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 * 4. 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 static struct lwkt_token tcp_port_token = 165 LWKT_TOKEN_INITIALIZER(tcp_port_token); 166 167 int tcp_mssdflt = TCP_MSS; 168 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 169 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size"); 170 171 #ifdef INET6 172 int tcp_v6mssdflt = TCP6_MSS; 173 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW, 174 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6"); 175 #endif 176 177 /* 178 * Minimum MSS we accept and use. This prevents DoS attacks where 179 * we are forced to a ridiculous low MSS like 20 and send hundreds 180 * of packets instead of one. The effect scales with the available 181 * bandwidth and quickly saturates the CPU and network interface 182 * with packet generation and sending. Set to zero to disable MINMSS 183 * checking. This setting prevents us from sending too small packets. 184 */ 185 int tcp_minmss = TCP_MINMSS; 186 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW, 187 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size"); 188 189 #if 0 190 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ; 191 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW, 192 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time"); 193 #endif 194 195 int tcp_do_rfc1323 = 1; 196 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW, 197 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions"); 198 199 static int tcp_tcbhashsize = 0; 200 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD, 201 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable"); 202 203 static int do_tcpdrain = 1; 204 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0, 205 "Enable tcp_drain routine for extra help when low on mbufs"); 206 207 static int icmp_may_rst = 1; 208 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0, 209 "Certain ICMP unreachable messages may abort connections in SYN_SENT"); 210 211 static int tcp_isn_reseed_interval = 0; 212 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW, 213 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret"); 214 215 /* 216 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on 217 * by default, but with generous values which should allow maximal 218 * bandwidth. In particular, the slop defaults to 50 (5 packets). 219 * 220 * The reason for doing this is that the limiter is the only mechanism we 221 * have which seems to do a really good job preventing receiver RX rings 222 * on network interfaces from getting blown out. Even though GigE/10GigE 223 * is supposed to flow control it looks like either it doesn't actually 224 * do it or Open Source drivers do not properly enable it. 225 * 226 * People using the limiter to reduce bottlenecks on slower WAN connections 227 * should set the slop to 20 (2 packets). 228 */ 229 static int tcp_inflight_enable = 1; 230 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW, 231 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting"); 232 233 static int tcp_inflight_debug = 0; 234 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW, 235 &tcp_inflight_debug, 0, "Debug TCP inflight calculations"); 236 237 static int tcp_inflight_min = 6144; 238 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW, 239 &tcp_inflight_min, 0, "Lower bound for TCP inflight window"); 240 241 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT; 242 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW, 243 &tcp_inflight_max, 0, "Upper bound for TCP inflight window"); 244 245 static int tcp_inflight_stab = 50; 246 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW, 247 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 3 packets)"); 248 249 static int tcp_do_rfc3390 = 1; 250 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_RW, 251 &tcp_do_rfc3390, 0, 252 "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)"); 253 254 static u_long tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 255 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwmaxsegs, CTLFLAG_RW, 256 &tcp_iw_maxsegs, 0, "TCP IW segments max"); 257 258 static u_long tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 259 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwcapsegs, CTLFLAG_RW, 260 &tcp_iw_capsegs, 0, "TCP IW segments"); 261 262 int tcp_low_rtobase = 1; 263 SYSCTL_INT(_net_inet_tcp, OID_AUTO, low_rtobase, CTLFLAG_RW, 264 &tcp_low_rtobase, 0, "Lowering the Initial RTO (RFC 6298)"); 265 266 static int tcp_do_ncr = 1; 267 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr, CTLFLAG_RW, 268 &tcp_do_ncr, 0, "Non-Congestion Robustness (RFC 4653)"); 269 270 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives"); 271 static struct malloc_pipe tcptemp_mpipe; 272 273 static void tcp_willblock(void); 274 static void tcp_notify (struct inpcb *, int); 275 276 struct tcp_stats tcpstats_percpu[MAXCPU] __cachealign; 277 278 static int 279 sysctl_tcpstats(SYSCTL_HANDLER_ARGS) 280 { 281 int cpu, error = 0; 282 283 for (cpu = 0; cpu < ncpus; ++cpu) { 284 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu], 285 sizeof(struct tcp_stats)))) 286 break; 287 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu], 288 sizeof(struct tcp_stats)))) 289 break; 290 } 291 292 return (error); 293 } 294 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW), 295 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics"); 296 297 /* 298 * Target size of TCP PCB hash tables. Must be a power of two. 299 * 300 * Note that this can be overridden by the kernel environment 301 * variable net.inet.tcp.tcbhashsize 302 */ 303 #ifndef TCBHASHSIZE 304 #define TCBHASHSIZE 512 305 #endif 306 307 /* 308 * This is the actual shape of what we allocate using the zone 309 * allocator. Doing it this way allows us to protect both structures 310 * using the same generation count, and also eliminates the overhead 311 * of allocating tcpcbs separately. By hiding the structure here, 312 * we avoid changing most of the rest of the code (although it needs 313 * to be changed, eventually, for greater efficiency). 314 */ 315 #define ALIGNMENT 32 316 #define ALIGNM1 (ALIGNMENT - 1) 317 struct inp_tp { 318 union { 319 struct inpcb inp; 320 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1]; 321 } inp_tp_u; 322 struct tcpcb tcb; 323 struct tcp_callout inp_tp_rexmt; 324 struct tcp_callout inp_tp_persist; 325 struct tcp_callout inp_tp_keep; 326 struct tcp_callout inp_tp_2msl; 327 struct tcp_callout inp_tp_delack; 328 struct netmsg_tcp_timer inp_tp_timermsg; 329 struct netmsg_base inp_tp_sndmore; 330 }; 331 #undef ALIGNMENT 332 #undef ALIGNM1 333 334 /* 335 * Tcp initialization 336 */ 337 void 338 tcp_init(void) 339 { 340 struct inpcbporthead *porthashbase; 341 struct inpcbinfo *ticb; 342 u_long porthashmask; 343 int hashsize = TCBHASHSIZE; 344 int cpu; 345 346 /* 347 * note: tcptemp is used for keepalives, and it is ok for an 348 * allocation to fail so do not specify MPF_INT. 349 */ 350 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp), 351 25, -1, 0, NULL, NULL, NULL); 352 353 tcp_delacktime = TCPTV_DELACK; 354 tcp_keepinit = TCPTV_KEEP_INIT; 355 tcp_keepidle = TCPTV_KEEP_IDLE; 356 tcp_keepintvl = TCPTV_KEEPINTVL; 357 tcp_maxpersistidle = TCPTV_KEEP_IDLE; 358 tcp_msl = TCPTV_MSL; 359 tcp_rexmit_min = TCPTV_MIN; 360 tcp_rexmit_slop = TCPTV_CPU_VAR; 361 362 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize); 363 if (!powerof2(hashsize)) { 364 kprintf("WARNING: TCB hash size not a power of 2\n"); 365 hashsize = 512; /* safe default */ 366 } 367 tcp_tcbhashsize = hashsize; 368 porthashbase = hashinit(hashsize, M_PCB, &porthashmask); 369 370 for (cpu = 0; cpu < ncpus2; cpu++) { 371 ticb = &tcbinfo[cpu]; 372 in_pcbinfo_init(ticb); 373 ticb->cpu = cpu; 374 ticb->hashbase = hashinit(hashsize, M_PCB, 375 &ticb->hashmask); 376 ticb->porthashbase = porthashbase; 377 ticb->porthashmask = porthashmask; 378 ticb->porttoken = &tcp_port_token; 379 #if 0 380 ticb->porthashbase = hashinit(hashsize, M_PCB, 381 &ticb->porthashmask); 382 #endif 383 ticb->wildcardhashbase = hashinit(hashsize, M_PCB, 384 &ticb->wildcardhashmask); 385 ticb->ipi_size = sizeof(struct inp_tp); 386 TAILQ_INIT(&tcpcbackq[cpu]); 387 } 388 389 tcp_reass_maxseg = nmbclusters / 16; 390 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg); 391 392 #ifdef INET6 393 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) 394 #else 395 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) 396 #endif 397 if (max_protohdr < TCP_MINPROTOHDR) 398 max_protohdr = TCP_MINPROTOHDR; 399 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) 400 panic("tcp_init"); 401 #undef TCP_MINPROTOHDR 402 403 /* 404 * Initialize TCP statistics counters for each CPU. 405 */ 406 for (cpu = 0; cpu < ncpus; ++cpu) { 407 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats)); 408 } 409 410 syncache_init(); 411 netisr_register_rollup(tcp_willblock, NETISR_ROLLUP_PRIO_TCP); 412 } 413 414 static void 415 tcp_willblock(void) 416 { 417 struct tcpcb *tp; 418 int cpu = mycpu->gd_cpuid; 419 420 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) { 421 KKASSERT(tp->t_flags & TF_ONOUTPUTQ); 422 tp->t_flags &= ~TF_ONOUTPUTQ; 423 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq); 424 tcp_output(tp); 425 } 426 } 427 428 /* 429 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. 