks[into].start)) { /* src has a lower starting point */ sf->sf_blks[into].start = sf->sf_blks[src].start; } if (SEQ_GT(sf->sf_blks[src].end, sf->sf_blks[into].end)) { /* src has a higher ending point */ sf->sf_blks[into].end = sf->sf_blks[src].end; } sf->sf_bits = sack_blk_clr(sf, src); sf->sf_used--; } static void sack_board_collapse(struct sack_filter *sf) { int32_t i, j, i_d, j_d; for(i=0; isf_blks[i], i); if (j == -1) { /* No overlap */ continue; } /* * Ok j and i overlap with each other, collapse the * one out furthest away from the current position. */ if (sf->sf_cur > i) i_d = sf->sf_cur - i; else i_d = i - sf->sf_cur; if (sf->sf_cur > j) j_d = sf->sf_cur - j; else j_d = j - sf->sf_cur; if (j_d > i_d) { sack_collapse(sf, j, i); } else sack_collapse(sf, i, j); } } #ifndef _KERNEL uint64_t saved=0; uint64_t tot_sack_blks=0; static void sack_filter_dump(FILE *out, struct sack_filter *sf) { int i; fprintf(out, " sf_ack:%u sf_bits:0x%x c:%d used:%d\n", sf->sf_ack, sf->sf_bits, sf->sf_cur, sf->sf_used); for(i=0; isf_blks[i].start, sf->sf_blks[i].end); } } } #endif #ifndef _KERNEL static #endif int sack_filter_blks(struct sack_filter *sf, struct sackblk *in, int numblks, tcp_seq th_ack) { int32_t i, ret; if (numblks > TCP_MAX_SACK) { #ifdef _KERNEL panic("sf:%p sb:%p Impossible number of sack blocks %d > 4\n", sf, in, numblks); #endif return(numblks); } #ifndef _KERNEL if ((sf->sf_used > 1) && (no_collapse == 0)) sack_board_collapse(sf); #else if (sf->sf_used > 1) sack_board_collapse(sf); #endif if ((sf->sf_used == 0) && numblks) { /* * We are brand new add the blocks in * reverse order. Note we can see more * than one in new, since ack's could be lost. */ int cnt_added = 0; sf->sf_ack = th_ack; for(i=(numblks-1), sf->sf_cur=0; i >= 0; i--) { memcpy(&sf->sf_blks[sf->sf_cur], &in[i], sizeof(struct sackblk)); sf->sf_bits = sack_blk_set(sf, sf->sf_cur); sf->sf_cur++; sf->sf_cur %= SACK_FILTER_BLOCKS; sf->sf_used++; cnt_added++; #ifndef _KERNEL if (sf->sf_used > highest_used) highest_used = sf->sf_used; #endif } if (sf->sf_cur) sf->sf_cur--; return (cnt_added); } if (SEQ_GT(th_ack, sf->sf_ack)) { sack_filter_prune(sf, th_ack); } if (numblks) { if (SEQ_GEQ(th_ack, sf->sf_ack)) { ret = sack_filter_new(sf, in, numblks, th_ack); } else { ret = sack_filter_old(sf, in, numblks); } } else ret = 0; return (ret); } void sack_filter_reject(struct sack_filter *sf, struct sackblk *in) { /* * Given a specified block (that had made * it past the sack filter). Reject that * block triming it off any sack-filter block * that has it. Usually because the block was * too small and did not cover a whole send. * * This function will only "undo" sack-blocks * that are fresh and touch the edges of * blocks in our filter. */ int i; for(i=0; isf_blks[i].end == in->end) { /* The end moves back to start */ if (SEQ_GT(in->start, sf->sf_blks[i].start)) /* in-blk |----| */ /* sf-blk |---------| */ sf->sf_blks[i].end = in->start; else { /* It consumes this block */ /* in-blk |---------| */ /* sf-blk |------| */ /* */ /* sf-blk |---------| */ sf->sf_bits = sack_blk_clr(sf, i); sf->sf_used--; } continue; } if (sf->sf_blks[i].start == in->start) { if (SEQ_LT(in->end, sf->sf_blks[i].end)) { /* in-blk |----| */ /* sf-blk |---------| */ sf->sf_blks[i].start = in->end; } else { /* It consumes this block */ /* in-blk |----------| */ /* sf-blk |-------| */ /* */ /* sf-blk |----------| */ sf->sf_bits = sack_blk_clr(sf, i); sf->sf_used--; } continue; } } } #ifndef _KERNEL int main(int argc, char **argv) { char buffer[512]; struct sackblk blks[TCP_MAX_SACK]; FILE *err; tcp_seq th_ack, snd_una, snd_max = 0; struct sack_filter sf; int32_t numblks,i; int snd_una_set=0; double a, b, c; int invalid_sack_print = 0; uint32_t chg_remembered=0; uint32_t sack_chg=0; char line_buf[10][256]; int line_buf_at=0; in = stdin; out = stdout; while ((i = getopt(argc, argv, "ndIi:o:?h")) != -1) { switch (i) { case 'n': no_collapse = 1; break; case 'd': detailed_dump = 1; break; case'I': invalid_sack_print = 1; break; case 'i': in = fopen(optarg, "r"); if (in == NULL) { fprintf(stderr, "Fatal error can't open %s for input\n", optarg); exit(-1); } break; case 'o': out = fopen(optarg, "w"); if (out == NULL) { fprintf(stderr, "Fatal error can't open %s for output\n", optarg); exit(-1); } break; default: case '?': case 'h': fprintf(stderr, "Use %s [ -i infile -o outfile -I]\n", argv[0]); return(0); break; }; } sack_filter_clear(&sf, 0); memset(buffer, 0, sizeof(buffer)); memset(blks, 0, sizeof(blks)); numblks = 0; fprintf(out, "************************************\n"); while (fgets(buffer, sizeof(buffer), in) != NULL) { sprintf(line_buf[line_buf_at], "%s", buffer); line_buf_at++; if (strncmp(buffer, "QUIT", 4) == 0) { break; } else if (strncmp(buffer, "DUMP", 4) == 0) { sack_filter_dump(out, &sf); } else if (strncmp(buffer, "MAX:", 4) == 0) { snd_max = strtoul(&buffer[4], NULL, 0); } else if (strncmp(buffer, "COMMIT", 6) == 0) { int nn, ii; if (numblks) { uint32_t szof, tot_chg; for(ii=0; ii chg_remembered)){ fprintf(out,"***WARNING WILL RODGERS DANGER!! sack_chg:%u last:%u\n", sack_chg, chg_remembered ); } sack_chg = chg_remembered = 0; } else if (strncmp(buffer, "RXT", 3) == 0) { sack_filter_clear(&sf, snd_una); } else if (strncmp(buffer, "ACK:", 4) == 0) { th_ack = strtoul(&buffer[4], NULL, 0); if (snd_una_set == 0) { snd_una = th_ack; snd_una_set = 1; } else if (SEQ_GT(th_ack, snd_una)) { snd_una = th_ack; } } else if (strncmp(buffer, "EXIT", 4) == 0) { sack_filter_clear(&sf, snd_una); sack_chg = chg_remembered = 0; } else if (strncmp(buffer, "SACK:", 5) == 0) { char *end=NULL; uint32_t start; uint32_t endv; start = strtoul(&buffer[5], &end, 0); if (end) { endv = strtoul(&end[1], NULL, 0); } else { fprintf(out, "--Sack invalid skip 0 start:%u : ??\n", start); continue; } if (SEQ_GT(endv, snd_max)) snd_max = endv; if (SEQ_LT(endv, start)) { fprintf(out, "--Sack invalid skip 1 endv:%u < start:%u\n", endv, start); continue; } if (numblks == TCP_MAX_SACK) { fprintf(out, "--Exceeded max %d\n", numblks); exit(0); } blks[numblks].start = start; blks[numblks].end = endv; numblks++; } else if (strncmp(buffer, "REJ:n:n", 4) == 0) { struct sackblk in; char *end=NULL; in.start = strtoul(&buffer[4], &end, 0); if (end) { in.end = strtoul(&end[1], NULL, 0); sack_filter_reject(&sf, &in); } else fprintf(out, "Invalid input END:A:B\n"); } else if (strncmp(buffer, "HELP", 4) == 0) { fprintf(out, "You can input:\n"); fprintf(out, "SACK:S:E -- to define a sack block\n"); fprintf(out, "RXT -- to clear the filter without changing the remembered\n"); fprintf(out, "EXIT -- To clear the sack filter and start all fresh\n"); fprintf(out, "ACK:N -- To advance the cum-ack to N\n"); fprintf(out, "MAX:N -- To set send-max to N\n"); fprintf(out, "COMMIT -- To apply the sack you built to the filter and dump the filter\n"); fprintf(out, "DUMP -- To display the current contents of the sack filter\n"); fprintf(out, "QUIT -- To exit this program\n"); } else { fprintf(out, "Command %s unknown\n", buffer); } memset(buffer, 0, sizeof(buffer)); } if (in != stdin) { fclose(in); } if (out != stdout) { fclose(out); } a = saved * 100.0; b = tot_sack_blks * 1.0; if (b > 0.0) c = a/b; else c = 0.0; if (out != stdout) err = stdout; else err = stderr; fprintf(err, "Saved %lu sack blocks out of %lu (%2.3f%%) old_skip:%lu old_usd:%lu high_cnt:%d ow:%d ea:%d\n", saved, tot_sack_blks, c, cnt_skipped_oldsack, cnt_used_oldsack, highest_used, over_written, empty_avail); return(0); } #endif diff --git a/sys/netinet/tcp_subr.c b/sys/netinet/tcp_subr.c index 6e85d5a77340..a132d9cf7e96 100644 --- a/sys/netinet/tcp_subr.c +++ b/sys/netinet/tcp_subr.c @@ -1,4075 +1,4031 @@ /*- * 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 #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 #ifdef INVARIANTS #define TCPSTATES #endif #include #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 #ifdef INET6 #include #endif #include #include #include #include #ifdef INET6 static ip6proto_ctlinput_t tcp6_ctlinput; static udp_tun_icmp_t tcp6_ctlinput_viaudp; #endif 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 uint32_t tcp_ack_war_time_window = 1000; SYSCTL_UINT(_net_inet_tcp, OID_AUTO, ack_war_timewindow, CTLFLAG_RW, &tcp_ack_war_time_window, 1000, "If the tcp_stack does ack-war prevention how many milliseconds are in its time window?"); uint32_t tcp_ack_war_cnt = 5; SYSCTL_UINT(_net_inet_tcp, OID_AUTO, ack_war_cnt, CTLFLAG_RW, &tcp_ack_war_cnt, 5, "If the tcp_stack does ack-war prevention how many acks can be sent in its time window?"); 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"); /* * As of June 2021, several TCP stacks violate RFC 7323 from September 2014. * Some stacks negotiate TS, but never send them after connection setup. Some * stacks negotiate TS, but don't send them when sending keep-alive segments. * These include modern widely deployed TCP stacks. * Therefore tolerating violations for now... */ VNET_DEFINE(int, tcp_tolerate_missing_ts) = 1; 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"); /* How many connections are pacing */ static volatile uint32_t number_of_tcp_connections_pacing = 0; static uint32_t shadow_num_connections = 0; static int tcp_pacing_limit = 10000; SYSCTL_INT(_net_inet_tcp, OID_AUTO, pacing_limit, CTLFLAG_RW, &tcp_pacing_limit, 1000, "If the TCP stack does pacing, is there a limit (-1 = no, 0 = no pacing N = number of connections)"); SYSCTL_UINT(_net_inet_tcp, OID_AUTO, pacing_count, CTLFLAG_RD, &shadow_num_connections, 0, "Number of TCP connections being paced"); 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 struct inpcb *tcp_mtudisc(struct inpcb *, int); static struct inpcb *tcp_drop_syn_sent(struct inpcb *, int); static char * tcp_log_addr(struct in_conninfo *inc, struct tcphdr *th, const void *ip4hdr, const void *ip6hdr); static ipproto_ctlinput_t tcp_ctlinput; static udp_tun_icmp_t tcp_ctlinput_viaudp; static struct tcp_function_block tcp_def_funcblk = { .tfb_tcp_block_name = "freebsd", .tfb_tcp_output = tcp_default_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; void tcp_record_dsack(struct tcpcb *tp, tcp_seq start, tcp_seq end, int tlp) { TCPSTAT_INC(tcps_dsack_count); tp->t_dsack_pack++; if (tlp == 0) { if (SEQ_GT(end, start)) { tp->t_dsack_bytes += (end - start); TCPSTAT_ADD(tcps_dsack_bytes, (end - start)); } else { tp->t_dsack_tlp_bytes += (start - end); TCPSTAT_ADD(tcps_dsack_bytes, (start - end)); } } else { if (SEQ_GT(end, start)) { tp->t_dsack_bytes += (end - start); TCPSTAT_ADD(tcps_dsack_tlp_bytes, (end - start)); } else { tp->t_dsack_tlp_bytes += (start - end); TCPSTAT_ADD(tcps_dsack_tlp_bytes, (start - end)); } } } 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); } /* Find a matching alias for the given tcp_function_block. */ int find_tcp_function_alias(struct tcp_function_block *blk, struct tcp_function_set *fs) { struct tcp_function *f; int found; found = 0; rw_rlock(&tcp_function_lock); TAILQ_FOREACH(f, &t_functions, tf_next) { if ((f->tf_fb == blk) && (strncmp(f->tf_name, blk->tfb_tcp_block_name, TCP_FUNCTION_NAME_LEN_MAX) != 0)) { /* Matching function block with different name. */ strncpy(fs->function_set_name, f->tf_name, TCP_FUNCTION_NAME_LEN_MAX); found = 1; break; } } /* Null terminate the string appropriately. */ if (found) { fs->function_set_name[TCP_FUNCTION_NAME_LEN_MAX - 1] = '\0'; } else { fs->function_set_name[0] = '\0'; } rw_runlock(&tcp_function_lock); return (found); } 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 bool 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 (true); out: m_freem(m); return (true); } 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 writing! */ #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 writing! */ 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 = tptosocket(tp); INP_WLOCK_ASSERT(tptoinpcb(tp)); 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. */ 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(tptoinpcb(tp)); } /* * 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) -INPCBSTORAGE_DEFINE(tcpcbstor, inpcb, "tcpinp", "tcp_inpcb", "tcp", "tcphash"); +INPCBSTORAGE_DEFINE(tcpcbstor, tcpcb, "tcpinp", "tcp_inpcb", "tcp", "tcphash"); /* * 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_iterator inpi = INP_ALL_ITERATOR(&V_tcbinfo, INPLOOKUP_WLOCKPCB); 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); while ((inp = inp_next(&inpi)) != NULL) { tp = intotcpcb(inp); if (tp == NULL || tp->t_fb != blk) continue; tcp_switch_back_to_default(tp); } 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); } static void tcp_drain(void) { struct epoch_tracker et; VNET_ITERATOR_DECL(vnet_iter); if (!do_tcpdrain) return; NET_EPOCH_ENTER(et); VNET_LIST_RLOCK_NOSLEEP(); VNET_FOREACH(vnet_iter) { CURVNET_SET(vnet_iter); struct inpcb_iterator inpi = INP_ALL_ITERATOR(&V_tcbinfo, INPLOOKUP_WLOCKPCB); 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. */ while ((inpb = inp_next(&inpi)) != NULL) { 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 } } CURVNET_RESTORE(); } VNET_LIST_RUNLOCK_NOSLEEP(); NET_EPOCH_EXIT(et); } static void tcp_vnet_init(void *arg __unused) { #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 in_pcbinfo_init(&V_tcbinfo, &tcpcbstor, tcp_tcbhashsize, tcp_tcbhashsize); - /* - * 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"); - 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(); COUNTER_ARRAY_ALLOC(V_tcps_states, TCP_NSTATES, M_WAITOK); VNET_PCPUSTAT_ALLOC(tcpstat, M_WAITOK); V_tcp_msl = TCPTV_MSL; } VNET_SYSINIT(tcp_vnet_init, SI_SUB_PROTO_DOMAIN, SI_ORDER_FOURTH, tcp_vnet_init, NULL); static void tcp_init(void *arg __unused) { const char *tcbhash_tuneable; int hashsize; tcp_reass_global_init(); /* XXX virtualize those below? */ tcp_delacktime = TCPTV_DELACK; tcp_keepinit = TCPTV_KEEP_INIT; tcp_keepidle = TCPTV_KEEP_IDLE; tcp_keepintvl = TCPTV_KEEPINTVL; tcp_maxpersistidle = TCPTV_KEEP_IDLE; 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; /* 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_protosw.pr_soreceive = soreceive_stream; #endif #ifdef INET6 tcp6_protosw.pr_soreceive = soreceive_stream; #endif /* INET6 */ } #ifdef INET6 max_protohdr_grow(sizeof(struct ip6_hdr) + sizeof(struct tcphdr)); #else /* INET6 */ max_protohdr_grow(sizeof(struct tcpiphdr)); #endif /* INET6 */ ISN_LOCK_INIT(); EVENTHANDLER_REGISTER(shutdown_pre_sync, tcp_fini, NULL, SHUTDOWN_PRI_DEFAULT); EVENTHANDLER_REGISTER(vm_lowmem, tcp_drain, NULL, LOWMEM_PRI_DEFAULT); EVENTHANDLER_REGISTER(mbuf_lowmem, tcp_drain, NULL, LOWMEM_PRI_DEFAULT); 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_locks_taken = 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_bad_csums = counter_u64_alloc(M_WAITOK); #ifdef TCPPCAP tcp_pcap_init(); #endif hashsize = TCBHASHSIZE; tcbhash_tuneable = "net.inet.tcp.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) 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); } tcp_tcbhashsize = hashsize; #ifdef INET IPPROTO_REGISTER(IPPROTO_TCP, tcp_input, tcp_ctlinput); #endif #ifdef INET6 IP6PROTO_REGISTER(IPPROTO_TCP, tcp6_input, tcp6_ctlinput); #endif } SYSINIT(tcp_init, SI_SUB_PROTO_DOMAIN, SI_ORDER_THIRD, tcp_init, NULL); #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_INFO_WLOCK(&V_tcbinfo); n = V_tcbinfo.ipi_count; INP_INFO_WUNLOCK(&V_tcbinfo); if (n == 0) break; pause("tcpdes", hz / 10); } tcp_hc_destroy(); syncache_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(); COUNTER_ARRAY_FREE(V_tcps_states, TCP_NSTATES); VNET_PCPUSTAT_FREE(tcpstat); #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, 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); 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; 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_off = 5; tcp_set_flags(th, 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, 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, uint16_t flags) { struct tcpopt to; struct inpcb *inp; struct ip *ip; struct mbuf *optm; struct udphdr *uh = NULL; struct tcphdr *nth; struct tcp_log_buffer *lgb; u_char *optp; #ifdef INET6 struct ip6_hdr *ip6; int isipv6; #endif /* INET6 */ int optlen, tlen, win, ulen; int ect = 0; bool incl_opts; uint16_t port; int output_ret; #ifdef INVARIANTS int thflags = tcp_get_flags(th); #endif 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 = tptoinpcb(tp); 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)) || (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) { ect = tcp_ecn_output_established(tp, &flags, 0, false); /* 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 = htonl(ect << 20); ip6->ip6_vfc = IPV6_VERSION; 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_tos = ect; 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_off = (sizeof (struct tcphdr) + optlen) >> 2; tcp_set_flags(nth, 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 #ifdef INET6 if (isipv6) { 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(inp, NULL); } #endif /* INET6 */ #if defined(INET6) && defined(INET) else #endif #ifdef INET { 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); lgb = NULL; if ((tp != NULL) && (tp->t_logstate != TCP_LOG_STATE_OFF)) { if (INP_WLOCKED(inp)) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = inp->inp_in_hpts; log.u_bbr.flex8 = 4; log.u_bbr.pkts_out = tp->t_maxseg; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.delivered = 0; lgb = tcp_log_event_(tp, nth, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK, 0, &log, false, NULL, NULL, 0, &tv); } else { /* * We can not log the packet, since we only own the * read lock, but a write lock is needed. The read lock * is not upgraded to a write lock, since only getting * the read lock was done intentionally to improve the * handling of SYN flooding attacks. * This happens only for pure SYN segments received in * the initial CLOSED state, or received in a more * advanced state than listen and the UDP encapsulation * port is unexpected. * The incoming SYN segments do not really belong to * the TCP connection and the handling does not change * the state of the TCP connection. Therefore, the * sending of the RST segments is not logged. Please * note that also the incoming SYN segments are not * logged. * * The following code ensures that the above description * is and stays correct. */ KASSERT((thflags & (TH_ACK|TH_SYN)) == TH_SYN && (tp->t_state == TCPS_CLOSED || (tp->t_state > TCPS_LISTEN && tp->t_port != port)), ("%s: Logging of TCP segment with flags 0x%b and " "UDP encapsulation port %u skipped in state %s", __func__, thflags, PRINT_TH_FLAGS, ntohs(port), tcpstates[tp->t_state])); } } if (flags & TH_ACK) TCPSTAT_INC(tcps_sndacks); else if (flags & (TH_SYN|TH_FIN|TH_RST)) TCPSTAT_INC(tcps_sndctrl); TCPSTAT_INC(tcps_sndtotal); #ifdef INET6 if (isipv6) { TCP_PROBE5(send, NULL, tp, ip6, tp, nth); output_ret = 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); output_ret = ip_output(m, NULL, NULL, 0, NULL, inp); } #endif if (lgb != NULL) lgb->tlb_errno = output_ret; } /* - * 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(). + * 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 by tcpcbstor declaration. */ struct tcpcb * tcp_newtcpcb(struct inpcb *inp) { - struct tcpcb_mem *tm; - struct tcpcb *tp; + struct tcpcb *tp = intotcpcb(inp); #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; + /* + * Historically allocation was done with M_ZERO. There is a lot of + * code that rely on that. For now take safe approach and zero whole + * tcpcb. This definitely can be optimized. + */ + bzero(&tp->t_start_zero, t_zero_size); /* Initialise cc_var struct for this tcpcb. */ - tp->ccv = &tm->ccv; - tp->ccv->type = IPPROTO_TCP; - tp->ccv->ccvc.tcp = tp; + tp->t_ccv.type = IPPROTO_TCP; + tp->t_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_attach(tp, CC_DEFAULT_ALGO()); - /* - * 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, NULL) > 0) { + if (CC_ALGO(tp)->cb_init(&tp->t_ccv, NULL) > 0) { cc_detach(tp); 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 (khelp_init_osd(HELPER_CLASS_TCP, &tp->t_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 - 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); + callout_init(&tp->tt_rexmt, 1); + callout_init(&tp->tt_persist, 1); + callout_init(&tp->tt_keep, 1); + callout_init(&tp->tt_2msl, 1); + callout_init(&tp->tt_delack, 1); switch (V_tcp_do_rfc1323) { case 0: break; default: case 1: tp->t_flags = (TF_REQ_SCALE|TF_REQ_TSTMP); break; case 2: tp->t_flags = TF_REQ_SCALE; break; case 3: tp->t_flags = TF_REQ_TSTMP; break; } 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 TCPHPTS /* * If using hpts lets drop a random number in so * not all new connections fall on the same CPU. */ inp->inp_hpts_cpu = hpts_random_cpu(inp); #endif #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 if (V_tcp_do_lrd) tp->t_flags |= TF_LRD; - return (tp); /* XXX */ + + /* + * XXXGL: this self-reference might be pointless. It will go away + * when the TCP timers are properly locked and could never fire after + * tcp_discardcb(). + */ + in_pcbref(inp); + + return (tp); } /* * 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 = tptosocket(tp); NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(tptoinpcb(tp)); if (TCPS_HAVERCVDSYN(tp->t_state)) { tcp_state_change(tp, TCPS_CLOSED); /* Don't use tcp_output() here due to possible recursion. */ (void)tcp_output_nodrop(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 = tptoinpcb(tp); 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; + tp->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); } /* 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_ALGO(tp)->cb_destroy(&tp->t_ccv); CC_DATA(tp) = NULL; /* Detach from the CC algorithm */ cc_detach(tp); #ifdef TCP_HHOOK - khelp_destroy_osd(tp->osd); + khelp_destroy_osd(&tp->t_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) { - bool released __diagused; - - released = tcp_freecb(tp); - KASSERT(!released, ("%s: inp %p should not have been released " - "here", __func__, inp)); - } + if (tp->tt_draincnt == 0) + tcp_freecb(tp); } bool tcp_freecb(struct tcpcb *tp) { struct inpcb *inp = tptoinpcb(tp); struct socket *so = tptosocket(tp); #ifdef INET6 bool isipv6 = (inp->inp_vflag & INP_IPV6) != 0; #endif INP_WLOCK_ASSERT(inp); - MPASS(tp->t_timers->tt_draincnt == 0); + MPASS(tp->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); /* * 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! * * XXXRRS: Updating must be after the stack fini() since * that may be converting some internal representation of * say srtt etc into the general one used by other stacks. * Lets also at least protect against the so being NULL * as RW stated below. */ if ((tp->t_rttupdated >= 4) && (so != NULL)) { 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); } refcount_release(&tp->t_fb->tfb_refcnt); - uma_zfree(V_tcpcb_zone, tp); return (in_pcbrele_wlocked(inp)); } /* * 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 = tptoinpcb(tp); struct socket *so = tptosocket(tp); 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; } #ifdef TCPHPTS tcp_hpts_remove(inp); #endif 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")); soisdisconnected(so); if (inp->inp_flags & INP_SOCKREF) { inp->inp_flags &= ~INP_SOCKREF; INP_WUNLOCK(inp); sorele(so); return (NULL); } return (tp); } /* * 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_WLOCK_ASSERT(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 inpcb_iterator inpi = INP_ALL_ITERATOR(&V_tcbinfo, INPLOOKUP_RLOCKPCB); struct xinpgen xig; struct inpcb *inp; 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); while ((inp = inp_next(&inpi)) != NULL) { if (inp->inp_gencnt <= xig.xig_gen && cr_canseeinpcb(req->td->td_ucred, inp) == 0) { struct xtcpcb xt; tcp_inptoxtp(inp, &xt); error = SYSCTL_OUT(req, &xt, sizeof xt); if (error) { INP_RUNLOCK(inp); break; } else continue; } } 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 (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 (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 /* Path MTU to try next when a fragmentation-needed message is received. */ static inline int tcp_next_pmtu(const struct icmp *icp, const struct ip *ip) { int 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); return (mtu); } static void tcp_ctlinput_with_port(struct icmp *icp, uint16_t port) { struct ip *ip; struct tcphdr *th; struct inpcb *inp; struct tcpcb *tp; struct inpcb *(*notify)(struct inpcb *, int); struct in_conninfo inc; tcp_seq icmp_tcp_seq; int errno, mtu; errno = icmp_errmap(icp); switch (errno) { case 0: return; case EMSGSIZE: notify = tcp_mtudisc_notify; break; case ECONNREFUSED: if (V_icmp_may_rst) notify = tcp_drop_syn_sent; else notify = tcp_notify; break; case EHOSTUNREACH: if (V_icmp_may_rst && icp->icmp_type == ICMP_TIMXCEED) notify = tcp_drop_syn_sent; else notify = tcp_notify; break; default: notify = tcp_notify; } ip = &icp->icmp_ip; th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2)); icmp_tcp_seq = th->th_seq; inp = in_pcblookup(&V_tcbinfo, ip->ip_dst, th->th_dport, ip->ip_src, th->th_sport, INPLOOKUP_WLOCKPCB, NULL); if (inp != NULL) { tp = intotcpcb(inp); #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE && errno == EMSGSIZE) { /* * MTU discovery for offloaded connections. Let * the TOE driver verify seq# and process it. */ mtu = tcp_next_pmtu(icp, ip); tcp_offload_pmtu_update(tp, icmp_tcp_seq, mtu); goto out; } #endif 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 (errno == EMSGSIZE) { /* * MTU discovery: we got a needfrag and * will potentially try a lower MTU. */ mtu = tcp_next_pmtu(icp, ip); /* * 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 = ip->ip_dst; inc.inc_fibnum = inp->inp_inc.inc_fibnum; tcp_hc_updatemtu(&inc, mtu); inp = tcp_mtudisc(inp, mtu); } } else inp = (*notify)(inp, errno); } } else { bzero(&inc, sizeof(inc)); inc.inc_fport = th->th_dport; inc.inc_lport = th->th_sport; inc.inc_faddr = ip->ip_dst; inc.inc_laddr = ip->ip_src; syncache_unreach(&inc, icmp_tcp_seq, port); } out: if (inp != NULL) INP_WUNLOCK(inp); } static void tcp_ctlinput(struct icmp *icmp) { tcp_ctlinput_with_port(icmp, htons(0)); } static void tcp_ctlinput_viaudp(udp_tun_icmp_param_t param) { /* Its a tunneled TCP over UDP icmp */ struct icmp *icmp = param.icmp; struct ip *outer_ip, *inner_ip; struct udphdr *udp; struct tcphdr *th, ttemp; int i_hlen, o_len; uint16_t port; outer_ip = (struct ip *)((caddr_t)icmp - sizeof(struct ip)); inner_ip = &icmp->icmp_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(icmp, port); } #endif /* INET */ #ifdef INET6 static inline int tcp6_next_pmtu(const struct icmp6_hdr *icmp6) { int 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; /* XXXNP: what is the adjustment for? */ return (mtu); } static void tcp6_ctlinput_with_port(struct ip6ctlparam *ip6cp, uint16_t port) { struct in6_addr *dst; struct inpcb *(*notify)(struct inpcb *, int); struct ip6_hdr *ip6; struct mbuf *m; struct inpcb *inp; struct tcpcb *tp; struct icmp6_hdr *icmp6; 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; int errno; icmp6 = ip6cp->ip6c_icmp6; m = ip6cp->ip6c_m; ip6 = ip6cp->ip6c_ip6; off = ip6cp->ip6c_off; dst = &ip6cp->ip6c_finaldst->sin6_addr; errno = icmp6_errmap(icmp6); switch (errno) { case 0: return; case EMSGSIZE: notify = tcp_mtudisc_notify; break; case ECONNREFUSED: if (V_icmp_may_rst) notify = tcp_drop_syn_sent; else notify = tcp_notify; break; case EHOSTUNREACH: /* * There are only four ICMPs that may reset connection: * - administratively prohibited * - port unreachable * - time exceeded in transit * - unknown next header */ if (V_icmp_may_rst && ((icmp6->icmp6_type == ICMP6_DST_UNREACH && (icmp6->icmp6_code == ICMP6_DST_UNREACH_ADMIN || icmp6->icmp6_code == ICMP6_DST_UNREACH_NOPORT)) || (icmp6->icmp6_type == ICMP6_TIME_EXCEEDED && icmp6->icmp6_code == ICMP6_TIME_EXCEED_TRANSIT) || (icmp6->icmp6_type == ICMP6_PARAM_PROB && icmp6->icmp6_code == ICMP6_PARAMPROB_NEXTHEADER))) notify = tcp_drop_syn_sent; else notify = tcp_notify; break; default: notify = tcp_notify; } /* 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); 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) { tp = intotcpcb(inp); #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE && errno == EMSGSIZE) { /* MTU discovery for offloaded connections. */ mtu = tcp6_next_pmtu(icmp6); tcp_offload_pmtu_update(tp, icmp_tcp_seq, mtu); goto out; } #endif 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 (errno == EMSGSIZE) { /* * MTU discovery: * If we got a needfrag set the MTU * in the route to the suggested new * value (if given) and then notify. */ mtu = tcp6_next_pmtu(icmp6); 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, errno); } } 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, port); } out: if (inp != NULL) INP_WUNLOCK(inp); } static void tcp6_ctlinput(struct ip6ctlparam *ctl) { tcp6_ctlinput_with_port(ctl, htons(0)); } static void tcp6_ctlinput_viaudp(udp_tun_icmp_param_t param) { struct ip6ctlparam *ip6cp = param.ip6cp; struct mbuf *m; struct udphdr *udp; uint16_t port; 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(ip6cp, 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. */ static struct inpcb * tcp_drop_syn_sent(struct inpcb *inp, int errno) { struct tcpcb *tp; NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(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) { return (tcp_mtudisc(inp, -1)); } static struct inpcb * tcp_mtudisc(struct inpcb *inp, int mtuoffer) { struct tcpcb *tp; struct socket *so; INP_WLOCK_ASSERT(inp); 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); if (tp->t_fb->tfb_tcp_mtu_chg != NULL) { /* * Conceptually the snd_nxt setting * and freeing sack holes should * be done by the default stacks * own tfb_tcp_mtu_chg(). */ tp->t_fb->tfb_tcp_mtu_chg(tp); } if (tcp_output(tp) < 0) return (NULL); else return (inp); } #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); } /* * Handle setsockopt(IPV6_USE_MIN_MTU) by a TCP stack. * * XXXGL: we are updating inpcb here with INC_IPV6MINMTU flag. * The right place to do that is ip6_setpktopt() that has just been * executed. By the way it just filled ip6po_minmtu for us. */ void tcp6_use_min_mtu(struct tcpcb *tp) { struct inpcb *inp = tptoinpcb(tp); INP_WLOCK_ASSERT(inp); /* * 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. */ inp->inp_inc.inc_flags |= INC_IPV6MINMTU; 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 && tp->t_maxseg > TCP6_MSS) tp->t_maxseg = TCP6_MSS; } } #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); } u_int tcp_fixed_maxseg(const struct tcpcb *tp) { 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 only consider fixed options that we would send every * time I.e. SACK is not considered. This is important * for cc modules to figure out what the modulo of the * cwnd should be. */ #define PAD(len) ((((len) / 4) + !!