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. All advertising materials mentioning features or use of this software 47 * must display the following acknowledgement: 48 * This product includes software developed by the University of 49 * California, Berkeley and its contributors. 50 * 4. Neither the name of the University nor the names of its contributors 51 * may be used to endorse or promote products derived from this software 52 * without specific prior written permission. 53 * 54 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 55 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 56 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 57 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 58 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 59 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 60 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 61 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 62 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 63 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 64 * SUCH DAMAGE. 65 * 66 * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95 67 * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $ 68 * $DragonFly: src/sys/netinet/tcp_subr.c,v 1.59 2007/07/06 11:59:03 sephe Exp $ 69 */ 70 71 #include "opt_compat.h" 72 #include "opt_inet6.h" 73 #include "opt_ipsec.h" 74 #include "opt_tcpdebug.h" 75 76 #include <sys/param.h> 77 #include <sys/systm.h> 78 #include <sys/callout.h> 79 #include <sys/kernel.h> 80 #include <sys/sysctl.h> 81 #include <sys/malloc.h> 82 #include <sys/mpipe.h> 83 #include <sys/mbuf.h> 84 #ifdef INET6 85 #include <sys/domain.h> 86 #endif 87 #include <sys/proc.h> 88 #include <sys/socket.h> 89 #include <sys/socketvar.h> 90 #include <sys/protosw.h> 91 #include <sys/random.h> 92 #include <sys/in_cksum.h> 93 #include <sys/ktr.h> 94 95 #include <vm/vm_zone.h> 96 97 #include <net/route.h> 98 #include <net/if.h> 99 #include <net/netisr.h> 100 101 #define _IP_VHL 102 #include <netinet/in.h> 103 #include <netinet/in_systm.h> 104 #include <netinet/ip.h> 105 #include <netinet/ip6.h> 106 #include <netinet/in_pcb.h> 107 #include <netinet6/in6_pcb.h> 108 #include <netinet/in_var.h> 109 #include <netinet/ip_var.h> 110 #include <netinet6/ip6_var.h> 111 #include <netinet/ip_icmp.h> 112 #ifdef INET6 113 #include <netinet/icmp6.h> 114 #endif 115 #include <netinet/tcp.h> 116 #include <netinet/tcp_fsm.h> 117 #include <netinet/tcp_seq.h> 118 #include <netinet/tcp_timer.h> 119 #include <netinet/tcp_var.h> 120 #include <netinet6/tcp6_var.h> 121 #include <netinet/tcpip.h> 122 #ifdef TCPDEBUG 123 #include <netinet/tcp_debug.h> 124 #endif 125 #include <netinet6/ip6protosw.h> 126 127 #ifdef IPSEC 128 #include <netinet6/ipsec.h> 129 #ifdef INET6 130 #include <netinet6/ipsec6.h> 131 #endif 132 #endif 133 134 #ifdef FAST_IPSEC 135 #include <netproto/ipsec/ipsec.h> 136 #ifdef INET6 137 #include <netproto/ipsec/ipsec6.h> 138 #endif 139 #define IPSEC 140 #endif 141 142 #include <sys/md5.h> 143 #include <sys/msgport2.h> 144 #include <machine/smp.h> 145 146 #include <net/netmsg2.h> 147 148 #if !defined(KTR_TCP) 149 #define KTR_TCP KTR_ALL 150 #endif 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 struct inpcbinfo tcbinfo[MAXCPU]; 158 struct tcpcbackqhead tcpcbackq[MAXCPU]; 159 160 int tcp_mssdflt = TCP_MSS; 161 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 162 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size"); 163 164 #ifdef INET6 165 int tcp_v6mssdflt = TCP6_MSS; 166 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW, 167 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6"); 168 #endif 169 170 #if 0 171 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ; 172 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW, 173 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time"); 174 #endif 175 176 int tcp_do_rfc1323 = 1; 177 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW, 178 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions"); 179 180 int tcp_do_rfc1644 = 0; 181 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1644, rfc1644, CTLFLAG_RW, 182 &tcp_do_rfc1644, 0, "Enable rfc1644 (TTCP) extensions"); 183 184 static int tcp_tcbhashsize = 0; 185 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD, 186 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable"); 187 188 static int do_tcpdrain = 1; 189 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0, 190 "Enable tcp_drain routine for extra help when low on mbufs"); 191 192 /* XXX JH */ 193 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD, 194 &tcbinfo[0].ipi_count, 0, "Number of active PCBs"); 195 196 static int icmp_may_rst = 1; 197 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0, 198 "Certain ICMP unreachable messages may abort connections in SYN_SENT"); 199 200 static int tcp_isn_reseed_interval = 0; 201 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW, 202 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret"); 203 204 /* 205 * TCP bandwidth limiting sysctls. Note that the default lower bound of 206 * 1024 exists only for debugging. A good production default would be 207 * something like 6100. 208 */ 209 static int tcp_inflight_enable = 0; 210 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW, 211 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting"); 212 213 static int tcp_inflight_debug = 0; 214 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW, 215 &tcp_inflight_debug, 0, "Debug TCP inflight calculations"); 216 217 static int tcp_inflight_min = 6144; 218 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW, 219 &tcp_inflight_min, 0, "Lower bound for TCP inflight window"); 220 221 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT; 222 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW, 223 &tcp_inflight_max, 0, "Upper bound for TCP inflight window"); 224 225 static int tcp_inflight_stab = 20; 226 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW, 227 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 2 packets)"); 228 229 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives"); 230 static struct malloc_pipe tcptemp_mpipe; 231 232 static void tcp_willblock(void); 233 static void tcp_cleartaocache (void); 234 static void tcp_notify (struct inpcb *, int); 235 236 struct tcp_stats tcpstats_percpu[MAXCPU]; 237 #ifdef SMP 238 static int 239 sysctl_tcpstats(SYSCTL_HANDLER_ARGS) 240 { 241 int cpu, error = 0; 242 243 for (cpu = 0; cpu < ncpus; ++cpu) { 244 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu], 245 sizeof(struct tcp_stats)))) 246 break; 247 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu], 248 sizeof(struct tcp_stats)))) 249 break; 250 } 251 252 return (error); 253 } 254 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW), 255 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics"); 256 #else 257 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW, 258 &tcpstat, tcp_stats, "TCP statistics"); 259 #endif 260 261 /* 262 * Target size of TCP PCB hash tables. Must be a power of two. 263 * 264 * Note that this can be overridden by the kernel environment 265 * variable net.inet.tcp.tcbhashsize 266 */ 267 #ifndef TCBHASHSIZE 268 #define TCBHASHSIZE 512 269 #endif 270 271 /* 272 * This is the actual shape of what we allocate using the zone 273 * allocator. Doing it this way allows us to protect both structures 274 * using the same generation count, and also eliminates the overhead 275 * of allocating tcpcbs separately. By hiding the structure here, 276 * we avoid changing most of the rest of the code (although it needs 277 * to be changed, eventually, for greater efficiency). 