430 * tcp_template used to store this data in mbufs, but we now recopy it out 431 * of the tcpcb each time to conserve mbufs. 432 */ 433 void 434 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr, boolean_t tso) 435 { 436 struct inpcb *inp = tp->t_inpcb; 437 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr; 438 439 #ifdef INET6 440 if (inp->inp_vflag & INP_IPV6) { 441 struct ip6_hdr *ip6; 442 443 ip6 = (struct ip6_hdr *)ip_ptr; 444 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | 445 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK); 446 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | 447 (IPV6_VERSION & IPV6_VERSION_MASK); 448 ip6->ip6_nxt = IPPROTO_TCP; 449 ip6->ip6_plen = sizeof(struct tcphdr); 450 ip6->ip6_src = inp->in6p_laddr; 451 ip6->ip6_dst = inp->in6p_faddr; 452 tcp_hdr->th_sum = 0; 453 } else 454 #endif 455 { 456 struct ip *ip = (struct ip *) ip_ptr; 457 u_int plen; 458 459 ip->ip_vhl = IP_VHL_BORING; 460 ip->ip_tos = 0; 461 ip->ip_len = 0; 462 ip->ip_id = 0; 463 ip->ip_off = 0; 464 ip->ip_ttl = 0; 465 ip->ip_sum = 0; 466 ip->ip_p = IPPROTO_TCP; 467 ip->ip_src = inp->inp_laddr; 468 ip->ip_dst = inp->inp_faddr; 469 470 if (tso) 471 plen = htons(IPPROTO_TCP); 472 else 473 plen = htons(sizeof(struct tcphdr) + IPPROTO_TCP); 474 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr, 475 ip->ip_dst.s_addr, plen); 476 } 477 478 tcp_hdr->th_sport = inp->inp_lport; 479 tcp_hdr->th_dport = inp->inp_fport; 480 tcp_hdr->th_seq = 0; 481 tcp_hdr->th_ack = 0; 482 tcp_hdr->th_x2 = 0; 483 tcp_hdr->th_off = 5; 484 tcp_hdr->th_flags = 0; 485 tcp_hdr->th_win = 0; 486 tcp_hdr->th_urp = 0; 487 } 488 489 /* 490 * Create template to be used to send tcp packets on a connection. 491 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only 492 * use for this function is in keepalives, which use tcp_respond. 493 */ 494 struct tcptemp * 495 tcp_maketemplate(struct tcpcb *tp) 496 { 497 struct tcptemp *tmp; 498 499 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL) 500 return (NULL); 501 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t, FALSE); 502 return (tmp); 503 } 504 505 void 506 tcp_freetemplate(struct tcptemp *tmp) 507 { 508 mpipe_free(&tcptemp_mpipe, tmp); 509 } 510 511 /* 512 * Send a single message to the TCP at address specified by 513 * the given TCP/IP header. If m == NULL, then we make a copy 514 * of the tcpiphdr at ti and send directly to the addressed host. 515 * This is used to force keep alive messages out using the TCP 516 * template for a connection. If flags are given then we send 517 * a message back to the TCP which originated the * segment ti, 518 * and discard the mbuf containing it and any other attached mbufs. 519 * 520 * In any case the ack and sequence number of the transmitted 521 * segment are as specified by the parameters. 522 * 523 * NOTE: If m != NULL, then ti must point to *inside* the mbuf. 524 */ 525 void 526 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, 527 tcp_seq ack, tcp_seq seq, int flags) 528 { 529 int tlen; 530 int win = 0; 531 struct route *ro = NULL; 532 struct route sro; 533 struct ip *ip = ipgen; 534 struct tcphdr *nth; 535 int ipflags = 0; 536 struct route_in6 *ro6 = NULL; 537 struct route_in6 sro6; 538 struct ip6_hdr *ip6 = ipgen; 539 boolean_t use_tmpro = TRUE; 540 #ifdef INET6 541 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6); 542 #else 543 const boolean_t isipv6 = FALSE; 544 #endif 545 546 if (tp != NULL) { 547 if (!(flags & TH_RST)) { 548 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv); 549 if (win < 0) 550 win = 0; 551 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 552 win = (long)TCP_MAXWIN << tp->rcv_scale; 553 } 554 /* 555 * Don't use the route cache of a listen socket, 556 * it is not MPSAFE; use temporary route cache. 557 */ 558 if (tp->t_state != TCPS_LISTEN) { 559 if (isipv6) 560 ro6 = &tp->t_inpcb->in6p_route; 561 else 562 ro = &tp->t_inpcb->inp_route; 563 use_tmpro = FALSE; 564 } 565 } 566 if (use_tmpro) { 567 if (isipv6) { 568 ro6 = &sro6; 569 bzero(ro6, sizeof *ro6); 570 } else { 571 ro = &sro; 572 bzero(ro, sizeof *ro); 573 } 574 } 575 if (m == NULL) { 576 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 577 if (m == NULL) 578 return; 579 tlen = 0; 580 m->m_data += max_linkhdr; 581 if (isipv6) { 582 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); 583 ip6 = mtod(m, struct ip6_hdr *); 584 nth = (struct tcphdr *)(ip6 + 1); 585 } else { 586 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip)); 587 ip = mtod(m, struct ip *); 588 nth = (struct tcphdr *)(ip + 1); 589 } 590 bcopy(th, nth, sizeof(struct tcphdr)); 591 flags = TH_ACK; 592 } else { 593 m_freem(m->m_next); 594 m->m_next = NULL; 595 m->m_data = (caddr_t)ipgen; 596 /* m_len is set later */ 597 tlen = 0; 598 #define xchg(a, b, type) { type t; t = a; a = b; b = t; } 599 if (isipv6) { 600 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); 601 nth = (struct tcphdr *)(ip6 + 1); 602 } else { 603 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long); 604 nth = (struct tcphdr *)(ip + 1); 605 } 606 if (th != nth) { 607 /* 608 * this is usually a case when an extension header 609 * exists between the IPv6 header and the 610 * TCP header. 611 */ 612 nth->th_sport = th->th_sport; 613 nth->th_dport = th->th_dport; 614 } 615 xchg(nth->th_dport, nth->th_sport, n_short); 616 #undef xchg 617 } 618 if (isipv6) { 619 ip6->ip6_flow = 0; 620 ip6->ip6_vfc = IPV6_VERSION; 621 ip6->ip6_nxt = IPPROTO_TCP; 622 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen)); 623 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 624 } else { 625 tlen += sizeof(struct tcpiphdr); 626 ip->ip_len = tlen; 627 ip->ip_ttl = ip_defttl; 628 } 629 m->m_len = tlen; 630 m->m_pkthdr.len = tlen; 631 m->m_pkthdr.rcvif = NULL; 632 nth->th_seq = htonl(seq); 633 nth->th_ack = htonl(ack); 634 nth->th_x2 = 0; 635 nth->th_off = sizeof(struct tcphdr) >> 2; 636 nth->th_flags = flags; 637 if (tp != NULL) 638 nth->th_win = htons((u_short) (win >> tp->rcv_scale)); 639 else 640 nth->th_win = htons((u_short)win); 641 nth->th_urp = 0; 642 if (isipv6) { 643 nth->th_sum = 0; 644 nth->th_sum = in6_cksum(m, IPPROTO_TCP, 645 sizeof(struct ip6_hdr), 646 tlen - sizeof(struct ip6_hdr)); 647 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL, 648 (ro6 && ro6->ro_rt) ? 649 ro6->ro_rt->rt_ifp : NULL); 650 } else { 651 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 652 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); 653 m->m_pkthdr.csum_flags = CSUM_TCP; 654 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 655 m->m_pkthdr.csum_thlen = sizeof(struct tcphdr); 656 } 657 #ifdef TCPDEBUG 658 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) 659 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); 660 #endif 661 if (isipv6) { 662 ip6_output(m, NULL, ro6, ipflags, NULL, NULL, 663 tp ? tp->t_inpcb : NULL); 664 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) { 665 RTFREE(ro6->ro_rt); 666 ro6->ro_rt = NULL; 667 } 668 } else { 669 ipflags |= IP_DEBUGROUTE; 670 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL); 671 if ((ro == &sro) && (ro->ro_rt != NULL)) { 672 RTFREE(ro->ro_rt); 673 ro->ro_rt = NULL; 674 } 675 } 676 } 677 678 /* 679 * Create a new TCP control block, making an 680 * empty reassembly queue and hooking it to the argument 681 * protocol control block. The `inp' parameter must have 682 * come from the zone allocator set up in tcp_init(). 