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 1.63 2008/11/11 10:46:58 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_timer2.h> 120 #include <netinet/tcp_var.h> 121 #include <netinet6/tcp6_var.h> 122 #include <netinet/tcpip.h> 123 #ifdef TCPDEBUG 124 #include <netinet/tcp_debug.h> 125 #endif 126 #include <netinet6/ip6protosw.h> 127 128 #ifdef IPSEC 129 #include <netinet6/ipsec.h> 130 #ifdef INET6 131 #include <netinet6/ipsec6.h> 132 #endif 133 #endif 134 135 #ifdef FAST_IPSEC 136 #include <netproto/ipsec/ipsec.h> 137 #ifdef INET6 138 #include <netproto/ipsec/ipsec6.h> 139 #endif 140 #define IPSEC 141 #endif 142 143 #include <sys/md5.h> 144 #include <sys/msgport2.h> 145 #include <machine/smp.h> 146 147 #include <net/netmsg2.h> 148 149 #if !defined(KTR_TCP) 150 #define KTR_TCP KTR_ALL 151 #endif 152 KTR_INFO_MASTER(tcp); 153 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0); 154 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0); 155 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0); 156 #define logtcp(name) KTR_LOG(tcp_ ## name) 157 158 struct inpcbinfo tcbinfo[MAXCPU]; 159 struct tcpcbackqhead tcpcbackq[MAXCPU]; 160 161 int tcp_mpsafe_proto = 0; 162 TUNABLE_INT("net.inet.tcp.mpsafe_proto", &tcp_mpsafe_proto); 163 164 static int tcp_mpsafe_thread = 0; 165 TUNABLE_INT("net.inet.tcp.mpsafe_thread", &tcp_mpsafe_thread); 166 SYSCTL_INT(_net_inet_tcp, OID_AUTO, mpsafe_thread, CTLFLAG_RW, 167 &tcp_mpsafe_thread, 0, 168 "0:BGL, 1:Adaptive BGL, 2:No BGL(experimental)"); 169 170 int tcp_mssdflt = TCP_MSS; 171 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 172 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size"); 173 174 #ifdef INET6 175 int tcp_v6mssdflt = TCP6_MSS; 176 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW, 177 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6"); 178 #endif 179 180 #if 0 181 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ; 182 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW, 183 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time"); 184 #endif 185 186 int tcp_do_rfc1323 = 1; 187 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW, 188 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions"); 189 190 int tcp_do_rfc1644 = 0; 191 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1644, rfc1644, CTLFLAG_RW, 192 &tcp_do_rfc1644, 0, "Enable rfc1644 (TTCP) extensions"); 193 194 static int tcp_tcbhashsize = 0; 195 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD, 196 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable"); 197 198 static int do_tcpdrain = 1; 199 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0, 200 "Enable tcp_drain routine for extra help when low on mbufs"); 201 202 /* XXX JH */ 203 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD, 204 &tcbinfo[0].ipi_count, 0, "Number of active PCBs"); 205 206 static int icmp_may_rst = 1; 207 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0, 208 "Certain ICMP unreachable messages may abort connections in SYN_SENT"); 209 210 static int tcp_isn_reseed_interval = 0; 211 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW, 212 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret"); 213 214 /* 215 * TCP bandwidth limiting sysctls. Note that the default lower bound of 216 * 1024 exists only for debugging. A good production default would be 217 * something like 6100. 218 */ 219 static int tcp_inflight_enable = 0; 220 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW, 221 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting"); 222 223 static int tcp_inflight_debug = 0; 224 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW, 225 &tcp_inflight_debug, 0, "Debug TCP inflight calculations"); 226 227 static int tcp_inflight_min = 6144; 228 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW, 229 &tcp_inflight_min, 0, "Lower bound for TCP inflight window"); 230 231 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT; 232 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW, 233 &tcp_inflight_max, 0, "Upper bound for TCP inflight window"); 234 235 static int tcp_inflight_stab = 20; 236 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW, 237 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 2 packets)"); 238 239 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives"); 240 static struct malloc_pipe tcptemp_mpipe; 241 242 static void tcp_willblock(int); 243 static void tcp_cleartaocache (void); 244 static void tcp_notify (struct inpcb *, int); 245 246 struct tcp_stats tcpstats_percpu[MAXCPU]; 247 #ifdef SMP 248 static int 249 sysctl_tcpstats(SYSCTL_HANDLER_ARGS) 250 { 251 int cpu, error = 0; 252 253 for (cpu = 0; cpu < ncpus; ++cpu) { 254 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu], 255 sizeof(struct tcp_stats)))) 256 break; 257 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu], 258 sizeof(struct tcp_stats)))) 259 break; 260 } 261 262 return (error); 263 } 264 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW), 265 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics"); 266 #else 267 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW, 268 &tcpstat, tcp_stats, "TCP statistics"); 269 #endif 270 271 /* 272 * Target size of TCP PCB hash tables. Must be a power of two. 273 * 274 * Note that this can be overridden by the kernel environment 275 * variable net.inet.tcp.tcbhashsize 276 */ 277 #ifndef TCBHASHSIZE 278 #define TCBHASHSIZE 512 279 #endif 280 281 /* 282 * This is the actual shape of what we allocate using the zone 283 * allocator. Doing it this way allows us to protect both structures 284 * using the same generation count, and also eliminates the overhead 285 * of allocating tcpcbs separately. By hiding the structure here, 286 * we avoid changing most of the rest of the code (although it needs 287 * to be changed, eventually, for greater efficiency). 288 */ 289 #define ALIGNMENT 32 290 #define ALIGNM1 (ALIGNMENT - 1) 291 struct inp_tp { 292 union { 293 struct inpcb inp; 294 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1]; 295 } inp_tp_u; 296 struct tcpcb tcb; 297 struct tcp_callout inp_tp_rexmt; 298 struct tcp_callout inp_tp_persist; 299 struct tcp_callout inp_tp_keep; 300 struct tcp_callout inp_tp_2msl; 301 struct tcp_callout inp_tp_delack; 302 struct netmsg_tcp_timer inp_tp_timermsg; 303 }; 304 #undef ALIGNMENT 305 #undef ALIGNM1 306 307 /* 308 * Tcp initialization 309 */ 310 void 311 tcp_init(void) 312 { 313 struct inpcbporthead *porthashbase; 314 u_long porthashmask; 315 struct vm_zone *ipi_zone; 316 int hashsize = TCBHASHSIZE; 317 int cpu; 318 319 /* 320 * note: tcptemp is used for keepalives, and it is ok for an 321 * allocation to fail so do not specify MPF_INT. 322 */ 323 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp), 324 25, -1, 0, NULL); 325 326 tcp_ccgen = 1; 327 tcp_cleartaocache(); 328 329 tcp_delacktime = TCPTV_DELACK; 330 tcp_keepinit = TCPTV_KEEP_INIT; 331 tcp_keepidle = TCPTV_KEEP_IDLE; 332 tcp_keepintvl = TCPTV_KEEPINTVL; 333 tcp_maxpersistidle = TCPTV_KEEP_IDLE; 334 tcp_msl = TCPTV_MSL; 335 tcp_rexmit_min = TCPTV_MIN; 336 tcp_rexmit_slop = TCPTV_CPU_VAR; 337 338 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize); 339 if (!powerof2(hashsize)) { 340 kprintf("WARNING: TCB hash size