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 4 160 161 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_ISIPV6(inp)) { 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(M_NOWAIT, 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_ISIPV6(inp); 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_ISIPV6(inp); 839 #else 840 const boolean_t isipv6 = FALSE; 841 #endif 842 843 if (tp->t_flags & TF_LISTEN) { 844 /* 845 * Pending socket/syncache inheritance 846 * 847 * If this is a listen(2) socket, find another listen(2) 848 * socket in the same local group, which could inherit 849 * the syncache and sockets pending on the completion 850 * and incompletion queues. 851 * 852 * NOTE: 853 * Currently the inheritance could only happen on the 854 * listen(2) sockets w/ SO_REUSEPORT set. 855 */ 856 KASSERT(&curthread->td_msgport == netisr_cpuport(0), 857 ("listen socket close not in netisr0")); 858 inp_inh = in_pcblocalgroup_last(&tcbinfo[0], inp); 859 if (inp_inh != NULL) 860 tp_inh = intotcpcb(inp_inh); 861 } 862 863 /* 864 * INP_WILDCARD indicates that listen(2) has been called on 865 * this socket. This implies: 866 * - A wildcard inp's hash is replicated for each protocol thread. 867 * - Syncache for this inp grows independently in each protocol 868 * thread. 869 * - There is more than one cpu 870 * 871 * We have to chain a message to the rest of the protocol threads 872 * to cleanup the wildcard hash and the syncache. The cleanup 873 * in the current protocol thread is defered till the end of this 874 * function (syncache_destroy and in_pcbdetach). 875 * 876 * NOTE: 877 * After cleanup the inp's hash and syncache entries, this inp will 878 * no longer be available to the rest of the protocol threads, so we 879 * are safe to whack the inp in the following code. 880 */ 881 if ((inp->inp_flags & INP_WILDCARD) && ncpus2 > 1) { 882 struct netmsg_listen_detach nmsg; 883 884 KKASSERT(so->so_port == netisr_cpuport(0)); 885 KKASSERT(&curthread->td_msgport == netisr_cpuport(0)); 886 KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]); 887 888 netmsg_init(&nmsg.base, NULL, &curthread->td_msgport, 889 MSGF_PRIORITY, tcp_listen_detach_handler); 890 nmsg.nm_tp = tp; 891 nmsg.nm_tp_inh = tp_inh; 892 lwkt_domsg(netisr_cpuport(1), &nmsg.base.lmsg, 0); 893 } 894 895 KKASSERT(tp->t_state != TCPS_TERMINATING); 896 tp->t_state = TCPS_TERMINATING; 897 898 /* 899 * Make sure that all of our timers are stopped before we 900 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL), 901 * timers are never used. If timer message is never created 902 * (tp->tt_msg->tt_tcb == NULL), timers are never used too. 903 */ 904 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) { 905 tcp_callout_stop(tp, tp->tt_rexmt); 906 tcp_callout_stop(tp, tp->tt_persist); 907 tcp_callout_stop(tp, tp->tt_keep); 908 tcp_callout_stop(tp, tp->tt_2msl); 909 tcp_callout_stop(tp, tp->tt_delack); 910 } 911 912 if (tp->t_flags & TF_ONOUTPUTQ) { 913 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 914 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 915 tp->t_flags &= ~TF_ONOUTPUTQ; 916 } 917 918 /* 919 * If we got enough samples through the srtt filter, 920 * save the rtt and rttvar in the routing entry. 921 * 'Enough' is arbitrarily defined as the 16 samples. 922 * 16 samples is enough for the srtt filter to converge 923 * to within 5% of the correct value; fewer samples and 924 * we could save a very bogus rtt. 925 * 926 * Don't update the default route's characteristics and don't 927 * update anything that the user "locked". 928 */ 929 if (tp->t_rttupdated >= 16) { 930 u_long i = 0; 931 932 if (isipv6) { 933 struct sockaddr_in6 *sin6; 934 935 if ((rt = inp->in6p_route.ro_rt) == NULL) 936 goto no_valid_rt; 937 sin6 = (struct sockaddr_in6 *)rt_key(rt); 938 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 939 goto no_valid_rt; 940 } else 941 if ((rt = inp->inp_route.ro_rt) == NULL || 942 ((struct sockaddr_in *)rt_key(rt))-> 943 sin_addr.s_addr == INADDR_ANY) 944 goto no_valid_rt; 945 946 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 947 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 948 if (rt->rt_rmx.rmx_rtt && i) 949 /* 950 * filter this update to half the old & half 951 * the new values, converting scale. 952 * See route.h and tcp_var.h for a 953 * description of the scaling constants. 954 */ 955 rt->rt_rmx.rmx_rtt = 956 (rt->rt_rmx.rmx_rtt + i) / 2; 957 else 958 rt->rt_rmx.rmx_rtt = i; 959 tcpstat.tcps_cachedrtt++; 960 } 961 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 962 i = tp->t_rttvar * 963 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 964 if (rt->rt_rmx.rmx_rttvar && i) 965 rt->rt_rmx.rmx_rttvar = 966 (rt->rt_rmx.rmx_rttvar + i) / 2; 967 else 968 rt->rt_rmx.rmx_rttvar = i; 969 tcpstat.tcps_cachedrttvar++; 970 } 971 /* 972 * The old comment here said: 973 * update the pipelimit (ssthresh) if it has been updated 974 * already or if a pipesize was specified & the threshhold 975 * got below half the pipesize. I.e., wait for bad news 976 * before we start updating, then update on both good 977 * and bad news. 978 * 979 * But we want to save the ssthresh even if no pipesize is 980 * specified explicitly in the route, because such 981 * connections still have an implicit pipesize specified 982 * by the global tcp_sendspace. In the absence of a reliable 983 * way to calculate the pipesize, it will have to do. 984 */ 985 i = tp->snd_ssthresh; 986 if (rt->rt_rmx.rmx_sendpipe != 0) 987 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 988 else 989 dosavessthresh = (i < so->so_snd.ssb_hiwat/2); 990 if (dosavessthresh || 991 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 992 (rt->rt_rmx.rmx_ssthresh != 0))) { 993 /* 994 * convert the limit from user data bytes to 995 * packets then to packet data bytes. 