278 */ 279 #define ALIGNMENT 32 280 #define ALIGNM1 (ALIGNMENT - 1) 281 struct inp_tp { 282 union { 283 struct inpcb inp; 284 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1]; 285 } inp_tp_u; 286 struct tcpcb tcb; 287 struct callout inp_tp_rexmt, inp_tp_persist, inp_tp_keep, inp_tp_2msl; 288 struct callout inp_tp_delack; 289 }; 290 #undef ALIGNMENT 291 #undef ALIGNM1 292 293 /* 294 * Tcp initialization 295 */ 296 void 297 tcp_init(void) 298 { 299 struct inpcbporthead *porthashbase; 300 u_long porthashmask; 301 struct vm_zone *ipi_zone; 302 int hashsize = TCBHASHSIZE; 303 int cpu; 304 305 /* 306 * note: tcptemp is used for keepalives, and it is ok for an 307 * allocation to fail so do not specify MPF_INT. 308 */ 309 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp), 310 25, -1, 0, NULL); 311 312 tcp_ccgen = 1; 313 tcp_cleartaocache(); 314 315 tcp_delacktime = TCPTV_DELACK; 316 tcp_keepinit = TCPTV_KEEP_INIT; 317 tcp_keepidle = TCPTV_KEEP_IDLE; 318 tcp_keepintvl = TCPTV_KEEPINTVL; 319 tcp_maxpersistidle = TCPTV_KEEP_IDLE; 320 tcp_msl = TCPTV_MSL; 321 tcp_rexmit_min = TCPTV_MIN; 322 tcp_rexmit_slop = TCPTV_CPU_VAR; 323 324 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize); 325 if (!powerof2(hashsize)) { 326 kprintf("WARNING: TCB hash size not a power of 2\n"); 327 hashsize = 512; /* safe default */ 328 } 329 tcp_tcbhashsize = hashsize; 330 porthashbase = hashinit(hashsize, M_PCB, &porthashmask); 331 ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets, 332 ZONE_INTERRUPT, 0); 333 334 for (cpu = 0; cpu < ncpus2; cpu++) { 335 in_pcbinfo_init(&tcbinfo[cpu]); 336 tcbinfo[cpu].cpu = cpu; 337 tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB, 338 &tcbinfo[cpu].hashmask); 339 tcbinfo[cpu].porthashbase = porthashbase; 340 tcbinfo[cpu].porthashmask = porthashmask; 341 tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB, 342 &tcbinfo[cpu].wildcardhashmask); 343 tcbinfo[cpu].ipi_zone = ipi_zone; 344 TAILQ_INIT(&tcpcbackq[cpu]); 345 } 346 347 tcp_reass_maxseg = nmbclusters / 16; 348 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg); 349 350 #ifdef INET6 351 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) 352 #else 353 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) 354 #endif 355 if (max_protohdr < TCP_MINPROTOHDR) 356 max_protohdr = TCP_MINPROTOHDR; 357 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) 358 panic("tcp_init"); 359 #undef TCP_MINPROTOHDR 360 361 /* 362 * Initialize TCP statistics counters for each CPU. 363 */ 364 #ifdef SMP 365 for (cpu = 0; cpu < ncpus; ++cpu) { 366 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats)); 367 } 368 #else 369 bzero(&tcpstat, sizeof(struct tcp_stats)); 370 #endif 371 372 syncache_init(); 373 tcp_sack_init(); 374 tcp_thread_init(); 375 } 376 377 void 378 tcpmsg_service_loop(void *dummy) 379 { 380 struct netmsg *msg; 381 382 while ((msg = lwkt_waitport(&curthread->td_msgport, 0))) { 383 do { 384 logtcp(rxmsg); 385 msg->nm_dispatch(msg); 386 } while ((msg = lwkt_getport(&curthread->td_msgport)) != NULL); 387 logtcp(delayed); 388 tcp_willblock(); 389 logtcp(wait); 390 } 391 } 392 393 static void 394 tcp_willblock(void) 395 { 396 struct tcpcb *tp; 397 int cpu = mycpu->gd_cpuid; 398 399 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) { 400 KKASSERT(tp->t_flags & TF_ONOUTPUTQ); 401 tp->t_flags &= ~TF_ONOUTPUTQ; 402 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq); 403 tcp_output(tp); 404 } 405 } 406 407 408 /* 409 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. 410 * tcp_template used to store this data in mbufs, but we now recopy it out 411 * of the tcpcb each time to conserve mbufs. 412 */ 413 void 414 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr) 415 { 416 struct inpcb *inp = tp->t_inpcb; 417 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr; 418 419 #ifdef INET6 420 if (inp->inp_vflag & INP_IPV6) { 421 struct ip6_hdr *ip6; 422 423 ip6 = (struct ip6_hdr *)ip_ptr; 424 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | 425 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK); 426 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | 427 (IPV6_VERSION & IPV6_VERSION_MASK); 428 ip6->ip6_nxt = IPPROTO_TCP; 429 ip6->ip6_plen = sizeof(struct tcphdr); 430 ip6->ip6_src = inp->in6p_laddr; 431 ip6->ip6_dst = inp->in6p_faddr; 432 tcp_hdr->th_sum = 0; 433 } else 434 #endif 435 { 436 struct ip *ip = (struct ip *) ip_ptr; 437 438 ip->ip_vhl = IP_VHL_BORING; 439 ip->ip_tos = 0; 440 ip->ip_len = 0; 441 ip->ip_id = 0; 442 ip->ip_off = 0; 443 ip->ip_ttl = 0; 444 ip->ip_sum = 0; 445 ip->ip_p = IPPROTO_TCP; 446 ip->ip_src = inp->inp_laddr; 447 ip->ip_dst = inp->inp_faddr; 448 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr, 449 ip->ip_dst.s_addr, 450 htons(sizeof(struct tcphdr) + IPPROTO_TCP)); 451 } 452 453 tcp_hdr->th_sport = inp->inp_lport; 454 tcp_hdr->th_dport = inp->inp_fport; 455 tcp_hdr->th_seq = 0; 456 tcp_hdr->th_ack = 0; 457 tcp_hdr->th_x2 = 0; 458 tcp_hdr->th_off = 5; 459 tcp_hdr->th_flags = 0; 460 tcp_hdr->th_win = 0; 461 tcp_hdr->th_urp = 0; 462 } 463 464 /* 465 * Create template to be used to send tcp packets on a connection. 466 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only 467 * use for this function is in keepalives, which use tcp_respond. 468 */ 469 struct tcptemp * 470 tcp_maketemplate(struct tcpcb *tp) 471 { 472 struct tcptemp *tmp; 473 474 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL) 475 return (NULL); 476 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t); 477 return (tmp); 478 } 479 480 void 481 tcp_freetemplate(struct tcptemp *tmp) 482 { 483 mpipe_free(&tcptemp_mpipe, tmp); 484 } 485 486 /* 487 * Send a single message to the TCP at address specified by 488 * the given TCP/IP header. If m == NULL, then we make a copy 489 * of the tcpiphdr at ti and send directly to the addressed host. 490 * This is used to force keep alive messages out using the TCP 491 * template for a connection. If flags are given then we send 492 * a message back to the TCP which originated the * segment ti, 493 * and discard the mbuf containing it and any other attached mbufs. 494 * 495 * In any case the ack and sequence number of the transmitted 496 * segment are as specified by the parameters. 497 * 498 * NOTE: If m != NULL, then ti must point to *inside* the mbuf. 499 */ 500 void 501 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, 502 tcp_seq ack, tcp_seq seq, int flags) 503 { 504 int tlen; 505 int win = 0; 506 struct route *ro = NULL; 507 struct route sro; 508 struct ip *ip = ipgen; 509 struct tcphdr *nth; 510 int ipflags = 0; 511 struct route_in6 *ro6 = NULL; 512 struct route_in6 sro6; 513 struct ip6_hdr *ip6 = ipgen; 514 #ifdef INET6 515 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6); 516 #else 517 const boolean_t isipv6 = FALSE; 518 #endif 519 520 if (tp != NULL) { 521 if (!