683 */ 684 struct tcpcb * 685 tcp_newtcpcb(struct inpcb *inp) 686 { 687 struct inp_tp *it; 688 struct tcpcb *tp; 689 #ifdef INET6 690 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 691 #else 692 const boolean_t isipv6 = FALSE; 693 #endif 694 695 it = (struct inp_tp *)inp; 696 tp = &it->tcb; 697 bzero(tp, sizeof(struct tcpcb)); 698 TAILQ_INIT(&tp->t_segq); 699 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 700 tp->t_rxtthresh = tcprexmtthresh; 701 702 /* Set up our timeouts. */ 703 tp->tt_rexmt = &it->inp_tp_rexmt; 704 tp->tt_persist = &it->inp_tp_persist; 705 tp->tt_keep = &it->inp_tp_keep; 706 tp->tt_2msl = &it->inp_tp_2msl; 707 tp->tt_delack = &it->inp_tp_delack; 708 tcp_inittimers(tp); 709 710 /* 711 * Zero out timer message. We don't create it here, 712 * since the current CPU may not be the owner of this 713 * inpcb. 714 */ 715 tp->tt_msg = &it->inp_tp_timermsg; 716 bzero(tp->tt_msg, sizeof(*tp->tt_msg)); 717 718 tp->t_keepinit = tcp_keepinit; 719 tp->t_keepidle = tcp_keepidle; 720 tp->t_keepintvl = tcp_keepintvl; 721 tp->t_keepcnt = tcp_keepcnt; 722 tp->t_maxidle = tp->t_keepintvl * tp->t_keepcnt; 723 724 if (tcp_do_ncr) 725 tp->t_flags |= TF_NCR; 726 if (tcp_do_rfc1323) 727 tp->t_flags |= (TF_REQ_SCALE | TF_REQ_TSTMP); 728 729 tp->t_inpcb = inp; /* XXX */ 730 tp->t_state = TCPS_CLOSED; 731 /* 732 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 733 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 734 * reasonable initial retransmit time. 735 */ 736 tp->t_srtt = TCPTV_SRTTBASE; 737 tp->t_rttvar = 738 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 739 tp->t_rttmin = tcp_rexmit_min; 740 tp->t_rxtcur = TCPTV_RTOBASE; 741 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 742 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 743 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 744 tp->snd_last = ticks; 745 tp->t_rcvtime = ticks; 746 /* 747 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 748 * because the socket may be bound to an IPv6 wildcard address, 749 * which may match an IPv4-mapped IPv6 address. 750 */ 751 inp->inp_ip_ttl = ip_defttl; 752 inp->inp_ppcb = tp; 753 tcp_sack_tcpcb_init(tp); 754 755 tp->tt_sndmore = &it->inp_tp_sndmore; 756 tcp_output_init(tp); 757 758 return (tp); /* XXX */ 759 } 760 761 /* 762 * Drop a TCP connection, reporting the specified error. 763 * If connection is synchronized, then send a RST to peer. 764 */ 765 struct tcpcb * 766 tcp_drop(struct tcpcb *tp, int error) 767 { 768 struct socket *so = tp->t_inpcb->inp_socket; 769 770 if (TCPS_HAVERCVDSYN(tp->t_state)) { 771 tp->t_state = TCPS_CLOSED; 772 tcp_output(tp); 773 tcpstat.tcps_drops++; 774 } else 775 tcpstat.tcps_conndrops++; 776 if (error == ETIMEDOUT && tp->t_softerror) 777 error = tp->t_softerror; 778 so->so_error = error; 779 return (tcp_close(tp)); 780 } 781 782 struct netmsg_listen_detach { 783 struct netmsg_base base; 784 struct tcpcb *nm_tp; 785 }; 786 787 static void 788 tcp_listen_detach_handler(netmsg_t msg) 789 { 790 struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg; 791 struct tcpcb *tp = nmsg->nm_tp; 792 int cpu = mycpuid, nextcpu; 793 794 if (tp->t_flags & TF_LISTEN) 795 syncache_destroy(tp); 796 797 in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]); 798 799 nextcpu = cpu + 1; 800 if (nextcpu < ncpus2) 801 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nmsg->base.lmsg); 802 else 803 lwkt_replymsg(&nmsg->base.lmsg, 0); 804 } 805 806 /* 807 * Close a TCP control block: 808 * discard all space held by the tcp 809 * discard internet protocol block 810 * wake up any sleepers 811 */ 812 struct tcpcb * 813 tcp_close(struct tcpcb *tp) 814 { 815 struct tseg_qent *q; 816 struct inpcb *inp = tp->t_inpcb; 817 struct socket *so = inp->inp_socket; 818 struct rtentry *rt; 819 boolean_t dosavessthresh; 820 #ifdef INET6 821 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 822 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0); 823 #else 824 const boolean_t isipv6 = FALSE; 825 #endif 826 827 /* 828 * INP_WILDCARD_MP indicates that listen(2) has been called on 829 * this socket. This implies: 830 * - A wildcard inp's hash is replicated for each protocol thread. 831 * - Syncache for this inp grows independently in each protocol 832 * thread. 833 * - There is more than one cpu 834 * 835 * We have to chain a message to the rest of the protocol threads 836 * to cleanup the wildcard hash and the syncache. The cleanup 837 * in the current protocol thread is defered till the end of this 838 * function. 839 * 840 * NOTE: 841 * After cleanup the inp's hash and syncache entries, this inp will 842 * no longer be available to the rest of the protocol threads, so we 843 * are safe to whack the inp in the following code. 844 */ 845 if (inp->inp_flags & INP_WILDCARD_MP) { 846 struct netmsg_listen_detach nmsg; 847 848 KKASSERT(so->so_port == netisr_cpuport(0)); 849 KKASSERT(&curthread->td_msgport == netisr_cpuport(0)); 850 KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]); 851 852 netmsg_init(&nmsg.base, NULL, &curthread->td_msgport, 853 MSGF_PRIORITY, tcp_listen_detach_handler); 854 nmsg.nm_tp = tp; 855 lwkt_domsg(netisr_cpuport(1), &nmsg.base.lmsg, 0); 856 857 inp->inp_flags &= ~INP_WILDCARD_MP; 858 } 859 860 KKASSERT(tp->t_state != TCPS_TERMINATING); 861 tp->t_state = TCPS_TERMINATING; 862 863 /* 864 * Make sure that all of our timers are stopped before we 865 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL), 866 * timers are never used. If timer message is never created 867 * (tp->tt_msg->tt_tcb == NULL), timers are never used too. 868 */ 869 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) { 870 tcp_callout_stop(tp, tp->tt_rexmt); 871 tcp_callout_stop(tp, tp->tt_persist); 872 tcp_callout_stop(tp, tp->tt_keep); 873 tcp_callout_stop(tp, tp->tt_2msl); 874 tcp_callout_stop(tp, tp->tt_delack); 875 } 876 877 if (tp->t_flags & TF_ONOUTPUTQ) { 878 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 879 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 880 tp->t_flags &= ~TF_ONOUTPUTQ; 881 } 882 883 /* 884 * If we got enough samples through the srtt filter, 885 * save the rtt and rttvar in the routing entry. 886 * 'Enough' is arbitrarily defined as the 16 samples. 887 * 16 samples is enough for the srtt filter to converge 888 * to within 5% of the correct value; fewer samples and 889 * we could save a very bogus rtt. 890 * 891 * Don't update the default route's characteristics and don't 892 * update anything that the user "locked". 893 */ 894 if (tp->t_rttupdated >= 16) { 895 u_long i = 0; 896 897 if (isipv6) { 898 struct sockaddr_in6 *sin6; 899 900 if ((rt = inp->in6p_route.ro_rt) == NULL) 901 goto no_valid_rt; 902 sin6 = (struct sockaddr_in6 *)rt_key(rt); 903 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 904 goto no_valid_rt; 905 } else 906 if ((rt = inp->inp_route.ro_rt) == NULL || 907 ((struct sockaddr_in *)rt_key(rt))-> 908 sin_addr.s_addr == INADDR_ANY) 909 goto no_valid_rt; 910 911 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 912 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 913 if (rt->rt_rmx.rmx_rtt && i) 914 /* 915 * filter this update to half the old & half 916 * the new values, converting scale. 917 * See route.h and tcp_var.h for a 918 * description of the scaling constants. 919 */ 920 rt->rt_rmx.rmx_rtt = 921 (rt->rt_rmx.rmx_rtt + i) / 2; 922 else 923 rt->rt_rmx.rmx_rtt = i; 924 tcpstat.tcps_cachedrtt++; 925 } 926 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 927 i = tp->t_rttvar * 928 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 929 if (rt->rt_rmx.rmx_rttvar && i) 930 rt->rt_rmx.rmx_rttvar = 931 (rt->rt_rmx.rmx_rttvar + i) / 2; 932 else 933 rt->rt_rmx.rmx_rttvar = i; 934 tcpstat.tcps_cachedrttvar++; 935 } 936 /* 937 * The old comment here said: 938 * update the pipelimit (ssthresh) if it has been updated 939 * already or if a pipesize was specified & the threshhold 940 * got below half the pipesize. I.e., wait for bad news 941 * before we start updating, then update on both good 942 * and bad news. 943 * 944 * But we want to save the ssthresh even if no pipesize is 945 * specified explicitly in the route, because such 946 * connections still have an implicit pipesize specified 947 * by the global tcp_sendspace. In the absence of a reliable 948 * way to calculate the pipesize, it will have to do. 949 */ 950 i = tp->snd_ssthresh; 951 if (rt->rt_rmx.rmx_sendpipe != 0) 952 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 953 else 954 dosavessthresh = (i < so->so_snd.ssb_hiwat/2); 955 if (dosavessthresh || 956 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 957 (rt->rt_rmx.rmx_ssthresh != 0))) { 958 /* 959 * convert the limit from user data bytes to 960 * packets then to packet data bytes. 