Index: head/sys/netinet/ip_input.c =================================================================== --- head/sys/netinet/ip_input.c (revision 359380) +++ head/sys/netinet/ip_input.c (revision 359381) @@ -1,1429 +1,1429 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1988, 1993 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)ip_input.c 8.2 (Berkeley) 1/4/94 */ #include __FBSDID("$FreeBSD$"); #include "opt_bootp.h" #include "opt_ipstealth.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 #include #include #include #include #include #include #ifdef CTASSERT CTASSERT(sizeof(struct ip) == 20); #endif /* IP reassembly functions are defined in ip_reass.c. */ extern void ipreass_init(void); extern void ipreass_drain(void); extern void ipreass_slowtimo(void); #ifdef VIMAGE extern void ipreass_destroy(void); #endif struct rmlock in_ifaddr_lock; RM_SYSINIT(in_ifaddr_lock, &in_ifaddr_lock, "in_ifaddr_lock"); VNET_DEFINE(int, rsvp_on); VNET_DEFINE(int, ipforwarding); SYSCTL_INT(_net_inet_ip, IPCTL_FORWARDING, forwarding, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(ipforwarding), 0, "Enable IP forwarding between interfaces"); VNET_DEFINE_STATIC(int, ipsendredirects) = 1; /* XXX */ #define V_ipsendredirects VNET(ipsendredirects) SYSCTL_INT(_net_inet_ip, IPCTL_SENDREDIRECTS, redirect, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(ipsendredirects), 0, "Enable sending IP redirects"); /* * XXX - Setting ip_checkinterface mostly implements the receive side of * the Strong ES model described in RFC 1122, but since the routing table * and transmit implementation do not implement the Strong ES model, * setting this to 1 results in an odd hybrid. * * XXX - ip_checkinterface currently must be disabled if you use ipnat * to translate the destination address to another local interface. * * XXX - ip_checkinterface must be disabled if you add IP aliases * to the loopback interface instead of the interface where the * packets for those addresses are received. */ VNET_DEFINE_STATIC(int, ip_checkinterface); #define V_ip_checkinterface VNET(ip_checkinterface) SYSCTL_INT(_net_inet_ip, OID_AUTO, check_interface, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(ip_checkinterface), 0, "Verify packet arrives on correct interface"); VNET_DEFINE(pfil_head_t, inet_pfil_head); /* Packet filter hooks */ static struct netisr_handler ip_nh = { .nh_name = "ip", .nh_handler = ip_input, .nh_proto = NETISR_IP, #ifdef RSS .nh_m2cpuid = rss_soft_m2cpuid_v4, .nh_policy = NETISR_POLICY_CPU, .nh_dispatch = NETISR_DISPATCH_HYBRID, #else .nh_policy = NETISR_POLICY_FLOW, #endif }; #ifdef RSS /* * Directly dispatched frames are currently assumed * to have a flowid already calculated. * * It should likely have something that assert it * actually has valid flow details. */ static struct netisr_handler ip_direct_nh = { .nh_name = "ip_direct", .nh_handler = ip_direct_input, .nh_proto = NETISR_IP_DIRECT, .nh_m2cpuid = rss_soft_m2cpuid_v4, .nh_policy = NETISR_POLICY_CPU, .nh_dispatch = NETISR_DISPATCH_HYBRID, }; #endif extern struct domain inetdomain; extern struct protosw inetsw[]; u_char ip_protox[IPPROTO_MAX]; VNET_DEFINE(struct in_ifaddrhead, in_ifaddrhead); /* first inet address */ VNET_DEFINE(struct in_ifaddrhashhead *, in_ifaddrhashtbl); /* inet addr hash table */ VNET_DEFINE(u_long, in_ifaddrhmask); /* mask for hash table */ #ifdef IPCTL_DEFMTU SYSCTL_INT(_net_inet_ip, IPCTL_DEFMTU, mtu, CTLFLAG_RW, &ip_mtu, 0, "Default MTU"); #endif #ifdef IPSTEALTH VNET_DEFINE(int, ipstealth); SYSCTL_INT(_net_inet_ip, OID_AUTO, stealth, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(ipstealth), 0, "IP stealth mode, no TTL decrementation on forwarding"); #endif /* * IP statistics are stored in the "array" of counter(9)s. */ VNET_PCPUSTAT_DEFINE(struct ipstat, ipstat); VNET_PCPUSTAT_SYSINIT(ipstat); SYSCTL_VNET_PCPUSTAT(_net_inet_ip, IPCTL_STATS, stats, struct ipstat, ipstat, "IP statistics (struct ipstat, netinet/ip_var.h)"); #ifdef VIMAGE VNET_PCPUSTAT_SYSUNINIT(ipstat); #endif /* VIMAGE */ /* * Kernel module interface for updating ipstat. The argument is an index * into ipstat treated as an array. */ void kmod_ipstat_inc(int statnum) { counter_u64_add(VNET(ipstat)[statnum], 1); } void kmod_ipstat_dec(int statnum) { counter_u64_add(VNET(ipstat)[statnum], -1); } static int sysctl_netinet_intr_queue_maxlen(SYSCTL_HANDLER_ARGS) { int error, qlimit; netisr_getqlimit(&ip_nh, &qlimit); error = sysctl_handle_int(oidp, &qlimit, 0, req); if (error || !req->newptr) return (error); if (qlimit < 1) return (EINVAL); return (netisr_setqlimit(&ip_nh, qlimit)); } SYSCTL_PROC(_net_inet_ip, IPCTL_INTRQMAXLEN, intr_queue_maxlen, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 0, sysctl_netinet_intr_queue_maxlen, "I", "Maximum size of the IP input queue"); static int sysctl_netinet_intr_queue_drops(SYSCTL_HANDLER_ARGS) { u_int64_t qdrops_long; int error, qdrops; netisr_getqdrops(&ip_nh, &qdrops_long); qdrops = qdrops_long; error = sysctl_handle_int(oidp, &qdrops, 0, req); if (error || !req->newptr) return (error); if (qdrops != 0) return (EINVAL); netisr_clearqdrops(&ip_nh); return (0); } SYSCTL_PROC(_net_inet_ip, IPCTL_INTRQDROPS, intr_queue_drops, CTLTYPE_INT | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0, sysctl_netinet_intr_queue_drops, "I", "Number of packets dropped from the IP input queue"); #ifdef RSS static int sysctl_netinet_intr_direct_queue_maxlen(SYSCTL_HANDLER_ARGS) { int error, qlimit; netisr_getqlimit(&ip_direct_nh, &qlimit); error = sysctl_handle_int(oidp, &qlimit, 0, req); if (error || !req->newptr) return (error); if (qlimit < 1) return (EINVAL); return (netisr_setqlimit(&ip_direct_nh, qlimit)); } SYSCTL_PROC(_net_inet_ip, IPCTL_INTRDQMAXLEN, intr_direct_queue_maxlen, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 0, sysctl_netinet_intr_direct_queue_maxlen, "I", "Maximum size of the IP direct input queue"); static int sysctl_netinet_intr_direct_queue_drops(SYSCTL_HANDLER_ARGS) { u_int64_t qdrops_long; int error, qdrops; netisr_getqdrops(&ip_direct_nh, &qdrops_long); qdrops = qdrops_long; error = sysctl_handle_int(oidp, &qdrops, 0, req); if (error || !req->newptr) return (error); if (qdrops != 0) return (EINVAL); netisr_clearqdrops(&ip_direct_nh); return (0); } SYSCTL_PROC(_net_inet_ip, IPCTL_INTRDQDROPS, intr_direct_queue_drops, CTLTYPE_INT | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0, sysctl_netinet_intr_direct_queue_drops, "I", "Number of packets dropped from the IP direct input queue"); #endif /* RSS */ /* * IP initialization: fill in IP protocol switch table. * All protocols not implemented in kernel go to raw IP protocol handler. */ void ip_init(void) { struct pfil_head_args args; struct protosw *pr; int i; CK_STAILQ_INIT(&V_in_ifaddrhead); V_in_ifaddrhashtbl = hashinit(INADDR_NHASH, M_IFADDR, &V_in_ifaddrhmask); /* Initialize IP reassembly queue. */ ipreass_init(); /* Initialize packet filter hooks. */ args.pa_version = PFIL_VERSION; args.pa_flags = PFIL_IN | PFIL_OUT; args.pa_type = PFIL_TYPE_IP4; args.pa_headname = PFIL_INET_NAME; V_inet_pfil_head = pfil_head_register(&args); if (hhook_head_register(HHOOK_TYPE_IPSEC_IN, AF_INET, &V_ipsec_hhh_in[HHOOK_IPSEC_INET], HHOOK_WAITOK | HHOOK_HEADISINVNET) != 0) printf("%s: WARNING: unable to register input helper hook\n", __func__); if (hhook_head_register(HHOOK_TYPE_IPSEC_OUT, AF_INET, &V_ipsec_hhh_out[HHOOK_IPSEC_INET], HHOOK_WAITOK | HHOOK_HEADISINVNET) != 0) printf("%s: WARNING: unable to register output helper hook\n", __func__); /* Skip initialization of globals for non-default instances. */ #ifdef VIMAGE if (!IS_DEFAULT_VNET(curvnet)) { netisr_register_vnet(&ip_nh); #ifdef RSS netisr_register_vnet(&ip_direct_nh); #endif return; } #endif pr = pffindproto(PF_INET, IPPROTO_RAW, SOCK_RAW); if (pr == NULL) panic("ip_init: PF_INET not found"); /* Initialize the entire ip_protox[] array to IPPROTO_RAW. */ for (i = 0; i < IPPROTO_MAX; i++) ip_protox[i] = pr - inetsw; /* * Cycle through IP protocols and put them into the appropriate place * in ip_protox[]. */ for (pr = inetdomain.dom_protosw; pr < inetdomain.dom_protoswNPROTOSW; pr++) if (pr->pr_domain->dom_family == PF_INET && pr->pr_protocol && pr->pr_protocol != IPPROTO_RAW) { /* Be careful to only index valid IP protocols. */ if (pr->pr_protocol < IPPROTO_MAX) ip_protox[pr->pr_protocol] = pr - inetsw; } netisr_register(&ip_nh); #ifdef RSS netisr_register(&ip_direct_nh); #endif } #ifdef VIMAGE static void ip_destroy(void *unused __unused) { struct ifnet *ifp; int error; #ifdef RSS netisr_unregister_vnet(&ip_direct_nh); #endif netisr_unregister_vnet(&ip_nh); pfil_head_unregister(V_inet_pfil_head); error = hhook_head_deregister(V_ipsec_hhh_in[HHOOK_IPSEC_INET]); if (error != 0) { printf("%s: WARNING: unable to deregister input helper hook " "type HHOOK_TYPE_IPSEC_IN, id HHOOK_IPSEC_INET: " "error %d returned\n", __func__, error); } error = hhook_head_deregister(V_ipsec_hhh_out[HHOOK_IPSEC_INET]); if (error != 0) { printf("%s: WARNING: unable to deregister output helper hook " "type