(flags & TH_RST)) { 522 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv); 523 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 524 win = (long)TCP_MAXWIN << tp->rcv_scale; 525 } 526 if (isipv6) 527 ro6 = &tp->t_inpcb->in6p_route; 528 else 529 ro = &tp->t_inpcb->inp_route; 530 } else { 531 if (isipv6) { 532 ro6 = &sro6; 533 bzero(ro6, sizeof *ro6); 534 } else { 535 ro = &sro; 536 bzero(ro, sizeof *ro); 537 } 538 } 539 if (m == NULL) { 540 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 541 if (m == NULL) 542 return; 543 tlen = 0; 544 m->m_data += max_linkhdr; 545 if (isipv6) { 546 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); 547 ip6 = mtod(m, struct ip6_hdr *); 548 nth = (struct tcphdr *)(ip6 + 1); 549 } else { 550 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip)); 551 ip = mtod(m, struct ip *); 552 nth = (struct tcphdr *)(ip + 1); 553 } 554 bcopy(th, nth, sizeof(struct tcphdr)); 555 flags = TH_ACK; 556 } else { 557 m_freem(m->m_next); 558 m->m_next = NULL; 559 m->m_data = (caddr_t)ipgen; 560 /* m_len is set later */ 561 tlen = 0; 562 #define xchg(a, b, type) { type t; t = a; a = b; b = t; } 563 if (isipv6) { 564 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); 565 nth = (struct tcphdr *)(ip6 + 1); 566 } else { 567 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long); 568 nth = (struct tcphdr *)(ip + 1); 569 } 570 if (th != nth) { 571 /* 572 * this is usually a case when an extension header 573 * exists between the IPv6 header and the 574 * TCP header. 575 */ 576 nth->th_sport = th->th_sport; 577 nth->th_dport = th->th_dport; 578 } 579 xchg(nth->th_dport, nth->th_sport, n_short); 580 #undef xchg 581 } 582 if (isipv6) { 583 ip6->ip6_flow = 0; 584 ip6->ip6_vfc = IPV6_VERSION; 585 ip6->ip6_nxt = IPPROTO_TCP; 586 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen)); 587 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 588 } else { 589 tlen += sizeof(struct tcpiphdr); 590 ip->ip_len = tlen; 591 ip->ip_ttl = ip_defttl; 592 } 593 m->m_len = tlen; 594 m->m_pkthdr.len = tlen; 595 m->m_pkthdr.rcvif = (struct ifnet *) NULL; 596 nth->th_seq = htonl(seq); 597 nth->th_ack = htonl(ack); 598 nth->th_x2 = 0; 599 nth->th_off = sizeof(struct tcphdr) >> 2; 600 nth->th_flags = flags; 601 if (tp != NULL) 602 nth->th_win = htons((u_short) (win >> tp->rcv_scale)); 603 else 604 nth->th_win = htons((u_short)win); 605 nth->th_urp = 0; 606 if (isipv6) { 607 nth->th_sum = 0; 608 nth->th_sum = in6_cksum(m, IPPROTO_TCP, 609 sizeof(struct ip6_hdr), 610 tlen - sizeof(struct ip6_hdr)); 611 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL, 612 (ro6 && ro6->ro_rt) ? 613 ro6->ro_rt->rt_ifp : NULL); 614 } else { 615 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 616 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); 617 m->m_pkthdr.csum_flags = CSUM_TCP; 618 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 619 } 620 #ifdef TCPDEBUG 621 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) 622 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); 623 #endif 624 if (isipv6) { 625 ip6_output(m, NULL, ro6, ipflags, NULL, NULL, 626 tp ? tp->t_inpcb : NULL); 627 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) { 628 RTFREE(ro6->ro_rt); 629 ro6->ro_rt = NULL; 630 } 631 } else { 632 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL); 633 if ((ro == &sro) && (ro->ro_rt != NULL)) { 634 RTFREE(ro->ro_rt); 635 ro->ro_rt = NULL; 636 } 637 } 638 } 639 640 /* 641 * Create a new TCP control block, making an 642 * empty reassembly queue and hooking it to the argument 643 * protocol control block. The `inp' parameter must have 644 * come from the zone allocator set up in tcp_init(). 645 */ 646 struct tcpcb * 647 tcp_newtcpcb(struct inpcb *inp) 648 { 649 struct inp_tp *it; 650 struct tcpcb *tp; 651 #ifdef INET6 652 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 653 #else 654 const boolean_t isipv6 = FALSE; 655 #endif 656 657 it = (struct inp_tp *)inp; 658 tp = &it->tcb; 659 bzero(tp, sizeof(struct tcpcb)); 660 LIST_INIT(&tp->t_segq); 661 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 662 663 /* Set up our timeouts. */ 664 callout_init(tp->tt_rexmt = &it->inp_tp_rexmt); 665 callout_init(tp->tt_persist = &it->inp_tp_persist); 666 callout_init(tp->tt_keep = &it->inp_tp_keep); 667 callout_init(tp->tt_2msl = &it->inp_tp_2msl); 668 callout_init(tp->tt_delack = &it->inp_tp_delack); 669 670 if (tcp_do_rfc1323) 671 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP); 672 if (tcp_do_rfc1644) 673 tp->t_flags |= TF_REQ_CC; 674 tp->t_inpcb = inp; /* XXX */ 675 tp->t_state = TCPS_CLOSED; 676 /* 677 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 678 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 679 * reasonable initial retransmit time. 680 */ 681 tp->t_srtt = TCPTV_SRTTBASE; 682 tp->t_rttvar = 683 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 684 tp->t_rttmin = tcp_rexmit_min; 685 tp->t_rxtcur = TCPTV_RTOBASE; 686 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 687 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 688 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 689 tp->t_rcvtime = ticks; 690 /* 691 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 692 * because the socket may be bound to an IPv6 wildcard address, 693 * which may match an IPv4-mapped IPv6 address. 694 */ 695 inp->inp_ip_ttl = ip_defttl; 696 inp->inp_ppcb = tp; 697 tcp_sack_tcpcb_init(tp); 698 return (tp); /* XXX */ 699 } 700 701 /* 702 * Drop a TCP connection, reporting the specified error. 703 * If connection is synchronized, then send a RST to peer. 704 */ 705 struct tcpcb * 706 tcp_drop(struct tcpcb *tp, int error) 707 { 708 struct socket *so = tp->t_inpcb->inp_socket; 709 710 if (TCPS_HAVERCVDSYN(tp->t_state)) { 711 tp->t_state = TCPS_CLOSED; 712 tcp_output(tp); 713 tcpstat.tcps_drops++; 714 } else 715 tcpstat.tcps_conndrops++; 716 if (error == ETIMEDOUT && tp->t_softerror) 717 error = tp->t_softerror; 718 so->so_error = error; 719 return (tcp_close(tp)); 720 } 721 722 #ifdef SMP 723 724 struct netmsg_remwildcard { 725 struct netmsg nm_netmsg; 726 struct inpcb *nm_inp; 727 struct inpcbinfo *nm_pcbinfo; 728 #if defined(INET6) 729 int nm_isinet6; 730 #else 731 int nm_unused01; 732 #endif 733 }; 734 735 /* 736 * Wildcard inpcb's on SMP boxes must be removed from all cpus before the 737 * inp can be detached. We do this by cycling through the cpus, ending up 738 * on the cpu controlling the inp last and then doing the disconnect. 