1 /* 2 * Copyright (c) 2003, 2004 Jeffrey M. Hsu. All rights reserved. 3 * Copyright (c) 2003, 2004 The DragonFly Project. All rights reserved. 4 * 5 * This code is derived from software contributed to The DragonFly Project 6 * by Jeffrey M. Hsu. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. Neither the name of The DragonFly Project nor the names of its 17 * contributors may be used to endorse or promote products derived 18 * from this software without specific, prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 22 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 23 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 24 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 25 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 26 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 27 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 28 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 29 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 30 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 */ 33 34 /* 35 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 36 * The Regents of the University of California. All rights reserved. 37 * 38 * Redistribution and use in source and binary forms, with or without 39 * modification, are permitted provided that the following conditions 40 * are met: 41 * 1. Redistributions of source code must retain the above copyright 42 * notice, this list of conditions and the following disclaimer. 43 * 2. Redistributions in binary form must reproduce the above copyright 44 * notice, this list of conditions and the following disclaimer in the 45 * documentation and/or other materials provided with the distribution. 46 * 3. Neither the name of the University nor the names of its contributors 47 * may be used to endorse or promote products derived from this software 48 * without specific prior written permission. 49 * 50 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 51 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 52 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 53 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 54 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 55 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 56 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 57 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 58 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 59 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 60 * SUCH DAMAGE. 61 * 62 * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95 63 * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $ 64 */ 65 66 #include "opt_compat.h" 67 #include "opt_inet.h" 68 #include "opt_inet6.h" 69 #include "opt_ipsec.h" 70 #include "opt_tcpdebug.h" 71 72 #include <sys/param.h> 73 #include <sys/systm.h> 74 #include <sys/callout.h> 75 #include <sys/kernel.h> 76 #include <sys/sysctl.h> 77 #include <sys/malloc.h> 78 #include <sys/mpipe.h> 79 #include <sys/mbuf.h> 80 #ifdef INET6 81 #include <sys/domain.h> 82 #endif 83 #include <sys/proc.h> 84 #include <sys/priv.h> 85 #include <sys/socket.h> 86 #include <sys/socketops.h> 87 #include <sys/socketvar.h> 88 #include <sys/protosw.h> 89 #include <sys/random.h> 90 #include <sys/in_cksum.h> 91 #include <sys/ktr.h> 92 93 #include <net/route.h> 94 #include <net/if.h> 95 #include <net/netisr2.h> 96 97 #define _IP_VHL 98 #include <netinet/in.h> 99 #include <netinet/in_systm.h> 100 #include <netinet/ip.h> 101 #include <netinet/ip6.h> 102 #include <netinet/in_pcb.h> 103 #include <netinet6/in6_pcb.h> 104 #include <netinet/in_var.h> 105 #include <netinet/ip_var.h> 106 #include <netinet6/ip6_var.h> 107 #include <netinet/ip_icmp.h> 108 #ifdef INET6 109 #include <netinet/icmp6.h> 110 #endif 111 #include <netinet/tcp.h> 112 #include <netinet/tcp_fsm.h> 113 #include <netinet/tcp_seq.h> 114 #include <netinet/tcp_timer.h> 115 #include <netinet/tcp_timer2.h> 116 #include <netinet/tcp_var.h> 117 #include <netinet6/tcp6_var.h> 118 #include <netinet/tcpip.h> 119 #ifdef TCPDEBUG 120 #include <netinet/tcp_debug.h> 121 #endif 122 #include <netinet6/ip6protosw.h> 123 124 #ifdef IPSEC 125 #include <netinet6/ipsec.h> 126 #include <netproto/key/key.h> 127 #ifdef INET6 128 #include <netinet6/ipsec6.h> 129 #endif 130 #endif 131 132 #ifdef FAST_IPSEC 133 #include <netproto/ipsec/ipsec.h> 134 #ifdef INET6 135 #include <netproto/ipsec/ipsec6.h> 136 #endif 137 #define IPSEC 138 #endif 139 140 #include <sys/md5.h> 141 #include <machine/smp.h> 142 143 #include <sys/msgport2.h> 144 #include <sys/mplock2.h> 145 #include <net/netmsg2.h> 146 147 #if !defined(KTR_TCP) 148 #define KTR_TCP KTR_ALL 149 #endif 150 /* 151 KTR_INFO_MASTER(tcp); 152 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0); 153 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0); 154 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0); 155 #define logtcp(name) KTR_LOG(tcp_ ## name) 156 */ 157 158 #define TCP_IW_MAXSEGS_DFLT 4 159 #define TCP_IW_CAPSEGS_DFLT 3 160 161 struct inpcbinfo tcbinfo[MAXCPU]; 162 struct tcpcbackqhead tcpcbackq[MAXCPU]; 163 164 int tcp_mssdflt = TCP_MSS; 165 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 166 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size"); 167 168 #ifdef INET6 169 int tcp_v6mssdflt = TCP6_MSS; 170 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW, 171 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6"); 172 #endif 173 174 /* 175 * Minimum MSS we accept and use. This prevents DoS attacks where 176 * we are forced to a ridiculous low MSS like 20 and send hundreds 177 * of packets instead of one. The effect scales with the available 178 * bandwidth and quickly saturates the CPU and network interface 179 * with packet generation and sending. Set to zero to disable MINMSS 180 * checking. This setting prevents us from sending too small packets. 181 */ 182 int tcp_minmss = TCP_MINMSS; 183 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW, 184 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size"); 185 186 #if 0 187 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ; 188 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW, 189 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time"); 190 #endif 191 192 int tcp_do_rfc1323 = 1; 193 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW, 194 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions"); 195 196 static int tcp_tcbhashsize = 0; 197 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD, 198 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable"); 199 200 static int do_tcpdrain = 1; 201 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0, 202 "Enable tcp_drain routine for extra help when low on mbufs"); 203 204 static int icmp_may_rst = 1; 205 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0, 206 "Certain ICMP unreachable messages may abort connections in SYN_SENT"); 207 208 static int tcp_isn_reseed_interval = 0; 209 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW, 210 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret"); 211 212 /* 213 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on 214 * by default, but with generous values which should allow maximal 215 * bandwidth. In particular, the slop defaults to 50 (5 packets). 216 * 217 * The reason for doing this is that the limiter is the only mechanism we 218 * have which seems to do a really good job preventing receiver RX rings 219 * on network interfaces from getting blown out. Even though GigE/10GigE 220 * is supposed to flow control it looks like either it doesn't actually 221 * do it or Open Source drivers do not properly enable it. 222 * 223 * People using the limiter to reduce bottlenecks on slower WAN connections 224 * should set the slop to 20 (2 packets). 