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/priv.h> 89 #include <sys/socket.h> 90 #include <sys/socketvar.h> 91 #include <sys/protosw.h> 92 #include <sys/random.h> 93 #include <sys/in_cksum.h> 94 #include <sys/ktr.h> 95 96 #include <vm/vm_zone.h> 97 98 #include <net/route.h> 99 #include <net/if.h> 100 #include <net/netisr.h> 101 102 #define _IP_VHL 103 #include <netinet/in.h> 104 #include <netinet/in_systm.h> 105 #include <netinet/ip.h> 106 #include <netinet/ip6.h> 107 #include <netinet/in_pcb.h> 108 #include <netinet6/in6_pcb.h> 109 #include <netinet/in_var.h> 110 #include <netinet/ip_var.h> 111 #include <netinet6/ip6_var.h> 112 #include <netinet/ip_icmp.h> 113 #ifdef INET6 114 #include <netinet/icmp6.h> 115 #endif 116 #include <netinet/tcp.h> 117 #include <netinet/tcp_fsm.h> 118 #include <netinet/tcp_seq.h> 119 #include <netinet/tcp_timer.h> 120 #include <netinet/tcp_timer2.h> 121 #include <netinet/tcp_var.h> 122 #include <netinet6/tcp6_var.h> 123 #include <netinet/tcpip.h> 124 #ifdef TCPDEBUG 125 #include <netinet/tcp_debug.h> 126 #endif 127 #include <netinet6/ip6protosw.h> 128 129 #ifdef IPSEC 130 #include <netinet6/ipsec.h> 131 #ifdef INET6 132 #include <netinet6/ipsec6.h> 133 #endif 134 #endif 135 136 #ifdef FAST_IPSEC 137 #include <netproto/ipsec/ipsec.h> 138 #ifdef INET6 139 #include <netproto/ipsec/ipsec6.h> 140 #endif 141 #define IPSEC 142 #endif 143 144 #include <sys/md5.h> 145 #include <sys/msgport2.h> 146 #include <machine/smp.h> 147 148 #include <net/netmsg2.h> 149 150 #if !defined(KTR_TCP) 151 #define KTR_TCP KTR_ALL 152 #endif 153 KTR_INFO_MASTER(tcp); 154 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0); 155 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0); 156 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0); 157 #define logtcp(name) KTR_LOG(tcp_ ## name) 158 159 struct inpcbinfo tcbinfo[MAXCPU]; 160 struct tcpcbackqhead tcpcbackq[MAXCPU]; 161 162 int tcp_mpsafe_proto = 0; 163 TUNABLE_INT("net.inet.tcp.mpsafe_proto", &tcp_mpsafe_proto); 164 165 static int tcp_mpsafe_thread = NETMSG_SERVICE_ADAPTIVE; 166 TUNABLE_INT("net.inet.tcp.mpsafe_thread", &tcp_mpsafe_thread); 167 SYSCTL_INT(_net_inet_tcp, OID_AUTO, mpsafe_thread, CTLFLAG_RW, 168 &tcp_mpsafe_thread, 0, 169 "0:BGL, 1:Adaptive BGL, 2:No BGL(experimental)"); 170 171 int tcp_mssdflt = TCP_MSS; 172 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 173 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size"); 174 175 #ifdef INET6 176 int tcp_v6mssdflt = TCP6_MSS; 177 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW, 178 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6"); 179 #endif 180 181 #if 0 182 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ; 183 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW, 184 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time"); 185 #endif 186 187 int tcp_do_rfc1323 = 1; 188 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW, 189 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions"); 190 191 int tcp_do_rfc1644 = 0; 192 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1644, rfc1644, CTLFLAG_RW, 193 &tcp_do_rfc1644, 0, "Enable rfc1644 (TTCP) extensions"); 194 195 static int tcp_tcbhashsize = 0; 196 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD, 197 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable"); 198 199 static int do_tcpdrain = 1; 200 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0, 201 "Enable tcp_drain routine for extra help when low on mbufs"); 202 203 /* XXX JH */ 204 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD, 205 &tcbinfo[0].ipi_count, 0, "Number of active PCBs"); 206 207 static int icmp_may_rst = 1; 208 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0, 209 "Certain ICMP unreachable messages may abort connections in SYN_SENT"); 210 211 static int tcp_isn_reseed_interval = 0; 212 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW, 213 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret"); 214 215 /* 216 * TCP bandwidth limiting sysctls. Note that the default lower bound of 217 * 1024 exists only for debugging. A good production default would be 218 * something like 6100. 219 */ 220 static int tcp_inflight_enable = 0; 221 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW, 222 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting"); 223 224 static int tcp_inflight_debug = 0; 225 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW, 226 &tcp_inflight_debug, 0, "Debug TCP inflight calculations"); 227 228 static int tcp_inflight_min = 6144; 229 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW, 230 &tcp_inflight_min, 0, "Lower bound for TCP inflight window"); 231 232 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT; 233 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW, 234 &tcp_inflight_max, 0, "Upper bound for TCP inflight window"); 235 236 static int tcp_inflight_stab = 20; 237 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW, 238 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 2 packets)"); 239 240 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives"); 241 static struct malloc_pipe tcptemp_mpipe; 242 243 