739 */ 740 static void 741 in_pcbremwildcardhash_handler(struct netmsg *msg0) 742 { 743 struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0; 744 int cpu; 745 746 cpu = msg->nm_pcbinfo->cpu; 747 748 if (cpu == msg->nm_inp->inp_pcbinfo->cpu) { 749 /* note: detach removes any wildcard hash entry */ 750 #ifdef INET6 751 if (msg->nm_isinet6) 752 in6_pcbdetach(msg->nm_inp); 753 else 754 #endif 755 in_pcbdetach(msg->nm_inp); 756 lwkt_replymsg(&msg->nm_netmsg.nm_lmsg, 0); 757 } else { 758 in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo); 759 cpu = (cpu + 1) % ncpus2; 760 msg->nm_pcbinfo = &tcbinfo[cpu]; 761 lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg); 762 } 763 } 764 765 #endif 766 767 /* 768 * Close a TCP control block: 769 * discard all space held by the tcp 770 * discard internet protocol block 771 * wake up any sleepers 772 */ 773 struct tcpcb * 774 tcp_close(struct tcpcb *tp) 775 { 776 struct tseg_qent *q; 777 struct inpcb *inp = tp->t_inpcb; 778 struct socket *so = inp->inp_socket; 779 struct rtentry *rt; 780 boolean_t dosavessthresh; 781 #ifdef SMP 782 int cpu; 783 #endif 784 #ifdef INET6 785 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 786 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0); 787 #else 788 const boolean_t isipv6 = FALSE; 789 #endif 790 791 /* 792 * The tp is not instantly destroyed in the wildcard case. Setting 793 * the state to TCPS_TERMINATING will prevent the TCP stack from 794 * messing with it, though it should be noted that this change may 795 * not take effect on other cpus until we have chained the wildcard 796 * hash removal. 797 * 798 * XXX we currently depend on the BGL to synchronize the tp->t_state 799 * update and prevent other tcp protocol threads from accepting new 800 * connections on the listen socket we might be trying to close down. 801 */ 802 KKASSERT(tp->t_state != TCPS_TERMINATING); 803 tp->t_state = TCPS_TERMINATING; 804 805 /* 806 * Make sure that all of our timers are stopped before we 807 * delete the PCB. 808 */ 809 callout_stop(tp->tt_rexmt); 810 callout_stop(tp->tt_persist); 811 callout_stop(tp->tt_keep); 812 callout_stop(tp->tt_2msl); 813 callout_stop(tp->tt_delack); 814 815 if (tp->t_flags & TF_ONOUTPUTQ) { 816 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 817 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 818 tp->t_flags &= ~TF_ONOUTPUTQ; 819 } 820 821 /* 822 * If we got enough samples through the srtt filter, 823 * save the rtt and rttvar in the routing entry. 824 * 'Enough' is arbitrarily defined as the 16 samples. 825 * 16 samples is enough for the srtt filter to converge 826 * to within 5% of the correct value; fewer samples and 827 * we could save a very bogus rtt. 828 * 829 * Don't update the default route's characteristics and don't 830 * update anything that the user "locked". 831 */ 832 if (tp->t_rttupdated >= 16) { 833 u_long i = 0; 834 835 if (isipv6) { 836 struct sockaddr_in6 *sin6; 837 838 if ((rt = inp->in6p_route.ro_rt) == NULL) 839 goto no_valid_rt; 840 sin6 = (struct sockaddr_in6 *)rt_key(rt); 841 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 842 goto no_valid_rt; 843 } else 844 if ((rt = inp->inp_route.ro_rt) == NULL || 845 ((struct sockaddr_in *)rt_key(rt))-> 846 sin_addr.s_addr == INADDR_ANY) 847 goto no_valid_rt; 848 849 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 850 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 851 if (rt->rt_rmx.rmx_rtt && i) 852 /* 853 * filter this update to half the old & half 854 * the new values, converting scale. 855 * See route.h and tcp_var.h for a 856 * description of the scaling constants. 857 */ 858 rt->rt_rmx.rmx_rtt = 859 (rt->rt_rmx.rmx_rtt + i) / 2; 860 else 861 rt->rt_rmx.rmx_rtt = i; 862 tcpstat.tcps_cachedrtt++; 863 } 864 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 865 i = tp->t_rttvar * 866 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 867 if (rt->rt_rmx.rmx_rttvar && i) 868 rt->rt_rmx.rmx_rttvar = 869 (rt->rt_rmx.rmx_rttvar + i) / 2; 870 else 871 rt->rt_rmx.rmx_rttvar = i; 872 tcpstat.tcps_cachedrttvar++; 873 } 874 /* 875 * The old comment here said: 876 * update the pipelimit (ssthresh) if it has been updated 877 * already or if a pipesize was specified & the threshhold 878 * got below half the pipesize. I.e., wait for bad news 879 * before we start updating, then update on both good 880 * and bad news. 881 * 882 * But we want to save the ssthresh even if no pipesize is 883 * specified explicitly in the route, because such 884 * connections still have an implicit pipesize specified 885 * by the global tcp_sendspace. In the absence of a reliable 886 * way to calculate the pipesize, it will have to do. 887 */ 888 i = tp->snd_ssthresh; 889 if (rt->rt_rmx.rmx_sendpipe != 0) 890 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 891 else 892 dosavessthresh = (i < so->so_snd.ssb_hiwat/2); 893 if (dosavessthresh || 894 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 895 (rt->rt_rmx.rmx_ssthresh != 0))) { 896 /* 897 * convert the limit from user data bytes to 898 * packets then to packet data bytes. 899 */ 900 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 901 if (i < 2) 902 i = 2; 903 i *= tp->t_maxseg + 904 (isipv6 ? 905 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 906 sizeof(struct tcpiphdr)); 907 if (rt->rt_rmx.rmx_ssthresh) 908 rt->rt_rmx.rmx_ssthresh = 909 (rt->rt_rmx.rmx_ssthresh + i) / 2; 910 else 911 rt->rt_rmx.rmx_ssthresh = i; 912 tcpstat.tcps_cachedssthresh++; 913 } 914 } 915 916 no_valid_rt: 917 /* free the reassembly queue, if any */ 918 while((q = LIST_FIRST(&tp->t_segq)) != NULL) { 919 LIST_REMOVE(q, tqe_q); 920 m_freem(q->tqe_m); 921 FREE(q, M_TSEGQ); 922 tcp_reass_qsize--; 923 } 924 /* throw away SACK blocks in scoreboard*/ 925 if (TCP_DO_SACK(tp)) 926 tcp_sack_cleanup(&tp->scb); 927 928 inp->inp_ppcb = NULL; 929 soisdisconnected(so); 930 /* 931 * Discard the inp. In the SMP case a wildcard inp's hash (created 932 * by a listen socket or an INADDR_ANY udp socket) is replicated 933 * for each protocol thread and must be removed in the context of 934 * that thread. This is accomplished by chaining the message 935 * through the cpus. 936 * 937 * If the inp is not wildcarded we simply detach, which will remove 938 * the any hashes still present for this inp. 939 */ 940 #ifdef SMP 941 if (inp->inp_flags & INP_WILDCARD_MP) { 942 struct netmsg_remwildcard *msg; 943 944 cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2; 945 msg = kmalloc(sizeof(struct netmsg_remwildcard), 946 M_LWKTMSG, M_INTWAIT); 947 netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0, 948 in_pcbremwildcardhash_handler); 949 #ifdef INET6 950 msg->nm_isinet6 = isafinet6; 951 #endif 952 msg->nm_inp = inp; 953 msg->nm_pcbinfo = &tcbinfo[cpu]; 954 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg); 955 } else 956 #endif 957 { 958 /* note: detach removes any wildcard hash entry */ 959 #ifdef INET6 960 if (isafinet6) 961 in6_pcbdetach(inp); 962 else 963 #endif 964 in_pcbdetach(inp); 965 } 966 tcpstat.tcps_closed++; 967 return (NULL); 968 } 969 970 static __inline void 971 tcp_drain_oncpu(struct inpcbhead *head) 972 { 973 struct inpcb *inpb; 974 struct tcpcb *tcpb; 975 struct tseg_qent *te; 976 977 LIST_FOREACH(inpb, head, inp_list) { 978 if (inpb->inp_flags & INP_PLACEMARKER) 979 continue; 980 if ((tcpb = intotcpcb(inpb))) { 981 while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) { 982 LIST_REMOVE(te, tqe_q); 983 m_freem(te->tqe_m); 984 FREE(te, M_TSEGQ); 985 tcp_reass_qsize--; 986 } 987 } 988 } 989 } 990 991 #ifdef SMP 992 struct netmsg_tcp_drain { 993 struct netmsg nm_netmsg; 994 struct inpcbhead *nm_head; 995 }; 996 997 static void 998 tcp_drain_handler(netmsg_t netmsg) 999 { 1000 struct netmsg_tcp_drain *nm = (void *)netmsg; 1001 1002 tcp_drain_oncpu(nm->nm_head); 1003 lwkt_replymsg(&nm->nm_netmsg.nm_lmsg, 0); 1004 } 1005 #endif 1006 1007 void 1008 tcp_drain(void) 1009 { 1010 #ifdef SMP 1011 int cpu; 1012 #endif 1013 1014 if (!do_tcpdrain) 1015 return; 1016 1017 /* 1018 * Walk the tcpbs, if existing, and flush the reassembly queue, 1019 * if there is one... 1020 * XXX: The "Net/3" implementation doesn't imply that the TCP 1021 * reassembly queue should be flushed, but in a situation 1022 * where we're really low on mbufs, this is potentially 1023 * useful. 