225 */ 226 static int tcp_inflight_enable = 1; 227 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW, 228 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting"); 229 230 static int tcp_inflight_debug = 0; 231 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW, 232 &tcp_inflight_debug, 0, "Debug TCP inflight calculations"); 233 234 static int tcp_inflight_min = 6144; 235 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW, 236 &tcp_inflight_min, 0, "Lower bound for TCP inflight window"); 237 238 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT; 239 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW, 240 &tcp_inflight_max, 0, "Upper bound for TCP inflight window"); 241 242 static int tcp_inflight_stab = 50; 243 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW, 244 &tcp_inflight_stab, 0, "Fudge bw 1/10% (50=5%)"); 245 246 static int tcp_inflight_adjrtt = 2; 247 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_adjrtt, CTLFLAG_RW, 248 &tcp_inflight_adjrtt, 0, "Slop for rtt 1/(hz*32)"); 249 250 static int tcp_do_rfc3390 = 1; 251 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_RW, 252 &tcp_do_rfc3390, 0, 253 "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)"); 254 255 static u_long tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 256 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwmaxsegs, CTLFLAG_RW, 257 &tcp_iw_maxsegs, 0, "TCP IW segments max"); 258 259 static u_long tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 260 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwcapsegs, CTLFLAG_RW, 261 &tcp_iw_capsegs, 0, "TCP IW segments"); 262 263 int tcp_low_rtobase = 1; 264 SYSCTL_INT(_net_inet_tcp, OID_AUTO, low_rtobase, CTLFLAG_RW, 265 &tcp_low_rtobase, 0, "Lowering the Initial RTO (RFC 6298)"); 266 267 static int tcp_do_ncr = 1; 268 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr, CTLFLAG_RW, 269 &tcp_do_ncr, 0, "Non-Congestion Robustness (RFC 4653)"); 270 271 int tcp_ncr_rxtthresh_max = 16; 272 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr_rxtthresh_max, CTLFLAG_RW, 273 &tcp_ncr_rxtthresh_max, 0, 274 "Non-Congestion Robustness (RFC 4653), DupThresh upper limit"); 275 276 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives"); 277 static struct malloc_pipe tcptemp_mpipe; 278 279 static void tcp_willblock(void); 280 static void tcp_notify (struct inpcb *, int); 281 282 struct tcp_stats tcpstats_percpu[MAXCPU] __cachealign; 283 284 static struct netmsg_base tcp_drain_netmsg[MAXCPU]; 285 static void tcp_drain_dispatch(netmsg_t nmsg); 286 287 static int 288 sysctl_tcpstats(SYSCTL_HANDLER_ARGS) 289 { 290 int cpu, error = 0; 291 292 for (cpu = 0; cpu < ncpus2; ++cpu) { 293 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu], 294 sizeof(struct tcp_stats)))) 295 break; 296 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu], 297 sizeof(struct tcp_stats)))) 298 break; 299 } 300 301 return (error); 302 } 303 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW), 304 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics"); 305 306 /* 307 * Target size of TCP PCB hash tables. Must be a power of two. 308 * 309 * Note that this can be overridden by the kernel environment 310 * variable net.inet.tcp.tcbhashsize 311 */ 312 #ifndef TCBHASHSIZE 313 #define TCBHASHSIZE 512 314 #endif 315 316 /* 317 * This is the actual shape of what we allocate using the zone 318 * allocator. Doing it this way allows us to protect both structures 319 * using the same generation count, and also eliminates the overhead 320 * of allocating tcpcbs separately. By hiding the structure here, 321 * we avoid changing most of the rest of the code (although it needs 322 * to be changed, eventually, for greater efficiency). 323 */ 324 #define ALIGNMENT 32 325 #define ALIGNM1 (ALIGNMENT - 1) 326 struct inp_tp { 327 union { 328 struct inpcb inp; 329 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1]; 330 } inp_tp_u; 331 struct tcpcb tcb; 332 struct tcp_callout inp_tp_rexmt; 333 struct tcp_callout inp_tp_persist; 334 struct tcp_callout inp_tp_keep; 335 struct tcp_callout inp_tp_2msl; 336 struct tcp_callout inp_tp_delack; 337 struct netmsg_tcp_timer inp_tp_timermsg; 338 struct netmsg_base inp_tp_sndmore; 339 }; 340 #undef ALIGNMENT 341 #undef ALIGNM1 342 343 /* 344 * Tcp initialization 345 */ 346 void 347 tcp_init(void) 348 { 349 struct inpcbportinfo *portinfo; 350 struct inpcbinfo *ticb; 351 int hashsize = TCBHASHSIZE; 352 int cpu; 353 354 /* 355 * note: tcptemp is used for keepalives, and it is ok for an 356 * allocation to fail so do not specify MPF_INT. 357 */ 358 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp), 359 25, -1, 0, NULL, NULL, NULL); 360 361 tcp_delacktime = TCPTV_DELACK; 362 tcp_keepinit = TCPTV_KEEP_INIT; 363 tcp_keepidle = TCPTV_KEEP_IDLE; 364 tcp_keepintvl = TCPTV_KEEPINTVL; 365 tcp_maxpersistidle = TCPTV_KEEP_IDLE; 366 tcp_msl = TCPTV_MSL; 367 tcp_rexmit_min = TCPTV_MIN; 368 tcp_rexmit_slop = TCPTV_CPU_VAR; 369 370 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize); 371 if (!powerof2(hashsize)) { 372 kprintf("WARNING: TCB hash size not a power of 2\n"); 373 hashsize = 512; /* safe default */ 374 } 375 tcp_tcbhashsize = hashsize; 376 377 portinfo = kmalloc_cachealign(sizeof(*portinfo) * ncpus2, M_PCB, 378 M_WAITOK); 379 380 for (cpu = 0; cpu < ncpus2; cpu++) { 381 ticb = &tcbinfo[cpu]; 382 in_pcbinfo_init(ticb, cpu, FALSE); 383 ticb->hashbase = hashinit(hashsize, M_PCB, 384 &ticb->hashmask); 385 in_pcbportinfo_init(&portinfo[cpu], hashsize, TRUE, cpu); 386 ticb->portinfo = portinfo; 387 ticb->portinfo_mask = ncpus2_mask; 388 ticb->wildcardhashbase = hashinit(hashsize, M_PCB, 389 &ticb->wildcardhashmask); 390 ticb->localgrphashbase = hashinit(hashsize, M_PCB, 391 &ticb->localgrphashmask); 392 ticb->ipi_size = sizeof(struct inp_tp); 393 TAILQ_INIT(&tcpcbackq[cpu]); 394 } 395 396 tcp_reass_maxseg = nmbclusters / 16; 397 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg); 398 399 #ifdef INET6 400 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) 401 #else 402 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) 403 #endif 404 if (max_protohdr < TCP_MINPROTOHDR) 405 max_protohdr = TCP_MINPROTOHDR; 406 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) 407 panic("tcp_init"); 408 #undef TCP_MINPROTOHDR 409 410 /* 411 * Initialize TCP statistics counters for each CPU. 412 */ 413 for (cpu = 0; cpu < ncpus2; ++cpu) 414 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats)); 415 416 /* 417 * Initialize netmsgs for TCP drain 418 */ 419 for (cpu = 0; cpu < ncpus2; ++cpu) { 420 netmsg_init(&tcp_drain_netmsg[cpu], NULL, &netisr_adone_rport, 421 MSGF_PRIORITY, tcp_drain_dispatch); 422 } 423 424 syncache_init(); 425 netisr_register_rollup(tcp_willblock, NETISR_ROLLUP_PRIO_TCP); 426 } 427 428 static void 429 tcp_willblock(void) 430 { 431 struct tcpcb *tp; 432 int cpu = mycpu->gd_cpuid; 433 434 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) { 435 KKASSERT(tp->t_flags & TF_ONOUTPUTQ); 436 tp->t_flags &= ~TF_ONOUTPUTQ; 437 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq); 438 tcp_output(tp); 439 } 440 } 441 442 /* 443 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. 444 * tcp_template used to store this data in mbufs, but we now recopy it out 445 * of the tcpcb each time to conserve mbufs. 446 */ 447 void 448 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr, boolean_t tso) 449 { 450 struct inpcb *inp = tp->t_inpcb; 451 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr; 452 453 #ifdef INET6 454 if (inp->inp_vflag & INP_IPV6) { 455 struct ip6_hdr *ip6; 456 457 ip6 = (struct ip6_hdr *)ip_ptr; 458 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | 459 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK); 460 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | 461 (IPV6_VERSION & IPV6_VERSION_MASK); 462 ip6->ip6_nxt = IPPROTO_TCP; 463 ip6->ip6_plen = sizeof(struct tcphdr); 464 ip6->ip6_src = inp->in6p_laddr; 465 ip6->ip6_dst = inp->in6p_faddr; 466 tcp_hdr->th_sum = 0; 467 } else 468 #endif 469 { 470 struct ip *ip = (struct ip *) ip_ptr; 471 u_int plen; 472 473 ip->ip_vhl = IP_VHL_BORING; 474 ip->ip_tos = 0; 475 ip->ip_len = 0; 476 ip->ip_id = 0; 477 ip->ip_off = 0; 478 ip->ip_ttl = 0; 479 ip->ip_sum = 0; 480 ip->ip_p = IPPROTO_TCP; 481 ip->ip_src = inp->inp_laddr; 482 ip->ip_dst = inp->inp_faddr; 483 484 if (tso) 485 plen = htons(IPPROTO_TCP); 486 else 487 plen = htons(sizeof(struct tcphdr) + IPPROTO_TCP); 488 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr, 489 ip->ip_dst.s_addr, plen); 490 } 491 492 tcp_hdr->th_sport = inp->inp_lport; 493 tcp_hdr->th_dport = inp->inp_fport; 494 tcp_hdr->th_seq = 0; 495 tcp_hdr->th_ack = 0; 496 tcp_hdr->th_x2 = 0; 497 tcp_hdr->th_off = 5; 498 tcp_hdr->th_flags = 0; 499 tcp_hdr->th_win = 0; 500 tcp_hdr->th_urp = 0; 501 } 502 503 /* 504 * Create template to be used to send tcp packets on a connection. 505 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only 506 * use for this function is in keepalives, which use tcp_respond. 507 */ 508 struct tcptemp * 509 tcp_maketemplate(struct tcpcb *tp) 510 { 511 struct tcptemp *tmp; 512 513 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL) 514 return (NULL); 515 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t, FALSE); 516 return (tmp); 517 } 518 519 void 520 tcp_freetemplate(struct tcptemp *tmp) 521 { 522 mpipe_free(&tcptemp_mpipe, tmp); 523 } 524 525 /* 526 * Send a single message to the TCP at address specified by 527 * the given TCP/IP header. If m == NULL, then we make a copy 528 * of the tcpiphdr at ti and send directly to the addressed host. 529 * This is used to force keep alive messages out using the TCP 530 * template for a connection. If flags are given then we send 531 * a message back to the TCP which originated the * segment ti, 532 * and discard the mbuf containing it and any other attached mbufs. 533 * 534 * In any case the ack and sequence number of the transmitted 535 * segment are as specified by the parameters. 536 * 537 * NOTE: If m != NULL, then ti must point to *inside* the mbuf. 538 */ 539 void 540 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, 541 tcp_seq ack, tcp_seq seq, int flags) 542 { 543 int tlen; 544 long win = 0; 545 struct route *ro = NULL; 546 struct route sro; 547 struct ip *ip = ipgen; 548 struct tcphdr *nth; 549 int ipflags = 0; 550 struct route_in6 *ro6 = NULL; 551 struct route_in6 sro6; 552 struct ip6_hdr *ip6 = ipgen; 553 boolean_t use_tmpro = TRUE; 554 #ifdef INET6 555 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6); 556 #else 557 const boolean_t isipv6 = FALSE; 558 #endif 559 560 if (tp != NULL) { 561 if (!