not a power of 2\n"); 341 hashsize = 512; /* safe default */ 342 } 343 tcp_tcbhashsize = hashsize; 344 porthashbase = hashinit(hashsize, M_PCB, &porthashmask); 345 ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets, 346 ZONE_INTERRUPT, 0); 347 348 for (cpu = 0; cpu < ncpus2; cpu++) { 349 in_pcbinfo_init(&tcbinfo[cpu]); 350 tcbinfo[cpu].cpu = cpu; 351 tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB, 352 &tcbinfo[cpu].hashmask); 353 tcbinfo[cpu].porthashbase = porthashbase; 354 tcbinfo[cpu].porthashmask = porthashmask; 355 tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB, 356 &tcbinfo[cpu].wildcardhashmask); 357 tcbinfo[cpu].ipi_zone = ipi_zone; 358 TAILQ_INIT(&tcpcbackq[cpu]); 359 } 360 361 tcp_reass_maxseg = nmbclusters / 16; 362 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg); 363 364 #ifdef INET6 365 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) 366 #else 367 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) 368 #endif 369 if (max_protohdr < TCP_MINPROTOHDR) 370 max_protohdr = TCP_MINPROTOHDR; 371 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) 372 panic("tcp_init"); 373 #undef TCP_MINPROTOHDR 374 375 /* 376 * Initialize TCP statistics counters for each CPU. 377 */ 378 #ifdef SMP 379 for (cpu = 0; cpu < ncpus; ++cpu) { 380 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats)); 381 } 382 #else 383 bzero(&tcpstat, sizeof(struct tcp_stats)); 384 #endif 385 386 syncache_init(); 387 tcp_thread_init(); 388 } 389 390 void 391 tcpmsg_service_loop(void *dummy) 392 { 393 struct netmsg *msg; 394 int mplocked; 395 396 /* 397 * Thread was started with TDF_MPSAFE 398 */ 399 mplocked = 0; 400 401 while ((msg = lwkt_waitport(&curthread->td_msgport, 0))) { 402 do { 403 logtcp(rxmsg); 404 mplocked = netmsg_service(msg, tcp_mpsafe_thread, 405 mplocked); 406 } while ((msg = lwkt_getport(&curthread->td_msgport)) != NULL); 407 408 logtcp(delayed); 409 tcp_willblock(mplocked); 410 logtcp(wait); 411 } 412 } 413 414 static void 415 tcp_willblock(int mplocked) 416 { 417 struct tcpcb *tp; 418 int cpu = mycpu->gd_cpuid; 419 int unlock = 0; 420 421 if (!mplocked && !tcp_mpsafe_proto) { 422 if (TAILQ_EMPTY(&tcpcbackq[cpu])) 423 return; 424 425 get_mplock(); 426 mplocked = 1; 427 unlock = 1; 428 } 429 430 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) { 431 KKASSERT(tp->t_flags & TF_ONOUTPUTQ); 432 tp->t_flags &= ~TF_ONOUTPUTQ; 433 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq); 434 tcp_output(tp); 435 } 436 437 if (unlock) 438 rel_mplock(); 439 } 440 441 442 /* 443 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. 444 * tcp_template used to store this data in mbufs, but we now recopy it out 445 * of the tcpcb each time to conserve mbufs. 446 */ 447 void 448 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr) 449 { 450 struct inpcb *inp = tp->t_inpcb; 451 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr; 452 453 #ifdef INET6 454 if (inp->inp_vflag & INP_IPV6) { 455 struct ip6_hdr *ip6; 456 457 ip6 = (struct ip6_hdr *)ip_ptr; 458 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | 459 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK); 460 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | 461 (IPV6_VERSION & IPV6_VERSION_MASK); 462 ip6->ip6_nxt = IPPROTO_TCP; 463 ip6->ip6_plen = sizeof(struct tcphdr); 464 ip6->ip6_src = inp->in6p_laddr; 465 ip6->ip6_dst = inp->in6p_faddr; 466 tcp_hdr->th_sum = 0; 467 } else 468 #endif 469 { 470 struct ip *ip = (struct ip *) ip_ptr; 471 472 ip->ip_vhl = IP_VHL_BORING; 473 ip->ip_tos = 0; 474 ip->ip_len = 0; 475 ip->ip_id = 0; 476 ip->ip_off = 0; 477 ip->ip_ttl = 0; 478 ip->ip_sum = 0; 479 ip->ip_p = IPPROTO_TCP; 480 ip->ip_src = inp->inp_laddr; 481 ip->ip_dst = inp->inp_faddr; 482 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr, 483 ip->ip_dst.s_addr, 484 htons(sizeof(struct tcphdr) + IPPROTO_TCP)); 485 } 486 487 tcp_hdr->th_sport = inp->inp_lport; 488 tcp_hdr->th_dport = inp->inp_fport; 489 tcp_hdr->th_seq = 0; 490 tcp_hdr->th_ack = 0; 491 tcp_hdr->th_x2 = 0; 492 tcp_hdr->th_off = 5; 493 tcp_hdr->th_flags = 0; 494 tcp_hdr->th_win = 0; 495 tcp_hdr->th_urp = 0; 496 } 497 498 /* 499 * Create template to be used to send tcp packets on a connection. 500 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only 501 * use for this function is in keepalives, which use tcp_respond. 502 */ 503 struct tcptemp * 504 tcp_maketemplate(struct tcpcb *tp) 505 { 506 struct tcptemp *tmp; 507 508 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL) 509 return (NULL); 510 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t); 511 return (tmp); 512 } 513 514 void 515 tcp_freetemplate(struct tcptemp *tmp) 516 { 517 mpipe_free(&tcptemp_mpipe, tmp); 518 } 519 520 /* 521 * Send a single message to the TCP at address specified by 522 * the given TCP/IP header. If m == NULL, then we make a copy 523 * of the tcpiphdr at ti and send directly to the addressed host. 524 * This is used to force keep alive messages out using the TCP 525 * template for a connection. If flags are given then we send 526 * a message back to the TCP which originated the * segment ti, 527 * and discard the mbuf containing it and any other attached mbufs. 528 * 529 * In any case the ack and sequence number of the transmitted 530 * segment are as specified by the parameters. 531 * 532 * NOTE: If m != NULL, then ti must point to *inside* the mbuf. 533 */ 534 void 535 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, 536 tcp_seq ack, tcp_seq seq, int flags) 537 { 538 int tlen; 539 int win = 0; 540 struct route *ro = NULL; 541 struct route sro; 542 struct ip *ip = ipgen; 543 struct tcphdr *nth; 544 int ipflags = 0; 545 struct route_in6 *ro6 = NULL; 546 struct route_in6 sro6; 547 struct ip6_hdr *ip6 = ipgen; 548 boolean_t use_tmpro = TRUE; 549 #ifdef INET6 550 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6); 551 #else 552 const boolean_t isipv6 = FALSE; 553 #endif 554 555 if (tp != NULL) { 556 if (!(flags & TH_RST)) { 557 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv); 558 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 559 win = (long)TCP_MAXWIN << tp->rcv_scale; 560 } 561 /* 562 * Don't use the route cache of a listen socket, 563 * it is not MPSAFE; use temporary route cache. 564 */ 565 if (tp->t_state != TCPS_LISTEN) { 566 if (isipv6) 567 ro6 = &tp->t_inpcb->in6p_route; 568 else 569 ro = &tp->t_inpcb->inp_route; 570 use_tmpro = FALSE; 571 } 572 } 573 if (use_tmpro) { 574 if (isipv6) { 575 ro6 = &sro6; 576 bzero(ro6, sizeof *ro6); 577 } else { 578 ro = &sro; 579 bzero(ro, sizeof *ro); 580 } 581 } 582 if (m == NULL) { 583 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 584 if (m == NULL) 585 return; 586 tlen = 0; 587 m->m_data += max_linkhdr; 588 if (isipv6) { 589 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); 590 ip6 = mtod(m, struct ip6_hdr *); 591 nth = (struct tcphdr *)(ip6 + 1); 592 } else { 593 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip)); 594 ip = mtod(m, struct ip *); 595 nth = (struct tcphdr *)(ip + 1); 596 } 597 bcopy(th, nth, sizeof(struct tcphdr)); 598 flags = TH_ACK; 599 } else { 600 m_freem(m->m_next); 601 m->m_next = NULL; 602 m->m_data = (caddr_t)ipgen; 603 /* m_len is set later */ 604 tlen = 0; 605 #define xchg(a, b, type) { type t; t = a; a = b; b = t; } 606 if (isipv6) { 607 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); 608 nth = (struct tcphdr *)(ip6 + 1); 609 } else { 610 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long); 611 nth = (struct tcphdr *)(ip + 1); 612 } 613 if (th != nth) { 614 /* 615 * this is usually a case when an extension header 616 * exists between the IPv6 header and the 617 * TCP header. 