961 */ 962 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 963 if (i < 2) 964 i = 2; 965 i *= tp->t_maxseg + 966 (isipv6 ? 967 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 968 sizeof(struct tcpiphdr)); 969 if (rt->rt_rmx.rmx_ssthresh) 970 rt->rt_rmx.rmx_ssthresh = 971 (rt->rt_rmx.rmx_ssthresh + i) / 2; 972 else 973 rt->rt_rmx.rmx_ssthresh = i; 974 tcpstat.tcps_cachedssthresh++; 975 } 976 } 977 978 no_valid_rt: 979 /* free the reassembly queue, if any */ 980 while((q = TAILQ_FIRST(&tp->t_segq)) != NULL) { 981 TAILQ_REMOVE(&tp->t_segq, q, tqe_q); 982 m_freem(q->tqe_m); 983 kfree(q, M_TSEGQ); 984 atomic_add_int(&tcp_reass_qsize, -1); 985 } 986 /* throw away SACK blocks in scoreboard*/ 987 if (TCP_DO_SACK(tp)) 988 tcp_sack_destroy(&tp->scb); 989 990 inp->inp_ppcb = NULL; 991 soisdisconnected(so); 992 /* note: pcb detached later on */ 993 994 tcp_destroy_timermsg(tp); 995 tcp_output_cancel(tp); 996 997 if (tp->t_flags & TF_LISTEN) 998 syncache_destroy(tp); 999 1000 so_async_rcvd_drop(so); 1001 /* Drop the reference for the asynchronized pru_rcvd */ 1002 sofree(so); 1003 1004 /* 1005 * NOTE: 1006 * pcbdetach removes any wildcard hash entry on the current CPU. 1007 */ 1008 #ifdef INET6 1009 if (isafinet6) 1010 in6_pcbdetach(inp); 1011 else 1012 #endif 1013 in_pcbdetach(inp); 1014 1015 tcpstat.tcps_closed++; 1016 return (NULL); 1017 } 1018 1019 static __inline void 1020 tcp_drain_oncpu(struct inpcbhead *head) 1021 { 1022 struct inpcb *marker; 1023 struct inpcb *inpb; 1024 struct tcpcb *tcpb; 1025 struct tseg_qent *te; 1026 1027 /* 1028 * Allows us to block while running the list 1029 */ 1030 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1031 marker->inp_flags |= INP_PLACEMARKER; 1032 LIST_INSERT_HEAD(head, marker, inp_list); 1033 1034 while ((inpb = LIST_NEXT(marker, inp_list)) != NULL) { 1035 if ((inpb->inp_flags & INP_PLACEMARKER) == 0 && 1036 (tcpb = intotcpcb(inpb)) != NULL && 1037 (te = TAILQ_FIRST(&tcpb->t_segq)) != NULL) { 1038 TAILQ_REMOVE(&tcpb->t_segq, te, tqe_q); 1039 if (te->tqe_th->th_flags & TH_FIN) 1040 tcpb->t_flags &= ~TF_QUEDFIN; 1041 m_freem(te->tqe_m); 1042 kfree(te, M_TSEGQ); 1043 atomic_add_int(&tcp_reass_qsize, -1); 1044 /* retry */ 1045 } else { 1046 LIST_REMOVE(marker, inp_list); 1047 LIST_INSERT_AFTER(inpb, marker, inp_list); 1048 } 1049 } 1050 LIST_REMOVE(marker, inp_list); 1051 kfree(marker, M_TEMP); 1052 } 1053 1054 struct netmsg_tcp_drain { 1055 struct netmsg_base base; 1056 struct inpcbhead *nm_head; 1057 }; 1058 1059 static void 1060 tcp_drain_handler(netmsg_t msg) 1061 { 1062 struct netmsg_tcp_drain *nm = (void *)msg; 1063 1064 tcp_drain_oncpu(nm->nm_head); 1065 lwkt_replymsg(&nm->base.lmsg, 0); 1066 } 1067 1068 void 1069 tcp_drain(void) 1070 { 1071 int cpu; 1072 1073 if (!do_tcpdrain) 1074 return; 1075 1076 /* 1077 * Walk the tcpbs, if existing, and flush the reassembly queue, 1078 * if there is one... 1079 * XXX: The "Net/3" implementation doesn't imply that the TCP 1080 * reassembly queue should be flushed, but in a situation 1081 * where we're really low on mbufs, this is potentially 1082 * useful. 1083 */ 1084 for (cpu = 0; cpu < ncpus2; cpu++) { 1085 struct netmsg_tcp_drain *nm; 1086 1087 if (cpu == mycpu->gd_cpuid) { 1088 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead); 1089 } else { 1090 nm = kmalloc(sizeof(struct netmsg_tcp_drain), 1091 M_LWKTMSG, M_NOWAIT); 1092 if (nm == NULL) 1093 continue; 1094 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1095 0, tcp_drain_handler); 1096 nm->nm_head = &tcbinfo[cpu].pcblisthead; 1097 lwkt_sendmsg(netisr_cpuport(cpu), &nm->base.lmsg); 1098 } 1099 } 1100 } 1101 1102 /* 1103 * Notify a tcp user of an asynchronous error; 1104 * store error as soft error, but wake up user 1105 * (for now, won't do anything until can select for soft error). 1106 * 1107 * Do not wake up user since there currently is no mechanism for 1108 * reporting soft errors (yet - a kqueue filter may be added). 1109 */ 1110 static void 1111 tcp_notify(struct inpcb *inp, int error) 1112 { 1113 struct tcpcb *tp = intotcpcb(inp); 1114 1115 /* 1116 * Ignore some errors if we are hooked up. 1117 * If connection hasn't completed, has retransmitted several times, 1118 * and receives a second error, give up now. This is better 1119 * than waiting a long time to establish a connection that 1120 * can never complete. 1121 */ 1122 if (tp->t_state == TCPS_ESTABLISHED && 1123 (error == EHOSTUNREACH || error == ENETUNREACH || 1124 error == EHOSTDOWN)) { 1125 return; 1126 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1127 tp->t_softerror) 1128 tcp_drop(tp, error); 1129 else 1130 tp->t_softerror = error; 1131 #if 0 1132 wakeup(&so->so_timeo); 1133 sorwakeup(so); 1134 sowwakeup(so); 1135 #endif 1136 } 1137 1138 static int 1139 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1140 { 1141 int error, i, n; 1142 struct inpcb *marker; 1143 struct inpcb *inp; 1144 globaldata_t gd; 1145 int origcpu, ccpu; 1146 1147 error = 0; 1148 n = 0; 1149 1150 /* 1151 * The process of preparing the TCB list is too time-consuming and 1152 * resource-intensive to repeat twice on every request. 1153 */ 1154 if (req->oldptr == NULL) { 1155 for (ccpu = 0; ccpu < ncpus; ++ccpu) { 1156 gd = globaldata_find(ccpu); 1157 n += tcbinfo[gd->gd_cpuid].ipi_count; 1158 } 1159 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1160 return (0); 1161 } 1162 1163 if (req->newptr != NULL) 1164 return (EPERM); 1165 1166 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1167 marker->inp_flags |= INP_PLACEMARKER; 1168 1169 /* 1170 * OK, now we're committed to doing something. Run the inpcb list 1171 * for each cpu in the system and construct the output. Use a 1172 * list placemarker to deal with list changes occuring during 1173 * copyout blockages (but otherwise depend on being on the correct 1174 * cpu to avoid races). 1175 */ 1176 origcpu = mycpu->gd_cpuid; 1177 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) { 1178 globaldata_t rgd; 1179 caddr_t inp_ppcb; 1180 struct xtcpcb xt; 1181 int cpu_id; 1182 1183 cpu_id = (origcpu + ccpu) % ncpus; 1184 if ((smp_active_mask & CPUMASK(cpu_id)) == 0) 1185 continue; 1186 rgd = globaldata_find(cpu_id); 1187 lwkt_setcpu_self(rgd); 1188 1189 n = tcbinfo[cpu_id].ipi_count; 1190 1191 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list); 1192 i = 0; 1193 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1194 /* 1195 * process a snapshot of pcbs, ignoring placemarkers 1196 * and using our own to allow SYSCTL_OUT to block. 1197 */ 1198 LIST_REMOVE(marker, inp_list); 1199 LIST_INSERT_AFTER(inp, marker, inp_list); 1200 1201 if (inp->inp_flags & INP_PLACEMARKER) 1202 continue; 1203 if (prison_xinpcb(req->td, inp)) 1204 continue; 1205 1206 xt.xt_len = sizeof xt; 1207 bcopy(inp, &xt.xt_inp, sizeof *inp); 1208 inp_ppcb = inp->inp_ppcb; 1209 if (inp_ppcb != NULL) 1210 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1211 else 1212 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1213 if (inp->inp_socket) 1214 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1215 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1216 break; 1217 ++i; 1218 } 1219 LIST_REMOVE(marker, inp_list); 1220 if (error == 0 && i < n) { 1221 bzero(&xt, sizeof xt); 1222 xt.xt_len = sizeof xt; 1223 while (i < n) { 1224 error = SYSCTL_OUT(req, &xt, sizeof xt); 1225 if (error) 1226 break; 1227 ++i; 1228 } 1229 } 1230 } 1231 1232 /* 1233 * Make sure we are on the same cpu we were on originally, since 1234 * higher level callers expect this. Also don't pollute caches with 1235 * migrated userland data by (eventually) returning to userland 1236 * on a different cpu. 1237 */ 1238 lwkt_setcpu_self(globaldata_find(origcpu)); 1239 kfree(marker, M_TEMP); 1240 return (error); 1241 } 1242 1243 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1244 