(flags & TH_RST)) { 562 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv); 563 if (win < 0) 564 win = 0; 565 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 566 win = (long)TCP_MAXWIN << tp->rcv_scale; 567 } 568 /* 569 * Don't use the route cache of a listen socket, 570 * it is not MPSAFE; use temporary route cache. 571 */ 572 if (tp->t_state != TCPS_LISTEN) { 573 if (isipv6) 574 ro6 = &tp->t_inpcb->in6p_route; 575 else 576 ro = &tp->t_inpcb->inp_route; 577 use_tmpro = FALSE; 578 } 579 } 580 if (use_tmpro) { 581 if (isipv6) { 582 ro6 = &sro6; 583 bzero(ro6, sizeof *ro6); 584 } else { 585 ro = &sro; 586 bzero(ro, sizeof *ro); 587 } 588 } 589 if (m == NULL) { 590 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 591 if (m == NULL) 592 return; 593 tlen = 0; 594 m->m_data += max_linkhdr; 595 if (isipv6) { 596 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); 597 ip6 = mtod(m, struct ip6_hdr *); 598 nth = (struct tcphdr *)(ip6 + 1); 599 } else { 600 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip)); 601 ip = mtod(m, struct ip *); 602 nth = (struct tcphdr *)(ip + 1); 603 } 604 bcopy(th, nth, sizeof(struct tcphdr)); 605 flags = TH_ACK; 606 } else { 607 m_freem(m->m_next); 608 m->m_next = NULL; 609 m->m_data = (caddr_t)ipgen; 610 /* m_len is set later */ 611 tlen = 0; 612 #define xchg(a, b, type) { type t; t = a; a = b; b = t; } 613 if (isipv6) { 614 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); 615 nth = (struct tcphdr *)(ip6 + 1); 616 } else { 617 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long); 618 nth = (struct tcphdr *)(ip + 1); 619 } 620 if (th != nth) { 621 /* 622 * this is usually a case when an extension header 623 * exists between the IPv6 header and the 624 * TCP header. 625 */ 626 nth->th_sport = th->th_sport; 627 nth->th_dport = th->th_dport; 628 } 629 xchg(nth->th_dport, nth->th_sport, n_short); 630 #undef xchg 631 } 632 if (isipv6) { 633 ip6->ip6_flow = 0; 634 ip6->ip6_vfc = IPV6_VERSION; 635 ip6->ip6_nxt = IPPROTO_TCP; 636 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen)); 637 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 638 } else { 639 tlen += sizeof(struct tcpiphdr); 640 ip->ip_len = tlen; 641 ip->ip_ttl = ip_defttl; 642 } 643 m->m_len = tlen; 644 m->m_pkthdr.len = tlen; 645 m->m_pkthdr.rcvif = NULL; 646 nth->th_seq = htonl(seq); 647 nth->th_ack = htonl(ack); 648 nth->th_x2 = 0; 649 nth->th_off = sizeof(struct tcphdr) >> 2; 650 nth->th_flags = flags; 651 if (tp != NULL) 652 nth->th_win = htons((u_short) (win >> tp->rcv_scale)); 653 else 654 nth->th_win = htons((u_short)win); 655 nth->th_urp = 0; 656 if (isipv6) { 657 nth->th_sum = 0; 658 nth->th_sum = in6_cksum(m, IPPROTO_TCP, 659 sizeof(struct ip6_hdr), 660 tlen - sizeof(struct ip6_hdr)); 661 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL, 662 (ro6 && ro6->ro_rt) ? 663 ro6->ro_rt->rt_ifp : NULL); 664 } else { 665 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 666 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); 667 m->m_pkthdr.csum_flags = CSUM_TCP; 668 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 669 m->m_pkthdr.csum_thlen = sizeof(struct tcphdr); 670 } 671 #ifdef TCPDEBUG 672 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) 673 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); 674 #endif 675 if (isipv6) { 676 ip6_output(m, NULL, ro6, ipflags, NULL, NULL, 677 tp ? tp->t_inpcb : NULL); 678 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) { 679 RTFREE(ro6->ro_rt); 680 ro6->ro_rt = NULL; 681 } 682 } else { 683 ipflags |= IP_DEBUGROUTE; 684 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL); 685 if ((ro == &sro) && (ro->ro_rt != NULL)) { 686 RTFREE(ro->ro_rt); 687 ro->ro_rt = NULL; 688 } 689 } 690 } 691 692 /* 693 * Create a new TCP control block, making an 694 * empty reassembly queue and hooking it to the argument 695 * protocol control block. The `inp' parameter must have 696 * come from the zone allocator set up in tcp_init(). 697 */ 698 struct tcpcb * 699 tcp_newtcpcb(struct inpcb *inp) 700 { 701 struct inp_tp *it; 702 struct tcpcb *tp; 703 #ifdef INET6 704 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 705 #else 706 const boolean_t isipv6 = FALSE; 707 #endif 708 709 it = (struct inp_tp *)inp; 710 tp = &it->tcb; 711 bzero(tp, sizeof(struct tcpcb)); 712 TAILQ_INIT(&tp->t_segq); 713 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 714 tp->t_rxtthresh = tcprexmtthresh; 715 716 /* Set up our timeouts. */ 717 tp->tt_rexmt = &it->inp_tp_rexmt; 718 tp->tt_persist = &it->inp_tp_persist; 719 tp->tt_keep = &it->inp_tp_keep; 720 tp->tt_2msl = &it->inp_tp_2msl; 721 tp->tt_delack = &it->inp_tp_delack; 722 tcp_inittimers(tp); 723 724 /* 725 * Zero out timer message. We don't create it here, 726 * since the current CPU may not be the owner of this 727 * inpcb. 728 */ 729 tp->tt_msg = &it->inp_tp_timermsg; 730 bzero(tp->tt_msg, sizeof(*tp->tt_msg)); 731 732 tp->t_keepinit = tcp_keepinit; 733 tp->t_keepidle = tcp_keepidle; 734 tp->t_keepintvl = tcp_keepintvl; 735 tp->t_keepcnt = tcp_keepcnt; 736 tp->t_maxidle = tp->t_keepintvl * tp->t_keepcnt; 737 738 if (tcp_do_ncr) 739 tp->t_flags |= TF_NCR; 740 if (tcp_do_rfc1323) 741 tp->t_flags |= (TF_REQ_SCALE | TF_REQ_TSTMP); 742 743 tp->t_inpcb = inp; /* XXX */ 744 tp->t_state = TCPS_CLOSED; 745 /* 746 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 747 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 748 * reasonable initial retransmit time. 749 */ 750 tp->t_srtt = TCPTV_SRTTBASE; 751 tp->t_rttvar = 752 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 753 tp->t_rttmin = tcp_rexmit_min; 754 tp->t_rxtcur = TCPTV_RTOBASE; 755 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 756 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 757 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 758 tp->snd_last = ticks; 759 tp->t_rcvtime = ticks; 760 /* 761 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 762 * because the socket may be bound to an IPv6 wildcard address, 763 * which may match an IPv4-mapped IPv6 address. 764 */ 765 inp->inp_ip_ttl = ip_defttl; 766 inp->inp_ppcb = tp; 767 tcp_sack_tcpcb_init(tp); 768 769 tp->tt_sndmore = &it->inp_tp_sndmore; 770 tcp_output_init(tp); 771 772 return (tp); /* XXX */ 773 } 774 775 /* 776 * Drop a TCP connection, reporting the specified error. 777 * If connection is synchronized, then send a RST to peer. 778 */ 779 struct tcpcb * 780 tcp_drop(struct tcpcb *tp, int error) 781 { 782 struct socket *so = tp->t_inpcb->inp_socket; 783 784 if (TCPS_HAVERCVDSYN(tp->t_state)) { 785 tp->t_state = TCPS_CLOSED; 786 tcp_output(tp); 787 tcpstat.tcps_drops++; 788 } else 789 tcpstat.tcps_conndrops++; 790 if (error == ETIMEDOUT && tp->t_softerror) 791 error = tp->t_softerror; 792 so->so_error = error; 793 return (tcp_close(tp)); 794 } 795 796 struct netmsg_listen_detach { 797 struct netmsg_base base; 798 struct tcpcb *nm_tp; 799 struct tcpcb *nm_tp_inh; 800 }; 801 802 static void 803 tcp_listen_detach_handler(netmsg_t msg) 804 { 805 struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg; 806 struct tcpcb *tp = nmsg->nm_tp; 807 int cpu = mycpuid, nextcpu; 808 809 if (tp->t_flags & TF_LISTEN) 810 syncache_destroy(tp, nmsg->nm_tp_inh); 811 812 in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]); 813 814 nextcpu = cpu + 1; 815 if (nextcpu < ncpus2) 816 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nmsg->base.lmsg); 817 else 818 lwkt_replymsg(&nmsg->base.lmsg, 0); 819 } 820 821 /* 822 * Close a TCP control block: 823 * discard all space held by the tcp 824 * discard internet protocol block 825 * wake up any sleepers 826 */ 827 struct tcpcb * 828 tcp_close(struct tcpcb *tp) 829 { 830 struct tseg_qent *q; 831 struct inpcb *inp = tp->t_inpcb; 832 struct inpcb *inp_inh = NULL; 833 struct tcpcb *tp_inh = NULL; 834 struct socket *so = inp->inp_socket; 835 struct rtentry *rt; 836 boolean_t dosavessthresh; 837 #ifdef INET6 838 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 839 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0); 840 #else 841 const boolean_t isipv6 = FALSE; 842 #endif 843 844 if (tp->t_flags & TF_LISTEN) { 845 /* 846 * Pending socket/syncache inheritance 847 * 848 * If this is a listen(2) socket, find another listen(2) 849 * socket in the same local group, which could inherit 850 * the syncache and sockets pending on the completion 851 * and incompletion queues. 852 * 853 * NOTE: 854 * Currently the inheritance could only happen on the 855 * listen(2) sockets w/ SO_REUSEPORT set. 856 */ 857 KASSERT(&curthread->td_msgport == netisr_cpuport(0), 858 ("listen socket close not in netisr0")); 859 inp_inh = in_pcblocalgroup_last(&tcbinfo[0], inp); 860 if (inp_inh != NULL) 861 tp_inh = intotcpcb(inp_inh); 862 } 863 864 /* 865 * INP_WILDCARD indicates that listen(2) has been called on 866 * this socket. This implies: 867 * - A wildcard inp's hash is replicated for each protocol thread. 