996 */ 997 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 998 if (i < 2) 999 i = 2; 1000 i *= tp->t_maxseg + 1001 (isipv6 ? 1002 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1003 sizeof(struct tcpiphdr)); 1004 if (rt->rt_rmx.rmx_ssthresh) 1005 rt->rt_rmx.rmx_ssthresh = 1006 (rt->rt_rmx.rmx_ssthresh + i) / 2; 1007 else 1008 rt->rt_rmx.rmx_ssthresh = i; 1009 tcpstat.tcps_cachedssthresh++; 1010 } 1011 } 1012 1013 no_valid_rt: 1014 /* free the reassembly queue, if any */ 1015 while((q = TAILQ_FIRST(&tp->t_segq)) != NULL) { 1016 TAILQ_REMOVE(&tp->t_segq, q, tqe_q); 1017 m_freem(q->tqe_m); 1018 kfree(q, M_TSEGQ); 1019 atomic_add_int(&tcp_reass_qsize, -1); 1020 } 1021 /* throw away SACK blocks in scoreboard*/ 1022 if (TCP_DO_SACK(tp)) 1023 tcp_sack_destroy(&tp->scb); 1024 1025 inp->inp_ppcb = NULL; 1026 soisdisconnected(so); 1027 /* note: pcb detached later on */ 1028 1029 tcp_destroy_timermsg(tp); 1030 tcp_output_cancel(tp); 1031 1032 if (tp->t_flags & TF_LISTEN) { 1033 syncache_destroy(tp, tp_inh); 1034 if (inp_inh != NULL && inp_inh->inp_socket != NULL) { 1035 /* 1036 * Pending sockets inheritance only needs 1037 * to be done once in the current thread, 1038 * i.e. netisr0. 1039 */ 1040 soinherit(so, inp_inh->inp_socket); 1041 } 1042 } 1043 1044 so_async_rcvd_drop(so); 1045 /* Drop the reference for the asynchronized pru_rcvd */ 1046 sofree(so); 1047 1048 /* 1049 * NOTE: 1050 * pcbdetach removes any wildcard hash entry on the current CPU. 1051 */ 1052 #ifdef INET6 1053 if (isipv6) 1054 in6_pcbdetach(inp); 1055 else 1056 #endif 1057 in_pcbdetach(inp); 1058 1059 tcpstat.tcps_closed++; 1060 return (NULL); 1061 } 1062 1063 static __inline void 1064 tcp_drain_oncpu(struct inpcbinfo *pcbinfo) 1065 { 1066 struct inpcbhead *head = &pcbinfo->pcblisthead; 1067 struct inpcb *inpb; 1068 1069 /* 1070 * Since we run in netisr, it is MP safe, even if 1071 * we block during the inpcb list iteration, i.e. 1072 * we don't need to use inpcb marker here. 1073 */ 1074 KASSERT(&curthread->td_msgport == netisr_cpuport(pcbinfo->cpu), 1075 ("not in correct netisr")); 1076 1077 LIST_FOREACH(inpb, head, inp_list) { 1078 struct tcpcb *tcpb; 1079 struct tseg_qent *te; 1080 1081 if ((inpb->inp_flags & INP_PLACEMARKER) == 0 && 1082 (tcpb = intotcpcb(inpb)) != NULL && 1083 (te = TAILQ_FIRST(&tcpb->t_segq)) != NULL) { 1084 TAILQ_REMOVE(&tcpb->t_segq, te, tqe_q); 1085 if (te->tqe_th->th_flags & TH_FIN) 1086 tcpb->t_flags &= ~TF_QUEDFIN; 1087 m_freem(te->tqe_m); 1088 kfree(te, M_TSEGQ); 1089 atomic_add_int(&tcp_reass_qsize, -1); 1090 /* retry */ 1091 } 1092 } 1093 } 1094 1095 static void 1096 tcp_drain_dispatch(netmsg_t nmsg) 1097 { 1098 crit_enter(); 1099 lwkt_replymsg(&nmsg->lmsg, 0); /* reply ASAP */ 1100 crit_exit(); 1101 1102 tcp_drain_oncpu(&tcbinfo[mycpuid]); 1103 } 1104 1105 static void 1106 tcp_drain_ipi(void *arg __unused) 1107 { 1108 int cpu = mycpuid; 1109 struct lwkt_msg *msg = &tcp_drain_netmsg[cpu].lmsg; 1110 1111 crit_enter(); 1112 if (msg->ms_flags & MSGF_DONE) 1113 lwkt_sendmsg_oncpu(netisr_cpuport(cpu), msg); 1114 crit_exit(); 1115 } 1116 1117 void 1118 tcp_drain(void) 1119 { 1120 cpumask_t mask; 1121 1122 if (!do_tcpdrain) 1123 return; 1124 1125 /* 1126 * Walk the tcpbs, if existing, and flush the reassembly queue, 1127 * if there is one... 1128 * XXX: The "Net/3" implementation doesn't imply that the TCP 1129 * reassembly queue should be flushed, but in a situation 1130 * where we're really low on mbufs, this is potentially 1131 * useful. 1132 * YYY: We may consider run tcp_drain_oncpu directly here, 1133 * however, that will require M_WAITOK memory allocation 1134 * for the inpcb marker. 1135 */ 1136 CPUMASK_ASSBMASK(mask, ncpus2); 1137 CPUMASK_ANDMASK(mask, smp_active_mask); 1138 if (CPUMASK_TESTNZERO(mask)) 1139 lwkt_send_ipiq_mask(mask, tcp_drain_ipi, NULL); 1140 } 1141 1142 /* 1143 * Notify a tcp user of an asynchronous error; 1144 * store error as soft error, but wake up user 1145 * (for now, won't do anything until can select for soft error). 1146 * 1147 * Do not wake up user since there currently is no mechanism for 1148 * reporting soft errors (yet - a kqueue filter may be added). 1149 */ 1150 static void 1151 tcp_notify(struct inpcb *inp, int error) 1152 { 1153 struct tcpcb *tp = intotcpcb(inp); 1154 1155 /* 1156 * Ignore some errors if we are hooked up. 1157 * If connection hasn't completed, has retransmitted several times, 1158 * and receives a second error, give up now. This is better 1159 * than waiting a long time to establish a connection that 1160 * can never complete. 1161 */ 1162 if (tp->t_state == TCPS_ESTABLISHED && 1163 (error == EHOSTUNREACH || error == ENETUNREACH || 1164 error == EHOSTDOWN)) { 1165 return; 1166 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1167 tp->t_softerror) 1168 tcp_drop(tp, error); 1169 else 1170 tp->t_softerror = error; 1171 #if 0 1172 wakeup(&so->so_timeo); 1173 sorwakeup(so); 1174 sowwakeup(so); 1175 #endif 1176 } 1177 1178 static int 1179 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1180 { 1181 int error, i, n; 1182 struct inpcb *marker; 1183 struct inpcb *inp; 1184 int origcpu, ccpu; 1185 1186 error = 0; 1187 n = 0; 1188 1189 /* 1190 * The process of preparing the TCB list is too time-consuming and 1191 * resource-intensive to repeat twice on every request. 1192 */ 1193 if (req->oldptr == NULL) { 1194 for (ccpu = 0; ccpu < ncpus2; ++ccpu) 1195 n += tcbinfo[ccpu].ipi_count; 1196 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1197 return (0); 1198 } 1199 1200 if (req->newptr != NULL) 1201 return (EPERM); 1202 1203 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1204 marker->inp_flags |= INP_PLACEMARKER; 1205 1206 /* 1207 * OK, now we're committed to doing something. Run the inpcb list 1208 * for each cpu in the system and construct the output. Use a 1209 * list placemarker to deal with list changes occuring during 1210 * copyout blockages (but otherwise depend on being on the correct 1211 * cpu to avoid races). 