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1245 1246 static int 1247 tcp_getcred(SYSCTL_HANDLER_ARGS) 1248 { 1249 struct sockaddr_in addrs[2]; 1250 struct inpcb *inp; 1251 int cpu; 1252 int error; 1253 1254 error = priv_check(req->td, PRIV_ROOT); 1255 if (error != 0) 1256 return (error); 1257 error = SYSCTL_IN(req, addrs, sizeof addrs); 1258 if (error != 0) 1259 return (error); 1260 crit_enter(); 1261 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1262 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1263 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1264 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1265 if (inp == NULL || inp->inp_socket == NULL) { 1266 error = ENOENT; 1267 goto out; 1268 } 1269 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1270 out: 1271 crit_exit(); 1272 return (error); 1273 } 1274 1275 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1276 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1277 1278 #ifdef INET6 1279 static int 1280 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1281 { 1282 struct sockaddr_in6 addrs[2]; 1283 struct inpcb *inp; 1284 int error; 1285 boolean_t mapped = FALSE; 1286 1287 error = priv_check(req->td, PRIV_ROOT); 1288 if (error != 0) 1289 return (error); 1290 error = SYSCTL_IN(req, addrs, sizeof addrs); 1291 if (error != 0) 1292 return (error); 1293 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { 1294 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) 1295 mapped = TRUE; 1296 else 1297 return (EINVAL); 1298 } 1299 crit_enter(); 1300 if (mapped) { 1301 inp = in_pcblookup_hash(&tcbinfo[0], 1302 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], 1303 addrs[1].sin6_port, 1304 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], 1305 addrs[0].sin6_port, 1306 0, NULL); 1307 } else { 1308 inp = in6_pcblookup_hash(&tcbinfo[0], 1309 &addrs[1].sin6_addr, addrs[1].sin6_port, 1310 &addrs[0].sin6_addr, addrs[0].sin6_port, 1311 0, NULL); 1312 } 1313 if (inp == NULL || inp->inp_socket == NULL) { 1314 error = ENOENT; 1315 goto out; 1316 } 1317 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1318 out: 1319 crit_exit(); 1320 return (error); 1321 } 1322 1323 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1324 0, 0, 1325 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1326 #endif 1327 1328 struct netmsg_tcp_notify { 1329 struct netmsg_base base; 1330 void (*nm_notify)(struct inpcb *, int); 1331 struct in_addr nm_faddr; 1332 int nm_arg; 1333 }; 1334 1335 static void 1336 tcp_notifyall_oncpu(netmsg_t msg) 1337 { 1338 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg; 1339 int nextcpu; 1340 1341 in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nm->nm_faddr, 1342 nm->nm_arg, nm->nm_notify); 1343 1344 nextcpu = mycpuid + 1; 1345 if (nextcpu < ncpus2) 1346 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nm->base.lmsg); 1347 else 1348 lwkt_replymsg(&nm->base.lmsg, 0); 1349 } 1350 1351 void 1352 tcp_ctlinput(netmsg_t msg) 1353 { 1354 int cmd = msg->ctlinput.nm_cmd; 1355 struct sockaddr *sa = msg->ctlinput.nm_arg; 1356 struct ip *ip = msg->ctlinput.nm_extra; 1357 struct tcphdr *th; 1358 struct in_addr faddr; 1359 struct inpcb *inp; 1360 struct tcpcb *tp; 1361 void (*notify)(struct inpcb *, int) = tcp_notify; 1362 tcp_seq icmpseq; 1363 int arg, cpu; 1364 1365 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) { 1366 goto done; 1367 } 1368 1369 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1370 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1371 goto done; 1372 1373 arg = inetctlerrmap[cmd]; 1374 if (cmd == PRC_QUENCH) { 1375 notify = tcp_quench; 1376 } else if (icmp_may_rst && 1377 (cmd == PRC_UNREACH_ADMIN_PROHIB || 1378 cmd == PRC_UNREACH_PORT || 1379 cmd == PRC_TIMXCEED_INTRANS) && 1380 ip != NULL) { 1381 notify = tcp_drop_syn_sent; 1382 } else if (cmd == PRC_MSGSIZE) { 1383 struct icmp *icmp = (struct icmp *) 1384 ((caddr_t)ip - offsetof(struct icmp, icmp_ip)); 1385 1386 arg = ntohs(icmp->icmp_nextmtu); 1387 notify = tcp_mtudisc; 1388 } else if (PRC_IS_REDIRECT(cmd)) { 1389 ip = NULL; 1390 notify = in_rtchange; 1391 } else if (cmd == PRC_HOSTDEAD) { 1392 ip = NULL; 1393 } 1394 1395 if (ip != NULL) { 1396 crit_enter(); 1397 th = (struct tcphdr *)((caddr_t)ip + 1398 (IP_VHL_HL(ip->ip_vhl) << 2)); 1399 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport, 1400 ip->ip_src.s_addr, th->th_sport); 1401 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport, 1402 ip->ip_src, th->th_sport, 0, NULL); 1403 if ((inp != NULL) && (inp->inp_socket != NULL)) { 1404 icmpseq = htonl(th->th_seq); 1405 tp = intotcpcb(inp); 1406 if (SEQ_GEQ(icmpseq, tp->snd_una) && 1407 SEQ_LT(icmpseq, tp->snd_max)) 1408 (*notify)(inp, arg); 1409 } else { 1410 struct in_conninfo inc; 1411 1412 inc.inc_fport = th->th_dport; 1413 inc.inc_lport = th->th_sport; 1414 inc.inc_faddr = faddr; 1415 inc.inc_laddr = ip->ip_src; 1416 #ifdef INET6 1417 inc.inc_isipv6 = 0; 1418 #endif 1419 syncache_unreach(&inc, th); 1420 } 1421 crit_exit(); 1422 } else { 1423 struct netmsg_tcp_notify *nm; 1424 1425 KKASSERT(&curthread->td_msgport == netisr_cpuport(0)); 1426 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT); 1427 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1428 0, tcp_notifyall_oncpu); 1429 nm->nm_faddr = faddr; 1430 nm->nm_arg = arg; 1431 nm->nm_notify = notify; 1432 1433 lwkt_sendmsg(netisr_cpuport(0), &nm->base.lmsg); 1434 } 1435 done: 1436 lwkt_replymsg(&msg->lmsg, 0); 1437 } 1438 1439 #ifdef INET6 1440 1441 void 1442 tcp6_ctlinput(netmsg_t msg) 1443 { 1444 int cmd = msg->ctlinput.nm_cmd; 1445 struct sockaddr *sa = msg->ctlinput.nm_arg; 1446 void *d = msg->ctlinput.nm_extra; 1447 struct tcphdr th; 1448 void (*notify) (struct inpcb *, int) = tcp_notify; 1449 struct ip6_hdr *ip6; 1450 struct mbuf *m; 1451 struct ip6ctlparam *ip6cp = NULL; 1452 const struct sockaddr_in6 *sa6_src = NULL; 1453 int off; 1454 struct tcp_portonly { 1455 u_int16_t th_sport; 1456 u_int16_t th_dport; 1457 } *thp; 1458 int arg; 1459 1460 if (sa->sa_family != AF_INET6 || 1461 sa->sa_len != sizeof(struct sockaddr_in6)) { 1462 goto out; 1463 } 1464 1465 arg = 0; 1466 if (cmd == PRC_QUENCH) 1467 notify = tcp_quench; 1468 else if (cmd == PRC_MSGSIZE) { 1469 struct ip6ctlparam *ip6cp = d; 1470 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6; 1471 1472 arg = ntohl(icmp6->icmp6_mtu); 1473 notify = tcp_mtudisc; 1474 } else if (!PRC_IS_REDIRECT(cmd) && 1475 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) { 1476 goto out; 1477 } 1478 1479 /* if the parameter is from icmp6, decode it. */ 1480 if (d != NULL) { 1481 ip6cp = (struct ip6ctlparam *)d; 1482 m = ip6cp->ip6c_m; 1483 ip6 = ip6cp->ip6c_ip6; 1484 off = ip6cp->ip6c_off; 1485 sa6_src = ip6cp->ip6c_src; 1486 } else { 1487 m = NULL; 1488 ip6 = NULL; 1489 off = 0; /* fool gcc */ 1490 sa6_src = &sa6_any; 1491 } 1492 1493 if (ip6 != NULL) { 1494 struct in_conninfo inc; 1495 /* 1496 * XXX: We assume that when IPV6 is non NULL, 1497 * M and OFF are valid. 1498 */ 1499 1500 /* check if we can safely examine src and dst ports */ 1501 if (m->m_pkthdr.len < off + sizeof *thp) 1502 goto out; 1503 1504 bzero(&th, sizeof th); 1505 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1506 1507 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport, 1508 (struct sockaddr *)ip6cp->ip6c_src, 1509 th.th_sport, cmd, arg, notify); 1510 1511 inc.inc_fport = th.th_dport; 1512 inc.inc_lport = th.th_sport; 1513 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1514 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1515 inc.inc_isipv6 = 1; 1516 syncache_unreach(&inc, &th); 1517 } else { 1518 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0, 1519 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify); 1520 } 1521 out: 1522 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0); 1523 } 1524 1525 #endif 1526 1527 /* 1528 * Following is where TCP initial sequence number generation occurs. 1529 * 1530 * There are two places where we must use initial sequence numbers: 1531 * 1. In SYN-ACK packets. 1532 * 2. In SYN packets. 1533 * 1534 * All ISNs for SYN-ACK packets are generated by the syncache. See 1535 * tcp_syncache.c for details. 1536 * 1537 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1538 * depends on this property. In addition, these ISNs should be 1539 * unguessable so as to prevent connection hijacking. To satisfy 1540 * the requirements of this situation, the algorithm outlined in 1541 * RFC 1948 is used to generate sequence numbers. 