868 * - Syncache for this inp grows independently in each protocol 869 * thread. 870 * - There is more than one cpu 871 * 872 * We have to chain a message to the rest of the protocol threads 873 * to cleanup the wildcard hash and the syncache. The cleanup 874 * in the current protocol thread is defered till the end of this 875 * function (syncache_destroy and in_pcbdetach). 876 * 877 * NOTE: 878 * After cleanup the inp's hash and syncache entries, this inp will 879 * no longer be available to the rest of the protocol threads, so we 880 * are safe to whack the inp in the following code. 881 */ 882 if ((inp->inp_flags & INP_WILDCARD) && ncpus2 > 1) { 883 struct netmsg_listen_detach nmsg; 884 885 KKASSERT(so->so_port == netisr_cpuport(0)); 886 KKASSERT(&curthread->td_msgport == netisr_cpuport(0)); 887 KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]); 888 889 netmsg_init(&nmsg.base, NULL, &curthread->td_msgport, 890 MSGF_PRIORITY, tcp_listen_detach_handler); 891 nmsg.nm_tp = tp; 892 nmsg.nm_tp_inh = tp_inh; 893 lwkt_domsg(netisr_cpuport(1), &nmsg.base.lmsg, 0); 894 } 895 896 KKASSERT(tp->t_state != TCPS_TERMINATING); 897 tp->t_state = TCPS_TERMINATING; 898 899 /* 900 * Make sure that all of our timers are stopped before we 901 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL), 902 * timers are never used. If timer message is never created 903 * (tp->tt_msg->tt_tcb == NULL), timers are never used too. 904 */ 905 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) { 906 tcp_callout_stop(tp, tp->tt_rexmt); 907 tcp_callout_stop(tp, tp->tt_persist); 908 tcp_callout_stop(tp, tp->tt_keep); 909 tcp_callout_stop(tp, tp->tt_2msl); 910 tcp_callout_stop(tp, tp->tt_delack); 911 } 912 913 if (tp->t_flags & TF_ONOUTPUTQ) { 914 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 915 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 916 tp->t_flags &= ~TF_ONOUTPUTQ; 917 } 918 919 /* 920 * If we got enough samples through the srtt filter, 921 * save the rtt and rttvar in the routing entry. 922 * 'Enough' is arbitrarily defined as the 16 samples. 923 * 16 samples is enough for the srtt filter to converge 924 * to within 5% of the correct value; fewer samples and 925 * we could save a very bogus rtt. 926 * 927 * Don't update the default route's characteristics and don't 928 * update anything that the user "locked". 929 */ 930 if (tp->t_rttupdated >= 16) { 931 u_long i = 0; 932 933 if (isipv6) { 934 struct sockaddr_in6 *sin6; 935 936 if ((rt = inp->in6p_route.ro_rt) == NULL) 937 goto no_valid_rt; 938 sin6 = (struct sockaddr_in6 *)rt_key(rt); 939 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 940 goto no_valid_rt; 941 } else 942 if ((rt = inp->inp_route.ro_rt) == NULL || 943 ((struct sockaddr_in *)rt_key(rt))-> 944 sin_addr.s_addr == INADDR_ANY) 945 goto no_valid_rt; 946 947 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 948 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 949 if (rt->rt_rmx.rmx_rtt && i) 950 /* 951 * filter this update to half the old & half 952 * the new values, converting scale. 953 * See route.h and tcp_var.h for a 954 * description of the scaling constants. 955 */ 956 rt->rt_rmx.rmx_rtt = 957 (rt->rt_rmx.rmx_rtt + i) / 2; 958 else 959 rt->rt_rmx.rmx_rtt = i; 960 tcpstat.tcps_cachedrtt++; 961 } 962 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 963 i = tp->t_rttvar * 964 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 965 if (rt->rt_rmx.rmx_rttvar && i) 966 rt->rt_rmx.rmx_rttvar = 967 (rt->rt_rmx.rmx_rttvar + i) / 2; 968 else 969 rt->rt_rmx.rmx_rttvar = i; 970 tcpstat.tcps_cachedrttvar++; 971 } 972 /* 973 * The old comment here said: 974 * update the pipelimit (ssthresh) if it has been updated 975 * already or if a pipesize was specified & the threshhold 976 * got below half the pipesize. I.e., wait for bad news 977 * before we start updating, then update on both good 978 * and bad news. 979 * 980 * But we want to save the ssthresh even if no pipesize is 981 * specified explicitly in the route, because such 982 * connections still have an implicit pipesize specified 983 * by the global tcp_sendspace. In the absence of a reliable 984 * way to calculate the pipesize, it will have to do. 985 */ 986 i = tp->snd_ssthresh; 987 if (rt->rt_rmx.rmx_sendpipe != 0) 988 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 989 else 990 dosavessthresh = (i < so->so_snd.ssb_hiwat/2); 991 if (dosavessthresh || 992 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 993 (rt->rt_rmx.rmx_ssthresh != 0))) { 994 /* 995 * convert the limit from user data bytes to 996 * packets then to packet data bytes. 997 */ 998 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 999 if (i < 2) 1000 i = 2; 1001 i *= tp->t_maxseg + 1002 (isipv6 ? 1003 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1004 sizeof(struct tcpiphdr)); 1005 if (rt->rt_rmx.rmx_ssthresh) 1006 rt->rt_rmx.rmx_ssthresh = 1007 (rt->rt_rmx.rmx_ssthresh + i) / 2; 1008 else 1009 rt->rt_rmx.rmx_ssthresh = i; 1010 tcpstat.tcps_cachedssthresh++; 1011 } 1012 } 1013 1014 no_valid_rt: 1015 /* free the reassembly queue, if any */ 1016 while((q = TAILQ_FIRST(&tp->t_segq)) != NULL) { 1017 TAILQ_REMOVE(&tp->t_segq, q, tqe_q); 1018 m_freem(q->tqe_m); 1019 kfree(q, M_TSEGQ); 1020 atomic_add_int(&tcp_reass_qsize, -1); 1021 } 1022 /* throw away SACK blocks in scoreboard*/ 1023 if (TCP_DO_SACK(tp)) 1024 tcp_sack_destroy(&tp->scb); 1025 1026 inp->inp_ppcb = NULL; 1027 soisdisconnected(so); 1028 /* note: pcb detached later on */ 1029 1030 tcp_destroy_timermsg(tp); 1031 tcp_output_cancel(tp); 1032 1033 if (tp->t_flags & TF_LISTEN) { 1034 syncache_destroy(tp, tp_inh); 1035 if (inp_inh != NULL && inp_inh->inp_socket != NULL) { 1036 /* 1037 * Pending sockets inheritance only needs 1038 * to be done once in the current thread, 1039 * i.e. netisr0. 1040 */ 1041 soinherit(so, inp_inh->inp_socket); 1042 } 1043 } 1044 1045 so_async_rcvd_drop(so); 1046 /* Drop the reference for the asynchronized pru_rcvd */ 1047 sofree(so); 1048 1049 /* 1050 * NOTE: 1051 * pcbdetach removes any wildcard hash entry on the current CPU. 1052 */ 1053 #ifdef INET6 1054 if (isafinet6) 1055 in6_pcbdetach(inp); 1056 else 1057 #endif 1058 in_pcbdetach(inp); 1059 1060 tcpstat.tcps_closed++; 1061 return (NULL); 1062 } 1063 1064 static __inline void 1065 tcp_drain_oncpu(struct inpcbinfo *pcbinfo) 1066 { 1067 struct inpcbhead *head = &pcbinfo->pcblisthead; 1068 struct inpcb *inpb; 1069 1070 /* 1071 * Since we run in netisr, it is MP safe, even if 1072 * we block during the inpcb list iteration, i.e. 1073 * we don't need to use inpcb marker here. 1074 */ 1075 KASSERT(&curthread->td_msgport == netisr_cpuport(pcbinfo->cpu), 1076 ("not in correct netisr")); 1077 1078 LIST_FOREACH(inpb, head, inp_list) { 1079 struct tcpcb *tcpb; 1080 struct tseg_qent *te; 1081 1082 if ((inpb->inp_flags & INP_PLACEMARKER) == 0 && 1083 (tcpb = intotcpcb(inpb)) != NULL && 1084 (te = TAILQ_FIRST(&tcpb->t_segq)) != NULL) { 1085 TAILQ_REMOVE(&tcpb->t_segq, te, tqe_q); 1086 if (te->tqe_th->th_flags & TH_FIN) 1087 tcpb->t_flags &= ~TF_QUEDFIN; 1088 m_freem(te->tqe_m); 1089 kfree(te, M_TSEGQ); 1090 atomic_add_int(&tcp_reass_qsize, -1); 1091 /* retry */ 1092 } 1093 } 1094 } 1095 1096 static void 1097 tcp_drain_dispatch(netmsg_t nmsg) 1098 { 1099 crit_enter(); 1100 lwkt_replymsg(&nmsg->lmsg, 0); /* reply ASAP */ 1101 crit_exit(); 1102 1103 tcp_drain_oncpu(&tcbinfo[mycpuid]); 1104 } 1105 1106 static void 1107 tcp_drain_ipi(void *arg __unused) 1108 { 1109 int cpu = mycpuid; 1110 struct lwkt_msg *msg = &tcp_drain_netmsg[cpu].lmsg; 1111 1112 crit_enter(); 1113 if (msg->ms_flags & MSGF_DONE) 1114 lwkt_sendmsg_oncpu(netisr_cpuport(cpu), msg); 1115 crit_exit(); 1116 } 1117 1118 void 1119 tcp_drain(void) 1120 { 1121 cpumask_t mask; 1122 1123 if (!do_tcpdrain) 1124 return; 1125 1126 /* 1127 * Walk the tcpbs, if existing, and flush the reassembly queue, 1128 * if there is one... 1129 * XXX: The "Net/3" implementation doesn't imply that the TCP 1130 * reassembly queue should be flushed, but in a situation 1131 * where we're really low on mbufs, this is potentially 1132 * useful. 1133 * YYY: We may consider run tcp_drain_oncpu directly here, 1134 * however, that will require M_WAITOK memory allocation 1135 * for the inpcb marker. 1136 */ 1137 CPUMASK_ASSBMASK(mask, ncpus2); 1138 CPUMASK_ANDMASK(mask, smp_active_mask); 1139 if (CPUMASK_TESTNZERO(mask)) 1140 lwkt_send_ipiq_mask(mask, tcp_drain_ipi, NULL); 1141 } 1142 1143 /* 1144 * Notify a tcp user of an asynchronous error; 1145 * store error as soft error, but wake up user 1146 * (for now, won't do anything until can select for soft error). 1147 * 1148 * Do not wake up user since there currently is no mechanism for 1149 * reporting soft errors (yet - a kqueue filter may be added). 1150 */ 1151 static void 1152 tcp_notify(struct inpcb *inp, int error) 1153 { 1154 struct tcpcb *tp = intotcpcb(inp); 1155 1156 /* 1157 * Ignore some errors if we are hooked up. 1158 * If connection hasn't completed, has retransmitted several times, 1159 * and receives a second error, give up now. This is better 1160 * than waiting a long time to establish a connection that 1161 * can never complete. 