((len) % 4)) * 4) 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 += PAD(TCPOLEN_SIGNATURE); #endif } else { if (tp->t_flags & TF_REQ_TSTMP) optlen = TCPOLEN_TSTAMP_APPA; else optlen = PAD(TCPOLEN_MAXSEG); if (tp->t_flags & TF_REQ_SCALE) optlen += PAD(TCPOLEN_WINDOW); #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (tp->t_flags & TF_SIGNATURE) optlen += PAD(TCPOLEN_SIGNATURE); #endif if (tp->t_flags & TF_SACK_PERMIT) optlen += PAD(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; #ifdef INET struct sockaddr_in *fin = NULL, *lin = NULL; #endif struct epoch_tracker et; #ifdef INET6 struct sockaddr_in6 *fin6, *lin6; #endif int error; inp = 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]); #ifdef INET fin = (struct sockaddr_in *)&addrs[0]; lin = (struct sockaddr_in *)&addrs[1]; #endif 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 (!SOLISTENING(inp->inp_socket)) { 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"); static int tcp_sysctl_setsockopt(SYSCTL_HANDLER_ARGS) { return (sysctl_setsockopt(oidp, arg1, arg2, req, &V_tcbinfo, &tcp_ctloutput_set)); } SYSCTL_PROC(_net_inet_tcp, OID_AUTO, setsockopt, CTLFLAG_VNET | CTLTYPE_STRUCT | CTLFLAG_WR | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0, tcp_sysctl_setsockopt, "", "Set socket option for TCP endpoint"); #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; #ifdef INET struct sockaddr_in *fin = NULL, *lin = NULL; #endif struct epoch_tracker et; #ifdef INET6 struct sockaddr_in6 *fin6, *lin6; #endif int error; inp = 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]); #ifdef INET fin = (struct sockaddr_in *)&addrs[0]; lin = (struct sockaddr_in *)&addrs[1]; #endif 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) { 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); 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, const 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, const 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, const void *ip4hdr, const void *ip6hdr) { char *s, *sp; size_t size; #ifdef INET const struct ip *ip = (const struct ip *)ip4hdr; #endif #ifdef INET6 const struct ip6_hdr *ip6 = (const struct ip6_hdr *)ip6hdr; #endif /* INET6 */ /* * 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", tcp_get_flags(th), 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); sbintime_t now; bzero(xt, sizeof(*xt)); 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_dsack_bytes = tp->t_dsack_bytes; xt->t_dsack_tlp_bytes = tp->t_dsack_tlp_bytes; xt->t_dsack_pack = tp->t_dsack_pack; 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; \ + if (callout_active(&tp->ttt)) \ + xt->ttt = (tp->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); } 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; } } } int tcp_can_enable_pacing(void) { if ((tcp_pacing_limit == -1) || (tcp_pacing_limit > number_of_tcp_connections_pacing)) { atomic_fetchadd_int(&number_of_tcp_connections_pacing, 1); shadow_num_connections = number_of_tcp_connections_pacing; return (1); } else { return (0); } } static uint8_t tcp_pacing_warning = 0; void tcp_decrement_paced_conn(void) { uint32_t ret; ret = atomic_fetchadd_int(&number_of_tcp_connections_pacing, -1); shadow_num_connections = number_of_tcp_connections_pacing; KASSERT(ret != 0, ("tcp_paced_connection_exits -1 would cause wrap?")); if (ret == 0) { if (tcp_pacing_limit != -1) { printf("Warning all pacing is now disabled, count decrements invalidly!\n"); tcp_pacing_limit = 0; } else if (tcp_pacing_warning == 0) { printf("Warning pacing count is invalid, invalid decrement\n"); tcp_pacing_warning = 1; } } } diff --git a/sys/netinet/tcp_timer.c b/sys/netinet/tcp_timer.c index 9cfdc04fc6f5..a0eb4b0aad2d 100644 --- a/sys/netinet/tcp_timer.c +++ b/sys/netinet/tcp_timer.c @@ -1,1148 +1,1149 @@ /*- * 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_timer.c 8.2 (Berkeley) 5/24/95 */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_tcpdebug.h" #include "opt_rss.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef INET6 #include #endif #include #include #include #include #include #include #include #include #ifdef INET6 #include #endif #include #include int tcp_persmin; SYSCTL_PROC(_net_inet_tcp, OID_AUTO, persmin, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_persmin, 0, sysctl_msec_to_ticks, "I", "minimum persistence interval"); int tcp_persmax; SYSCTL_PROC(_net_inet_tcp, OID_AUTO, persmax, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_persmax, 0, sysctl_msec_to_ticks, "I", "maximum persistence interval"); int tcp_keepinit; SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINIT, keepinit, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_keepinit, 0, sysctl_msec_to_ticks, "I", "time to establish connection"); int tcp_keepidle; SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPIDLE, keepidle, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_keepidle, 0, sysctl_msec_to_ticks, "I", "time before keepalive probes begin"); int tcp_keepintvl; SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINTVL, keepintvl, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_keepintvl, 0, sysctl_msec_to_ticks, "I", "time between keepalive probes"); int tcp_delacktime; SYSCTL_PROC(_net_inet_tcp, TCPCTL_DELACKTIME, delacktime, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_delacktime, 0, sysctl_msec_to_ticks, "I", "Time before a delayed ACK is sent"); VNET_DEFINE(int, tcp_msl); SYSCTL_PROC(_net_inet_tcp, OID_AUTO, msl, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_VNET, &VNET_NAME(tcp_msl), 0, sysctl_msec_to_ticks, "I", "Maximum segment lifetime"); int tcp_rexmit_initial; SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_initial, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_rexmit_initial, 0, sysctl_msec_to_ticks, "I", "Initial Retransmission Timeout"); int tcp_rexmit_min; SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_min, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_rexmit_min, 0, sysctl_msec_to_ticks, "I", "Minimum Retransmission Timeout"); int tcp_rexmit_slop; SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_slop, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_rexmit_slop, 0, sysctl_msec_to_ticks, "I", "Retransmission Timer Slop"); VNET_DEFINE(int, tcp_always_keepalive) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, always_keepalive, CTLFLAG_VNET|CTLFLAG_RW, &VNET_NAME(tcp_always_keepalive) , 0, "Assume SO_KEEPALIVE on all TCP connections"); int tcp_fast_finwait2_recycle = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, fast_finwait2_recycle, CTLFLAG_RW, &tcp_fast_finwait2_recycle, 0, "Recycle closed FIN_WAIT_2 connections faster"); int tcp_finwait2_timeout; SYSCTL_PROC(_net_inet_tcp, OID_AUTO, finwait2_timeout, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_finwait2_timeout, 0, sysctl_msec_to_ticks, "I", "FIN-WAIT2 timeout"); int tcp_keepcnt = TCPTV_KEEPCNT; SYSCTL_INT(_net_inet_tcp, OID_AUTO, keepcnt, CTLFLAG_RW, &tcp_keepcnt, 0, "Number of keepalive probes to send"); /* max idle probes */ int tcp_maxpersistidle; int tcp_rexmit_drop_options = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, rexmit_drop_options, CTLFLAG_RW, &tcp_rexmit_drop_options, 0, "Drop TCP options from 3rd and later retransmitted SYN"); int tcp_maxunacktime = TCPTV_MAXUNACKTIME; SYSCTL_PROC(_net_inet_tcp, OID_AUTO, maxunacktime, CTLTYPE_INT|CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_maxunacktime, 0, sysctl_msec_to_ticks, "I", "Maximum time (in ms) that a session can linger without making progress"); VNET_DEFINE(int, tcp_pmtud_blackhole_detect); SYSCTL_INT(_net_inet_tcp, OID_AUTO, pmtud_blackhole_detection, CTLFLAG_RW|CTLFLAG_VNET, &VNET_NAME(tcp_pmtud_blackhole_detect), 0, "Path MTU Discovery Black Hole Detection Enabled"); #ifdef INET VNET_DEFINE(int, tcp_pmtud_blackhole_mss) = 1200; SYSCTL_INT(_net_inet_tcp, OID_AUTO, pmtud_blackhole_mss, CTLFLAG_RW|CTLFLAG_VNET, &VNET_NAME(tcp_pmtud_blackhole_mss), 0, "Path MTU Discovery Black Hole Detection lowered MSS"); #endif #ifdef INET6 VNET_DEFINE(int, tcp_v6pmtud_blackhole_mss) = 1220; SYSCTL_INT(_net_inet_tcp, OID_AUTO, v6pmtud_blackhole_mss, CTLFLAG_RW|CTLFLAG_VNET, &VNET_NAME(tcp_v6pmtud_blackhole_mss), 0, "Path MTU Discovery IPv6 Black Hole Detection lowered MSS"); #endif #ifdef RSS static int per_cpu_timers = 1; #else static int per_cpu_timers = 0; #endif SYSCTL_INT(_net_inet_tcp, OID_AUTO, per_cpu_timers, CTLFLAG_RW, &per_cpu_timers , 0, "run tcp timers on all cpus"); /* * Map the given inp to a CPU id. * * This queries RSS if it's compiled in, else it defaults to the current * CPU ID. */ inline int inp_to_cpuid(struct inpcb *inp) { u_int cpuid; if (per_cpu_timers) { #ifdef RSS cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype); if (cpuid == NETISR_CPUID_NONE) return (curcpu); /* XXX */ else return (cpuid); #endif /* * We don't have a flowid -> cpuid mapping, so cheat and * just map unknown cpuids to curcpu. Not the best, but * apparently better than defaulting to swi 0. */ cpuid = inp->inp_flowid % (mp_maxid + 1); if (! CPU_ABSENT(cpuid)) return (cpuid); return (curcpu); } else { return (0); } } int tcp_backoff[TCP_MAXRXTSHIFT + 1] = { 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 512, 512, 512 }; int tcp_totbackoff = 2559; /* sum of tcp_backoff[] */ /* * TCP timer processing. */ void tcp_timer_delack(void *xtp) { struct epoch_tracker et; struct tcpcb *tp = xtp; struct inpcb *inp = tptoinpcb(tp); INP_WLOCK(inp); CURVNET_SET(inp->inp_vnet); - if (callout_pending(&tp->t_timers->tt_delack) || - !callout_active(&tp->t_timers->tt_delack)) { + if (callout_pending(&tp->tt_delack) || + !callout_active(&tp->tt_delack)) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } - callout_deactivate(&tp->t_timers->tt_delack); + callout_deactivate(&tp->tt_delack); if ((inp->inp_flags & INP_DROPPED) != 0) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } tp->t_flags |= TF_ACKNOW; TCPSTAT_INC(tcps_delack); NET_EPOCH_ENTER(et); (void) tcp_output_unlock(tp); NET_EPOCH_EXIT(et); CURVNET_RESTORE(); } /* * Call tcp_close() from a callout context. */ static void tcp_timer_close(struct tcpcb *tp) { struct epoch_tracker et; struct inpcb *inp = tptoinpcb(tp); INP_WLOCK_ASSERT(inp); NET_EPOCH_ENTER(et); tp = tcp_close(tp); NET_EPOCH_EXIT(et); if (tp != NULL) INP_WUNLOCK(inp); } /* * Call tcp_drop() from a callout context. */ static void tcp_timer_drop(struct tcpcb *tp) { struct epoch_tracker et; struct inpcb *inp = tptoinpcb(tp); INP_WLOCK_ASSERT(inp); NET_EPOCH_ENTER(et); tp = tcp_drop(tp, ETIMEDOUT); NET_EPOCH_EXIT(et); if (tp != NULL) INP_WUNLOCK(inp); } void tcp_timer_2msl(void *xtp) { struct tcpcb *tp = xtp; struct inpcb *inp = tptoinpcb(tp); #ifdef TCPDEBUG int ostate; ostate = tp->t_state; #endif INP_WLOCK(inp); CURVNET_SET(inp->inp_vnet); tcp_log_end_status(tp, TCP_EI_STATUS_2MSL); tcp_free_sackholes(tp); - if (callout_pending(&tp->t_timers->tt_2msl) || - !callout_active(&tp->t_timers->tt_2msl)) { + if (callout_pending(&tp->tt_2msl) || + !callout_active(&tp->tt_2msl)) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } - callout_deactivate(&tp->t_timers->tt_2msl); + callout_deactivate(&tp->tt_2msl); if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } - KASSERT((tp->t_timers->tt_flags & TT_STOPPED) == 0, + KASSERT((tp->tt_flags & TT_STOPPED) == 0, ("%s: tp %p tcpcb can't be stopped here", __func__, tp)); /* * 2 MSL timeout in shutdown went off. If we're closed but * still waiting for peer to close and connection has been idle * too long delete connection control block. Otherwise, check * again in a bit. * * If fastrecycle of FIN_WAIT_2, in FIN_WAIT_2 and receiver has closed, * there's no point in hanging onto FIN_WAIT_2 socket. Just close it. * Ignore fact that there were recent incoming segments. * * XXXGL: check if inp_socket shall always be !NULL here? */ if (tp->t_state == TCPS_TIME_WAIT) { tcp_timer_close(tp); CURVNET_RESTORE(); return; } else if (tp->t_state == TCPS_FIN_WAIT_2 && tcp_fast_finwait2_recycle && inp->inp_socket && (inp->inp_socket->so_rcv.sb_state & SBS_CANTRCVMORE)) { TCPSTAT_INC(tcps_finwait2_drops); tcp_timer_close(tp); CURVNET_RESTORE(); return; } else { if (ticks - tp->t_rcvtime <= TP_MAXIDLE(tp)) { - callout_reset(&tp->t_timers->tt_2msl, + callout_reset(&tp->tt_2msl, TP_KEEPINTVL(tp), tcp_timer_2msl, tp); } else { tcp_timer_close(tp); CURVNET_RESTORE(); return; } } #ifdef TCPDEBUG if (tptosocket(tp)->so_options & SO_DEBUG) tcp_trace(TA_USER, ostate, tp, (void *)0, (struct tcphdr *)0, PRU_SLOWTIMO); #endif TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO); INP_WUNLOCK(inp); CURVNET_RESTORE(); } void tcp_timer_keep(void *xtp) { struct epoch_tracker et; struct tcpcb *tp = xtp; struct inpcb *inp = tptoinpcb(tp); struct tcptemp *t_template; #ifdef TCPDEBUG int ostate; ostate = tp->t_state; #endif INP_WLOCK(inp); CURVNET_SET(inp->inp_vnet); - if (callout_pending(&tp->t_timers->tt_keep) || - !callout_active(&tp->t_timers->tt_keep)) { + if (callout_pending(&tp->tt_keep) || + !callout_active(&tp->tt_keep)) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } - callout_deactivate(&tp->t_timers->tt_keep); + callout_deactivate(&tp->tt_keep); if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } - KASSERT((tp->t_timers->tt_flags & TT_STOPPED) == 0, + KASSERT((tp->tt_flags & TT_STOPPED) == 0, ("%s: tp %p tcpcb can't be stopped here", __func__, tp)); /* * Because we don't regularly reset the keepalive callout in * the ESTABLISHED state, it may be that we don't actually need * to send a keepalive yet. If that occurs, schedule another * call for the next time the keepalive timer might expire. */ if (TCPS_HAVEESTABLISHED(tp->t_state)) { u_int idletime; idletime = ticks - tp->t_rcvtime; if (idletime < TP_KEEPIDLE(tp)) { - callout_reset(&tp->t_timers->tt_keep, + callout_reset(&tp->tt_keep, TP_KEEPIDLE(tp) - idletime, tcp_timer_keep, tp); INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } } /* * Keep-alive timer went off; send something * or drop connection if idle for too long. */ 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. */ TCPSTAT_INC(tcps_keepprobe); t_template = tcpip_maketemplate(inp); if (t_template) { NET_EPOCH_ENTER(et); tcp_respond(tp, t_template->tt_ipgen, &t_template->tt_t, (struct mbuf *)NULL, tp->rcv_nxt, tp->snd_una - 1, 0); NET_EPOCH_EXIT(et); free(t_template, M_TEMP); } - callout_reset(&tp->t_timers->tt_keep, TP_KEEPINTVL(tp), + callout_reset(&tp->tt_keep, TP_KEEPINTVL(tp), tcp_timer_keep, tp); } else - callout_reset(&tp->t_timers->tt_keep, TP_KEEPIDLE(tp), + callout_reset(&tp->tt_keep, TP_KEEPIDLE(tp), tcp_timer_keep, tp); #ifdef TCPDEBUG if (inp->inp_socket->so_options & SO_DEBUG) tcp_trace(TA_USER, ostate, tp, (void *)0, (struct tcphdr *)0, PRU_SLOWTIMO); #endif TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO); INP_WUNLOCK(inp); CURVNET_RESTORE(); return; dropit: TCPSTAT_INC(tcps_keepdrops); NET_EPOCH_ENTER(et); tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX); tp = tcp_drop(tp, ETIMEDOUT); #ifdef TCPDEBUG if (tp != NULL && (tptosocket(tp)->so_options & SO_DEBUG)) tcp_trace(TA_USER, ostate, tp, (void *)0, (struct tcphdr *)0, PRU_SLOWTIMO); #endif TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO); NET_EPOCH_EXIT(et); if (tp != NULL) INP_WUNLOCK(inp); CURVNET_RESTORE(); } /* * Has this session exceeded the maximum time without seeing a substantive * acknowledgement? If so, return true; otherwise false. */ static bool tcp_maxunacktime_check(struct tcpcb *tp) { /* Are we tracking this timer for this session? */ if (TP_MAXUNACKTIME(tp) == 0) return false; /* Do we have a current measurement. */ if (tp->t_acktime == 0) return false; /* Are we within the acceptable range? */ if (TSTMP_GT(TP_MAXUNACKTIME(tp) + tp->t_acktime, (u_int)ticks)) return false; /* We exceeded the timer. */ TCPSTAT_INC(tcps_progdrops); return true; } void tcp_timer_persist(void *xtp) { struct epoch_tracker et; struct tcpcb *tp = xtp; struct inpcb *inp = tptoinpcb(tp); bool progdrop; int outrv; #ifdef TCPDEBUG int ostate; ostate = tp->t_state; #endif INP_WLOCK(inp); CURVNET_SET(inp->inp_vnet); - if (callout_pending(&tp->t_timers->tt_persist) || - !callout_active(&tp->t_timers->tt_persist)) { + if (callout_pending(&tp->tt_persist) || + !callout_active(&tp->tt_persist)) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } - callout_deactivate(&tp->t_timers->tt_persist); + callout_deactivate(&tp->tt_persist); if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } - KASSERT((tp->t_timers->tt_flags & TT_STOPPED) == 0, + KASSERT((tp->tt_flags & TT_STOPPED) == 0, ("%s: tp %p tcpcb can't be stopped here", __func__, tp)); /* * Persistence timer into zero window. * Force a byte to be output, if possible. */ 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. * Also, drop the connection if we haven't been making * progress. */ progdrop = tcp_maxunacktime_check(tp); if (progdrop || (tp->t_rxtshift == TCP_MAXRXTSHIFT && (ticks - tp->t_rcvtime >= tcp_maxpersistidle || ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff))) { if (!progdrop) TCPSTAT_INC(tcps_persistdrop); tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); tcp_timer_drop(tp); CURVNET_RESTORE(); return; } /* * 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) { TCPSTAT_INC(tcps_persistdrop); tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); tcp_timer_drop(tp); CURVNET_RESTORE(); return; } tcp_setpersist(tp); tp->t_flags |= TF_FORCEDATA; NET_EPOCH_ENTER(et); outrv = tcp_output_nodrop(tp); tp->t_flags &= ~TF_FORCEDATA; #ifdef TCPDEBUG if (tp != NULL && tptosocket(tp)->so_options & SO_DEBUG) tcp_trace(TA_USER, ostate, tp, NULL, NULL, PRU_SLOWTIMO); #endif TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO); (void) tcp_unlock_or_drop(tp, outrv); NET_EPOCH_EXIT(et); CURVNET_RESTORE(); } void tcp_timer_rexmt(void * xtp) { struct epoch_tracker et; struct tcpcb *tp = xtp; struct inpcb *inp = tptoinpcb(tp); int rexmt, outrv; bool isipv6; #ifdef TCPDEBUG int ostate; ostate = tp->t_state; #endif INP_WLOCK(inp); CURVNET_SET(inp->inp_vnet); - if (callout_pending(&tp->t_timers->tt_rexmt) || - !callout_active(&tp->t_timers->tt_rexmt)) { + if (callout_pending(&tp->tt_rexmt) || + !callout_active(&tp->tt_rexmt)) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } - callout_deactivate(&tp->t_timers->tt_rexmt); + callout_deactivate(&tp->tt_rexmt); if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } - KASSERT((tp->t_timers->tt_flags & TT_STOPPED) == 0, + KASSERT((tp->tt_flags & TT_STOPPED) == 0, ("%s: tp %p tcpcb can't be stopped here", __func__, tp)); tcp_free_sackholes(tp); TCP_LOG_EVENT(tp, NULL, NULL, NULL, TCP_LOG_RTO, 0, 0, NULL, false); if (tp->t_fb->tfb_tcp_rexmit_tmr) { /* The stack has a timer action too. */ (*tp->t_fb->tfb_tcp_rexmit_tmr)(tp); } /* * Retransmission timer went off. Message has not * been acked within retransmit interval. Back off * to a longer retransmit interval and retransmit one segment. * * If we've either exceeded the maximum number of retransmissions, * or we've gone long enough without making progress, then drop * the session. */ if (++tp->t_rxtshift > TCP_MAXRXTSHIFT || tcp_maxunacktime_check(tp)) { if (tp->t_rxtshift > TCP_MAXRXTSHIFT) TCPSTAT_INC(tcps_timeoutdrop); tp->t_rxtshift = TCP_MAXRXTSHIFT; tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN); tcp_timer_drop(tp); CURVNET_RESTORE(); return; } 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; if ((tp->t_flags & TF_RCVD_TSTMP) == 0) tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); /* In the event that we've negotiated timestamps * badrxtwin will be set to the value that we set * the retransmitted packet's to_tsval to by tcp_output */ tp->t_flags |= TF_PREVVALID; } else tp->t_flags &= ~TF_PREVVALID; TCPSTAT_INC(tcps_rexmttimeo); if ((tp->t_state == TCPS_SYN_SENT) || (tp->t_state == TCPS_SYN_RECEIVED)) rexmt = tcp_rexmit_initial * tcp_backoff[tp->t_rxtshift]; else rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; TCPT_RANGESET(tp->t_rxtcur, rexmt, tp->t_rttmin, TCPTV_REXMTMAX); /* * 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 = (inp->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))) { if (tp->t_rxtshift == 1) { /* * We enter blackhole detection after the first * unsuccessful timer based retransmission. * Then we reduce up to two times the MSS, each * candidate giving two tries of retransmissions. * But we give a candidate only two tries, if it * actually reduces the MSS. */ tp->t_blackhole_enter = 2; tp->t_blackhole_exit = tp->t_blackhole_enter; if (isipv6) { #ifdef INET6 if (tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) tp->t_blackhole_exit += 2; if (tp->t_maxseg > V_tcp_v6mssdflt && V_tcp_v6pmtud_blackhole_mss > V_tcp_v6mssdflt) tp->t_blackhole_exit += 2; #endif } else { #ifdef INET if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) tp->t_blackhole_exit += 2; if (tp->t_maxseg > V_tcp_mssdflt && V_tcp_pmtud_blackhole_mss > V_tcp_mssdflt) tp->t_blackhole_exit += 2; #endif } } if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD|TF2_PLPMTU_MAXSEGSNT)) == (TF2_PLPMTU_PMTUD|TF2_PLPMTU_MAXSEGSNT)) && (tp->t_rxtshift >= tp->t_blackhole_enter && tp->t_rxtshift < tp->t_blackhole_exit && (tp->t_rxtshift - tp->t_blackhole_enter) % 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 && V_tcp_v6pmtud_blackhole_mss > V_tcp_v6mssdflt) { /* Use the sysctl tuneable blackhole MSS. */ tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; 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; 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 && V_tcp_pmtud_blackhole_mss > V_tcp_mssdflt) { /* Use the sysctl tuneable blackhole MSS. */ tp->t_maxseg = V_tcp_pmtud_blackhole_mss; 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; TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); } #endif /* * Reset the slow-start flight size * as it may depend on the new MSS. */ if (CC_ALGO(tp)->conn_init != NULL) - CC_ALGO(tp)->conn_init(tp->ccv); + CC_ALGO(tp)->conn_init(&tp->t_ccv); } 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. */ if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && (tp->t_rxtshift >= tp->t_blackhole_exit)) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; tp->t_maxseg = tp->t_pmtud_saved_maxseg; TCPSTAT_INC(tcps_pmtud_blackhole_failed); /* * Reset the slow-start flight size as it * may depend on the new MSS. */ if (CC_ALGO(tp)->conn_init != NULL) - CC_ALGO(tp)->conn_init(tp->ccv); + CC_ALGO(tp)->conn_init(&tp->t_ccv); } } } /* * 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, notify the L3 protocol that we're having * connection problems. */ if (tp->t_rxtshift > TCP_RTT_INVALIDATE) { #ifdef INET6 if ((inp->inp_vflag & INP_IPV6) != 0) in6_losing(inp); else #endif in_losing(inp); } tp->snd_nxt = tp->snd_una; tp->snd_recover = tp->snd_max; /* * Force a segment to be sent. */ tp->t_flags |= TF_ACKNOW; /* * If timing a segment in this window, stop the timer. */ tp->t_rtttime = 0; cc_cong_signal(tp, NULL, CC_RTO); NET_EPOCH_ENTER(et); outrv = tcp_output_nodrop(tp); #ifdef TCPDEBUG if (tp != NULL && (tptosocket(tp)->so_options & SO_DEBUG)) tcp_trace(TA_USER, ostate, tp, (void *)0, (struct tcphdr *)0, PRU_SLOWTIMO); #endif TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO); (void) tcp_unlock_or_drop(tp, outrv); NET_EPOCH_EXIT(et); CURVNET_RESTORE(); } void tcp_timer_activate(struct tcpcb *tp, uint32_t timer_type, u_int delta) { struct callout *t_callout; callout_func_t *f_callout; struct inpcb *inp = tptoinpcb(tp); int cpu = inp_to_cpuid(inp); #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) return; #endif - if (tp->t_timers->tt_flags & TT_STOPPED) + if (tp->tt_flags & TT_STOPPED) return; switch (timer_type) { case TT_DELACK: - t_callout = &tp->t_timers->tt_delack; + t_callout = &tp->tt_delack; f_callout = tcp_timer_delack; break; case TT_REXMT: - t_callout = &tp->t_timers->tt_rexmt; + t_callout = &tp->tt_rexmt; f_callout = tcp_timer_rexmt; break; case TT_PERSIST: - t_callout = &tp->t_timers->tt_persist; + t_callout = &tp->tt_persist; f_callout = tcp_timer_persist; break; case TT_KEEP: - t_callout = &tp->t_timers->tt_keep; + t_callout = &tp->tt_keep; f_callout = tcp_timer_keep; break; case TT_2MSL: - t_callout = &tp->t_timers->tt_2msl; + t_callout = &tp->tt_2msl; f_callout = tcp_timer_2msl; break; default: if (tp->t_fb->tfb_tcp_timer_activate) { tp->t_fb->tfb_tcp_timer_activate(tp, timer_type, delta); return; } panic("tp %p bad timer_type %#x", tp, timer_type); } if (delta == 0) { callout_stop(t_callout); } else { callout_reset_on(t_callout, delta, f_callout, tp, cpu); } } int tcp_timer_active(struct tcpcb *tp, uint32_t timer_type) { struct callout *t_callout; switch (timer_type) { case TT_DELACK: - t_callout = &tp->t_timers->tt_delack; + t_callout = &tp->tt_delack; break; case TT_REXMT: - t_callout = &tp->t_timers->tt_rexmt; + t_callout = &tp->tt_rexmt; break; case TT_PERSIST: - t_callout = &tp->t_timers->tt_persist; + t_callout = &tp->tt_persist; break; case TT_KEEP: - t_callout = &tp->t_timers->tt_keep; + t_callout = &tp->tt_keep; break; case TT_2MSL: - t_callout = &tp->t_timers->tt_2msl; + t_callout = &tp->tt_2msl; break; default: if (tp->t_fb->tfb_tcp_timer_active) { return(tp->t_fb->tfb_tcp_timer_active(tp, timer_type)); } panic("tp %p bad timer_type %#x", tp, timer_type); } return callout_active(t_callout); } /* * Stop the timer from running, and apply a flag * against the timer_flags that will force the * timer never to run. The flag is needed to assure * a race does not leave it running and cause * the timer to possibly restart itself (keep and persist * especially do this). */ int tcp_timer_suspend(struct tcpcb *tp, uint32_t timer_type) { struct callout *t_callout; uint32_t t_flags; switch (timer_type) { case TT_DELACK: t_flags = TT_DELACK_SUS; - t_callout = &tp->t_timers->tt_delack; + t_callout = &tp->tt_delack; break; case TT_REXMT: t_flags = TT_REXMT_SUS; - t_callout = &tp->t_timers->tt_rexmt; + t_callout = &tp->tt_rexmt; break; case TT_PERSIST: t_flags = TT_PERSIST_SUS; - t_callout = &tp->t_timers->tt_persist; + t_callout = &tp->tt_persist; break; case TT_KEEP: t_flags = TT_KEEP_SUS; - t_callout = &tp->t_timers->tt_keep; + t_callout = &tp->tt_keep; break; case TT_2MSL: t_flags = TT_2MSL_SUS; - t_callout = &tp->t_timers->tt_2msl; + t_callout = &tp->tt_2msl; break; default: panic("tp:%p bad timer_type 0x%x", tp, timer_type); } - tp->t_timers->tt_flags |= t_flags; + tp->tt_flags |= t_flags; return (callout_stop(t_callout)); } void tcp_timers_unsuspend(struct tcpcb *tp, uint32_t timer_type) { + switch (timer_type) { case TT_DELACK: - if (tp->t_timers->tt_flags & TT_DELACK_SUS) { - tp->t_timers->tt_flags &= ~TT_DELACK_SUS; + if (tp->tt_flags & TT_DELACK_SUS) { + tp->tt_flags &= ~TT_DELACK_SUS; if (tp->t_flags & TF_DELACK) { /* Delayed ack timer should be up activate a timer */ tp->t_flags &= ~TF_DELACK; tcp_timer_activate(tp, TT_DELACK, tcp_delacktime); } } break; case TT_REXMT: - if (tp->t_timers->tt_flags & TT_REXMT_SUS) { - tp->t_timers->tt_flags &= ~TT_REXMT_SUS; + if (tp->tt_flags & TT_REXMT_SUS) { + tp->tt_flags &= ~TT_REXMT_SUS; if (SEQ_GT(tp->snd_max, tp->snd_una) && (tcp_timer_active((tp), TT_PERSIST) == 0) && tp->snd_wnd) { /* We have outstanding data activate a timer */ tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); } } break; case TT_PERSIST: - if (tp->t_timers->tt_flags & TT_PERSIST_SUS) { - tp->t_timers->tt_flags &= ~TT_PERSIST_SUS; + if (tp->tt_flags & TT_PERSIST_SUS) { + tp->tt_flags &= ~TT_PERSIST_SUS; if (tp->snd_wnd == 0) { /* Activate the persists timer */ tp->t_rxtshift = 0; tcp_setpersist(tp); } } break; case TT_KEEP: - if (tp->t_timers->tt_flags & TT_KEEP_SUS) { - tp->t_timers->tt_flags &= ~TT_KEEP_SUS; + if (tp->tt_flags & TT_KEEP_SUS) { + tp->tt_flags &= ~TT_KEEP_SUS; tcp_timer_activate(tp, TT_KEEP, TCPS_HAVEESTABLISHED(tp->t_state) ? TP_KEEPIDLE(tp) : TP_KEEPINIT(tp)); } break; case TT_2MSL: - if (tp->t_timers->tt_flags &= TT_2MSL_SUS) { + if (tp->tt_flags &= TT_2MSL_SUS) { struct socket *so = tptosocket(tp); - tp->t_timers->tt_flags &= ~TT_2MSL_SUS; + tp->tt_flags &= ~TT_2MSL_SUS; if ((tp->t_state == TCPS_FIN_WAIT_2) && (so == NULL || /* XXXGL: needed? */ (so->so_rcv.sb_state & SBS_CANTRCVMORE))) { /* Star the 2MSL timer */ tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle) ? tcp_finwait2_timeout : TP_MAXIDLE(tp)); } } break; default: panic("tp:%p bad timer_type 0x%x", tp, timer_type); } } static void tcp_timer_discard(void *ptp) { struct epoch_tracker et; struct tcpcb *tp = (struct tcpcb *)ptp; struct inpcb *inp = tptoinpcb(tp); INP_WLOCK(inp); CURVNET_SET(inp->inp_vnet); NET_EPOCH_ENTER(et); - KASSERT((tp->t_timers->tt_flags & TT_STOPPED) != 0, + KASSERT((tp->tt_flags & TT_STOPPED) != 0, ("%s: tcpcb has to be stopped here", __func__)); - if (--tp->t_timers->tt_draincnt > 0 || + if (--tp->tt_draincnt > 0 || tcp_freecb(tp) == false) INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); CURVNET_RESTORE(); } void tcp_timer_stop(struct tcpcb *tp, uint32_t timer_type) { struct callout *t_callout; - tp->t_timers->tt_flags |= TT_STOPPED; + tp->tt_flags |= TT_STOPPED; switch (timer_type) { case TT_DELACK: - t_callout = &tp->t_timers->tt_delack; + t_callout = &tp->tt_delack; break; case TT_REXMT: - t_callout = &tp->t_timers->tt_rexmt; + t_callout = &tp->tt_rexmt; break; case TT_PERSIST: - t_callout = &tp->t_timers->tt_persist; + t_callout = &tp->tt_persist; break; case TT_KEEP: - t_callout = &tp->t_timers->tt_keep; + t_callout = &tp->tt_keep; break; case TT_2MSL: - t_callout = &tp->t_timers->tt_2msl; + t_callout = &tp->tt_2msl; break; default: if (tp->t_fb->tfb_tcp_timer_stop) { /* * XXXrrs we need to look at this with the * stop case below (flags). */ tp->t_fb->tfb_tcp_timer_stop(tp, timer_type); return; } panic("tp %p bad timer_type %#x", tp, timer_type); } if (callout_async_drain(t_callout, tcp_timer_discard) == 0) { /* * Can't stop the callout, defer tcpcb actual deletion * to the last one. We do this using the async drain * function and incrementing the count in */ - tp->t_timers->tt_draincnt++; + tp->tt_draincnt++; } } diff --git a/sys/netinet/tcp_timer.h b/sys/netinet/tcp_timer.h index 8742098855ab..e72ab5f90546 100644 --- a/sys/netinet/tcp_timer.h +++ b/sys/netinet/tcp_timer.h @@ -1,239 +1,227 @@ /*- * 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_timer.h 8.1 (Berkeley) 6/10/93 * $FreeBSD$ */ #ifndef _NETINET_TCP_TIMER_H_ #define _NETINET_TCP_TIMER_H_ /* * The TCPT_REXMT timer is used to force retransmissions. * The TCP has the TCPT_REXMT timer set whenever segments * have been sent for which ACKs are expected but not yet * received. If an ACK is received which advances tp->snd_una, * then the retransmit timer is cleared (if there are no more * outstanding segments) or reset to the base value (if there * are more ACKs expected). Whenever the retransmit timer goes off, * we retransmit one unacknowledged segment, and do a backoff * on the retransmit timer. * * The TCPT_PERSIST timer is used to keep window size information * flowing even if the window goes shut. If all previous transmissions * have been acknowledged (so that there are no retransmissions in progress), * and the window is too small to bother sending anything, then we start * the TCPT_PERSIST timer. When it expires, if the window is nonzero, * we go to transmit state. Otherwise, at intervals send a single byte * into the peer's window to force him to update our window information. * We do this at most as often as TCPT_PERSMIN time intervals, * but no more frequently than the current estimate of round-trip * packet time. The TCPT_PERSIST timer is cleared whenever we receive * a window update from the peer. * * The TCPT_KEEP timer is used to keep connections alive. If an * connection is idle (no segments received) for TCPTV_KEEP_INIT amount of time, * but not yet established, then we drop the connection. Once the connection * is established, if the connection is idle for TCPTV_KEEP_IDLE time * (and keepalives have been enabled on the socket), we begin to probe * the connection. We force the peer to send us a segment by sending: * * This segment is (deliberately) outside the window, and should elicit * an ack segment in response from the peer. If, despite the TCPT_KEEP * initiated segments we cannot elicit a response from a peer in TCPT_MAXIDLE * amount of time probing, then we drop the connection. */ /* * Time constants. */ #define TCPTV_MSL ( 30*hz) /* max seg lifetime (hah!) */ #define TCPTV_SRTTBASE 0 /* base roundtrip time; if 0, no idea yet */ #define TCPTV_RTOBASE ( 1*hz) /* assumed RTO if no info */ #define TCPTV_PERSMIN ( 5*hz) /* minimum persist interval */ #define TCPTV_PERSMAX ( 60*hz) /* maximum persist interval */ #define TCPTV_KEEP_INIT ( 75*hz) /* initial connect keepalive */ #define TCPTV_KEEP_IDLE (120*60*hz) /* dflt time before probing */ #define TCPTV_KEEPINTVL ( 75*hz) /* default probe interval */ #define TCPTV_KEEPCNT 8 /* max probes before drop */ #define TCPTV_MAXUNACKTIME 0 /* max time without making progress */ #define TCPTV_FINWAIT2_TIMEOUT (60*hz) /* FIN_WAIT_2 timeout if no receiver */ /* * Minimum retransmit timer is 3 ticks, for algorithmic stability. * TCPT_RANGESET() will add another TCPTV_CPU_VAR to deal with * the expected worst-case processing variances by the kernels * representing the end points. Such variances do not always show * up in the srtt because the timestamp is often calculated at * the interface rather then at the TCP layer. This value is * typically 50ms. However, it is also possible that delayed * acks (typically 100ms) could create issues so we set the slop * to 200ms to try to cover it. Note that, properly speaking, * delayed-acks should not create a major issue for interactive * environments which 'P'ush the last segment, at least as * long as implementations do the required 'at least one ack * for every two packets' for the non-interactive streaming case. * (maybe the RTO calculation should use 2*RTT instead of RTT * to handle the ack-every-other-packet case). * * The prior minimum of 1*hz (1 second) badly breaks throughput on any * networks faster then a modem that has minor (e.g. 1%) packet loss. */ #define TCPTV_MIN ( hz/33 ) /* minimum allowable value */ #define TCPTV_CPU_VAR ( hz/5 ) /* cpu variance allowed (200ms) */ #define TCPTV_REXMTMAX ( 64*hz) /* max allowable REXMT value */ #define TCPTV_TWTRUNC 8 /* RTO factor to truncate TW */ #define TCP_MAXRXTSHIFT 12 /* maximum retransmits */ #define TCPTV_DELACK ( hz/25 ) /* 40ms timeout */ /* * If we exceed this number of retransmits for a single segment, we'll consider * the current srtt measurement no longer valid and will recalculate from * scratch starting with the next ACK. */ #define TCP_RTT_INVALIDATE (TCP_MAXRXTSHIFT / 4) #ifdef TCPTIMERS static const char *tcptimers[] = { "REXMT", "PERSIST", "KEEP", "2MSL", "DELACK" }; #endif /* * Force a time value to be in a certain range. */ #define TCPT_RANGESET(tv, value, tvmin, tvmax) do { \ (tv) = (value) + tcp_rexmit_slop; \ if ((u_long)(tv) < (u_long)(tvmin)) \ (tv) = (tvmin); \ if ((u_long)(tv) > (u_long)(tvmax)) \ (tv) = (tvmax); \ } while(0) #ifdef _KERNEL -struct xtcp_timer; - -struct tcp_timer { - struct callout tt_rexmt; /* retransmit timer */ - struct callout tt_persist; /* retransmit persistence */ - struct callout tt_keep; /* keepalive */ - struct callout tt_2msl; /* 2*msl TIME_WAIT timer */ - struct callout tt_delack; /* delayed ACK timer */ - uint32_t tt_flags; /* Timers flags */ - uint32_t tt_draincnt; /* Count being drained */ -}; - /* - * Flags for the tt_flags field. + * Flags for the tcpcb's tt_flags field. */ #define TT_DELACK 0x0001 #define TT_REXMT 0x0002 #define TT_PERSIST 0x0004 #define TT_KEEP 0x0008 #define TT_2MSL 0x0010 #define TT_MASK (TT_DELACK|TT_REXMT|TT_PERSIST|TT_KEEP|TT_2MSL) /* * Suspend flags - used when suspending a timer * from ever running again. */ #define TT_DELACK_SUS 0x0100 #define TT_REXMT_SUS 0x0200 #define TT_PERSIST_SUS 0x0400 #define TT_KEEP_SUS 0x0800 #define TT_2MSL_SUS 0x1000 #define TT_STOPPED 0x00010000 #define TP_KEEPINIT(tp) ((tp)->t_keepinit ? (tp)->t_keepinit : tcp_keepinit) #define TP_KEEPIDLE(tp) ((tp)->t_keepidle ? (tp)->t_keepidle : tcp_keepidle) #define TP_KEEPINTVL(tp) ((tp)->t_keepintvl ? (tp)->t_keepintvl : tcp_keepintvl) #define TP_KEEPCNT(tp) ((tp)->t_keepcnt ? (tp)->t_keepcnt : tcp_keepcnt) #define TP_MAXIDLE(tp) (TP_KEEPCNT(tp) * TP_KEEPINTVL(tp)) #define TP_MAXUNACKTIME(tp) \ ((tp)->t_maxunacktime ? (tp)->t_maxunacktime : tcp_maxunacktime) /* * Obtain the time until the restransmit timer should fire. * This macro ensures the restransmit timer fires at the earlier of the * t_rxtcur value or the time the maxunacktime would be exceeded. */ #define TP_RXTCUR(tp) \ ((TP_MAXUNACKTIME(tp) == 0 || tp->t_acktime == 0) ? tp->t_rxtcur : \ max(1, min(tp->t_rxtcur, tp->t_acktime + TP_MAXUNACKTIME(tp) - ticks))) extern int tcp_persmin; /* minimum persist interval */ extern int tcp_persmax; /* maximum persist interval */ extern int tcp_keepinit; /* time to establish connection */ extern int tcp_keepidle; /* time before keepalive probes begin */ extern int tcp_keepintvl; /* time between keepalive probes */ extern int tcp_keepcnt; /* number of keepalives */ extern int tcp_delacktime; /* time before sending a delayed ACK */ extern int tcp_maxunacktime; /* max time without making progress */ extern int tcp_maxpersistidle; extern int tcp_rexmit_initial; extern int tcp_rexmit_min; extern int tcp_rexmit_slop; extern int tcp_ttl; /* time to live for TCP segs */ extern int tcp_backoff[]; extern int tcp_totbackoff; extern int tcp_rexmit_drop_options; extern int tcp_finwait2_timeout; extern int tcp_fast_finwait2_recycle; VNET_DECLARE(int, tcp_always_keepalive); #define V_tcp_always_keepalive VNET(tcp_always_keepalive) VNET_DECLARE(int, tcp_pmtud_blackhole_detect); #define V_tcp_pmtud_blackhole_detect VNET(tcp_pmtud_blackhole_detect) VNET_DECLARE(int, tcp_pmtud_blackhole_mss); #define V_tcp_pmtud_blackhole_mss VNET(tcp_pmtud_blackhole_mss) VNET_DECLARE(int, tcp_v6pmtud_blackhole_mss); #define V_tcp_v6pmtud_blackhole_mss VNET(tcp_v6pmtud_blackhole_mss) VNET_DECLARE(int, tcp_msl); #define V_tcp_msl VNET(tcp_msl) void tcp_timer_init(void); void tcp_timer_2msl(void *xtp); void tcp_timer_keep(void *xtp); void tcp_timer_persist(void *xtp); void tcp_timer_rexmt(void *xtp); void tcp_timer_delack(void *xtp); #endif /* _KERNEL */ #endif /* !_NETINET_TCP_TIMER_H_ */ diff --git a/sys/netinet/tcp_usrreq.c b/sys/netinet/tcp_usrreq.c index 4ac563d1fc8e..2d54ea3c8003 100644 --- a/sys/netinet/tcp_usrreq.c +++ b/sys/netinet/tcp_usrreq.c @@ -1,3190 +1,3188 @@ /*- * 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 #ifdef DDB #include #endif #include #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 #include #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(); 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); 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); KASSERT(inp->inp_flags & INP_DROPPED || tp->t_state < TCPS_SYN_SENT, ("%s: inp %p not dropped or embryonic", __func__, inp)); tcp_discardcb(tp); in_pcbdetach(inp); in_pcbfree(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; #ifdef KDTRACE_HOOKS struct tcpcb *tp = NULL; #endif struct sockaddr_in *sinp; sinp = (struct sockaddr_in *)nam; if (nam->sa_family != AF_INET) { /* * Preserve compatibility with old programs. */ if (nam->sa_family != AF_UNSPEC || nam->sa_len < offsetof(struct sockaddr_in, sin_zero) || sinp->sin_addr.s_addr != INADDR_ANY) return (EAFNOSUPPORT); nam->sa_family = AF_INET; } if (nam->sa_len != sizeof(*sinp)) return (EINVAL); /* * Must check for multicast addresses and disallow binding * to them. */ if (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_DROPPED) { error = EINVAL; goto out; } #ifdef KDTRACE_HOOKS tp = intotcpcb(inp); #endif 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; #ifdef KDTRACE_HOOKS struct tcpcb *tp = NULL; #endif struct sockaddr_in6 *sin6; u_char vflagsav; sin6 = (struct sockaddr_in6 *)nam; if (nam->sa_family != AF_INET6) return (EAFNOSUPPORT); if (nam->sa_len != sizeof(*sin6)) return (EINVAL); /* * Must check for multicast addresses and disallow binding * to them. */ if (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_DROPPED) { error = EINVAL; goto out; } #ifdef KDTRACE_HOOKS tp = intotcpcb(inp); #endif 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_DROPPED) { error = EINVAL; goto out; } tp = intotcpcb(inp); TCPDEBUG1(); SOCK_LOCK(so); error = solisten_proto_check(so); if (error != 0) { SOCK_UNLOCK(so); goto out; } if (inp->inp_lport == 0) { INP_HASH_WLOCK(&V_tcbinfo); error = in_pcbbind(inp, NULL, 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 } else { solisten_proto_abort(so); } 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_DROPPED) { error = EINVAL; goto out; } vflagsav = inp->inp_vflag; tp = intotcpcb(inp); TCPDEBUG1(); SOCK_LOCK(so); error = solisten_proto_check(so); if (error != 0) { SOCK_UNLOCK(so); goto out; } INP_HASH_WLOCK(&V_tcbinfo); if (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, NULL, 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 } else { solisten_proto_abort(so); } 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_family != AF_INET) return (EAFNOSUPPORT); if (nam->sa_len != sizeof (*sinp)) return (EINVAL); /* * Must disallow TCP ``connections'' to multicast addresses. */ if (IN_MULTICAST(ntohl(sinp->sin_addr.s_addr))) return (EAFNOSUPPORT); if (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_DROPPED) { error = ECONNREFUSED; goto out; } if (SOLISTENING(so)) { error = EOPNOTSUPP; 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 = tcp_output(tp); KASSERT(error >= 0, ("TCP stack %s requested tcp_drop(%p) at connect()" ", error code %d", tp->t_fb->tfb_tcp_block_name, tp, -error)); 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_family != AF_INET6) return (EAFNOSUPPORT); if (nam->sa_len != sizeof (*sin6)) return (EINVAL); /* * Must disallow TCP ``connections'' to multicast addresses. */ if (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_DROPPED) { error = ECONNREFUSED; goto out; } if (SOLISTENING(so)) { error = EINVAL; 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 = 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; NET_EPOCH_ENTER(et); if ((error = tcp6_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 = tcp_output(tp); out_in_epoch: NET_EPOCH_EXIT(et); out: KASSERT(error >= 0, ("TCP stack %s requested tcp_drop(%p) at connect()" ", error code %d", tp->t_fb->tfb_tcp_block_name, tp, -error)); /* * 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_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; #ifdef KDTRACE_HOOKS struct tcpcb *tp = NULL; #endif 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_DROPPED) { error = ECONNABORTED; goto out; } #ifdef KDTRACE_HOOKS tp = intotcpcb(inp); #endif 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; #ifdef KDTRACE_HOOKS struct tcpcb *tp = NULL; #endif 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_DROPPED) { error = ECONNABORTED; goto out; } #ifdef KDTRACE_HOOKS tp = intotcpcb(inp); #endif 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; inp = sotoinpcb(so); KASSERT(inp != NULL, ("inp == NULL")); INP_WLOCK(inp); if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); NET_EPOCH_ENTER(et); TCPDEBUG1(); socantsendmore(so); tcp_usrclosed(tp); if (!(inp->inp_flags & INP_DROPPED)) error = tcp_output_nodrop(tp); TCPDEBUG2(PRU_SHUTDOWN); TCP_PROBE2(debug__user, tp, PRU_SHUTDOWN); error = tcp_unlock_or_drop(tp, error); 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 outrv = 0, error = 0; TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_rcvd: inp == NULL")); INP_WLOCK(inp); if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); NET_EPOCH_ENTER(et); 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; #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) tcp_offload_rcvd(tp); else #endif outrv = tcp_output_nodrop(tp); out: TCPDEBUG2(PRU_RCVD); TCP_PROBE2(debug__user, tp, PRU_RCVD); (void) tcp_unlock_or_drop(tp, outrv); NET_EPOCH_EXIT(et); 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; if (control != NULL) { /* TCP doesn't do control messages (rights, creds, etc) */ if (control->m_len) { m_freem(control); return (EINVAL); } m_freem(control); /* empty control, just free it */ } inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_send: inp == NULL")); INP_WLOCK(inp); if (inp->inp_flags & INP_DROPPED) { if (m != NULL && (flags & PRUS_NOTREADY) == 0) m_freem(m); INP_WUNLOCK(inp); return (ECONNRESET); } vflagsav = inp->inp_vflag; incflagsav = inp->inp_inc.inc_flags; restoreflags = false; tp = intotcpcb(inp); NET_EPOCH_ENTER(et); if ((flags & PRUS_OOB) != 0 && (error = tcp_pru_options_support(tp, PRUS_OOB)) != 0) goto out; TCPDEBUG1(); if (nam != NULL && tp->t_state < TCPS_SYN_SENT) { if (tp->t_state == TCPS_LISTEN) { error = EINVAL; goto out; } switch (nam->sa_family) { #ifdef INET case AF_INET: sinp = (struct sockaddr_in *)nam; if (sinp->sin_len != sizeof(struct sockaddr_in)) { error = EINVAL; goto out; } if ((inp->inp_vflag & INP_IPV6) != 0) { error = EAFNOSUPPORT; goto out; } if (IN_MULTICAST(ntohl(sinp->sin_addr.s_addr))) { error = EAFNOSUPPORT; goto out; } if (ntohl(sinp->sin_addr.s_addr) == INADDR_BROADCAST) { error = EACCES; goto out; } if ((error = prison_remote_ip4(td->td_ucred, &sinp->sin_addr))) 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)) { error = EINVAL; goto out; } if ((inp->inp_vflag & INP_IPV6PROTO) == 0) { error = EAFNOSUPPORT; goto out; } if (IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) { error = EAFNOSUPPORT; goto out; } if (IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) { #ifdef INET if ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0) { error = EINVAL; goto out; } if ((inp->inp_vflag & INP_IPV4) == 0) { error = EAFNOSUPPORT; 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; goto out; } if ((error = prison_remote_ip4(td->td_ucred, &sinp->sin_addr))) goto out; isipv6 = 0; #else /* !INET */ error = EAFNOSUPPORT; goto out; #endif /* INET */ } else { if ((inp->inp_vflag & INP_IPV6) == 0) { 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))) goto out; isipv6 = 1; } break; } #endif /* INET6 */ default: error = EAFNOSUPPORT; goto out; } } if (!(flags & PRUS_OOB)) { if (tp->t_acktime == 0) tp->t_acktime = ticks; sbappendstream(&so->so_snd, m, flags); m = NULL; if (nam && tp->t_state < TCPS_SYN_SENT) { KASSERT(tp->t_state == TCPS_CLOSED, ("%s: tp %p is listening", __func__, tp)); /* * 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) { /* m is freed if PRUS_NOTREADY is unset. */ sbflush(&so->so_snd); 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 = tcp_output_nodrop(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); 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. */ if (tp->t_acktime == 0) tp->t_acktime = ticks; sbappendstream_locked(&so->so_snd, m, flags); SOCKBUF_UNLOCK(&so->so_snd); m = NULL; 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 != 0) { /* m is freed if PRUS_NOTREADY is unset. */ sbflush(&so->so_snd); 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) == 0) { tp->t_flags |= TF_FORCEDATA; error = tcp_output_nodrop(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: /* * In case of PRUS_NOTREADY, the caller or tcp_usr_ready() is * responsible for freeing memory. */ if (m != NULL && (flags & PRUS_NOTREADY) == 0) m_freem(m); /* * 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)); error = tcp_unlock_or_drop(tp, error); 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_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) { INP_WUNLOCK(inp); return (error); } NET_EPOCH_ENTER(et); error = tcp_output_unlock(tp); NET_EPOCH_EXIT(et); 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_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)) { soref(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_DROPPED)) { tp = intotcpcb(inp); tp->t_flags |= TF_CLOSED; TCPDEBUG1(); tcp_disconnect(tp); TCPDEBUG2(PRU_CLOSE); TCP_PROBE2(debug__user, tp, PRU_CLOSE); } if (!(inp->inp_flags & INP_DROPPED)) { soref(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_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 protosw tcp_protosw = { .pr_type = SOCK_STREAM, .pr_protocol = IPPROTO_TCP, .pr_flags = PR_CONNREQUIRED | PR_IMPLOPCL | PR_WANTRCVD | PR_CAPATTACH, .pr_ctloutput = tcp_ctloutput, .pr_abort = tcp_usr_abort, .pr_accept = tcp_usr_accept, .pr_attach = tcp_usr_attach, .pr_bind = tcp_usr_bind, .pr_connect = tcp_usr_connect, .pr_control = in_control, .pr_detach = tcp_usr_detach, .pr_disconnect = tcp_usr_disconnect, .pr_listen = tcp_usr_listen, .pr_peeraddr = in_getpeeraddr, .pr_rcvd = tcp_usr_rcvd, .pr_rcvoob = tcp_usr_rcvoob, .pr_send = tcp_usr_send, .pr_ready = tcp_usr_ready, .pr_shutdown = tcp_usr_shutdown, .pr_sockaddr = in_getsockaddr, .pr_sosetlabel = in_pcbsosetlabel, .pr_close = tcp_usr_close, }; #endif /* INET */ #ifdef INET6 struct protosw tcp6_protosw = { .pr_type = SOCK_STREAM, .pr_protocol = IPPROTO_TCP, .pr_flags = PR_CONNREQUIRED | PR_IMPLOPCL |PR_WANTRCVD | PR_CAPATTACH, .pr_ctloutput = tcp_ctloutput, .pr_abort = tcp_usr_abort, .pr_accept = tcp6_usr_accept, .pr_attach = tcp_usr_attach, .pr_bind = tcp6_usr_bind, .pr_connect = tcp6_usr_connect, .pr_control = in6_control, .pr_detach = tcp_usr_detach, .pr_disconnect = tcp_usr_disconnect, .pr_listen = tcp6_usr_listen, .pr_peeraddr = in6_mapped_peeraddr, .pr_rcvd = tcp_usr_rcvd, .pr_rcvoob = tcp_usr_rcvoob, .pr_send = tcp_usr_send, .pr_ready = tcp_usr_ready, .pr_shutdown = tcp_usr_shutdown, .pr_sockaddr = in6_mapped_sockaddr, .pr_sosetlabel = in_pcbsosetlabel, .pr_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 = tptoinpcb(tp), *oinp; struct socket *so = tptosocket(tp); 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 = tptoinpcb(tp); 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(tptoinpcb(tp)); 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 | TF2_ACE_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 /* * AccECN related counters. */ if ((tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT)) == (TF2_ECN_PERMIT | TF2_ACE_PERMIT)) /* * Internal counter starts at 5 for AccECN * but 0 for RFC3168 ECN. */ ti->tcpi_delivered_ce = tp->t_scep - 5; else ti->tcpi_delivered_ce = tp->t_scep; ti->tcpi_received_ce = tp->t_rcep; } /* * 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_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_set(struct inpcb *inp, struct sockopt *sopt) { struct socket *so = inp->inp_socket; struct tcpcb *tp = intotcpcb(inp); int error = 0; MPASS(sopt->sopt_dir == SOPT_SET); INP_WLOCK_ASSERT(inp); KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("inp_flags == %x", inp->inp_flags)); KASSERT(so != NULL, ("inp_socket == NULL")); if (sopt->sopt_level != IPPROTO_TCP) { INP_WUNLOCK(inp); #ifdef INET6 if (inp->inp_vflag & INP_IPV6PROTO) error = ip6_ctloutput(so, sopt); #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET error = ip_ctloutput(so, sopt); #endif /* * When an IP-level socket option affects TCP, pass control * down to stack tfb_tcp_ctloutput, otherwise return what * IP level returned. */ switch (sopt->sopt_level) { #ifdef INET6 case IPPROTO_IPV6: if ((inp->inp_vflag & INP_IPV6PROTO) == 0) return (error); switch (sopt->sopt_name) { case IPV6_TCLASS: /* Notify tcp stacks that care (e.g. RACK). */ break; case IPV6_USE_MIN_MTU: /* Update t_maxseg accordingly. */ break; default: return (error); } break; #endif #ifdef INET case IPPROTO_IP: switch (sopt->sopt_name) { case IP_TOS: inp->inp_ip_tos &= ~IPTOS_ECN_MASK; break; case IP_TTL: /* Notify tcp stacks that care (e.g. RACK). */ break; default: return (error); } break; #endif default: return (error); } INP_WLOCK(inp); if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } } else if (sopt->sopt_name == TCP_FUNCTION_BLK) { /* * Protect the TCP option TCP_FUNCTION_BLK so * that a sub-function can *never* overwrite this. */ struct tcp_function_set fsn; struct tcp_function_block *blk; INP_WUNLOCK(inp); error = sooptcopyin(sopt, &fsn, sizeof fsn, sizeof fsn); if (error) return (error); INP_WLOCK(inp); if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(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) { struct epoch_tracker et; /* * Tell the stack to cleanup with 0 i.e. * the tcb is not going away. */ NET_EPOCH_ENTER(et); (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); NET_EPOCH_EXIT(et); } #ifdef TCPHPTS /* Assure that we are not on any hpts */ tcp_hpts_remove(tptoinpcb(tp)); #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); } /* Pass in the INP locked, callee must unlock it. */ return (tp->t_fb->tfb_tcp_ctloutput(inp, sopt)); } static int tcp_ctloutput_get(struct inpcb *inp, struct sockopt *sopt) { struct socket *so = inp->inp_socket; struct tcpcb *tp = intotcpcb(inp); int error = 0; MPASS(sopt->sopt_dir == SOPT_GET); INP_WLOCK_ASSERT(inp); KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("inp_flags == %x", inp->inp_flags)); KASSERT(so != NULL, ("inp_socket == NULL")); if (sopt->sopt_level != IPPROTO_TCP) { INP_WUNLOCK(inp); #ifdef INET6 if (inp->inp_vflag & INP_IPV6PROTO) error = ip6_ctloutput(so, sopt); #endif /* INET6 */ #if defined(INET6) && defined(INET) else #endif #ifdef INET error = ip_ctloutput(so, sopt); #endif return (error); } if (((sopt->sopt_name == TCP_FUNCTION_BLK) || (sopt->sopt_name == TCP_FUNCTION_ALIAS))) { struct tcp_function_set fsn; if (sopt->sopt_name == TCP_FUNCTION_ALIAS) { memset(&fsn, 0, sizeof(fsn)); find_tcp_function_alias(tp->t_fb, &fsn); } else { 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, callee must unlock it. */ return (tp->t_fb->tfb_tcp_ctloutput(inp, sopt)); } int tcp_ctloutput(struct socket *so, struct sockopt *sopt) { struct inpcb *inp; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_ctloutput: inp == NULL")); INP_WLOCK(inp); if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } if (sopt->sopt_dir == SOPT_SET) return (tcp_ctloutput_set(inp, sopt)); else if (sopt->sopt_dir == SOPT_GET) return (tcp_ctloutput_get(inp, sopt)); else panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir); } /* * 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 extern struct cc_algo newreno_cc_algo; static int tcp_set_cc_mod(struct inpcb *inp, struct sockopt *sopt) { struct cc_algo *algo; void *ptr = NULL; struct tcpcb *tp; struct cc_var cc_mem; char buf[TCP_CA_NAME_MAX]; size_t mem_sz; int error; INP_WUNLOCK(inp); error = sooptcopyin(sopt, buf, TCP_CA_NAME_MAX - 1, 1); if (error) return(error); buf[sopt->sopt_valsize] = '\0'; CC_LIST_RLOCK(); STAILQ_FOREACH(algo, &cc_list, entries) { if (strncmp(buf, algo->name, TCP_CA_NAME_MAX) == 0) { if (algo->flags & CC_MODULE_BEING_REMOVED) { /* We can't "see" modules being unloaded */ continue; } break; } } if (algo == NULL) { CC_LIST_RUNLOCK(); return(ESRCH); } /* * With a reference the algorithm cannot be removed * so we hold a reference through the change process. */ cc_refer(algo); CC_LIST_RUNLOCK(); if (algo->cb_init != NULL) { /* We can now pre-get the memory for the CC */ mem_sz = (*algo->cc_data_sz)(); if (mem_sz == 0) { goto no_mem_needed; } ptr = malloc(mem_sz, M_CC_MEM, M_WAITOK); } else { no_mem_needed: mem_sz = 0; ptr = NULL; } /* * Make sure its