1024 */ 1025 #ifdef SMP 1026 for (cpu = 0; cpu < ncpus2; cpu++) { 1027 struct netmsg_tcp_drain *msg; 1028 1029 if (cpu == mycpu->gd_cpuid) { 1030 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead); 1031 } else { 1032 msg = kmalloc(sizeof(struct netmsg_tcp_drain), 1033 M_LWKTMSG, M_NOWAIT); 1034 if (msg == NULL) 1035 continue; 1036 netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0, 1037 tcp_drain_handler); 1038 msg->nm_head = &tcbinfo[cpu].pcblisthead; 1039 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg); 1040 } 1041 } 1042 #else 1043 tcp_drain_oncpu(&tcbinfo[0].pcblisthead); 1044 #endif 1045 } 1046 1047 /* 1048 * Notify a tcp user of an asynchronous error; 1049 * store error as soft error, but wake up user 1050 * (for now, won't do anything until can select for soft error). 1051 * 1052 * Do not wake up user since there currently is no mechanism for 1053 * reporting soft errors (yet - a kqueue filter may be added). 1054 */ 1055 static void 1056 tcp_notify(struct inpcb *inp, int error) 1057 { 1058 struct tcpcb *tp = intotcpcb(inp); 1059 1060 /* 1061 * Ignore some errors if we are hooked up. 1062 * If connection hasn't completed, has retransmitted several times, 1063 * and receives a second error, give up now. This is better 1064 * than waiting a long time to establish a connection that 1065 * can never complete. 1066 */ 1067 if (tp->t_state == TCPS_ESTABLISHED && 1068 (error == EHOSTUNREACH || error == ENETUNREACH || 1069 error == EHOSTDOWN)) { 1070 return; 1071 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1072 tp->t_softerror) 1073 tcp_drop(tp, error); 1074 else 1075 tp->t_softerror = error; 1076 #if 0 1077 wakeup(&so->so_timeo); 1078 sorwakeup(so); 1079 sowwakeup(so); 1080 #endif 1081 } 1082 1083 static int 1084 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1085 { 1086 int error, i, n; 1087 struct inpcb *marker; 1088 struct inpcb *inp; 1089 inp_gen_t gencnt; 1090 globaldata_t gd; 1091 int origcpu, ccpu; 1092 1093 error = 0; 1094 n = 0; 1095 1096 /* 1097 * The process of preparing the TCB list is too time-consuming and 1098 * resource-intensive to repeat twice on every request. 1099 */ 1100 if (req->oldptr == NULL) { 1101 for (ccpu = 0; ccpu < ncpus; ++ccpu) { 1102 gd = globaldata_find(ccpu); 1103 n += tcbinfo[gd->gd_cpuid].ipi_count; 1104 } 1105 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1106 return (0); 1107 } 1108 1109 if (req->newptr != NULL) 1110 return (EPERM); 1111 1112 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1113 marker->inp_flags |= INP_PLACEMARKER; 1114 1115 /* 1116 * OK, now we're committed to doing something. Run the inpcb list 1117 * for each cpu in the system and construct the output. Use a 1118 * list placemarker to deal with list changes occuring during 1119 * copyout blockages (but otherwise depend on being on the correct 1120 * cpu to avoid races). 1121 */ 1122 origcpu = mycpu->gd_cpuid; 1123 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) { 1124 globaldata_t rgd; 1125 caddr_t inp_ppcb; 1126 struct xtcpcb xt; 1127 int cpu_id; 1128 1129 cpu_id = (origcpu + ccpu) % ncpus; 1130 if ((smp_active_mask & (1 << cpu_id)) == 0) 1131 continue; 1132 rgd = globaldata_find(cpu_id); 1133 lwkt_setcpu_self(rgd); 1134 1135 gencnt = tcbinfo[cpu_id].ipi_gencnt; 1136 n = tcbinfo[cpu_id].ipi_count; 1137 1138 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list); 1139 i = 0; 1140 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1141 /* 1142 * process a snapshot of pcbs, ignoring placemarkers 1143 * and using our own to allow SYSCTL_OUT to block. 1144 */ 1145 LIST_REMOVE(marker, inp_list); 1146 LIST_INSERT_AFTER(inp, marker, inp_list); 1147 1148 if (inp->inp_flags & INP_PLACEMARKER) 1149 continue; 1150 if (inp->inp_gencnt > gencnt) 1151 continue; 1152 if (prison_xinpcb(req->td, inp)) 1153 continue; 1154 1155 xt.xt_len = sizeof xt; 1156 bcopy(inp, &xt.xt_inp, sizeof *inp); 1157 inp_ppcb = inp->inp_ppcb; 1158 if (inp_ppcb != NULL) 1159 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1160 else 1161 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1162 if (inp->inp_socket) 1163 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1164 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1165 break; 1166 ++i; 1167 } 1168 LIST_REMOVE(marker, inp_list); 1169 if (error == 0 && i < n) { 1170 bzero(&xt, sizeof xt); 1171 xt.xt_len = sizeof xt; 1172 while (i < n) { 1173 error = SYSCTL_OUT(req, &xt, sizeof xt); 1174 if (error) 1175 break; 1176 ++i; 1177 } 1178 } 1179 } 1180 1181 /* 1182 * Make sure we are on the same cpu we were on originally, since 1183 * higher level callers expect this. Also don't pollute caches with 1184 * migrated userland data by (eventually) returning to userland 1185 * on a different cpu. 1186 */ 1187 lwkt_setcpu_self(globaldata_find(origcpu)); 1188 kfree(marker, M_TEMP); 1189 return (error); 1190 } 1191 1192 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1193 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1194 1195 static int 1196 tcp_getcred(SYSCTL_HANDLER_ARGS) 1197 { 1198 struct sockaddr_in addrs[2]; 1199 struct inpcb *inp; 1200 int cpu; 1201 int error; 1202 1203 error = suser(req->td); 1204 if (error != 0) 1205 return (error); 1206 error = SYSCTL_IN(req, addrs, sizeof addrs); 1207 if (error != 0) 1208 return (error); 1209 crit_enter(); 1210 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1211 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1212 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1213 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1214 if (inp == NULL || inp->inp_socket == NULL) { 1215 error = ENOENT; 1216 goto out; 1217 } 1218 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1219 out: 1220 crit_exit(); 1221 return (error); 1222 } 1223 1224 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1225 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1226 1227 #ifdef INET6 1228 static int 1229 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1230 { 1231 struct sockaddr_in6 addrs[2]; 1232 struct inpcb *inp; 1233 int error; 1234 boolean_t mapped = FALSE; 1235 1236 error = suser(req->td); 1237 if (error != 0) 1238 return (error); 1239 error = SYSCTL_IN(req, addrs, sizeof addrs); 1240 if (error != 0) 1241 return (error); 1242 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { 1243 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) 1244 mapped = TRUE; 1245 else 1246 return (EINVAL); 1247 } 1248 crit_enter(); 1249 if (mapped) { 1250 inp = in_pcblookup_hash(&tcbinfo[0], 1251 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], 1252 addrs[1].sin6_port, 1253 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], 1254 addrs[0].sin6_port, 1255 0, NULL); 1256 } else { 1257 inp = in6_pcblookup_hash(&tcbinfo[0], 1258 &addrs[1].sin6_addr, addrs[1].sin6_port, 1259 &addrs[0].sin6_addr, addrs[0].sin6_port, 1260 0, NULL); 1261 } 1262 if (inp == NULL || inp->inp_socket == NULL) { 1263 error = ENOENT; 