1162 */ 1163 if (tp->t_state == TCPS_ESTABLISHED && 1164 (error == EHOSTUNREACH || error == ENETUNREACH || 1165 error == EHOSTDOWN)) { 1166 return; 1167 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1168 tp->t_softerror) 1169 tcp_drop(tp, error); 1170 else 1171 tp->t_softerror = error; 1172 #if 0 1173 wakeup(&so->so_timeo); 1174 sorwakeup(so); 1175 sowwakeup(so); 1176 #endif 1177 } 1178 1179 static int 1180 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1181 { 1182 int error, i, n; 1183 struct inpcb *marker; 1184 struct inpcb *inp; 1185 int origcpu, ccpu; 1186 1187 error = 0; 1188 n = 0; 1189 1190 /* 1191 * The process of preparing the TCB list is too time-consuming and 1192 * resource-intensive to repeat twice on every request. 1193 */ 1194 if (req->oldptr == NULL) { 1195 for (ccpu = 0; ccpu < ncpus2; ++ccpu) 1196 n += tcbinfo[ccpu].ipi_count; 1197 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1198 return (0); 1199 } 1200 1201 if (req->newptr != NULL) 1202 return (EPERM); 1203 1204 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1205 marker->inp_flags |= INP_PLACEMARKER; 1206 1207 /* 1208 * OK, now we're committed to doing something. Run the inpcb list 1209 * for each cpu in the system and construct the output. Use a 1210 * list placemarker to deal with list changes occuring during 1211 * copyout blockages (but otherwise depend on being on the correct 1212 * cpu to avoid races). 1213 */ 1214 origcpu = mycpu->gd_cpuid; 1215 for (ccpu = 0; ccpu < ncpus2 && error == 0; ++ccpu) { 1216 caddr_t inp_ppcb; 1217 struct xtcpcb xt; 1218 1219 lwkt_migratecpu(ccpu); 1220 1221 n = tcbinfo[ccpu].ipi_count; 1222 1223 LIST_INSERT_HEAD(&tcbinfo[ccpu].pcblisthead, marker, inp_list); 1224 i = 0; 1225 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1226 /* 1227 * process a snapshot of pcbs, ignoring placemarkers 1228 * and using our own to allow SYSCTL_OUT to block. 1229 */ 1230 LIST_REMOVE(marker, inp_list); 1231 LIST_INSERT_AFTER(inp, marker, inp_list); 1232 1233 if (inp->inp_flags & INP_PLACEMARKER) 1234 continue; 1235 if (prison_xinpcb(req->td, inp)) 1236 continue; 1237 1238 xt.xt_len = sizeof xt; 1239 bcopy(inp, &xt.xt_inp, sizeof *inp); 1240 inp_ppcb = inp->inp_ppcb; 1241 if (inp_ppcb != NULL) 1242 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1243 else 1244 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1245 if (inp->inp_socket) 1246 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1247 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1248 break; 1249 ++i; 1250 } 1251 LIST_REMOVE(marker, inp_list); 1252 if (error == 0 && i < n) { 1253 bzero(&xt, sizeof xt); 1254 xt.xt_len = sizeof xt; 1255 while (i < n) { 1256 error = SYSCTL_OUT(req, &xt, sizeof xt); 1257 if (error) 1258 break; 1259 ++i; 1260 } 1261 } 1262 } 1263 1264 /* 1265 * Make sure we are on the same cpu we were on originally, since 1266 * higher level callers expect this. Also don't pollute caches with 1267 * migrated userland data by (eventually) returning to userland 1268 * on a different cpu. 1269 */ 1270 lwkt_migratecpu(origcpu); 1271 kfree(marker, M_TEMP); 1272 return (error); 1273 } 1274 1275 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1276 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1277 1278 static int 1279 tcp_getcred(SYSCTL_HANDLER_ARGS) 1280 { 1281 struct sockaddr_in addrs[2]; 1282 struct ucred cred0, *cred = NULL; 1283 struct inpcb *inp; 1284 int cpu, origcpu, error; 1285 1286 error = priv_check(req->td, PRIV_ROOT); 1287 if (error != 0) 1288 return (error); 1289 error = SYSCTL_IN(req, addrs, sizeof addrs); 1290 if (error != 0) 1291 return (error); 1292 1293 origcpu = mycpuid; 1294 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1295 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1296 1297 lwkt_migratecpu(cpu); 1298 1299 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1300 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1301 if (inp == NULL || inp->inp_socket == NULL) { 1302 error = ENOENT; 1303 } else if (inp->inp_socket->so_cred != NULL) { 1304 cred0 = *(inp->inp_socket->so_cred); 1305 cred = &cred0; 1306 } 1307 1308 lwkt_migratecpu(origcpu); 1309 1310 if (error) 1311 return (error); 1312 1313 return SYSCTL_OUT(req, cred, sizeof(struct ucred)); 1314 } 1315 1316 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1317 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1318 1319 #ifdef INET6 1320 static int 1321 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1322 { 1323 struct sockaddr_in6 addrs[2]; 1324 struct inpcb *inp; 1325 int error; 1326 boolean_t mapped = FALSE; 1327 1328 error = priv_check(req->td, PRIV_ROOT); 1329 if (error != 0) 1330 return (error); 1331 error = SYSCTL_IN(req, addrs, sizeof addrs); 1332 if (error != 0) 1333 return (error); 1334 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { 1335 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) 1336 mapped = TRUE; 1337 else 1338 return (EINVAL); 1339 } 1340 crit_enter(); 1341 if (mapped) { 1342 inp = in_pcblookup_hash(&tcbinfo[0], 1343 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], 1344 addrs[1].sin6_port, 1345 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], 1346 addrs[0].sin6_port, 1347 0, NULL); 1348 } else { 1349 inp = in6_pcblookup_hash(&tcbinfo[0], 1350 &addrs[1].sin6_addr, addrs[1].sin6_port, 1351 &addrs[0].sin6_addr, addrs[0].sin6_port, 1352 0, NULL); 1353 } 1354 if (inp == NULL || inp->inp_socket == NULL) { 1355 error = ENOENT; 1356 goto out; 1357 } 1358 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1359 out: 1360 crit_exit(); 1361 return (error); 1362 } 1363 1364 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1365 0, 0, 1366 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1367 #endif 1368 1369 struct netmsg_tcp_notify { 1370 struct netmsg_base base; 1371 inp_notify_t nm_notify; 1372 struct in_addr nm_faddr; 1373 int nm_arg; 1374 }; 1375 1376 static void 1377 tcp_notifyall_oncpu(netmsg_t msg) 1378 { 1379 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg; 1380 int nextcpu; 1381 1382 in_pcbnotifyall(&tcbinfo[mycpuid], nm->nm_faddr, 1383 nm->nm_arg, nm->nm_notify); 1384 1385 nextcpu = mycpuid + 1; 1386 if (nextcpu < ncpus2) 1387 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nm->base.lmsg); 1388 else 1389 lwkt_replymsg(&nm->base.lmsg, 0); 1390 } 1391 1392 inp_notify_t 1393 tcp_get_inpnotify(int cmd, const struct sockaddr *sa, 1394 int *arg, struct ip **ip0, int *cpuid) 1395 { 1396 struct ip *ip = *ip0; 1397 struct in_addr faddr; 1398 inp_notify_t notify = tcp_notify; 1399 1400 faddr = ((const struct sockaddr_in *)sa)->sin_addr; 1401 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1402 return NULL; 1403 1404 *arg = inetctlerrmap[cmd]; 1405 if (cmd == PRC_QUENCH) { 1406 notify = tcp_quench; 1407 } else if (icmp_may_rst && 1408 (cmd == PRC_UNREACH_ADMIN_PROHIB || 1409 cmd == PRC_UNREACH_PORT || 1410 cmd == PRC_TIMXCEED_INTRANS) && 1411 ip != NULL) { 1412 notify = tcp_drop_syn_sent; 1413 } else if (cmd == PRC_MSGSIZE) { 1414 const struct icmp *icmp = (const struct icmp *) 1415 ((caddr_t)ip - offsetof(struct icmp, icmp_ip)); 1416 1417 *arg = ntohs(icmp->icmp_nextmtu); 1418 notify = tcp_mtudisc; 1419 } else if (PRC_IS_REDIRECT(cmd)) { 1420 ip = NULL; 1421 notify = in_rtchange; 1422 } else if (cmd == PRC_HOSTDEAD) { 1423 ip = NULL; 1424 } else if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) { 1425 return NULL; 1426 } 1427 1428 if (cpuid != NULL) { 1429 if (ip == NULL) { 1430 /* Go through all CPUs */ 1431 *cpuid = ncpus; 1432 } else { 1433 const struct tcphdr *th; 1434 1435 th = (const struct tcphdr *) 1436 ((caddr_t)ip + (IP_VHL_HL(ip->ip_vhl) << 2)); 1437 *cpuid = tcp_addrcpu(faddr.s_addr, th->th_dport, 1438 ip->ip_src.s_addr, th->th_sport); 1439 } 1440 } 1441 1442 *ip0 = ip; 1443 return notify; 1444 } 1445 1446 void 1447 tcp_ctlinput(netmsg_t msg) 1448 { 1449 int cmd = msg->ctlinput.nm_cmd; 1450 struct sockaddr *sa = msg->ctlinput.nm_arg; 1451 struct ip *ip = msg->ctlinput.nm_extra; 1452 struct in_addr faddr; 1453 inp_notify_t notify; 1454 int arg; 1455 1456 notify = tcp_get_inpnotify(cmd, sa, &arg, &ip, NULL); 1457 if (notify == NULL) 1458 goto done; 1459 1460 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1461 if (ip != NULL) { 1462 const struct tcphdr *th; 1463 struct inpcb *inp; 1464 1465 th = (const struct tcphdr *) 1466 ((caddr_t)ip + (IP_VHL_HL(ip->ip_vhl) << 2)); 1467 inp = in_pcblookup_hash(&tcbinfo[mycpuid], faddr, th->th_dport, 1468 ip->ip_src, th->th_sport, 0, NULL); 1469 if (inp != NULL && inp->inp_socket != NULL) { 1470 tcp_seq icmpseq = htonl(th->th_seq); 1471 struct tcpcb *tp = intotcpcb(inp); 1472 1473 if (SEQ_GEQ(icmpseq, tp->snd_una) && 1474 SEQ_LT(icmpseq, tp->snd_max)) 1475 notify(inp, arg); 1476 } else { 1477 struct in_conninfo inc; 1478 1479 inc.inc_fport = th->th_dport; 1480 inc.inc_lport = th->th_sport; 1481 inc.inc_faddr = faddr; 1482 inc.inc_laddr = ip->ip_src; 1483 #ifdef INET6 1484 inc.inc_isipv6 = 0; 1485 #endif 1486 syncache_unreach(&inc, th); 1487 } 1488 } else { 1489 struct netmsg_tcp_notify *nm; 1490 1491 