static void tcp_willblock(int); 244 static void tcp_cleartaocache (void); 245 static void tcp_notify (struct inpcb *, int); 246 247 struct tcp_stats tcpstats_percpu[MAXCPU]; 248 #ifdef SMP 249 static int 250 sysctl_tcpstats(SYSCTL_HANDLER_ARGS) 251 { 252 int cpu, error = 0; 253 254 for (cpu = 0; cpu < ncpus; ++cpu) { 255 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu], 256 sizeof(struct tcp_stats)))) 257 break; 258 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu], 259 sizeof(struct tcp_stats)))) 260 break; 261 } 262 263 return (error); 264 } 265 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW), 266 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics"); 267 #else 268 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW, 269 &tcpstat, tcp_stats, "TCP statistics"); 270 #endif 271 272 /* 273 * Target size of TCP PCB hash tables. Must be a power of two. 274 * 275 * Note that this can be overridden by the kernel environment 276 * variable net.inet.tcp.tcbhashsize 277 */ 278 #ifndef TCBHASHSIZE 279 #define TCBHASHSIZE 512 280 #endif 281 282 /* 283 * This is the actual shape of what we allocate using the zone 284 * allocator. Doing it this way allows us to protect both structures 285 * using the same generation count, and also eliminates the overhead 286 * of allocating tcpcbs separately. By hiding the structure here, 287 * we avoid changing most of the rest of the code (although it needs 288 * to be changed, eventually, for greater efficiency). 289 */ 290 #define ALIGNMENT 32 291 #define ALIGNM1 (ALIGNMENT - 1) 292 struct inp_tp { 293 union { 294 struct inpcb inp; 295 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1]; 296 } inp_tp_u; 297 struct tcpcb tcb; 298 struct tcp_callout inp_tp_rexmt; 299 struct tcp_callout inp_tp_persist; 300 struct tcp_callout inp_tp_keep; 301 struct tcp_callout inp_tp_2msl; 302 struct tcp_callout inp_tp_delack; 303 struct netmsg_tcp_timer inp_tp_timermsg; 304 }; 305 #undef ALIGNMENT 306 #undef ALIGNM1 307 308 /* 309 * Tcp initialization 310 */ 311 void 312 tcp_init(void) 313 { 314 struct inpcbporthead *porthashbase; 315 u_long porthashmask; 316 struct vm_zone *ipi_zone; 317 int hashsize = TCBHASHSIZE; 318 int cpu; 319 320 /* 321 * note: tcptemp is used for keepalives, and it is ok for an 322 * allocation to fail so do not specify MPF_INT. 323 */ 324 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp), 325 25, -1, 0, NULL); 326 327 tcp_ccgen = 1; 328 tcp_cleartaocache(); 329 330 tcp_delacktime = TCPTV_DELACK; 331 tcp_keepinit = TCPTV_KEEP_INIT; 332 tcp_keepidle = TCPTV_KEEP_IDLE; 333 tcp_keepintvl = TCPTV_KEEPINTVL; 334 tcp_maxpersistidle = TCPTV_KEEP_IDLE; 335 tcp_msl = TCPTV_MSL; 336 tcp_rexmit_min = TCPTV_MIN; 337 tcp_rexmit_slop = TCPTV_CPU_VAR; 338 339 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize); 340 if (!powerof2(hashsize)) { 341 kprintf("WARNING: TCB hash size not a power of 2\n"); 342 hashsize = 512; /* safe default */ 343 } 344 tcp_tcbhashsize = hashsize; 345 porthashbase = hashinit(hashsize, M_PCB, &porthashmask); 346 ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets, 347 ZONE_INTERRUPT, 0); 348 349 for (cpu = 0; cpu < ncpus2; cpu++) { 350 in_pcbinfo_init(&tcbinfo[cpu]); 351 tcbinfo[cpu].cpu = cpu; 352 tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB, 353 &tcbinfo[cpu].hashmask); 354 tcbinfo[cpu].porthashbase = porthashbase; 355 tcbinfo[cpu].porthashmask = porthashmask; 356 tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB, 357 &tcbinfo[cpu].wildcardhashmask); 358 tcbinfo[cpu].ipi_zone = ipi_zone; 359 TAILQ_INIT(&tcpcbackq[cpu]); 360 } 361 362 tcp_reass_maxseg = nmbclusters / 16; 363 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg); 364 365 #ifdef INET6 366 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) 367 #else 368 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) 369 #endif 370 if (max_protohdr < TCP_MINPROTOHDR) 371 max_protohdr = TCP_MINPROTOHDR; 372 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) 373 panic("tcp_init"); 374 #undef TCP_MINPROTOHDR 375 376 /* 377 * Initialize TCP statistics counters for each CPU. 378 */ 379 #ifdef SMP 380 for (cpu = 0; cpu < ncpus; ++cpu) { 381 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats)); 382 } 383 #else 384 bzero(&tcpstat, sizeof(struct tcp_stats)); 385 #endif 386 387 syncache_init(); 388 tcp_thread_init(); 389 } 390 391 void 392 tcpmsg_service_loop(void *dummy) 393 { 394 struct netmsg *msg; 395 int mplocked; 396 397 /* 398 * Thread was started with TDF_MPSAFE 399 */ 400 mplocked = 0; 401 402 while ((msg = lwkt_waitport(&curthread->td_msgport, 0))) { 403 do { 404 logtcp(rxmsg); 405 mplocked = netmsg_service(msg, tcp_mpsafe_thread, 406 mplocked); 407 } while ((msg = lwkt_getport(&curthread->td_msgport)) != NULL); 408 409 logtcp(delayed); 410 tcp_willblock(mplocked); 411 logtcp(wait); 412 } 413 } 414 415 static void 416 tcp_willblock(int mplocked) 417 { 418 struct tcpcb *tp; 419 int cpu = mycpu->gd_cpuid; 420 int unlock = 0; 421 422 if (!mplocked && !tcp_mpsafe_proto) { 423 if (TAILQ_EMPTY(&tcpcbackq[cpu])) 424 return; 425 426 get_mplock(); 427 mplocked = 1; 428 unlock = 1; 429 } 430 431 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) { 432 KKASSERT(tp->t_flags & TF_ONOUTPUTQ); 433 tp->t_flags &= ~TF_ONOUTPUTQ; 434 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq); 435 tcp_output(tp); 436 } 437 438 if (unlock) 439 rel_mplock(); 440 } 441 442 443 /* 444 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. 