1212 */ 1213 origcpu = mycpu->gd_cpuid; 1214 for (ccpu = 0; ccpu < ncpus2 && error == 0; ++ccpu) { 1215 caddr_t inp_ppcb; 1216 struct xtcpcb xt; 1217 1218 lwkt_migratecpu(ccpu); 1219 1220 n = tcbinfo[ccpu].ipi_count; 1221 1222 LIST_INSERT_HEAD(&tcbinfo[ccpu].pcblisthead, marker, inp_list); 1223 i = 0; 1224 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1225 /* 1226 * process a snapshot of pcbs, ignoring placemarkers 1227 * and using our own to allow SYSCTL_OUT to block. 1228 */ 1229 LIST_REMOVE(marker, inp_list); 1230 LIST_INSERT_AFTER(inp, marker, inp_list); 1231 1232 if (inp->inp_flags & INP_PLACEMARKER) 1233 continue; 1234 if (prison_xinpcb(req->td, inp)) 1235 continue; 1236 1237 xt.xt_len = sizeof xt; 1238 bcopy(inp, &xt.xt_inp, sizeof *inp); 1239 inp_ppcb = inp->inp_ppcb; 1240 if (inp_ppcb != NULL) 1241 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1242 else 1243 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1244 if (inp->inp_socket) 1245 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1246 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1247 break; 1248 ++i; 1249 } 1250 LIST_REMOVE(marker, inp_list); 1251 if (error == 0 && i < n) { 1252 bzero(&xt, sizeof xt); 1253 xt.xt_len = sizeof xt; 1254 while (i < n) { 1255 error = SYSCTL_OUT(req, &xt, sizeof xt); 1256 if (error) 1257 break; 1258 ++i; 1259 } 1260 } 1261 } 1262 1263 /* 1264 * Make sure we are on the same cpu we were on originally, since 1265 * higher level callers expect this. Also don't pollute caches with 1266 * migrated userland data by (eventually) returning to userland 1267 * on a different cpu. 1268 */ 1269 lwkt_migratecpu(origcpu); 1270 kfree(marker, M_TEMP); 1271 return (error); 1272 } 1273 1274 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1275 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1276 1277 static int 1278 tcp_getcred(SYSCTL_HANDLER_ARGS) 1279 { 1280 struct sockaddr_in addrs[2]; 1281 struct ucred cred0, *cred = NULL; 1282 struct inpcb *inp; 1283 int cpu, origcpu, error; 1284 1285 error = priv_check(req->td, PRIV_ROOT); 1286 if (error != 0) 1287 return (error); 1288 error = SYSCTL_IN(req, addrs, sizeof addrs); 1289 if (error != 0) 1290 return (error); 1291 1292 origcpu = mycpuid; 1293 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1294 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1295 1296 lwkt_migratecpu(cpu); 1297 1298 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1299 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1300 if (inp == NULL || inp->inp_socket == NULL) { 1301 error = ENOENT; 1302 } else if (inp->inp_socket->so_cred != NULL) { 1303 cred0 = *(inp->inp_socket->so_cred); 1304 cred = &cred0; 1305 } 1306 1307 lwkt_migratecpu(origcpu); 1308 1309 if (error) 1310 return (error); 1311 1312 return SYSCTL_OUT(req, cred, sizeof(struct ucred)); 1313 } 1314 1315 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1316 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1317 1318 #ifdef INET6 1319 static int 1320 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1321 { 1322 struct sockaddr_in6 addrs[2]; 1323 struct inpcb *inp; 1324 int error; 1325 1326 error = priv_check(req->td, PRIV_ROOT); 1327 if (error != 0) 1328 return (error); 1329 error = SYSCTL_IN(req, addrs, sizeof addrs); 1330 if (error != 0) 1331 return (error); 1332 crit_enter(); 1333 inp = in6_pcblookup_hash(&tcbinfo[0], 1334 &addrs[1].sin6_addr, addrs[1].sin6_port, 1335 &addrs[0].sin6_addr, addrs[0].sin6_port, 0, NULL); 1336 if (inp == NULL || inp->inp_socket == NULL) { 1337 error = ENOENT; 1338 goto out; 1339 } 1340 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1341 out: 1342 crit_exit(); 1343 return (error); 1344 } 1345 1346 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1347 0, 0, 1348 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1349 #endif 1350 1351 struct netmsg_tcp_notify { 1352 struct netmsg_base base; 1353 inp_notify_t nm_notify; 1354 struct in_addr nm_faddr; 1355 int nm_arg; 1356 }; 1357 1358 static void 1359 tcp_notifyall_oncpu(netmsg_t msg) 1360 { 1361 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg; 1362 int nextcpu; 1363 1364 in_pcbnotifyall(&tcbinfo[mycpuid], nm->nm_faddr, 1365 nm->nm_arg, nm->nm_notify); 1366 1367 nextcpu = mycpuid + 1; 1368 if (nextcpu < ncpus2) 1369 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nm->base.lmsg); 1370 else 1371 lwkt_replymsg(&nm->base.lmsg, 0); 1372 } 1373 1374 inp_notify_t 1375 tcp_get_inpnotify(int cmd, const struct sockaddr *sa, 1376 int *arg, struct ip **ip0, int *cpuid) 1377 { 1378 struct ip *ip = *ip0; 1379 struct in_addr faddr; 1380 inp_notify_t notify = tcp_notify; 1381 1382 faddr = ((const struct sockaddr_in *)sa)->sin_addr; 1383 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1384 return NULL; 1385 1386 *arg = inetctlerrmap[cmd]; 1387 if (cmd == PRC_QUENCH) { 1388 notify = tcp_quench; 1389 } else if (icmp_may_rst && 1390 (cmd == PRC_UNREACH_ADMIN_PROHIB || 1391 cmd == PRC_UNREACH_PORT || 1392 cmd == PRC_TIMXCEED_INTRANS) && 1393 ip != NULL) { 1394 notify = tcp_drop_syn_sent; 1395 } else if (cmd == PRC_MSGSIZE) { 1396 const struct icmp *icmp = (const struct icmp *) 1397 ((caddr_t)ip - offsetof(struct icmp, icmp_ip)); 1398 1399 *arg = ntohs(icmp->icmp_nextmtu); 1400 notify = tcp_mtudisc; 1401 } else if (PRC_IS_REDIRECT(cmd)) { 1402 ip = NULL; 1403 notify = in_rtchange; 1404 } else if (cmd == PRC_HOSTDEAD) { 1405 ip = NULL; 1406 } else if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) { 1407 return NULL; 1408 } 1409 1410 if (cpuid != NULL) { 1411 if (ip == NULL) { 1412 /* Go through all CPUs */ 1413 *cpuid = ncpus; 1414 } else { 1415 const struct