618 */ 619 nth->th_sport = th->th_sport; 620 nth->th_dport = th->th_dport; 621 } 622 xchg(nth->th_dport, nth->th_sport, n_short); 623 #undef xchg 624 } 625 if (isipv6) { 626 ip6->ip6_flow = 0; 627 ip6->ip6_vfc = IPV6_VERSION; 628 ip6->ip6_nxt = IPPROTO_TCP; 629 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen)); 630 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 631 } else { 632 tlen += sizeof(struct tcpiphdr); 633 ip->ip_len = tlen; 634 ip->ip_ttl = ip_defttl; 635 } 636 m->m_len = tlen; 637 m->m_pkthdr.len = tlen; 638 m->m_pkthdr.rcvif = (struct ifnet *) NULL; 639 nth->th_seq = htonl(seq); 640 nth->th_ack = htonl(ack); 641 nth->th_x2 = 0; 642 nth->th_off = sizeof(struct tcphdr) >> 2; 643 nth->th_flags = flags; 644 if (tp != NULL) 645 nth->th_win = htons((u_short) (win >> tp->rcv_scale)); 646 else 647 nth->th_win = htons((u_short)win); 648 nth->th_urp = 0; 649 if (isipv6) { 650 nth->th_sum = 0; 651 nth->th_sum = in6_cksum(m, IPPROTO_TCP, 652 sizeof(struct ip6_hdr), 653 tlen - sizeof(struct ip6_hdr)); 654 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL, 655 (ro6 && ro6->ro_rt) ? 656 ro6->ro_rt->rt_ifp : NULL); 657 } else { 658 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 659 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); 660 m->m_pkthdr.csum_flags = CSUM_TCP; 661 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 662 } 663 #ifdef TCPDEBUG 664 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) 665 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); 666 #endif 667 if (isipv6) { 668 ip6_output(m, NULL, ro6, ipflags, NULL, NULL, 669 tp ? tp->t_inpcb : NULL); 670 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) { 671 RTFREE(ro6->ro_rt); 672 ro6->ro_rt = NULL; 673 } 674 } else { 675 ipflags |= IP_DEBUGROUTE; 676 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL); 677 if ((ro == &sro) && (ro->ro_rt != NULL)) { 678 RTFREE(ro->ro_rt); 679 ro->ro_rt = NULL; 680 } 681 } 682 } 683 684 /* 685 * Create a new TCP control block, making an 686 * empty reassembly queue and hooking it to the argument 687 * protocol control block. The `inp' parameter must have 688 * come from the zone allocator set up in tcp_init(). 689 */ 690 struct tcpcb * 691 tcp_newtcpcb(struct inpcb *inp) 692 { 693 struct inp_tp *it; 694 struct tcpcb *tp; 695 #ifdef INET6 696 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 697 #else 698 const boolean_t isipv6 = FALSE; 699 #endif 700 701 it = (struct inp_tp *)inp; 702 tp = &it->tcb; 703 bzero(tp, sizeof(struct tcpcb)); 704 LIST_INIT(&tp->t_segq); 705 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 706 707 /* Set up our timeouts. */ 708 tp->tt_rexmt = &it->inp_tp_rexmt; 709 tp->tt_persist = &it->inp_tp_persist; 710 tp->tt_keep = &it->inp_tp_keep; 711 tp->tt_2msl = &it->inp_tp_2msl; 712 tp->tt_delack = &it->inp_tp_delack; 713 tcp_inittimers(tp); 714 715 tp->tt_msg = &it->inp_tp_timermsg; 716 if (isipv6) { 717 /* Don't mess with IPv6; always create timer message */ 718 tcp_create_timermsg(tp); 719 } else { 720 /* 721 * Zero out timer message. We don't create it here, 722 * since the current CPU may not be the owner of this 723 * inpcb. 724 */ 725 bzero(tp->tt_msg, sizeof(*tp->tt_msg)); 726 } 727 728 if (tcp_do_rfc1323) 729 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP); 730 if (tcp_do_rfc1644) 731 tp->t_flags |= TF_REQ_CC; 732 tp->t_inpcb = inp; /* XXX */ 733 tp->t_state = TCPS_CLOSED; 734 /* 735 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 736 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 737 * reasonable initial retransmit time. 738 */ 739 tp->t_srtt = TCPTV_SRTTBASE; 740 tp->t_rttvar = 741 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 742 tp->t_rttmin = tcp_rexmit_min; 743 tp->t_rxtcur = TCPTV_RTOBASE; 744 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 745 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 746 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 747 tp->t_rcvtime = ticks; 748 /* 749 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 750 * because the socket may be bound to an IPv6 wildcard address, 751 * which may match an IPv4-mapped IPv6 address. 752 */ 753 inp->inp_ip_ttl = ip_defttl; 754 inp->inp_ppcb = tp; 755 tcp_sack_tcpcb_init(tp); 756 return (tp); /* XXX */ 757 } 758 759 /* 760 * Drop a TCP connection, reporting the specified error. 761 * If connection is synchronized, then send a RST to peer. 762 */ 763 struct tcpcb * 764 tcp_drop(struct tcpcb *tp, int error) 765 { 766 struct socket *so = tp->t_inpcb->inp_socket; 767 768 if (TCPS_HAVERCVDSYN(tp->t_state)) { 769 tp->t_state = TCPS_CLOSED; 770 tcp_output(tp); 771 tcpstat.tcps_drops++; 772 } else 773 tcpstat.tcps_conndrops++; 774 if (error == ETIMEDOUT && tp->t_softerror) 775 error = tp->t_softerror; 776 so->so_error = error; 777 return (tcp_close(tp)); 778 } 779 780 #ifdef SMP 781 782 struct netmsg_remwildcard { 783 struct netmsg nm_netmsg; 784 struct inpcb *nm_inp; 785 struct inpcbinfo *nm_pcbinfo; 786 #if defined(INET6) 787 int nm_isinet6; 788 #else 789 int nm_unused01; 790 #endif 791 }; 792 793 /* 794 * Wildcard inpcb's on SMP boxes must be removed from all cpus before the 795 * inp can be detached. We do this by cycling through the cpus, ending up 796 * on the cpu controlling the inp last and then doing the disconnect. 797 */ 798 static void 799 in_pcbremwildcardhash_handler(struct netmsg *msg0) 800 { 801 struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0; 802 int cpu; 803 804 cpu = msg->nm_pcbinfo->cpu; 805 806 if (cpu == msg->nm_inp->inp_pcbinfo->cpu) { 807 /* note: detach removes any wildcard hash entry */ 808 #ifdef INET6 809 if (msg->nm_isinet6) 810 in6_pcbdetach(msg->nm_inp); 811 else 812 #endif 813 in_pcbdetach(msg->nm_inp); 814 lwkt_replymsg(&msg->nm_netmsg.nm_lmsg, 0); 815 } else { 816 in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo); 817 cpu = (cpu + 1) % ncpus2; 818 msg->nm_pcbinfo = &tcbinfo[cpu]; 819 lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg); 820 } 821 } 822 823 #endif 824 825 /* 826 * Close a TCP control block: 827 * discard all space held by the tcp 828 * discard internet protocol block 829 * wake up any sleepers 830 */ 831 struct tcpcb * 832 tcp_close(struct tcpcb *tp) 833 { 834 struct tseg_qent *q; 835 struct inpcb *inp = tp->t_inpcb; 836 struct socket *so = inp->inp_socket; 837 struct rtentry *rt; 838 boolean_t dosavessthresh; 839 #ifdef SMP 840 int cpu; 841 #endif 842 #ifdef INET6 843 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 844 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0); 845 #else 846 const boolean_t isipv6 = FALSE; 847 #endif 848 849 /* 850 * The tp is not instantly destroyed in the wildcard case. Setting 851 * the state to TCPS_TERMINATING will prevent the TCP stack from 852 * messing with it, though it should be noted that this change may 853 * not take effect on other cpus until we have chained the wildcard 854 * hash removal. 