KKASSERT(&curthread->td_msgport == netisr_cpuport(0)); 1492 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT); 1493 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1494 0, tcp_notifyall_oncpu); 1495 nm->nm_faddr = faddr; 1496 nm->nm_arg = arg; 1497 nm->nm_notify = notify; 1498 1499 lwkt_sendmsg(netisr_cpuport(0), &nm->base.lmsg); 1500 } 1501 done: 1502 lwkt_replymsg(&msg->lmsg, 0); 1503 } 1504 1505 #ifdef INET6 1506 1507 void 1508 tcp6_ctlinput(netmsg_t msg) 1509 { 1510 int cmd = msg->ctlinput.nm_cmd; 1511 struct sockaddr *sa = msg->ctlinput.nm_arg; 1512 void *d = msg->ctlinput.nm_extra; 1513 struct tcphdr th; 1514 inp_notify_t notify = tcp_notify; 1515 struct ip6_hdr *ip6; 1516 struct mbuf *m; 1517 struct ip6ctlparam *ip6cp = NULL; 1518 const struct sockaddr_in6 *sa6_src = NULL; 1519 int off; 1520 struct tcp_portonly { 1521 u_int16_t th_sport; 1522 u_int16_t th_dport; 1523 } *thp; 1524 int arg; 1525 1526 if (sa->sa_family != AF_INET6 || 1527 sa->sa_len != sizeof(struct sockaddr_in6)) { 1528 goto out; 1529 } 1530 1531 arg = 0; 1532 if (cmd == PRC_QUENCH) 1533 notify = tcp_quench; 1534 else if (cmd == PRC_MSGSIZE) { 1535 struct ip6ctlparam *ip6cp = d; 1536 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6; 1537 1538 arg = ntohl(icmp6->icmp6_mtu); 1539 notify = tcp_mtudisc; 1540 } else if (!PRC_IS_REDIRECT(cmd) && 1541 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) { 1542 goto out; 1543 } 1544 1545 /* if the parameter is from icmp6, decode it. */ 1546 if (d != NULL) { 1547 ip6cp = (struct ip6ctlparam *)d; 1548 m = ip6cp->ip6c_m; 1549 ip6 = ip6cp->ip6c_ip6; 1550 off = ip6cp->ip6c_off; 1551 sa6_src = ip6cp->ip6c_src; 1552 } else { 1553 m = NULL; 1554 ip6 = NULL; 1555 off = 0; /* fool gcc */ 1556 sa6_src = &sa6_any; 1557 } 1558 1559 if (ip6 != NULL) { 1560 struct in_conninfo inc; 1561 /* 1562 * XXX: We assume that when IPV6 is non NULL, 1563 * M and OFF are valid. 1564 */ 1565 1566 /* check if we can safely examine src and dst ports */ 1567 if (m->m_pkthdr.len < off + sizeof *thp) 1568 goto out; 1569 1570 bzero(&th, sizeof th); 1571 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1572 1573 in6_pcbnotify(&tcbinfo[0], sa, th.th_dport, 1574 (struct sockaddr *)ip6cp->ip6c_src, 1575 th.th_sport, cmd, arg, notify); 1576 1577 inc.inc_fport = th.th_dport; 1578 inc.inc_lport = th.th_sport; 1579 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1580 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1581 inc.inc_isipv6 = 1; 1582 syncache_unreach(&inc, &th); 1583 } else { 1584 in6_pcbnotify(&tcbinfo[0], sa, 0, 1585 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify); 1586 } 1587 out: 1588 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0); 1589 } 1590 1591 #endif 1592 1593 /* 1594 * Following is where TCP initial sequence number generation occurs. 1595 * 1596 * There are two places where we must use initial sequence numbers: 1597 * 1. In SYN-ACK packets. 1598 * 2. In SYN packets. 1599 * 1600 * All ISNs for SYN-ACK packets are generated by the syncache. See 1601 * tcp_syncache.c for details. 1602 * 1603 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1604 * depends on this property. In addition, these ISNs should be 1605 * unguessable so as to prevent connection hijacking. To satisfy 1606 * the requirements of this situation, the algorithm outlined in 1607 * RFC 1948 is used to generate sequence numbers. 1608 * 1609 * Implementation details: 1610 * 1611 * Time is based off the system timer, and is corrected so that it 1612 * increases by one megabyte per second. This allows for proper 1613 * recycling on high speed LANs while still leaving over an hour 1614 * before rollover. 1615 * 1616 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1617 * between seeding of isn_secret. This is normally set to zero, 1618 * as reseeding should not be necessary. 1619 * 1620 */ 1621 1622 #define ISN_BYTES_PER_SECOND 1048576 1623 1624 u_char isn_secret[32]; 1625 int isn_last_reseed; 1626 MD5_CTX isn_ctx; 1627 1628 tcp_seq 1629 tcp_new_isn(struct tcpcb *tp) 1630 { 1631 u_int32_t md5_buffer[4]; 1632 tcp_seq new_isn; 1633 1634 /* Seed if this is the first use, reseed if requested. */ 1635 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1636 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1637 < (u_int)ticks))) { 1638 read_random_unlimited(&isn_secret, sizeof isn_secret); 1639 isn_last_reseed = ticks; 1640 } 1641 1642 /* Compute the md5 hash and return the ISN. */ 1643 MD5Init(&isn_ctx); 1644 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1645 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1646 #ifdef INET6 1647 if (tp->t_inpcb->inp_vflag & INP_IPV6) { 1648 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1649 sizeof(struct in6_addr)); 1650 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1651 sizeof(struct in6_addr)); 1652 } else 1653 #endif 1654 { 1655 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1656 sizeof(struct in_addr)); 1657 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1658 sizeof(struct in_addr)); 1659 } 1660 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1661 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1662 new_isn = (tcp_seq) md5_buffer[0]; 1663 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1664 return (new_isn); 1665 } 1666 1667 /* 1668 * When a source quench is received, close congestion window 1669 * to one segment. We will gradually open it again as we proceed. 1670 */ 1671 void 1672 tcp_quench(struct inpcb *inp, int error) 1673 { 1674 struct tcpcb *tp = intotcpcb(inp); 1675 1676 if (tp != NULL) { 1677 tp->snd_cwnd = tp->t_maxseg; 1678 tp->snd_wacked = 0; 1679 } 1680 } 1681 1682 /* 1683 * When a specific ICMP unreachable message is received and the 1684 * connection state is SYN-SENT, drop the connection. This behavior 1685 * is controlled by the icmp_may_rst sysctl. 1686 */ 1687 void 1688 tcp_drop_syn_sent(struct inpcb *inp, int error) 1689 { 1690 struct tcpcb *tp = intotcpcb(inp); 1691 1692 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT)) 1693 tcp_drop(tp, error); 1694 } 1695 1696 /* 1697 * When a `need fragmentation' ICMP is received, update our idea of the MSS 1698 * based on the new value in the route. Also nudge TCP to send something, 1699 * since we know the packet we just sent was dropped. 1700 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1701 */ 1702 void 1703 tcp_mtudisc(struct inpcb *inp, int mtu) 1704 { 1705 struct tcpcb *tp = intotcpcb(inp); 1706 struct rtentry *rt; 1707 struct socket *so = inp->inp_socket; 1708 int maxopd, mss; 1709 #ifdef INET6 1710 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0); 1711 #else 1712 const boolean_t isipv6 = FALSE; 1713 #endif 1714 1715 if (tp == NULL) 1716 return; 1717 1718 /* 1719 * If no MTU is provided in the ICMP message, use the 1720 * next lower likely value, as specified in RFC 1191. 1721 */ 1722 if (mtu == 0) { 1723 int oldmtu; 1724 1725 oldmtu = tp->t_maxopd + 1726 (isipv6 ? 1727 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1728 sizeof(struct tcpiphdr)); 1729 mtu = ip_next_mtu(oldmtu, 0); 1730 } 1731 1732 if (isipv6) 1733 rt = tcp_rtlookup6(&inp->inp_inc); 1734 else 1735 rt = tcp_rtlookup(&inp->inp_inc); 1736 if (rt != NULL) { 1737 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu) 1738 mtu = rt->rt_rmx.rmx_mtu; 1739 1740 maxopd = mtu - 1741 (isipv6 ? 1742 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1743 sizeof(struct tcpiphdr)); 1744 1745 /* 1746 * XXX - The following conditional probably violates the TCP 1747 * spec. The problem is that, since we don't know the 1748 * other end's MSS, we are supposed to use a conservative 1749 * default. But, if we do that, then MTU discovery will 1750 * never actually take place, because the conservative 1751 * default is much less than the MTUs typically seen 1752 * on the Internet today. For the moment, we'll sweep 1753 * this under the carpet. 1754 * 1755 * The conservative default might not actually be a problem 1756 * if the only case this occurs is when sending an initial 1757 * SYN with options and data to a host we've never talked 1758 * to before. Then, they will reply with an MSS value which 1759 * will get recorded and the new parameters should get 1760 * recomputed. For Further Study. 1761 */ 1762 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd) 1763 maxopd = rt->rt_rmx.rmx_mssopt; 1764 } else 1765 maxopd = mtu - 1766 (isipv6 ? 