tcphdr *th; 1416 1417 th = (const struct tcphdr *) 1418 ((caddr_t)ip + (IP_VHL_HL(ip->ip_vhl) << 2)); 1419 *cpuid = tcp_addrcpu(faddr.s_addr, th->th_dport, 1420 ip->ip_src.s_addr, th->th_sport); 1421 } 1422 } 1423 1424 *ip0 = ip; 1425 return notify; 1426 } 1427 1428 void 1429 tcp_ctlinput(netmsg_t msg) 1430 { 1431 int cmd = msg->ctlinput.nm_cmd; 1432 struct sockaddr *sa = msg->ctlinput.nm_arg; 1433 struct ip *ip = msg->ctlinput.nm_extra; 1434 struct in_addr faddr; 1435 inp_notify_t notify; 1436 int arg, cpuid; 1437 1438 notify = tcp_get_inpnotify(cmd, sa, &arg, &ip, &cpuid); 1439 if (notify == NULL) 1440 goto done; 1441 1442 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1443 if (ip != NULL) { 1444 const struct tcphdr *th; 1445 struct inpcb *inp; 1446 1447 if (cpuid != mycpuid) 1448 goto done; 1449 1450 th = (const struct tcphdr *) 1451 ((caddr_t)ip + (IP_VHL_HL(ip->ip_vhl) << 2)); 1452 inp = in_pcblookup_hash(&tcbinfo[mycpuid], faddr, th->th_dport, 1453 ip->ip_src, th->th_sport, 0, NULL); 1454 if (inp != NULL && inp->inp_socket != NULL) { 1455 tcp_seq icmpseq = htonl(th->th_seq); 1456 struct tcpcb *tp = intotcpcb(inp); 1457 1458 if (SEQ_GEQ(icmpseq, tp->snd_una) && 1459 SEQ_LT(icmpseq, tp->snd_max)) 1460 notify(inp, arg); 1461 } else { 1462 struct in_conninfo inc; 1463 1464 inc.inc_fport = th->th_dport; 1465 inc.inc_lport = th->th_sport; 1466 inc.inc_faddr = faddr; 1467 inc.inc_laddr = ip->ip_src; 1468 #ifdef INET6 1469 inc.inc_isipv6 = 0; 1470 #endif 1471 syncache_unreach(&inc, th); 1472 } 1473 } else if (msg->ctlinput.nm_direct) { 1474 if (cpuid != ncpus && cpuid != mycpuid) 1475 goto done; 1476 if (mycpuid >= ncpus2) 1477 goto done; 1478 1479 in_pcbnotifyall(&tcbinfo[mycpuid], faddr, arg, notify); 1480 } else { 1481 struct netmsg_tcp_notify *nm; 1482 1483 KKASSERT(&curthread->td_msgport == netisr_cpuport(0)); 1484 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT); 1485 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1486 0, tcp_notifyall_oncpu); 1487 nm->nm_faddr = faddr; 1488 nm->nm_arg = arg; 1489 nm->nm_notify = notify; 1490 1491 lwkt_sendmsg(netisr_cpuport(0), &nm->base.lmsg); 1492 } 1493 done: 1494 lwkt_replymsg(&msg->lmsg, 0); 1495 } 1496 1497 #ifdef INET6 1498 1499 void 1500 tcp6_ctlinput(netmsg_t msg) 1501 { 1502 int cmd = msg->ctlinput.nm_cmd; 1503 struct sockaddr *sa = msg->ctlinput.nm_arg; 1504 void *d = msg->ctlinput.nm_extra; 1505 struct tcphdr th; 1506 inp_notify_t notify = tcp_notify; 1507 struct ip6_hdr *ip6; 1508 struct mbuf *m; 1509 struct ip6ctlparam *ip6cp = NULL; 1510 const struct sockaddr_in6 *sa6_src = NULL; 1511 int off; 1512 struct tcp_portonly { 1513 u_int16_t th_sport; 1514 u_int16_t th_dport; 1515 } *thp; 1516 int arg; 1517 1518 if (sa->sa_family != AF_INET6 || 1519 sa->sa_len != sizeof(struct sockaddr_in6)) { 1520 goto out; 1521 } 1522 1523 arg = 0; 1524 if (cmd == PRC_QUENCH) 1525 notify = tcp_quench; 1526 else if (cmd == PRC_MSGSIZE) { 1527 struct ip6ctlparam *ip6cp = d; 1528 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6; 1529 1530 arg = ntohl(icmp6->icmp6_mtu); 1531 notify = tcp_mtudisc; 1532 } else if (!PRC_IS_REDIRECT(cmd) && 1533 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) { 1534 goto out; 1535 } 1536 1537 /* if the parameter is from icmp6, decode it. */ 1538 if (d != NULL) { 1539 ip6cp = (struct ip6ctlparam *)d; 1540 m = ip6cp->ip6c_m; 1541 ip6 = ip6cp->ip6c_ip6; 1542 off = ip6cp->ip6c_off; 1543 sa6_src = ip6cp->ip6c_src; 1544 } else { 1545 m = NULL; 1546 ip6 = NULL; 1547 off = 0; /* fool gcc */ 1548 sa6_src = &sa6_any; 1549 } 1550 1551 if (ip6 != NULL) { 1552 struct in_conninfo inc; 1553 /* 1554 * XXX: We assume that when IPV6 is non NULL, 1555 * M and OFF are valid. 1556 */ 1557 1558 /* check if we can safely examine src and dst ports */ 1559 if (m->m_pkthdr.len < off + sizeof *thp) 1560 goto out; 1561 1562 bzero(&th, sizeof th); 1563 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1564 1565 in6_pcbnotify(&tcbinfo[0], sa, th.th_dport, 1566 (struct sockaddr *)ip6cp->ip6c_src, 1567 th.th_sport, cmd, arg, notify); 1568 1569 inc.inc_fport = th.th_dport; 1570 inc.inc_lport = th.th_sport; 1571 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1572 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1573 inc.inc_isipv6 = 1; 1574 syncache_unreach(&inc, &th); 1575 } else { 1576 in6_pcbnotify(&tcbinfo[0], sa, 0, 1577 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify); 1578 } 1579 out: 1580 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0); 1581 } 1582 1583 #endif 1584 1585 /* 1586 * Following is where TCP initial sequence number generation occurs. 1587 * 1588 * There are two places where we must use initial sequence numbers: 1589 * 1. In SYN-ACK packets. 1590 * 2. In SYN packets. 1591 * 1592 * All ISNs for SYN-ACK packets are generated by the syncache. See 1593 * tcp_syncache.c for details. 1594 * 1595 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1596 * depends on this property. In addition, these ISNs should be 1597 * unguessable so as to prevent connection hijacking. To satisfy 1598 * the requirements of this situation, the algorithm outlined in 1599 * RFC 1948 is used to generate sequence numbers. 1600 * 1601 * Implementation details: 1602 * 1603 * Time is based off the system timer, and is corrected so that it 1604 * increases by one megabyte per second. This allows for proper 1605 * recycling on high speed LANs while still leaving over an hour 1606 * before rollover. 1607 * 1608 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1609 * between seeding of isn_secret. This is normally set to zero, 1610 * as reseeding should not be necessary. 