855 * 856 * XXX we currently depend on the BGL to synchronize the tp->t_state 857 * update and prevent other tcp protocol threads from accepting new 858 * connections on the listen socket we might be trying to close down. 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. 867 */ 868 if (tp->tt_msg != NULL) { 869 tcp_callout_stop(tp, tp->tt_rexmt); 870 tcp_callout_stop(tp, tp->tt_persist); 871 tcp_callout_stop(tp, tp->tt_keep); 872 tcp_callout_stop(tp, tp->tt_2msl); 873 tcp_callout_stop(tp, tp->tt_delack); 874 } 875 876 if (tp->t_flags & TF_ONOUTPUTQ) { 877 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 878 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 879 tp->t_flags &= ~TF_ONOUTPUTQ; 880 } 881 882 /* 883 * If we got enough samples through the srtt filter, 884 * save the rtt and rttvar in the routing entry. 885 * 'Enough' is arbitrarily defined as the 16 samples. 886 * 16 samples is enough for the srtt filter to converge 887 * to within 5% of the correct value; fewer samples and 888 * we could save a very bogus rtt. 889 * 890 * Don't update the default route's characteristics and don't 891 * update anything that the user "locked". 892 */ 893 if (tp->t_rttupdated >= 16) { 894 u_long i = 0; 895 896 if (isipv6) { 897 struct sockaddr_in6 *sin6; 898 899 if ((rt = inp->in6p_route.ro_rt) == NULL) 900 goto no_valid_rt; 901 sin6 = (struct sockaddr_in6 *)rt_key(rt); 902 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 903 goto no_valid_rt; 904 } else 905 if ((rt = inp->inp_route.ro_rt) == NULL || 906 ((struct sockaddr_in *)rt_key(rt))-> 907 sin_addr.s_addr == INADDR_ANY) 908 goto no_valid_rt; 909 910 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 911 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 912 if (rt->rt_rmx.rmx_rtt && i) 913 /* 914 * filter this update to half the old & half 915 * the new values, converting scale. 916 * See route.h and tcp_var.h for a 917 * description of the scaling constants. 918 */ 919 rt->rt_rmx.rmx_rtt = 920 (rt->rt_rmx.rmx_rtt + i) / 2; 921 else 922 rt->rt_rmx.rmx_rtt = i; 923 tcpstat.tcps_cachedrtt++; 924 } 925 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 926 i = tp->t_rttvar * 927 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 928 if (rt->rt_rmx.rmx_rttvar && i) 929 rt->rt_rmx.rmx_rttvar = 930 (rt->rt_rmx.rmx_rttvar + i) / 2; 931 else 932 rt->rt_rmx.rmx_rttvar = i; 933 tcpstat.tcps_cachedrttvar++; 934 } 935 /* 936 * The old comment here said: 937 * update the pipelimit (ssthresh) if it has been updated 938 * already or if a pipesize was specified & the threshhold 939 * got below half the pipesize. I.e., wait for bad news 940 * before we start updating, then update on both good 941 * and bad news. 942 * 943 * But we want to save the ssthresh even if no pipesize is 944 * specified explicitly in the route, because such 945 * connections still have an implicit pipesize specified 946 * by the global tcp_sendspace. In the absence of a reliable 947 * way to calculate the pipesize, it will have to do. 948 */ 949 i = tp->snd_ssthresh; 950 if (rt->rt_rmx.rmx_sendpipe != 0) 951 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 952 else 953 dosavessthresh = (i < so->so_snd.ssb_hiwat/2); 954 if (dosavessthresh || 955 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 956 (rt->rt_rmx.rmx_ssthresh != 0))) { 957 /* 958 * convert the limit from user data bytes to 959 * packets then to packet data bytes. 960 */ 961 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 962 if (i < 2) 963 i = 2; 964 i *= tp->t_maxseg + 965 (isipv6 ? 966 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 967 sizeof(struct tcpiphdr)); 968 if (rt->rt_rmx.rmx_ssthresh) 969 rt->rt_rmx.rmx_ssthresh = 970 (rt->rt_rmx.rmx_ssthresh + i) / 2; 971 else 972 rt->rt_rmx.rmx_ssthresh = i; 973 tcpstat.tcps_cachedssthresh++; 974 } 975 } 976 977 no_valid_rt: 978 /* free the reassembly queue, if any */ 979 while((q = LIST_FIRST(&tp->t_segq)) != NULL) { 980 LIST_REMOVE(q, tqe_q); 981 m_freem(q->tqe_m); 982 FREE(q, M_TSEGQ); 983 tcp_reass_qsize--; 984 } 985 /* throw away SACK blocks in scoreboard*/ 986 if (TCP_DO_SACK(tp)) 987 tcp_sack_cleanup(&tp->scb); 988 989 inp->inp_ppcb = NULL; 990 soisdisconnected(so); 991 992 tcp_destroy_timermsg(tp); 993 994 /* 995 * Discard the inp. In the SMP case a wildcard inp's hash (created 996 * by a listen socket or an INADDR_ANY udp socket) is replicated 997 * for each protocol thread and must be removed in the context of 998 * that thread. This is accomplished by chaining the message 999 * through the cpus. 1000 * 1001 * If the inp is not wildcarded we simply detach, which will remove 1002 * the any hashes still present for this inp. 1003 */ 1004 #ifdef SMP 1005 if (inp->inp_flags & INP_WILDCARD_MP) { 1006 struct netmsg_remwildcard *msg; 1007 1008 cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2; 1009 msg = kmalloc(sizeof(struct netmsg_remwildcard), 1010 M_LWKTMSG, M_INTWAIT); 1011 netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0, 1012 in_pcbremwildcardhash_handler); 1013 #ifdef INET6 1014 msg->nm_isinet6 = isafinet6; 1015 #endif 1016 msg->nm_inp = inp; 1017 msg->nm_pcbinfo = &tcbinfo[cpu]; 1018 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg); 1019 } else 1020 #endif 1021 { 1022 /* note: detach removes any wildcard hash entry */ 1023 #ifdef INET6 1024 if (isafinet6) 1025 in6_pcbdetach(inp); 1026 else 1027 #endif 1028 in_pcbdetach(inp); 1029 } 1030 tcpstat.tcps_closed++; 1031 return (NULL); 1032 } 1033 1034 static __inline void 1035 tcp_drain_oncpu(struct inpcbhead *head) 1036 { 1037 struct inpcb *inpb; 1038 struct tcpcb *tcpb; 1039 struct tseg_qent *te; 1040 1041 LIST_FOREACH(inpb, head, inp_list) { 1042 if (inpb->inp_flags & INP_PLACEMARKER) 1043 continue; 1044 if ((tcpb = intotcpcb(inpb))) { 1045 while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) { 1046 LIST_REMOVE(te, tqe_q); 1047 m_freem(te->tqe_m); 1048 FREE(te, M_TSEGQ); 1049 tcp_reass_qsize--; 1050 } 1051 } 1052 } 1053 } 1054 1055 #ifdef SMP 1056 struct netmsg_tcp_drain { 1057 struct netmsg nm_netmsg; 1058 struct inpcbhead *nm_head; 1059 }; 1060 1061 static void 1062 tcp_drain_handler(netmsg_t netmsg) 1063 { 1064 struct netmsg_tcp_drain *nm = (void *)netmsg; 1065 1066 tcp_drain_oncpu(nm->nm_head); 1067 lwkt_replymsg(&nm->nm_netmsg.nm_lmsg, 0); 1068 } 1069 #endif 1070 1071 void 1072 tcp_drain(void) 1073 { 1074 #ifdef SMP 1075 int cpu; 1076 #endif 1077 1078 if (!do_tcpdrain) 1079 return; 1080 1081 /* 1082 * Walk the tcpbs, if existing, and flush the reassembly queue, 1083 * if there is one... 1084 * XXX: The "Net/3" implementation doesn't imply that the TCP 1085 * reassembly queue should be flushed, but in a situation 1086 * where we're really low on mbufs, this is potentially 1087 * useful. 1088 */ 1089 #ifdef SMP 1090 for (cpu = 0; cpu < ncpus2; cpu++) { 1091 struct netmsg_tcp_drain *msg; 1092 1093 if (cpu == mycpu->gd_cpuid) { 1094 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead); 1095 } else { 1096 msg = kmalloc(sizeof(struct netmsg_tcp_drain), 1097 M_LWKTMSG, M_NOWAIT); 1098 if (msg == NULL) 1099 continue; 1100 netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0, 1101 tcp_drain_handler); 1102 msg->nm_head = &tcbinfo[cpu].pcblisthead; 1103 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg); 1104 } 1105 } 1106 #else 1107 tcp_drain_oncpu(&tcbinfo[0].pcblisthead); 1108 #endif 1109 } 1110 1111 /* 1112 * Notify a tcp user of an asynchronous error; 1113 * store error as soft error, but wake up user 1114 * (for now, won't do anything until can select for soft error). 