445 * tcp_template used to store this data in mbufs, but we now recopy it out 446 * of the tcpcb each time to conserve mbufs. 447 */ 448 void 449 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr) 450 { 451 struct inpcb *inp = tp->t_inpcb; 452 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr; 453 454 #ifdef INET6 455 if (inp->inp_vflag & INP_IPV6) { 456 struct ip6_hdr *ip6; 457 458 ip6 = (struct ip6_hdr *)ip_ptr; 459 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | 460 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK); 461 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | 462 (IPV6_VERSION & IPV6_VERSION_MASK); 463 ip6->ip6_nxt = IPPROTO_TCP; 464 ip6->ip6_plen = sizeof(struct tcphdr); 465 ip6->ip6_src = inp->in6p_laddr; 466 ip6->ip6_dst = inp->in6p_faddr; 467 tcp_hdr->th_sum = 0; 468 } else 469 #endif 470 { 471 struct ip *ip = (struct ip *) ip_ptr; 472 473 ip->ip_vhl = IP_VHL_BORING; 474 ip->ip_tos = 0; 475 ip->ip_len = 0; 476 ip->ip_id = 0; 477 ip->ip_off = 0; 478 ip->ip_ttl = 0; 479 ip->ip_sum = 0; 480 ip->ip_p = IPPROTO_TCP; 481 ip->ip_src = inp->inp_laddr; 482 ip->ip_dst = inp->inp_faddr; 483 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr, 484 ip->ip_dst.s_addr, 485 htons(sizeof(struct tcphdr) + IPPROTO_TCP)); 486 } 487 488 tcp_hdr->th_sport = inp->inp_lport; 489 tcp_hdr->th_dport = inp->inp_fport; 490 tcp_hdr->th_seq = 0; 491 tcp_hdr->th_ack = 0; 492 tcp_hdr->th_x2 = 0; 493 tcp_hdr->th_off = 5; 494 tcp_hdr->th_flags = 0; 495 tcp_hdr->th_win = 0; 496 tcp_hdr->th_urp = 0; 497 } 498 499 /* 500 * Create template to be used to send tcp packets on a connection. 501 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only 502 * use for this function is in keepalives, which use tcp_respond. 503 */ 504 struct tcptemp * 505 tcp_maketemplate(struct tcpcb *tp) 506 { 507 struct tcptemp *tmp; 508 509 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL) 510 return (NULL); 511 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t); 512 return (tmp); 513 } 514 515 void 516 tcp_freetemplate(struct tcptemp *tmp) 517 { 518 mpipe_free(&tcptemp_mpipe, tmp); 519 } 520 521 /* 522 * Send a single message to the TCP at address specified by 523 * the given TCP/IP header. If m == NULL, then we make a copy 524 * of the tcpiphdr at ti and send directly to the addressed host. 525 * This is used to force keep alive messages out using the TCP 526 * template for a connection. If flags are given then we send 527 * a message back to the TCP which originated the * segment ti, 528 * and discard the mbuf containing it and any other attached mbufs. 529 * 530 * In any case the ack and sequence number of the transmitted 531 * segment are as specified by the parameters. 532 * 533 * NOTE: If m != NULL, then ti must point to *inside* the mbuf. 534 */ 535 void 536 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, 537 tcp_seq ack, tcp_seq seq, int flags) 538 { 539 int tlen; 540 int win = 0; 541 struct route *ro = NULL; 542 struct route sro; 543 struct ip *ip = ipgen; 544 struct tcphdr *nth; 545 int ipflags = 0; 546 struct route_in6 *ro6 = NULL; 547 struct route_in6 sro6; 548 struct ip6_hdr *ip6 = ipgen; 549 boolean_t use_tmpro = TRUE; 550 #ifdef INET6 551 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6); 552 #else 553 const boolean_t isipv6 = FALSE; 554 #endif 555 556 if (tp != NULL) { 557 if (!(flags & TH_RST)) { 558 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv); 559 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 560 win = (long)TCP_MAXWIN << tp->rcv_scale; 561 } 562 /* 563 * Don't use the route cache of a listen socket, 564 * it is not MPSAFE; use temporary route cache. 565 */ 566 if (tp->t_state != TCPS_LISTEN) { 567 if (isipv6) 568 ro6 = &tp->t_inpcb->in6p_route; 569 else 570 ro = &tp->t_inpcb->inp_route; 571 use_tmpro = FALSE; 572 } 573 } 574 if (use_tmpro) { 575 if (isipv6) { 576 ro6 = &sro6; 577 bzero(ro6, sizeof *ro6); 578 } else { 579 ro = &sro; 580 bzero(ro, sizeof *ro); 581 } 582 } 583 if (m == NULL) { 584 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 585 if (m == NULL) 586 return; 587 tlen = 0; 588 m->m_data += max_linkhdr; 589 if (isipv6) { 590 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); 591 ip6 = mtod(m, struct ip6_hdr *); 592 nth = (struct tcphdr *)(ip6 + 1); 593 } else { 594 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip)); 595 ip = mtod(m, struct ip *); 596 nth = (struct tcphdr *)(ip + 1); 597 } 598 bcopy(th, nth, sizeof(struct tcphdr)); 599 flags = TH_ACK; 600 } else { 601 m_freem(m->m_next); 602 m->m_next = NULL; 603 m->m_data = (caddr_t)ipgen; 604 /* m_len is set later */ 605 tlen = 0; 606 #define xchg(a, b, type) { type t; t = a; a = b; b = t; } 607 if (isipv6) { 608 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); 609 nth = (struct tcphdr *)(ip6 + 1); 610 } else { 611 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long); 612 nth = (struct tcphdr *)(ip + 1); 613 } 614 if (th != nth) { 615 /* 616 * this is usually a case when an extension header 617 * exists between the IPv6 header and the 618 * TCP header. 