1611 * 1612 */ 1613 1614 #define ISN_BYTES_PER_SECOND 1048576 1615 1616 u_char isn_secret[32]; 1617 int isn_last_reseed; 1618 MD5_CTX isn_ctx; 1619 1620 tcp_seq 1621 tcp_new_isn(struct tcpcb *tp) 1622 { 1623 u_int32_t md5_buffer[4]; 1624 tcp_seq new_isn; 1625 1626 /* Seed if this is the first use, reseed if requested. */ 1627 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1628 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1629 < (u_int)ticks))) { 1630 read_random_unlimited(&isn_secret, sizeof isn_secret); 1631 isn_last_reseed = ticks; 1632 } 1633 1634 /* Compute the md5 hash and return the ISN. */ 1635 MD5Init(&isn_ctx); 1636 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1637 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1638 #ifdef INET6 1639 if (INP_ISIPV6(tp->t_inpcb)) { 1640 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1641 sizeof(struct in6_addr)); 1642 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1643 sizeof(struct in6_addr)); 1644 } else 1645 #endif 1646 { 1647 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1648 sizeof(struct in_addr)); 1649 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1650 sizeof(struct in_addr)); 1651 } 1652 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1653 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1654 new_isn = (tcp_seq) md5_buffer[0]; 1655 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1656 return (new_isn); 1657 } 1658 1659 /* 1660 * When a source quench is received, close congestion window 1661 * to one segment. We will gradually open it again as we proceed. 1662 */ 1663 void 1664 tcp_quench(struct inpcb *inp, int error) 1665 { 1666 struct tcpcb *tp = intotcpcb(inp); 1667 1668 if (tp != NULL) { 1669 tp->snd_cwnd = tp->t_maxseg; 1670 tp->snd_wacked = 0; 1671 } 1672 } 1673 1674 /* 1675 * When a specific ICMP unreachable message is received and the 1676 * connection state is SYN-SENT, drop the connection. This behavior 1677 * is controlled by the icmp_may_rst sysctl. 1678 */ 1679 void 1680 tcp_drop_syn_sent(struct inpcb *inp, int error) 1681 { 1682 struct tcpcb *tp = intotcpcb(inp); 1683 1684 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT)) 1685 tcp_drop(tp, error); 1686 } 1687 1688 /* 1689 * When a `need fragmentation' ICMP is received, update our idea of the MSS 1690 * based on the new value in the route. Also nudge TCP to send something, 1691 * since we know the packet we just sent was dropped. 1692 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1693 */ 1694 void 1695 tcp_mtudisc(struct inpcb *inp, int mtu) 1696 { 1697 struct tcpcb *tp = intotcpcb(inp); 1698 struct rtentry *rt; 1699 struct socket *so = inp->inp_socket; 1700 int maxopd, mss; 1701 #ifdef INET6 1702 boolean_t isipv6 = INP_ISIPV6(inp); 1703 #else 1704 const boolean_t isipv6 = FALSE; 1705 #endif 1706 1707 if (tp == NULL) 1708 return; 1709 1710 /* 1711 * If no MTU is provided in the ICMP message, use the 1712 * next lower likely value, as specified in RFC 1191. 1713 */ 1714 if (mtu == 0) { 1715 int oldmtu; 1716 1717 oldmtu = tp->t_maxopd + 1718 (isipv6 ? 1719 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1720 sizeof(struct tcpiphdr)); 1721 mtu = ip_next_mtu(oldmtu, 0); 1722 } 1723 1724 if (isipv6) 1725 rt = tcp_rtlookup6(&inp->inp_inc); 1726 else 1727 rt = tcp_rtlookup(&inp->inp_inc); 1728 if (rt != NULL) { 1729 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu) 1730 mtu = rt->rt_rmx.rmx_mtu; 1731 1732 maxopd = mtu - 1733 (isipv6 ? 1734 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1735 sizeof(struct tcpiphdr)); 1736 1737 /* 1738 * XXX - The following conditional probably violates the TCP 1739 * spec. The problem is that, since we don't know the 1740 * other end's MSS, we are supposed to use a conservative 1741 * default. But, if we do that, then MTU discovery will 1742 * never actually take place, because the conservative 1743 * default is much less than the MTUs typically seen 1744 * on the Internet today. For the moment, we'll sweep 1745 * this under the carpet. 1746 * 1747 * The conservative default might not actually be a problem 1748 * if the only case this occurs is when sending an initial 1749 * SYN with options and data to a host we've never talked 1750 * to before. Then, they will reply with an MSS value which 1751 * will get recorded and the new parameters should get 1752 * recomputed. For Further Study. 1753 */ 1754 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd) 1755 maxopd = rt->rt_rmx.rmx_mssopt; 1756 } else 1757 maxopd = mtu - 1758 (isipv6 ? 1759 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1760 sizeof(struct tcpiphdr)); 1761 1762 if (tp->t_maxopd <= maxopd) 1763 return; 1764 tp->t_maxopd = maxopd; 1765 1766 mss = maxopd; 1767 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) == 1768 (TF_REQ_TSTMP | TF_RCVD_TSTMP)) 1769 mss -= TCPOLEN_TSTAMP_APPA; 1770 1771 /* round down to multiple of MCLBYTES */ 1772 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */ 1773 if (mss > MCLBYTES) 1774 mss &= ~(MCLBYTES - 1); 1775 #else 1776 if (mss > MCLBYTES) 1777 mss = (mss / MCLBYTES) * MCLBYTES; 1778 #endif 1779 1780 if (so->so_snd.ssb_hiwat < mss) 1781 mss = so->so_snd.ssb_hiwat; 1782 1783 tp->t_maxseg = mss; 1784 tp->t_rtttime = 0; 1785 tp->snd_nxt = tp->snd_una; 1786 tcp_output(tp); 1787 tcpstat.tcps_mturesent++; 1788 } 1789 1790 /* 1791 * Look-up the routing entry to the peer of this inpcb. If no route 1792 * is found and it cannot be allocated the return NULL. This routine 1793 * is called by TCP routines that access the rmx structure and by tcp_mss 1794 * to get the interface MTU. 1795 */ 1796 struct rtentry * 1797 tcp_rtlookup(struct in_conninfo *inc) 1798 { 1799 struct route *ro = &inc->inc_route; 1800 1801 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1802 /* No route yet, so try to acquire one */ 1803 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1804 /* 1805 * unused portions of the structure MUST be zero'd 1806 * out because rtalloc() treats