1542 * 1543 * Implementation details: 1544 * 1545 * Time is based off the system timer, and is corrected so that it 1546 * increases by one megabyte per second. This allows for proper 1547 * recycling on high speed LANs while still leaving over an hour 1548 * before rollover. 1549 * 1550 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1551 * between seeding of isn_secret. This is normally set to zero, 1552 * as reseeding should not be necessary. 1553 * 1554 */ 1555 1556 #define ISN_BYTES_PER_SECOND 1048576 1557 1558 u_char isn_secret[32]; 1559 int isn_last_reseed; 1560 MD5_CTX isn_ctx; 1561 1562 tcp_seq 1563 tcp_new_isn(struct tcpcb *tp) 1564 { 1565 u_int32_t md5_buffer[4]; 1566 tcp_seq new_isn; 1567 1568 /* Seed if this is the first use, reseed if requested. */ 1569 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1570 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1571 < (u_int)ticks))) { 1572 read_random_unlimited(&isn_secret, sizeof isn_secret); 1573 isn_last_reseed = ticks; 1574 } 1575 1576 /* Compute the md5 hash and return the ISN. */ 1577 MD5Init(&isn_ctx); 1578 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1579 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1580 #ifdef INET6 1581 if (tp->t_inpcb->inp_vflag & INP_IPV6) { 1582 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1583 sizeof(struct in6_addr)); 1584 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1585 sizeof(struct in6_addr)); 1586 } else 1587 #endif 1588 { 1589 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1590 sizeof(struct in_addr)); 1591 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1592 sizeof(struct in_addr)); 1593 } 1594 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1595 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1596 new_isn = (tcp_seq) md5_buffer[0]; 1597 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1598 return (new_isn); 1599 } 1600 1601 /* 1602 * When a source quench is received, close congestion window 1603 * to one segment. We will gradually open it again as we proceed. 1604 */ 1605 void 1606 tcp_quench(struct inpcb *inp, int error) 1607 { 1608 struct tcpcb *tp = intotcpcb(inp); 1609 1610 if (tp != NULL) { 1611 tp->snd_cwnd = tp->t_maxseg; 1612 tp->snd_wacked = 0; 1613 } 1614 } 1615 1616 /* 1617 * When a specific ICMP unreachable message is received and the 1618 * connection state is SYN-SENT, drop the connection. This behavior 1619 * is controlled by the icmp_may_rst sysctl. 1620 */ 1621 void 1622 tcp_drop_syn_sent(struct inpcb *inp, int error) 1623 { 1624 struct tcpcb *tp = intotcpcb(inp); 1625 1626 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT)) 1627 tcp_drop(tp, error); 1628 } 1629 1630 /* 1631 * When a `need fragmentation' ICMP is received, update our idea of the MSS 1632 * based on the new value in the route. Also nudge TCP to send something, 1633 * since we know the packet we just sent was dropped. 1634 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1635 */ 1636 void 1637 tcp_mtudisc(struct inpcb *inp, int mtu) 1638 { 1639 struct tcpcb *tp = intotcpcb(inp); 1640 struct rtentry *rt; 1641 struct socket *so = inp->inp_socket; 1642 int maxopd, mss; 1643 #ifdef INET6 1644 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0); 1645 #else 1646 const boolean_t isipv6 = FALSE; 1647 #endif 1648 1649 if (tp == NULL) 1650 return; 1651 1652 /* 1653 * If no MTU is provided in the ICMP message, use the 1654 * next lower likely value, as specified in RFC 1191. 1655 */ 1656 if (mtu == 0) { 1657 int oldmtu; 1658 1659 oldmtu = tp->t_maxopd + 1660 (isipv6 ? 1661 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1662 sizeof(struct tcpiphdr)); 1663 mtu = ip_next_mtu(oldmtu, 0); 1664 } 1665 1666 if (isipv6) 1667 rt = tcp_rtlookup6(&inp->inp_inc); 1668 else 1669 rt = tcp_rtlookup(&inp->inp_inc); 1670 if (rt != NULL) { 1671 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu) 1672 mtu = rt->rt_rmx.rmx_mtu; 1673 1674 maxopd = mtu - 1675 (isipv6 ? 1676 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1677 sizeof(struct tcpiphdr)); 1678 1679 /* 1680 * XXX - The following conditional probably violates the TCP 1681 * spec. The problem is that, since we don't know the 1682 * other end's MSS, we are supposed to use a conservative 1683 * default. But, if we do that, then MTU discovery will 1684 * never actually take place, because the conservative 1685 * default is much less than the MTUs typically seen 1686 * on the Internet today. For the moment, we'll sweep 1687 * this under the carpet. 1688 * 1689 * The conservative default might not actually be a problem 1690 * if the only case this occurs is when sending an initial 1691 * SYN with options and data to a host we've never talked 1692 * to before. Then, they will reply with an MSS value which 1693 * will get recorded and the new parameters should get 1694 * recomputed. For Further Study. 1695 */ 1696 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd) 1697 maxopd = rt->rt_rmx.rmx_mssopt; 1698 } else 1699 maxopd = mtu - 1700 (isipv6 ? 1701 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1702 sizeof(struct tcpiphdr)); 1703 1704 if (tp->t_maxopd <= maxopd) 1705 return; 1706 tp->t_maxopd = maxopd; 1707 1708 mss = maxopd; 1709 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) == 1710 (TF_REQ_TSTMP | TF_RCVD_TSTMP)) 1711 mss -= TCPOLEN_TSTAMP_APPA; 1712 1713 /* round down to multiple of MCLBYTES */ 1714 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */ 1715 if (mss > MCLBYTES) 1716 mss &= ~(MCLBYTES - 1); 1717 #else 1718 if (mss > MCLBYTES) 1719 mss = (mss / MCLBYTES) * MCLBYTES; 1720 #endif 1721 1722 if (so->so_snd.ssb_hiwat < mss) 1723 mss = so->so_snd.ssb_hiwat; 1724 1725 tp->t_maxseg = mss; 1726 tp->t_rtttime = 0; 1727 tp->snd_nxt = tp->snd_una; 1728 tcp_output(tp); 1729 tcpstat.tcps_mturesent++; 1730 } 1731 1732 /* 1733 * Look-up the routing entry to the peer of this inpcb. If no route 1734 * is found and it cannot be allocated the return NULL. This routine 1735 * is called by TCP routines that access the rmx structure and by tcp_mss 1736 * to get the interface MTU. 1737 */ 1738 struct rtentry * 1739 tcp_rtlookup(struct in_conninfo *inc) 1740 { 1741 struct route *ro = &inc->inc_route; 1742 1743 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1744 /* No route yet, so try to acquire one */ 1745 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1746 /* 1747 * unused portions of the structure MUST be zero'd 1748 * out because rtalloc() treats it as opaque data 1749 */ 1750 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1751 ro->ro_dst.sa_family = AF_INET; 1752 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1753 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1754 inc->inc_faddr; 1755 rtalloc(ro); 1756 } 1757 } 1758 return (ro->ro_rt); 1759 } 1760 1761 #ifdef INET6 1762 struct rtentry * 1763 tcp_rtlookup6(struct in_conninfo *inc) 1764 { 1765 struct route_in6 *ro6 = &inc->inc6_route; 1766 1767 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1768 /* No route yet, so try to acquire one */ 1769 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1770 /* 1771 * unused portions of the structure MUST be zero'd 1772 * out because rtalloc() treats it as opaque data 1773 */ 1774 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1775 ro6->ro_dst.sin6_family = AF_INET6; 1776 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1777 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1778 rtalloc((struct route *)ro6); 1779 } 1780 } 1781 return (ro6->ro_rt); 1782 } 1783 #endif 1784 1785 #ifdef IPSEC 1786 /* compute ESP/AH header size for TCP, including outer IP header. */ 1787 size_t 1788 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1789 { 1790 struct inpcb *inp; 1791 struct mbuf *m; 1792 size_t hdrsiz; 1793 struct ip *ip; 1794 struct tcphdr *th; 1795 1796 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1797 return (0); 1798 MGETHDR(m, MB_DONTWAIT, MT_DATA); 1799 if (!m) 1800 return (0); 1801 1802 #ifdef INET6 1803 if (inp->inp_vflag & INP_IPV6) { 1804 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1805 1806 th = (struct tcphdr *)(ip6 + 1); 1807 m->m_pkthdr.len = m->m_len = 1808 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1809 tcp_fillheaders(tp, ip6, th, FALSE); 1810 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1811 } else 1812 #endif 1813 { 1814 ip = mtod(m, struct ip *); 1815 th = (struct tcphdr *)(ip + 1); 1816 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1817 tcp_fillheaders(tp, ip, th, FALSE); 1818 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1819 } 1820 1821 m_free(m); 1822 return (hdrsiz); 1823 } 1824 #endif 1825 1826 /* 1827 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1828 * 1829 * This code attempts to calculate the bandwidth-delay product as a 1830 * means of determining the optimal window size to maximize bandwidth, 1831 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1832 * routers. This code also does a fairly good job keeping RTTs in check 1833 * across slow links like modems. We implement an algorithm which is very 1834 * similar (but not meant to be) TCP/Vegas. The code operates on the 1835 * transmitter side of a TCP connection and so only effects the transmit 1836 * side of the connection. 