1264 goto out; 1265 } 1266 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1267 out: 1268 crit_exit(); 1269 return (error); 1270 } 1271 1272 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1273 0, 0, 1274 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1275 #endif 1276 1277 void 1278 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip) 1279 { 1280 struct ip *ip = vip; 1281 struct tcphdr *th; 1282 struct in_addr faddr; 1283 struct inpcb *inp; 1284 struct tcpcb *tp; 1285 void (*notify)(struct inpcb *, int) = tcp_notify; 1286 tcp_seq icmpseq; 1287 int arg, cpu; 1288 1289 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) { 1290 return; 1291 } 1292 1293 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1294 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1295 return; 1296 1297 arg = inetctlerrmap[cmd]; 1298 if (cmd == PRC_QUENCH) { 1299 notify = tcp_quench; 1300 } else if (icmp_may_rst && 1301 (cmd == PRC_UNREACH_ADMIN_PROHIB || 1302 cmd == PRC_UNREACH_PORT || 1303 cmd == PRC_TIMXCEED_INTRANS) && 1304 ip != NULL) { 1305 notify = tcp_drop_syn_sent; 1306 } else if (cmd == PRC_MSGSIZE) { 1307 struct icmp *icmp = (struct icmp *) 1308 ((caddr_t)ip - offsetof(struct icmp, icmp_ip)); 1309 1310 arg = ntohs(icmp->icmp_nextmtu); 1311 notify = tcp_mtudisc; 1312 } else if (PRC_IS_REDIRECT(cmd)) { 1313 ip = NULL; 1314 notify = in_rtchange; 1315 } else if (cmd == PRC_HOSTDEAD) { 1316 ip = NULL; 1317 } 1318 1319 if (ip != NULL) { 1320 crit_enter(); 1321 th = (struct tcphdr *)((caddr_t)ip + 1322 (IP_VHL_HL(ip->ip_vhl) << 2)); 1323 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport, 1324 ip->ip_src.s_addr, th->th_sport); 1325 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport, 1326 ip->ip_src, th->th_sport, 0, NULL); 1327 if ((inp != NULL) && (inp->inp_socket != NULL)) { 1328 icmpseq = htonl(th->th_seq); 1329 tp = intotcpcb(inp); 1330 if (SEQ_GEQ(icmpseq, tp->snd_una) && 1331 SEQ_LT(icmpseq, tp->snd_max)) 1332 (*notify)(inp, arg); 1333 } else { 1334 struct in_conninfo inc; 1335 1336 inc.inc_fport = th->th_dport; 1337 inc.inc_lport = th->th_sport; 1338 inc.inc_faddr = faddr; 1339 inc.inc_laddr = ip->ip_src; 1340 #ifdef INET6 1341 inc.inc_isipv6 = 0; 1342 #endif 1343 syncache_unreach(&inc, th); 1344 } 1345 crit_exit(); 1346 } else { 1347 for (cpu = 0; cpu < ncpus2; cpu++) { 1348 in_pcbnotifyall(&tcbinfo[cpu].pcblisthead, faddr, arg, 1349 notify); 1350 } 1351 } 1352 } 1353 1354 #ifdef INET6 1355 void 1356 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d) 1357 { 1358 struct tcphdr th; 1359 void (*notify) (struct inpcb *, int) = tcp_notify; 1360 struct ip6_hdr *ip6; 1361 struct mbuf *m; 1362 struct ip6ctlparam *ip6cp = NULL; 1363 const struct sockaddr_in6 *sa6_src = NULL; 1364 int off; 1365 struct tcp_portonly { 1366 u_int16_t th_sport; 1367 u_int16_t th_dport; 1368 } *thp; 1369 int arg; 1370 1371 if (sa->sa_family != AF_INET6 || 1372 sa->sa_len != sizeof(struct sockaddr_in6)) 1373 return; 1374 1375 arg = 0; 1376 if (cmd == PRC_QUENCH) 1377 notify = tcp_quench; 1378 else if (cmd == PRC_MSGSIZE) { 1379 struct ip6ctlparam *ip6cp = d; 1380 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6; 1381 1382 arg = ntohl(icmp6->icmp6_mtu); 1383 notify = tcp_mtudisc; 1384 } else if (!PRC_IS_REDIRECT(cmd) && 1385 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) { 1386 return; 1387 } 1388 1389 /* if the parameter is from icmp6, decode it. */ 1390 if (d != NULL) { 1391 ip6cp = (struct ip6ctlparam *)d; 1392 m = ip6cp->ip6c_m; 1393 ip6 = ip6cp->ip6c_ip6; 1394 off = ip6cp->ip6c_off; 1395 sa6_src = ip6cp->ip6c_src; 1396 } else { 1397 m = NULL; 1398 ip6 = NULL; 1399 off = 0; /* fool gcc */ 1400 sa6_src = &sa6_any; 1401 } 1402 1403 if (ip6 != NULL) { 1404 struct in_conninfo inc; 1405 /* 1406 * XXX: We assume that when IPV6 is non NULL, 1407 * M and OFF are valid. 1408 */ 1409 1410 /* check if we can safely examine src and dst ports */ 1411 if (m->m_pkthdr.len < off + sizeof *thp) 1412 return; 1413 1414 bzero(&th, sizeof th); 1415 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1416 1417 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport, 1418 (struct sockaddr *)ip6cp->ip6c_src, 1419 th.th_sport, cmd, arg, notify); 1420 1421 inc.inc_fport = th.th_dport; 1422 inc.inc_lport = th.th_sport; 1423 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1424 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1425 inc.inc_isipv6 = 1; 1426 syncache_unreach(&inc, &th); 1427 } else 1428 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0, 1429 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify); 1430 } 1431 #endif 1432 1433 /* 1434 * Following is where TCP initial sequence number generation occurs. 1435 * 1436 * There are two places where we must use initial sequence numbers: 1437 * 1. In SYN-ACK packets. 1438 * 2. In SYN packets. 1439 * 1440 * All ISNs for SYN-ACK packets are generated by the syncache. See 1441 * tcp_syncache.c for details. 1442 * 1443 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1444 * depends on this property. In addition, these ISNs should be 1445 * unguessable so as to prevent connection hijacking. To satisfy 1446 * the requirements of this situation, the algorithm outlined in 1447 * RFC 1948 is used to generate sequence numbers. 1448 * 1449 * Implementation details: 1450 * 1451 * Time is based off the system timer, and is corrected so that it 1452 * increases by one megabyte per second. This allows for proper 1453 * recycling on high speed LANs while still leaving over an hour 1454 * before rollover. 1455 * 1456 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1457 * between seeding of isn_secret. This is normally set to zero, 1458 * as reseeding should not be necessary. 1459 * 1460 */ 1461 1462 #define ISN_BYTES_PER_SECOND 1048576 1463 1464 u_char isn_secret[32]; 1465 int isn_last_reseed; 1466 MD5_CTX isn_ctx; 1467 1468 tcp_seq 1469 tcp_new_isn(struct tcpcb *tp) 1470 { 1471 u_int32_t md5_buffer[4]; 1472 tcp_seq new_isn; 1473 1474 /* Seed if this is the first use, reseed if requested. */ 1475 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1476 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1477 < (u_int)ticks))) { 1478 read_random_unlimited(&isn_secret, sizeof isn_secret); 1479 isn_last_reseed = ticks; 1480 } 1481 1482 /* Compute the md5 hash and return the ISN. */ 1483 MD5Init(&isn_ctx); 1484 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1485 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1486 #ifdef INET6 1487 if (tp->t_inpcb->inp_vflag & INP_IPV6) { 1488 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1489 sizeof(struct in6_addr)); 1490 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1491 sizeof(struct in6_addr)); 1492 } else 1493 #endif 1494 { 1495 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1496 sizeof(struct in_addr)); 1497 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1498 sizeof(struct in_addr)); 1499 } 1500 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1501 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1502 new_isn = (tcp_seq) md5_buffer[0]; 1503 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1504 return (new_isn); 1505 } 1506 1507 /* 1508 * When a source quench is received, close congestion window 1509 * to one segment. We will gradually open it again as we proceed. 