1767 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1768 sizeof(struct tcpiphdr)); 1769 1770 if (tp->t_maxopd <= maxopd) 1771 return; 1772 tp->t_maxopd = maxopd; 1773 1774 mss = maxopd; 1775 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) == 1776 (TF_REQ_TSTMP | TF_RCVD_TSTMP)) 1777 mss -= TCPOLEN_TSTAMP_APPA; 1778 1779 /* round down to multiple of MCLBYTES */ 1780 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */ 1781 if (mss > MCLBYTES) 1782 mss &= ~(MCLBYTES - 1); 1783 #else 1784 if (mss > MCLBYTES) 1785 mss = (mss / MCLBYTES) * MCLBYTES; 1786 #endif 1787 1788 if (so->so_snd.ssb_hiwat < mss) 1789 mss = so->so_snd.ssb_hiwat; 1790 1791 tp->t_maxseg = mss; 1792 tp->t_rtttime = 0; 1793 tp->snd_nxt = tp->snd_una; 1794 tcp_output(tp); 1795 tcpstat.tcps_mturesent++; 1796 } 1797 1798 /* 1799 * Look-up the routing entry to the peer of this inpcb. If no route 1800 * is found and it cannot be allocated the return NULL. This routine 1801 * is called by TCP routines that access the rmx structure and by tcp_mss 1802 * to get the interface MTU. 1803 */ 1804 struct rtentry * 1805 tcp_rtlookup(struct in_conninfo *inc) 1806 { 1807 struct route *ro = &inc->inc_route; 1808 1809 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1810 /* No route yet, so try to acquire one */ 1811 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1812 /* 1813 * unused portions of the structure MUST be zero'd 1814 * out because rtalloc() treats it as opaque data 1815 */ 1816 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1817 ro->ro_dst.sa_family = AF_INET; 1818 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1819 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1820 inc->inc_faddr; 1821 rtalloc(ro); 1822 } 1823 } 1824 return (ro->ro_rt); 1825 } 1826 1827 #ifdef INET6 1828 struct rtentry * 1829 tcp_rtlookup6(struct in_conninfo *inc) 1830 { 1831 struct route_in6 *ro6 = &inc->inc6_route; 1832 1833 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1834 /* No route yet, so try to acquire one */ 1835 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1836 /* 1837 * unused portions of the structure MUST be zero'd 1838 * out because rtalloc() treats it as opaque data 1839 */ 1840 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1841 ro6->ro_dst.sin6_family = AF_INET6; 1842 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1843 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1844 rtalloc((struct route *)ro6); 1845 } 1846 } 1847 return (ro6->ro_rt); 1848 } 1849 #endif 1850 1851 #ifdef IPSEC 1852 /* compute ESP/AH header size for TCP, including outer IP header. */ 1853 size_t 1854 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1855 { 1856 struct inpcb *inp; 1857 struct mbuf *m; 1858 size_t hdrsiz; 1859 struct ip *ip; 1860 struct tcphdr *th; 1861 1862 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1863 return (0); 1864 MGETHDR(m, MB_DONTWAIT, MT_DATA); 1865 if (!m) 1866 return (0); 1867 1868 #ifdef INET6 1869 if (inp->inp_vflag & INP_IPV6) { 1870 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1871 1872 th = (struct tcphdr *)(ip6 + 1); 1873 m->m_pkthdr.len = m->m_len = 1874 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1875 tcp_fillheaders(tp, ip6, th, FALSE); 1876 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1877 } else 1878 #endif 1879 { 1880 ip = mtod(m, struct ip *); 1881 th = (struct tcphdr *)(ip + 1); 1882 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1883 tcp_fillheaders(tp, ip, th, FALSE); 1884 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1885 } 1886 1887 m_free(m); 1888 return (hdrsiz); 1889 } 1890 #endif 1891 1892 /* 1893 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1894 * 1895 * This code attempts to calculate the bandwidth-delay product as a 1896 * means of determining the optimal window size to maximize bandwidth, 1897 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1898 * routers. This code also does a fairly good job keeping RTTs in check 1899 * across slow links like modems. We implement an algorithm which is very 1900 * similar (but not meant to be) TCP/Vegas. The code operates on the 1901 * transmitter side of a TCP connection and so only effects the transmit 1902 * side of the connection. 1903 * 1904 * BACKGROUND: TCP makes no provision for the management of buffer space 1905 * at the end points or at the intermediate routers and switches. A TCP 1906 * stream, whether using NewReno or not, will eventually buffer as 1907 * many packets as it is able and the only reason this typically works is 1908 * due to the fairly small default buffers made available for a connection 1909 * (typicaly 16K or 32K). As machines use larger windows and/or window 1910 * scaling it is now fairly easy for even a single TCP connection to blow-out 1911 * all available buffer space not only on the local interface, but on 1912 * intermediate routers and switches as well. NewReno makes a misguided 1913 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1914 * then backing off, then steadily increasing the window again until another 1915 * failure occurs, ad-infinitum. This results in terrible oscillation that 1916 * is only made worse as network loads increase and the idea of intentionally 1917 * blowing out network buffers is, frankly, a terrible way to manage network 1918 * resources. 1919 * 1920 * It is far better to limit the transmit window prior to the failure 1921 * condition being achieved. There are two general ways to do this: First 1922 * you can 'scan' through different transmit window sizes and locate the 1923 * point where the RTT stops increasing, indicating that you have filled the 1924 * pipe, then scan backwards until you note that RTT stops decreasing, then 1925 * repeat ad-infinitum. This method works in principle but has severe 1926 * implementation issues due to RTT variances, timer granularity, and 1927 * instability in the algorithm which can lead to many false positives and 1928 * create oscillations as well as interact badly with other TCP streams 1929 * implementing the same algorithm. 1930 * 1931 * The second method is to limit the window to the bandwidth delay product 1932 * of the link. This is the method we implement. RTT variances and our 1933 * own manipulation of the congestion window, bwnd, can potentially 1934 * destabilize the algorithm. For this reason we have to stabilize the 1935 * elements used to calculate the window. We do this by using the minimum 1936 * observed RTT, the long term average of the observed bandwidth, and 1937 * by adding two segments worth of slop. It isn't perfect but it is able 1938 * to react to changing conditions and gives us a very stable basis on 1939 * which to extend the algorithm. 1940 */ 1941 void 1942 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1943 { 1944 u_long bw; 1945 u_long ibw; 1946 u_long bwnd; 1947 int save_ticks; 1948 int delta_ticks; 1949 1950 /* 1951 * If inflight_enable is disabled in the middle of a tcp connection, 1952 * make sure snd_bwnd is effectively disabled. 1953 */ 1954 if (!tcp_inflight_enable) { 1955 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1956 tp->snd_bandwidth = 0; 1957 return; 1958 } 1959 1960 /* 1961 * Validate the delta time. If a connection is new or has been idle 1962 * a long time we have to reset the bandwidth calculator. 1963 */ 1964 save_ticks = ticks; 1965 cpu_ccfence(); 1966 delta_ticks = save_ticks - tp->t_bw_rtttime; 1967 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1968 tp->t_bw_rtttime = save_ticks; 1969 tp->t_bw_rtseq = ack_seq; 1970 if (tp->snd_bandwidth == 0) 1971 tp->snd_bandwidth = tcp_inflight_min; 1972 return; 1973 } 1974 1975 /* 1976 * A delta of at least 1 tick is required. Waiting 2 ticks will 1977 * result in better (bw) accuracy. More than that and the ramp-up 1978 * will be too slow. 1979 */ 1980 if (delta_ticks == 0 || delta_ticks == 1) 1981 return; 1982 1983 /* 1984 * Sanity check, plus ignore pure window update acks. 1985 */ 1986 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1987 return; 1988 1989 /* 1990 * Figure out the bandwidth. Due to the tick granularity this 1991 * is a very rough number and it MUST be averaged over a fairly 1992 * long period of time. XXX we need to take into account a link 1993 * that is not using all available bandwidth, but for now our 1994 * slop will ramp us up if this case occurs and the bandwidth later 1995 * increases. 1996 */ 1997 ibw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 1998 tp->t_bw_rtttime = save_ticks; 1999 tp->t_bw_rtseq = ack_seq; 2000 bw = ((int64_t)tp->snd_bandwidth * 15 + ibw) >> 4; 2001 2002 tp->snd_bandwidth = bw; 2003 2004 /* 2005 * Calculate the semi-static bandwidth delay product, plus two maximal 2006 * segments. The additional slop puts us squarely in the sweet 2007 * spot and also handles the bandwidth run-up case. Without the 2008 * slop we could be locking ourselves into a lower bandwidth. 2009 * 2010 * At very high speeds the bw calculation can become overly sensitive 2011 * and error prone when delta_ticks is low (e.g. usually 1). To deal 2012 * with the problem the stab must be scaled to the bw. A stab of 50 2013 * (the default) increases the bw for the purposes of the bwnd 2014 * calculation by 5%. 2015 * 2016 * Situations Handled: 2017 * (1) Prevents over-queueing of packets on LANs, especially on 2018 * high speed LANs, allowing larger TCP buffers to be 2019 * specified, and also does a good job preventing 2020 * over-queueing of packets over choke points like modems 2021 * (at least for the transmit side). 2022 * 2023 * (2) Is able to handle changing network loads (bandwidth 2024 * drops so bwnd drops, bandwidth increases so bwnd 2025 * increases). 2026 * 2027 * (3) Theoretically should stabilize in the face of multiple 2028 * connections implementing the same algorithm (this may need 2029 * a little work). 