1837 * 1838 * BACKGROUND: TCP makes no provision for the management of buffer space 1839 * at the end points or at the intermediate routers and switches. A TCP 1840 * stream, whether using NewReno or not, will eventually buffer as 1841 * many packets as it is able and the only reason this typically works is 1842 * due to the fairly small default buffers made available for a connection 1843 * (typicaly 16K or 32K). As machines use larger windows and/or window 1844 * scaling it is now fairly easy for even a single TCP connection to blow-out 1845 * all available buffer space not only on the local interface, but on 1846 * intermediate routers and switches as well. NewReno makes a misguided 1847 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1848 * then backing off, then steadily increasing the window again until another 1849 * failure occurs, ad-infinitum. This results in terrible oscillation that 1850 * is only made worse as network loads increase and the idea of intentionally 1851 * blowing out network buffers is, frankly, a terrible way to manage network 1852 * resources. 1853 * 1854 * It is far better to limit the transmit window prior to the failure 1855 * condition being achieved. There are two general ways to do this: First 1856 * you can 'scan' through different transmit window sizes and locate the 1857 * point where the RTT stops increasing, indicating that you have filled the 1858 * pipe, then scan backwards until you note that RTT stops decreasing, then 1859 * repeat ad-infinitum. This method works in principle but has severe 1860 * implementation issues due to RTT variances, timer granularity, and 1861 * instability in the algorithm which can lead to many false positives and 1862 * create oscillations as well as interact badly with other TCP streams 1863 * implementing the same algorithm. 1864 * 1865 * The second method is to limit the window to the bandwidth delay product 1866 * of the link. This is the method we implement. RTT variances and our 1867 * own manipulation of the congestion window, bwnd, can potentially 1868 * destabilize the algorithm. For this reason we have to stabilize the 1869 * elements used to calculate the window. We do this by using the minimum 1870 * observed RTT, the long term average of the observed bandwidth, and 1871 * by adding two segments worth of slop. It isn't perfect but it is able 1872 * to react to changing conditions and gives us a very stable basis on 1873 * which to extend the algorithm. 1874 */ 1875 void 1876 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1877 { 1878 u_long bw; 1879 u_long bwnd; 1880 int save_ticks; 1881 int delta_ticks; 1882 1883 /* 1884 * If inflight_enable is disabled in the middle of a tcp connection, 1885 * make sure snd_bwnd is effectively disabled. 1886 */ 1887 if (!tcp_inflight_enable) { 1888 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1889 tp->snd_bandwidth = 0; 1890 return; 1891 } 1892 1893 /* 1894 * Validate the delta time. If a connection is new or has been idle 1895 * a long time we have to reset the bandwidth calculator. 1896 */ 1897 save_ticks = ticks; 1898 delta_ticks = save_ticks - tp->t_bw_rtttime; 1899 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1900 tp->t_bw_rtttime = ticks; 1901 tp->t_bw_rtseq = ack_seq; 1902 if (tp->snd_bandwidth == 0) 1903 tp->snd_bandwidth = tcp_inflight_min; 1904 return; 1905 } 1906 if (delta_ticks == 0) 1907 return; 1908 1909 /* 1910 * Sanity check, plus ignore pure window update acks. 1911 */ 1912 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1913 return; 1914 1915 /* 1916 * Figure out the bandwidth. Due to the tick granularity this 1917 * is a very rough number and it MUST be averaged over a fairly 1918 * long period of time. XXX we need to take into account a link 1919 * that is not using all available bandwidth, but for now our 1920 * slop will ramp us up if this case occurs and the bandwidth later 1921 * increases. 1922 */ 1923 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 1924 tp->t_bw_rtttime = save_ticks; 1925 tp->t_bw_rtseq = ack_seq; 1926 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4; 1927 1928 tp->snd_bandwidth = bw; 1929 1930 /* 1931 * Calculate the semi-static bandwidth delay product, plus two maximal 1932 * segments. The additional slop puts us squarely in the sweet 1933 * spot and also handles the bandwidth run-up case. Without the 1934 * slop we could be locking ourselves into a lower bandwidth. 1935 * 1936 * Situations Handled: 1937 * (1) Prevents over-queueing of packets on LANs, especially on 1938 * high speed LANs, allowing larger TCP buffers to be 1939 * specified, and also does a good job preventing 1940 * over-queueing of packets over choke points like modems 1941 * (at least for the transmit side). 1942 * 1943 * (2) Is able to handle changing network loads (bandwidth 1944 * drops so bwnd drops, bandwidth increases so bwnd 1945 * increases). 1946 * 1947 * (3) Theoretically should stabilize in the face of multiple 1948 * connections implementing the same algorithm (this may need 1949 * a little work). 1950 * 1951 * (4) Stability value (defaults to 20 = 2 maximal packets) can 1952 * be adjusted with a sysctl but typically only needs to be on 1953 * very slow connections. A value no smaller then 5 should 1954 * be used, but only reduce this default if you have no other 1955 * choice. 1956 */ 1957 1958 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2) 1959 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 1960 tcp_inflight_stab * (int)tp->t_maxseg / 10; 1961 #undef USERTT 1962 1963 if (tcp_inflight_debug > 0) { 1964 static int ltime; 1965 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) { 1966 ltime = ticks; 1967 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n", 1968 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd); 1969 } 1970 } 1971 if ((long)bwnd < tcp_inflight_min) 1972 bwnd = tcp_inflight_min; 1973 if (bwnd > tcp_inflight_max) 1974 bwnd = tcp_inflight_max; 1975 if ((long)bwnd < tp->t_maxseg * 2) 1976 bwnd = tp->t_maxseg * 2; 1977 tp->snd_bwnd = bwnd; 1978 } 1979 1980 static void 1981 tcp_rmx_iwsegs(struct tcpcb *tp, u_long *maxsegs, u_long *capsegs) 1982 { 1983 struct rtentry *rt; 1984 struct inpcb *inp = tp->t_inpcb; 1985 #ifdef INET6 1986 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE); 1987 #else 1988 const boolean_t isipv6 = FALSE; 1989 #endif 1990 1991 /* XXX */ 1992 if (tcp_iw_maxsegs < TCP_IW_MAXSEGS_DFLT) 1993 tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 1994 if (tcp_iw_capsegs < TCP_IW_CAPSEGS_DFLT) 1995 tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 1996 1997 if (isipv6) 1998 rt = tcp_rtlookup6(&inp->inp_inc); 1999 else 2000 rt = tcp_rtlookup(&inp->inp_inc); 2001 if (rt == NULL || 2002 rt->rt_rmx.rmx_iwmaxsegs < TCP_IW_MAXSEGS_DFLT || 2003 rt->rt_rmx.rmx_iwcapsegs < TCP_IW_CAPSEGS_DFLT) { 2004 *maxsegs = tcp_iw_maxsegs; 2005 *capsegs = tcp_iw_capsegs; 2006 return; 2007 } 2008 *maxsegs = rt->rt_rmx.rmx_iwmaxsegs; 2009 *capsegs = rt->rt_rmx.rmx_iwcapsegs; 2010 } 2011 2012 u_long 2013 tcp_initial_window(struct tcpcb *tp) 2014 { 2015 if (tcp_do_rfc3390) { 2016 /* 2017 * RFC3390: 2018 * "If the SYN or SYN/ACK is lost, the initial window 2019 * used by a sender after a correctly transmitted SYN 2020 * MUST be one segment consisting of MSS bytes." 