HHOOK_TYPE_IPSEC_OUT, id HHOOK_IPSEC_INET: " "error %d returned\n", __func__, error); } /* Remove the IPv4 addresses from all interfaces. */ in_ifscrub_all(); /* Make sure the IPv4 routes are gone as well. */ IFNET_RLOCK(); CK_STAILQ_FOREACH(ifp, &V_ifnet, if_link) rt_flushifroutes_af(ifp, AF_INET); IFNET_RUNLOCK(); /* Destroy IP reassembly queue. */ ipreass_destroy(); /* Cleanup in_ifaddr hash table; should be empty. */ hashdestroy(V_in_ifaddrhashtbl, M_IFADDR, V_in_ifaddrhmask); } VNET_SYSUNINIT(ip, SI_SUB_PROTO_DOMAIN, SI_ORDER_THIRD, ip_destroy, NULL); #endif #ifdef RSS /* * IP direct input routine. * * This is called when reinjecting completed fragments where * all of the previous checking and book-keeping has been done. */ void ip_direct_input(struct mbuf *m) { struct ip *ip; int hlen; ip = mtod(m, struct ip *); hlen = ip->ip_hl << 2; #if defined(IPSEC) || defined(IPSEC_SUPPORT) if (IPSEC_ENABLED(ipv4)) { if (IPSEC_INPUT(ipv4, m, hlen, ip->ip_p) != 0) return; } #endif /* IPSEC */ IPSTAT_INC(ips_delivered); (*inetsw[ip_protox[ip->ip_p]].pr_input)(&m, &hlen, ip->ip_p); return; } #endif /* * Ip input routine. Checksum and byte swap header. If fragmented * try to reassemble. Process options. Pass to next level. */ void ip_input(struct mbuf *m) { struct rm_priotracker in_ifa_tracker; struct ip *ip = NULL; struct in_ifaddr *ia = NULL; struct ifaddr *ifa; struct ifnet *ifp; int checkif, hlen = 0; uint16_t sum, ip_len; int dchg = 0; /* dest changed after fw */ struct in_addr odst; /* original dst address */ M_ASSERTPKTHDR(m); NET_EPOCH_ASSERT(); if (m->m_flags & M_FASTFWD_OURS) { m->m_flags &= ~M_FASTFWD_OURS; /* Set up some basics that will be used later. */ ip = mtod(m, struct ip *); hlen = ip->ip_hl << 2; ip_len = ntohs(ip->ip_len); goto ours; } IPSTAT_INC(ips_total); if (m->m_pkthdr.len < sizeof(struct ip)) goto tooshort; if (m->m_len < sizeof (struct ip) && (m = m_pullup(m, sizeof (struct ip))) == NULL) { IPSTAT_INC(ips_toosmall); return; } ip = mtod(m, struct ip *); if (ip->ip_v != IPVERSION) { IPSTAT_INC(ips_badvers); goto bad; } hlen = ip->ip_hl << 2; if (hlen < sizeof(struct ip)) { /* minimum header length */ IPSTAT_INC(ips_badhlen); goto bad; } if (hlen > m->m_len) { if ((m = m_pullup(m, hlen)) == NULL) { IPSTAT_INC(ips_badhlen); return; } ip = mtod(m, struct ip *); } IP_PROBE(receive, NULL, NULL, ip, m->m_pkthdr.rcvif, ip, NULL); /* IN_LOOPBACK must not appear on the wire - RFC1122 */ ifp = m->m_pkthdr.rcvif; if (IN_LOOPBACK(ntohl(ip->ip_dst.s_addr)) || IN_LOOPBACK(ntohl(ip->ip_src.s_addr))) { if ((ifp->if_flags & IFF_LOOPBACK) == 0) { IPSTAT_INC(ips_badaddr); goto bad; } } if (m->m_pkthdr.csum_flags & CSUM_IP_CHECKED) { sum = !(m->m_pkthdr.csum_flags & CSUM_IP_VALID); } else { if (hlen == sizeof(struct ip)) { sum = in_cksum_hdr(ip); } else { sum = in_cksum(m, hlen); } } if (sum) { IPSTAT_INC(ips_badsum); goto bad; } #ifdef ALTQ if (altq_input != NULL && (*altq_input)(m, AF_INET) == 0) /* packet is dropped by traffic conditioner */ return; #endif ip_len = ntohs(ip->ip_len); if (ip_len < hlen) { IPSTAT_INC(ips_badlen); goto bad; } /* * Check that the amount of data in the buffers * is as at least much as the IP header would have us expect. * Trim mbufs if longer than we expect. * Drop packet if shorter than we expect. */ if (m->m_pkthdr.len < ip_len) { tooshort: IPSTAT_INC(ips_tooshort); goto bad; } if (m->m_pkthdr.len > ip_len) { if (m->m_len == m->m_pkthdr.len) { m->m_len = ip_len; m->m_pkthdr.len = ip_len; } else m_adj(m, ip_len - m->m_pkthdr.len); } /* * Try to forward the packet, but if we fail continue. * ip_tryforward() does not generate redirects, so fall * through to normal processing if redirects are required. * ip_tryforward() does inbound and outbound packet firewall * processing. If firewall has decided that destination becomes * our local address, it sets M_FASTFWD_OURS flag. In this * case skip another inbound firewall processing and update * ip pointer. */ if (V_ipforwarding != 0 && V_ipsendredirects == 0 #if defined(IPSEC) || defined(IPSEC_SUPPORT) && (!IPSEC_ENABLED(ipv4) || IPSEC_CAPS(ipv4, m, IPSEC_CAP_OPERABLE) == 0) #endif ) { if ((m = ip_tryforward(m)) == NULL) return; if (m->m_flags & M_FASTFWD_OURS) { m->m_flags &= ~M_FASTFWD_OURS; ip = mtod(m, struct ip *); goto ours; } } #if defined(IPSEC) || defined(IPSEC_SUPPORT) /* * Bypass packet filtering for packets previously handled by IPsec. */ if (IPSEC_ENABLED(ipv4) && IPSEC_CAPS(ipv4, m, IPSEC_CAP_BYPASS_FILTER) != 0) goto passin; #endif /* * Run through list of hooks for input packets. * * NB: Beware of the destination address changing (e.g. * by NAT rewriting). When this happens, tell * ip_forward to do the right thing. */ /* Jump over all PFIL processing if hooks are not active. */ if (!PFIL_HOOKED_IN(V_inet_pfil_head)) goto passin; odst = ip->ip_dst; if (pfil_run_hooks(V_inet_pfil_head, &m, ifp, PFIL_IN, NULL) != PFIL_PASS) return; if (m == NULL) /* consumed by filter */ return; ip = mtod(m, struct ip *); dchg = (odst.s_addr != ip->ip_dst.s_addr); ifp = m->m_pkthdr.rcvif; if (m->m_flags & M_FASTFWD_OURS) { m->m_flags &= ~M_FASTFWD_OURS; goto ours; } if (m->m_flags & M_IP_NEXTHOP) { if (m_tag_find(m, PACKET_TAG_IPFORWARD, NULL) != NULL) { /* * Directly ship the packet on. This allows * forwarding packets originally destined to us * to some other directly connected host. */ ip_forward(m, 1); return; } } passin: /* * Process options and, if not destined for us, * ship it on. ip_dooptions returns 1 when an * error was detected (causing an icmp message * to be sent and the original packet to be freed). */ if (hlen > sizeof (struct ip) && ip_dooptions(m, 0)) return; /* greedy RSVP, snatches any PATH packet of the RSVP protocol and no * matter if it is destined to another node, or whether it is * a multicast one, RSVP wants it! and prevents it from being forwarded * anywhere else. Also checks if the rsvp daemon is running before * grabbing the packet. */ if (V_rsvp_on && ip->ip_p==IPPROTO_RSVP) goto ours; /* * Check our list of addresses, to see if the packet is for us. * If we don't have any addresses, assume any unicast packet * we receive might be for us (and let the upper layers deal * with it). */ if (CK_STAILQ_EMPTY(&V_in_ifaddrhead) && (m->m_flags & (M_MCAST|M_BCAST)) == 0) goto ours; /* * Enable a consistency check between the destination address * and the arrival interface for a unicast packet (the RFC 1122 * strong ES model) if IP forwarding is disabled and the packet * is not locally generated and the packet is not subject to * 'ipfw fwd'. * * XXX - Checking also should be disabled if the destination * address is ipnat'ed to a different interface. * * XXX - Checking is incompatible with IP aliases added * to the loopback interface instead of the interface where * the packets are received. * * XXX - This is the case for carp vhost IPs as well so we * insert a workaround. If the packet got here, we already * checked with carp_iamatch() and carp_forus(). */ checkif = V_ip_checkinterface && (V_ipforwarding == 0) && ifp != NULL && ((ifp->if_flags & IFF_LOOPBACK) == 0) && ifp->if_carp == NULL && (dchg == 0); /* * Check for exact addresses in the hash bucket. */ IN_IFADDR_RLOCK(&in_ifa_tracker); LIST_FOREACH(ia, INADDR_HASH(ip->ip_dst.s_addr), ia_hash) { /* * If the address matches, verify that the packet * arrived via the correct interface if checking is * enabled. */ if (IA_SIN(ia)->sin_addr.s_addr == ip->ip_dst.s_addr && (!checkif || ia->ia_ifp == ifp)) { counter_u64_add(ia->ia_ifa.ifa_ipackets, 1); counter_u64_add(ia->ia_ifa.ifa_ibytes, m->m_pkthdr.len); IN_IFADDR_RUNLOCK(&in_ifa_tracker); goto ours; } } IN_IFADDR_RUNLOCK(&in_ifa_tracker); /* * Check for broadcast addresses. * * Only accept broadcast packets that arrive via the matching * interface. Reception of forwarded directed broadcasts would * be handled via ip_forward() and ether_output() with the loopback * into the stack for SIMPLEX interfaces handled by ether_output(). */ if (ifp != NULL && ifp->if_flags & IFF_BROADCAST) { CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) { if (ifa->ifa_addr->sa_family != AF_INET) continue; ia = ifatoia(ifa); if (satosin(&ia->ia_broadaddr)->sin_addr.s_addr == ip->ip_dst.s_addr) { counter_u64_add(ia->ia_ifa.ifa_ipackets, 1); counter_u64_add(ia->ia_ifa.ifa_ibytes, m->m_pkthdr.len); goto ours; } #ifdef BOOTP_COMPAT if (IA_SIN(ia)->sin_addr.s_addr == INADDR_ANY) { counter_u64_add(ia->ia_ifa.ifa_ipackets, 1); counter_u64_add(ia->ia_ifa.ifa_ibytes, m->m_pkthdr.len); goto ours; } #endif } ia = NULL; } /* RFC 3927 2.7: Do not forward datagrams for 169.254.0.0/16. */ if (IN_LINKLOCAL(ntohl(ip->ip_dst.s_addr))) { IPSTAT_INC(ips_cantforward); m_freem(m); return; } if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr))) { if (V_ip_mrouter) { /* * If we are acting as a multicast router, all * incoming multicast packets are passed to the * kernel-level multicast forwarding function. * The packet is returned (relatively) intact; if * ip_mforward() returns a non-zero value, the packet * must be discarded, else it may be accepted