619 */ 620 nth->th_sport = th->th_sport; 621 nth->th_dport = th->th_dport; 622 } 623 xchg(nth->th_dport, nth->th_sport, n_short); 624 #undef xchg 625 } 626 if (isipv6) { 627 ip6->ip6_flow = 0; 628 ip6->ip6_vfc = IPV6_VERSION; 629 ip6->ip6_nxt = IPPROTO_TCP; 630 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen)); 631 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 632 } else { 633 tlen += sizeof(struct tcpiphdr); 634 ip->ip_len = tlen; 635 ip->ip_ttl = ip_defttl; 636 } 637 m->m_len = tlen; 638 m->m_pkthdr.len = tlen; 639 m->m_pkthdr.rcvif = (struct ifnet *) NULL; 640 nth->th_seq = htonl(seq); 641 nth->th_ack = htonl(ack); 642 nth->th_x2 = 0; 643 nth->th_off = sizeof(struct tcphdr) >> 2; 644 nth->th_flags = flags; 645 if (tp != NULL) 646 nth->th_win = htons((u_short) (win >> tp->rcv_scale)); 647 else 648 nth->th_win = htons((u_short)win); 649 nth->th_urp = 0; 650 if (isipv6) { 651 nth->th_sum = 0; 652 nth->th_sum = in6_cksum(m, IPPROTO_TCP, 653 sizeof(struct ip6_hdr), 654 tlen - sizeof(struct ip6_hdr)); 655 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL, 656 (ro6 && ro6->ro_rt) ? 657 ro6->ro_rt->rt_ifp : NULL); 658 } else { 659 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 660 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); 661 m->m_pkthdr.csum_flags = CSUM_TCP; 662 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 663 } 664 #ifdef TCPDEBUG 665 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) 666 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); 667 #endif 668 if (isipv6) { 669 ip6_output(m, NULL, ro6, ipflags, NULL, NULL, 670 tp ? tp->t_inpcb : NULL); 671 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) { 672 RTFREE(ro6->ro_rt); 673 ro6->ro_rt = NULL; 674 } 675 } else { 676 ipflags |= IP_DEBUGROUTE; 677 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL); 678 if ((ro == &sro) && (ro->ro_rt != NULL)) { 679 RTFREE(ro->ro_rt); 680 ro->ro_rt = NULL; 681 } 682 } 683 } 684 685 /* 686 * Create a new TCP control block, making an 687 * empty reassembly queue and hooking it to the argument 688 * protocol control block. The `inp' parameter must have 689 * come from the zone allocator set up in tcp_init(). 690 */ 691 struct tcpcb * 692 tcp_newtcpcb(struct inpcb *inp) 693 { 694 struct inp_tp *it; 695 struct tcpcb *tp; 696 #ifdef INET6 697 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 698 #else 699 const boolean_t isipv6 = FALSE; 700 #endif 701 702 it = (struct inp_tp *)inp; 703 tp = &it->tcb; 704 bzero(tp, sizeof(struct tcpcb)); 705 LIST_INIT(&tp->t_segq); 706 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 707 708 /* Set up our timeouts. */ 709 tp->tt_rexmt = &it->inp_tp_rexmt; 710 tp->tt_persist = &it->inp_tp_persist; 711 tp->tt_keep = &it->inp_tp_keep; 712 tp->tt_2msl = &it->inp_tp_2msl; 713 tp->tt_delack = &it->inp_tp_delack; 714 tcp_inittimers(tp); 715 716 tp->tt_msg = &it->inp_tp_timermsg; 717 if (isipv6) { 718 /* Don't mess with IPv6; always create timer message */ 719 tcp_create_timermsg(tp); 720 } else { 721 /* 722 * Zero out timer message. We don't create it here, 723 * since the current CPU may not be the owner of this 724 * inpcb. 725 */ 726 bzero(tp->tt_msg, sizeof(*tp->tt_msg)); 727 } 728 729 if (tcp_do_rfc1323) 730 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP); 731 if (tcp_do_rfc1644) 732 tp->t_flags |= TF_REQ_CC; 733 tp->t_inpcb = inp; /* XXX */ 734 tp->t_state = TCPS_CLOSED; 735 /* 736 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 737 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 738 * reasonable initial retransmit time. 739 */ 740 tp->t_srtt = TCPTV_SRTTBASE; 741 tp->t_rttvar = 742 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 743 tp->t_rttmin = tcp_rexmit_min; 744 tp->t_rxtcur = TCPTV_RTOBASE; 745 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 746 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 747 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 748 tp->t_rcvtime = ticks; 749 /* 750 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 751 * because the socket may be bound to an IPv6 wildcard address, 752 * which may match an IPv4-mapped IPv6 address. 753 */ 754 inp->inp_ip_ttl = ip_defttl; 755 inp->inp_ppcb = tp; 756 tcp_sack_tcpcb_init(tp); 757 return (tp); /* XXX */ 758 } 759 760 /* 761 * Drop a TCP connection, reporting the specified error. 762 * If connection is synchronized, then send a RST to peer. 763 */ 764 struct tcpcb * 765 tcp_drop(struct tcpcb *tp, int error) 766 { 767 struct socket *so = tp->t_inpcb->inp_socket; 768 769 if (TCPS_HAVERCVDSYN(tp->t_state)) { 770 tp->t_state = TCPS_CLOSED; 771 tcp_output(tp); 772 tcpstat.tcps_drops++; 773 } else 774 tcpstat.tcps_conndrops++; 775 if (error == ETIMEDOUT && tp->t_softerror) 776 error = tp->t_softerror; 777 so->so_error = error; 778 return (tcp_close(tp)); 779 } 780 781 #ifdef SMP 782 783 struct netmsg_remwildcard { 784 struct netmsg nm_netmsg; 785 struct inpcb *nm_inp; 786 struct inpcbinfo *nm_pcbinfo; 787 #if defined(INET6) 788 int nm_isinet6; 789 #else 790 int nm_unused01; 791 #endif 792 }; 793 794 /* 795 * Wildcard inpcb's on SMP boxes must be removed from all cpus before the 796 * inp can be detached. We do this by cycling through the cpus, ending up 797 * on the cpu controlling the inp last and then doing the disconnect. 