it as opaque data 1807 */ 1808 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1809 ro->ro_dst.sa_family = AF_INET; 1810 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1811 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1812 inc->inc_faddr; 1813 rtalloc(ro); 1814 } 1815 } 1816 return (ro->ro_rt); 1817 } 1818 1819 #ifdef INET6 1820 struct rtentry * 1821 tcp_rtlookup6(struct in_conninfo *inc) 1822 { 1823 struct route_in6 *ro6 = &inc->inc6_route; 1824 1825 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1826 /* No route yet, so try to acquire one */ 1827 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1828 /* 1829 * unused portions of the structure MUST be zero'd 1830 * out because rtalloc() treats it as opaque data 1831 */ 1832 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1833 ro6->ro_dst.sin6_family = AF_INET6; 1834 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1835 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1836 rtalloc((struct route *)ro6); 1837 } 1838 } 1839 return (ro6->ro_rt); 1840 } 1841 #endif 1842 1843 #ifdef IPSEC 1844 /* compute ESP/AH header size for TCP, including outer IP header. */ 1845 size_t 1846 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1847 { 1848 struct inpcb *inp; 1849 struct mbuf *m; 1850 size_t hdrsiz; 1851 struct ip *ip; 1852 struct tcphdr *th; 1853 1854 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1855 return (0); 1856 MGETHDR(m, M_NOWAIT, MT_DATA); 1857 if (!m) 1858 return (0); 1859 1860 #ifdef INET6 1861 if (INP_ISIPV6(inp)) { 1862 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1863 1864 th = (struct tcphdr *)(ip6 + 1); 1865 m->m_pkthdr.len = m->m_len = 1866 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1867 tcp_fillheaders(tp, ip6, th, FALSE); 1868 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1869 } else 1870 #endif 1871 { 1872 ip = mtod(m, struct ip *); 1873 th = (struct tcphdr *)(ip + 1); 1874 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1875 tcp_fillheaders(tp, ip, th, FALSE); 1876 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1877 } 1878 1879 m_free(m); 1880 return (hdrsiz); 1881 } 1882 #endif 1883 1884 /* 1885 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1886 * 1887 * This code attempts to calculate the bandwidth-delay product as a 1888 * means of determining the optimal window size to maximize bandwidth, 1889 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1890 * routers. This code also does a fairly good job keeping RTTs in check 1891 * across slow links like modems. We implement an algorithm which is very 1892 * similar (but not meant to be) TCP/Vegas. The code operates on the 1893 * transmitter side of a TCP connection and so only effects the transmit 1894 * side of the connection. 1895 * 1896 * BACKGROUND: TCP makes no provision for the management of buffer space 1897 * at the end points or at the intermediate routers and switches. A TCP 1898 * stream, whether using NewReno or not, will eventually buffer as 1899 * many packets as it is able and the only reason this typically works is 1900 * due to the fairly small default buffers made available for a connection 1901 * (typicaly 16K or 32K). As machines use larger windows and/or window 1902 * scaling it is now fairly easy for even a single TCP connection to blow-out 1903 * all available buffer space not only on the local interface, but on 1904 * intermediate routers and switches as well. NewReno makes a misguided 1905 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1906 * then backing off, then steadily increasing the window again until another 1907 * failure occurs, ad-infinitum. This results in terrible oscillation that 1908 * is only made worse as network loads increase and the idea of intentionally 1909 * blowing out network buffers is, frankly, a terrible way to manage network 1910 * resources. 1911 * 1912 * It is far better to limit the transmit window prior to the failure 1913 * condition being achieved. There are two general ways to do this: First 1914 * you can 'scan' through different transmit window sizes and locate the 1915 * point where the RTT stops increasing, indicating that you have filled the 1916 * pipe, then scan backwards until you note that RTT stops decreasing, then 1917 * repeat ad-infinitum. This method works in principle but has severe 1918 * implementation issues due to RTT variances, timer granularity, and 1919 * instability in the algorithm which can lead to many false positives and 1920 * create oscillations as well as interact badly with other TCP streams 1921 * implementing the same algorithm. 1922 * 1923 * The second method is to limit the window to the bandwidth delay product 1924 * of the link. This is the method we implement. RTT variances and our 1925 * own manipulation of the congestion window, bwnd, can potentially 1926 * destabilize the algorithm. For this reason we have to stabilize the 1927 * elements used to calculate the window. We do this by using the minimum 1928 * observed RTT, the long term average of the observed bandwidth, and 1929 * by adding two segments worth of slop. It isn't perfect but it is able 1930 * to react to changing conditions and gives us a very stable basis on 1931 * which to extend the algorithm. 1932 */ 1933 void 1934 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1935 { 1936 u_long bw; 1937 u_long ibw; 1938 u_long bwnd; 1939 int save_ticks; 1940 int delta_ticks; 1941 1942 /* 1943 * If inflight_enable is disabled in the middle of a tcp connection, 1944 * make sure snd_bwnd is effectively disabled. 1945 */ 1946 if (!tcp_inflight_enable) { 1947 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1948 tp->snd_bandwidth = 0; 1949 return; 1950 } 1951 1952 /* 1953 * Validate the delta time. If a connection is new or has been idle 1954 * a long time we have to reset the bandwidth calculator. 