below. */ if (ip_mforward && ip_mforward(ip, ifp, m, 0) != 0) { IPSTAT_INC(ips_cantforward); m_freem(m); return; } /* * The process-level routing daemon needs to receive * all multicast IGMP packets, whether or not this * host belongs to their destination groups. */ if (ip->ip_p == IPPROTO_IGMP) goto ours; IPSTAT_INC(ips_forward); } /* * Assume the packet is for us, to avoid prematurely taking * a lock on the in_multi hash. Protocols must perform * their own filtering and update statistics accordingly. */ goto ours; } if (ip->ip_dst.s_addr == (u_long)INADDR_BROADCAST) goto ours; if (ip->ip_dst.s_addr == INADDR_ANY) goto ours; /* * Not for us; forward if possible and desirable. */ if (V_ipforwarding == 0) { IPSTAT_INC(ips_cantforward); m_freem(m); } else { ip_forward(m, dchg); } return; ours: #ifdef IPSTEALTH /* * IPSTEALTH: Process non-routing options only * if the packet is destined for us. */ if (V_ipstealth && hlen > sizeof (struct ip) && ip_dooptions(m, 1)) return; #endif /* IPSTEALTH */ /* * Attempt reassembly; if it succeeds, proceed. * ip_reass() will return a different mbuf. */ if (ip->ip_off & htons(IP_MF | IP_OFFMASK)) { /* XXXGL: shouldn't we save & set m_flags? */ m = ip_reass(m); if (m == NULL) return; ip = mtod(m, struct ip *); /* Get the header length of the reassembled packet */ hlen = ip->ip_hl << 2; } #if defined(IPSEC) || defined(IPSEC_SUPPORT) if (IPSEC_ENABLED(ipv4)) { if (IPSEC_INPUT(ipv4, m, hlen, ip->ip_p) != 0) return; } #endif /* IPSEC */ /* * Switch out to protocol's input routine. */ IPSTAT_INC(ips_delivered); (*inetsw[ip_protox[ip->ip_p]].pr_input)(&m, &hlen, ip->ip_p); return; bad: m_freem(m); } /* * IP timer processing; * if a timer expires on a reassembly * queue, discard it. */ void ip_slowtimo(void) { VNET_ITERATOR_DECL(vnet_iter); VNET_LIST_RLOCK_NOSLEEP(); VNET_FOREACH(vnet_iter) { CURVNET_SET(vnet_iter); ipreass_slowtimo(); CURVNET_RESTORE(); } VNET_LIST_RUNLOCK_NOSLEEP(); } void ip_drain(void) { VNET_ITERATOR_DECL(vnet_iter); VNET_LIST_RLOCK_NOSLEEP(); VNET_FOREACH(vnet_iter) { CURVNET_SET(vnet_iter); ipreass_drain(); CURVNET_RESTORE(); } VNET_LIST_RUNLOCK_NOSLEEP(); } /* * The protocol to be inserted into ip_protox[] must be already registered * in inetsw[], either statically or through pf_proto_register(). */ int ipproto_register(short ipproto) { struct protosw *pr; /* Sanity checks. */ if (ipproto <= 0 || ipproto >= IPPROTO_MAX) return (EPROTONOSUPPORT); /* * The protocol slot must not be occupied by another protocol * already. An index pointing to IPPROTO_RAW is unused. */ pr = pffindproto(PF_INET, IPPROTO_RAW, SOCK_RAW); if (pr == NULL) return (EPFNOSUPPORT); if (ip_protox[ipproto] != pr - inetsw) /* IPPROTO_RAW */ return (EEXIST); /* Find the protocol position in inetsw[] and set the index. */ for (pr = inetdomain.dom_protosw; pr < inetdomain.dom_protoswNPROTOSW; pr++) { if (pr->pr_domain->dom_family == PF_INET && pr->pr_protocol && pr->pr_protocol == ipproto) { ip_protox[pr->pr_protocol] = pr - inetsw; return (0); } } return (EPROTONOSUPPORT); } int ipproto_unregister(short ipproto) { struct protosw *pr; /* Sanity checks. */ if (ipproto <= 0 || ipproto >= IPPROTO_MAX) return (EPROTONOSUPPORT); /* Check if the protocol was indeed registered. */ pr = pffindproto(PF_INET, IPPROTO_RAW, SOCK_RAW); if (pr == NULL) return (EPFNOSUPPORT); if (ip_protox[ipproto] == pr - inetsw) /* IPPROTO_RAW */ return (ENOENT); /* Reset the protocol slot to IPPROTO_RAW. */ ip_protox[ipproto] = pr - inetsw; return (0); } u_char inetctlerrmap[PRC_NCMDS] = { 0, 0, 0, 0, 0, EMSGSIZE, EHOSTDOWN, EHOSTUNREACH, EHOSTUNREACH, EHOSTUNREACH, ECONNREFUSED, ECONNREFUSED, EMSGSIZE, EHOSTUNREACH, 0, 0, 0, 0, EHOSTUNREACH, 0, ENOPROTOOPT, ECONNREFUSED }; /* * Forward a packet. If some error occurs return the sender * an icmp packet. Note we can't always generate a meaningful * icmp message because icmp doesn't have a large enough repertoire * of codes and types. * * If not forwarding, just drop the packet. This could be confusing * if ipforwarding was zero but some routing protocol was advancing * us as a gateway to somewhere. However, we must let the routing * protocol deal with that. * * The srcrt parameter indicates whether the packet is being forwarded * via a source route. */ void ip_forward(struct mbuf *m, int srcrt) { struct ip *ip = mtod(m, struct ip *); struct in_ifaddr *ia; struct mbuf *mcopy; struct sockaddr_in *sin; struct in_addr dest; struct route ro; int error, type = 0, code = 0, mtu = 0; NET_EPOCH_ASSERT(); if (m->m_flags & (M_BCAST|M_MCAST) || in_canforward(ip->ip_dst) == 0) { IPSTAT_INC(ips_cantforward); m_freem(m); return; } if ( #ifdef IPSTEALTH V_ipstealth == 0 && #endif ip->ip_ttl <= IPTTLDEC) { icmp_error(m, ICMP_TIMXCEED, ICMP_TIMXCEED_INTRANS, 0, 0); return; } bzero(&ro, sizeof(ro)); sin = (struct sockaddr_in *)&ro.ro_dst; sin->sin_family = AF_INET; sin->sin_len = sizeof(*sin); sin->sin_addr = ip->ip_dst; #ifdef RADIX_MPATH rtalloc_mpath_fib(&ro, ntohl(ip->ip_src.s_addr ^ ip->ip_dst.s_addr), M_GETFIB(m)); #else in_rtalloc_ign(&ro, 0, M_GETFIB(m)); #endif if (ro.ro_rt != NULL) { ia = ifatoia(ro.ro_rt->rt_ifa); } else ia = NULL; /* * Save the IP header and at most 8 bytes of the payload, * in case we need to generate an ICMP message to the src. * * XXX this can be optimized a lot by saving the data in a local * buffer on the stack (72 bytes at most), and only allocating the * mbuf if really necessary. The vast majority of the packets * are forwarded without having to send an ICMP back (either * because unnecessary, or because rate limited), so we are * really we are wasting a lot of work here. * * We don't use m_copym() because it might return a reference * to a shared cluster. Both this function and ip_output() * assume exclusive access to the IP header in `m', so any * data in a cluster may change before we reach icmp_error(). */ mcopy = m_gethdr(M_NOWAIT, m->m_type); if (mcopy != NULL && !m_dup_pkthdr(mcopy, m, M_NOWAIT)) { /* * It's probably ok if the pkthdr dup fails (because * the deep copy of the tag chain failed), but for now * be conservative and just discard the copy since * code below may some day want the tags. */ m_free(mcopy); mcopy = NULL; } if (mcopy != NULL) { mcopy->m_len = min(ntohs(ip->ip_len), M_TRAILINGSPACE(mcopy)); mcopy->m_pkthdr.len = mcopy->m_len; m_copydata(m, 0, mcopy->m_len, mtod(mcopy, caddr_t)); } #ifdef IPSTEALTH if (V_ipstealth == 0) #endif ip->ip_ttl -= IPTTLDEC; #if defined(IPSEC) || defined(IPSEC_SUPPORT) if (IPSEC_ENABLED(ipv4)) { if ((error = IPSEC_FORWARD(ipv4, m)) != 0) { /* mbuf consumed by IPsec */ m_freem(mcopy); if (error != EINPROGRESS) IPSTAT_INC(ips_cantforward); return; } /* No IPsec processing required */ } #endif /* IPSEC */ /* * If forwarding packet using same interface that it came in on, * perhaps should send a redirect to sender to shortcut a hop. * Only send redirect if source is sending directly to us, * and if packet was not source routed (or has any options). * Also, don't send redirect if forwarding using a default route * or a route modified by a redirect. */ dest.s_addr = 0; if (!srcrt && V_ipsendredirects && ia != NULL && ia->ia_ifp == m->m_pkthdr.rcvif) { struct rtentry *rt; rt = ro.ro_rt; if (rt && (rt->rt_flags & (RTF_DYNAMIC|RTF_MODIFIED)) == 0 && satosin(rt_key(rt))->sin_addr.s_addr != 0) { #define RTA(rt) ((struct in_ifaddr *)(rt->rt_ifa)) u_long src = ntohl(ip->ip_src.s_addr); if (RTA(rt) && (src & RTA(rt)->ia_subnetmask) == RTA(rt)->ia_subnet) { if (rt->rt_flags & RTF_GATEWAY) dest.s_addr = satosin(rt->rt_gateway)->sin_addr.s_addr; else dest.s_addr = ip->ip_dst.s_addr; /* Router requirements says to only send host redirects */ type = ICMP_REDIRECT; code = ICMP_REDIRECT_HOST; } } } error = ip_output(m, NULL, &ro, IP_FORWARDING, NULL, NULL); if (error == EMSGSIZE && ro.ro_rt) mtu = ro.ro_rt->rt_mtu; RO_RTFREE(&ro); if (error) IPSTAT_INC(ips_cantforward); else { IPSTAT_INC(ips_forward); if (type) IPSTAT_INC(ips_redirectsent); else { if (mcopy) m_freem(mcopy); return; } } if (mcopy == NULL) return; switch (error) { case 0: /* forwarded, but need redirect */ /* type, code set above */ break; case ENETUNREACH: case EHOSTUNREACH: case ENETDOWN: case EHOSTDOWN: default: type = ICMP_UNREACH; code = ICMP_UNREACH_HOST; break; case EMSGSIZE: type = ICMP_UNREACH; code = ICMP_UNREACH_NEEDFRAG; /* * If the MTU was set before make sure we are below the * interface MTU. * If the MTU wasn't set before use the interface mtu or * fall back to the next smaller mtu step compared to the * current packet size. */ if (mtu != 0) { if (ia != NULL) mtu = min(mtu, ia->ia_ifp->if_mtu); } else { if (ia != NULL) mtu = ia->ia_ifp->if_mtu; else mtu = ip_next_mtu(ntohs(ip->ip_len), 0); } IPSTAT_INC(ips_cantfrag); break; case ENOBUFS: case EACCES: /* ipfw denied packet */ m_freem(mcopy); return; } icmp_error(mcopy, type, code, dest.s_addr, mtu); } #define CHECK_SO_CT(sp, ct) \ (((sp->so_options & SO_TIMESTAMP) && (sp->so_ts_clock == ct)) ? 