1510 */ 1511 void 1512 tcp_quench(struct inpcb *inp, int error) 1513 { 1514 struct tcpcb *tp = intotcpcb(inp); 1515 1516 if (tp != NULL) { 1517 tp->snd_cwnd = tp->t_maxseg; 1518 tp->snd_wacked = 0; 1519 } 1520 } 1521 1522 /* 1523 * When a specific ICMP unreachable message is received and the 1524 * connection state is SYN-SENT, drop the connection. This behavior 1525 * is controlled by the icmp_may_rst sysctl. 1526 */ 1527 void 1528 tcp_drop_syn_sent(struct inpcb *inp, int error) 1529 { 1530 struct tcpcb *tp = intotcpcb(inp); 1531 1532 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT)) 1533 tcp_drop(tp, error); 1534 } 1535 1536 /* 1537 * When a `need fragmentation' ICMP is received, update our idea of the MSS 1538 * based on the new value in the route. Also nudge TCP to send something, 1539 * since we know the packet we just sent was dropped. 1540 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1541 */ 1542 void 1543 tcp_mtudisc(struct inpcb *inp, int mtu) 1544 { 1545 struct tcpcb *tp = intotcpcb(inp); 1546 struct rtentry *rt; 1547 struct socket *so = inp->inp_socket; 1548 int maxopd, mss; 1549 #ifdef INET6 1550 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0); 1551 #else 1552 const boolean_t isipv6 = FALSE; 1553 #endif 1554 1555 if (tp == NULL) 1556 return; 1557 1558 /* 1559 * If no MTU is provided in the ICMP message, use the 1560 * next lower likely value, as specified in RFC 1191. 1561 */ 1562 if (mtu == 0) { 1563 int oldmtu; 1564 1565 oldmtu = tp->t_maxopd + 1566 (isipv6 ? 1567 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1568 sizeof(struct tcpiphdr)); 1569 mtu = ip_next_mtu(oldmtu, 0); 1570 } 1571 1572 if (isipv6) 1573 rt = tcp_rtlookup6(&inp->inp_inc); 1574 else 1575 rt = tcp_rtlookup(&inp->inp_inc); 1576 if (rt != NULL) { 1577 struct rmxp_tao *taop = rmx_taop(rt->rt_rmx); 1578 1579 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu) 1580 mtu = rt->rt_rmx.rmx_mtu; 1581 1582 maxopd = mtu - 1583 (isipv6 ? 1584 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1585 sizeof(struct tcpiphdr)); 1586 1587 /* 1588 * XXX - The following conditional probably violates the TCP 1589 * spec. The problem is that, since we don't know the 1590 * other end's MSS, we are supposed to use a conservative 1591 * default. But, if we do that, then MTU discovery will 1592 * never actually take place, because the conservative 1593 * default is much less than the MTUs typically seen 1594 * on the Internet today. For the moment, we'll sweep 1595 * this under the carpet. 1596 * 1597 * The conservative default might not actually be a problem 1598 * if the only case this occurs is when sending an initial 1599 * SYN with options and data to a host we've never talked 1600 * to before. Then, they will reply with an MSS value which 1601 * will get recorded and the new parameters should get 1602 * recomputed. For Further Study. 1603 */ 1604 if (taop->tao_mssopt != 0 && taop->tao_mssopt < maxopd) 1605 maxopd = taop->tao_mssopt; 1606 } else 1607 maxopd = mtu - 1608 (isipv6 ? 1609 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1610 sizeof(struct tcpiphdr)); 1611 1612 if (tp->t_maxopd <= maxopd) 1613 return; 1614 tp->t_maxopd = maxopd; 1615 1616 mss = maxopd; 1617 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) == 1618 (TF_REQ_TSTMP | TF_RCVD_TSTMP)) 1619 mss -= TCPOLEN_TSTAMP_APPA; 1620 1621 if ((tp->t_flags & (TF_REQ_CC | TF_RCVD_CC | TF_NOOPT)) == 1622 (TF_REQ_CC | TF_RCVD_CC)) 1623 mss -= TCPOLEN_CC_APPA; 1624 1625 /* round down to multiple of MCLBYTES */ 1626 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */ 1627 if (mss > MCLBYTES) 1628 mss &= ~(MCLBYTES - 1); 1629 #else 1630 if (mss > MCLBYTES) 1631 mss = (mss / MCLBYTES) * MCLBYTES; 1632 #endif 1633 1634 if (so->so_snd.ssb_hiwat < mss) 1635 mss = so->so_snd.ssb_hiwat; 1636 1637 tp->t_maxseg = mss; 1638 tp->t_rtttime = 0; 1639 tp->snd_nxt = tp->snd_una; 1640 tcp_output(tp); 1641 tcpstat.tcps_mturesent++; 1642 } 1643 1644 /* 1645 * Look-up the routing entry to the peer of this inpcb. If no route 1646 * is found and it cannot be allocated the return NULL. This routine 1647 * is called by TCP routines that access the rmx structure and by tcp_mss 1648 * to get the interface MTU. 1649 */ 1650 struct rtentry * 1651 tcp_rtlookup(struct in_conninfo *inc) 1652 { 1653 struct route *ro = &inc->inc_route; 1654 1655 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1656 /* No route yet, so try to acquire one */ 1657 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1658 /* 1659 * unused portions of the structure MUST be zero'd 1660 * out because rtalloc() treats it as opaque data 1661 */ 1662 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1663 ro->ro_dst.sa_family = AF_INET; 1664 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1665 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1666 inc->inc_faddr; 1667 rtalloc(ro); 1668 } 1669 } 1670 return (ro->ro_rt); 1671 } 1672 1673 #ifdef INET6 1674 struct rtentry * 1675 tcp_rtlookup6(struct in_conninfo *inc) 1676 { 1677 struct route_in6 *ro6 = &inc->inc6_route; 1678 1679 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1680 /* No route yet, so try to acquire one */ 1681 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1682 /* 1683 * unused portions of the structure MUST be zero'd 1684 * out because rtalloc() treats it as opaque data 1685 */ 1686 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1687 ro6->ro_dst.sin6_family = AF_INET6; 1688 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1689 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1690 rtalloc((struct route *)ro6); 1691 } 1692 } 1693 return (ro6->ro_rt); 1694 } 1695 #endif 1696 1697 #ifdef IPSEC 1698 /* compute ESP/AH header size for TCP, including outer IP header. */ 1699 size_t 1700 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1701 { 1702 struct inpcb *inp; 1703 struct mbuf *m; 1704 size_t hdrsiz; 1705 struct ip *ip; 1706 struct tcphdr *th; 1707 1708 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1709 return (0); 1710 MGETHDR(m, MB_DONTWAIT, MT_DATA); 1711 if (!m) 1712 return (0); 1713 1714 #ifdef INET6 1715 if (inp->inp_vflag & INP_IPV6) { 1716 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1717 1718 th = (struct tcphdr *)(ip6 + 1); 1719 m->m_pkthdr.len = m->m_len = 1720 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1721 tcp_fillheaders(tp, ip6, th); 1722 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1723 } else 1724 #endif 1725 { 1726 ip = mtod(m, struct ip *); 1727 th = (struct tcphdr *)(ip + 1); 1728 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1729 tcp_fillheaders(tp, ip, th); 1730 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1731 } 1732 1733 m_free(m); 1734 return (hdrsiz); 1735 } 1736 #endif 1737 1738 /* 1739 * Return a pointer to the cached information about the remote host. 1740 * The cached information is stored in the protocol specific part of 1741 * the route metrics. 