798 */ 799 static void 800 in_pcbremwildcardhash_handler(struct netmsg *msg0) 801 { 802 struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0; 803 int cpu; 804 805 cpu = msg->nm_pcbinfo->cpu; 806 807 if (cpu == msg->nm_inp->inp_pcbinfo->cpu) { 808 /* note: detach removes any wildcard hash entry */ 809 #ifdef INET6 810 if (msg->nm_isinet6) 811 in6_pcbdetach(msg->nm_inp); 812 else 813 #endif 814 in_pcbdetach(msg->nm_inp); 815 lwkt_replymsg(&msg->nm_netmsg.nm_lmsg, 0); 816 } else { 817 in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo); 818 cpu = (cpu + 1) % ncpus2; 819 msg->nm_pcbinfo = &tcbinfo[cpu]; 820 lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg); 821 } 822 } 823 824 #endif 825 826 /* 827 * Close a TCP control block: 828 * discard all space held by the tcp 829 * discard internet protocol block 830 * wake up any sleepers 831 */ 832 struct tcpcb * 833 tcp_close(struct tcpcb *tp) 834 { 835 struct tseg_qent *q; 836 struct inpcb *inp = tp->t_inpcb; 837 struct socket *so = inp->inp_socket; 838 struct rtentry *rt; 839 boolean_t dosavessthresh; 840 #ifdef SMP 841 int cpu; 842 #endif 843 #ifdef INET6 844 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 845 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0); 846 #else 847 const boolean_t isipv6 = FALSE; 848 #endif 849 850 /* 851 * The tp is not instantly destroyed in the wildcard case. Setting 852 * the state to TCPS_TERMINATING will prevent the TCP stack from 853 * messing with it, though it should be noted that this change may 854 * not take effect on other cpus until we have chained the wildcard 855 * hash removal. 856 * 857 * XXX we currently depend on the BGL to synchronize the tp->t_state 858 * update and prevent other tcp protocol threads from accepting new 859 * connections on the listen socket we might be trying to close down. 860 */ 861 KKASSERT(tp->t_state != TCPS_TERMINATING); 862 tp->t_state = TCPS_TERMINATING; 863 864 /* 865 * Make sure that all of our timers are stopped before we 866 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL), 867 * timers are never used. 868 */ 869 if (tp->tt_msg != NULL) { 870 tcp_callout_stop(tp, tp->tt_rexmt); 871 tcp_callout_stop(tp, tp->tt_persist); 872 tcp_callout_stop(tp, tp->tt_keep); 873 tcp_callout_stop(tp, tp->tt_2msl); 874 tcp_callout_stop(tp, tp->tt_delack); 875 } 876 877 if (tp->t_flags & TF_ONOUTPUTQ) { 878 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 879 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 880 tp->t_flags &= ~TF_ONOUTPUTQ; 881 } 882 883 /* 884 * If we got enough samples through the srtt filter, 885 * save the rtt and rttvar in the routing entry. 886 * 'Enough' is arbitrarily defined as the 16 samples. 887 * 16 samples is enough for the srtt filter to converge 888 * to within 5% of the correct value; fewer samples and 889 * we could save a very bogus rtt. 890 * 891 * Don't update the default route's characteristics and don't 892 * update anything that the user "locked". 893 */ 894 if (tp->t_rttupdated >= 16) { 895 u_long i = 0; 896 897 if (isipv6) { 898 struct sockaddr_in6 *sin6; 899 900 if ((rt = inp->in6p_route.ro_rt) == NULL) 901 goto no_valid_rt; 902 sin6 = (struct sockaddr_in6 *)rt_key(rt); 903 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 904 goto no_valid_rt; 905 } else 906 if ((rt = inp->inp_route.ro_rt) == NULL || 907 ((struct sockaddr_in *)rt_key(rt))-> 908 sin_addr.s_addr == INADDR_ANY) 909 goto no_valid_rt; 910 911 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 912 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 913 if (rt->rt_rmx.rmx_rtt && i) 914 /* 915 * filter this update to half the old & half 916 * the new values, converting scale. 917 * See route.h and tcp_var.h for a 918 * description of the scaling constants. 919 */ 920 rt->rt_rmx.rmx_rtt = 921 (rt->rt_rmx.rmx_rtt + i) / 2; 922 else 923 rt->rt_rmx.rmx_rtt = i; 924 tcpstat.tcps_cachedrtt++; 925 } 926 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 927 i = tp->t_rttvar * 928 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 929 if (rt->rt_rmx.rmx_rttvar && i) 930 rt->rt_rmx.rmx_rttvar = 931 (rt->rt_rmx.rmx_rttvar + i) / 2; 932 else 933 rt->rt_rmx.rmx_rttvar = i; 934 tcpstat.tcps_cachedrttvar++; 935 } 936 /* 937 * The old comment here said: 938 * update the pipelimit (ssthresh) if it has been updated 939 * already or if a pipesize was specified & the threshhold 940 * got below half the pipesize. I.e., wait for bad news 941 * before we start updating, then update on both good 942 * and bad news. 943 * 944 * But we want to save the ssthresh even if no pipesize is 945 * specified explicitly in the route, because such 946 * connections still have an implicit pipesize specified 947 * by the global tcp_sendspace. In the absence of a reliable 948 * way to calculate the pipesize, it will have to do. 949 */ 950 i = tp->snd_ssthresh; 951 if (rt->rt_rmx.rmx_sendpipe != 0) 952 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 953 else 954 dosavessthresh = (i < so->so_snd.ssb_hiwat/2); 955 if (dosavessthresh || 956 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 957 (rt->rt_rmx.rmx_ssthresh != 0))) { 958 /* 959 * convert the limit from user data bytes to 960 * packets then to packet data bytes. 