1115 * 1116 * Do not wake up user since there currently is no mechanism for 1117 * reporting soft errors (yet - a kqueue filter may be added). 1118 */ 1119 static void 1120 tcp_notify(struct inpcb *inp, int error) 1121 { 1122 struct tcpcb *tp = intotcpcb(inp); 1123 1124 /* 1125 * Ignore some errors if we are hooked up. 1126 * If connection hasn't completed, has retransmitted several times, 1127 * and receives a second error, give up now. This is better 1128 * than waiting a long time to establish a connection that 1129 * can never complete. 1130 */ 1131 if (tp->t_state == TCPS_ESTABLISHED && 1132 (error == EHOSTUNREACH || error == ENETUNREACH || 1133 error == EHOSTDOWN)) { 1134 return; 1135 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1136 tp->t_softerror) 1137 tcp_drop(tp, error); 1138 else 1139 tp->t_softerror = error; 1140 #if 0 1141 wakeup(&so->so_timeo); 1142 sorwakeup(so); 1143 sowwakeup(so); 1144 #endif 1145 } 1146 1147 static int 1148 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1149 { 1150 int error, i, n; 1151 struct inpcb *marker; 1152 struct inpcb *inp; 1153 inp_gen_t gencnt; 1154 globaldata_t gd; 1155 int origcpu, ccpu; 1156 1157 error = 0; 1158 n = 0; 1159 1160 /* 1161 * The process of preparing the TCB list is too time-consuming and 1162 * resource-intensive to repeat twice on every request. 1163 */ 1164 if (req->oldptr == NULL) { 1165 for (ccpu = 0; ccpu < ncpus; ++ccpu) { 1166 gd = globaldata_find(ccpu); 1167 n += tcbinfo[gd->gd_cpuid].ipi_count; 1168 } 1169 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1170 return (0); 1171 } 1172 1173 if (req->newptr != NULL) 1174 return (EPERM); 1175 1176 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1177 marker->inp_flags |= INP_PLACEMARKER; 1178 1179 /* 1180 * OK, now we're committed to doing something. Run the inpcb list 1181 * for each cpu in the system and construct the output. Use a 1182 * list placemarker to deal with list changes occuring during 1183 * copyout blockages (but otherwise depend on being on the correct 1184 * cpu to avoid races). 1185 */ 1186 origcpu = mycpu->gd_cpuid; 1187 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) { 1188 globaldata_t rgd; 1189 caddr_t inp_ppcb; 1190 struct xtcpcb xt; 1191 int cpu_id; 1192 1193 cpu_id = (origcpu + ccpu) % ncpus; 1194 if ((smp_active_mask & (1 << cpu_id)) == 0) 1195 continue; 1196 rgd = globaldata_find(cpu_id); 1197 lwkt_setcpu_self(rgd); 1198 1199 gencnt = tcbinfo[cpu_id].ipi_gencnt; 1200 n = tcbinfo[cpu_id].ipi_count; 1201 1202 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list); 1203 i = 0; 1204 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1205 /* 1206 * process a snapshot of pcbs, ignoring placemarkers 1207 * and using our own to allow SYSCTL_OUT to block. 1208 */ 1209 LIST_REMOVE(marker, inp_list); 1210 LIST_INSERT_AFTER(inp, marker, inp_list); 1211 1212 if (inp->inp_flags & INP_PLACEMARKER) 1213 continue; 1214 if (inp->inp_gencnt > gencnt) 1215 continue; 1216 if (prison_xinpcb(req->td, inp)) 1217 continue; 1218 1219 xt.xt_len = sizeof xt; 1220 bcopy(inp, &xt.xt_inp, sizeof *inp); 1221 inp_ppcb = inp->inp_ppcb; 1222 if (inp_ppcb != NULL) 1223 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1224 else 1225 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1226 if (inp->inp_socket) 1227 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1228 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1229 break; 1230 ++i; 1231 } 1232 LIST_REMOVE(marker, inp_list); 1233 if (error == 0 && i < n) { 1234 bzero(&xt, sizeof xt); 1235 xt.xt_len = sizeof xt; 1236 while (i < n) { 1237 error = SYSCTL_OUT(req, &xt, sizeof xt); 1238 if (error) 1239 break; 1240 ++i; 1241 } 1242 } 1243 } 1244 1245 /* 1246 * Make sure we are on the same cpu we were on originally, since 1247 * higher level callers expect this. Also don't pollute caches with 1248 * migrated userland data by (eventually) returning to userland 1249 * on a different cpu. 1250 */ 1251 lwkt_setcpu_self(globaldata_find(origcpu)); 1252 kfree(marker, M_TEMP); 1253 return (error); 1254 } 1255 1256 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1257 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1258 1259 static int 1260 tcp_getcred(SYSCTL_HANDLER_ARGS) 1261 { 1262 struct sockaddr_in addrs[2]; 1263 struct inpcb *inp; 1264 int cpu; 1265 int error; 1266 1267 error = suser(req->td); 1268 if (error != 0) 1269 return (error); 1270 error = SYSCTL_IN(req, addrs, sizeof addrs); 1271 if (error != 0) 1272 return (error); 1273 crit_enter(); 1274 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1275 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1276 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1277 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1278 if (inp == NULL || inp->inp_socket == NULL) { 1279 error = ENOENT; 1280 goto out; 1281 } 1282 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1283 out: 1284 crit_exit(); 1285 return (error); 1286 } 1287 1288 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1289 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1290 1291 #ifdef INET6 1292 static int 1293 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1294 { 1295 struct sockaddr_in6 addrs[2]; 1296 struct inpcb *inp; 1297 int error; 1298 boolean_t mapped = FALSE; 1299 1300 error = suser(req->td); 1301 if (error != 0) 1302 return (error); 1303 error = SYSCTL_IN(req, addrs, sizeof addrs); 1304 if (error != 0) 1305 return (error); 1306 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { 1307 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) 1308 mapped = TRUE; 1309 else 1310 return (EINVAL); 1311 } 1312 crit_enter(); 1313 if (mapped) { 1314 inp = in_pcblookup_hash(&tcbinfo[0], 1315 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], 1316 addrs[1].sin6_port, 1317 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], 1318 addrs[0].sin6_port, 1319 0, NULL); 1320 } else { 1321 inp = in6_pcblookup_hash(&tcbinfo[0], 1322 &addrs[1].sin6_addr, addrs[1].sin6_port, 1323 &addrs[0].sin6_addr, addrs[0].sin6_port, 1324 0, NULL); 1325 } 1326 if (inp == NULL || inp->inp_socket == NULL) { 1327 error = ENOENT; 1328 goto out; 1329 } 1330 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1331 out: 1332 crit_exit(); 1333 return (error); 1334 } 1335 1336 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1337 0, 0, 1338 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1339 #endif 1340 1341 struct netmsg_tcp_notify { 1342 struct netmsg nm_nmsg; 1343 void (*nm_notify)(struct inpcb *, int); 1344 struct in_addr nm_faddr; 1345 int nm_arg; 1346 }; 1347 1348 static void 1349 tcp_notifyall_oncpu(struct netmsg *netmsg) 1350 { 1351 struct netmsg_tcp_notify *nmsg = (struct netmsg_tcp_notify *)netmsg; 1352 int nextcpu; 1353 1354 in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nmsg->nm_faddr, 1355 nmsg->nm_arg, nmsg->nm_notify); 1356 1357 nextcpu = mycpuid + 1; 1358 if (nextcpu < ncpus2) 1359 lwkt_forwardmsg(tcp_cport(nextcpu), &netmsg->nm_lmsg); 1360 else 1361 lwkt_replymsg(&netmsg->nm_lmsg, 0); 1362 } 1363 1364 void 1365 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip) 1366 { 1367 struct ip *ip = vip; 1368 struct tcphdr *th; 1369 struct in_addr faddr; 1370 struct inpcb *inp; 1371 struct tcpcb *tp; 1372 void (*notify)(struct inpcb *, int) = tcp_notify; 1373 tcp_seq icmpseq; 1374 int arg, cpu; 1375 1376 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) { 1377 return; 1378 } 1379 1380 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1381 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1382 return; 1383 1384 arg = inetctlerrmap[cmd]; 1385 if (cmd == PRC_QUENCH) { 1386 notify = tcp_quench; 1387 } else if (icmp_may_rst && 1388 (cmd == PRC_UNREACH_ADMIN_PROHIB || 1389 cmd == PRC_UNREACH_PORT || 1390 cmd == PRC_TIMXCEED_INTRANS) && 1391 ip != NULL) { 1392 notify = tcp_drop_syn_sent; 1393 } else if (cmd == PRC_MSGSIZE) { 1394 struct icmp *icmp = (struct icmp *) 1395 ((caddr_t)ip - offsetof(struct icmp, icmp_ip)); 1396 1397 arg = ntohs(icmp->icmp_nextmtu); 1398 notify = tcp_mtudisc; 1399 } else if (PRC_IS_REDIRECT(cmd)) { 1400 ip = NULL; 1401 notify = in_rtchange; 1402 } else if (cmd == PRC_HOSTDEAD) { 1403 ip = NULL; 1404 } 1405 1406 if (ip != NULL) { 1407 crit_enter(); 1408 th = (struct tcphdr *)((caddr_t)ip + 1409 (IP_VHL_HL(ip->ip_vhl) << 2)); 1410 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport, 1411 ip->ip_src.s_addr, th->th_sport); 1412 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport, 1413 ip->ip_src, th->th_sport, 0, NULL); 1414 if ((inp != NULL) && (inp->inp_socket != NULL)) { 1415 icmpseq = htonl(th->th_seq); 1416 tp = intotcpcb(inp); 1417 if (SEQ_GEQ(icmpseq, tp->snd_una) && 1418 SEQ_LT(icmpseq, tp->snd_max)) 1419 (*notify)(inp, arg); 1420 } else { 1421 struct in_conninfo inc; 1422 1423 inc.inc_fport = th->th_dport; 1424 inc.inc_lport = th->th_sport; 1425 inc.inc_faddr = faddr; 1426 inc.inc_laddr = ip->ip_src; 1427 #ifdef INET6 1428 inc.inc_isipv6 = 0; 1429 #endif 1430 syncache_unreach(&inc, th); 1431 } 1432 crit_exit(); 1433 } else { 1434 struct netmsg_tcp_notify nmsg; 1435 1436 KKASSERT(&curthread->td_msgport == cpu_portfn(0)); 1437 netmsg_init(&nmsg.nm_nmsg, &curthread->td_msgport, 0, 1438 tcp_notifyall_oncpu); 1439 nmsg.nm_faddr = faddr; 1440 nmsg.nm_arg = arg; 1441 nmsg.nm_notify = notify; 1442 1443 lwkt_domsg(tcp_cport(0), &nmsg.nm_nmsg.nm_lmsg, 0); 1444 } 1445 } 1446 1447 #ifdef INET6 1448 void 1449 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d) 1450 { 1451 struct tcphdr th; 1452 void (*notify) (struct inpcb *, int) = tcp_notify; 1453 struct ip6_hdr *ip6; 1454 struct mbuf *m; 1455 struct ip6ctlparam *ip6cp = NULL; 1456 const struct sockaddr_in6 *sa6_src = NULL; 1457 int off; 1458 struct tcp_portonly { 1459 u_int16_t th_sport; 1460 u_int16_t th_dport; 1461 } *thp; 1462 int arg; 1463 1464 if (sa->sa_family != AF_INET6 || 1465 sa->sa_len != sizeof(struct sockaddr_in6)) 1466 return; 1467 1468 arg = 0; 1469 if (cmd == PRC_QUENCH) 1470 notify = tcp_quench; 1471 else if (cmd == PRC_MSGSIZE) { 1472 struct ip6ctlparam *ip6cp = d; 1473 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6; 1474 1475 arg = ntohl(icmp6->icmp6_mtu); 1476 notify = tcp_mtudisc; 1477 } else if (!PRC_IS_REDIRECT(cmd) && 1478 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) { 1479 return; 1480 } 1481 1482 /* if the parameter is from icmp6, decode it. */ 1483 if (d != NULL) { 1484 ip6cp = (struct ip6ctlparam *)d; 1485 m = ip6cp->ip6c_m; 1486 ip6 = ip6cp->ip6c_ip6; 1487 off = ip6cp->ip6c_off; 1488 sa6_src = ip6cp->ip6c_src; 1489 } else { 1490 m = NULL; 1491 ip6 = NULL; 1492 off = 0; /* fool gcc */ 1493 sa6_src = &sa6_any; 1494 } 1495 1496 if (ip6 != NULL) { 1497 struct in_conninfo inc; 1498 /* 1499 * XXX: We assume that when IPV6 is non NULL, 1500 * M and OFF are valid. 1501 */ 1502 1503 /* check if we can safely examine src and dst ports */ 1504 if (m->m_pkthdr.len < off + sizeof *thp) 1505 return; 1506 1507 bzero(&th, sizeof th); 1508 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1509 1510 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport, 1511 (struct sockaddr *)ip6cp->ip6c_src, 1512 th.th_sport, cmd, arg, notify); 1513 1514 inc.inc_fport = th.th_dport; 1515 inc.inc_lport = th.th_sport; 1516 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1517 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1518 inc.inc_isipv6 = 1; 1519 syncache_unreach(&inc, &th); 1520 } else 1521 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0, 1522 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify); 1523 } 1524 #endif 1525 1526 /* 1527 * Following is where TCP initial sequence number generation occurs. 1528 * 1529 * There are two places where we must use initial sequence numbers: 1530 * 1. In SYN-ACK packets. 1531 * 2. In SYN packets. 1532 * 1533 * All ISNs for SYN-ACK packets are generated by the syncache. See 1534 * tcp_syncache.c for details. 1535 * 1536 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1537 * depends on this property. In addition, these ISNs should be 1538 * unguessable so as to prevent connection hijacking. To satisfy 1539 * the requirements of this situation, the algorithm outlined in 1540 * RFC 1948 is used to generate sequence numbers. 1541 * 1542 * Implementation details: 1543 * 1544 * Time is based off the system timer, and is corrected so that it 1545 * increases by one megabyte per second. This allows for proper 1546 * recycling on high speed LANs while still leaving over an hour 1547 * before rollover. 1548 * 1549 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1550 * between seeding of isn_secret. This is normally set to zero, 1551 * as reseeding should not be necessary. 1552 * 1553 */ 1554 1555 #define ISN_BYTES_PER_SECOND 1048576 1556 1557 u_char isn_secret[32]; 1558 int isn_last_reseed; 1559 MD5_CTX isn_ctx; 1560 1561 tcp_seq 1562 tcp_new_isn(struct tcpcb *tp) 1563 { 1564 u_int32_t md5_buffer[4]; 1565 tcp_seq new_isn; 1566 1567 /* Seed if this is the first use, reseed if requested. */ 1568 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1569 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1570 < (u_int)ticks))) { 1571 read_random_unlimited(&isn_secret, sizeof isn_secret); 1572 isn_last_reseed = ticks; 1573 } 1574 1575 /* Compute the md5 hash and return the ISN. */ 1576 MD5Init(&isn_ctx); 1577 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1578 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1579 #ifdef INET6 1580 if (tp->t_inpcb->inp_vflag & INP_IPV6) { 1581 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1582 sizeof(struct in6_addr)); 1583 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1584 sizeof(struct in6_addr)); 1585 } else 1586 #endif 1587 { 1588 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1589 sizeof(struct in_addr)); 1590 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1591 sizeof(struct in_addr)); 1592 } 1593 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1594 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1595 new_isn = (tcp_seq) md5_buffer[0]; 1596 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1597 return (new_isn); 1598 } 1599 1600 /* 1601 * When a source quench is received, close congestion window 1602 * to one segment. We will gradually open it again as we proceed. 