all clean and zero and also get * back the inplock. */ memset(&cc_mem, 0, sizeof(cc_mem)); INP_WLOCK(inp); if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); if (ptr) free(ptr, M_CC_MEM); /* Release our temp reference */ CC_LIST_RLOCK(); cc_release(algo); CC_LIST_RUNLOCK(); return (ECONNRESET); } tp = intotcpcb(inp); if (ptr != NULL) memset(ptr, 0, mem_sz); cc_mem.ccvc.tcp = tp; /* * We once again hold a write lock over the tcb so it's * safe to do these things without ordering concerns. * Note here we init into stack memory. */ if (algo->cb_init != NULL) error = algo->cb_init(&cc_mem, ptr); else error = 0; /* * The CC algorithms, when given their memory * should not fail we could in theory have a * KASSERT here. */ if (error == 0) { /* * Touchdown, lets go ahead and move the * connection to the new CC module by * copying in the cc_mem after we call * the old ones cleanup (if any). */ if (CC_ALGO(tp)->cb_destroy != NULL) - CC_ALGO(tp)->cb_destroy(tp->ccv); + CC_ALGO(tp)->cb_destroy(&tp->t_ccv); /* Detach the old CC from the tcpcb */ cc_detach(tp); /* Copy in our temp memory that was inited */ - memcpy(tp->ccv, &cc_mem, sizeof(struct cc_var)); + memcpy(&tp->t_ccv, &cc_mem, sizeof(struct cc_var)); /* Now attach the new, which takes a reference */ cc_attach(tp, algo); /* Ok now are we where we have gotten past any conn_init? */ if (TCPS_HAVEESTABLISHED(tp->t_state) && (CC_ALGO(tp)->conn_init != NULL)) { /* Yep run the connection init for the new CC */ - CC_ALGO(tp)->conn_init(tp->ccv); + CC_ALGO(tp)->conn_init(&tp->t_ccv); } } else if (ptr) free(ptr, M_CC_MEM); INP_WUNLOCK(inp); /* Now lets release our temp reference */ CC_LIST_RLOCK(); cc_release(algo); CC_LIST_RUNLOCK(); return (error); } int tcp_default_ctloutput(struct inpcb *inp, struct sockopt *sopt) { struct tcpcb *tp = intotcpcb(inp); int error, opt, optval; u_int ui; struct tcp_info ti; #ifdef KERN_TLS struct tls_enable tls; struct socket *so = inp->inp_socket; #endif char *pbuf, buf[TCP_LOG_ID_LEN]; #ifdef STATS struct statsblob *sbp; #endif size_t len; INP_WLOCK_ASSERT(inp); KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("inp_flags == %x", inp->inp_flags)); KASSERT(inp->inp_socket != NULL, ("inp_socket == NULL")); switch (sopt->sopt_level) { #ifdef INET6 case IPPROTO_IPV6: MPASS(inp->inp_vflag & INP_IPV6PROTO); switch (sopt->sopt_name) { case IPV6_USE_MIN_MTU: tcp6_use_min_mtu(tp); /* FALLTHROUGH */ } INP_WUNLOCK(inp); return (0); #endif #ifdef INET case IPPROTO_IP: INP_WUNLOCK(inp); return (0); #endif } /* * 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); + error = CC_ALGO(tp)->ctl_output(&tp->t_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: INP_WUNLOCK(inp); if (!TCPMD5_ENABLED()) return (ENOPROTOOPT); error = TCPMD5_PCBCTL(inp, sopt); if (error) return (error); INP_WLOCK_RECHECK(inp); goto unlock_and_done; #endif /* IPSEC */ case TCP_NODELAY: case TCP_NOOPT: case TCP_LRD: 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; case TCP_LRD: opt = TF_LRD; 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 = tcp_output_nodrop(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: error = tcp_set_cc_mod(inp, sopt); 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_MAXUNACKTIME: 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_MAXUNACKTIME: tp->t_maxunacktime = ui; break; 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: INP_WUNLOCK(inp); if (!TCPMD5_ENABLED()) 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_MAXUNACKTIME: case TCP_KEEPIDLE: case TCP_KEEPINTVL: case TCP_KEEPINIT: case TCP_KEEPCNT: switch (sopt->sopt_name) { case TCP_MAXUNACKTIME: ui = TP_MAXUNACKTIME(tp) / hz; break; 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: error = ktls_get_tx_mode(so, &optval); INP_WUNLOCK(inp); if (error == 0) error = sooptcopyout(sopt, &optval, sizeof(optval)); break; case TCP_RXTLS_MODE: error = ktls_get_rx_mode(so, &optval); INP_WUNLOCK(inp); if (error == 0) error = sooptcopyout(sopt, &optval, sizeof(optval)); break; #endif case TCP_LRD: optval = tp->t_flags & TF_LRD; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof optval); break; 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 = tptoinpcb(tp); struct socket *so = tptosocket(tp); 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)) /* Ignore stack's drop request, we already at it. */ (void)tcp_output_nodrop(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(tptoinpcb(tp)); 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_acktime == 0) tp->t_acktime = ticks; if (tp->t_state >= TCPS_FIN_WAIT_2) { soisdisconnected(tptosocket(tp)); /* 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_PREVVALID) { db_printf("%sTF_PREVVALID", comma ? ", " : ""); comma = 1; } if (t_flags & TF_MORETOCOME) { db_printf("%sTF_MORETOCOME", comma ? ", " : ""); comma = 1; } if (t_flags & TF_SONOTCONN) { db_printf("%sTF_SONOTCONN", 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_WASCRECOVERY) { db_printf("%sTF_WASCRECOVERY", 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_PLPMTU_BLACKHOLE) { db_printf("%sTF2_PLPMTU_BLACKHOLE", comma ? ", " : ""); comma = 1; } if (t_flags2 & TF2_PLPMTU_PMTUD) { db_printf("%sTF2_PLPMTU_PMTUD", comma ? ", " : ""); comma = 1; } if (t_flags2 & TF2_PLPMTU_MAXSEGSNT) { db_printf("%sTF2_PLPMTU_MAXSEGSNT", comma ? ", " : ""); comma = 1; } if (t_flags2 & TF2_LOG_AUTO) { db_printf("%sTF2_LOG_AUTO", comma ? ", " : ""); comma = 1; } if (t_flags2 & TF2_DROP_AF_DATA) { db_printf("%sTF2_DROP_AF_DATA", comma ? ", " : ""); comma = 1; } if (t_flags2 & TF2_ECN_PERMIT) { db_printf("%sTF2_ECN_PERMIT", comma ? ", " : ""); comma = 1; } if (t_flags2 & TF2_ECN_SND_CWR) { db_printf("%sTF2_ECN_SND_CWR", comma ? ", " : ""); comma = 1; } if (t_flags2 & TF2_ECN_SND_ECE) { db_printf("%sTF2_ECN_SND_ECE", comma ? ", " : ""); comma = 1; } if (t_flags2 & TF2_ACE_PERMIT) { db_printf("%sTF2_ACE_PERMIT", comma ? ", " : ""); comma = 1; } if (t_flags2 & TF2_FBYTES_COMPLETE) { db_printf("%sTF2_FBYTES_COMPLETE", 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); + &tp->tt_rexmt, &tp->tt_persist, &tp->tt_keep); + db_printf("tt_2msl: %p tt_delack: %p\n", &tp->tt_2msl, + &tp->tt_delack); 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\n", tp->t_rttvar, tp->t_rxtshift, tp->t_rttmin); 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 46670a04e42c..3f18f0af39cd 100644 --- a/sys/netinet/tcp_var.h +++ b/sys/netinet/tcp_var.h @@ -1,1295 +1,1302 @@ /*- * 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 "opt_kern_tls.h" #include #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) +#include + /* 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 tcp_get_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. + * Tcp control block, one per tcp connection. */ struct tcpcb { - /* Cache line 1 */ - struct inpcb *t_inpcb; /* back pointer to internet pcb */ + struct inpcb t_inpcb; /* embedded protocol indepenent cb */ +#define t_start_zero t_fb +#define t_zero_size (sizeof(struct tcpcb) - \ + offsetof(struct tcpcb, t_start_zero)) struct tcp_function_block *t_fb;/* TCP function call block */ void *t_fb_ptr; /* Pointer to t_fb specific data */ + + struct callout tt_rexmt; /* retransmit timer */ + struct callout tt_persist; /* retransmit persistence */ + struct callout tt_keep; /* keepalive */ + struct callout tt_2msl; /* 2*msl TIME_WAIT timer */ + struct callout tt_delack; /* delayed ACK timer */ + uint32_t tt_flags; /* Timers flags */ + uint32_t tt_draincnt; /* Count being drained */ + 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 */ uint32_t ts_offset; /* our timestamp offset */ uint32_t rfbuf_ts; /* recv buffer autoscaling timestamp */ int rcv_numsacks; /* # distinct sack blks present */ u_int t_tsomax; /* TSO total burst length limit */ 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 */ uint32_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; /* total reassembly queue byte length */ 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 */ 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 first byte queued in */ u_int t_fbyte_out; /* ticks time 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 */ 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 */ + struct cc_algo *t_cc; /* congestion control algorithm */ + struct cc_var t_ccv; /* congestion control specific vars */ 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 */ uint32_t t_rcep; /* Number of received CE marked pkts */ uint32_t t_scep; /* Synced number of delivered CE pkts */ 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 */ uint32_t t_maxpeakrate; /* max peak rate set by user, bytes/s */ uint32_t t_sndtlppack; /* tail loss probe packets sent */ uint64_t t_sndtlpbyte; /* total tail loss probe bytes sent */ uint64_t t_sndbytes; /* total bytes sent */ uint64_t t_snd_rxt_bytes; /* total bytes retransmitted */ uint32_t t_dsack_bytes; /* dsack bytes received */ uint32_t t_dsack_tlp_bytes; /* dsack bytes received for TLPs sent */ uint32_t t_dsack_pack; /* dsack packets we have eceived */ /* TCP Fast Open */ uint8_t t_tfo_client_cookie_len; /* TFO client cookie length */ uint32_t t_end_info_status; /* Status flag of end info */ unsigned int *t_tfo_pending; /* TFO 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 +#ifdef TCP_HHOOK + struct osd t_osd; /* storage for Khelp module data */ +#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 */ #define TCP_FUNC_OUTPUT_CANDROP 0x02 /* tfb_tcp_output may ask tcp_drop */ /* * 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 inpcb *inp, struct sockopt *sopt); /* 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); void (*tfb_hwtls_change)(struct tcpcb *, int); int (*tfb_compute_pipe)(struct tcpcb *tp); 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); struct tcpcb * tcp_drop(struct tcpcb *, int); #ifdef _NETINET_IN_PCB_H_ -#define intotcpcb(inp) ((struct tcpcb *)(inp)->inp_ppcb) +#define intotcpcb(inp) __containerof((inp), struct tcpcb, t_inpcb) #define sototcpcb(so) intotcpcb(sotoinpcb(so)) -#define tptoinpcb(tp) tp->t_inpcb -#define tptosocket(tp) tp->t_inpcb->inp_socket +#define tptoinpcb(tp) (&(tp)->t_inpcb) +#define tptosocket(tp) (tp)->t_inpcb.inp_socket /* * tcp_output() * Handles tcp_drop request from advanced stacks and reports that inpcb is * gone with negative return code. * Drop in replacement for the default stack. */ static inline int tcp_output(struct tcpcb *tp) { struct inpcb *inp = tptoinpcb(tp); int rv; INP_WLOCK_ASSERT(inp); rv = tp->t_fb->tfb_tcp_output(tp); if (rv < 0) { KASSERT(tp->t_fb->tfb_flags & TCP_FUNC_OUTPUT_CANDROP, ("TCP stack %s requested tcp_drop(%p)", tp->t_fb->tfb_tcp_block_name, tp)); tp = tcp_drop(tp, -rv); if (tp) INP_WUNLOCK(inp); } return (rv); } /* * tcp_output_unlock() * Always returns unlocked, handles drop request from advanced stacks. * Always returns positive error code. */ static inline int tcp_output_unlock(struct tcpcb *tp) { struct inpcb *inp = tptoinpcb(tp); int rv; INP_WLOCK_ASSERT(inp); rv = tp->t_fb->tfb_tcp_output(tp); if (rv < 0) { KASSERT(tp->t_fb->tfb_flags & TCP_FUNC_OUTPUT_CANDROP, ("TCP stack %s requested tcp_drop(%p)", tp->t_fb->tfb_tcp_block_name, tp)); rv = -rv; tp = tcp_drop(tp, rv); if (tp) INP_WUNLOCK(inp); } else INP_WUNLOCK(inp); return (rv); } /* * tcp_output_nodrop() * Always returns locked. It is caller's responsibility to run tcp_drop()! * Useful in syscall implementations, when we want to perform some logging * and/or tracing with tcpcb before calling tcp_drop(). To be used with * tcp_unlock_or_drop() later. * * XXXGL: maybe don't allow stacks to return a drop request at certain * TCP states? Why would it do in connect(2)? In recv(2)? */ static inline int tcp_output_nodrop(struct tcpcb *tp) { int rv; INP_WLOCK_ASSERT(tptoinpcb(tp)); rv = tp->t_fb->tfb_tcp_output(tp); KASSERT(rv >= 0 || tp->t_fb->tfb_flags & TCP_FUNC_OUTPUT_CANDROP, ("TCP stack %s requested tcp_drop(%p)", tp->t_fb->tfb_tcp_block_name, tp)); return (rv); } /* * tcp_unlock_or_drop() * Handle return code from tfb_tcp_output() after we have logged/traced, * to be used with tcp_output_nodrop(). */ static inline int tcp_unlock_or_drop(struct tcpcb *tp, int tcp_output_retval) { struct inpcb *inp = tptoinpcb(tp); INP_WLOCK_ASSERT(inp); if (tcp_output_retval < 0) { tcp_output_retval = -tcp_output_retval; if (tcp_drop(tp, tcp_output_retval) != NULL) INP_WUNLOCK(inp); } else INP_WUNLOCK(inp); return (tcp_output_retval); } #endif /* _NETINET_IN_PCB_H_ */ static int inline tcp_packets_this_ack(struct tcpcb *tp, tcp_seq ack) { return ((ack - tp->snd_una) / tp->t_maxseg + ((((ack - tp->snd_una) % tp->t_maxseg) != 0) ? 1 : 0)); } #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 * Note: accessing and restoring from * these may only be done in the 1st * RTO recovery round (t_rxtshift == 1) */ #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_SONOTCONN 0x00020000 /* needs soisconnected() on ESTAB */ #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_CLOSED 0x04000000 /* close(2) called on socket */ #define TF_UNUSED1 0x08000000 /* unused */ #define TF_LRD 0x10000000 /* Lost Retransmission Detection */ #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 */ #define TF2_ECN_USE_ECT1 0x00000800 /* Use ECT(1) marking on session */ /* * 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_ */ /* * 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_progdrops; /* drops due to no progress */ 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_lostrexmt; /* SACK lost retransmission recovered */ uint64_t tcps_sack_sboverflow; /* times scoreboard overflowed */ /* ECN related stats */ uint64_t tcps_ecn_rcvce; /* ECN Congestion Experienced */ uint64_t tcps_ecn_rcvect0; /* ECN Capable Transport */ uint64_t tcps_ecn_rcvect1; /* 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 tcps_tunneled_pkts; /* Packets encap's in UDP received */ uint64_t tcps_tunneled_errs; /* Packets that had errors that were UDP encaped */ /* Dsack related stats */ uint64_t tcps_dsack_count; /* Number of ACKs arriving with DSACKs */ uint64_t tcps_dsack_bytes; /* Number of bytes DSACK'ed no TLP */ uint64_t tcps_dsack_tlp_bytes; /* Number of bytes DSACK'ed due to TLPs */ /* TCPS_TIME_WAIT usage stats */ uint64_t tcps_tw_recycles; /* Times time-wait was recycled. */ uint64_t tcps_tw_resets; /* Times time-wait sent a reset. */ uint64_t tcps_tw_responds; /* Times time-wait sent a valid ack. */ /* Accurate ECN Handshake stats */ uint64_t tcps_ace_nect; /* ACE SYN packet with Non-ECT */ uint64_t tcps_ace_ect1; /* ACE SYN packet with ECT1 */ uint64_t tcps_ace_ect0; /* ACE SYN packet with ECT0 */ uint64_t tcps_ace_ce; /* ACE SYN packet with CE */ /* ECN related stats */ uint64_t tcps_ecn_sndect0; /* ECN Capable Transport */ uint64_t tcps_ecn_sndect1; /* ECN Capable Transport */ uint64_t _pad[4]; /* 4 TBD placeholder for STABLE */ }; #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) */ uint32_t t_dsack_bytes; /* (n) */ uint32_t t_dsack_tlp_bytes; /* (n) */ uint32_t t_dsack_pack; /* (n) */ uint16_t xt_encaps_port; /* (s) */ int16_t spare16; int32_t spare32[22]; } __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_lrd); 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 inpcbinfo, tcbinfo); #define V_tcp_do_lrd VNET(tcp_do_lrd) #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_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 *); struct tcpcb * tcp_close(struct tcpcb *); void tcp_discardcb(struct tcpcb *); bool tcp_freecb(struct tcpcb *); void tcp_twstart(struct tcpcb *); int tcp_ctloutput(struct socket *, struct sockopt *); void tcp_fini(void *); char *tcp_log_addrs(struct in_conninfo *, struct tcphdr *, const void *, const void *); char *tcp_log_vain(struct in_conninfo *, struct tcphdr *, const 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_ecnpkt_handler_flags(struct tcpcb *tp, uint16_t flags, 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_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); int find_tcp_function_alias(struct tcp_function_block *blk, 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 inpcb *inp, struct sockopt *sopt); int tcp_ctloutput_set(struct inpcb *inp, struct sockopt *sopt); 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_locks_taken; 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_bad_csums; #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 extern uint32_t tcp_ack_war_time_window; extern uint32_t tcp_ack_war_cnt; uint32_t tcp_maxmtu(struct in_conninfo *, struct tcp_ifcap *); uint32_t tcp_maxmtu6(struct in_conninfo *, struct tcp_ifcap *); void tcp6_use_min_mtu(struct tcpcb *); u_int tcp_maxseg(const struct tcpcb *); u_int tcp_fixed_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 tcpcb * tcp_newtcpcb(struct inpcb *); int tcp_default_output(struct tcpcb *); void tcp_state_change(struct tcpcb *, int); void tcp_respond(struct tcpcb *, void *, struct tcphdr *, struct mbuf *, tcp_seq, tcp_seq, uint16_t); bool tcp_twcheck(struct inpcb *, struct tcpopt *, struct tcphdr *, struct mbuf *, int); void tcp_setpersist(struct tcpcb *); void tcp_record_dsack(struct tcpcb *tp, tcp_seq start, tcp_seq end, int tlp); struct tcptemp * tcpip_maketemplate(struct inpcb *); 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 *); void cc_after_idle(struct tcpcb *tp); extern struct protosw tcp_protosw; /* shared for TOE */ extern struct protosw tcp6_protosw; /* shared for TOE */ 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); int tcp_dsack_block_exists(struct tcpcb *); 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_lost_retransmission(struct tcpcb *, struct tcphdr *); void tcp_sack_partialack(struct tcpcb *, struct tcphdr *); void tcp_free_sackholes(struct tcpcb *tp); void tcp_sack_lost_retransmission(struct tcpcb *, struct tcphdr *); 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); int tcp_can_enable_pacing(void); void tcp_decrement_paced_conn(void); 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); } static inline uint16_t tcp_get_flags(const struct tcphdr *th) { return (((uint16_t)th->th_x2 << 8) | th->th_flags); } static inline void tcp_set_flags(struct tcphdr *th, uint16_t flags) { th->th_x2 = (flags >> 8) & 0x0f; th->th_flags = flags & 0xff; } static inline void tcp_account_for_send(struct tcpcb *tp, uint32_t len, uint8_t is_rxt, uint8_t is_tlp, int hw_tls) { if (is_tlp) { tp->t_sndtlppack++; tp->t_sndtlpbyte += len; } /* To get total bytes sent you must add t_snd_rxt_bytes to t_sndbytes */ if (is_rxt) tp->t_snd_rxt_bytes += len; else tp->t_sndbytes += len; #ifdef KERN_TLS if (hw_tls && is_rxt && len != 0) { uint64_t rexmit_percent = (1000ULL * tp->t_snd_rxt_bytes) / (10ULL * (tp->t_snd_rxt_bytes + tp->t_sndbytes)); if (rexmit_percent > ktls_ifnet_max_rexmit_pct) ktls_disable_ifnet(tp); } #endif } #endif /* _KERNEL */ #endif /* _NETINET_TCP_VAR_H_ */ diff --git a/sys/netinet6/in6_pcb.c b/sys/netinet6/in6_pcb.c index 3e829f00f51a..1f9bd1445479 100644 --- a/sys/netinet6/in6_pcb.c +++ b/sys/netinet6/in6_pcb.c @@ -1,1160 +1,1159 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project. * 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 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. * * $KAME: in6_pcb.c,v 1.31 2001/05/21 05:45:10 jinmei Exp $ */ /*- * Copyright (c) 1982, 1986, 1991, 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. * * @(#)in_pcb.c 8.2 (Berkeley) 1/4/94 */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_route.h" #include "opt_rss.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 -#include #include #include #include #include #include #include #include #include #include int in6_pcbsetport(struct in6_addr *laddr, struct inpcb *inp, struct ucred *cred) { struct socket *so = inp->inp_socket; u_int16_t lport = 0; int error, lookupflags = 0; #ifdef INVARIANTS struct inpcbinfo *pcbinfo = inp->inp_pcbinfo; #endif INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK_ASSERT(pcbinfo); error = prison_local_ip6(cred, laddr, ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0)); if (error) return(error); /* XXX: this is redundant when called from in6_pcbbind */ if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT|SO_REUSEPORT_LB)) == 0) lookupflags = INPLOOKUP_WILDCARD; inp->inp_flags |= INP_ANONPORT; error = in_pcb_lport(inp, NULL, &lport, cred, lookupflags); if (error != 0) return (error); inp->inp_lport = lport; if (in_pcbinshash(inp) != 0) { inp->in6p_laddr = in6addr_any; inp->inp_lport = 0; return (EAGAIN); } return (0); } int in6_pcbbind(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred) { struct socket *so = inp->inp_socket; struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)NULL; struct inpcbinfo *pcbinfo = inp->inp_pcbinfo; u_short lport = 0; int error, lookupflags = 0; int reuseport = (so->so_options & SO_REUSEPORT); /* * XXX: Maybe we could let SO_REUSEPORT_LB set SO_REUSEPORT bit here * so that we don't have to add to the (already messy) code below. */ int reuseport_lb = (so->so_options & SO_REUSEPORT_LB); INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK_ASSERT(pcbinfo); if (inp->inp_lport || !IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) return (EINVAL); if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT|SO_REUSEPORT_LB)) == 0) lookupflags = INPLOOKUP_WILDCARD; if (nam == NULL) { if ((error = prison_local_ip6(cred, &inp->in6p_laddr, ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0) return (error); } else { sin6 = (struct sockaddr_in6 *)nam; KASSERT(sin6->sin6_family == AF_INET6, ("%s: invalid address family for %p", __func__, sin6)); KASSERT(sin6->sin6_len == sizeof(*sin6), ("%s: invalid address length for %p", __func__, sin6)); if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0) return(error); if ((error = prison_local_ip6(cred, &sin6->sin6_addr, ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0) return (error); lport = sin6->sin6_port; if (IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) { /* * Treat SO_REUSEADDR as SO_REUSEPORT for multicast; * allow compepte duplication of binding if * SO_REUSEPORT is set, or if SO_REUSEADDR is set * and a multicast address is bound on both * new and duplicated sockets. */ if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT)) != 0) reuseport = SO_REUSEADDR|SO_REUSEPORT; /* * XXX: How to deal with SO_REUSEPORT_LB here? * Treat same as SO_REUSEPORT for now. */ if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT_LB)) != 0) reuseport_lb = SO_REUSEADDR|SO_REUSEPORT_LB; } else if (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) { struct epoch_tracker et; struct ifaddr *ifa; sin6->sin6_port = 0; /* yech... */ NET_EPOCH_ENTER(et); if ((ifa = ifa_ifwithaddr((struct sockaddr *)sin6)) == NULL && (inp->inp_flags & INP_BINDANY) == 0) { NET_EPOCH_EXIT(et); return (EADDRNOTAVAIL); } /* * XXX: bind to an anycast address might accidentally * cause sending a packet with anycast source address. * We should allow to bind to a deprecated address, since * the application dares to use it. */ if (ifa != NULL && ((struct in6_ifaddr *)ifa)->ia6_flags & (IN6_IFF_ANYCAST|IN6_IFF_NOTREADY|IN6_IFF_DETACHED)) { NET_EPOCH_EXIT(et); return (EADDRNOTAVAIL); } NET_EPOCH_EXIT(et); } if (lport) { struct inpcb *t; /* GROSS */ if (ntohs(lport) <= V_ipport_reservedhigh && ntohs(lport) >= V_ipport_reservedlow && priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT)) return (EACCES); if (!IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr) && priv_check_cred(inp->inp_cred, PRIV_NETINET_REUSEPORT) != 0) { t = in6_pcblookup_local(pcbinfo, &sin6->sin6_addr, lport, INPLOOKUP_WILDCARD, cred); if (t && ((inp->inp_flags2 & INP_BINDMULTI) == 0) && (so->so_type != SOCK_STREAM || IN6_IS_ADDR_UNSPECIFIED(&t->in6p_faddr)) && (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr) || !IN6_IS_ADDR_UNSPECIFIED(&t->in6p_laddr) || (t->inp_flags2 & INP_REUSEPORT) || (t->inp_flags2 & INP_REUSEPORT_LB) == 0) && (inp->inp_cred->cr_uid != t->inp_cred->cr_uid)) return (EADDRINUSE); /* * If the socket is a BINDMULTI socket, then * the credentials need to match and the * original socket also has to have been bound * with BINDMULTI. */ if (t && (! in_pcbbind_check_bindmulti(inp, t))) return (EADDRINUSE); #ifdef INET if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 && IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) { struct sockaddr_in sin; in6_sin6_2_sin(&sin, sin6); t = in_pcblookup_local(pcbinfo, sin.sin_addr, lport, INPLOOKUP_WILDCARD, cred); if (t && ((inp->inp_flags2 & INP_BINDMULTI) == 0) && (so->so_type != SOCK_STREAM || ntohl(t->inp_faddr.s_addr) == INADDR_ANY) && (inp->inp_cred->cr_uid != t->inp_cred->cr_uid)) return (EADDRINUSE); if (t && (! in_pcbbind_check_bindmulti(inp, t))) return (EADDRINUSE); } #endif } t = in6_pcblookup_local(pcbinfo, &sin6->sin6_addr, lport, lookupflags, cred); if (t && (reuseport & inp_so_options(t)) == 0 && (reuseport_lb & inp_so_options(t)) == 0) { return (EADDRINUSE); } #ifdef INET if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 && IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) { struct sockaddr_in sin; in6_sin6_2_sin(&sin, sin6); t = in_pcblookup_local(pcbinfo, sin.sin_addr, lport, lookupflags, cred); if (t && (reuseport & inp_so_options(t)) == 0 && (reuseport_lb & inp_so_options(t)) == 0 && (ntohl(t->inp_laddr.s_addr) != INADDR_ANY || (t->inp_vflag & INP_IPV6PROTO) != 0)) { return (EADDRINUSE); } } #endif } inp->in6p_laddr = sin6->sin6_addr; } if (lport == 0) { if ((error = in6_pcbsetport(&inp->in6p_laddr, inp, cred)) != 0) { /* Undo an address bind that may have occurred. */ inp->in6p_laddr = in6addr_any; return (error); } } else { inp->inp_lport = lport; if (in_pcbinshash(inp) != 0) { inp->in6p_laddr = in6addr_any; inp->inp_lport = 0; return (EAGAIN); } } return (0); } /* * Transform old in6_pcbconnect() into an inner subroutine for new * in6_pcbconnect(): Do some validity-checking on the remote * address (in mbuf 'nam') and then determine local host address * (i.e., which interface) to use to access that remote host. * * This preserves definition of in6_pcbconnect(), while supporting a * slightly different version for T/TCP. (This is more than * a bit of a kludge, but cleaning up the internal interfaces would * have forced minor changes in every protocol). */ static int in6_pcbladdr(struct inpcb *inp, struct sockaddr_in6 *sin6, struct in6_addr *plocal_addr6) { int error = 0; int scope_ambiguous = 0; struct in6_addr in6a; struct epoch_tracker et; INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo); /* XXXRW: why? */ if (sin6->sin6_port == 0) return (EADDRNOTAVAIL); if (sin6->sin6_scope_id == 0 && !V_ip6_use_defzone) scope_ambiguous = 1; if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0) return(error); if (!CK_STAILQ_EMPTY(&V_in6_ifaddrhead)) { /* * If the destination address is UNSPECIFIED addr, * use the loopback addr, e.g ::1. */ if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) sin6->sin6_addr = in6addr_loopback; } if ((error = prison_remote_ip6(inp->inp_cred, &sin6->sin6_addr)) != 0) return (error); NET_EPOCH_ENTER(et); error = in6_selectsrc_socket(sin6, inp->in6p_outputopts, inp, inp->inp_cred, scope_ambiguous, &in6a, NULL); NET_EPOCH_EXIT(et); if (error) return (error); /* * Do not update this earlier, in case we return with an error. * * XXX: this in6_selectsrc_socket result might replace the bound local * address with the address specified by setsockopt(IPV6_PKTINFO). * Is it the intended behavior? */ *plocal_addr6 = in6a; /* * Don't do pcblookup call here; return interface in * plocal_addr6 * and exit to caller, that will do the lookup. */ return (0); } /* * Outer subroutine: * Connect from a socket to a specified address. * Both address and port must be specified in argument sin. * If don't have a local address for this socket yet, * then pick one. */ int in6_pcbconnect_mbuf(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred, struct mbuf *m, bool rehash) { struct inpcbinfo *pcbinfo = inp->inp_pcbinfo; struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam; struct sockaddr_in6 laddr6; int error; KASSERT(sin6->sin6_family == AF_INET6, ("%s: invalid address family for %p", __func__, sin6)); KASSERT(sin6->sin6_len == sizeof(*sin6), ("%s: invalid address length for %p", __func__, sin6)); bzero(&laddr6, sizeof(laddr6)); laddr6.sin6_family = AF_INET6; INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK_ASSERT(pcbinfo); #ifdef ROUTE_MPATH if (CALC_FLOWID_OUTBOUND) { uint32_t hash_type, hash_val; hash_val = fib6_calc_software_hash(&inp->in6p_laddr, &sin6->sin6_addr, 0, sin6->sin6_port, inp->inp_socket->so_proto->pr_protocol, &hash_type); inp->inp_flowid = hash_val; inp->inp_flowtype = hash_type; } #endif /* * Call inner routine, to assign local interface address. * in6_pcbladdr() may automatically