1 : 0) void ip_savecontrol(struct inpcb *inp, struct mbuf **mp, struct ip *ip, struct mbuf *m) { bool stamped; stamped = false; if ((inp->inp_socket->so_options & SO_BINTIME) || CHECK_SO_CT(inp->inp_socket, SO_TS_BINTIME)) { struct bintime boottimebin, bt; struct timespec ts1; if ((m->m_flags & (M_PKTHDR | M_TSTMP)) == (M_PKTHDR | M_TSTMP)) { mbuf_tstmp2timespec(m, &ts1); timespec2bintime(&ts1, &bt); getboottimebin(&boottimebin); bintime_add(&bt, &boottimebin); } else { bintime(&bt); } *mp = sbcreatecontrol((caddr_t)&bt, sizeof(bt), SCM_BINTIME, SOL_SOCKET); if (*mp != NULL) { mp = &(*mp)->m_next; stamped = true; } } if (CHECK_SO_CT(inp->inp_socket, SO_TS_REALTIME_MICRO)) { struct bintime boottimebin, bt1; - struct timespec ts1;; + struct timespec ts1; struct timeval tv; if ((m->m_flags & (M_PKTHDR | M_TSTMP)) == (M_PKTHDR | M_TSTMP)) { mbuf_tstmp2timespec(m, &ts1); timespec2bintime(&ts1, &bt1); getboottimebin(&boottimebin); bintime_add(&bt1, &boottimebin); bintime2timeval(&bt1, &tv); } else { microtime(&tv); } *mp = sbcreatecontrol((caddr_t)&tv, sizeof(tv), SCM_TIMESTAMP, SOL_SOCKET); if (*mp != NULL) { mp = &(*mp)->m_next; stamped = true; } } else if (CHECK_SO_CT(inp->inp_socket, SO_TS_REALTIME)) { struct bintime boottimebin; struct timespec ts, ts1; if ((m->m_flags & (M_PKTHDR | M_TSTMP)) == (M_PKTHDR | M_TSTMP)) { mbuf_tstmp2timespec(m, &ts); getboottimebin(&boottimebin); bintime2timespec(&boottimebin, &ts1); timespecadd(&ts, &ts1, &ts); } else { nanotime(&ts); } *mp = sbcreatecontrol((caddr_t)&ts, sizeof(ts), SCM_REALTIME, SOL_SOCKET); if (*mp != NULL) { mp = &(*mp)->m_next; stamped = true; } } else if (CHECK_SO_CT(inp->inp_socket, SO_TS_MONOTONIC)) { struct timespec ts; if ((m->m_flags & (M_PKTHDR | M_TSTMP)) == (M_PKTHDR | M_TSTMP)) mbuf_tstmp2timespec(m, &ts); else nanouptime(&ts); *mp = sbcreatecontrol((caddr_t)&ts, sizeof(ts), SCM_MONOTONIC, SOL_SOCKET); if (*mp != NULL) { mp = &(*mp)->m_next; stamped = true; } } if (stamped && (m->m_flags & (M_PKTHDR | M_TSTMP)) == (M_PKTHDR | M_TSTMP)) { struct sock_timestamp_info sti; bzero(&sti, sizeof(sti)); sti.st_info_flags = ST_INFO_HW; if ((m->m_flags & M_TSTMP_HPREC) != 0) sti.st_info_flags |= ST_INFO_HW_HPREC; *mp = sbcreatecontrol((caddr_t)&sti, sizeof(sti), SCM_TIME_INFO, SOL_SOCKET); if (*mp != NULL) mp = &(*mp)->m_next; } if (inp->inp_flags & INP_RECVDSTADDR) { *mp = sbcreatecontrol((caddr_t)&ip->ip_dst, sizeof(struct in_addr), IP_RECVDSTADDR, IPPROTO_IP); if (*mp) mp = &(*mp)->m_next; } if (inp->inp_flags & INP_RECVTTL) { *mp = sbcreatecontrol((caddr_t)&ip->ip_ttl, sizeof(u_char), IP_RECVTTL, IPPROTO_IP); if (*mp) mp = &(*mp)->m_next; } #ifdef notyet /* XXX * Moving these out of udp_input() made them even more broken * than they already were. */ /* options were tossed already */ if (inp->inp_flags & INP_RECVOPTS) { *mp = sbcreatecontrol((caddr_t)opts_deleted_above, sizeof(struct in_addr), IP_RECVOPTS, IPPROTO_IP); if (*mp) mp = &(*mp)->m_next; } /* ip_srcroute doesn't do what we want here, need to fix */ if (inp->inp_flags & INP_RECVRETOPTS) { *mp = sbcreatecontrol((caddr_t)ip_srcroute(m), sizeof(struct in_addr), IP_RECVRETOPTS, IPPROTO_IP); if (*mp) mp = &(*mp)->m_next; } #endif if (inp->inp_flags & INP_RECVIF) { struct ifnet *ifp; struct sdlbuf { struct sockaddr_dl sdl; u_char pad[32]; } sdlbuf; struct sockaddr_dl *sdp; struct sockaddr_dl *sdl2 = &sdlbuf.sdl; if ((ifp = m->m_pkthdr.rcvif) && ifp->if_index && ifp->if_index <= V_if_index) { sdp = (struct sockaddr_dl *)ifp->if_addr->ifa_addr; /* * Change our mind and don't try copy. */ if (sdp->sdl_family != AF_LINK || sdp->sdl_len > sizeof(sdlbuf)) { goto makedummy; } bcopy(sdp, sdl2, sdp->sdl_len); } else { makedummy: sdl2->sdl_len = offsetof(struct sockaddr_dl, sdl_data[0]); sdl2->sdl_family = AF_LINK; sdl2->sdl_index = 0; sdl2->sdl_nlen = sdl2->sdl_alen = sdl2->sdl_slen = 0; } *mp = sbcreatecontrol((caddr_t)sdl2, sdl2->sdl_len, IP_RECVIF, IPPROTO_IP); if (*mp) mp = &(*mp)->m_next; } if (inp->inp_flags & INP_RECVTOS) { *mp = sbcreatecontrol((caddr_t)&ip->ip_tos, sizeof(u_char), IP_RECVTOS, IPPROTO_IP); if (*mp) mp = &(*mp)->m_next; } if (inp->inp_flags2 & INP_RECVFLOWID) { uint32_t flowid, flow_type; flowid = m->m_pkthdr.flowid; flow_type = M_HASHTYPE_GET(m); /* * XXX should handle the failure of one or the * other - don't populate both? */ *mp = sbcreatecontrol((caddr_t) &flowid, sizeof(uint32_t), IP_FLOWID, IPPROTO_IP); if (*mp) mp = &(*mp)->m_next; *mp = sbcreatecontrol((caddr_t) &flow_type, sizeof(uint32_t), IP_FLOWTYPE, IPPROTO_IP); if (*mp) mp = &(*mp)->m_next; } #ifdef RSS if (inp->inp_flags2 & INP_RECVRSSBUCKETID) { uint32_t flowid, flow_type; uint32_t rss_bucketid; flowid = m->m_pkthdr.flowid; flow_type = M_HASHTYPE_GET(m); if (rss_hash2bucket(flowid, flow_type, &rss_bucketid) == 0) { *mp = sbcreatecontrol((caddr_t) &rss_bucketid, sizeof(uint32_t), IP_RSSBUCKETID, IPPROTO_IP); if (*mp) mp = &(*mp)->m_next; } } #endif } /* * XXXRW: Multicast routing code in ip_mroute.c is generally MPSAFE, but the * ip_rsvp and ip_rsvp_on variables need to be interlocked with rsvp_on * locking. This code remains in ip_input.c as ip_mroute.c is optionally * compiled. */ VNET_DEFINE_STATIC(int, ip_rsvp_on); VNET_DEFINE(struct socket *, ip_rsvpd); #define V_ip_rsvp_on VNET(ip_rsvp_on) int ip_rsvp_init(struct socket *so) { if (so->so_type != SOCK_RAW || so->so_proto->pr_protocol != IPPROTO_RSVP) return EOPNOTSUPP; if (V_ip_rsvpd != NULL) return EADDRINUSE; V_ip_rsvpd = so; /* * This may seem silly, but we need to be sure we don't over-increment * the RSVP counter, in case something slips up. */ if (!V_ip_rsvp_on) { V_ip_rsvp_on = 1; V_rsvp_on++; } return 0; } int ip_rsvp_done(void) { V_ip_rsvpd = NULL; /* * This may seem silly, but we need to be sure we don't over-decrement * the RSVP counter, in case something slips up. */ if (V_ip_rsvp_on) { V_ip_rsvp_on = 0; V_rsvp_on--; } return 0; } int rsvp_input(struct mbuf **mp, int *offp, int proto) { struct mbuf *m; m = *mp; *mp = NULL; if (rsvp_input_p) { /* call the real one if loaded */ *mp = m; rsvp_input_p(mp, offp, proto); return (IPPROTO_DONE); } /* Can still get packets with rsvp_on = 0 if there is a local member * of the group to which the RSVP packet is addressed. But in this * case we want to throw the packet away. */ if (!V_rsvp_on) { m_freem(m); return (IPPROTO_DONE); } if (V_ip_rsvpd != NULL) { *mp = m; rip_input(mp, offp, proto); return (IPPROTO_DONE); } /* Drop the packet */ m_freem(m); return (IPPROTO_DONE); } Index: head/sys/netinet/tcp_ratelimit.c =================================================================== --- head/sys/netinet/tcp_ratelimit.c (revision 359380) +++ head/sys/netinet/tcp_ratelimit.c (revision 359381) @@ -1,1503 +1,1503 @@ /*- * * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 2018-2019 * Netflix 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 REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * */ /** * Author: Randall Stewart */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_tcpdebug.h" #include "opt_ratelimit.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define TCPSTATES /* for logging */ #include #ifdef INET6 #include #endif #include #ifndef USECS_IN_SECOND #define USECS_IN_SECOND 1000000 #endif /* * For the purposes of each send, what is the size * of an ethernet frame. */ MALLOC_DEFINE(M_TCPPACE, "tcp_hwpace", "TCP Hardware pacing memory"); #ifdef RATELIMIT /* * The following preferred table will seem weird to * the casual viewer. Why do we not have any rates below * 1Mbps? Why do we have a rate at 1.44Mbps called common? * Why do the rates cluster in the 1-100Mbps range more * than others? Why does the table jump around at the beginnign * and then be more consistently raising? * * Let me try to answer those questions. A lot of * this is dependant on the hardware. We have three basic * supporters of rate limiting * * Chelsio - Supporting 16 configurable rates. * Mlx - c4 supporting 13 fixed rates. * Mlx - c5 & c6 supporting 127 configurable rates. * * The c4 is why we have a common rate that is available * in all rate tables. This is a selected rate from the * c4 table and we assure its available in all ratelimit * tables. This way the tcp_ratelimit code has an assured * rate it should always be able to get. This answers a * couple of the questions above. * * So what about the rest, well the table is built to * try to get the most out of a joint hardware/software * pacing system. The software pacer will always pick * a rate higher than the b/w that it is estimating * * on the path. This is done for two reasons. * a) So we can discover more b/w * and * b) So we can send a block of MSS's down and then * have the software timer go off after the previous * send is completely out of the hardware. * * But when we do we don't want to have the delay * between the last packet sent by the hardware be * excessively long (to reach our desired rate). * * So let me give an example for clarity. * * Lets assume that the tcp stack sees that 29,110,000 bps is * what the bw of the path is. The stack would select the * rate 31Mbps. 31Mbps means that each send that is done * by the