1955 */ 1956 save_ticks = ticks; 1957 cpu_ccfence(); 1958 delta_ticks = save_ticks - tp->t_bw_rtttime; 1959 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1960 tp->t_bw_rtttime = save_ticks; 1961 tp->t_bw_rtseq = ack_seq; 1962 if (tp->snd_bandwidth == 0) 1963 tp->snd_bandwidth = tcp_inflight_min; 1964 return; 1965 } 1966 1967 /* 1968 * A delta of at least 1 tick is required. Waiting 2 ticks will 1969 * result in better (bw) accuracy. More than that and the ramp-up 1970 * will be too slow. 1971 */ 1972 if (delta_ticks == 0 || delta_ticks == 1) 1973 return; 1974 1975 /* 1976 * Sanity check, plus ignore pure window update acks. 1977 */ 1978 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1979 return; 1980 1981 /* 1982 * Figure out the bandwidth. Due to the tick granularity this 1983 * is a very rough number and it MUST be averaged over a fairly 1984 * long period of time. XXX we need to take into account a link 1985 * that is not using all available bandwidth, but for now our 1986 * slop will ramp us up if this case occurs and the bandwidth later 1987 * increases. 1988 */ 1989 ibw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 1990 tp->t_bw_rtttime = save_ticks; 1991 tp->t_bw_rtseq = ack_seq; 1992 bw = ((int64_t)tp->snd_bandwidth * 15 + ibw) >> 4; 1993 1994 tp->snd_bandwidth = bw; 1995 1996 /* 1997 * Calculate the semi-static bandwidth delay product, plus two maximal 1998 * segments. The additional slop puts us squarely in the sweet 1999 * spot and also handles the bandwidth run-up case. Without the 2000 * slop we could be locking ourselves into a lower bandwidth. 2001 * 2002 * At very high speeds the bw calculation can become overly sensitive 2003 * and error prone when delta_ticks is low (e.g. usually 1). To deal 2004 * with the problem the stab must be scaled to the bw. A stab of 50 2005 * (the default) increases the bw for the purposes of the bwnd 2006 * calculation by 5%. 2007 * 2008 * Situations Handled: 2009 * (1) Prevents over-queueing of packets on LANs, especially on 2010 * high speed LANs, allowing larger TCP buffers to be 2011 * specified, and also does a good job preventing 2012 * over-queueing of packets over choke points like modems 2013 * (at least for the transmit side). 2014 * 2015 * (2) Is able to handle changing network loads (bandwidth 2016 * drops so bwnd drops, bandwidth increases so bwnd 2017 * increases). 2018 * 2019 * (3) Theoretically should stabilize in the face of multiple 2020 * connections implementing the same algorithm (this may need 2021 * a little work). 2022 * 2023 * (4) Stability value (defaults to 20 = 2 maximal packets) can 2024 * be adjusted with a sysctl but typically only needs to be on 2025 * very slow connections. A value no smaller then 5 should 2026 * be used, but only reduce this default if you have no other 2027 * choice. 2028 */ 2029 2030 #define USERTT ((tp->t_srtt + tp->t_rttvar) + tcp_inflight_adjrtt) 2031 bw += bw * tcp_inflight_stab / 1000; 2032 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 2033 (int)tp->t_maxseg * 2; 2034 #undef USERTT 2035 2036 if (tcp_inflight_debug > 0) { 2037 static int ltime; 2038 if ((u_int)(save_ticks - ltime) >= hz / tcp_inflight_debug) { 2039 ltime = save_ticks; 2040 kprintf("%p ibw %ld bw %ld rttvar %d srtt %d " 2041 "bwnd %ld delta %d snd_win %ld\n", 2042 tp, ibw, bw, tp->t_rttvar, tp->t_srtt, 2043 bwnd, delta_ticks, tp->snd_wnd); 2044 } 2045 } 2046 if ((long)bwnd < tcp_inflight_min) 2047 bwnd = tcp_inflight_min; 2048 if (bwnd > tcp_inflight_max) 2049 bwnd = tcp_inflight_max; 2050 if ((long)bwnd < tp->t_maxseg * 2) 2051 bwnd = tp->t_maxseg * 2; 2052 tp->snd_bwnd = bwnd; 2053 } 2054 2055 static void 2056 tcp_rmx_iwsegs(struct tcpcb *tp, u_long *maxsegs, u_long *capsegs) 2057 { 2058 struct rtentry *rt; 2059 struct inpcb *inp = tp->t_inpcb; 2060 #ifdef INET6 2061 boolean_t isipv6 = INP_ISIPV6(inp); 2062 #else 2063 const boolean_t isipv6 = FALSE; 2064 #endif 2065 2066 /* XXX */ 2067 if (tcp_iw_maxsegs < TCP_IW_MAXSEGS_DFLT) 2068 tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 2069 if (tcp_iw_capsegs < TCP_IW_CAPSEGS_DFLT) 2070 tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 2071 2072 if (isipv6) 2073 rt = tcp_rtlookup6(&inp->inp_inc); 2074 else 2075 rt = tcp_rtlookup(&inp->inp_inc); 2076 if (rt == NULL || 2077 rt->rt_rmx.rmx_iwmaxsegs < TCP_IW_MAXSEGS_DFLT || 2078 rt->rt_rmx.rmx_iwcapsegs < TCP_IW_CAPSEGS_DFLT) { 2079 *maxsegs = tcp_iw_maxsegs; 2080 *capsegs = tcp_iw_capsegs; 2081 return; 2082 } 2083 *maxsegs = rt->rt_rmx.rmx_iwmaxsegs; 2084 *capsegs = rt->rt_rmx.rmx_iwcapsegs; 2085 } 2086 2087 u_long 2088 tcp_initial_window(struct tcpcb *tp) 2089 { 2090 if (tcp_do_rfc3390) { 2091 /* 2092 * RFC3390: 2093 * "If the SYN or SYN/ACK is lost, the initial window 2094 * used by a sender after a correctly transmitted SYN 2095 * MUST be one segment consisting of MSS bytes." 2096 * 2097 * However, we do something a little bit more aggressive 2098 * then RFC3390 here: 2099 * - Only if time spent in the SYN or SYN|ACK retransmition 2100 * >= 3 seconds, the IW is reduced. We do this mainly 2101 * because when RFC3390 is published, the initial RTO is 2102 * still 3 seconds (the threshold we test here), while 2103 * after RFC6298, the initial RTO is 1 second. This 2104 * behaviour probably still falls within the spirit of 2105 * RFC3390. 2106 * - When IW is reduced, 2*MSS is used instead of 1*MSS. 2107 * Mainly to avoid sender and receiver deadlock until 2108 * delayed ACK timer expires. And even RFC2581 does not 2109 * try to reduce IW upon SYN or SYN|ACK retransmition 2110 * timeout. 