961 */ 962 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 963 if (i < 2) 964 i = 2; 965 i *= tp->t_maxseg + 966 (isipv6 ? 967 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 968 sizeof(struct tcpiphdr)); 969 if (rt->rt_rmx.rmx_ssthresh) 970 rt->rt_rmx.rmx_ssthresh = 971 (rt->rt_rmx.rmx_ssthresh + i) / 2; 972 else 973 rt->rt_rmx.rmx_ssthresh = i; 974 tcpstat.tcps_cachedssthresh++; 975 } 976 } 977 978 no_valid_rt: 979 /* free the reassembly queue, if any */ 980 while((q = LIST_FIRST(&tp->t_segq)) != NULL) { 981 LIST_REMOVE(q, tqe_q); 982 m_freem(q->tqe_m); 983 FREE(q, M_TSEGQ); 984 tcp_reass_qsize--; 985 } 986 /* throw away SACK blocks in scoreboard*/ 987 if (TCP_DO_SACK(tp)) 988 tcp_sack_cleanup(&tp->scb); 989 990 inp->inp_ppcb = NULL; 991 soisdisconnected(so); 992 993 tcp_destroy_timermsg(tp); 994 995 /* 996 * Discard the inp. In the SMP case a wildcard inp's hash (created 997 * by a listen socket or an INADDR_ANY udp socket) is replicated 998 * for each protocol thread and must be removed in the context of 999 * that thread. This is accomplished by chaining the message 1000 * through the cpus. 1001 * 1002 * If the inp is not wildcarded we simply detach, which will remove 1003 * the any hashes still present for this inp. 1004 */ 1005 #ifdef SMP 1006 if (inp->inp_flags & INP_WILDCARD_MP) { 1007 struct netmsg_remwildcard *msg; 1008 1009 cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2; 1010 msg = kmalloc(sizeof(struct netmsg_remwildcard), 1011 M_LWKTMSG, M_INTWAIT); 1012 netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0, 1013 in_pcbremwildcardhash_handler); 1014 #ifdef INET6 1015 msg->nm_isinet6 = isafinet6; 1016 #endif 1017 msg->nm_inp = inp; 1018 msg->nm_pcbinfo = &tcbinfo[cpu]; 1019 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg); 1020 } else 1021 #endif 1022 { 1023 /* note: detach removes any wildcard hash entry */ 1024 #ifdef INET6 1025 if (isafinet6) 1026 in6_pcbdetach(inp); 1027 else 1028 #endif 1029 in_pcbdetach(inp); 1030 } 1031 tcpstat.tcps_closed++; 1032 return (NULL); 1033 } 1034 1035 static __inline void 1036 tcp_drain_oncpu(struct inpcbhead *head) 1037 { 1038 struct inpcb *inpb; 1039 struct tcpcb *tcpb; 1040 struct tseg_qent *te; 1041 1042 LIST_FOREACH(inpb, head, inp_list) { 1043 if (inpb->inp_flags & INP_PLACEMARKER) 1044 continue; 1045 if ((tcpb = intotcpcb(inpb))) { 1046 while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) { 1047 LIST_REMOVE(te, tqe_q); 1048 m_freem(te->tqe_m); 1049 FREE(te, M_TSEGQ); 1050 tcp_reass_qsize--; 1051 } 1052 } 1053 } 1054 } 1055 1056 #ifdef SMP 1057 struct netmsg_tcp_drain { 1058 struct netmsg nm_netmsg; 1059 struct inpcbhead *nm_head; 1060 }; 1061 1062 static void 1063 tcp_drain_handler(netmsg_t netmsg) 1064 { 1065 struct netmsg_tcp_drain *nm = (void *)netmsg; 1066 1067 tcp_drain_oncpu(nm->nm_head); 1068 lwkt_replymsg(&nm->nm_netmsg.nm_lmsg, 0); 1069 } 1070 #endif 1071 1072 void 1073 tcp_drain(void) 1074 { 1075 #ifdef SMP 1076 int cpu; 1077 #endif 1078 1079 if (!do_tcpdrain) 1080 return; 1081 1082 /* 1083 * Walk the tcpbs, if existing, and flush the reassembly queue, 1084 * if there is one... 1085 * XXX: The "Net/3" implementation doesn't imply that the TCP 1086 * reassembly queue should be flushed, but in a situation 1087 * where we're really low on mbufs, this is potentially 1088 * useful. 1089 */ 1090 #ifdef SMP 1091 for (cpu = 0; cpu < ncpus2; cpu++) { 1092 struct netmsg_tcp_drain *msg; 1093 1094 if (cpu == mycpu->gd_cpuid) { 1095 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead); 1096 } else { 1097 msg = kmalloc(sizeof(struct netmsg_tcp_drain), 1098 M_LWKTMSG, M_NOWAIT); 1099 if (msg == NULL) 1100 continue; 1101 netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0, 1102 tcp_drain_handler); 1103 msg->nm_head = &tcbinfo[cpu].pcblisthead; 1104 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg); 1105 } 1106 } 1107 #else 1108 tcp_drain_oncpu(&tcbinfo[0].pcblisthead); 1109 #endif 1110 } 1111 1112 /* 1113 * Notify a tcp user of an asynchronous error; 1114 * store error as soft error, but wake up user 1115 * (for now, won't do anything until can select for soft error). 1116 * 1117 * Do not wake up user since there currently is no mechanism for 1118 * reporting soft errors (yet - a kqueue filter may be added). 1119 */ 1120 static void 1121 tcp_notify(struct inpcb *inp, int error) 1122 { 1123 struct tcpcb *tp = intotcpcb(inp); 1124 1125 /* 1126 * Ignore some errors if we are hooked up. 1127 * If connection hasn't completed, has retransmitted several times, 1128 * and receives a second error, give up now. This is better 1129 * than waiting a long time to establish a connection that 1130 * can never complete. 