fill in sin6_scope_id. */ if ((error = in6_pcbladdr(inp, sin6, &laddr6.sin6_addr)) != 0) return (error); if (in6_pcblookup_hash_locked(pcbinfo, &sin6->sin6_addr, sin6->sin6_port, IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr) ? &laddr6.sin6_addr : &inp->in6p_laddr, inp->inp_lport, 0, NULL, M_NODOM) != NULL) { return (EADDRINUSE); } if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) { if (inp->inp_lport == 0) { /* * rehash was required to be true in the past for * this case; retain that convention. However, * we now call in_pcb_lport_dest rather than * in6_pcbbind; the former does not insert into * the hash table, the latter does. Change rehash * to false to do the in_pcbinshash below. */ KASSERT(rehash == true, ("Rehashing required for unbound inps")); rehash = false; error = in_pcb_lport_dest(inp, (struct sockaddr *) &laddr6, &inp->inp_lport, (struct sockaddr *) sin6, sin6->sin6_port, cred, INPLOOKUP_WILDCARD); if (error) return (error); } inp->in6p_laddr = laddr6.sin6_addr; } inp->in6p_faddr = sin6->sin6_addr; inp->inp_fport = sin6->sin6_port; /* update flowinfo - draft-itojun-ipv6-flowlabel-api-00 */ inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; if (inp->inp_flags & IN6P_AUTOFLOWLABEL) inp->inp_flow |= (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK); if (rehash) { in_pcbrehash(inp); } else { in_pcbinshash(inp); } return (0); } int in6_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred) { return (in6_pcbconnect_mbuf(inp, nam, cred, NULL, true)); } void in6_pcbdisconnect(struct inpcb *inp) { INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo); bzero((caddr_t)&inp->in6p_faddr, sizeof(inp->in6p_faddr)); inp->inp_fport = 0; /* clear flowinfo - draft-itojun-ipv6-flowlabel-api-00 */ inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; in_pcbrehash(inp); } struct sockaddr * in6_sockaddr(in_port_t port, struct in6_addr *addr_p) { struct sockaddr_in6 *sin6; sin6 = malloc(sizeof *sin6, M_SONAME, M_WAITOK); bzero(sin6, sizeof *sin6); sin6->sin6_family = AF_INET6; sin6->sin6_len = sizeof(*sin6); sin6->sin6_port = port; sin6->sin6_addr = *addr_p; (void)sa6_recoverscope(sin6); /* XXX: should catch errors */ return (struct sockaddr *)sin6; } struct sockaddr * in6_v4mapsin6_sockaddr(in_port_t port, struct in_addr *addr_p) { struct sockaddr_in sin; struct sockaddr_in6 *sin6_p; bzero(&sin, sizeof sin); sin.sin_family = AF_INET; sin.sin_len = sizeof(sin); sin.sin_port = port; sin.sin_addr = *addr_p; sin6_p = malloc(sizeof *sin6_p, M_SONAME, M_WAITOK); in6_sin_2_v4mapsin6(&sin, sin6_p); return (struct sockaddr *)sin6_p; } int in6_getsockaddr(struct socket *so, struct sockaddr **nam) { struct inpcb *inp; struct in6_addr addr; in_port_t port; inp = sotoinpcb(so); KASSERT(inp != NULL, ("in6_getsockaddr: inp == NULL")); INP_RLOCK(inp); port = inp->inp_lport; addr = inp->in6p_laddr; INP_RUNLOCK(inp); *nam = in6_sockaddr(port, &addr); return 0; } int in6_getpeeraddr(struct socket *so, struct sockaddr **nam) { struct inpcb *inp; struct in6_addr addr; in_port_t port; inp = sotoinpcb(so); KASSERT(inp != NULL, ("in6_getpeeraddr: inp == NULL")); INP_RLOCK(inp); port = inp->inp_fport; addr = inp->in6p_faddr; INP_RUNLOCK(inp); *nam = in6_sockaddr(port, &addr); return 0; } int in6_mapped_sockaddr(struct socket *so, struct sockaddr **nam) { struct inpcb *inp; int error; inp = sotoinpcb(so); KASSERT(inp != NULL, ("in6_mapped_sockaddr: inp == NULL")); #ifdef INET if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) { error = in_getsockaddr(so, nam); if (error == 0) in6_sin_2_v4mapsin6_in_sock(nam); } else #endif { /* scope issues will be handled in in6_getsockaddr(). */ error = in6_getsockaddr(so, nam); } return error; } int in6_mapped_peeraddr(struct socket *so, struct sockaddr **nam) { struct inpcb *inp; int error; inp = sotoinpcb(so); KASSERT(inp != NULL, ("in6_mapped_peeraddr: inp == NULL")); #ifdef INET if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) { error = in_getpeeraddr(so, nam); if (error == 0) in6_sin_2_v4mapsin6_in_sock(nam); } else #endif /* scope issues will be handled in in6_getpeeraddr(). */ error = in6_getpeeraddr(so, nam); return error; } /* * Pass some notification to all connections of a protocol * associated with address dst. The local address and/or port numbers * may be specified to limit the search. The "usual action" will be * taken, depending on the ctlinput cmd. The caller must filter any * cmds that are uninteresting (e.g., no error in the map). * Call the protocol specific routine (if any) to report * any errors for each matching socket. */ static bool inp_match6(const struct inpcb *inp, void *v __unused) { return ((inp->inp_vflag & INP_IPV6) != 0); } void in6_pcbnotify(struct inpcbinfo *pcbinfo, struct sockaddr_in6 *sa6_dst, u_int fport_arg, const struct sockaddr_in6 *src, u_int lport_arg, int errno, void *cmdarg, struct inpcb *(*notify)(struct inpcb *, int)) { struct inpcb_iterator inpi = INP_ITERATOR(pcbinfo, INPLOOKUP_WLOCKPCB, inp_match6, NULL); struct inpcb *inp; struct sockaddr_in6 sa6_src; u_short fport = fport_arg, lport = lport_arg; u_int32_t flowinfo; if (IN6_IS_ADDR_UNSPECIFIED(&sa6_dst->sin6_addr)) return; /* * note that src can be NULL when we get notify by local fragmentation. */ sa6_src = (src == NULL) ? sa6_any : *src; flowinfo = sa6_src.sin6_flowinfo; while ((inp = inp_next(&inpi)) != NULL) { INP_WLOCK_ASSERT(inp); /* * If the error designates a new path MTU for a destination * and the application (associated with this socket) wanted to * know the value, notify. * XXX: should we avoid to notify the value to TCP sockets? */ if (errno == EMSGSIZE && cmdarg != NULL) ip6_notify_pmtu(inp, sa6_dst, *(uint32_t *)cmdarg); /* * Detect if we should notify the error. If no source and * destination ports are specified, but non-zero flowinfo and * local address match, notify the error. This is the case * when the error is delivered with an encrypted buffer * by ESP. Otherwise, just compare addresses and ports * as usual. */ if (lport == 0 && fport == 0 && flowinfo && inp->inp_socket != NULL && flowinfo == (inp->inp_flow & IPV6_FLOWLABEL_MASK) && IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, &sa6_src.sin6_addr)) goto do_notify; else if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, &sa6_dst->sin6_addr) || inp->inp_socket == 0 || (lport && inp->inp_lport != lport) || (!IN6_IS_ADDR_UNSPECIFIED(&sa6_src.sin6_addr) && !IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, &sa6_src.sin6_addr)) || (fport && inp->inp_fport != fport)) { continue; } do_notify: if (notify) (*notify)(inp, errno); } } /* * Lookup a PCB based on the local address and port. Caller must hold the * hash lock. No inpcb locks or references are acquired. */ struct inpcb * in6_pcblookup_local(struct inpcbinfo *pcbinfo, struct in6_addr *laddr, u_short lport, int lookupflags, struct ucred *cred) { struct inpcb *inp; int matchwild = 3, wildcard; KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0, ("%s: invalid lookup flags %d", __func__, lookupflags)); INP_HASH_LOCK_ASSERT(pcbinfo); if ((lookupflags & INPLOOKUP_WILDCARD) == 0) { struct inpcbhead *head; /* * Look for an unconnected (wildcard foreign addr) PCB that * matches the local address and port we're looking for. */ head = &pcbinfo->ipi_hashbase[INP_PCBHASH_WILD(lport, pcbinfo->ipi_hashmask)]; CK_LIST_FOREACH(inp, head, inp_hash) { /* XXX inp locking */ if ((inp->inp_vflag & INP_IPV6) == 0) continue; if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) && IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) && inp->inp_lport == lport) { /* Found. */ if (prison_equal_ip6(cred->cr_prison, inp->inp_cred->cr_prison)) return (inp); } } /* * Not found. */ return (NULL); } else { struct inpcbporthead *porthash; struct inpcbport *phd; struct inpcb *match = NULL; /* * Best fit PCB lookup. * * First see if this local port is in use by looking on the * port hash list. */ porthash = &pcbinfo->ipi_porthashbase[INP_PCBPORTHASH(lport, pcbinfo->ipi_porthashmask)]; CK_LIST_FOREACH(phd, porthash, phd_hash) { if (phd->phd_port == lport) break; } if (phd != NULL) { /* * Port is in use by one or more PCBs. Look for best * fit. */ CK_LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) { wildcard = 0; if (!prison_equal_ip6(cred->cr_prison, inp->inp_cred->cr_prison)) continue; /* XXX inp locking */ if ((inp->inp_vflag & INP_IPV6) == 0) continue; if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) wildcard++; if (!IN6_IS_ADDR_UNSPECIFIED( &inp->in6p_laddr)) { if (IN6_IS_ADDR_UNSPECIFIED(laddr)) wildcard++; else if (!IN6_ARE_ADDR_EQUAL( &inp->in6p_laddr, laddr)) continue; } else { if (!IN6_IS_ADDR_UNSPECIFIED(laddr)) wildcard++; } if (wildcard < matchwild) { match = inp; matchwild = wildcard; if (matchwild == 0) break; } } } return (match); } } static bool in6_multi_match(const struct inpcb *inp, void *v __unused) { if ((inp->inp_vflag & INP_IPV6) && inp->in6p_moptions != NULL) return (true); else return (false); } void in6_pcbpurgeif0(struct inpcbinfo *pcbinfo, struct ifnet *ifp) { struct inpcb_iterator inpi = INP_ITERATOR(pcbinfo, INPLOOKUP_RLOCKPCB, in6_multi_match, NULL); struct inpcb *inp; struct in6_multi *inm; struct in6_mfilter *imf; struct ip6_moptions *im6o; IN6_MULTI_LOCK_ASSERT(); while ((inp = inp_next(&inpi)) != NULL) { INP_RLOCK_ASSERT(inp); im6o = inp->in6p_moptions; /* * Unselect the outgoing ifp for multicast if it * is being detached. */ if (im6o->im6o_multicast_ifp == ifp) im6o->im6o_multicast_ifp = NULL; /* * Drop multicast group membership if we joined * through the interface being detached. */ restart: IP6_MFILTER_FOREACH(imf, &im6o->im6o_head) { if ((inm = imf->im6f_in6m) == NULL) continue; if (inm->in6m_ifp != ifp) continue; ip6_mfilter_remove(&im6o->im6o_head, imf); in6_leavegroup_locked(inm, NULL); ip6_mfilter_free(imf); goto restart; } } } /* * Check for alternatives when higher level complains * about service problems. For now, invalidate cached * routing information. If the route was created dynamically * (by a redirect), time to try a default gateway again. */ void in6_losing(struct inpcb *inp) { RO_INVALIDATE_CACHE(&inp->inp_route6); } /* * After a routing change, flush old routing * and allocate a (hopefully) better one. */ struct inpcb * in6_rtchange(struct inpcb *inp, int errno __unused) { RO_INVALIDATE_CACHE(&inp->inp_route6); return inp; } static bool in6_pcblookup_lb_numa_match(const struct inpcblbgroup *grp, int domain) { return (domain == M_NODOM || domain == grp->il_numa_domain); } static struct inpcb * in6_pcblookup_lbgroup(const struct inpcbinfo *pcbinfo, const struct in6_addr *laddr, uint16_t lport, const struct in6_addr *faddr, uint16_t fport, int lookupflags, uint8_t domain) { const struct inpcblbgrouphead *hdr; struct inpcblbgroup *grp; struct inpcblbgroup *jail_exact, *jail_wild, *local_exact, *local_wild; INP_HASH_LOCK_ASSERT(pcbinfo); hdr = &pcbinfo->ipi_lbgrouphashbase[ INP_PCBPORTHASH(lport, pcbinfo->ipi_lbgrouphashmask)]; /* * Search for an LB group match based on the following criteria: * - prefer jailed groups to non-jailed groups * - prefer exact source address matches to wildcard matches * - prefer groups bound to the specified NUMA domain */ jail_exact = jail_wild = local_exact = local_wild = NULL; CK_LIST_FOREACH(grp, hdr, il_list) { bool injail; #ifdef INET if (!(grp->il_vflag & INP_IPV6)) continue; #endif if (grp->il_lport != lport) continue; injail = prison_flag(grp->il_cred, PR_IP6) != 0; if (injail && prison_check_ip6_locked(grp->il_cred->cr_prison, laddr) != 0) continue; if (IN6_ARE_ADDR_EQUAL(&grp->il6_laddr, laddr)) { if (injail) { jail_exact = grp; if (in6_pcblookup_lb_numa_match(grp, domain)) /* This is a perfect match. */ goto out; } else if (local_exact == NULL || in6_pcblookup_lb_numa_match(grp, domain)) { local_exact = grp; } } else if (IN6_IS_ADDR_UNSPECIFIED(&grp->il6_laddr) && (lookupflags & INPLOOKUP_WILDCARD) != 0) { if (injail) { if (jail_wild == NULL || in6_pcblookup_lb_numa_match(grp, domain)) jail_wild = grp; } else if (local_wild == NULL || in6_pcblookup_lb_numa_match(grp, domain)) { local_wild = grp; } } } if (jail_exact != NULL) grp = jail_exact; else if (jail_wild != NULL) grp = jail_wild; else if (local_exact != NULL) grp = local_exact; else grp = local_wild; if (grp == NULL) return (NULL); out: return (grp->il_inp[INP6_PCBLBGROUP_PKTHASH(faddr, lport, fport) % grp->il_inpcnt]); } /* * Lookup PCB in hash list. Used in in_pcb.c as well as here. */ struct inpcb * in6_pcblookup_hash_locked(struct inpcbinfo *pcbinfo, struct in6_addr *faddr, u_int fport_arg, struct in6_addr *laddr, u_int lport_arg, int lookupflags, struct ifnet *ifp, uint8_t numa_domain) { struct inpcbhead *head; struct inpcb *inp, *tmpinp; u_short fport = fport_arg, lport = lport_arg; KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0, ("%s: invalid lookup flags %d", __func__, lookupflags)); INP_HASH_LOCK_ASSERT(pcbinfo); /* * First look for an exact match. */ tmpinp = NULL; head = &pcbinfo->ipi_hashbase[INP6_PCBHASH(faddr, lport, fport, pcbinfo->ipi_hashmask)]; CK_LIST_FOREACH(inp, head, inp_hash) { /* XXX inp locking */ if ((inp->inp_vflag & INP_IPV6) == 0) continue; if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) && IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) && inp->inp_fport == fport && inp->inp_lport == lport) { /* * XXX We should be able to directly return * the inp here, without any checks. * Well unless both bound with SO_REUSEPORT? */ if (prison_flag(inp->inp_cred, PR_IP6)) return (inp); if (tmpinp == NULL) tmpinp = inp; } } if (tmpinp != NULL) return (tmpinp); /* * Then look for a wildcard match, if requested. */ if ((lookupflags & INPLOOKUP_WILDCARD) != 0) { struct inpcb *local_wild = NULL, *local_exact = NULL; struct inpcb *jail_wild = NULL; int injail; /* * First see if an LB group matches the request before scanning * all sockets on this port. */ inp = in6_pcblookup_lbgroup(pcbinfo, laddr, lport, faddr, fport, lookupflags, numa_domain); if (inp != NULL) return (inp); /* * Order of socket selection - we always prefer jails. * 1. jailed, non-wild. * 2. jailed, wild. * 3. non-jailed, non-wild. * 4. non-jailed, wild. */ head = &pcbinfo->ipi_hashbase[INP_PCBHASH_WILD(lport, pcbinfo->ipi_hashmask)]; CK_LIST_FOREACH(inp, head, inp_hash) { /* XXX inp locking */ if ((inp->inp_vflag & INP_IPV6) == 0) continue; if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) || inp->inp_lport != lport) { continue; } injail = prison_flag(inp->inp_cred, PR_IP6); if (injail) { if (prison_check_ip6_locked( inp->inp_cred->cr_prison, laddr) != 0) continue; } else { if (local_exact != NULL) continue; } if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) { if (injail) return (inp); else local_exact = inp; } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) { if (injail) jail_wild = inp; else local_wild = inp; } } /* LIST_FOREACH */ if (jail_wild != NULL) return (jail_wild); if (local_exact != NULL) return (local_exact); if (local_wild != NULL) return (local_wild); } /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */ /* * Not found. */ return (NULL); } /* * Lookup PCB in hash list, using pcbinfo tables. This variation locks the * hash list lock, and will return the inpcb locked (i.e., requires * INPLOOKUP_LOCKPCB). */ static struct inpcb * in6_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in6_addr *faddr, u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags, struct ifnet *ifp, uint8_t numa_domain) { struct inpcb *inp; smr_enter(pcbinfo->ipi_smr); inp = in6_pcblookup_hash_locked(pcbinfo, faddr, fport, laddr, lport, lookupflags & INPLOOKUP_WILDCARD, ifp, numa_domain); if (inp != NULL) { if (__predict_false(inp_smr_lock(inp, (lookupflags & INPLOOKUP_LOCKMASK)) == false)) inp = NULL; } else smr_exit(pcbinfo->ipi_smr); return (inp); } /* * Public inpcb lookup routines, accepting a 4-tuple, and optionally, an mbuf * from which a pre-calculated hash value may be extracted. */ struct inpcb * in6_pcblookup(struct inpcbinfo *pcbinfo, struct in6_addr *faddr, u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags, struct ifnet *ifp) { KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0, ("%s: invalid lookup flags %d", __func__, lookupflags)); KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0, ("%s: LOCKPCB not set", __func__)); return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport, lookupflags, ifp, M_NODOM)); } struct inpcb * in6_pcblookup_mbuf(struct inpcbinfo *pcbinfo, struct in6_addr *faddr, u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags, struct ifnet *ifp, struct mbuf *m) { KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0, ("%s: invalid lookup flags %d", __func__, lookupflags)); KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0, ("%s: LOCKPCB not set", __func__)); return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport, lookupflags, ifp, m->m_pkthdr.numa_domain)); } void init_sin6(struct sockaddr_in6 *sin6, struct mbuf *m, int srcordst) { struct ip6_hdr *ip; ip = mtod(m, struct ip6_hdr *); bzero(sin6, sizeof(*sin6)); sin6->sin6_len = sizeof(*sin6); sin6->sin6_family = AF_INET6; sin6->sin6_addr = srcordst ? ip->ip6_dst : ip->ip6_src; (void)sa6_recoverscope(sin6); /* XXX: should catch errors... */ return; } diff --git a/sys/netipsec/ipsec_input.c b/sys/netipsec/ipsec_input.c index f6a1723617a4..9feb0df8bd15 100644 --- a/sys/netipsec/ipsec_input.c +++ b/sys/netipsec/ipsec_input.c @@ -1,775 +1,776 @@ /* $OpenBSD: ipsec_input.c,v 1.63 2003/02/20 18:35:43 deraadt Exp $ */ /*- * The authors of this code are John Ioannidis (ji@tla.org), * Angelos D. Keromytis (kermit@csd.uch.gr) and * Niels Provos (provos@physnet.uni-hamburg.de). * * This code was written by John Ioannidis for BSD/OS in Athens, Greece, * in November 1995. * * Ported to OpenBSD and NetBSD, with additional transforms, in December 1996, * by Angelos D. Keromytis. * * Additional transforms and features in 1997 and 1998 by Angelos D. Keromytis * and Niels Provos. * * Additional features in 1999 by Angelos D. Keromytis. * * Copyright (C) 1995, 1996, 1997, 1998, 1999 by John Ioannidis, * Angelos D. Keromytis and Niels Provos. * Copyright (c) 2001, Angelos D. Keromytis. * Copyright (c) 2016 Andrey V. Elsukov * * Permission to use, copy, and modify this software with or without fee * is hereby granted, provided that this entire notice is included in * all copies of any software which is or includes a copy or * modification of this software. * You may use this code under the GNU public license if you so wish. Please * contribute changes back to the authors under this freer than GPL license * so that we may further the use of strong encryption without limitations to * all. * * THIS SOFTWARE IS BEING PROVIDED "AS IS", WITHOUT ANY EXPRESS OR * IMPLIED WARRANTY. IN PARTICULAR, NONE OF THE AUTHORS MAKES ANY * REPRESENTATION OR WARRANTY OF ANY KIND CONCERNING THE * MERCHANTABILITY OF THIS SOFTWARE OR ITS FITNESS FOR ANY PARTICULAR * PURPOSE. */ /* * IPsec input processing. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.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 #ifdef INET6 #include #endif #include #ifdef INET6 #include #endif #include #ifdef INET6 #include #endif #include #include #include #include #include #include #include #include #include #include #include #define IPSEC_ISTAT(proto, name) do { \ if ((proto) == IPPROTO_ESP) \ ESPSTAT_INC(esps_##name); \ else if ((proto) == IPPROTO_AH) \ AHSTAT_INC(ahs_##name); \ else \ IPCOMPSTAT_INC(ipcomps_##name); \ } while (0) /* * ipsec_common_input gets called when an IPsec-protected packet * is received by IPv4 or IPv6. Its job is to find the right SA * and call the appropriate transform. The transform callback * takes care of further processing (like ingress filtering). */ static int ipsec_common_input(struct mbuf *m, int skip, int protoff, int af, int sproto) { IPSEC_DEBUG_DECLARE(char buf[IPSEC_ADDRSTRLEN]); union sockaddr_union dst_address; struct secasvar *sav; uint32_t spi; int error; IPSEC_ISTAT(sproto, input); IPSEC_ASSERT(m != NULL, ("null packet")); IPSEC_ASSERT(sproto == IPPROTO_ESP || sproto == IPPROTO_AH || sproto == IPPROTO_IPCOMP, ("unexpected security protocol %u", sproto)); if ((sproto == IPPROTO_ESP && !V_esp_enable) || (sproto == IPPROTO_AH && !V_ah_enable) || (sproto == IPPROTO_IPCOMP && !V_ipcomp_enable)) { m_freem(m); IPSEC_ISTAT(sproto, pdrops); return EOPNOTSUPP; } if (m->m_pkthdr.len - skip < 2 * sizeof (u_int32_t)) { m_freem(m); IPSEC_ISTAT(sproto, hdrops); DPRINTF(("%s: packet too small\n", __func__)); return EINVAL; } /* Retrieve the SPI from the relevant IPsec header */ if (sproto == IPPROTO_ESP) m_copydata(m, skip, sizeof(u_int32_t), (caddr_t) &spi); else if (sproto == IPPROTO_AH) m_copydata(m, skip + sizeof(u_int32_t), sizeof(u_int32_t), (caddr_t) &spi); else if (sproto == IPPROTO_IPCOMP) { u_int16_t cpi; m_copydata(m, skip + sizeof(u_int16_t), sizeof(u_int16_t), (caddr_t) &cpi); spi = ntohl(htons(cpi)); } /* * Find the SA and (indirectly) call the appropriate * kernel crypto routine. The resulting mbuf chain is a valid * IP packet ready to go through input processing. */ bzero(&dst_address, sizeof (dst_address)); dst_address.sa.sa_family = af; switch (af) { #ifdef INET case AF_INET: dst_address.sin.sin_len = sizeof(struct sockaddr_in); m_copydata(m, offsetof(struct ip, ip_dst), sizeof(struct in_addr), (caddr_t) &dst_address.sin.sin_addr); break; #endif /* INET */ #ifdef INET6 case AF_INET6: dst_address.sin6.sin6_len = sizeof(struct sockaddr_in6); m_copydata(m, offsetof(struct ip6_hdr, ip6_dst), sizeof(struct in6_addr), (caddr_t) &dst_address.sin6.sin6_addr); /* We keep addresses in SADB without embedded scope id */ if (IN6_IS_SCOPE_LINKLOCAL(&dst_address.sin6.sin6_addr)) { /* XXX: sa6_recoverscope() */ dst_address.sin6.sin6_scope_id = ntohs(dst_address.sin6.sin6_addr.s6_addr16[1]); dst_address.sin6.sin6_addr.s6_addr16[1] = 0; } break; #endif /* INET6 */ default: DPRINTF(("%s: unsupported protocol family %u\n", __func__, af)); m_freem(m); IPSEC_ISTAT(sproto, nopf); return EPFNOSUPPORT; } /* NB: only pass dst since key_allocsa follows RFC2401 */ sav = key_allocsa(&dst_address, sproto, spi); if (sav == NULL) { DPRINTF(("%s: no key association found for SA %s/%08lx/%u\n", __func__, ipsec_address(&dst_address, buf, sizeof(buf)), (u_long) ntohl(spi), sproto)); IPSEC_ISTAT(sproto, notdb); m_freem(m); return ENOENT; } if (sav->tdb_xform == NULL) { DPRINTF(("%s: attempted to use uninitialized SA %s/%08lx/%u\n", __func__, ipsec_address(&dst_address, buf, sizeof(buf)), (u_long) ntohl(spi), sproto)); IPSEC_ISTAT(sproto, noxform); key_freesav(&sav); m_freem(m); return ENXIO; } /* * Call appropriate transform and return -- callback takes care of * everything else. */ error = (*sav->tdb_xform->xf_input)(m, sav, skip, protoff); return (error); } #ifdef INET /* * IPSEC_INPUT() method implementation for IPv4. * 0 - Permitted by inbound security policy for further processing. * EACCES - Forbidden by inbound security policy. * EINPROGRESS - consumed by IPsec. */ int ipsec4_input(struct mbuf *m, int offset, int proto) { switch (proto) { case IPPROTO_AH: case IPPROTO_ESP: case IPPROTO_IPCOMP: /* Do inbound IPsec processing for AH/ESP/IPCOMP */ ipsec_common_input(m, offset, offsetof(struct ip, ip_p), AF_INET, proto); return (EINPROGRESS); /* mbuf consumed by IPsec */ default: /* * Protocols with further headers get their IPsec treatment * within the protocol specific processing. */ switch (proto) { case IPPROTO_ICMP: case IPPROTO_IGMP: case IPPROTO_IPV4: case IPPROTO_IPV6: case IPPROTO_RSVP: case IPPROTO_GRE: case IPPROTO_MOBILE: case IPPROTO_ETHERIP: case IPPROTO_PIM: case IPPROTO_SCTP: break; default: return (0); } }; /* * Enforce IPsec policy checking if we are seeing last header. */ if (ipsec4_in_reject(m, NULL) != 0) { /* Forbidden by inbound security policy */ m_freem(m); return (EACCES); } return (0); } int ipsec4_ctlinput(ipsec_ctlinput_param_t param) { struct icmp *icp = param.icmp; struct ip *ip = &icp->icmp_ip; struct sockaddr_in icmpsrc = { .sin_len = sizeof(struct sockaddr_in), .sin_family = AF_INET, .sin_addr = ip->ip_dst, }; struct in_conninfo inc; struct secasvar *sav; uint32_t pmtu, spi; uint32_t max_pmtu; uint8_t proto; pmtu = ntohs(icp->icmp_nextmtu); if (pmtu < V_ip4_ipsec_min_pmtu) return (EINVAL); proto = ip->ip_p; if (proto != IPPROTO_ESP && proto != IPPROTO_AH && proto != IPPROTO_IPCOMP) return (EINVAL); memcpy(&spi, (caddr_t)ip + (ip->ip_hl << 2), sizeof(spi)); sav = key_allocsa((union sockaddr_union *)&icmpsrc, proto, spi); if (sav == NULL) return (ENOENT); key_freesav(&sav); memset(&inc, 0, sizeof(inc)); inc.inc_faddr = ip->ip_dst; /* Update pmtu only if its smaller than the current one. */ max_pmtu = tcp_hc_getmtu(&inc); if (max_pmtu == 0) max_pmtu = tcp_maxmtu(&inc, NULL); if (pmtu < max_pmtu) tcp_hc_updatemtu(&inc, pmtu); return (0); } /* * IPsec input callback for INET protocols. * This routine is called as the transform callback. * Takes care of filtering and other sanity checks on * the processed packet. */ int ipsec4_common_input_cb(struct mbuf *m, struct secasvar *sav, int skip, int protoff) { IPSEC_DEBUG_DECLARE(char buf[IPSEC_ADDRSTRLEN]); struct epoch_tracker et; struct ipsec_ctx_data ctx; struct xform_history *xh; struct secasindex *saidx; struct m_tag *mtag; struct ip *ip; int error, prot, af, sproto, isr_prot; IPSEC_ASSERT(sav != NULL, ("null SA")); IPSEC_ASSERT(sav->sah != NULL, ("null SAH")); saidx = &sav->sah->saidx; af = saidx->dst.sa.sa_family; IPSEC_ASSERT(af == AF_INET, ("unexpected af %u", af)); sproto = saidx->proto; IPSEC_ASSERT(sproto == IPPROTO_ESP || sproto == IPPROTO_AH || sproto == IPPROTO_IPCOMP, ("unexpected security protocol %u", sproto)); if (skip != 0) { /* * Fix IPv4 header */ if (m->m_len < skip && (m = m_pullup(m, skip)) == NULL) { DPRINTF(("%s: processing failed for SA %s/%08lx\n", __func__, ipsec_address(&sav->sah->saidx.dst, buf, sizeof(buf)), (u_long) ntohl(sav->spi))); IPSEC_ISTAT(sproto, hdrops); error = ENOBUFS; goto bad_noepoch; } ip = mtod(m, struct ip *); ip->ip_len = htons(m->m_pkthdr.len); ip->ip_sum = 0; ip->ip_sum = in_cksum(m, ip->ip_hl << 2); } else { ip = mtod(m, struct ip *); } prot = ip->ip_p; /* * Check that we have NAT-T enabled and apply transport mode * decapsulation NAT procedure (RFC3948). * Do this before invoking into the PFIL. */ if (sav->natt != NULL && (prot == IPPROTO_UDP || prot == IPPROTO_TCP)) udp_ipsec_adjust_cksum(m, sav, prot, skip); /* * Needed for ipsec_run_hooks and netisr_queue_src */ NET_EPOCH_ENTER(et); IPSEC_INIT_CTX(&ctx, &m, NULL, sav, AF_INET, IPSEC_ENC_BEFORE); if ((error = ipsec_run_hhooks(&ctx, HHOOK_TYPE_IPSEC_IN)) != 0) goto bad; ip = mtod(m, struct ip *); /* update pointer */ /* IP-in-IP encapsulation */ if (prot == IPPROTO_IPIP && saidx->mode != IPSEC_MODE_TRANSPORT) { if (m->m_pkthdr.len - skip < sizeof(struct ip)) { IPSEC_ISTAT(sproto, hdrops); error = EINVAL; goto bad; } /* enc0: strip outer IPv4 header */ m_striphdr(m, 0, ip->ip_hl << 2); } #ifdef INET6 /* IPv6-in-IP encapsulation. */ else if (prot == IPPROTO_IPV6 && saidx->mode != IPSEC_MODE_TRANSPORT) { if (m->m_pkthdr.len - skip < sizeof(struct ip6_hdr)) { IPSEC_ISTAT(sproto, hdrops); error = EINVAL; goto bad; } /* enc0: strip IPv4 header, keep IPv6 header only */ m_striphdr(m, 0, ip->ip_hl << 2); } #endif /* INET6 */ else if (prot != IPPROTO_IPV6 && saidx->mode == IPSEC_MODE_ANY) { /* * When mode is wildcard, inner protocol is IPv6 and * we have no INET6 support - drop this packet a bit later. * In other cases we assume transport mode. Set prot to * correctly choose netisr. */ prot = IPPROTO_IPIP; } /* * Record what we've done to the packet (under what SA it was * processed). */ if (sproto != IPPROTO_IPCOMP) { mtag = m_tag_get(PACKET_TAG_IPSEC_IN_DONE, sizeof(struct xform_history), M_NOWAIT); if (mtag == NULL) { DPRINTF(("%s: failed to get tag\n", __func__)); IPSEC_ISTAT(sproto, hdrops); error = ENOMEM; goto bad; } xh = (struct xform_history *)(mtag + 1); bcopy(&saidx->dst, &xh->dst, saidx->dst.sa.sa_len); xh->spi = sav->spi; xh->proto = sproto; xh->mode = saidx->mode; m_tag_prepend(m, mtag); } key_sa_recordxfer(sav, m); /* record data transfer */ /* * In transport mode requeue decrypted mbuf back to IPv4 protocol * handler. This is necessary to correctly expose rcvif. */ if (saidx->mode == IPSEC_MODE_TRANSPORT) prot = IPPROTO_IPIP; /* * Re-dispatch via software interrupt. */ switch (prot) { case IPPROTO_IPIP: isr_prot = NETISR_IP; af = AF_INET; break; #ifdef INET6 case IPPROTO_IPV6: isr_prot = NETISR_IPV6; af = AF_INET6; break; #endif default: DPRINTF(("%s: cannot handle inner ip proto %d\n", __func__, prot)); IPSEC_ISTAT(sproto, nopf); error = EPFNOSUPPORT; goto bad; } IPSEC_INIT_CTX(&ctx, &m, NULL, sav, af, IPSEC_ENC_AFTER); if ((error = ipsec_run_hhooks(&ctx, HHOOK_TYPE_IPSEC_IN)) != 0) goto bad; /* Handle virtual tunneling interfaces */ if (saidx->mode == IPSEC_MODE_TUNNEL) error = ipsec_if_input(m, sav, af); if (error == 0) { error = netisr_queue_src(isr_prot, (uintptr_t)sav->spi, m); if (error) { IPSEC_ISTAT(sproto, qfull); DPRINTF(("%s: queue full; proto %u packet dropped\n", __func__, sproto)); } } NET_EPOCH_EXIT(et); key_freesav(&sav); return (error); bad: NET_EPOCH_EXIT(et); bad_noepoch: key_freesav(&sav); if (m != NULL) m_freem(m); return (error); } #endif /* INET */ #ifdef INET6 static bool ipsec6_lasthdr(int proto) { switch (proto) { case IPPROTO_IPV4: case IPPROTO_IPV6: case IPPROTO_GRE: case IPPROTO_ICMPV6: case IPPROTO_ETHERIP: case IPPROTO_PIM: case IPPROTO_SCTP: return (true); default: return (false); }; } /* * IPSEC_INPUT() method implementation for IPv6. * 0 - Permitted by inbound security policy for further processing. * EACCES - Forbidden by inbound security policy. * EINPROGRESS - consumed by IPsec. */ int ipsec6_input(struct mbuf *m, int offset, int proto) { switch (proto) { case IPPROTO_AH: case IPPROTO_ESP: case IPPROTO_IPCOMP: /* Do inbound IPsec processing for AH/ESP/IPCOMP */ ipsec_common_input(m, offset, offsetof(struct ip6_hdr, ip6_nxt), AF_INET6, proto); return (EINPROGRESS); /* mbuf consumed by IPsec */ default: /* * Protocols with further headers get their IPsec treatment * within the protocol specific processing. */ if (!ipsec6_lasthdr(proto)) return (0); /* FALLTHROUGH */ }; /* * Enforce IPsec policy checking if we are seeing last header. */ if (ipsec6_in_reject(m, NULL) != 0) { /* Forbidden by inbound security policy */ m_freem(m); return (EACCES); } return (0); } int ipsec6_ctlinput(ipsec_ctlinput_param_t param) { return (0); } extern ipproto_input_t *ip6_protox[]; /* * IPsec input callback, called by the transform callback. Takes care of * filtering and other sanity checks on the processed packet. */ int ipsec6_common_input_cb(struct mbuf *m, struct secasvar *sav, int skip, int protoff) { IPSEC_DEBUG_DECLARE(char buf[IPSEC_ADDRSTRLEN]); struct epoch_tracker et; struct ipsec_ctx_data ctx; struct xform_history *xh; struct secasindex *saidx; struct ip6_hdr *ip6; struct m_tag *mtag; int prot, af, sproto; int nxt, isr_prot; int error, nest; uint8_t nxt8; IPSEC_ASSERT(sav != NULL, ("null SA")); IPSEC_ASSERT(sav->sah != NULL, ("null SAH")); saidx = &sav->sah->saidx; af = saidx->dst.sa.sa_family; IPSEC_ASSERT(af == AF_INET6, ("unexpected af %u", af)); sproto = saidx->proto; IPSEC_ASSERT(sproto == IPPROTO_ESP || sproto == IPPROTO_AH || sproto == IPPROTO_IPCOMP, ("unexpected security protocol %u", sproto)); NET_EPOCH_ENTER(et); /* Fix IPv6 header */ if (m->m_len < sizeof(struct ip6_hdr) && (m = m_pullup(m, sizeof(struct ip6_hdr))) == NULL) { DPRINTF(("%s: processing failed for SA %s/%08lx\n", __func__, ipsec_address(&sav->sah->saidx.dst, buf, sizeof(buf)), (u_long) ntohl(sav->spi))); IPSEC_ISTAT(sproto, hdrops); error = EACCES; goto bad; } IPSEC_INIT_CTX(&ctx, &m, NULL, sav, af, IPSEC_ENC_BEFORE); if ((error = ipsec_run_hhooks(&ctx, HHOOK_TYPE_IPSEC_IN)) != 0) goto bad; ip6 = mtod(m, struct ip6_hdr *); ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(struct ip6_hdr)); /* Save protocol */ m_copydata(m, protoff, 1, &nxt8); prot = nxt8; /* IPv6-in-IP encapsulation */ if (prot == IPPROTO_IPV6 && saidx->mode != IPSEC_MODE_TRANSPORT) { if (m->m_pkthdr.len - skip < sizeof(struct ip6_hdr)) { IPSEC_ISTAT(sproto, hdrops); error = EINVAL; goto bad; } /* ip6n will now contain the inner IPv6 header. */ m_striphdr(m, 0, skip); skip = 0; } #ifdef INET /* IP-in-IP encapsulation */ else if (prot == IPPROTO_IPIP && saidx->mode != IPSEC_MODE_TRANSPORT) { if (m->m_pkthdr.len - skip < sizeof(struct ip)) { IPSEC_ISTAT(sproto, hdrops); error = EINVAL; goto bad; } /* ipn will now contain the inner IPv4 header */ m_striphdr(m, 0, skip); skip = 0; } #endif /* INET */ else { prot = IPPROTO_IPV6; /* for correct BPF processing */ } /* * Record what we've done to the packet (under what SA it was * processed). */ if (sproto != IPPROTO_IPCOMP) { mtag = m_tag_get(PACKET_TAG_IPSEC_IN_DONE, sizeof(struct xform_history), M_NOWAIT); if (mtag == NULL) { DPRINTF(("%s: failed to get tag\n", __func__)); IPSEC_ISTAT(sproto, hdrops); error = ENOMEM; goto bad; } xh = (struct xform_history *)(mtag + 1); bcopy(&saidx->dst, &xh->dst, saidx->dst.sa.sa_len); xh->spi = sav->spi; xh->proto = sproto; xh->mode = saidx->mode; m_tag_prepend(m, mtag); } key_sa_recordxfer(sav, m); #ifdef INET if (prot == IPPROTO_IPIP) af = AF_INET; else #endif af = AF_INET6; IPSEC_INIT_CTX(&ctx, &m, NULL, sav, af, IPSEC_ENC_AFTER); if ((error = ipsec_run_hhooks(&ctx, HHOOK_TYPE_IPSEC_IN)) != 0) goto bad; if (skip == 0) { /* * We stripped outer IPv6 header. * Now we should requeue decrypted packet via netisr. */ switch (prot) { #ifdef INET case IPPROTO_IPIP: isr_prot = NETISR_IP; break; #endif case IPPROTO_IPV6: isr_prot = NETISR_IPV6; break; default: DPRINTF(("%s: cannot handle inner ip proto %d\n", __func__, prot)); IPSEC_ISTAT(sproto, nopf); error = EPFNOSUPPORT; goto bad; } /* Handle virtual tunneling interfaces */ if (saidx->mode == IPSEC_MODE_TUNNEL) error = ipsec_if_input(m, sav, af); if (error == 0) { error = netisr_queue_src(isr_prot, (uintptr_t)sav->spi, m); if (error) { IPSEC_ISTAT(sproto, qfull); DPRINTF(("%s: queue full; proto %u packet" " dropped\n", __func__, sproto)); } } NET_EPOCH_EXIT(et); key_freesav(&sav); return (error); } /* * See the end of ip6_input for this logic. * IPPROTO_IPV[46] case will be processed just like other ones */ nest = 0; nxt = nxt8; while (nxt != IPPROTO_DONE) { if (V_ip6_hdrnestlimit && (++nest > V_ip6_hdrnestlimit)) { IP6STAT_INC(ip6s_toomanyhdr); error = EINVAL; goto bad; } /* * Protection against faulty packet - there should be * more sanity checks in header chain processing. */ if (m->m_pkthdr.len < skip) { IP6STAT_INC(ip6s_tooshort); in6_ifstat_inc(m->m_pkthdr.rcvif, ifs6_in_truncated); error = EINVAL; goto bad; } /* * Enforce IPsec policy checking if we are seeing last header. * note that we do not visit this with protocols with pcb layer * code - like udp/tcp/raw ip. */ if (ipsec6_lasthdr(nxt) && ipsec6_in_reject(m, NULL)) { error = EINVAL; goto bad; } nxt = ip6_protox[nxt](&m, &skip, nxt); } NET_EPOCH_EXIT(et); key_freesav(&sav); return (0); bad: NET_EPOCH_EXIT(et); key_freesav(&sav); if (m) m_freem(m); return (error); } #endif /* INET6 */ diff --git a/sys/netipsec/ipsec_output.c b/sys/netipsec/ipsec_output.c index ff0175e417cd..31a6e90ee283 100644 --- a/sys/netipsec/ipsec_output.c +++ b/sys/netipsec/ipsec_output.c @@ -1,1181 +1,1182 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2002, 2003 Sam Leffler, Errno Consulting * Copyright (c) 2016 Andrey V. Elsukov * 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. * * 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. * * $FreeBSD$ */ /* * IPsec output processing. */ #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_sctp.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include +#include #include #include #include #include #include #ifdef INET6 #include #endif #include #include #include #ifdef INET6 #include #include #endif #include #ifdef INET6 #include #endif #if defined(SCTP) || defined(SCTP_SUPPORT) #include #endif #include #include #include #include #ifdef INET6 #include #endif #include #include #include #include #include #include #include #include #include #define IPSEC_OSTAT_INC(proto, name) do { \ if ((proto) == IPPROTO_ESP) \ ESPSTAT_INC(esps_##name); \ else if ((proto) == IPPROTO_AH)\ AHSTAT_INC(ahs_##name); \ else \ IPCOMPSTAT_INC(ipcomps_##name); \ } while (0) static int ipsec_encap(struct mbuf **mp, struct secasindex *saidx); static size_t ipsec_get_pmtu(struct secasvar *sav); #ifdef INET static struct secasvar * ipsec4_allocsa(struct mbuf *m, struct secpolicy *sp, u_int *pidx, int *error) { struct secasindex *saidx, tmpsaidx; struct ipsecrequest *isr; struct sockaddr_in *sin; struct secasvar *sav; struct ip *ip; /* * Check system global policy controls. */ next: isr = sp->req[*pidx]; if ((isr->saidx.proto == IPPROTO_ESP && !V_esp_enable) || (isr->saidx.proto == IPPROTO_AH && !V_ah_enable) || (isr->saidx.proto == IPPROTO_IPCOMP && !V_ipcomp_enable)) { DPRINTF(("%s: IPsec outbound packet dropped due" " to policy (check your sysctls)\n", __func__)); IPSEC_OSTAT_INC(isr->saidx.proto, pdrops); *error = EHOSTUNREACH; return (NULL); } /* * Craft SA index to search for proper SA. Note that * we only initialize unspecified SA peers for transport * mode; for tunnel mode they must already be filled in. */ if (isr->saidx.mode == IPSEC_MODE_TRANSPORT) { saidx = &tmpsaidx; *saidx = isr->saidx; ip = mtod(m, struct ip *); if (saidx->src.sa.sa_len == 0) { sin = &saidx->src.sin; sin->sin_len = sizeof(*sin); sin->sin_family = AF_INET; sin->sin_port = IPSEC_PORT_ANY; sin->sin_addr = ip->ip_src; } if (saidx->dst.sa.sa_len == 0) { sin = &saidx->dst.sin; sin->sin_len = sizeof(*sin); sin->sin_family = AF_INET; sin->sin_port = IPSEC_PORT_ANY; sin->sin_addr = ip->ip_dst; } } else saidx = &sp->req[*pidx]->saidx; /* * Lookup SA and validate it. */ sav = key_allocsa_policy(sp, saidx, error); if (sav == NULL) { IPSECSTAT_INC(ips_out_nosa); if (*error != 0) return (NULL); if (ipsec_get_reqlevel(sp, *pidx) != IPSEC_LEVEL_REQUIRE) { /* * We have no SA and policy that doesn't require * this IPsec transform, thus we can continue w/o * IPsec processing, i.e. return EJUSTRETURN. * But first check if there is some bundled transform. */ if (sp->tcount > ++(*pidx)) goto next; *error = EJUSTRETURN; } return (NULL); } IPSEC_ASSERT(sav->tdb_xform != NULL, ("SA with NULL tdb_xform")); return (sav); } /* * IPsec output logic for IPv4. */ static int ipsec4_perform_request(struct mbuf *m, struct secpolicy *sp, struct inpcb *inp, u_int idx) { struct ipsec_ctx_data ctx; union sockaddr_union *dst; struct secasvar *sav; struct ip *ip; int error, i, off; IPSEC_ASSERT(idx < sp->tcount, ("Wrong IPsec request index %d", idx)); /* * We hold the reference to SP. Content of SP couldn't be changed. * Craft secasindex and do lookup for suitable SA. * Then do encapsulation if needed and call xform's output. * We need to store SP in the xform callback parameters. * In xform callback we will extract SP and it can be used to * determine next transform. At the end of transform we can * release reference to SP. */ sav = ipsec4_allocsa(m, sp, &idx, &error); if (sav == NULL) { if (error == EJUSTRETURN) { /* No IPsec required */ key_freesp(&sp); return (error); } goto bad; } /* * XXXAE: most likely ip_sum at this point is wrong. */ IPSEC_INIT_CTX(&ctx, &m, inp, sav, AF_INET, IPSEC_ENC_BEFORE); if ((error = ipsec_run_hhooks(&ctx, HHOOK_TYPE_IPSEC_OUT)) != 0) goto bad; ip = mtod(m, struct ip *); dst = &sav->sah->saidx.dst; /* Do the appropriate encapsulation, if necessary */ if (sp->req[idx]->saidx.mode == IPSEC_MODE_TUNNEL || /* Tunnel requ'd */ dst->sa.sa_family != AF_INET || /* PF mismatch */ (dst->sa.sa_family == AF_INET && /* Proxy */ dst->sin.sin_addr.s_addr != INADDR_ANY && dst->sin.sin_addr.s_addr != ip->ip_dst.s_addr)) { /* Fix IPv4 header checksum and length */ ip->ip_len = htons(m->m_pkthdr.len); ip->ip_sum = 0; ip->ip_sum = in_cksum(m, ip->ip_hl << 2); error = ipsec_encap(&m, &sav->sah->saidx); if (error != 0) { DPRINTF(("%s: encapsulation for SPI 0x%08x failed " "with error %d\n", __func__, ntohl(sav->spi), error)); /* XXXAE: IPSEC_OSTAT_INC(tunnel); */ goto bad; } inp = NULL; } IPSEC_INIT_CTX(&ctx, &m, inp, sav, dst->sa.sa_family, IPSEC_ENC_AFTER); if ((error = ipsec_run_hhooks(&ctx, HHOOK_TYPE_IPSEC_OUT)) != 0) goto bad; /* * Dispatch to the appropriate IPsec transform logic. The * packet will be returned for transmission after crypto * processing, etc. are completed. * * NB: m & sav are ``passed to caller'' who's responsible for * reclaiming their resources. */ switch(dst->sa.sa_family) { case AF_INET: ip = mtod(m, struct ip *); i = ip->ip_hl << 2; off = offsetof(struct ip, ip_p); break; #ifdef INET6 case AF_INET6: i = sizeof(struct ip6_hdr); off = offsetof(struct ip6_hdr, ip6_nxt); break; #endif /* INET6 */ default: DPRINTF(("%s: unsupported protocol family %u\n", __func__, dst->sa.sa_family)); error = EPFNOSUPPORT; IPSEC_OSTAT_INC(sav->sah->saidx.proto, nopf); goto bad; } error = (*sav->tdb_xform->xf_output)(m, sp, sav, idx, i, off); return (error); bad: IPSECSTAT_INC(ips_out_inval); if (m != NULL) m_freem(m); if (sav != NULL) key_freesav(&sav); key_freesp(&sp); return (error); } int ipsec4_process_packet(struct mbuf *m, struct secpolicy *sp, struct inpcb *inp) { return (ipsec4_perform_request(m, sp, inp, 0)); } int ipsec4_check_pmtu(struct mbuf *m, struct secpolicy *sp, int forwarding) { struct secasvar *sav; struct ip *ip; size_t hlen, pmtu; uint32_t idx; int error; /* Don't check PMTU if the frame won't have DF bit set. */ if (!V_ip4_ipsec_dfbit) return (0); if (V_ip4_ipsec_dfbit == 1) goto setdf; /* V_ip4_ipsec_dfbit > 1 - we will copy it from inner header. */ ip = mtod(m, struct ip *); if (!(ip->ip_off & htons(IP_DF))) return (0); setdf: idx = sp->tcount - 1; sav = ipsec4_allocsa(m, sp, &idx, &error); if (sav == NULL) { key_freesp(&sp); /* * No matching SA was found and SADB_ACQUIRE message was generated. * Since we have matched a SP to this packet drop it silently. */ if (error == 0) error = EINPROGRESS; if (error != EJUSTRETURN) m_freem(m); return (error); } pmtu = ipsec_get_pmtu(sav); if (pmtu == 0) { key_freesav(&sav); return (0); } hlen = ipsec_hdrsiz_internal(sp); key_freesav(&sav); if (m_length(m, NULL) + hlen > pmtu) { /* * If we're forwarding generate ICMP message here, * so that it contains pmtu subtracted by header size. * Set error to EINPROGRESS, in order for the frame * to be dropped silently. */ if (forwarding) { if (pmtu > hlen) icmp_error(m, ICMP_UNREACH, ICMP_UNREACH_NEEDFRAG, 0, pmtu - hlen); else m_freem(m); key_freesp(&sp); return (EINPROGRESS); /* Pretend that we consumed it. */ } else { m_freem(m); key_freesp(&sp); return (EMSGSIZE); } } return (0); } static int ipsec4_common_output(struct mbuf *m, struct inpcb *inp, int forwarding) { struct secpolicy *sp; int error; /* Lookup for the corresponding outbound security policy */ sp = ipsec4_checkpolicy(m, inp, &error, !forwarding); if (sp == NULL) { if (error == -EINVAL) { /* Discarded by policy. */ m_freem(m); return (EACCES); } return (0); /* No IPsec required. */ } /* * Usually we have to have tunnel mode IPsec security policy * when we are forwarding a packet. Otherwise we could not handle * encrypted replies, because they are not destined for us. But * some users are doing source address translation for forwarded * packets, and thus, even if they are forwarded, the replies will * return back to us. */ if (!forwarding) { /* * Do delayed checksums now because we send before * this is done in the normal processing path. */ if (m->m_pkthdr.csum_flags & CSUM_DELAY_DATA) { in_delayed_cksum(m); m->m_pkthdr.csum_flags &= ~CSUM_DELAY_DATA; } #if defined(SCTP) || defined(SCTP_SUPPORT) if (m->m_pkthdr.csum_flags & CSUM_SCTP) { struct ip *ip; ip = mtod(m, struct ip *); sctp_delayed_cksum(m, (uint32_t)(ip->ip_hl << 2)); m->m_pkthdr.csum_flags &= ~CSUM_SCTP; } #endif } /* NB: callee frees mbuf and releases reference to SP */ error = ipsec4_check_pmtu(m, sp, forwarding); if (error != 0) { if (error == EJUSTRETURN) return (0); return (error); } error = ipsec4_process_packet(m, sp, inp); if (error == EJUSTRETURN) { /* * We had a SP with a level of 'use' and no SA. We * will just continue to process the packet without * IPsec processing and return without error. */ return (0); } if (error == 0) return (EINPROGRESS); /* consumed by IPsec */ return (error); } /* * IPSEC_OUTPUT() method implementation for IPv4. * 0 - no IPsec handling needed * other values - mbuf consumed by IPsec. */ int ipsec4_output(struct mbuf *m, struct inpcb *inp) { /* * If the packet is resubmitted to ip_output (e.g. after * AH, ESP, etc. processing), there will be a tag to bypass * the lookup and related policy checking. */ if (m_tag_find(m, PACKET_TAG_IPSEC_OUT_DONE, NULL) != NULL) return (0); return (ipsec4_common_output(m, inp, 0)); } /* * IPSEC_FORWARD() method implementation for IPv4. * 0 - no IPsec handling needed * other values - mbuf consumed by IPsec. */ int ipsec4_forward(struct mbuf *m) { /* * Check if this packet has an active inbound SP and needs to be * dropped instead of forwarded. */ if (ipsec4_in_reject(m, NULL) != 0) { m_freem(m); return (EACCES); } return (ipsec4_common_output(m, NULL, 1)); } #endif #ifdef INET6 static int in6_sa_equal_addrwithscope(const struct sockaddr_in6 *sa, const struct in6_addr *ia) { struct in6_addr ia2; if (IN6_IS_SCOPE_LINKLOCAL(&sa->sin6_addr)) { memcpy(&ia2, &sa->sin6_addr, sizeof(ia2)); ia2.s6_addr16[1] = htons(sa->sin6_scope_id); return (IN6_ARE_ADDR_EQUAL(ia, &ia2)); } return (IN6_ARE_ADDR_EQUAL(&sa->sin6_addr, ia)); } static struct secasvar * ipsec6_allocsa(struct mbuf *m, struct secpolicy *sp, u_int *pidx, int *error) { struct secasindex *saidx, tmpsaidx; struct ipsecrequest *isr; struct sockaddr_in6 *sin6; struct secasvar *sav; struct ip6_hdr *ip6; /* * Check system global policy controls. */ next: isr = sp->req[*pidx]; if ((isr->saidx.proto == IPPROTO_ESP && !V_esp_enable) || (isr->saidx.proto == IPPROTO_AH && !V_ah_enable) || (isr->saidx.proto == IPPROTO_IPCOMP && !V_ipcomp_enable)) { DPRINTF(("%s: IPsec outbound packet dropped due" " to policy (check your sysctls)\n", __func__)); IPSEC_OSTAT_INC(isr->saidx.proto, pdrops); *error = EHOSTUNREACH; return (NULL); } /* * Craft SA index to search for proper SA. Note that * we only fillin unspecified SA peers for transport * mode; for tunnel mode they must already be filled in. */ if (isr->saidx.mode == IPSEC_MODE_TRANSPORT) { saidx = &tmpsaidx; *saidx = isr->saidx; ip6 = mtod(m, struct ip6_hdr *); if (saidx->src.sin6.sin6_len == 0) { sin6 = (struct sockaddr_in6 *)&saidx->src; sin6->sin6_len = sizeof(*sin6); sin6->sin6_family = AF_INET6; sin6->sin6_port = IPSEC_PORT_ANY; sin6->sin6_addr = ip6->ip6_src; if (IN6_IS_SCOPE_LINKLOCAL(&ip6->ip6_src)) { /* fix scope id for comparing SPD */ sin6->sin6_addr.s6_addr16[1] = 0; sin6->sin6_scope_id = ntohs(ip6->ip6_src.s6_addr16[1]); } } if (saidx->dst.sin6.sin6_len == 0) { sin6 = (struct sockaddr_in6 *)&saidx->dst; sin6->sin6_len = sizeof(*sin6); sin6->sin6_family = AF_INET6; sin6->sin6_port = IPSEC_PORT_ANY; sin6->sin6_addr = ip6->ip6_dst; if (IN6_IS_SCOPE_LINKLOCAL(&ip6->ip6_dst)) { /* fix scope id for comparing SPD */ sin6->sin6_addr.s6_addr16[1] = 0; sin6->sin6_scope_id = ntohs(ip6->ip6_dst.s6_addr16[1]); } } } else saidx = &sp->req[*pidx]->saidx; /* * Lookup SA and validate it. */ sav = key_allocsa_policy(sp, saidx, error); if (sav == NULL) { IPSEC6STAT_INC(ips_out_nosa); if (*error != 0) return (NULL); if (ipsec_get_reqlevel(sp, *pidx) != IPSEC_LEVEL_REQUIRE) { /* * We have no SA and policy that doesn't require * this IPsec transform, thus we can continue w/o * IPsec processing, i.e. return EJUSTRETURN. * But first check if there is some bundled transform. */ if (sp->tcount > ++(*pidx)) goto next; *error = EJUSTRETURN; } return (NULL); } IPSEC_ASSERT(sav->tdb_xform != NULL, ("SA with NULL tdb_xform")); return (sav); } /* * IPsec output logic for IPv6. */ static int ipsec6_perform_request(struct mbuf *m, struct secpolicy *sp, struct inpcb *inp, u_int idx) { struct ipsec_ctx_data ctx; union sockaddr_union *dst; struct secasvar *sav; struct ip6_hdr *ip6; int error, i, off; IPSEC_ASSERT(idx < sp->tcount, ("Wrong IPsec request index %d", idx)); sav = ipsec6_allocsa(m, sp, &idx, &error); if (sav == NULL) { if (error == EJUSTRETURN) { /* No IPsec required */ key_freesp(&sp); return (error); } goto bad; } /* Fix IP length in case if it is not set yet. */ ip6 = mtod(m, struct ip6_hdr *); ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); IPSEC_INIT_CTX(&ctx, &m, inp, sav, AF_INET6, IPSEC_ENC_BEFORE); if ((error = ipsec_run_hhooks(&ctx, HHOOK_TYPE_IPSEC_OUT)) != 0) goto bad; ip6 = mtod(m, struct ip6_hdr *); /* pfil can change mbuf */ dst = &sav->sah->saidx.dst; /* Do the appropriate encapsulation, if necessary */ if (sp->req[idx]->saidx.mode == IPSEC_MODE_TUNNEL || /* Tunnel requ'd */ dst->sa.sa_family != AF_INET6 || /* PF mismatch */ ((dst->sa.sa_family == AF_INET6) && (!IN6_IS_ADDR_UNSPECIFIED(&dst->sin6.sin6_addr)) && (!in6_sa_equal_addrwithscope(&dst->sin6, &ip6->ip6_dst)))) { if (m->m_pkthdr.len - sizeof(*ip6) > IPV6_MAXPACKET) { /* No jumbogram support. */ error = ENXIO; /*XXX*/ goto bad; } error = ipsec_encap(&m, &sav->sah->saidx); if (error != 0) { DPRINTF(("%s: encapsulation for SPI 0x%08x failed " "with error %d\n", __func__, ntohl(sav->spi), error)); /* XXXAE: IPSEC_OSTAT_INC(tunnel); */ goto bad; } inp = NULL; } IPSEC_INIT_CTX(&ctx, &m, inp, sav, dst->sa.sa_family, IPSEC_ENC_AFTER); if ((error = ipsec_run_hhooks(&ctx, HHOOK_TYPE_IPSEC_OUT)) != 0) goto bad; switch(dst->sa.sa_family) { #ifdef INET case AF_INET: { struct ip *ip; ip = mtod(m, struct ip *); i = ip->ip_hl << 2; off = offsetof(struct ip, ip_p); } break; #endif /* AF_INET */ case AF_INET6: i = sizeof(struct ip6_hdr); off = offsetof(struct ip6_hdr, ip6_nxt); break; default: DPRINTF(("%s: unsupported protocol family %u\n", __func__, dst->sa.sa_family)); error = EPFNOSUPPORT; IPSEC_OSTAT_INC(sav->sah->saidx.proto, nopf); goto bad; } error = (*sav->tdb_xform->xf_output)(m, sp, sav, idx, i, off); return (error); bad: IPSEC6STAT_INC(ips_out_inval); if (m != NULL) m_freem(m); if (sav != NULL) key_freesav(&sav); key_freesp(&sp); return (error); } int ipsec6_process_packet(struct mbuf *m, struct secpolicy *sp, struct inpcb *inp) { return (ipsec6_perform_request(m, sp, inp, 0)); } /* * IPv6 implementation is based on IPv4 implementation. */ int ipsec6_check_pmtu(struct mbuf *m, struct secpolicy *sp, int forwarding) { struct secasvar *sav; size_t hlen, pmtu; uint32_t idx; int error; /* * According to RFC8200 L3 fragmentation is supposed to be done only on * locally generated packets. During L3 forwarding packets that are too * big are always supposed to be dropped, with an ICMPv6 packet being * sent back. */ if (!forwarding) return (0); idx = sp->tcount - 1; sav = ipsec6_allocsa(m, sp, &idx, &error); if (sav == NULL) { key_freesp(&sp); /* * No matching SA was found and SADB_ACQUIRE message was generated. * Since we have matched a SP to this packet drop it silently. */ if (error == 0) error = EINPROGRESS; if (error != EJUSTRETURN) m_freem(m); return (error); } pmtu = ipsec_get_pmtu(sav); if (pmtu == 0) { key_freesav(&sav); return (0); } hlen = ipsec_hdrsiz_internal(sp); key_freesav(&sav); if (m_length(m, NULL) + hlen > pmtu) { /* * If we're forwarding generate ICMPv6 message here, * so that it contains pmtu subtracted by header size. * Set error to EINPROGRESS, in order for the frame * to be dropped silently. */ if (forwarding) { if (pmtu > hlen) icmp6_error(m, ICMP6_PACKET_TOO_BIG, 0, pmtu - hlen); else m_freem(m); key_freesp(&sp); return (EINPROGRESS); /* Pretend that we consumed it. */ } } return (0); } static int ipsec6_common_output(struct mbuf *m, struct inpcb *inp, int forwarding) { struct secpolicy *sp; int error; /* Lookup for the corresponding outbound security policy */ sp = ipsec6_checkpolicy(m, inp, &error, !forwarding); if (sp == NULL) { if (error == -EINVAL) { /* Discarded by policy. */ m_freem(m); return (EACCES); } return (0); /* No IPsec required. */ } if (!forwarding) { /* * Do delayed checksums now because we send before * this is done in the normal processing path. */ if (m->m_pkthdr.csum_flags & CSUM_DELAY_DATA_IPV6) { in6_delayed_cksum(m, m->m_pkthdr.len - sizeof(struct ip6_hdr), sizeof(struct ip6_hdr)); m->m_pkthdr.csum_flags &= ~CSUM_DELAY_DATA_IPV6; } #if defined(SCTP) || defined(SCTP_SUPPORT) if (m->m_pkthdr.csum_flags & CSUM_SCTP_IPV6) { sctp_delayed_cksum(m, sizeof(struct ip6_hdr)); m->m_pkthdr.csum_flags &= ~CSUM_SCTP_IPV6; } #endif } error = ipsec6_check_pmtu(m, sp, forwarding); if (error != 0) { if (error == EJUSTRETURN) return (0); return (error); } /* NB: callee frees mbuf and releases reference to SP */ error = ipsec6_process_packet(m, sp, inp); if (error == EJUSTRETURN) { /* * We had a SP with a level of 'use' and no SA. We * will just continue to process the packet without * IPsec processing and return without error. */ return (0); } if (error == 0) return (EINPROGRESS); /* consumed by IPsec */ return (error); } /* * IPSEC_OUTPUT() method implementation for IPv6. * 0 - no IPsec handling needed * other values - mbuf consumed by IPsec. */ int ipsec6_output(struct mbuf *m, struct inpcb *inp) { /* * If the packet is resubmitted to ip_output (e.g. after * AH, ESP, etc. processing), there will be a tag to bypass * the lookup and related policy checking. */ if (m_tag_find(m, PACKET_TAG_IPSEC_OUT_DONE, NULL) != NULL) return (0); return (ipsec6_common_output(m, inp, 0)); } /* * IPSEC_FORWARD() method implementation for IPv6. * 0 - no IPsec handling needed * other values - mbuf consumed by IPsec. */ int ipsec6_forward(struct mbuf *m) { /* * Check if this packet has an active inbound SP and needs to be * dropped instead of forwarded. */ if (ipsec6_in_reject(m, NULL) != 0) { m_freem(m); return (EACCES); } return (ipsec6_common_output(m, NULL, 1)); } #endif /* INET6 */ int ipsec_process_done(struct mbuf *m, struct secpolicy *sp, struct secasvar *sav, u_int idx) { struct epoch_tracker et; struct xform_history *xh; struct secasindex *saidx; struct m_tag *mtag; int error; saidx = &sav->sah->saidx; switch (saidx->dst.sa.sa_family) { #ifdef INET case AF_INET: /* Fix the header length, for AH processing. */ mtod(m, struct ip *)->ip_len = htons(m->m_pkthdr.len); break; #endif /* INET */ #ifdef INET6 case AF_INET6: /* Fix the header length, for AH processing. */ if (m->m_pkthdr.len < sizeof (struct ip6_hdr)) { error = ENXIO; goto bad; } if (m->m_pkthdr.len - sizeof (struct ip6_hdr) > IPV6_MAXPACKET) { /* No jumbogram support. */ error = ENXIO; /*?