hardware will cause a 390 micro-second gap between * the packets sent at that rate. For 29,110,000 bps we * would need 416 micro-seconds gap between each send. * * Note that are calculating a complete time for pacing * which includes the ethernet, IP and TCP overhead. So * a full 1514 bytes is used for the above calculations. * My testing has shown that both cards are also using this * as their basis i.e. full payload size of the ethernet frame. * The TCP stack caller needs to be aware of this and make the * appropriate overhead calculations be included in its choices. * * Now, continuing our example, we pick a MSS size based on the * delta between the two rates (416 - 390) divided into the rate * we really wish to send at rounded up. That results in a MSS * send of 17 mss's at once. The hardware then will * run out of data in a single 17MSS send in 6,630 micro-seconds. * * On the other hand the software pacer will send more data * in 7,072 micro-seconds. This means that we will refill * the hardware 52 microseconds after it would have sent * next if it had not ran out of data. This is a win since we are * only sending every 7ms or so and yet all the packets are spaced on * the wire with 94% of what they should be and only * the last packet is delayed extra to make up for the * difference. * * Note that the above formula has two important caveat. * If we are above (b/w wise) over 100Mbps we double the result * of the MSS calculation. The second caveat is if we are 500Mbps * or more we just send the maximum MSS at once i.e. 45MSS. At * the higher b/w's even the cards have limits to what times (timer granularity) * they can insert between packets and start to send more than one * packet at a time on the wire. * */ #define COMMON_RATE 180500 const uint64_t desired_rates[] = { 122500, /* 1Mbps - rate 1 */ 180500, /* 1.44Mpbs - rate 2 common rate */ 375000, /* 3Mbps - rate 3 */ 625000, /* 5Mbps - rate 4 */ 875000, /* 7Mbps - rate 5 */ 1125000, /* 9Mbps - rate 6 */ 1375000, /* 11Mbps - rate 7 */ 1625000, /* 13Mbps - rate 8 */ 2625000, /* 21Mbps - rate 9 */ 3875000, /* 31Mbps - rate 10 */ 5125000, /* 41Meg - rate 11 */ 12500000, /* 100Mbps - rate 12 */ 25000000, /* 200Mbps - rate 13 */ 50000000, /* 400Mbps - rate 14 */ 63750000, /* 51Mbps - rate 15 */ 100000000, /* 800Mbps - rate 16 */ 1875000, /* 15Mbps - rate 17 */ 2125000, /* 17Mbps - rate 18 */ 2375000, /* 19Mbps - rate 19 */ 2875000, /* 23Mbps - rate 20 */ 3125000, /* 25Mbps - rate 21 */ 3375000, /* 27Mbps - rate 22 */ 3625000, /* 29Mbps - rate 23 */ 4125000, /* 33Mbps - rate 24 */ 4375000, /* 35Mbps - rate 25 */ 4625000, /* 37Mbps - rate 26 */ 4875000, /* 39Mbps - rate 27 */ 5375000, /* 43Mbps - rate 28 */ 5625000, /* 45Mbps - rate 29 */ 5875000, /* 47Mbps - rate 30 */ 6125000, /* 49Mbps - rate 31 */ 6625000, /* 53Mbps - rate 32 */ 6875000, /* 55Mbps - rate 33 */ 7125000, /* 57Mbps - rate 34 */ 7375000, /* 59Mbps - rate 35 */ 7625000, /* 61Mbps - rate 36 */ 7875000, /* 63Mbps - rate 37 */ 8125000, /* 65Mbps - rate 38 */ 8375000, /* 67Mbps - rate 39 */ 8625000, /* 69Mbps - rate 40 */ 8875000, /* 71Mbps - rate 41 */ 9125000, /* 73Mbps - rate 42 */ 9375000, /* 75Mbps - rate 43 */ 9625000, /* 77Mbps - rate 44 */ 9875000, /* 79Mbps - rate 45 */ 10125000, /* 81Mbps - rate 46 */ 10375000, /* 83Mbps - rate 47 */ 10625000, /* 85Mbps - rate 48 */ 10875000, /* 87Mbps - rate 49 */ 11125000, /* 89Mbps - rate 50 */ 11375000, /* 91Mbps - rate 51 */ 11625000, /* 93Mbps - rate 52 */ 11875000, /* 95Mbps - rate 53 */ 13125000, /* 105Mbps - rate 54 */ 13750000, /* 110Mbps - rate 55 */ 14375000, /* 115Mbps - rate 56 */ 15000000, /* 120Mbps - rate 57 */ 15625000, /* 125Mbps - rate 58 */ 16250000, /* 130Mbps - rate 59 */ 16875000, /* 135Mbps - rate 60 */ 17500000, /* 140Mbps - rate 61 */ 18125000, /* 145Mbps - rate 62 */ 18750000, /* 150Mbps - rate 64 */ 20000000, /* 160Mbps - rate 65 */ 21250000, /* 170Mbps - rate 66 */ 22500000, /* 180Mbps - rate 67 */ 23750000, /* 190Mbps - rate 68 */ 26250000, /* 210Mbps - rate 69 */ 27500000, /* 220Mbps - rate 70 */ 28750000, /* 230Mbps - rate 71 */ 30000000, /* 240Mbps - rate 72 */ 31250000, /* 250Mbps - rate 73 */ 34375000, /* 275Mbps - rate 74 */ 37500000, /* 300Mbps - rate 75 */ 40625000, /* 325Mbps - rate 76 */ 43750000, /* 350Mbps - rate 77 */ 46875000, /* 375Mbps - rate 78 */ 53125000, /* 425Mbps - rate 79 */ 56250000, /* 450Mbps - rate 80 */ 59375000, /* 475Mbps - rate 81 */ 62500000, /* 500Mbps - rate 82 */ 68750000, /* 550Mbps - rate 83 */ 75000000, /* 600Mbps - rate 84 */ 81250000, /* 650Mbps - rate 85 */ 87500000, /* 700Mbps - rate 86 */ 93750000, /* 750Mbps - rate 87 */ 106250000, /* 850Mbps - rate 88 */ 112500000, /* 900Mbps - rate 89 */ 125000000, /* 1Gbps - rate 90 */ 156250000, /* 1.25Gps - rate 91 */ 187500000, /* 1.5Gps - rate 92 */ 218750000, /* 1.75Gps - rate 93 */ 250000000, /* 2Gbps - rate 94 */ 281250000, /* 2.25Gps - rate 95 */ 312500000, /* 2.5Gbps - rate 96 */ 343750000, /* 2.75Gbps - rate 97 */ 375000000, /* 3Gbps - rate 98 */ 500000000, /* 4Gbps - rate 99 */ 625000000, /* 5Gbps - rate 100 */ 750000000, /* 6Gbps - rate 101 */ 875000000, /* 7Gbps - rate 102 */ 1000000000, /* 8Gbps - rate 103 */ 1125000000, /* 9Gbps - rate 104 */ 1250000000, /* 10Gbps - rate 105 */ 1875000000, /* 15Gbps - rate 106 */ 2500000000 /* 20Gbps - rate 107 */ }; #define MAX_HDWR_RATES (sizeof(desired_rates)/sizeof(uint64_t)) #define RS_ORDERED_COUNT 16 /* * Number that are in order * at the beginning of the table, * over this a sort is required. */ #define RS_NEXT_ORDER_GROUP 16 /* * The point in our table where * we come fill in a second ordered * group (index wise means -1). */ #define ALL_HARDWARE_RATES 1004 /* * 1Meg - 1Gig in 1 Meg steps * plus 100, 200k and 500k and * 10Gig */ #define RS_ONE_MEGABIT_PERSEC 1000000 #define RS_ONE_GIGABIT_PERSEC 1000000000 #define RS_TEN_GIGABIT_PERSEC 10000000000 static struct head_tcp_rate_set int_rs; static struct mtx rs_mtx; uint32_t rs_number_alive; uint32_t rs_number_dead; SYSCTL_NODE(_net_inet_tcp, OID_AUTO, rl, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "TCP Ratelimit stats"); SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, alive, CTLFLAG_RW, &rs_number_alive, 0, "Number of interfaces initialized for ratelimiting"); SYSCTL_UINT(_net_inet_tcp_rl, OID_AUTO, dead, CTLFLAG_RW, &rs_number_dead, 0, "Number of interfaces departing from ratelimiting"); static void rl_add_syctl_entries(struct sysctl_oid *rl_sysctl_root, struct tcp_rate_set *rs) { /* * Add sysctl entries for thus interface. */ if (rs->rs_flags & RS_INTF_NO_SUP) { SYSCTL_ADD_S32(&rs->sysctl_ctx, SYSCTL_CHILDREN(rl_sysctl_root), OID_AUTO, "disable", CTLFLAG_RD, &rs->rs_disable, 0, "Disable this interface from new hdwr limiting?"); } else { SYSCTL_ADD_S32(&rs->sysctl_ctx, SYSCTL_CHILDREN(rl_sysctl_root), OID_AUTO, "disable", CTLFLAG_RW, &rs->rs_disable, 0, "Disable this interface from new hdwr limiting?"); } SYSCTL_ADD_S32(&rs->sysctl_ctx, SYSCTL_CHILDREN(rl_sysctl_root), OID_AUTO, "minseg", CTLFLAG_RW, &rs->rs_min_seg, 0, "What is the minimum we need to send on this interface?"); SYSCTL_ADD_U64(&rs->sysctl_ctx, SYSCTL_CHILDREN(rl_sysctl_root), OID_AUTO, "flow_limit", CTLFLAG_RW, &rs->rs_flow_limit, 0, "What is the limit for number of flows (0=unlimited)?"); SYSCTL_ADD_S32(&rs->sysctl_ctx, SYSCTL_CHILDREN(rl_sysctl_root), OID_AUTO, "highest", CTLFLAG_RD, &rs->rs_highest_valid, 0, "Highest valid rate"); SYSCTL_ADD_S32(&rs->sysctl_ctx, SYSCTL_CHILDREN(rl_sysctl_root), OID_AUTO, "lowest", CTLFLAG_RD, &rs->rs_lowest_valid, 0, "Lowest valid rate"); SYSCTL_ADD_S32(&rs->sysctl_ctx, SYSCTL_CHILDREN(rl_sysctl_root), OID_AUTO, "flags", CTLFLAG_RD, &rs->rs_flags, 0, "What lags are on the entry?"); SYSCTL_ADD_S32(&rs->sysctl_ctx, SYSCTL_CHILDREN(rl_sysctl_root), OID_AUTO, "numrates", CTLFLAG_RD, &rs->rs_rate_cnt, 0, "How many rates re there?"); SYSCTL_ADD_U64(&rs->sysctl_ctx, SYSCTL_CHILDREN(rl_sysctl_root), OID_AUTO, "flows_using", CTLFLAG_RD, &rs->rs_flows_using, 0, "How many flows are using this interface now?"); #ifdef DETAILED_RATELIMIT_SYSCTL if (rs->rs_rlt && rs->rs_rate_cnt > 0) { /* Lets display the rates */ int i; struct sysctl_oid *rl_rates; struct sysctl_oid *rl_rate_num; char rate_num[16]; rl_rates = SYSCTL_ADD_NODE(&rs->sysctl_ctx, SYSCTL_CHILDREN(rl_sysctl_root), OID_AUTO, "rate", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Ratelist"); for( i = 0; i < rs->rs_rate_cnt; i++) { sprintf(rate_num, "%d", i); rl_rate_num = SYSCTL_ADD_NODE(&rs->sysctl_ctx, SYSCTL_CHILDREN(rl_rates), OID_AUTO, rate_num, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Individual Rate"); SYSCTL_ADD_U32(&rs->sysctl_ctx, SYSCTL_CHILDREN(rl_rate_num), OID_AUTO, "flags", CTLFLAG_RD, &rs->rs_rlt[i].flags, 0, "Flags on this rate"); SYSCTL_ADD_U32(&rs->sysctl_ctx, SYSCTL_CHILDREN(rl_rate_num), OID_AUTO, "pacetime", CTLFLAG_RD, &rs->rs_rlt[i].time_between, 0, "Time hardware inserts between 1500 byte sends"); SYSCTL_ADD_U64(&rs->sysctl_ctx, SYSCTL_CHILDREN(rl_rate_num), OID_AUTO, "rate", CTLFLAG_RD, &rs->rs_rlt[i].rate, 0, "Rate in bytes