1131 */ 1132 if (tp->t_state == TCPS_ESTABLISHED && 1133 (error == EHOSTUNREACH || error == ENETUNREACH || 1134 error == EHOSTDOWN)) { 1135 return; 1136 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1137 tp->t_softerror) 1138 tcp_drop(tp, error); 1139 else 1140 tp->t_softerror = error; 1141 #if 0 1142 wakeup(&so->so_timeo); 1143 sorwakeup(so); 1144 sowwakeup(so); 1145 #endif 1146 } 1147 1148 static int 1149 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1150 { 1151 int error, i, n; 1152 struct inpcb *marker; 1153 struct inpcb *inp; 1154 inp_gen_t gencnt; 1155 globaldata_t gd; 1156 int origcpu, ccpu; 1157 1158 error = 0; 1159 n = 0; 1160 1161 /* 1162 * The process of preparing the TCB list is too time-consuming and 1163 * resource-intensive to repeat twice on every request. 1164 */ 1165 if (req->oldptr == NULL) { 1166 for (ccpu = 0; ccpu < ncpus; ++ccpu) { 1167 gd = globaldata_find(ccpu); 1168 n += tcbinfo[gd->gd_cpuid].ipi_count; 1169 } 1170 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1171 return (0); 1172 } 1173 1174 if (req->newptr != NULL) 1175 return (EPERM); 1176 1177 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1178 marker->inp_flags |= INP_PLACEMARKER; 1179 1180 /* 1181 * OK, now we're committed to doing something. Run the inpcb list 1182 * for each cpu in the system and construct the output. Use a 1183 * list placemarker to deal with list changes occuring during 1184 * copyout blockages (but otherwise depend on being on the correct 1185 * cpu to avoid races). 1186 */ 1187 origcpu = mycpu->gd_cpuid; 1188 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) { 1189 globaldata_t rgd; 1190 caddr_t inp_ppcb; 1191 struct xtcpcb xt; 1192 int cpu_id; 1193 1194 cpu_id = (origcpu + ccpu) % ncpus; 1195 if ((smp_active_mask & (1 << cpu_id)) == 0) 1196 continue; 1197 rgd = globaldata_find(cpu_id); 1198 lwkt_setcpu_self(rgd); 1199 1200 gencnt = tcbinfo[cpu_id].ipi_gencnt; 1201 n = tcbinfo[cpu_id].ipi_count; 1202 1203 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list); 1204 i = 0; 1205 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1206 /* 1207 * process a snapshot of pcbs, ignoring placemarkers 1208 * and using our own to allow SYSCTL_OUT to block. 1209 */ 1210 LIST_REMOVE(marker, inp_list); 1211 LIST_INSERT_AFTER(inp, marker, inp_list); 1212 1213 if (inp->inp_flags & INP_PLACEMARKER) 1214 continue; 1215 if (inp->inp_gencnt > gencnt) 1216 continue; 1217 if (prison_xinpcb(req->td, inp)) 1218 continue; 1219 1220 xt.xt_len = sizeof xt; 1221 bcopy(inp, &xt.xt_inp, sizeof *inp); 1222 inp_ppcb = inp->inp_ppcb; 1223 if (inp_ppcb != NULL) 1224 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1225 else 1226 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1227 if (inp->inp_socket) 1228 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1229 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1230 break; 1231 ++i; 1232 } 1233 LIST_REMOVE(marker, inp_list); 1234 if (error == 0 && i < n) { 1235 bzero(&xt, sizeof xt); 1236 xt.xt_len = sizeof xt; 1237 while (i < n) { 1238 error = SYSCTL_OUT(req, &xt, sizeof xt); 1239 if (error) 1240 break; 1241 ++i; 1242 } 1243 } 1244 } 1245 1246 /* 1247 * Make sure we are on the same cpu we were on originally, since 1248 * higher level callers expect this. Also don't pollute caches with 1249 * migrated userland data by (eventually) returning to userland 1250 * on a different cpu. 1251 */ 1252 lwkt_setcpu_self(globaldata_find(origcpu)); 1253 kfree(marker, M_TEMP); 1254 return (error); 1255 } 1256 1257 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1258 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1259 1260 static int 1261 tcp_getcred(SYSCTL_HANDLER_ARGS) 1262 { 1263 struct sockaddr_in addrs[2]; 1264 struct inpcb *inp; 1265 int cpu; 1266 int error; 1267 1268 error = priv_check(req->td, PRIV_ROOT); 1269 if (error != 0) 1270 return (error); 1271 error = SYSCTL_IN(req, addrs, sizeof addrs); 1272 if (error != 0) 1273 return (error); 1274 crit_enter(); 1275 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1276 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1277 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1278 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1279 if (inp == NULL || inp->inp_socket == NULL) { 1280 error = ENOENT; 1281 goto out; 1282 } 1283 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1284 out: 1285 crit_exit(); 1286 return (error); 1287 } 1288 1289 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1290 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1291 1292 #ifdef INET6 1293 static int 1294 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1295 { 1296 struct sockaddr_in6 addrs[2]; 1297 struct inpcb *inp; 1298 int error; 1299 boolean_t mapped = FALSE; 1300 1301 error = priv_check(req->td, PRIV_ROOT); 1302 if (error != 0) 1303 return (error); 1304 error = SYSCTL_IN(req, addrs, sizeof addrs); 1305 if (error != 0) 1306 return (error); 1307 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { 1308 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) 1309 mapped = TRUE; 1310 else 1311 return (EINVAL); 1312 } 1313 crit_enter(); 1314 if (mapped) { 1315 inp = in_pcblookup_hash(&tcbinfo[0], 1316 