*/ goto bad; } mtod(m, struct ip6_hdr *)->ip6_plen = htons(m->m_pkthdr.len - sizeof(struct ip6_hdr)); break; #endif /* INET6 */ default: DPRINTF(("%s: unknown protocol family %u\n", __func__, saidx->dst.sa.sa_family)); error = ENXIO; goto bad; } /* * Add a record of what we've done to the packet. */ mtag = m_tag_get(PACKET_TAG_IPSEC_OUT_DONE, sizeof(*xh), M_NOWAIT); if (mtag == NULL) { DPRINTF(("%s: could not get packet tag\n", __func__)); error = ENOMEM; goto bad; } xh = (struct xform_history *)(mtag + 1); xh->dst = saidx->dst; xh->proto = saidx->proto; xh->mode = saidx->mode; xh->spi = sav->spi; m_tag_prepend(m, mtag); key_sa_recordxfer(sav, m); /* record data transfer */ /* * If there's another (bundled) SA to apply, do so. * Note that this puts a burden on the kernel stack size. * If this is a problem we'll need to introduce a queue * to set the packet on so we can unwind the stack before * doing further processing. */ if (++idx < sp->tcount) { switch (saidx->dst.sa.sa_family) { #ifdef INET case AF_INET: key_freesav(&sav); IPSECSTAT_INC(ips_out_bundlesa); return (ipsec4_perform_request(m, sp, NULL, idx)); /* NOTREACHED */ #endif #ifdef INET6 case AF_INET6: key_freesav(&sav); IPSEC6STAT_INC(ips_out_bundlesa); return (ipsec6_perform_request(m, sp, NULL, idx)); /* NOTREACHED */ #endif /* INET6 */ default: DPRINTF(("%s: unknown protocol family %u\n", __func__, saidx->dst.sa.sa_family)); error = EPFNOSUPPORT; goto bad; } } key_freesp(&sp), sp = NULL; /* Release reference to SP */ #ifdef INET /* * Do UDP encapsulation if SA requires it. */ if (sav->natt != NULL) { error = udp_ipsec_output(m, sav); if (error != 0) goto bad; } #endif /* INET */ /* * We're done with IPsec processing, transmit the packet using the * appropriate network protocol (IP or IPv6). */ NET_EPOCH_ENTER(et); switch (saidx->dst.sa.sa_family) { #ifdef INET case AF_INET: key_freesav(&sav); error = ip_output(m, NULL, NULL, IP_RAWOUTPUT, NULL, NULL); break; #endif /* INET */ #ifdef INET6 case AF_INET6: key_freesav(&sav); error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); break; #endif /* INET6 */ default: panic("ipsec_process_done"); } NET_EPOCH_EXIT(et); return (error); bad: m_freem(m); key_freesav(&sav); if (sp != NULL) key_freesp(&sp); return (error); } /* * ipsec_prepend() is optimized version of M_PREPEND(). * ipsec_encap() is called by IPsec output routine for tunnel mode SA. * It is expected that after IP encapsulation some IPsec transform will * be performed. Each IPsec transform inserts its variable length header * just after outer IP header using m_makespace(). If given mbuf has not * enough free space at the beginning, we allocate new mbuf and reserve * some space at the beginning and at the end. * This helps avoid allocating of new mbuf and data copying in m_makespace(), * we place outer header in the middle of mbuf's data with reserved leading * and trailing space: * [ LEADINGSPACE ][ Outer IP header ][ TRAILINGSPACE ] * LEADINGSPACE will be used to add ethernet header, TRAILINGSPACE will * be used to inject AH/ESP/IPCOMP header. */ #define IPSEC_TRAILINGSPACE (sizeof(struct udphdr) +/* NAT-T */ \ max(sizeof(struct newesp) + EALG_MAX_BLOCK_LEN, /* ESP + IV */ \ sizeof(struct newah) + HASH_MAX_LEN /* AH + ICV */)) static struct mbuf * ipsec_prepend(struct mbuf *m, int len, int how) { struct mbuf *n; M_ASSERTPKTHDR(m); IPSEC_ASSERT(len < MHLEN, ("wrong length")); if (M_LEADINGSPACE(m) >= len) { /* No need to allocate new mbuf. */ m->m_data -= len; m->m_len += len; m->m_pkthdr.len += len; return (m); } n = m_gethdr(how, m->m_type); if (n == NULL) { m_freem(m); return (NULL); } m_move_pkthdr(n, m); n->m_next = m; if (len + IPSEC_TRAILINGSPACE < M_SIZE(n)) m_align(n, len + IPSEC_TRAILINGSPACE); n->m_len = len; n->m_pkthdr.len += len; return (n); } static size_t ipsec_get_pmtu(struct secasvar *sav) { union sockaddr_union *dst; struct in_conninfo inc; size_t pmtu; dst = &sav->sah->saidx.dst; memset(&inc, 0, sizeof(inc)); switch (dst->sa.sa_family) { #ifdef INET case AF_INET: inc.inc_faddr = satosin(&dst->sa)->sin_addr; break; #endif #ifdef INET6 case AF_INET6: inc.inc6_faddr = satosin6(&dst->sa)->sin6_addr; inc.inc_flags |= INC_ISIPV6; break; #endif default: return (0); } pmtu = tcp_hc_getmtu(&inc); if (pmtu != 0) return (pmtu); /* No entry in hostcache. Assume that PMTU is equal to link's MTU */ switch (dst->sa.sa_family) { #ifdef INET case AF_INET: pmtu = tcp_maxmtu(&inc, NULL); break; #endif #ifdef INET6 case AF_INET6: pmtu = tcp_maxmtu6(&inc, NULL); break; #endif default: return (0); } if (pmtu == 0) return (0); tcp_hc_updatemtu(&inc, pmtu); return (pmtu); } static int ipsec_encap(struct mbuf **mp, struct secasindex *saidx) { #ifdef INET6 struct ip6_hdr *ip6; #endif struct ip *ip; #ifdef INET int setdf; #endif uint8_t itos, proto; ip = mtod(*mp, struct ip *); switch (ip->ip_v) { #ifdef INET case IPVERSION: proto = IPPROTO_IPIP; /* * Collect IP_DF state from the inner header * and honor system-wide control of how to handle it. */ switch (V_ip4_ipsec_dfbit) { case 0: /* clear in outer header */ case 1: /* set in outer header */ setdf = V_ip4_ipsec_dfbit; break; default:/* propagate to outer header */ setdf = (ip->ip_off & htons(IP_DF)) != 0; } itos = ip->ip_tos; break; #endif #ifdef INET6 case (IPV6_VERSION >> 4): proto = IPPROTO_IPV6; ip6 = mtod(*mp, struct ip6_hdr *); itos = (ntohl(ip6->ip6_flow) >> 20) & 0xff; /* scoped address handling */ in6_clearscope(&ip6->ip6_src); in6_clearscope(&ip6->ip6_dst); break; #endif default: return (EAFNOSUPPORT); } switch (saidx->dst.sa.sa_family) { #ifdef INET case AF_INET: if (saidx->src.sa.sa_family != AF_INET || saidx->src.sin.sin_addr.s_addr == INADDR_ANY || saidx->dst.sin.sin_addr.s_addr == INADDR_ANY) return (EINVAL); *mp = ipsec_prepend(*mp, sizeof(struct ip), M_NOWAIT); if (*mp == NULL) return (ENOBUFS); ip = mtod(*mp, struct ip *); ip->ip_v = IPVERSION; ip->ip_hl = sizeof(struct ip) >> 2; ip->ip_p = proto; ip->ip_len = htons((*mp)->m_pkthdr.len); ip->ip_ttl = V_ip_defttl; ip->ip_sum = 0; ip->ip_off = setdf ? htons(IP_DF): 0; ip->ip_src = saidx->src.sin.sin_addr; ip->ip_dst = saidx->dst.sin.sin_addr; ip_ecn_ingress(V_ip4_ipsec_ecn, &ip->ip_tos, &itos); ip_fillid(ip); break; #endif /* INET */ #ifdef INET6 case AF_INET6: if (saidx->src.sa.sa_family != AF_INET6 || IN6_IS_ADDR_UNSPECIFIED(&saidx->src.sin6.sin6_addr) || IN6_IS_ADDR_UNSPECIFIED(&saidx->dst.sin6.sin6_addr)) return (EINVAL); *mp = ipsec_prepend(*mp, sizeof(struct ip6_hdr), M_NOWAIT); if (*mp == NULL) return (ENOBUFS); ip6 = mtod(*mp, struct ip6_hdr *); ip6->ip6_flow = 0; ip6->ip6_vfc = IPV6_VERSION; ip6->ip6_hlim = V_ip6_defhlim; ip6->ip6_nxt = proto; ip6->ip6_dst = saidx->dst.sin6.sin6_addr; /* For link-local address embed scope zone id */ if (IN6_IS_SCOPE_LINKLOCAL(&ip6->ip6_dst)) ip6->ip6_dst.s6_addr16[1] = htons(saidx->dst.sin6.sin6_scope_id & 0xffff); ip6->ip6_src = saidx->src.sin6.sin6_addr; if (IN6_IS_SCOPE_LINKLOCAL(&ip6->ip6_src)) ip6->ip6_src.s6_addr16[1] = htons(saidx->src.sin6.sin6_scope_id & 0xffff); ip6->ip6_plen = htons((*mp)->m_pkthdr.len - sizeof(*ip6)); ip_ecn_ingress(V_ip6_ipsec_ecn, &proto, &itos); ip6->ip6_flow |= htonl((uint32_t)proto << 20); break; #endif /* INET6 */ default: return (EAFNOSUPPORT); } (*mp)->m_flags &= ~(M_BCAST | M_MCAST); return (0); } diff --git a/sys/netpfil/ipfilter/netinet/ip_fil_freebsd.c b/sys/netpfil/ipfilter/netinet/ip_fil_freebsd.c index 212e6e2af6a7..b2ee855c3854 100644 --- a/sys/netpfil/ipfilter/netinet/ip_fil_freebsd.c +++ b/sys/netpfil/ipfilter/netinet/ip_fil_freebsd.c @@ -1,1465 +1,1465 @@ /* $FreeBSD$ */ /* * Copyright (C) 2012 by Darren Reed. * * See the IPFILTER.LICENCE file for details on licencing. */ #if !defined(lint) static const char sccsid[] = "@(#)ip_fil.c 2.41 6/5/96 (C) 1993-2000 Darren Reed"; static const char rcsid[] = "@(#)$Id$"; #endif #if defined(KERNEL) || defined(_KERNEL) # undef KERNEL # undef _KERNEL # define KERNEL 1 # define _KERNEL 1 #endif #if defined(__FreeBSD__) && \ !defined(KLD_MODULE) && !defined(IPFILTER_LKM) # include "opt_inet6.h" #endif #include #include #include #include #include #include #include #include #include #include #include #if defined(__FreeBSD__) # include #endif #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 "netinet/ip_compat.h" #ifdef USE_INET6 # include #endif #include "netinet/ip_fil.h" #include "netinet/ip_nat.h" #include "netinet/ip_frag.h" #include "netinet/ip_state.h" #include "netinet/ip_proxy.h" #include "netinet/ip_auth.h" #include "netinet/ip_sync.h" #include "netinet/ip_lookup.h" #include "netinet/ip_dstlist.h" #ifdef IPFILTER_SCAN # include "netinet/ip_scan.h" #endif #include "netinet/ip_pool.h" #include #include #ifdef CSUM_DATA_VALID # include #endif extern int ip_optcopy(struct ip *, struct ip *); #ifdef IPFILTER_M_IPFILTER MALLOC_DEFINE(M_IPFILTER, "ipfilter", "IP Filter packet filter data structures"); #endif static int ipf_send_ip(fr_info_t *, mb_t *); static void ipf_timer_func(void *arg); VNET_DEFINE(ipf_main_softc_t, ipfmain) = { .ipf_running = -2, }; #define V_ipfmain VNET(ipfmain) #include #include VNET_DEFINE_STATIC(eventhandler_tag, ipf_arrivetag); VNET_DEFINE_STATIC(eventhandler_tag, ipf_departtag); #define V_ipf_arrivetag VNET(ipf_arrivetag) #define V_ipf_departtag VNET(ipf_departtag) #if 0 /* * Disable the "cloner" event handler; we are getting interface * events before the firewall is fully initiallized and also no vnet * information thus leading to uninitialised memory accesses. * In addition it is unclear why we need it in first place. * If it turns out to be needed, well need a dedicated event handler * for it to deal with the ifc and the correct vnet. */ VNET_DEFINE_STATIC(eventhandler_tag, ipf_clonetag); #define V_ipf_clonetag VNET(ipf_clonetag) #endif static void ipf_ifevent(void *arg, struct ifnet *ifp); static void ipf_ifevent(void *arg, struct ifnet *ifp) { CURVNET_SET(ifp->if_vnet); if (V_ipfmain.ipf_running > 0) ipf_sync(&V_ipfmain, NULL); CURVNET_RESTORE(); } static pfil_return_t ipf_check_wrapper(struct mbuf **mp, struct ifnet *ifp, int flags, void *ruleset __unused, struct inpcb *inp) { struct ip *ip = mtod(*mp, struct ip *); pfil_return_t rv; CURVNET_SET(ifp->if_vnet); rv = ipf_check(&V_ipfmain, ip, ip->ip_hl << 2, ifp, !!(flags & PFIL_OUT), mp); CURVNET_RESTORE(); return (rv == 0 ? PFIL_PASS : PFIL_DROPPED); } #ifdef USE_INET6 static pfil_return_t ipf_check_wrapper6(struct mbuf **mp, struct ifnet *ifp, int flags, void *ruleset __unused, struct inpcb *inp) { pfil_return_t rv; CURVNET_SET(ifp->if_vnet); rv = ipf_check(&V_ipfmain, mtod(*mp, struct ip *), sizeof(struct ip6_hdr), ifp, !!(flags & PFIL_OUT), mp); CURVNET_RESTORE(); return (rv == 0 ? PFIL_PASS : PFIL_DROPPED); } # endif #if defined(IPFILTER_LKM) int ipf_identify(char *s) { if (strcmp(s, "ipl") == 0) return (1); return (0); } #endif /* IPFILTER_LKM */ static void ipf_timer_func(void *arg) { ipf_main_softc_t *softc = arg; SPL_INT(s); SPL_NET(s); READ_ENTER(&softc->ipf_global); if (softc->ipf_running > 0) ipf_slowtimer(softc); if (softc->ipf_running == -1 || softc->ipf_running == 1) { #if 0 softc->ipf_slow_ch = timeout(ipf_timer_func, softc, hz/2); #endif callout_init(&softc->ipf_slow_ch, 1); callout_reset(&softc->ipf_slow_ch, (hz / IPF_HZ_DIVIDE) * IPF_HZ_MULT, ipf_timer_func, softc); } RWLOCK_EXIT(&softc->ipf_global); SPL_X(s); } int ipfattach(ipf_main_softc_t *softc) { #ifdef USE_SPL int s; #endif SPL_NET(s); if (softc->ipf_running > 0) { SPL_X(s); return (EBUSY); } if (ipf_init_all(softc) < 0) { SPL_X(s); return (EIO); } bzero((char *)V_ipfmain.ipf_selwait, sizeof(V_ipfmain.ipf_selwait)); softc->ipf_running = 1; if (softc->ipf_control_forwarding & 1) V_ipforwarding = 1; SPL_X(s); #if 0 softc->ipf_slow_ch = timeout(ipf_timer_func, softc, (hz / IPF_HZ_DIVIDE) * IPF_HZ_MULT); #endif callout_init(&softc->ipf_slow_ch, 1); callout_reset(&softc->ipf_slow_ch, (hz / IPF_HZ_DIVIDE) * IPF_HZ_MULT, ipf_timer_func, softc); return (0); } /* * Disable the filter by removing the hooks from the IP input/output * stream. */ int ipfdetach(ipf_main_softc_t *softc) { #ifdef USE_SPL int s; #endif if (softc->ipf_control_forwarding & 2) V_ipforwarding = 0; SPL_NET(s); #if 0 if (softc->ipf_slow_ch.callout != NULL) untimeout(ipf_timer_func, softc, softc->ipf_slow_ch); bzero(&softc->ipf_slow, sizeof(softc->ipf_slow)); #endif callout_drain(&softc->ipf_slow_ch); ipf_fini_all(softc); softc->ipf_running = -2; SPL_X(s); return (0); } /* * Filter ioctl interface. */ int ipfioctl(struct cdev *dev, ioctlcmd_t cmd, caddr_t data, int mode, struct thread *p) #define p_cred td_ucred #define p_uid td_ucred->cr_ruid { int error = 0, unit = 0; SPL_INT(s); CURVNET_SET(TD_TO_VNET(p)); if (securelevel_ge(p->p_cred, 3) && (mode & FWRITE)) { V_ipfmain.ipf_interror = 130001; CURVNET_RESTORE(); return (EPERM); } if (jailed_without_vnet(p->p_cred)) { V_ipfmain.ipf_interror = 130018; CURVNET_RESTORE(); return (EOPNOTSUPP); } unit = GET_MINOR(dev); if ((IPL_LOGMAX < unit) || (unit < 0)) { V_ipfmain.ipf_interror = 130002; CURVNET_RESTORE(); return (ENXIO); } if (V_ipfmain.ipf_running <= 0) { if (unit != IPL_LOGIPF && cmd != SIOCIPFINTERROR) { V_ipfmain.ipf_interror = 130003; CURVNET_RESTORE(); return (EIO); } if (cmd != SIOCIPFGETNEXT && cmd != SIOCIPFGET && cmd != SIOCIPFSET && cmd != SIOCFRENB && cmd != SIOCGETFS && cmd != SIOCGETFF && cmd != SIOCIPFINTERROR) { V_ipfmain.ipf_interror = 130004; CURVNET_RESTORE(); return (EIO); } } SPL_NET(s); error = ipf_ioctlswitch(&V_ipfmain, unit, data, cmd, mode, p->p_uid, p); CURVNET_RESTORE(); if (error != -1) { SPL_X(s); return (error); } SPL_X(s); return (error); } /* * ipf_send_reset - this could conceivably be a call to tcp_respond(), but that * requires a large amount of setting up and isn't any more efficient. */ int ipf_send_reset(fr_info_t *fin) { struct tcphdr *tcp, *tcp2; int tlen = 0, hlen; struct mbuf *m; #ifdef USE_INET6 ip6_t *ip6; #endif ip_t *ip; tcp = fin->fin_dp; if (tcp->th_flags & TH_RST) return (-1); /* feedback loop */ if (ipf_checkl4sum(fin) == -1) return (-1); tlen = fin->fin_dlen - (TCP_OFF(tcp) << 2) + ((tcp->th_flags & TH_SYN) ? 1 : 0) + ((tcp->th_flags & TH_FIN) ? 1 : 0); #ifdef USE_INET6 hlen = (fin->fin_v == 6) ? sizeof(ip6_t) : sizeof(ip_t); #else hlen = sizeof(ip_t); #endif #ifdef MGETHDR MGETHDR(m, M_NOWAIT, MT_HEADER); #else MGET(m, M_NOWAIT, MT_HEADER); #endif if (m == NULL) return (-1); if (sizeof(*tcp2) + hlen > MLEN) { if (!(MCLGET(m, M_NOWAIT))) { FREE_MB_T(m); return (-1); } } m->m_len = sizeof(*tcp2) + hlen; m->m_data += max_linkhdr; m->m_pkthdr.len = m->m_len; m->m_pkthdr.rcvif = (struct ifnet *)0; ip = mtod(m, struct ip *); bzero((char *)ip, hlen); #ifdef USE_INET6 ip6 = (ip6_t *)ip; #endif tcp2 = (struct tcphdr *)((char *)ip + hlen); tcp2->th_sport = tcp->th_dport; tcp2->th_dport = tcp->th_sport; if (tcp->th_flags & TH_ACK) { tcp2->th_seq = tcp->th_ack; tcp2->th_flags = TH_RST; tcp2->th_ack = 0; } else { tcp2->th_seq = 0; tcp2->th_ack = ntohl(tcp->th_seq); tcp2->th_ack += tlen; tcp2->th_ack = htonl(tcp2->th_ack); tcp2->th_flags = TH_RST|TH_ACK; } TCP_X2_A(tcp2, 0); TCP_OFF_A(tcp2, sizeof(*tcp2) >> 2); tcp2->th_win = tcp->th_win; tcp2->th_sum = 0; tcp2->th_urp = 0; #ifdef USE_INET6 if (fin->fin_v == 6) { ip6->ip6_flow = ((ip6_t *)fin->fin_ip)->ip6_flow; ip6->ip6_plen = htons(sizeof(struct tcphdr)); ip6->ip6_nxt = IPPROTO_TCP; ip6->ip6_hlim = 0; ip6->ip6_src = fin->fin_dst6.in6; ip6->ip6_dst = fin->fin_src6.in6; tcp2->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(*ip6), sizeof(*tcp2)); return (ipf_send_ip(fin, m)); } #endif ip->ip_p = IPPROTO_TCP; ip->ip_len = htons(sizeof(struct tcphdr)); ip->ip_src.s_addr = fin->fin_daddr; ip->ip_dst.s_addr = fin->fin_saddr; tcp2->th_sum = in_cksum(m, hlen + sizeof(*tcp2)); ip->ip_len = htons(hlen + sizeof(*tcp2)); return (ipf_send_ip(fin, m)); } /* * ip_len must be in network byte order when called. */ static int ipf_send_ip(fr_info_t *fin, mb_t *m) { fr_info_t fnew; ip_t *ip, *oip; int hlen; ip = mtod(m, ip_t *); bzero((char *)&fnew, sizeof(fnew)); fnew.fin_main_soft = fin->fin_main_soft; IP_V_A(ip, fin->fin_v); switch (fin->fin_v) { case 4 : oip = fin->fin_ip; hlen = sizeof(*oip); fnew.fin_v = 4; fnew.fin_p = ip->ip_p; fnew.fin_plen = ntohs(ip->ip_len); IP_HL_A(ip, sizeof(*oip) >> 2); ip->ip_tos = oip->ip_tos; ip->ip_id = fin->fin_ip->ip_id; ip->ip_off = htons(V_path_mtu_discovery ? IP_DF : 0); ip->ip_ttl = V_ip_defttl; ip->ip_sum = 0; break; #ifdef USE_INET6 case 6 : { ip6_t *ip6 = (ip6_t *)ip; ip6->ip6_vfc = 0x60; ip6->ip6_hlim = IPDEFTTL; hlen = sizeof(*ip6); fnew.fin_p = ip6->ip6_nxt; fnew.fin_v = 6; fnew.fin_plen = ntohs(ip6->ip6_plen) + hlen; break; } #endif default : return (EINVAL); } #ifdef IPSEC_SUPPORT m->m_pkthdr.rcvif = NULL; #endif fnew.fin_ifp = fin->fin_ifp; fnew.fin_flx = FI_NOCKSUM; fnew.fin_m = m; fnew.fin_ip = ip; fnew.fin_mp = &m; fnew.fin_hlen = hlen; fnew.fin_dp = (char *)ip + hlen; (void) ipf_makefrip(hlen, ip, &fnew); return (ipf_fastroute(m, &m, &fnew, NULL)); } int ipf_send_icmp_err(int type, fr_info_t *fin, int dst) { int err, hlen, xtra, iclen, ohlen, avail, code; struct in_addr dst4; struct icmp *icmp; struct mbuf *m; i6addr_t dst6; void *ifp; #ifdef USE_INET6 ip6_t *ip6; #endif ip_t *ip, *ip2; if ((type < 0) || (type >= ICMP_MAXTYPE)) return (-1); code = fin->fin_icode; #ifdef USE_INET6 /* See NetBSD ip_fil_netbsd.c r1.4: */ if ((code < 0) || (code >= sizeof(icmptoicmp6unreach)/sizeof(int))) return (-1); #endif if (ipf_checkl4sum(fin) == -1) return (-1); #ifdef MGETHDR MGETHDR(m, M_NOWAIT, MT_HEADER); #else MGET(m, M_NOWAIT, MT_HEADER); #endif if (m == NULL) return (-1); avail = MHLEN; xtra = 0; hlen = 0; ohlen = 0; dst4.s_addr = 0; ifp = fin->fin_ifp; if (fin->fin_v == 4) { if ((fin->fin_p == IPPROTO_ICMP) && !(fin->fin_flx & FI_SHORT)) switch (ntohs(fin->fin_data[0]) >> 8) { case ICMP_ECHO : case ICMP_TSTAMP : case ICMP_IREQ : case ICMP_MASKREQ : break; default : FREE_MB_T(m); return (0); } if (dst == 0) { if (ipf_ifpaddr(&V_ipfmain, 4, FRI_NORMAL, ifp, &dst6, NULL) == -1) { FREE_MB_T(m); return (-1); } dst4 = dst6.in4; } else dst4.s_addr = fin->fin_daddr; hlen = sizeof(ip_t); ohlen = fin->fin_hlen; iclen = hlen + offsetof(struct icmp, icmp_ip) + ohlen; if (fin->fin_hlen < fin->fin_plen) xtra = MIN(fin->fin_dlen, 8); else xtra = 0; } #ifdef USE_INET6 else if (fin->fin_v == 6) { hlen = sizeof(ip6_t); ohlen = sizeof(ip6_t); iclen = hlen + offsetof(struct icmp, icmp_ip) + ohlen; type = icmptoicmp6types[type]; if (type == ICMP6_DST_UNREACH) code = icmptoicmp6unreach[code]; if (iclen + max_linkhdr + fin->fin_plen > avail) { if (!(MCLGET(m, M_NOWAIT))) { FREE_MB_T(m); return (-1); } avail = MCLBYTES; } xtra = MIN(fin->fin_plen, avail - iclen - max_linkhdr); xtra = MIN(xtra, IPV6_MMTU - iclen); if (dst == 0) { if (ipf_ifpaddr(&V_ipfmain, 6, FRI_NORMAL, ifp, &dst6, NULL) == -1) { FREE_MB_T(m); return (-1); } } else dst6 = fin->fin_dst6; } #endif else { FREE_MB_T(m); return (-1); } avail -= (max_linkhdr + iclen); if (avail < 0) { FREE_MB_T(m); return (-1); } if (xtra > avail) xtra = avail; iclen += xtra; m->m_data += max_linkhdr; m->m_pkthdr.rcvif = (struct ifnet *)0; m->m_pkthdr.len = iclen; m->m_len = iclen; ip = mtod(m, ip_t *); icmp = (struct icmp *)((char *)ip + hlen); ip2 = (ip_t *)&icmp->icmp_ip; icmp->icmp_type = type; icmp->icmp_code = fin->fin_icode; icmp->icmp_cksum = 0; #ifdef icmp_nextmtu if (type == ICMP_UNREACH && fin->fin_icode == ICMP_UNREACH_NEEDFRAG) { if (fin->fin_mtu != 0) { icmp->icmp_nextmtu = htons(fin->fin_mtu); } else if (ifp != NULL) { icmp->icmp_nextmtu = htons(GETIFMTU_4(ifp)); } else { /* make up a number... */ icmp->icmp_nextmtu = htons(fin->fin_plen - 20); } } #endif bcopy((char *)fin->fin_ip, (char *)ip2, ohlen); #ifdef USE_INET6 ip6 = (ip6_t *)ip; if (fin->fin_v == 6) { ip6->ip6_flow = ((ip6_t *)fin->fin_ip)->ip6_flow; ip6->ip6_plen = htons(iclen - hlen); ip6->ip6_nxt = IPPROTO_ICMPV6; ip6->ip6_hlim = 0; ip6->ip6_src = dst6.in6; ip6->ip6_dst = fin->fin_src6.in6; if (xtra > 0) bcopy((char *)fin->fin_ip + ohlen, (char *)&icmp->icmp_ip + ohlen, xtra); icmp->icmp_cksum = in6_cksum(m, IPPROTO_ICMPV6, sizeof(*ip6), iclen - hlen); } else #endif { ip->ip_p = IPPROTO_ICMP; ip->ip_src.s_addr = dst4.s_addr; ip->ip_dst.s_addr = fin->fin_saddr; if (xtra > 0) bcopy((char *)fin->fin_ip + ohlen, (char *)&icmp->icmp_ip + ohlen, xtra); icmp->icmp_cksum = ipf_cksum((u_short *)icmp, sizeof(*icmp) + 8); ip->ip_len = htons(iclen); ip->ip_p = IPPROTO_ICMP; } err = ipf_send_ip(fin, m); return (err); } /* * m0 - pointer to mbuf where the IP packet starts * mpp - pointer to the mbuf pointer that is the start of the mbuf chain */ int ipf_fastroute(mb_t *m0, mb_t **mpp, fr_info_t *fin, frdest_t *fdp) { register struct ip *ip, *mhip; register struct mbuf *m = *mpp; int len, off, error = 0, hlen, code; struct ifnet *ifp, *sifp; struct route ro; struct sockaddr_in *dst; const struct sockaddr *gw; struct nhop_object *nh; u_long fibnum = 0; u_short ip_off; frdest_t node; frentry_t *fr; #ifdef M_WRITABLE /* * HOT FIX/KLUDGE: * * If the mbuf we're about to send is not writable (because of * a cluster reference, for example) we'll need to make a copy * of it since this routine modifies the contents. * * If you have non-crappy network hardware that can transmit data * from the mbuf, rather than making a copy, this is gonna be a * problem. */ if (M_WRITABLE(m) == 0) { m0 = m_dup(m, M_NOWAIT); if (m0 != NULL) { FREE_MB_T(m); m = m0; *mpp = m; } else { error = ENOBUFS; FREE_MB_T(m); goto done; } } #endif #ifdef USE_INET6 if (fin->fin_v == 6) { /* * currently "to " and "to :ip#" are not supported * for IPv6 */ return (ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL)); } #endif hlen = fin->fin_hlen; ip = mtod(m0, struct ip *); ifp = NULL; /* * Route packet. */ bzero(&ro, sizeof (ro)); dst = (struct sockaddr_in *)&ro.ro_dst; dst->sin_family = AF_INET; dst->sin_addr = ip->ip_dst; dst->sin_len = sizeof(dst); gw = (const struct sockaddr *)dst; fr = fin->fin_fr; if ((fr != NULL) && !(fr->fr_flags & FR_KEEPSTATE) && (fdp != NULL) && (fdp->fd_type == FRD_DSTLIST)) { if (ipf_dstlist_select_node(fin, fdp->fd_ptr, NULL, &node) == 0) fdp = &node; } if (fdp != NULL) ifp = fdp->fd_ptr; else ifp = fin->fin_ifp; if ((ifp == NULL) && ((fr == NULL) || !(fr->fr_flags & FR_FASTROUTE))) { error = -2; goto bad; } if ((fdp != NULL) && (fdp->fd_ip.s_addr != 0)) dst->sin_addr = fdp->fd_ip; fibnum = M_GETFIB(m0); NET_EPOCH_ASSERT(); nh = fib4_lookup(fibnum, dst->sin_addr, 0, NHR_NONE, 0); if (nh == NULL) { if (in_localaddr(ip->ip_dst)) error = EHOSTUNREACH; else error = ENETUNREACH; goto bad; } if (ifp == NULL) ifp = nh->nh_ifp; if (nh->nh_flags & NHF_GATEWAY) { gw = &nh->gw_sa; ro.ro_flags |= RT_HAS_GW; } /* * For input packets which are being "fastrouted", they won't * go back through output filtering and miss their chance to get * NAT'd and counted. Duplicated packets aren't considered to be * part of the normal packet stream, so do not NAT them or pass * them through stateful checking, etc. */ if ((fdp != &fr->fr_dif) && (fin->fin_out == 0)) { sifp = fin->fin_ifp; fin->fin_ifp = ifp; fin->fin_out = 1; (void) ipf_acctpkt(fin, NULL); fin->fin_fr = NULL; if (!fr || !(fr->fr_flags & FR_RETMASK)) { u_32_t pass; (void) ipf_state_check(fin, &pass); } switch (ipf_nat_checkout(fin, NULL)) { case 0 : break; case 1 : ip->ip_sum = 0; break; case -1 : error = -1; goto bad; break; } fin->fin_ifp = sifp; fin->fin_out = 0; } else ip->ip_sum = 0; /* * If small enough for interface, can just send directly. */ if (ntohs(ip->ip_len) <= ifp->if_mtu) { if (!ip->ip_sum) ip->ip_sum = in_cksum(m, hlen); error = (*ifp->if_output)(ifp, m, gw, &ro); goto done; } /* * Too large for interface; fragment if possible. * Must be able to put at least 8 bytes per fragment. */ ip_off = ntohs(ip->ip_off); if (ip_off & IP_DF) { error = EMSGSIZE; goto bad; } len = (ifp->if_mtu - hlen) &~ 7; if (len < 8) { error = EMSGSIZE; goto bad; } { int mhlen, firstlen = len; struct mbuf **mnext = &m->m_act; /* * Loop through length of segment after first fragment, * make new header and copy data of each part and link onto chain. */ m0 = m; mhlen = sizeof (struct ip); for (off = hlen + len; off < ntohs(ip->ip_len); off += len) { #ifdef MGETHDR MGETHDR(m, M_NOWAIT, MT_HEADER); #else MGET(m, M_NOWAIT, MT_HEADER); #endif if (m == NULL) { m = m0; error = ENOBUFS; goto bad; } m->m_data += max_linkhdr; mhip = mtod(m, struct ip *); bcopy((char *)ip, (char *)mhip, sizeof(*ip)); if (hlen > sizeof (struct ip)) { mhlen = ip_optcopy(ip, mhip) + sizeof (struct ip); IP_HL_A(mhip, mhlen >> 2); } m->m_len = mhlen; mhip->ip_off = ((off - hlen) >> 3) + ip_off; if (off + len >= ntohs(ip->ip_len)) len = ntohs(ip->ip_len) - off; else mhip->ip_off |= IP_MF; mhip->ip_len = htons((u_short)(len + mhlen)); *mnext = m; m->m_next = m_copym(m0, off, len, M_NOWAIT); if (m->m_next == 0) { error = ENOBUFS; /* ??? */ goto sendorfree; } m->m_pkthdr.len = mhlen + len; m->m_pkthdr.rcvif = NULL; mhip->ip_off = htons((u_short)mhip->ip_off); mhip->ip_sum = 0; mhip->ip_sum = in_cksum(m, mhlen); mnext = &m->m_act; } /* * Update first fragment by trimming what's been copied out * and updating header, then send each fragment (in order). */ m_adj(m0, hlen + firstlen - ip->ip_len); ip->ip_len = htons((u_short)(hlen + firstlen)); ip->ip_off = htons((u_short)IP_MF); ip->ip_sum = 0; ip->ip_sum = in_cksum(m0, hlen); sendorfree: for (m = m0; m; m = m0) { m0 = m->m_act; m->m_act = 0; if (error == 0) error = (*ifp->if_output)(ifp, m, gw, &ro); else FREE_MB_T(m); } } done: if (!error) V_ipfmain.ipf_frouteok[0]++; else V_ipfmain.ipf_frouteok[1]++; return (0); bad: if (error == EMSGSIZE) { sifp = fin->fin_ifp; code = fin->fin_icode; fin->fin_icode = ICMP_UNREACH_NEEDFRAG; fin->fin_ifp = ifp; (void) ipf_send_icmp_err(ICMP_UNREACH, fin, 1); fin->fin_ifp = sifp; fin->fin_icode = code; } FREE_MB_T(m); goto done; } int ipf_verifysrc(fin) fr_info_t *fin; { struct nhop_object *nh; NET_EPOCH_ASSERT(); nh = fib4_lookup(RT_DEFAULT_FIB, fin->fin_src, 0, NHR_NONE, 0); if (nh == NULL) return (0); return (fin->fin_ifp == nh->nh_ifp); } /* * return the first IP Address associated with an interface */ int ipf_ifpaddr(ipf_main_softc_t *softc, int v, int atype, void *ifptr, i6addr_t *inp, i6addr_t *inpmask) { #ifdef USE_INET6 struct in6_addr *ia6 = NULL; #endif struct sockaddr *sock, *mask; struct sockaddr_in *sin; struct ifaddr *ifa; struct ifnet *ifp; if ((ifptr == NULL) || (ifptr == (void *)-1)) return (-1); sin = NULL; ifp = ifptr; if (v == 4) inp->in4.s_addr = 0; #ifdef USE_INET6 else if (v == 6) bzero((char *)inp, sizeof(*inp)); #endif ifa = CK_STAILQ_FIRST(&ifp->if_addrhead); sock = ifa->ifa_addr; while (sock != NULL && ifa != NULL) { sin = (struct sockaddr_in *)sock; if ((v == 4) && (sin->sin_family == AF_INET)) break; #ifdef USE_INET6 if ((v == 6) && (sin->sin_family == AF_INET6)) { ia6 = &((struct sockaddr_in6 *)sin)->sin6_addr; if (!IN6_IS_ADDR_LINKLOCAL(ia6) && !IN6_IS_ADDR_LOOPBACK(ia6)) break; } #endif ifa = CK_STAILQ_NEXT(ifa, ifa_link); if (ifa != NULL) sock = ifa->ifa_addr; } if (ifa == NULL || sin == NULL) return (-1); mask = ifa->ifa_netmask; if (atype == FRI_BROADCAST) sock = ifa->ifa_broadaddr; else if (atype == FRI_PEERADDR) sock = ifa->ifa_dstaddr; if (sock == NULL) return (-1); #ifdef USE_INET6 if (v == 6) { return (ipf_ifpfillv6addr(atype, (struct sockaddr_in6 *)sock, (struct sockaddr_in6 *)mask, inp, inpmask)); } #endif return (ipf_ifpfillv4addr(atype, (struct sockaddr_in *)sock, (struct sockaddr_in *)mask, &inp->in4, &inpmask->in4)); } u_32_t ipf_newisn(fin) fr_info_t *fin; { u_32_t newiss; newiss = arc4random(); return (newiss); } int ipf_checkv4sum(fr_info_t *fin) { #ifdef CSUM_DATA_VALID int manual = 0; u_short sum; ip_t *ip; mb_t *m; if ((fin->fin_flx & FI_NOCKSUM) != 0) return (0); if ((fin->fin_flx & FI_SHORT) != 0) return (1); if (fin->fin_cksum != FI_CK_NEEDED) return (fin->fin_cksum > FI_CK_NEEDED) ? 0 : -1; m = fin->fin_m; if (m == NULL) { manual = 1; goto skipauto; } ip = fin->fin_ip; if ((m->m_pkthdr.csum_flags & (CSUM_IP_CHECKED|CSUM_IP_VALID)) == CSUM_IP_CHECKED) { fin->fin_cksum = FI_CK_BAD; fin->fin_flx |= FI_BAD; DT2(ipf_fi_bad_checkv4sum_csum_ip_checked, fr_info_t *, fin, u_int, m->m_pkthdr.csum_flags & (CSUM_IP_CHECKED|CSUM_IP_VALID)); return (-1); } if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) { /* Depending on the driver, UDP may have zero checksum */ if (fin->fin_p == IPPROTO_UDP && (fin->fin_flx & (FI_FRAG|FI_SHORT|FI_BAD)) == 0) { udphdr_t *udp = fin->fin_dp; if (udp->uh_sum == 0) { /* * we're good no matter what the hardware * checksum flags and csum_data say (handling * of csum_data for zero UDP checksum is not * consistent across all drivers) */ fin->fin_cksum = 1; return (0); } } if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) sum = m->m_pkthdr.csum_data; else sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htonl(m->m_pkthdr.csum_data + fin->fin_dlen + fin->fin_p)); sum ^= 0xffff; if (sum != 0) { fin->fin_cksum = FI_CK_BAD; fin->fin_flx |= FI_BAD; DT2(ipf_fi_bad_checkv4sum_sum, fr_info_t *, fin, u_int, sum); } else { fin->fin_cksum = FI_CK_SUMOK; return (0); } } else { if (m->m_pkthdr.csum_flags == CSUM_DELAY_DATA) { fin->fin_cksum = FI_CK_L4FULL; return (0); } else if (m->m_pkthdr.csum_flags == CSUM_TCP || m->m_pkthdr.csum_flags == CSUM_UDP || m->m_pkthdr.csum_flags == CSUM_IP) { fin->fin_cksum = FI_CK_L4PART; return (0); } else { manual = 1; } } skipauto: if (manual != 0) { if (ipf_checkl4sum(fin) == -1) { fin->fin_flx |= FI_BAD; DT2(ipf_fi_bad_checkv4sum_manual, fr_info_t *, fin, u_int, manual); return (-1); } } #else if (ipf_checkl4sum(fin) == -1) { fin->fin_flx |= FI_BAD; DT2(ipf_fi_bad_checkv4sum_checkl4sum, fr_info_t *, fin, u_int, -1); return (-1); } #endif return (0); } #ifdef USE_INET6 int ipf_checkv6sum(fr_info_t *fin) { if ((fin->fin_flx & FI_NOCKSUM) != 0) { DT(ipf_checkv6sum_fi_nocksum); return (0); } if ((fin->fin_flx & FI_SHORT) != 0) { DT(ipf_checkv6sum_fi_short); return (1); } if (fin->fin_cksum != FI_CK_NEEDED) { DT(ipf_checkv6sum_fi_ck_needed); return (fin->fin_cksum > FI_CK_NEEDED) ? 