2021 * 2022 * However, we do something a little bit more aggressive 2023 * then RFC3390 here: 2024 * - Only if time spent in the SYN or SYN|ACK retransmition 2025 * >= 3 seconds, the IW is reduced. We do this mainly 2026 * because when RFC3390 is published, the initial RTO is 2027 * still 3 seconds (the threshold we test here), while 2028 * after RFC6298, the initial RTO is 1 second. This 2029 * behaviour probably still falls within the spirit of 2030 * RFC3390. 2031 * - When IW is reduced, 2*MSS is used instead of 1*MSS. 2032 * Mainly to avoid sender and receiver deadlock until 2033 * delayed ACK timer expires. And even RFC2581 does not 2034 * try to reduce IW upon SYN or SYN|ACK retransmition 2035 * timeout. 2036 * 2037 * See also: 2038 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03 2039 */ 2040 if (tp->t_rxtsyn >= TCPTV_RTOBASE3) { 2041 return (2 * tp->t_maxseg); 2042 } else { 2043 u_long maxsegs, capsegs; 2044 2045 tcp_rmx_iwsegs(tp, &maxsegs, &capsegs); 2046 return min(maxsegs * tp->t_maxseg, 2047 max(2 * tp->t_maxseg, capsegs * 1460)); 2048 } 2049 } else { 2050 /* 2051 * Even RFC2581 (back to 1999) allows 2*SMSS IW. 2052 * 2053 * Mainly to avoid sender and receiver deadlock 2054 * until delayed ACK timer expires. 2055 */ 2056 return (2 * tp->t_maxseg); 2057 } 2058 } 2059 2060 #ifdef TCP_SIGNATURE 2061 /* 2062 * Compute TCP-MD5 hash of a TCP segment. (RFC2385) 2063 * 2064 * We do this over ip, tcphdr, segment data, and the key in the SADB. 2065 * When called from tcp_input(), we can be sure that th_sum has been 2066 * zeroed out and verified already. 2067 * 2068 * Return 0 if successful, otherwise return -1. 2069 * 2070 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a 2071 * search with the destination IP address, and a 'magic SPI' to be 2072 * determined by the application. This is hardcoded elsewhere to 1179 2073 * right now. Another branch of this code exists which uses the SPD to 2074 * specify per-application flows but it is unstable. 2075 */ 2076 int 2077 tcpsignature_compute( 2078 struct mbuf *m, /* mbuf chain */ 2079 int len, /* length of TCP data */ 2080 int optlen, /* length of TCP options */ 2081 u_char *buf, /* storage for MD5 digest */ 2082 u_int direction) /* direction of flow */ 2083 { 2084 struct ippseudo ippseudo; 2085 MD5_CTX ctx; 2086 int doff; 2087 struct ip *ip; 2088 struct ipovly *ipovly; 2089 struct secasvar *sav; 2090 struct tcphdr *th; 2091 #ifdef INET6 2092 struct ip6_hdr *ip6; 2093 struct in6_addr in6; 2094 uint32_t plen; 2095 uint16_t nhdr; 2096 #endif /* INET6 */ 2097 u_short savecsum; 2098 2099 KASSERT(m != NULL, ("passed NULL mbuf. Game over.")); 2100 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature")); 2101 /* 2102 * Extract the destination from the IP header in the mbuf. 2103 */ 2104 ip = mtod(m, struct ip *); 2105 #ifdef INET6 2106 ip6 = NULL; /* Make the compiler happy. */ 2107 #endif /* INET6 */ 2108 /* 2109 * Look up an SADB entry which matches the address found in 2110 * the segment. 2111 */ 2112 switch (IP_VHL_V(ip->ip_vhl)) { 2113 case IPVERSION: 2114 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst, 2115 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2116 break; 2117 #ifdef INET6 2118 case (IPV6_VERSION >> 4): 2119 ip6 = mtod(m, struct ip6_hdr *); 2120 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst, 2121 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2122 break; 2123 #endif /* INET6 */ 2124 default: 2125 return (EINVAL); 2126 /* NOTREACHED */ 2127 break; 2128 } 2129 if (sav == NULL) { 2130 kprintf("%s: SADB lookup failed\n", __func__); 2131 return (EINVAL); 2132 } 2133 MD5Init(&ctx); 2134 2135 /* 2136 * Step 1: Update MD5 hash with IP pseudo-header. 2137 * 2138 * XXX The ippseudo header MUST be digested in network byte order, 2139 * or else we'll fail the regression test. Assume all fields we've 2140 * been doing arithmetic on have been in host byte order. 2141 * XXX One cannot depend on ipovly->ih_len here. When called from 2142 * tcp_output(), the underlying ip_len member has not yet been set. 2143 */ 2144 switch (IP_VHL_V(ip->ip_vhl)) { 2145 case IPVERSION: 2146 ipovly = (struct ipovly *)ip; 2147 ippseudo.ippseudo_src = ipovly->ih_src; 2148 ippseudo.ippseudo_dst = ipovly->ih_dst; 2149 ippseudo.ippseudo_pad = 0; 2150 ippseudo.ippseudo_p = IPPROTO_TCP; 2151 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen); 2152 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo)); 2153 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip)); 2154 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen; 2155 break; 2156 #ifdef INET6 2157 /* 2158 * RFC 2385, 2.0 Proposal 2159 * For IPv6, the pseudo-header is as described in RFC 2460, namely the 2160 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero- 2161 * extended next header value (to form 32 bits), and 32-bit segment 2162 * length. 2163 * Note: Upper-Layer Packet Length comes before Next Header. 2164 */ 2165 case (IPV6_VERSION >> 4): 2166 in6 = ip6->ip6_src; 2167 in6_clearscope(&in6); 2168 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2169 in6 = ip6->ip6_dst; 2170 in6_clearscope(&in6); 2171 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2172 plen = htonl(len + sizeof(struct tcphdr) + optlen); 2173 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t)); 2174 nhdr = 0; 2175 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2176 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2177 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2178 nhdr = IPPROTO_TCP; 2179 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2180 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr)); 2181 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen; 2182 break; 2183 #endif /* INET6 */ 2184 default: 2185 return (EINVAL); 2186 /* NOTREACHED */ 2187 break; 2188 } 2189 /* 2190 * Step 2: Update MD5 hash with TCP header, excluding options. 2191 * The TCP checksum must be set to zero. 2192 */ 2193 savecsum = th->th_sum; 2194 th->th_sum = 0; 2195 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr)); 2196 th->th_sum = savecsum; 2197 /* 2198 * Step 3: Update MD5 hash with TCP segment data. 2199 * Use m_apply() to avoid an early m_pullup(). 2200 */ 2201 if (len > 0) 2202 m_apply(m, doff, len, tcpsignature_apply, &ctx); 2203 /* 2204 * Step 4: Update MD5 hash with shared secret. 2205 */ 2206 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth)); 2207 MD5Final(buf, &ctx); 2208 key_sa_recordxfer(sav, m); 2209 key_freesav(sav); 2210 return (0); 2211 } 2212 2213 int 2214 tcpsignature_apply(void *fstate, void *data, unsigned int len) 2215 { 2216 2217 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len); 2218 return (0); 2219 } 2220 #endif /* TCP_SIGNATURE */ 2221 2222 boolean_t 2223 tcp_tso_pullup(struct mbuf **mp, int hoff, struct ip **ip0, int *iphlen0, 2224 struct tcphdr **th0, int *thoff0) 2225 { 2226 struct mbuf *m = *mp; 2227 struct ip *ip; 2228 int len, iphlen; 2229 struct tcphdr *th; 2230 int thoff; 2231 2232 len = hoff + sizeof(struct ip); 2233 2234 /* The fixed IP header must reside completely in the first mbuf. */ 2235 if (m->m_len < len) { 2236 m = m_pullup(m, len); 2237 if (m == NULL) 2238 goto fail; 2239 } 2240 2241 ip = mtodoff(m, struct ip *, hoff); 2242 iphlen = IP_VHL_HL(ip->ip_vhl) << 2; 2243 2244 /* The full IP header must reside completely in the one mbuf. */ 2245 if (m->m_len < hoff + iphlen) { 2246 m = m_pullup(m, hoff + iphlen); 2247 if (m == NULL) 2248 goto fail; 2249 ip = mtodoff(m, struct ip *, hoff); 2250 } 2251 2252 KASSERT(ip->ip_p == IPPROTO_TCP, ("not tcp %d", ip->ip_p)); 2253 2254 if (m->m_len < hoff + iphlen + sizeof(struct tcphdr)) { 2255 m = m_pullup(m, hoff + iphlen + sizeof(struct tcphdr)); 2256 if (m == NULL) 2257 goto fail; 2258 ip = mtodoff(m, struct ip *, hoff); 2259 } 2260 2261 th = (struct tcphdr *)((caddr_t)ip + iphlen); 2262 thoff = th->th_off << 2; 2263 2264 if (m->m_len < hoff + iphlen + thoff) { 2265 m = m_pullup(m, hoff + iphlen + thoff); 2266 if (m == NULL) 2267 goto fail; 2268 ip = mtodoff(m, struct ip *, hoff); 2269 th = (struct tcphdr *)((caddr_t)ip + iphlen); 2270 } 2271 2272 *mp = m; 2273 *ip0 = ip; 2274 *iphlen0 = iphlen; 2275 *th0 = th; 2276 *thoff0 = thoff; 2277 return TRUE; 2278 2279 fail: 2280 if (m != NULL) 2281 m_freem(m); 2282 *mp = NULL; 2283 return FALSE; 2284 } 2285