1603 */ 1604 void 1605 tcp_quench(struct inpcb *inp, int error) 1606 { 1607 struct tcpcb *tp = intotcpcb(inp); 1608 1609 if (tp != NULL) { 1610 tp->snd_cwnd = tp->t_maxseg; 1611 tp->snd_wacked = 0; 1612 } 1613 } 1614 1615 /* 1616 * When a specific ICMP unreachable message is received and the 1617 * connection state is SYN-SENT, drop the connection. This behavior 1618 * is controlled by the icmp_may_rst sysctl. 1619 */ 1620 void 1621 tcp_drop_syn_sent(struct inpcb *inp, int error) 1622 { 1623 struct tcpcb *tp = intotcpcb(inp); 1624 1625 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT)) 1626 tcp_drop(tp, error); 1627 } 1628 1629 /* 1630 * When a `need fragmentation' ICMP is received, update our idea of the MSS 1631 * based on the new value in the route. Also nudge TCP to send something, 1632 * since we know the packet we just sent was dropped. 1633 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1634 */ 1635 void 1636 tcp_mtudisc(struct inpcb *inp, int mtu) 1637 { 1638 struct tcpcb *tp = intotcpcb(inp); 1639 struct rtentry *rt; 1640 struct socket *so = inp->inp_socket; 1641 int maxopd, mss; 1642 #ifdef INET6 1643 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0); 1644 #else 1645 const boolean_t isipv6 = FALSE; 1646 #endif 1647 1648 if (tp == NULL) 1649 return; 1650 1651 /* 1652 * If no MTU is provided in the ICMP message, use the 1653 * next lower likely value, as specified in RFC 1191. 1654 */ 1655 if (mtu == 0) { 1656 int oldmtu; 1657 1658 oldmtu = tp->t_maxopd + 1659 (isipv6 ? 1660 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1661 sizeof(struct tcpiphdr)); 1662 mtu = ip_next_mtu(oldmtu, 0); 1663 } 1664 1665 if (isipv6) 1666 rt = tcp_rtlookup6(&inp->inp_inc); 1667 else 1668 rt = tcp_rtlookup(&inp->inp_inc); 1669 if (rt != NULL) { 1670 struct rmxp_tao *taop = rmx_taop(rt->rt_rmx); 1671 1672 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu) 1673 mtu = rt->rt_rmx.rmx_mtu; 1674 1675 maxopd = mtu - 1676 (isipv6 ? 1677 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1678 sizeof(struct tcpiphdr)); 1679 1680 /* 1681 * XXX - The following conditional probably violates the TCP 1682 * spec. The problem is that, since we don't know the 1683 * other end's MSS, we are supposed to use a conservative 1684 * default. But, if we do that, then MTU discovery will 1685 * never actually take place, because the conservative 1686 * default is much less than the MTUs typically seen 1687 * on the Internet today. For the moment, we'll sweep 1688 * this under the carpet. 1689 * 1690 * The conservative default might not actually be a problem 1691 * if the only case this occurs is when sending an initial 1692 * SYN with options and data to a host we've never talked 1693 * to before. Then, they will reply with an MSS value which 1694 * will get recorded and the new parameters should get 1695 * recomputed. For Further Study. 1696 */ 1697 if (taop->tao_mssopt != 0 && taop->tao_mssopt < maxopd) 1698 maxopd = taop->tao_mssopt; 1699 } else 1700 maxopd = mtu - 1701 (isipv6 ? 1702 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1703 sizeof(struct tcpiphdr)); 1704 1705 if (tp->t_maxopd <= maxopd) 1706 return; 1707 tp->t_maxopd = maxopd; 1708 1709 mss = maxopd; 1710 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) == 1711 (TF_REQ_TSTMP | TF_RCVD_TSTMP)) 1712 mss -= TCPOLEN_TSTAMP_APPA; 1713 1714 if ((tp->t_flags & (TF_REQ_CC | TF_RCVD_CC | TF_NOOPT)) == 1715 (TF_REQ_CC | TF_RCVD_CC)) 1716 mss -= TCPOLEN_CC_APPA; 1717 1718 /* round down to multiple of MCLBYTES */ 1719 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */ 1720 if (mss > MCLBYTES) 1721 mss &= ~(MCLBYTES - 1); 1722 #else 1723 if (mss > MCLBYTES) 1724 mss = (mss / MCLBYTES) * MCLBYTES; 1725 #endif 1726 1727 if (so->so_snd.ssb_hiwat < mss) 1728 mss = so->so_snd.ssb_hiwat; 1729 1730 tp->t_maxseg = mss; 1731 tp->t_rtttime = 0; 1732 tp->snd_nxt = tp->snd_una; 1733 tcp_output(tp); 1734 tcpstat.tcps_mturesent++; 1735 } 1736 1737 /* 1738 * Look-up the routing entry to the peer of this inpcb. If no route 1739 * is found and it cannot be allocated the return NULL. This routine 1740 * is called by TCP routines that access the rmx structure and by tcp_mss 1741 * to get the interface MTU. 1742 */ 1743 struct rtentry * 1744 tcp_rtlookup(struct in_conninfo *inc) 1745 { 1746 struct route *ro = &inc->inc_route; 1747 1748 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1749 /* No route yet, so try to acquire one */ 1750 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1751 /* 1752 * unused portions of the structure MUST be zero'd 1753 * out because rtalloc() treats it as opaque data 1754 */ 1755 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1756 ro->ro_dst.sa_family = AF_INET; 1757 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1758 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1759 inc->inc_faddr; 1760 rtalloc(ro); 1761 } 1762 } 1763 return (ro->ro_rt); 1764 } 1765 1766 #ifdef INET6 1767 struct rtentry * 1768 tcp_rtlookup6(struct in_conninfo *inc) 1769 { 1770 struct route_in6 *ro6 = &inc->inc6_route; 1771 1772 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1773 /* No route yet, so try to acquire one */ 1774 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1775 /* 1776 * unused portions of the structure MUST be zero'd 1777 * out because rtalloc() treats it as opaque data 1778 */ 1779 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1780 ro6->ro_dst.sin6_family = AF_INET6; 1781 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1782 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1783 rtalloc((struct route *)ro6); 1784 } 1785 } 1786 return (ro6->ro_rt); 1787 } 1788 #endif 1789 1790 #ifdef IPSEC 1791 /* compute ESP/AH header size for TCP, including outer IP header. */ 1792 size_t 1793 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1794 { 1795 struct inpcb *inp; 1796 struct mbuf *m; 1797 size_t hdrsiz; 1798 struct ip *ip; 1799 struct tcphdr *th; 1800 1801 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1802 return (0); 1803 MGETHDR(m, MB_DONTWAIT, MT_DATA); 1804 if (!m) 1805 return (0); 1806 1807 #ifdef INET6 1808 if (inp->inp_vflag & INP_IPV6) { 1809 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1810 1811 th = (struct tcphdr *)(ip6 + 1); 1812 m->m_pkthdr.len = m->m_len = 1813 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1814 tcp_fillheaders(tp, ip6, th); 1815 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1816 } else 1817 #endif 1818 { 1819 ip = mtod(m, struct ip *); 1820 th = (struct tcphdr *)(ip + 1); 1821 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1822 tcp_fillheaders(tp, ip, th); 1823 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1824 } 1825 1826 m_free(m); 1827 return (hdrsiz); 1828 } 1829 #endif 1830 1831 /* 1832 * Return a pointer to the cached information about the remote host. 1833 * The cached information is stored in the protocol specific part of 1834 * the route metrics. 