0 : -1; } if (ipf_checkl4sum(fin) == -1) { fin->fin_flx |= FI_BAD; DT2(ipf_fi_bad_checkv6sum_checkl4sum, fr_info_t *, fin, u_int, -1); return (-1); } return (0); } #endif /* USE_INET6 */ size_t mbufchainlen(struct mbuf *m0) { size_t len; if ((m0->m_flags & M_PKTHDR) != 0) { len = m0->m_pkthdr.len; } else { struct mbuf *m; for (m = m0, len = 0; m != NULL; m = m->m_next) len += m->m_len; } return (len); } /* ------------------------------------------------------------------------ */ /* Function: ipf_pullup */ /* Returns: NULL == pullup failed, else pointer to protocol header */ /* Parameters: xmin(I)- pointer to buffer where data packet starts */ /* fin(I) - pointer to packet information */ /* len(I) - number of bytes to pullup */ /* */ /* Attempt to move at least len bytes (from the start of the buffer) into a */ /* single buffer for ease of access. Operating system native functions are */ /* used to manage buffers - if necessary. If the entire packet ends up in */ /* a single buffer, set the FI_COALESCE flag even though ipf_coalesce() has */ /* not been called. Both fin_ip and fin_dp are updated before exiting _IF_ */ /* and ONLY if the pullup succeeds. */ /* */ /* We assume that 'xmin' is a pointer to a buffer that is part of the chain */ /* of buffers that starts at *fin->fin_mp. */ /* ------------------------------------------------------------------------ */ void * ipf_pullup(mb_t *xmin, fr_info_t *fin, int len) { int dpoff, ipoff; mb_t *m = xmin; char *ip; if (m == NULL) return (NULL); ip = (char *)fin->fin_ip; if ((fin->fin_flx & FI_COALESCE) != 0) return (ip); ipoff = fin->fin_ipoff; if (fin->fin_dp != NULL) dpoff = (char *)fin->fin_dp - (char *)ip; else dpoff = 0; if (M_LEN(m) < len) { mb_t *n = *fin->fin_mp; /* * Assume that M_PKTHDR is set and just work with what is left * rather than check.. * Should not make any real difference, anyway. */ if (m != n) { /* * Record the mbuf that points to the mbuf that we're * about to go to work on so that we can update the * m_next appropriately later. */ for (; n->m_next != m; n = n->m_next) ; } else { n = NULL; } #ifdef MHLEN if (len > MHLEN) #else if (len > MLEN) #endif { #ifdef HAVE_M_PULLDOWN if (m_pulldown(m, 0, len, NULL) == NULL) m = NULL; #else FREE_MB_T(*fin->fin_mp); m = NULL; n = NULL; #endif } else { m = m_pullup(m, len); } if (n != NULL) n->m_next = m; if (m == NULL) { /* * When n is non-NULL, it indicates that m pointed to * a sub-chain (tail) of the mbuf and that the head * of this chain has not yet been free'd. */ if (n != NULL) { FREE_MB_T(*fin->fin_mp); } *fin->fin_mp = NULL; fin->fin_m = NULL; return (NULL); } if (n == NULL) *fin->fin_mp = m; while (M_LEN(m) == 0) { m = m->m_next; } fin->fin_m = m; ip = MTOD(m, char *) + ipoff; fin->fin_ip = (ip_t *)ip; if (fin->fin_dp != NULL) fin->fin_dp = (char *)fin->fin_ip + dpoff; if (fin->fin_fraghdr != NULL) fin->fin_fraghdr = (char *)ip + ((char *)fin->fin_fraghdr - (char *)fin->fin_ip); } if (len == fin->fin_plen) fin->fin_flx |= FI_COALESCE; return (ip); } int ipf_inject(fr_info_t *fin, mb_t *m) { struct epoch_tracker et; int error = 0; NET_EPOCH_ENTER(et); if (fin->fin_out == 0) { netisr_dispatch(NETISR_IP, m); } else { fin->fin_ip->ip_len = ntohs(fin->fin_ip->ip_len); fin->fin_ip->ip_off = ntohs(fin->fin_ip->ip_off); error = ip_output(m, NULL, NULL, IP_FORWARDING, NULL, NULL); } NET_EPOCH_EXIT(et); return (error); } VNET_DEFINE_STATIC(pfil_hook_t, ipf_inet_hook); VNET_DEFINE_STATIC(pfil_hook_t, ipf_inet6_hook); #define V_ipf_inet_hook VNET(ipf_inet_hook) #define V_ipf_inet6_hook VNET(ipf_inet6_hook) int ipf_pfil_unhook(void) { pfil_remove_hook(V_ipf_inet_hook); #ifdef USE_INET6 pfil_remove_hook(V_ipf_inet6_hook); #endif return (0); } int ipf_pfil_hook(void) { struct pfil_hook_args pha; struct pfil_link_args pla; int error, error6; pha.pa_version = PFIL_VERSION; pha.pa_flags = PFIL_IN | PFIL_OUT; pha.pa_modname = "ipfilter"; pha.pa_rulname = "default-ip4"; pha.pa_func = ipf_check_wrapper; pha.pa_ruleset = NULL; pha.pa_type = PFIL_TYPE_IP4; V_ipf_inet_hook = pfil_add_hook(&pha); #ifdef USE_INET6 pha.pa_rulname = "default-ip6"; pha.pa_func = ipf_check_wrapper6; pha.pa_type = PFIL_TYPE_IP6; V_ipf_inet6_hook = pfil_add_hook(&pha); #endif pla.pa_version = PFIL_VERSION; pla.pa_flags = PFIL_IN | PFIL_OUT | PFIL_HEADPTR | PFIL_HOOKPTR; pla.pa_head = V_inet_pfil_head; pla.pa_hook = V_ipf_inet_hook; error = pfil_link(&pla); error6 = 0; #ifdef USE_INET6 pla.pa_head = V_inet6_pfil_head; pla.pa_hook = V_ipf_inet6_hook; error6 = pfil_link(&pla); #endif if (error || error6) error = ENODEV; else error = 0; return (error); } void ipf_event_reg(void) { V_ipf_arrivetag = EVENTHANDLER_REGISTER(ifnet_arrival_event, \ ipf_ifevent, NULL, \ EVENTHANDLER_PRI_ANY); V_ipf_departtag = EVENTHANDLER_REGISTER(ifnet_departure_event, \ ipf_ifevent, NULL, \ EVENTHANDLER_PRI_ANY); #if 0 V_ipf_clonetag = EVENTHANDLER_REGISTER(if_clone_event, ipf_ifevent, \ NULL, EVENTHANDLER_PRI_ANY); #endif } void ipf_event_dereg(void) { if (V_ipf_arrivetag != NULL) { EVENTHANDLER_DEREGISTER(ifnet_arrival_event, V_ipf_arrivetag); } if (V_ipf_departtag != NULL) { EVENTHANDLER_DEREGISTER(ifnet_departure_event, V_ipf_departtag); } #if 0 if (V_ipf_clonetag != NULL) { EVENTHANDLER_DEREGISTER(if_clone_event, V_ipf_clonetag); } #endif } u_32_t ipf_random(void) { return (arc4random()); } u_int ipf_pcksum(fr_info_t *fin, int hlen, u_int sum) { struct mbuf *m; u_int sum2; int off; m = fin->fin_m; off = (char *)fin->fin_dp - (char *)fin->fin_ip; m->m_data += hlen; m->m_len -= hlen; sum2 = in_cksum(fin->fin_m, fin->fin_plen - off); m->m_len += hlen; m->m_data -= hlen; /* * Both sum and sum2 are partial sums, so combine them together. */ sum += ~sum2 & 0xffff; while (sum > 0xffff) sum = (sum & 0xffff) + (sum >> 16); sum2 = ~sum & 0xffff; return (sum2); } #ifdef USE_INET6 u_int ipf_pcksum6(struct mbuf *m, ip6_t *ip6, u_int32_t off, u_int32_t len) { #ifdef _KERNEL int sum; if (m->m_len < sizeof(struct ip6_hdr)) { return (0xffff); } sum = in6_cksum(m, ip6->ip6_nxt, off, len); return (sum); #else u_short *sp; u_int sum; sp = (u_short *)&ip6->ip6_src; sum = *sp++; /* ip6_src */ sum += *sp++; sum += *sp++; sum += *sp++; sum += *sp++; sum += *sp++; sum += *sp++; sum += *sp++; sum += *sp++; /* ip6_dst */ sum += *sp++; sum += *sp++; sum += *sp++; sum += *sp++; sum += *sp++; sum += *sp++; sum += *sp++; return (ipf_pcksum(fin, off, sum)); #endif } #endif void ipf_fbsd_kenv_get(ipf_main_softc_t *softc) { TUNABLE_INT_FETCH("net.inet.ipf.large_nat", &softc->ipf_large_nat); } diff --git a/sys/netpfil/pf/pf_syncookies.c b/sys/netpfil/pf/pf_syncookies.c index 6a375411d8ea..0c9b5a459301 100644 --- a/sys/netpfil/pf/pf_syncookies.c +++ b/sys/netpfil/pf/pf_syncookies.c @@ -1,501 +1,502 @@ /* $OpenBSD: pf_syncookies.c,v 1.7 2018/09/10 15:54:28 henning Exp $ */ /* Copyright (c) 2016,2017 Henning Brauer * Copyright (c) 2016 Alexandr Nedvedicky * * syncookie parts based on FreeBSD sys/netinet/tcp_syncache.c * * 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. */ /* * when we're under synflood, we use syncookies to prevent state table * exhaustion. Trigger for the synflood mode is the number of half-open * connections in the state table. * We leave synflood mode when the number of half-open states - including * in-flight syncookies - drops far enough again */ /* * syncookie enabled Initial Sequence Number: * 24 bit MAC * 3 bit WSCALE index * 3 bit MSS index * 1 bit SACK permitted * 1 bit odd/even secret * * References: * RFC4987 TCP SYN Flooding Attacks and Common Mitigations * http://cr.yp.to/syncookies.html (overview) * http://cr.yp.to/syncookies/archive (details) */ //#include "pflog.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include +#include #include #include #include #include #include #define DPFPRINTF(n, x) if (V_pf_status.debug >= (n)) printf x union pf_syncookie { uint8_t cookie; struct { uint8_t oddeven:1, sack_ok:1, wscale_idx:3, mss_idx:3; } flags; }; #define PF_SYNCOOKIE_SECRET_SIZE SIPHASH_KEY_LENGTH #define PF_SYNCOOKIE_SECRET_LIFETIME 15 /* seconds */ /* Protected by PF_RULES_xLOCK. */ struct pf_syncookie_status { struct callout keytimeout; uint8_t oddeven; uint8_t key[2][SIPHASH_KEY_LENGTH]; uint32_t hiwat; /* absolute; # of states */ uint32_t lowat; }; VNET_DEFINE_STATIC(struct pf_syncookie_status, pf_syncookie_status); #define V_pf_syncookie_status VNET(pf_syncookie_status) static int pf_syncookies_setmode(u_int8_t); void pf_syncookie_rotate(void *); void pf_syncookie_newkey(void); uint32_t pf_syncookie_mac(struct pf_pdesc *, union pf_syncookie, uint32_t); uint32_t pf_syncookie_generate(struct mbuf *m, int off, struct pf_pdesc *, uint16_t); void pf_syncookies_init(void) { callout_init(&V_pf_syncookie_status.keytimeout, 1); PF_RULES_WLOCK(); pf_syncookies_setmode(PF_SYNCOOKIES_NEVER); PF_RULES_WUNLOCK(); } void pf_syncookies_cleanup(void) { callout_stop(&V_pf_syncookie_status.keytimeout); } int pf_get_syncookies(struct pfioc_nv *nv) { nvlist_t *nvl = NULL; void *nvlpacked = NULL; int error; #define ERROUT(x) ERROUT_FUNCTION(errout, x) nvl = nvlist_create(0); if (nvl == NULL) ERROUT(ENOMEM); nvlist_add_bool(nvl, "enabled", V_pf_status.syncookies_mode != PF_SYNCOOKIES_NEVER); nvlist_add_bool(nvl, "adaptive", V_pf_status.syncookies_mode == PF_SYNCOOKIES_ADAPTIVE); nvlist_add_number(nvl, "highwater", V_pf_syncookie_status.hiwat); nvlist_add_number(nvl, "lowwater", V_pf_syncookie_status.lowat); nvlpacked = nvlist_pack(nvl, &nv->len); if (nvlpacked == NULL) ERROUT(ENOMEM); if (nv->size == 0) { ERROUT(0); } else if (nv->size < nv->len) { ERROUT(ENOSPC); } error = copyout(nvlpacked, nv->data, nv->len); #undef ERROUT errout: nvlist_destroy(nvl); free(nvlpacked, M_NVLIST); return (error); } int pf_set_syncookies(struct pfioc_nv *nv) { nvlist_t *nvl = NULL; void *nvlpacked = NULL; int error; bool enabled, adaptive; uint32_t hiwat, lowat; uint8_t newmode; #define ERROUT(x) ERROUT_FUNCTION(errout, x) if (nv->len > pf_ioctl_maxcount) return (ENOMEM); nvlpacked = malloc(nv->len, M_NVLIST, M_WAITOK); if (nvlpacked == NULL) return (ENOMEM); error = copyin(nv->data, nvlpacked, nv->len); if (error) ERROUT(error); nvl = nvlist_unpack(nvlpacked, nv->len, 0); if (nvl == NULL) ERROUT(EBADMSG); if (! nvlist_exists_bool(nvl, "enabled") || ! nvlist_exists_bool(nvl, "adaptive")) ERROUT(EBADMSG); enabled = nvlist_get_bool(nvl, "enabled"); adaptive = nvlist_get_bool(nvl, "adaptive"); PFNV_CHK(pf_nvuint32_opt(nvl, "highwater", &hiwat, V_pf_syncookie_status.hiwat)); PFNV_CHK(pf_nvuint32_opt(nvl, "lowwater", &lowat, V_pf_syncookie_status.lowat)); if (lowat >= hiwat) ERROUT(EINVAL); newmode = PF_SYNCOOKIES_NEVER; if (enabled) newmode = adaptive ? PF_SYNCOOKIES_ADAPTIVE : PF_SYNCOOKIES_ALWAYS; PF_RULES_WLOCK(); error = pf_syncookies_setmode(newmode); V_pf_syncookie_status.lowat = lowat; V_pf_syncookie_status.hiwat = hiwat; PF_RULES_WUNLOCK(); #undef ERROUT errout: nvlist_destroy(nvl); free(nvlpacked, M_NVLIST); return (error); } static int pf_syncookies_setmode(u_int8_t mode) { if (mode > PF_SYNCOOKIES_MODE_MAX) return (EINVAL); if (V_pf_status.syncookies_mode == mode) return (0); V_pf_status.syncookies_mode = mode; if (V_pf_status.syncookies_mode == PF_SYNCOOKIES_ALWAYS) { pf_syncookie_newkey(); V_pf_status.syncookies_active = true; } return (0); } int pf_synflood_check(struct pf_pdesc *pd) { MPASS(pd->proto == IPPROTO_TCP); PF_RULES_RASSERT(); if (pd->pf_mtag && (pd->pf_mtag->tag & PF_TAG_SYNCOOKIE_RECREATED)) return (0); if (V_pf_status.syncookies_mode != PF_SYNCOOKIES_ADAPTIVE) return (V_pf_status.syncookies_mode); if (!V_pf_status.syncookies_active && atomic_load_32(&V_pf_status.states_halfopen) > V_pf_syncookie_status.hiwat) { /* We'd want to 'pf_syncookie_newkey()' here, but that requires * the rules write lock, which we can't get with the read lock * held. */ callout_reset(&V_pf_syncookie_status.keytimeout, 0, pf_syncookie_rotate, curvnet); V_pf_status.syncookies_active = true; DPFPRINTF(LOG_WARNING, ("synflood detected, enabling syncookies\n")); // XXXTODO V_pf_status.lcounters[LCNT_SYNFLOODS]++; } return (V_pf_status.syncookies_active); } void pf_syncookie_send(struct mbuf *m, int off, struct pf_pdesc *pd) { uint16_t mss; uint32_t iss; mss = max(V_tcp_mssdflt, pf_get_mss(m, off, pd->hdr.tcp.th_off, pd->af)); iss = pf_syncookie_generate(m, off, pd, mss); pf_send_tcp(NULL, pd->af, pd->dst, pd->src, *pd->dport, *pd->sport, iss, ntohl(pd->hdr.tcp.th_seq) + 1, TH_SYN|TH_ACK, 0, mss, 0, 1, 0); counter_u64_add(V_pf_status.lcounters[KLCNT_SYNCOOKIES_SENT], 1); /* XXX Maybe only in adaptive mode? */ atomic_add_64(&V_pf_status.syncookies_inflight[V_pf_syncookie_status.oddeven], 1); } uint8_t pf_syncookie_validate(struct pf_pdesc *pd) { uint32_t hash, ack, seq; union pf_syncookie cookie; MPASS(pd->proto == IPPROTO_TCP); PF_RULES_RASSERT(); seq = ntohl(pd->hdr.tcp.th_seq) - 1; ack = ntohl(pd->hdr.tcp.th_ack) - 1; cookie.cookie = (ack & 0xff) ^ (ack >> 24); /* we don't know oddeven before setting the cookie (union) */ if (atomic_load_64(&V_pf_status.syncookies_inflight[cookie.flags.oddeven]) == 0) return (0); hash = pf_syncookie_mac(pd, cookie, seq); if ((ack & ~0xff) != (hash & ~0xff)) return (0); counter_u64_add(V_pf_status.lcounters[KLCNT_SYNCOOKIES_VALID], 1); atomic_add_64(&V_pf_status.syncookies_inflight[cookie.flags.oddeven], -1); return (1); } /* * all following functions private */ void pf_syncookie_rotate(void *arg) { CURVNET_SET((struct vnet *)arg); /* do we want to disable syncookies? */ if (V_pf_status.syncookies_active && ((V_pf_status.syncookies_mode == PF_SYNCOOKIES_ADAPTIVE && (atomic_load_32(&V_pf_status.states_halfopen) + atomic_load_64(&V_pf_status.syncookies_inflight[0]) + atomic_load_64(&V_pf_status.syncookies_inflight[1])) < V_pf_syncookie_status.lowat) || V_pf_status.syncookies_mode == PF_SYNCOOKIES_NEVER) ) { V_pf_status.syncookies_active = false; DPFPRINTF(PF_DEBUG_MISC, ("syncookies disabled\n")); } /* nothing in flight any more? delete keys and return */ if (!V_pf_status.syncookies_active && atomic_load_64(&V_pf_status.syncookies_inflight[0]) == 0 && atomic_load_64(&V_pf_status.syncookies_inflight[1]) == 0) { memset(V_pf_syncookie_status.key[0], 0, PF_SYNCOOKIE_SECRET_SIZE); memset(V_pf_syncookie_status.key[1], 0, PF_SYNCOOKIE_SECRET_SIZE); CURVNET_RESTORE(); return; } PF_RULES_WLOCK(); /* new key, including timeout */ pf_syncookie_newkey(); PF_RULES_WUNLOCK(); CURVNET_RESTORE(); } void pf_syncookie_newkey(void) { PF_RULES_WASSERT(); MPASS(V_pf_syncookie_status.oddeven < 2); V_pf_syncookie_status.oddeven = (V_pf_syncookie_status.oddeven + 1) & 0x1; atomic_store_64(&V_pf_status.syncookies_inflight[V_pf_syncookie_status.oddeven], 0); arc4random_buf(V_pf_syncookie_status.key[V_pf_syncookie_status.oddeven], PF_SYNCOOKIE_SECRET_SIZE); callout_reset(&V_pf_syncookie_status.keytimeout, PF_SYNCOOKIE_SECRET_LIFETIME * hz, pf_syncookie_rotate, curvnet); } /* * 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 pf_syncookie_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; /* * Distribution and probability of certain WSCALE values. * The absence of the WSCALE option 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 pf_syncookie_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; uint32_t pf_syncookie_mac(struct pf_pdesc *pd, union pf_syncookie cookie, uint32_t seq) { SIPHASH_CTX ctx; uint32_t siphash[2]; PF_RULES_RASSERT(); MPASS(pd->proto == IPPROTO_TCP); SipHash24_Init(&ctx); SipHash_SetKey(&ctx, V_pf_syncookie_status.key[cookie.flags.oddeven]); switch (pd->af) { case AF_INET: SipHash_Update(&ctx, pd->src, sizeof(pd->src->v4)); SipHash_Update(&ctx, pd->dst, sizeof(pd->dst->v4)); break; case AF_INET6: SipHash_Update(&ctx, pd->src, sizeof(pd->src->v6)); SipHash_Update(&ctx, pd->dst, sizeof(pd->dst->v6)); break; default: panic("unknown address family"); } SipHash_Update(&ctx, pd->sport, sizeof(*pd->sport)); SipHash_Update(&ctx, pd->dport, sizeof(*pd->dport)); SipHash_Update(&ctx, &seq, sizeof(seq)); SipHash_Update(&ctx, &cookie, sizeof(cookie)); SipHash_Final((uint8_t *)&siphash, &ctx); return (siphash[0] ^ siphash[1]); } uint32_t pf_syncookie_generate(struct mbuf *m, int off, struct pf_pdesc *pd, uint16_t mss) { uint8_t i, wscale; uint32_t iss, hash; union pf_syncookie cookie; PF_RULES_RASSERT(); cookie.cookie = 0; /* map MSS */ for (i = nitems(pf_syncookie_msstab) - 1; pf_syncookie_msstab[i] > mss && i > 0; i--) /* nada */; cookie.flags.mss_idx = i; /* map WSCALE */ wscale = pf_get_wscale(m, off, pd->hdr.tcp.th_off, pd->af); for (i = nitems(pf_syncookie_wstab) - 1; pf_syncookie_wstab[i] > wscale && i > 0; i--) /* nada */; cookie.flags.wscale_idx = i; cookie.flags.sack_ok = 0; /* XXX */ cookie.flags.oddeven = V_pf_syncookie_status.oddeven; hash = pf_syncookie_mac(pd, cookie, ntohl(pd->hdr.tcp.th_seq)); /* * 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); return (iss); } struct mbuf * pf_syncookie_recreate_syn(uint8_t ttl, int off, struct pf_pdesc *pd) { uint8_t wscale; uint16_t mss; uint32_t ack, seq; union pf_syncookie cookie; seq = ntohl(pd->hdr.tcp.th_seq) - 1; ack = ntohl(pd->hdr.tcp.th_ack) - 1; cookie.cookie = (ack & 0xff) ^ (ack >> 24); if (cookie.flags.mss_idx >= nitems(pf_syncookie_msstab) || cookie.flags.wscale_idx >= nitems(pf_syncookie_wstab)) return (NULL); mss = pf_syncookie_msstab[cookie.flags.mss_idx]; wscale = pf_syncookie_wstab[cookie.flags.wscale_idx]; return (pf_build_tcp(NULL, pd->af, pd->src, pd->dst, *pd->sport, *pd->dport, seq, 0, TH_SYN, wscale, mss, ttl, 0, PF_TAG_SYNCOOKIE_RECREATED)); } diff --git a/usr.bin/systat/netstat.c b/usr.bin/systat/netstat.c index dc1e08c590dd..33e13cc5c906 100644 --- a/usr.bin/systat/netstat.c +++ b/usr.bin/systat/netstat.c @@ -1,641 +1,642 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1980, 1992, 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. */ #include __FBSDID("$FreeBSD$"); #ifdef lint static const char sccsid[] = "@(#)netstat.c 8.1 (Berkeley) 6/6/93"; #endif /* * netstat */ #include #include #include +#include #define _WANT_SOCKET #include #include #include #include #include #include #include #ifdef INET6 #include #endif #define _WANT_INPCB #include #include #include #include #include #include #include #define TCPSTATES #include #include #define _WANT_TCPCB #include #include #include #include #include #include #include #include #include #include "systat.h" #include "extern.h" static struct netinfo *enter(struct in_conninfo *, uint8_t, int, const char *); static void enter_kvm(struct inpcb *, struct socket *, int, const char *); static void enter_sysctl(struct xinpcb *, struct xsocket *, int, const char *); static void fetchnetstat_kvm(void); static void fetchnetstat_sysctl(void); static char *inetname(struct sockaddr *); static void inetprint(struct sockaddr *, const char *); #define streq(a,b) (strcmp(a,b)==0) #define YMAX(w) (getmaxy(w)-2) WINDOW * opennetstat(void) { sethostent(1); setnetent(1); return (subwin(stdscr, LINES-3-1, 0, MAINWIN_ROW, 0)); } struct netinfo { TAILQ_ENTRY(netinfo) chain; short ni_line; /* line on screen */ short ni_seen; /* 0 when not present in list */ short ni_flags; #define NIF_LACHG 0x1 /* local address changed */ #define NIF_FACHG 0x2 /* foreign address changed */ short ni_state; /* tcp state */ const char *ni_proto; /* protocol */ struct sockaddr_storage ni_lsa; /* local address */ struct sockaddr_storage ni_fsa; /* foreign address */ u_int ni_rcvcc; /* rcv buffer character count */ u_int ni_sndcc; /* snd buffer character count */ }; static TAILQ_HEAD(netinfohead, netinfo) netcb = TAILQ_HEAD_INITIALIZER(netcb); static int aflag = 0; static int nflag = 0; static int lastrow = 1; void closenetstat(WINDOW *w) { struct netinfo *p; endhostent(); endnetent(); TAILQ_FOREACH(p, &netcb, chain) { if (p->ni_line != -1) lastrow--; p->ni_line = -1; } if (w != NULL) { wclear(w); wrefresh(w); delwin(w); } } static const char *miblist[] = { "net.inet.tcp.pcblist", "net.inet.udp.pcblist" }; static char tcb[] = "tcb", udb[] = "udb"; struct nlist namelist[] = { #define X_TCB 0 { .n_name = tcb }, #define X_UDB 1 { .n_name = udb }, { .n_name = NULL }, }; int initnetstat(void) { protos = TCP|UDP; return(1); } void fetchnetstat(void) { if (use_kvm) fetchnetstat_kvm(); else fetchnetstat_sysctl(); } static void fetchnetstat_kvm(void) { struct inpcb *next; struct netinfo *p; struct inpcbhead head; struct inpcb inpcb; struct socket sockb; struct tcpcb tcpcb; void *off; int istcp; if (namelist[X_TCB].n_value == 0) return; TAILQ_FOREACH(p, &netcb, chain) p->ni_seen = 0; if (protos&TCP) { off = NPTR(X_TCB); istcp = 1; } else if (protos&UDP) { off = NPTR(X_UDB); istcp = 0; } else { error("No protocols to display"); return; } again: KREAD(off, &head, sizeof (struct inpcbhead)); LIST_FOREACH(next, &head, inp_list) { KREAD(next, &inpcb, sizeof (inpcb)); next = &inpcb; if (!aflag) { if (inpcb.inp_vflag & INP_IPV4) { if (inpcb.inp_laddr.s_addr == INADDR_ANY) continue; } #ifdef INET6 else if (inpcb.inp_vflag & INP_IPV6) { if (memcmp(&inpcb.in6p_laddr, &in6addr_any, sizeof(in6addr_any)) == 0) continue; } #endif } if (nhosts && !checkhost(&inpcb.inp_inc)) continue; if (nports && !checkport(&inpcb.inp_inc)) continue; if (istcp) { KREAD(inpcb.inp_socket, &sockb, sizeof (sockb)); KREAD(inpcb.inp_ppcb, &tcpcb, sizeof (tcpcb)); enter_kvm(&inpcb, &sockb, tcpcb.t_state, "tcp"); } else enter_kvm(&inpcb, &sockb, 0, "udp"); } if (istcp && (protos&UDP)) { istcp = 0; off = NPTR(X_UDB); goto again; } } static void fetchnetstat_sysctl(void) { struct netinfo *p; int idx; struct xinpgen *inpg; char *cur, *end; struct xinpcb *xip = NULL; struct xtcpcb *xtp = NULL; int plen; size_t lsz; TAILQ_FOREACH(p, &netcb, chain) p->ni_seen = 0; if (protos&TCP) { idx = 0; } else if (protos&UDP) { idx = 1; } else { error("No protocols to display"); return; } for (;idx < 2; idx++) { if (idx == 1 && !(protos&UDP)) break; inpg = (struct xinpgen *)sysctl_dynread(miblist[idx], &lsz); if (inpg == NULL) { error("sysctl(%s...) failed", miblist[idx]); continue; } /* * We currently do no require a consistent pcb list. * Try to be robust in case of struct size changes */ cur = ((char *)inpg) + inpg->xig_len; /* There is also a trailing struct xinpgen */ end = ((char *)inpg) + lsz - inpg->xig_len; if (end <= cur) { free(inpg); continue; } if (idx == 0) { /* TCP */ xtp = (struct xtcpcb *)cur; plen = xtp->xt_len; } else { xip = (struct xinpcb *)cur; plen = xip->xi_len; } while (cur + plen <= end) { if (idx == 0) { /* TCP */ xtp = (struct xtcpcb *)cur; xip = &xtp->xt_inp; } else { xip = (struct xinpcb *)cur; } cur += plen; if (!aflag) { if (xip->inp_vflag & INP_IPV4) { if (xip->inp_laddr.s_addr == INADDR_ANY) continue; } #ifdef INET6 else if (xip->inp_vflag & INP_IPV6) { if (memcmp(&xip->in6p_laddr, &in6addr_any, sizeof(in6addr_any)) == 0) continue; } #endif } if (nhosts && !checkhost(&xip->inp_inc)) continue; if (nports && !checkport(&xip->inp_inc)) continue; if (idx == 0) enter_sysctl(xip, &xip->xi_socket, xtp->t_state, "tcp"); else enter_sysctl(xip, &xip->xi_socket, 0, "udp"); } free(inpg); } } static void enter_kvm(struct inpcb *inp, struct socket *so, int state, const char *proto) { struct netinfo *p; if ((p = enter(&inp->inp_inc, inp->inp_vflag, state, proto)) != NULL) { p->ni_rcvcc = so->so_rcv.sb_ccc; p->ni_sndcc = so->so_snd.sb_ccc; } } static void enter_sysctl(struct xinpcb *xip, struct xsocket *so, int state, const char *proto) { struct netinfo *p; if ((p = enter(&xip->inp_inc, xip->inp_vflag, state, proto)) != NULL) { p->ni_rcvcc = so->so_rcv.sb_cc; p->ni_sndcc = so->so_snd.sb_cc; } } static struct netinfo * enter(struct in_conninfo *inc, uint8_t vflag, int state, const char *proto) { struct netinfo *p; struct sockaddr_storage lsa, fsa; struct sockaddr_in *sa4; #ifdef INET6 struct sockaddr_in6 *sa6; #endif memset(&lsa, 0, sizeof(lsa)); memset(&fsa, 0, sizeof(fsa)); if (vflag & INP_IPV4) { sa4 = (struct sockaddr_in *)&lsa; sa4->sin_addr = inc->inc_laddr; sa4->sin_port = inc->inc_lport; sa4->sin_family = AF_INET; sa4->sin_len = sizeof(struct sockaddr_in); sa4 = (struct sockaddr_in *)&fsa; sa4->sin_addr = inc->inc_faddr; sa4->sin_port = inc->inc_fport; sa4->sin_family = AF_INET; sa4->sin_len = sizeof(struct sockaddr_in); } #ifdef INET6 else if (vflag & INP_IPV6) { sa6 = (struct sockaddr_in6 *)&lsa; memcpy(&sa6->sin6_addr, &inc->inc6_laddr, sizeof(struct in6_addr)); sa6->sin6_port = inc->inc_lport; sa6->sin6_family = AF_INET6; sa6->sin6_len = sizeof(struct sockaddr_in6); sa6 = (struct sockaddr_in6 *)&fsa; memcpy(&sa6->sin6_addr, &inc->inc6_faddr, sizeof(struct in6_addr)); sa6->sin6_port = inc->inc_fport; sa6->sin6_family = AF_INET6; sa6->sin6_len = sizeof(struct sockaddr_in6); } #endif else return NULL; /* * Only take exact matches, any sockets with * previously unbound addresses will be deleted * below in the display routine because they * will appear as ``not seen'' in the kernel * data structures. */ TAILQ_FOREACH(p, &netcb, chain) { if (!streq(proto, p->ni_proto)) continue; if (p->ni_lsa.ss_family != lsa.ss_family || memcmp(&p->ni_lsa, &lsa, lsa.ss_len) != 0) continue; if (p->ni_fsa.ss_family == fsa.ss_family && memcmp(&p->ni_fsa, &fsa, fsa.ss_len) == 0) break; } if (p == NULL) { if ((p = malloc(sizeof(*p))) == NULL) { error("Out of memory"); return NULL; } TAILQ_INSERT_HEAD(&netcb, p, chain); p->ni_line = -1; memcpy(&p->ni_lsa, &lsa, lsa.ss_len); memcpy(&p->ni_fsa, &fsa, fsa.ss_len); p->ni_proto = strdup(proto); p->ni_flags = NIF_LACHG|NIF_FACHG; } p->ni_state = state; p->ni_seen = 1; return p; } /* column locations */ #define LADDR 0 #define FADDR LADDR+23 #define PROTO FADDR+23 #define RCVCC PROTO+6 #define SNDCC RCVCC+7 #define STATE SNDCC+7 void labelnetstat(void) { if (use_kvm && namelist[X_TCB].n_type == 0) return; wmove(wnd, 0, 0); wclrtobot(wnd); mvwaddstr(wnd, 0, LADDR, "Local Address"); mvwaddstr(wnd, 0, FADDR, "Foreign Address"); mvwaddstr(wnd, 0, PROTO, "Proto"); mvwaddstr(wnd, 0, RCVCC, "Recv-Q"); mvwaddstr(wnd, 0, SNDCC, "Send-Q"); mvwaddstr(wnd, 0, STATE, "(state)"); } void shownetstat(void) { struct netinfo *p, *q; char proto[6]; const char *family = ""; /* * First, delete any connections that have gone * away and adjust the position of connections * below to reflect the deleted line. */ p = TAILQ_FIRST(&netcb); while (p != NULL) { if (p->ni_line == -1 || p->ni_seen) { p = TAILQ_NEXT(p, chain); continue; } wmove(wnd, p->ni_line, 0); wdeleteln(wnd); TAILQ_FOREACH(q, &netcb, chain) if (q != p && q->ni_line > p->ni_line) { q->ni_line--; /* this shouldn't be necessary */ q->ni_flags |= NIF_LACHG|NIF_FACHG; } lastrow--; q = TAILQ_NEXT(p, chain); TAILQ_REMOVE(&netcb, p, chain); free(p); p = q; } /* * Update existing connections and add new ones. */ TAILQ_FOREACH(p, &netcb, chain) { if (p->ni_line == -1) { /* * Add a new entry if possible. */ if (lastrow > YMAX(wnd)) continue; p->ni_line = lastrow++; p->ni_flags |= NIF_LACHG|NIF_FACHG; } if (p->ni_flags & NIF_LACHG) { wmove(wnd, p->ni_line, LADDR); inetprint((struct sockaddr *)&p->ni_lsa, p->ni_proto); p->ni_flags &= ~NIF_LACHG; } if (p->ni_flags & NIF_FACHG) { wmove(wnd, p->ni_line, FADDR); inetprint((struct sockaddr *)&p->ni_fsa, p->ni_proto); p->ni_flags &= ~NIF_FACHG; } #ifdef INET6 family = (p->ni_lsa.ss_family == AF_INET) ? "4" : "6"; #endif snprintf(proto, sizeof(proto), "%s%s", p->ni_proto, family); mvwaddstr(wnd, p->ni_line, PROTO, proto); mvwprintw(wnd, p->ni_line, RCVCC, "%6u", p->ni_rcvcc); mvwprintw(wnd, p->ni_line, SNDCC, "%6u", p->ni_sndcc); if (streq(p->ni_proto, "tcp")) { if (p->ni_state < 0 || p->ni_state >= TCP_NSTATES) mvwprintw(wnd, p->ni_line, STATE, "%d", p->ni_state); else mvwaddstr(wnd, p->ni_line, STATE, tcpstates[p->ni_state]); } wclrtoeol(wnd); } if (lastrow < YMAX(wnd)) { wmove(wnd, lastrow, 0); wclrtobot(wnd); wmove(wnd, YMAX(wnd), 0); wdeleteln(wnd); /* XXX */ } } /* * Pretty print an Internet address (net address + port). * If the nflag was specified, use numbers instead of names. */ static void inetprint(struct sockaddr *sa, const char *proto) { struct servent *sp = 0; char line[80], *cp; int port; switch (sa->sa_family) { case AF_INET: port = ((struct sockaddr_in *)sa)->sin_port; break; #ifdef INET6 case AF_INET6: port = ((struct sockaddr_in6 *)sa)->sin6_port; break; #endif default: port = 0; break; } snprintf(line, sizeof(line), "%.*s.", 16, inetname(sa)); cp = strchr(line, '\0'); if (!nflag && port) sp = getservbyport(port, proto); if (sp || port == 0) snprintf(cp, sizeof(line) - (cp - line), "%.8s", sp ? sp->s_name : "*"); else snprintf(cp, sizeof(line) - (cp - line), "%d", ntohs((u_short)port)); /* pad to full column to clear any garbage */ cp = strchr(line, '\0'); while (cp - line < 22) *cp++ = ' '; line[22] = '\0'; waddstr(wnd, line); } /* * Construct an Internet address representation. * If the nflag has been supplied, give * numeric value, otherwise try for symbolic name. */ static char * inetname(struct sockaddr *sa) { char *cp = 0; static char line[NI_MAXHOST]; struct hostent *hp; struct in_addr in; #ifdef INET6 if (sa->sa_family == AF_INET6) { if (memcmp(&((struct sockaddr_in6 *)sa)->sin6_addr, &in6addr_any, sizeof(in6addr_any)) == 0) strcpy(line, "*"); else getnameinfo(sa, sa->sa_len, line, sizeof(line), NULL, 0, nflag ? NI_NUMERICHOST : 0); return (line); } #endif in = ((struct sockaddr_in *)sa)->sin_addr; if (!nflag && in.s_addr != INADDR_ANY) { hp = gethostbyaddr((char *)&in, sizeof (in), AF_INET); if (hp) cp = hp->h_name; } if (in.s_addr == INADDR_ANY) strcpy(line, "*"); else if (cp) snprintf(line, sizeof(line), "%s", cp); else { in.s_addr = ntohl(in.s_addr); #define C(x) ((x) & 0xff) snprintf(line, sizeof(line), "%u.%u.%u.%u", C(in.s_addr >> 24), C(in.s_addr >> 16), C(in.s_addr >> 8), C(in.s_addr)); } return (line); } int cmdnetstat(const char *cmd, const char *args) { if (prefix(cmd, "all")) { aflag = !aflag; goto fixup; } if (prefix(cmd, "numbers") || prefix(cmd, "names")) { struct netinfo *p; int new; new = prefix(cmd, "numbers"); if (new == nflag) return (1); TAILQ_FOREACH(p, &netcb, chain) { if (p->ni_line == -1) continue; p->ni_flags |= NIF_LACHG|NIF_FACHG; } nflag = new; goto redisplay; } if (!netcmd(cmd, args)) return (0); fixup: fetchnetstat(); redisplay: shownetstat(); refresh(); return (1); }