1742 */ 1743 struct rmxp_tao * 1744 tcp_gettaocache(struct in_conninfo *inc) 1745 { 1746 struct rtentry *rt; 1747 1748 #ifdef INET6 1749 if (inc->inc_isipv6) 1750 rt = tcp_rtlookup6(inc); 1751 else 1752 #endif 1753 rt = tcp_rtlookup(inc); 1754 1755 /* Make sure this is a host route and is up. */ 1756 if (rt == NULL || 1757 (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST)) 1758 return (NULL); 1759 1760 return (rmx_taop(rt->rt_rmx)); 1761 } 1762 1763 /* 1764 * Clear all the TAO cache entries, called from tcp_init. 1765 * 1766 * XXX 1767 * This routine is just an empty one, because we assume that the routing 1768 * routing tables are initialized at the same time when TCP, so there is 1769 * nothing in the cache left over. 1770 */ 1771 static void 1772 tcp_cleartaocache(void) 1773 { 1774 } 1775 1776 /* 1777 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1778 * 1779 * This code attempts to calculate the bandwidth-delay product as a 1780 * means of determining the optimal window size to maximize bandwidth, 1781 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1782 * routers. This code also does a fairly good job keeping RTTs in check 1783 * across slow links like modems. We implement an algorithm which is very 1784 * similar (but not meant to be) TCP/Vegas. The code operates on the 1785 * transmitter side of a TCP connection and so only effects the transmit 1786 * side of the connection. 1787 * 1788 * BACKGROUND: TCP makes no provision for the management of buffer space 1789 * at the end points or at the intermediate routers and switches. A TCP 1790 * stream, whether using NewReno or not, will eventually buffer as 1791 * many packets as it is able and the only reason this typically works is 1792 * due to the fairly small default buffers made available for a connection 1793 * (typicaly 16K or 32K). As machines use larger windows and/or window 1794 * scaling it is now fairly easy for even a single TCP connection to blow-out 1795 * all available buffer space not only on the local interface, but on 1796 * intermediate routers and switches as well. NewReno makes a misguided 1797 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1798 * then backing off, then steadily increasing the window again until another 1799 * failure occurs, ad-infinitum. This results in terrible oscillation that 1800 * is only made worse as network loads increase and the idea of intentionally 1801 * blowing out network buffers is, frankly, a terrible way to manage network 1802 * resources. 1803 * 1804 * It is far better to limit the transmit window prior to the failure 1805 * condition being achieved. There are two general ways to do this: First 1806 * you can 'scan' through different transmit window sizes and locate the 1807 * point where the RTT stops increasing, indicating that you have filled the 1808 * pipe, then scan backwards until you note that RTT stops decreasing, then 1809 * repeat ad-infinitum. This method works in principle but has severe 1810 * implementation issues due to RTT variances, timer granularity, and 1811 * instability in the algorithm which can lead to many false positives and 1812 * create oscillations as well as interact badly with other TCP streams 1813 * implementing the same algorithm. 1814 * 1815 * The second method is to limit the window to the bandwidth delay product 1816 * of the link. This is the method we implement. RTT variances and our 1817 * own manipulation of the congestion window, bwnd, can potentially 1818 * destabilize the algorithm. For this reason we have to stabilize the 1819 * elements used to calculate the window. We do this by using the minimum 1820 * observed RTT, the long term average of the observed bandwidth, and 1821 * by adding two segments worth of slop. It isn't perfect but it is able 1822 * to react to changing conditions and gives us a very stable basis on 1823 * which to extend the algorithm. 1824 */ 1825 void 1826 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1827 { 1828 u_long bw; 1829 u_long bwnd; 1830 int save_ticks; 1831 int delta_ticks; 1832 1833 /* 1834 * If inflight_enable is disabled in the middle of a tcp connection, 1835 * make sure snd_bwnd is effectively disabled. 1836 */ 1837 if (!tcp_inflight_enable) { 1838 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1839 tp->snd_bandwidth = 0; 1840 return; 1841 } 1842 1843 /* 1844 * Validate the delta time. If a connection is new or has been idle 1845 * a long time we have to reset the bandwidth calculator. 1846 */ 1847 save_ticks = ticks; 1848 delta_ticks = save_ticks - tp->t_bw_rtttime; 1849 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1850 tp->t_bw_rtttime = ticks; 1851 tp->t_bw_rtseq = ack_seq; 1852 if (tp->snd_bandwidth == 0) 1853 tp->snd_bandwidth = tcp_inflight_min; 1854 return; 1855 } 1856 if (delta_ticks == 0) 1857 return; 1858 1859 /* 1860 * Sanity check, plus ignore pure window update acks. 1861 */ 1862 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1863 return; 1864 1865 /* 1866 * Figure out the bandwidth. Due to the tick granularity this 1867 * is a very rough number and it MUST be averaged over a fairly 1868 * long period of time. XXX we need to take into account a link 1869 * that is not using all available bandwidth, but for now our 1870 * slop will ramp us up if this case occurs and the bandwidth later 1871 * increases. 1872 */ 1873 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 1874 tp->t_bw_rtttime = save_ticks; 1875 tp->t_bw_rtseq = ack_seq; 1876 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4; 1877 1878 tp->snd_bandwidth = bw; 1879 1880 /* 1881 * Calculate the semi-static bandwidth delay product, plus two maximal 1882 * segments. The additional slop puts us squarely in the sweet 1883 * spot and also handles the bandwidth run-up case. Without the 1884 * slop we could be locking ourselves into a lower bandwidth. 1885 * 1886 * Situations Handled: 1887 * (1) Prevents over-queueing of packets on LANs, especially on 1888 * high speed LANs, allowing larger TCP buffers to be 1889 * specified, and also does a good job preventing 1890 * over-queueing of packets over choke points like modems 1891 * (at least for the transmit side). 1892 * 1893 * (2) Is able to handle changing network loads (bandwidth 1894 * drops so bwnd drops, bandwidth increases so bwnd 1895 * increases). 1896 * 1897 * (3) Theoretically should stabilize in the face of multiple 1898 * connections implementing the same algorithm (this may need 1899 * a little work). 1900 * 1901 * (4) Stability value (defaults to 20 = 2 maximal packets) can 1902 * be adjusted with a sysctl but typically only needs to be on 1903 * very slow connections. A value no smaller then 5 should 1904 * be used, but only reduce this default if you have no other 1905 * choice. 1906 */ 1907 1908 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2) 1909 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 1910 tcp_inflight_stab * (int)tp->t_maxseg / 10; 1911 #undef USERTT 1912 1913 if (tcp_inflight_debug > 0) { 1914 static int ltime; 1915 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) { 1916 ltime = ticks; 1917 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n", 1918 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd); 1919 } 1920 } 1921 if ((long)bwnd < tcp_inflight_min) 1922 bwnd = tcp_inflight_min; 1923 if (bwnd > tcp_inflight_max) 1924 bwnd = tcp_inflight_max; 1925 if ((long)bwnd < tp->t_maxseg * 2) 1926 bwnd = tp->t_maxseg * 2; 1927 tp->snd_bwnd = bwnd; 1928 } 1929