1835 */ 1836 struct rmxp_tao * 1837 tcp_gettaocache(struct in_conninfo *inc) 1838 { 1839 struct rtentry *rt; 1840 1841 #ifdef INET6 1842 if (inc->inc_isipv6) 1843 rt = tcp_rtlookup6(inc); 1844 else 1845 #endif 1846 rt = tcp_rtlookup(inc); 1847 1848 /* Make sure this is a host route and is up. */ 1849 if (rt == NULL || 1850 (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST)) 1851 return (NULL); 1852 1853 return (rmx_taop(rt->rt_rmx)); 1854 } 1855 1856 /* 1857 * Clear all the TAO cache entries, called from tcp_init. 1858 * 1859 * XXX 1860 * This routine is just an empty one, because we assume that the routing 1861 * routing tables are initialized at the same time when TCP, so there is 1862 * nothing in the cache left over. 1863 */ 1864 static void 1865 tcp_cleartaocache(void) 1866 { 1867 } 1868 1869 /* 1870 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1871 * 1872 * This code attempts to calculate the bandwidth-delay product as a 1873 * means of determining the optimal window size to maximize bandwidth, 1874 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1875 * routers. This code also does a fairly good job keeping RTTs in check 1876 * across slow links like modems. We implement an algorithm which is very 1877 * similar (but not meant to be) TCP/Vegas. The code operates on the 1878 * transmitter side of a TCP connection and so only effects the transmit 1879 * side of the connection. 1880 * 1881 * BACKGROUND: TCP makes no provision for the management of buffer space 1882 * at the end points or at the intermediate routers and switches. A TCP 1883 * stream, whether using NewReno or not, will eventually buffer as 1884 * many packets as it is able and the only reason this typically works is 1885 * due to the fairly small default buffers made available for a connection 1886 * (typicaly 16K or 32K). As machines use larger windows and/or window 1887 * scaling it is now fairly easy for even a single TCP connection to blow-out 1888 * all available buffer space not only on the local interface, but on 1889 * intermediate routers and switches as well. NewReno makes a misguided 1890 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1891 * then backing off, then steadily increasing the window again until another 1892 * failure occurs, ad-infinitum. This results in terrible oscillation that 1893 * is only made worse as network loads increase and the idea of intentionally 1894 * blowing out network buffers is, frankly, a terrible way to manage network 1895 * resources. 1896 * 1897 * It is far better to limit the transmit window prior to the failure 1898 * condition being achieved. There are two general ways to do this: First 1899 * you can 'scan' through different transmit window sizes and locate the 1900 * point where the RTT stops increasing, indicating that you have filled the 1901 * pipe, then scan backwards until you note that RTT stops decreasing, then 1902 * repeat ad-infinitum. This method works in principle but has severe 1903 * implementation issues due to RTT variances, timer granularity, and 1904 * instability in the algorithm which can lead to many false positives and 1905 * create oscillations as well as interact badly with other TCP streams 1906 * implementing the same algorithm. 1907 * 1908 * The second method is to limit the window to the bandwidth delay product 1909 * of the link. This is the method we implement. RTT variances and our 1910 * own manipulation of the congestion window, bwnd, can potentially 1911 * destabilize the algorithm. For this reason we have to stabilize the 1912 * elements used to calculate the window. We do this by using the minimum 1913 * observed RTT, the long term average of the observed bandwidth, and 1914 * by adding two segments worth of slop. It isn't perfect but it is able 1915 * to react to changing conditions and gives us a very stable basis on 1916 * which to extend the algorithm. 1917 */ 1918 void 1919 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1920 { 1921 u_long bw; 1922 u_long bwnd; 1923 int save_ticks; 1924 int delta_ticks; 1925 1926 /* 1927 * If inflight_enable is disabled in the middle of a tcp connection, 1928 * make sure snd_bwnd is effectively disabled. 1929 */ 1930 if (!tcp_inflight_enable) { 1931 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1932 tp->snd_bandwidth = 0; 1933 return; 1934 } 1935 1936 /* 1937 * Validate the delta time. If a connection is new or has been idle 1938 * a long time we have to reset the bandwidth calculator. 1939 */ 1940 save_ticks = ticks; 1941 delta_ticks = save_ticks - tp->t_bw_rtttime; 1942 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1943 tp->t_bw_rtttime = ticks; 1944 tp->t_bw_rtseq = ack_seq; 1945 if (tp->snd_bandwidth == 0) 1946 tp->snd_bandwidth = tcp_inflight_min; 1947 return; 1948 } 1949 if (delta_ticks == 0) 1950 return; 1951 1952 /* 1953 * Sanity check, plus ignore pure window update acks. 1954 */ 1955 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1956 return; 1957 1958 /* 1959 * Figure out the bandwidth. Due to the tick granularity this 1960 * is a very rough number and it MUST be averaged over a fairly 1961 * long period of time. XXX we need to take into account a link 1962 * that is not using all available bandwidth, but for now our 1963 * slop will ramp us up if this case occurs and the bandwidth later 1964 * increases. 1965 */ 1966 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 1967 tp->t_bw_rtttime = save_ticks; 1968 tp->t_bw_rtseq = ack_seq; 1969 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4; 1970 1971 tp->snd_bandwidth = bw; 1972 1973 /* 1974 * Calculate the semi-static bandwidth delay product, plus two maximal 1975 * segments. The additional slop puts us squarely in the sweet 1976 * spot and also handles the bandwidth run-up case. Without the 1977 * slop we could be locking ourselves into a lower bandwidth. 1978 * 1979 * Situations Handled: 1980 * (1) Prevents over-queueing of packets on LANs, especially on 1981 * high speed LANs, allowing larger TCP buffers to be 1982 * specified, and also does a good job preventing 1983 * over-queueing of packets over choke points like modems 1984 * (at least for the transmit side). 1985 * 1986 * (2) Is able to handle changing network loads (bandwidth 1987 * drops so bwnd drops, bandwidth increases so bwnd 1988 * increases). 1989 * 1990 * (3) Theoretically should stabilize in the face of multiple 1991 * connections implementing the same algorithm (this may need 1992 * a little work). 1993 * 1994 * (4) Stability value (defaults to 20 = 2 maximal packets) can 1995 * be adjusted with a sysctl but typically only needs to be on 1996 * very slow connections. A value no smaller then 5 should 1997 * be used, but only reduce this default if you have no other 1998 * choice. 1999 */ 2000 2001 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2) 2002 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 2003 tcp_inflight_stab * (int)tp->t_maxseg / 10; 2004 #undef USERTT 2005 2006 if (tcp_inflight_debug > 0) { 2007 static int ltime; 2008 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) { 2009 ltime = ticks; 2010 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n", 2011 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd); 2012 } 2013 } 2014 if ((long)bwnd < tcp_inflight_min) 2015 bwnd = tcp_inflight_min; 2016 if (bwnd > tcp_inflight_max) 2017 bwnd = tcp_inflight_max; 2018 if ((long)bwnd < tp->t_maxseg * 2) 2019 bwnd = tp->t_maxseg * 2; 2020 tp->snd_bwnd = bwnd; 2021 } 2022