per second"); } } #endif } static void rs_destroy(epoch_context_t ctx) { struct tcp_rate_set *rs; bool do_free_rs; rs = __containerof(ctx, struct tcp_rate_set, rs_epoch_ctx); mtx_lock(&rs_mtx); rs->rs_flags &= ~RS_FUNERAL_SCHD; /* * In theory its possible (but unlikely) * that while the delete was occuring * and we were applying the DEAD flag * someone slipped in and found the * interface in a lookup. While we * decided rs_flows_using were 0 and * scheduling the epoch_call, the other * thread incremented rs_flow_using. This * is because users have a pointer and * we only use the rs_flows_using in an * atomic fashion, i.e. the other entities * are not protected. To assure this did * not occur, we check rs_flows_using here * before deleting. */ do_free_rs = (rs->rs_flows_using == 0); rs_number_dead--; mtx_unlock(&rs_mtx); if (do_free_rs) { sysctl_ctx_free(&rs->sysctl_ctx); free(rs->rs_rlt, M_TCPPACE); free(rs, M_TCPPACE); } } static void rs_defer_destroy(struct tcp_rate_set *rs) { mtx_assert(&rs_mtx, MA_OWNED); /* Check if already pending. */ if (rs->rs_flags & RS_FUNERAL_SCHD) return; rs_number_dead++; /* Set flag to only defer once. */ rs->rs_flags |= RS_FUNERAL_SCHD; NET_EPOCH_CALL(rs_destroy, &rs->rs_epoch_ctx); } #ifdef INET extern counter_u64_t rate_limit_set_ok; extern counter_u64_t rate_limit_active; extern counter_u64_t rate_limit_alloc_fail; #endif static int rl_attach_txrtlmt(struct ifnet *ifp, uint32_t flowtype, int flowid, uint64_t cfg_rate, struct m_snd_tag **tag) { int error; union if_snd_tag_alloc_params params = { .rate_limit.hdr.type = IF_SND_TAG_TYPE_RATE_LIMIT, .rate_limit.hdr.flowid = flowid, .rate_limit.hdr.flowtype = flowtype, .rate_limit.max_rate = cfg_rate, .rate_limit.flags = M_NOWAIT, }; if (ifp->if_snd_tag_alloc == NULL) { error = EOPNOTSUPP; } else { error = ifp->if_snd_tag_alloc(ifp, ¶ms, tag); #ifdef INET if (error == 0) { if_ref((*tag)->ifp); counter_u64_add(rate_limit_set_ok, 1); counter_u64_add(rate_limit_active, 1); } else counter_u64_add(rate_limit_alloc_fail, 1); #endif } return (error); } static void populate_canned_table(struct tcp_rate_set *rs, const uint64_t *rate_table_act) { /* * The internal table is "special", it * is two seperate ordered tables that * must be merged. We get here when the * adapter specifies a number of rates that * covers both ranges in the table in some * form. */ int i, at_low, at_high; uint8_t low_disabled = 0, high_disabled = 0; for(i = 0, at_low = 0, at_high = RS_NEXT_ORDER_GROUP; i < rs->rs_rate_cnt; i++) { rs->rs_rlt[i].flags = 0; rs->rs_rlt[i].time_between = 0; if ((low_disabled == 0) && (high_disabled || (rate_table_act[at_low] < rate_table_act[at_high]))) { rs->rs_rlt[i].rate = rate_table_act[at_low]; at_low++; if (at_low == RS_NEXT_ORDER_GROUP) low_disabled = 1; } else if (high_disabled == 0) { rs->rs_rlt[i].rate = rate_table_act[at_high]; at_high++; if (at_high == MAX_HDWR_RATES) high_disabled = 1; } } } static struct tcp_rate_set * rt_setup_new_rs(struct ifnet *ifp, int *error) { struct tcp_rate_set *rs; const uint64_t *rate_table_act; uint64_t lentim, res; size_t sz; uint32_t hash_type; int i; struct if_ratelimit_query_results rl; struct sysctl_oid *rl_sysctl_root; /* * We expect to enter with the * mutex locked. */ if (ifp->if_ratelimit_query == NULL) { /* * We can do nothing if we cannot * get a query back from the driver. */ printf("Warning:No query functions for %s:%d-- failed\n", ifp->if_dname, ifp->if_dunit); return (NULL); } rs = malloc(sizeof(struct tcp_rate_set), M_TCPPACE, M_NOWAIT | M_ZERO); if (rs == NULL) { if (error) *error = ENOMEM; printf("Warning:No memory for malloc of tcp_rate_set\n"); return (NULL); } memset(&rl, 0, sizeof(rl)); rl.flags = RT_NOSUPPORT; ifp->if_ratelimit_query(ifp, &rl); if (rl.flags & RT_IS_UNUSABLE) { /* * The interface does not really support * the rate-limiting. */ memset(rs, 0, sizeof(struct tcp_rate_set)); rs->rs_ifp = ifp; rs->rs_if_dunit = ifp->if_dunit; rs->rs_flags = RS_INTF_NO_SUP; rs->rs_disable = 1; rs_number_alive++; sysctl_ctx_init(&rs->sysctl_ctx); rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx, SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl), OID_AUTO, rs->rs_ifp->if_xname, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, ""); rl_add_syctl_entries(rl_sysctl_root, rs); mtx_lock(&rs_mtx); CK_LIST_INSERT_HEAD(&int_rs, rs, next); mtx_unlock(&rs_mtx); return (rs); } else if ((rl.flags & RT_IS_INDIRECT) == RT_IS_INDIRECT) { memset(rs, 0, sizeof(struct tcp_rate_set)); rs->rs_ifp = ifp; rs->rs_if_dunit = ifp->if_dunit; rs->rs_flags = RS_IS_DEFF; rs_number_alive++; sysctl_ctx_init(&rs->sysctl_ctx); rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx, SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl), OID_AUTO, rs->rs_ifp->if_xname, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, ""); rl_add_syctl_entries(rl_sysctl_root, rs); mtx_lock(&rs_mtx); CK_LIST_INSERT_HEAD(&int_rs, rs, next); mtx_unlock(&rs_mtx); return (rs); } else if ((rl.flags & RT_IS_FIXED_TABLE) == RT_IS_FIXED_TABLE) { /* Mellanox C4 likely */ rs->rs_ifp = ifp; rs->rs_if_dunit = ifp->if_dunit; rs->rs_rate_cnt = rl.number_of_rates; rs->rs_min_seg = rl.min_segment_burst; rs->rs_highest_valid = 0; rs->rs_flow_limit = rl.max_flows; rs->rs_flags = RS_IS_INTF | RS_NO_PRE; rs->rs_disable = 0; rate_table_act = rl.rate_table; } else if ((rl.flags & RT_IS_SELECTABLE) == RT_IS_SELECTABLE) { /* Chelsio, C5 and C6 of Mellanox? */ rs->rs_ifp = ifp; rs->rs_if_dunit = ifp->if_dunit; rs->rs_rate_cnt = rl.number_of_rates; rs->rs_min_seg = rl.min_segment_burst; rs->rs_disable = 0; rs->rs_flow_limit = rl.max_flows; rate_table_act = desired_rates; if ((rs->rs_rate_cnt > MAX_HDWR_RATES) && (rs->rs_rate_cnt < ALL_HARDWARE_RATES)) { /* * Our desired table is not big * enough, do what we can. */ rs->rs_rate_cnt = MAX_HDWR_RATES; } if (rs->rs_rate_cnt <= RS_ORDERED_COUNT) rs->rs_flags = RS_IS_INTF; else rs->rs_flags = RS_IS_INTF | RS_INT_TBL; if (rs->rs_rate_cnt >= ALL_HARDWARE_RATES) rs->rs_rate_cnt = ALL_HARDWARE_RATES; } else { free(rs, M_TCPPACE); return (NULL); } sz = sizeof(struct tcp_hwrate_limit_table) * rs->rs_rate_cnt; rs->rs_rlt = malloc(sz, M_TCPPACE, M_NOWAIT); if (rs->rs_rlt == NULL) { if (error) *error = ENOMEM; bail: free(rs, M_TCPPACE); return (NULL); } if (rs->rs_rate_cnt >= ALL_HARDWARE_RATES) { /* * The interface supports all * the rates we could possibly want. */ uint64_t rat; rs->rs_rlt[0].rate = 12500; /* 100k */ rs->rs_rlt[1].rate = 25000; /* 200k */ rs->rs_rlt[2].rate = 62500; /* 500k */ /* Note 125000 == 1Megabit * populate 1Meg - 1000meg. */ for(i = 3, rat = 125000; i< (ALL_HARDWARE_RATES-1); i++) { rs->rs_rlt[i].rate = rat; rat += 125000; } rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate = 1250000000; } else if (rs->rs_flags & RS_INT_TBL) { /* We populate this in a special way */ populate_canned_table(rs, rate_table_act); } else { /* * Just copy in the rates from * the table, it is in order. */ for (i=0; irs_rate_cnt; i++) { rs->rs_rlt[i].rate = rate_table_act[i]; rs->rs_rlt[i].time_between = 0; rs->rs_rlt[i].flags = 0; } } for (i = (rs->rs_rate_cnt - 1); i >= 0; i--) { /* * We go backwards through the list so that if we can't get * a rate and fail to init one, we have at least a chance of * getting the highest one. */ rs->rs_rlt[i].ptbl = rs; rs->rs_rlt[i].tag = NULL; /* * Calculate the time between. */ lentim = ETHERNET_SEGMENT_SIZE * USECS_IN_SECOND; res = lentim / rs->rs_rlt[i].rate; if (res > 0) rs->rs_rlt[i].time_between = res; else rs->rs_rlt[i].time_between = 1; if (rs->rs_flags & RS_NO_PRE) { rs->rs_rlt[i].flags = HDWRPACE_INITED; rs->rs_lowest_valid = i; } else { int err; if ((rl.flags & RT_IS_SETUP_REQ) && (ifp->if_ratelimit_query)) { err = ifp->if_ratelimit_setup(ifp, rs->rs_rlt[i].rate, i); if (err) goto handle_err; } #ifdef RSS hash_type = M_HASHTYPE_RSS_TCP_IPV4; #else hash_type = M_HASHTYPE_OPAQUE_HASH; #endif err = rl_attach_txrtlmt(ifp, hash_type, (i + 1), rs->rs_rlt[i].rate, &rs->rs_rlt[i].tag); if (err) { handle_err: if (i == (rs->rs_rate_cnt - 1)) { /* * Huh - first rate and we can't get * it? */ free(rs->rs_rlt, M_TCPPACE); if (error) *error = err; goto bail; } else { if (error) *error = err; } break; } else { rs->rs_rlt[i].flags = HDWRPACE_INITED | HDWRPACE_TAGPRESENT; rs->rs_lowest_valid = i; } } } /* Did we get at least 1 rate? */ if (rs->rs_rlt[(rs->rs_rate_cnt - 1)].flags & HDWRPACE_INITED) rs->rs_highest_valid = rs->rs_rate_cnt - 1; else { free(rs->rs_rlt, M_TCPPACE); goto bail; } rs_number_alive++; sysctl_ctx_init(&rs->sysctl_ctx); rl_sysctl_root = SYSCTL_ADD_NODE(&rs->sysctl_ctx, SYSCTL_STATIC_CHILDREN(_net_inet_tcp_rl), OID_AUTO, rs->rs_ifp->if_xname, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, ""); rl_add_syctl_entries(rl_sysctl_root, rs); mtx_lock(&rs_mtx); CK_LIST_INSERT_HEAD(&int_rs, rs, next); mtx_unlock(&rs_mtx); return (rs); } static const struct tcp_hwrate_limit_table * tcp_int_find_suitable_rate(const struct tcp_rate_set *rs, uint64_t bytes_per_sec, uint32_t flags) { struct tcp_hwrate_limit_table *arte = NULL, *rte = NULL; uint64_t mbits_per_sec, ind_calc; int i; mbits_per_sec = (bytes_per_sec * 8); if (flags & RS_PACING_LT) { if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) && (rs->rs_lowest_valid <= 2)){ /* * Smaller than 1Meg, only * 3 entries can match it. */ for(i = rs->rs_lowest_valid; i < 3; i++) { if (bytes_per_sec <= rs->rs_rlt[i].rate) { rte = &rs->rs_rlt[i]; break; } else if (rs->rs_rlt[i].flags & HDWRPACE_INITED) { arte = &rs->rs_rlt[i]; } } goto done; } else if ((mbits_per_sec > RS_ONE_GIGABIT_PERSEC) && (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)){ /* * Larger than 1G (the majority of * our table. */ if (mbits_per_sec < RS_TEN_GIGABIT_PERSEC) rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)]; else arte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)]; goto done; } /* * If we reach here its in our table (between 1Meg - 1000Meg), * just take the rounded down mbits per second, and add * 1Megabit to it, from this we can calculate * the index in the table. */ ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC; if ((ind_calc * RS_ONE_MEGABIT_PERSEC) != mbits_per_sec) ind_calc++; /* our table is offset by 3, we add 2 */ ind_calc += 2; if (ind_calc > (ALL_HARDWARE_RATES-1)) { /* This should not happen */ ind_calc = ALL_HARDWARE_RATES-1; } if ((ind_calc >= rs->rs_lowest_valid) && (ind_calc <= rs->rs_highest_valid)) rte = &rs->rs_rlt[ind_calc]; } else if (flags & RS_PACING_EXACT_MATCH) { if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) && (rs->rs_lowest_valid <= 2)){ for(i = rs->rs_lowest_valid; i < 3; i++) { if (bytes_per_sec == rs->rs_rlt[i].rate) { rte = &rs->rs_rlt[i]; break; } } } else if ((mbits_per_sec > RS_ONE_GIGABIT_PERSEC) && (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)) { /* > 1Gbps only one rate */ if (bytes_per_sec == rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) { /* Its 10G wow */ rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)]; } } else { /* Ok it must be a exact meg (its between 1G and 1Meg) */ ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC; if ((ind_calc * RS_ONE_MEGABIT_PERSEC) == mbits_per_sec) { /* its an exact Mbps */ ind_calc += 2; if (ind_calc > (ALL_HARDWARE_RATES-1)) { /* This should not happen */ ind_calc = ALL_HARDWARE_RATES-1; } if (rs->rs_rlt[ind_calc].flags & HDWRPACE_INITED) rte = &rs->rs_rlt[ind_calc]; } } } else { /* we want greater than the requested rate */ if ((mbits_per_sec < RS_ONE_MEGABIT_PERSEC) && (rs->rs_lowest_valid <= 2)){ arte = &rs->rs_rlt[3]; /* set alternate to 1Meg */ for (i=2; i>=rs->rs_lowest_valid; i--) { if (bytes_per_sec < rs->rs_rlt[i].rate) { rte = &rs->rs_rlt[i]; break; } else if ((flags & RS_PACING_GEQ) && (bytes_per_sec == rs->rs_rlt[i].rate)) { rte = &rs->rs_rlt[i]; break; } else { arte = &rs->rs_rlt[i]; /* new alternate */ } } } else if (mbits_per_sec > RS_ONE_GIGABIT_PERSEC) { if ((bytes_per_sec < rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) && (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)){ /* Our top rate is larger than the request */ rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)]; } else if ((flags & RS_PACING_GEQ) && (bytes_per_sec == rs->rs_rlt[(ALL_HARDWARE_RATES-1)].rate) && (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED)) { /* It matches our top rate */ rte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)]; } else if (rs->rs_rlt[(ALL_HARDWARE_RATES-1)].flags & HDWRPACE_INITED) { /* The top rate is an alternative */ arte = &rs->rs_rlt[(ALL_HARDWARE_RATES-1)]; } } else { /* Its in our range 1Meg - 1Gig */ if (flags & RS_PACING_GEQ) { ind_calc = mbits_per_sec/RS_ONE_MEGABIT_PERSEC; if ((ind_calc * RS_ONE_MEGABIT_PERSEC) == mbits_per_sec) { if (ind_calc > (ALL_HARDWARE_RATES-1)) { /* This should not happen */ ind_calc = (ALL_HARDWARE_RATES-1); } rte = &rs->rs_rlt[ind_calc]; } goto done; } ind_calc = (mbits_per_sec + (RS_ONE_MEGABIT_PERSEC-1))/RS_ONE_MEGABIT_PERSEC; ind_calc += 2; if (ind_calc > (ALL_HARDWARE_RATES-1)) { /* This should not happen */ ind_calc = ALL_HARDWARE_RATES-1; } if (rs->rs_rlt[ind_calc].flags & HDWRPACE_INITED) rte = &rs->rs_rlt[ind_calc]; } } done: if ((rte == NULL) && (arte != NULL) && (flags & RS_PACING_SUB_OK)) { /* We can use the substitute */ rte = arte; } return (rte); } static const struct tcp_hwrate_limit_table * tcp_find_suitable_rate(const struct tcp_rate_set *rs, uint64_t bytes_per_sec, uint32_t flags) { /** * Hunt the rate table with the restrictions in flags and find a * suitable rate if possible. * RS_PACING_EXACT_MATCH - look for an exact match to rate. * RS_PACING_GT - must be greater than. * RS_PACING_GEQ - must be greater than or equal. * RS_PACING_LT - must be less than. * RS_PACING_SUB_OK - If we don't meet criteria a * substitute is ok. */ int i, matched; struct tcp_hwrate_limit_table *rte = NULL; if ((rs->rs_flags & RS_INT_TBL) && (rs->rs_rate_cnt >= ALL_HARDWARE_RATES)) { /* * Here we don't want to paw thru * a big table, we have everything * from 1Meg - 1000Meg in 1Meg increments. * Use an alternate method to "lookup". */ return (tcp_int_find_suitable_rate(rs, bytes_per_sec, flags)); } if ((flags & RS_PACING_LT) || (flags & RS_PACING_EXACT_MATCH)) { /* * For exact and less than we go forward through the table. * This way when we find one larger we stop (exact was a * toss up). */ for (i = rs->rs_lowest_valid, matched = 0; i <= rs->rs_highest_valid; i++) { if ((flags & RS_PACING_EXACT_MATCH) && (bytes_per_sec == rs->rs_rlt[i].rate)) { rte = &rs->rs_rlt[i]; matched = 1; break; } else if ((flags & RS_PACING_LT) && (bytes_per_sec <= rs->rs_rlt[i].rate)) { rte = &rs->rs_rlt[i]; matched = 1; break; } if (bytes_per_sec > rs->rs_rlt[i].rate) break; } if ((matched == 0) && (flags & RS_PACING_LT) && (flags & RS_PACING_SUB_OK)) { /* Kick in a substitute (the lowest) */ rte = &rs->rs_rlt[rs->rs_lowest_valid]; } } else { /* * Here we go backward through the table so that we can find * the one greater in theory faster (but its probably a * wash). */ for (i = rs->rs_highest_valid, matched = 0; i >= rs->rs_lowest_valid; i--) { if (rs->rs_rlt[i].rate > bytes_per_sec) { /* A possible candidate */ rte = &rs->rs_rlt[i]; } if ((flags & RS_PACING_GEQ) && (bytes_per_sec == rs->rs_rlt[i].rate)) { /* An exact match and we want equal */ matched = 1; rte = &rs->rs_rlt[i]; break; } else if (rte) { /* * Found one that is larger than but don't * stop, there may be a more closer match. */ matched = 1; } if (rs->rs_rlt[i].rate < bytes_per_sec) { /* * We found a table entry that is smaller, * stop there will be none greater or equal. */ break; } } if ((matched == 0) && (flags & RS_PACING_SUB_OK)) { /* Kick in a substitute (the highest) */ rte = &rs->rs_rlt[rs->rs_highest_valid]; } } return (rte); } static struct ifnet * rt_find_real_interface(struct ifnet *ifp, struct inpcb *inp, int *error) { struct ifnet *tifp; struct m_snd_tag *tag; union if_snd_tag_alloc_params params = { .rate_limit.hdr.type = IF_SND_TAG_TYPE_RATE_LIMIT, .rate_limit.hdr.flowid = 1, .rate_limit.hdr.numa_domain = inp->inp_numa_domain, .rate_limit.max_rate = COMMON_RATE, .rate_limit.flags = M_NOWAIT, }; int err; #ifdef RSS params.rate_limit.hdr.flowtype = ((inp->inp_vflag & INP_IPV6) ? M_HASHTYPE_RSS_TCP_IPV6 : M_HASHTYPE_RSS_TCP_IPV4); #else params.rate_limit.hdr.flowtype = M_HASHTYPE_OPAQUE_HASH; #endif tag = NULL; if (ifp->if_snd_tag_alloc) { if (error) *error = ENODEV; return (NULL); } err = ifp->if_snd_tag_alloc(ifp, ¶ms, &tag); if (err) { /* Failed to setup a tag? */ if (error) *error = err; return (NULL); } tifp = tag->ifp; tifp->if_snd_tag_free(tag); return (tifp); } static const struct tcp_hwrate_limit_table * rt_setup_rate(struct inpcb *inp, struct ifnet *ifp, uint64_t bytes_per_sec, uint32_t flags, int *error) { /* First lets find the interface if it exists */ const struct tcp_hwrate_limit_table *rte; struct tcp_rate_set *rs; struct epoch_tracker et; int err; NET_EPOCH_ENTER(et); use_real_interface: CK_LIST_FOREACH(rs, &int_rs, next) { /* * Note we don't look with the lock since we either see a * new entry or will get one when we try to add it. */ if (rs->rs_flags & RS_IS_DEAD) { /* The dead are not looked at */ continue; } if ((rs->rs_ifp == ifp) && (rs->rs_if_dunit == ifp->if_dunit)) { /* Ok we found it */ break; } } if ((rs == NULL) || (rs->rs_flags & RS_INTF_NO_SUP) || (rs->rs_flags & RS_IS_DEAD)) { /* * This means we got a packet *before* * the IF-UP was processed below, * while or after we already received an interface * departed event. In either case we really don't * want to do anything with pacing, in * the departing case the packet is not * going to go very far. The new case * might be arguable, but its impossible * to tell from the departing case. */ if (rs->rs_disable && error) *error = ENODEV; NET_EPOCH_EXIT(et); return (NULL); } if ((rs == NULL) || (rs->rs_disable != 0)) { if (rs->rs_disable && error) *error = ENOSPC; NET_EPOCH_EXIT(et); return (NULL); } if (rs->rs_flags & RS_IS_DEFF) { /* We need to find the real interface */ struct ifnet *tifp; tifp = rt_find_real_interface(ifp, inp, error); if (tifp == NULL) { if (rs->rs_disable && error) *error = ENOTSUP; NET_EPOCH_EXIT(et); return (NULL); } goto use_real_interface; } if (rs->rs_flow_limit && ((rs->rs_flows_using + 1) > rs->rs_flow_limit)) { if (error) *error = ENOSPC; NET_EPOCH_EXIT(et); return (NULL); } rte = tcp_find_suitable_rate(rs, bytes_per_sec, flags); if (rte) { err = in_pcbattach_txrtlmt(inp, rs->rs_ifp, inp->inp_flowtype, inp->inp_flowid, rte->rate, &inp->inp_snd_tag); if (err) { /* Failed to attach */ if (error) *error = err; rte = NULL; } } if (rte) { /* * We use an atomic here for accounting so we don't have to * use locks when freeing. */ atomic_add_64(&rs->rs_flows_using, 1); } NET_EPOCH_EXIT(et); return (rte); } static void tcp_rl_ifnet_link(void *arg __unused, struct ifnet *ifp, int link_state) { int error; struct tcp_rate_set *rs; if (((ifp->if_capabilities & IFCAP_TXRTLMT) == 0) || (link_state != LINK_STATE_UP)) { /* * We only care on an interface going up that is rate-limit * capable. */ return; } mtx_lock(&rs_mtx); CK_LIST_FOREACH(rs, &int_rs, next) { if ((rs->rs_ifp == ifp) && (rs->rs_if_dunit == ifp->if_dunit)) { /* We already have initialized this guy */ mtx_unlock(&rs_mtx); return; } } mtx_unlock(&rs_mtx); rt_setup_new_rs(ifp, &error); } static void tcp_rl_ifnet_departure(void *arg __unused, struct ifnet *ifp) { struct tcp_rate_set *rs, *nrs; struct ifnet *tifp; int i; mtx_lock(&rs_mtx); CK_LIST_FOREACH_SAFE(rs, &int_rs, next, nrs) { if ((rs->rs_ifp == ifp) && (rs->rs_if_dunit == ifp->if_dunit)) { CK_LIST_REMOVE(rs, next); rs_number_alive--; rs->rs_flags |= RS_IS_DEAD; for (i = 0; i < rs->rs_rate_cnt; i++) { if (rs->rs_rlt[i].flags & HDWRPACE_TAGPRESENT) { tifp = rs->rs_rlt[i].tag->ifp; in_pcbdetach_tag(tifp, rs->rs_rlt[i].tag); rs->rs_rlt[i].tag = NULL; } rs->rs_rlt[i].flags = HDWRPACE_IFPDEPARTED; } if (rs->rs_flows_using == 0) rs_defer_destroy(rs); break; } } mtx_unlock(&rs_mtx); } static void tcp_rl_shutdown(void *arg __unused, int howto __unused) { struct tcp_rate_set *rs, *nrs; struct ifnet *tifp; int i; mtx_lock(&rs_mtx); CK_LIST_FOREACH_SAFE(rs, &int_rs, next, nrs) { CK_LIST_REMOVE(rs, next); rs_number_alive--; rs->rs_flags |= RS_IS_DEAD; for (i = 0; i < rs->rs_rate_cnt; i++) { if (rs->rs_rlt[i].flags & HDWRPACE_TAGPRESENT) { tifp = rs->rs_rlt[i].tag->ifp; in_pcbdetach_tag(tifp, rs->rs_rlt[i].tag); rs->rs_rlt[i].tag = NULL; } rs->rs_rlt[i].flags = HDWRPACE_IFPDEPARTED; } if (rs->rs_flows_using == 0) rs_defer_destroy(rs); } mtx_unlock(&rs_mtx); } const struct tcp_hwrate_limit_table * tcp_set_pacing_rate(struct tcpcb *tp, struct ifnet *ifp, uint64_t bytes_per_sec, int flags, int *error) { const struct tcp_hwrate_limit_table *rte; if (tp->t_inpcb->inp_snd_tag == NULL) { /* * We are setting up a rate for the first time. */ if ((ifp->if_capabilities & IFCAP_TXRTLMT) == 0) { /* Not supported by the egress */ if (error) *error = ENODEV; return (NULL); } #ifdef KERN_TLS if (tp->t_inpcb->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) { /* * We currently can't do both TLS and hardware * pacing */ if (error) *error = EINVAL; return (NULL); } #endif rte = rt_setup_rate(tp->t_inpcb, ifp, bytes_per_sec, flags, error); } else { /* * We are modifying a rate, wrong interface? */ if (error) *error = EINVAL; rte = NULL; } *error = 0; return (rte); } const struct tcp_hwrate_limit_table * tcp_chg_pacing_rate(const struct tcp_hwrate_limit_table *crte, struct tcpcb *tp, struct ifnet *ifp, uint64_t bytes_per_sec, int flags, int *error) { const struct tcp_hwrate_limit_table *nrte; const struct tcp_rate_set *rs; int is_indirect = 0; int err; if ((tp->t_inpcb->inp_snd_tag == NULL) || (crte == NULL)) { /* Wrong interface */ if (error) *error = EINVAL; return (NULL); } rs = crte->ptbl; if ((rs->rs_flags & RS_IS_DEAD) || (crte->flags & HDWRPACE_IFPDEPARTED)) { /* Release the rate, and try anew */ re_rate: tcp_rel_pacing_rate(crte, tp); nrte = tcp_set_pacing_rate(tp, ifp, bytes_per_sec, flags, error); return (nrte); } if ((rs->rs_flags & RT_IS_INDIRECT ) == RT_IS_INDIRECT) is_indirect = 1; else is_indirect = 0; if ((is_indirect == 0) && ((ifp != rs->rs_ifp) || (ifp->if_dunit != rs->rs_if_dunit))) { /* * Something changed, the user is not pointing to the same * ifp? Maybe a route updated on this guy? */ goto re_rate; } else if (is_indirect) { /* * For indirect we have to dig in and find the real interface. */ struct ifnet *rifp; rifp = rt_find_real_interface(ifp, tp->t_inpcb, error); if (rifp == NULL) { /* Can't find it? */ goto re_rate; } if ((rifp != rs->rs_ifp) || (ifp->if_dunit != rs->rs_if_dunit)) { goto re_rate; } } nrte = tcp_find_suitable_rate(rs, bytes_per_sec, flags); if (nrte == crte) { /* No change */ if (error) *error = 0; return (crte); } if (nrte == NULL) { /* Release the old rate */ tcp_rel_pacing_rate(crte, tp); return (NULL); } /* Change rates to our new entry */ err = in_pcbmodify_txrtlmt(tp->t_inpcb, nrte->rate); if (err) { if (error) *error = err; return (NULL); } if (error) *error = 0; return (nrte); } void tcp_rel_pacing_rate(const struct tcp_hwrate_limit_table *crte, struct tcpcb *tp) { const struct tcp_rate_set *crs; struct tcp_rate_set *rs; uint64_t pre; crs = crte->ptbl; /* * Now we must break the const * in order to release our refcount. */ rs = __DECONST(struct tcp_rate_set *, crs); pre = atomic_fetchadd_64(&rs->rs_flows_using, -1); if (pre == 1) { mtx_lock(&rs_mtx); /* * Is it dead? */ if (rs->rs_flags & RS_IS_DEAD) rs_defer_destroy(rs); mtx_unlock(&rs_mtx); } in_pcbdetach_txrtlmt(tp->t_inpcb); } #define ONE_POINT_TWO_MEG 150000 /* 1.2 megabits in bytes */ #define ONE_HUNDRED_MBPS 12500000 /* 100Mbps in bytes per second */ #define FIVE_HUNDRED_MBPS 62500000 /* 500Mbps in bytes per second */ #define MAX_MSS_SENT 43 /* 43 mss = 43 x 1500 = 64,500 bytes */ uint32_t tcp_get_pacing_burst_size (uint64_t bw, uint32_t segsiz, int can_use_1mss, const struct tcp_hwrate_limit_table *te, int *err) { /* * We use the google formula to calculate the * TSO size. I.E. * bw < 24Meg * tso = 2mss * else * tso = min(bw/1000, 64k) * * Note for these calculations we ignore the * packet overhead (enet hdr, ip hdr and tcp hdr). */ uint64_t lentim, res, bytes; uint32_t new_tso, min_tso_segs; bytes = bw / 1000; if (bytes > (64 * 1000)) bytes = 64 * 1000; /* Round up */ new_tso = (bytes + segsiz - 1) / segsiz; if (can_use_1mss && (bw < ONE_POINT_TWO_MEG)) min_tso_segs = 1; else min_tso_segs = 2; if (new_tso < min_tso_segs) new_tso = min_tso_segs; if (new_tso > MAX_MSS_SENT) new_tso = MAX_MSS_SENT; new_tso *= segsiz; /* * If we are not doing hardware pacing * then we are done. */ if (te == NULL) { if (err) *err = 0; return(new_tso); } /* * For hardware pacing we look at the * rate you are sending at and compare * that to the rate you have in hardware. * * If the hardware rate is slower than your * software rate then you are in error and * we will build a queue in our hardware whic * is probably not desired, in such a case * just return the non-hardware TSO size. * * If the rate in hardware is faster (which * it should be) then look at how long it * takes to send one ethernet segment size at * your b/w and compare that to the time it * takes to send at the rate you had selected. * * If your time is greater (which we hope it is) * we get the delta between the two, and then * divide that into your pacing time. This tells * us how many MSS you can send down at once (rounded up). * * Note we also double this value if the b/w is over * 100Mbps. If its over 500meg we just set you to the * max (43 segments). */ if (te->rate > FIVE_HUNDRED_MBPS) return (segsiz * MAX_MSS_SENT); if (te->rate == bw) { /* We are pacing at exactly the hdwr rate */ return (segsiz * MAX_MSS_SENT); } lentim = ETHERNET_SEGMENT_SIZE * USECS_IN_SECOND; res = lentim / bw; if (res > te->time_between) { uint32_t delta, segs; delta = res - te->time_between; segs = (res + delta - 1)/delta; if (te->rate > ONE_HUNDRED_MBPS) segs *= 2; if (segs < min_tso_segs) segs = min_tso_segs; if (segs > MAX_MSS_SENT) segs = MAX_MSS_SENT; segs *= segsiz; if (err) *err = 0; if (segs < new_tso) { /* unexpected ? */ return(new_tso); } else { return (segs); } } else { /* * Your time is smaller which means * we will grow a queue on our * hardware. Send back the non-hardware * rate. */ if (err) *err = -1; return (new_tso); } } static eventhandler_tag rl_ifnet_departs; static eventhandler_tag rl_ifnet_arrives; static eventhandler_tag rl_shutdown_start; static void tcp_rs_init(void *st __unused) { CK_LIST_INIT(&int_rs); rs_number_alive = 0; - rs_number_dead = 0;; + rs_number_dead = 0; mtx_init(&rs_mtx, "tcp_rs_mtx", "rsmtx", MTX_DEF); rl_ifnet_departs = EVENTHANDLER_REGISTER(ifnet_departure_event, tcp_rl_ifnet_departure, NULL, EVENTHANDLER_PRI_ANY); rl_ifnet_arrives = EVENTHANDLER_REGISTER(ifnet_link_event, tcp_rl_ifnet_link, NULL, EVENTHANDLER_PRI_ANY); rl_shutdown_start = EVENTHANDLER_REGISTER(shutdown_pre_sync, tcp_rl_shutdown, NULL, SHUTDOWN_PRI_FIRST); printf("TCP_ratelimit: Is now initialized\n"); } SYSINIT(tcp_rl_init, SI_SUB_SMP + 1, SI_ORDER_ANY, tcp_rs_init, NULL); #endif