2030 * 2031 * (4) Stability value (defaults to 20 = 2 maximal packets) can 2032 * be adjusted with a sysctl but typically only needs to be on 2033 * very slow connections. A value no smaller then 5 should 2034 * be used, but only reduce this default if you have no other 2035 * choice. 2036 */ 2037 2038 #define USERTT ((tp->t_srtt + tp->t_rttvar) + tcp_inflight_adjrtt) 2039 bw += bw * tcp_inflight_stab / 1000; 2040 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 2041 (int)tp->t_maxseg * 2; 2042 #undef USERTT 2043 2044 if (tcp_inflight_debug > 0) { 2045 static int ltime; 2046 if ((u_int)(save_ticks - ltime) >= hz / tcp_inflight_debug) { 2047 ltime = save_ticks; 2048 kprintf("%p ibw %ld bw %ld rttvar %d srtt %d " 2049 "bwnd %ld delta %d snd_win %ld\n", 2050 tp, ibw, bw, tp->t_rttvar, tp->t_srtt, 2051 bwnd, delta_ticks, tp->snd_wnd); 2052 } 2053 } 2054 if ((long)bwnd < tcp_inflight_min) 2055 bwnd = tcp_inflight_min; 2056 if (bwnd > tcp_inflight_max) 2057 bwnd = tcp_inflight_max; 2058 if ((long)bwnd < tp->t_maxseg * 2) 2059 bwnd = tp->t_maxseg * 2; 2060 tp->snd_bwnd = bwnd; 2061 } 2062 2063 static void 2064 tcp_rmx_iwsegs(struct tcpcb *tp, u_long *maxsegs, u_long *capsegs) 2065 { 2066 struct rtentry *rt; 2067 struct inpcb *inp = tp->t_inpcb; 2068 #ifdef INET6 2069 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE); 2070 #else 2071 const boolean_t isipv6 = FALSE; 2072 #endif 2073 2074 /* XXX */ 2075 if (tcp_iw_maxsegs < TCP_IW_MAXSEGS_DFLT) 2076 tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 2077 if (tcp_iw_capsegs < TCP_IW_CAPSEGS_DFLT) 2078 tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 2079 2080 if (isipv6) 2081 rt = tcp_rtlookup6(&inp->inp_inc); 2082 else 2083 rt = tcp_rtlookup(&inp->inp_inc); 2084 if (rt == NULL || 2085 rt->rt_rmx.rmx_iwmaxsegs < TCP_IW_MAXSEGS_DFLT || 2086 rt->rt_rmx.rmx_iwcapsegs < TCP_IW_CAPSEGS_DFLT) { 2087 *maxsegs = tcp_iw_maxsegs; 2088 *capsegs = tcp_iw_capsegs; 2089 return; 2090 } 2091 *maxsegs = rt->rt_rmx.rmx_iwmaxsegs; 2092 *capsegs = rt->rt_rmx.rmx_iwcapsegs; 2093 } 2094 2095 u_long 2096 tcp_initial_window(struct tcpcb *tp) 2097 { 2098 if (tcp_do_rfc3390) { 2099 /* 2100 * RFC3390: 2101 * "If the SYN or SYN/ACK is lost, the initial window 2102 * used by a sender after a correctly transmitted SYN 2103 * MUST be one segment consisting of MSS bytes." 2104 * 2105 * However, we do something a little bit more aggressive 2106 * then RFC3390 here: 2107 * - Only if time spent in the SYN or SYN|ACK retransmition 2108 * >= 3 seconds, the IW is reduced. We do this mainly 2109 * because when RFC3390 is published, the initial RTO is 2110 * still 3 seconds (the threshold we test here), while 2111 * after RFC6298, the initial RTO is 1 second. This 2112 * behaviour probably still falls within the spirit of 2113 * RFC3390. 2114 * - When IW is reduced, 2*MSS is used instead of 1*MSS. 2115 * Mainly to avoid sender and receiver deadlock until 2116 * delayed ACK timer expires. And even RFC2581 does not 2117 * try to reduce IW upon SYN or SYN|ACK retransmition 2118 * timeout. 2119 * 2120 * See also: 2121 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03 2122 */ 2123 if (tp->t_rxtsyn >= TCPTV_RTOBASE3) { 2124 return (2 * tp->t_maxseg); 2125 } else { 2126 u_long maxsegs, capsegs; 2127 2128 tcp_rmx_iwsegs(tp, &maxsegs, &capsegs); 2129 return min(maxsegs * tp->t_maxseg, 2130 max(2 * tp->t_maxseg, capsegs * 1460)); 2131 } 2132 } else { 2133 /* 2134 * Even RFC2581 (back to 1999) allows 2*SMSS IW. 2135 * 2136 * Mainly to avoid sender and receiver deadlock 2137 * until delayed ACK timer expires. 2138 */ 2139 return (2 * tp->t_maxseg); 2140 } 2141 } 2142 2143 #ifdef TCP_SIGNATURE 2144 /* 2145 * Compute TCP-MD5 hash of a TCP segment. (RFC2385) 2146 * 2147 * We do this over ip, tcphdr, segment data, and the key in the SADB. 2148 * When called from tcp_input(), we can be sure that th_sum has been 2149 * zeroed out and verified already. 2150 * 2151 * Return 0 if successful, otherwise return -1. 2152 * 2153 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a 2154 * search with the destination IP address, and a 'magic SPI' to be 2155 * determined by the application. This is hardcoded elsewhere to 1179 2156 * right now. Another branch of this code exists which uses the SPD to 2157 * specify per-application flows but it is unstable. 2158 */ 2159 int 2160 tcpsignature_compute( 2161 struct mbuf *m, /* mbuf chain */ 2162 int len, /* length of TCP data */ 2163 int optlen, /* length of TCP options */ 2164 u_char *buf, /* storage for MD5 digest */ 2165 u_int direction) /* direction of flow */ 2166 { 2167 struct ippseudo ippseudo; 2168 MD5_CTX ctx; 2169 int doff; 2170 struct ip *ip; 2171 struct ipovly *ipovly; 2172 struct secasvar *sav; 2173 struct tcphdr *th; 2174 #ifdef INET6 2175 struct ip6_hdr *ip6; 2176 struct in6_addr in6; 2177 uint32_t plen; 2178 uint16_t nhdr; 2179 #endif /* INET6 */ 2180 u_short savecsum; 2181 2182 KASSERT(m != NULL, ("passed NULL mbuf. Game over.")); 2183 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature")); 2184 /* 2185 * Extract the destination from the IP header in the mbuf. 2186 */ 2187 ip = mtod(m, struct ip *); 2188 #ifdef INET6 2189 ip6 = NULL; /* Make the compiler happy. */ 2190 #endif /* INET6 */ 2191 /* 2192 * Look up an SADB entry which matches the address found in 2193 * the segment. 2194 */ 2195 switch (IP_VHL_V(ip->ip_vhl)) { 2196 case IPVERSION: 2197 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst, 2198 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2199 break; 2200 #ifdef INET6 2201 case (IPV6_VERSION >> 4): 2202 ip6 = mtod(m, struct ip6_hdr *); 2203 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst, 2204 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2205 break; 2206 #endif /* INET6 */ 2207 default: 2208 return (EINVAL); 2209 /* NOTREACHED */ 2210 break; 2211 } 2212 if (sav == NULL) { 2213 kprintf("%s: SADB lookup failed\n", __func__); 2214 return (EINVAL); 2215 } 2216 MD5Init(&ctx); 2217 2218 /* 2219 * Step 1: Update MD5 hash with IP pseudo-header. 2220 * 2221 * XXX The ippseudo header MUST be digested in network byte order, 2222 * or else we'll fail the regression test. Assume all fields we've 2223 * been doing arithmetic on have been in host byte order. 2224 * XXX One cannot depend on ipovly->ih_len here. When called from 2225 * tcp_output(), the underlying ip_len member has not yet been set. 2226 */ 2227 switch (IP_VHL_V(ip->ip_vhl)) { 2228 case IPVERSION: 2229 ipovly = (struct ipovly *)ip; 2230 ippseudo.ippseudo_src = ipovly->ih_src; 2231 ippseudo.ippseudo_dst = ipovly->ih_dst; 2232 ippseudo.ippseudo_pad = 0; 2233 ippseudo.ippseudo_p = IPPROTO_TCP; 2234 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen); 2235 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo)); 2236 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip)); 2237 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen; 2238 break; 2239 #ifdef INET6 2240 /* 2241 * RFC 2385, 2.0 Proposal 2242 * For IPv6, the pseudo-header is as described in RFC 2460, namely the 2243 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero- 2244 * extended next header value (to form 32 bits), and 32-bit segment 2245 * length. 2246 * Note: Upper-Layer Packet Length comes before Next Header. 2247 */ 2248 case (IPV6_VERSION >> 4): 2249 in6 = ip6->ip6_src; 2250 in6_clearscope(&in6); 2251 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2252 in6 = ip6->ip6_dst; 2253 in6_clearscope(&in6); 2254 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2255 plen = htonl(len + sizeof(struct tcphdr) + optlen); 2256 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t)); 2257 nhdr = 0; 2258 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2259 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2260 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2261 nhdr = IPPROTO_TCP; 2262 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2263 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr)); 2264 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen; 2265 break; 2266 #endif /* INET6 */ 2267 default: 2268 return (EINVAL); 2269 /* NOTREACHED */ 2270 break; 2271 } 2272 /* 2273 * Step 2: Update MD5 hash with TCP header, excluding options. 2274 * The TCP checksum must be set to zero. 2275 */ 2276 savecsum = th->th_sum; 2277 th->th_sum = 0; 2278 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr)); 2279 th->th_sum = savecsum; 2280 /* 2281 * Step 3: Update MD5 hash with TCP segment data. 2282 * Use m_apply() to avoid an early m_pullup(). 2283 */ 2284 if (len > 0) 2285 m_apply(m, doff, len, tcpsignature_apply, &ctx); 2286 /* 2287 * Step 4: Update MD5 hash with shared secret. 2288 */ 2289 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth)); 2290 MD5Final(buf, &ctx); 2291 key_sa_recordxfer(sav, m); 2292 key_freesav(sav); 2293 return (0); 2294 } 2295 2296 int 2297 tcpsignature_apply(void *fstate, void *data, unsigned int len) 2298 { 2299 2300 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len); 2301 return (0); 2302 } 2303 #endif /* TCP_SIGNATURE */ 2304