2111 * 2112 * See also: 2113 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03 2114 */ 2115 if (tp->t_rxtsyn >= TCPTV_RTOBASE3) { 2116 return (2 * tp->t_maxseg); 2117 } else { 2118 u_long maxsegs, capsegs; 2119 2120 tcp_rmx_iwsegs(tp, &maxsegs, &capsegs); 2121 return min(maxsegs * tp->t_maxseg, 2122 max(2 * tp->t_maxseg, capsegs * 1460)); 2123 } 2124 } else { 2125 /* 2126 * Even RFC2581 (back to 1999) allows 2*SMSS IW. 2127 * 2128 * Mainly to avoid sender and receiver deadlock 2129 * until delayed ACK timer expires. 2130 */ 2131 return (2 * tp->t_maxseg); 2132 } 2133 } 2134 2135 #ifdef TCP_SIGNATURE 2136 /* 2137 * Compute TCP-MD5 hash of a TCP segment. (RFC2385) 2138 * 2139 * We do this over ip, tcphdr, segment data, and the key in the SADB. 2140 * When called from tcp_input(), we can be sure that th_sum has been 2141 * zeroed out and verified already. 2142 * 2143 * Return 0 if successful, otherwise return -1. 2144 * 2145 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a 2146 * search with the destination IP address, and a 'magic SPI' to be 2147 * determined by the application. This is hardcoded elsewhere to 1179 2148 * right now. Another branch of this code exists which uses the SPD to 2149 * specify per-application flows but it is unstable. 2150 */ 2151 int 2152 tcpsignature_compute( 2153 struct mbuf *m, /* mbuf chain */ 2154 int len, /* length of TCP data */ 2155 int optlen, /* length of TCP options */ 2156 u_char *buf, /* storage for MD5 digest */ 2157 u_int direction) /* direction of flow */ 2158 { 2159 struct ippseudo ippseudo; 2160 MD5_CTX ctx; 2161 int doff; 2162 struct ip *ip; 2163 struct ipovly *ipovly; 2164 struct secasvar *sav; 2165 struct tcphdr *th; 2166 #ifdef INET6 2167 struct ip6_hdr *ip6; 2168 struct in6_addr in6; 2169 uint32_t plen; 2170 uint16_t nhdr; 2171 #endif /* INET6 */ 2172 u_short savecsum; 2173 2174 KASSERT(m != NULL, ("passed NULL mbuf. Game over.")); 2175 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature")); 2176 /* 2177 * Extract the destination from the IP header in the mbuf. 2178 */ 2179 ip = mtod(m, struct ip *); 2180 #ifdef INET6 2181 ip6 = NULL; /* Make the compiler happy. */ 2182 #endif /* INET6 */ 2183 /* 2184 * Look up an SADB entry which matches the address found in 2185 * the segment. 2186 */ 2187 switch (IP_VHL_V(ip->ip_vhl)) { 2188 case IPVERSION: 2189 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst, 2190 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2191 break; 2192 #ifdef INET6 2193 case (IPV6_VERSION >> 4): 2194 ip6 = mtod(m, struct ip6_hdr *); 2195 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst, 2196 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2197 break; 2198 #endif /* INET6 */ 2199 default: 2200 return (EINVAL); 2201 /* NOTREACHED */ 2202 break; 2203 } 2204 if (sav == NULL) { 2205 kprintf("%s: SADB lookup failed\n", __func__); 2206 return (EINVAL); 2207 } 2208 MD5Init(&ctx); 2209 2210 /* 2211 * Step 1: Update MD5 hash with IP pseudo-header. 2212 * 2213 * XXX The ippseudo header MUST be digested in network byte order, 2214 * or else we'll fail the regression test. Assume all fields we've 2215 * been doing arithmetic on have been in host byte order. 2216 * XXX One cannot depend on ipovly->ih_len here. When called from 2217 * tcp_output(), the underlying ip_len member has not yet been set. 2218 */ 2219 switch (IP_VHL_V(ip->ip_vhl)) { 2220 case IPVERSION: 2221 ipovly = (struct ipovly *)ip; 2222 ippseudo.ippseudo_src = ipovly->ih_src; 2223 ippseudo.ippseudo_dst = ipovly->ih_dst; 2224 ippseudo.ippseudo_pad = 0; 2225 ippseudo.ippseudo_p = IPPROTO_TCP; 2226 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen); 2227 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo)); 2228 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip)); 2229 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen; 2230 break; 2231 #ifdef INET6 2232 /* 2233 * RFC 2385, 2.0 Proposal 2234 * For IPv6, the pseudo-header is as described in RFC 2460, namely the 2235 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero- 2236 * extended next header value (to form 32 bits), and 32-bit segment 2237 * length. 2238 * Note: Upper-Layer Packet Length comes before Next Header. 2239 */ 2240 case (IPV6_VERSION >> 4): 2241 in6 = ip6->ip6_src; 2242 in6_clearscope(&in6); 2243 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2244 in6 = ip6->ip6_dst; 2245 in6_clearscope(&in6); 2246 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2247 plen = htonl(len + sizeof(struct tcphdr) + optlen); 2248 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t)); 2249 nhdr = 0; 2250 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2251 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2252 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2253 nhdr = IPPROTO_TCP; 2254 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2255 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr)); 2256 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen; 2257 break; 2258 #endif /* INET6 */ 2259 default: 2260 return (EINVAL); 2261 /* NOTREACHED */ 2262 break; 2263 } 2264 /* 2265 * Step 2: Update MD5 hash with TCP header, excluding options. 2266 * The TCP checksum must be set to zero. 2267 */ 2268 savecsum = th->th_sum; 2269 th->th_sum = 0; 2270 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr)); 2271 th->th_sum = savecsum; 2272 /* 2273 * Step 3: Update MD5 hash with TCP segment data. 2274 * Use m_apply() to avoid an early m_pullup(). 2275 */ 2276 if (len > 0) 2277 m_apply(m, doff, len, tcpsignature_apply, &ctx); 2278 /* 2279 * Step 4: Update MD5 hash with shared secret. 2280 */ 2281 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth)); 2282 MD5Final(buf, &ctx); 2283 key_sa_recordxfer(sav, m); 2284 key_freesav(sav); 2285 return (0); 2286 } 2287 2288 int 2289 tcpsignature_apply(void *fstate, void *data, unsigned int len) 2290 { 2291 2292 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len); 2293 return (0); 2294 } 2295 #endif /* TCP_SIGNATURE */ 2296