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], 1317 addrs[1].sin6_port, 1318 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], 1319 addrs[0].sin6_port, 1320 0, NULL); 1321 } else { 1322 inp = in6_pcblookup_hash(&tcbinfo[0], 1323 &addrs[1].sin6_addr, addrs[1].sin6_port, 1324 &addrs[0].sin6_addr, addrs[0].sin6_port, 1325 0, NULL); 1326 } 1327 if (inp == NULL || inp->inp_socket == NULL) { 1328 error = ENOENT; 1329 goto out; 1330 } 1331 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1332 out: 1333 crit_exit(); 1334 return (error); 1335 } 1336 1337 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1338 0, 0, 1339 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1340 #endif 1341 1342 struct netmsg_tcp_notify { 1343 struct netmsg nm_nmsg; 1344 void (*nm_notify)(struct inpcb *, int); 1345 struct in_addr nm_faddr; 1346 int nm_arg; 1347 }; 1348 1349 static void 1350 tcp_notifyall_oncpu(struct netmsg *netmsg) 1351 { 1352 struct netmsg_tcp_notify *nmsg = (struct netmsg_tcp_notify *)netmsg; 1353 int nextcpu; 1354 1355 in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nmsg->nm_faddr, 1356 nmsg->nm_arg, nmsg->nm_notify); 1357 1358 nextcpu = mycpuid + 1; 1359 if (nextcpu < ncpus2) 1360 lwkt_forwardmsg(tcp_cport(nextcpu), &netmsg->nm_lmsg); 1361 else 1362 lwkt_replymsg(&netmsg->nm_lmsg, 0); 1363 } 1364 1365 void 1366 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip) 1367 { 1368 struct ip *ip = vip; 1369 struct tcphdr *th; 1370 struct in_addr faddr; 1371 struct inpcb *inp; 1372 struct tcpcb *tp; 1373 void (*notify)(struct inpcb *, int) = tcp_notify; 1374 tcp_seq icmpseq; 1375 int arg, cpu; 1376 1377 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) { 1378 return; 1379 } 1380 1381 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1382 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1383 return; 1384 1385 arg = inetctlerrmap[cmd]; 1386 if (cmd == PRC_QUENCH) { 1387 notify = tcp_quench; 1388 } else if (icmp_may_rst && 1389 (cmd == PRC_UNREACH_ADMIN_PROHIB || 1390 cmd == PRC_UNREACH_PORT || 1391 cmd == PRC_TIMXCEED_INTRANS) && 1392 ip != NULL) { 1393 notify = tcp_drop_syn_sent; 1394 } else if (cmd == PRC_MSGSIZE) { 1395 struct icmp *icmp = (struct icmp *) 1396 ((caddr_t)ip - offsetof(struct icmp, icmp_ip)); 1397 1398 arg = ntohs(icmp->icmp_nextmtu); 1399 notify = tcp_mtudisc; 1400 } else if (PRC_IS_REDIRECT(cmd)) { 1401 ip = NULL; 1402 notify = in_rtchange; 1403 } else if (cmd == PRC_HOSTDEAD) { 1404 ip = NULL; 1405 } 1406 1407 if (ip != NULL) { 1408 crit_enter(); 1409 th = (struct tcphdr *)((caddr_t)ip + 1410 (IP_VHL_HL(ip->ip_vhl) << 2)); 1411 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport, 1412 ip->ip_src.s_addr, th->th_sport); 1413 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport, 1414 ip->ip_src, th->th_sport, 0, NULL); 1415 if ((inp != NULL) && (inp->inp_socket != NULL)) { 1416 icmpseq = htonl(th->th_seq); 1417 tp = intotcpcb(inp); 1418 if (SEQ_GEQ(icmpseq, tp->snd_una) && 1419 SEQ_LT(icmpseq, tp->snd_max)) 1420 (*notify)(inp, arg); 1421 } else { 1422 struct in_conninfo inc; 1423 1424 inc.inc_fport = th->th_dport; 1425 inc.inc_lport = th->th_sport; 1426 inc.inc_faddr = faddr; 1427 inc.inc_laddr = ip->ip_src; 1428 #ifdef INET6 1429 inc.inc_isipv6 = 0; 1430 #endif 1431 syncache_unreach(&inc, th); 1432 } 1433 crit_exit(); 1434 } else { 1435 struct netmsg_tcp_notify nmsg; 1436 1437 KKASSERT(&curthread->td_msgport == cpu_portfn(0)); 1438 netmsg_init(&nmsg.nm_nmsg, &curthread->td_msgport, 0, 1439 tcp_notifyall_oncpu); 1440 nmsg.nm_faddr = faddr; 1441 nmsg.nm_arg = arg; 1442 nmsg.nm_notify = notify; 1443 1444 lwkt_domsg(tcp_cport(0), &nmsg.nm_nmsg.nm_lmsg, 0); 1445 } 1446 } 1447 1448 #ifdef INET6 1449 void 1450 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d) 1451 { 1452 struct tcphdr th; 1453 void (*notify) (struct inpcb *, int) = tcp_notify; 1454 struct ip6_hdr *ip6; 1455 struct mbuf *m; 1456 struct ip6ctlparam *ip6cp = NULL; 1457 const struct sockaddr_in6 *sa6_src = NULL; 1458 int off; 1459 struct tcp_portonly { 1460 u_int16_t th_sport; 1461 u_int16_t th_dport; 1462 } *thp; 1463 int arg; 1464 1465 if (sa->sa_family != AF_INET6 || 1466 sa->sa_len != sizeof(struct sockaddr_in6)) 1467 return; 1468 1469 arg = 0; 1470 if (cmd == PRC_QUENCH) 1471 notify = tcp_quench; 1472 else if (cmd == PRC_MSGSIZE) { 1473 struct ip6ctlparam *ip6cp = d; 1474 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6; 1475 1476 arg = ntohl(icmp6->icmp6_mtu); 1477 notify = tcp_mtudisc; 1478 } else if (!PRC_IS_REDIRECT(cmd) && 1479 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) { 1480 return; 1481 } 1482 1483 /* if the parameter is from icmp6, decode it. */ 1484 if (d != NULL) { 1485 ip6cp = (struct ip6ctlparam *)d; 1486 m = ip6cp->ip6c_m; 1487 ip6 = ip6cp->ip6c_ip6; 1488 off = ip6cp->ip6c_off; 1489 sa6_src = ip6cp->ip6c_src; 1490 } else { 1491 m = NULL; 1492 ip6 = NULL; 1493 off = 0; /* fool gcc */ 1494 sa6_src = &sa6_any; 1495 } 1496 1497 if (ip6 != NULL) { 1498 struct in_conninfo inc; 1499 /* 1500 * XXX: We assume that when IPV6 is non NULL, 1501 * M and OFF are valid. 1502 */ 1503 1504 /* check if we can safely examine src and dst ports */ 1505 if (m->m_pkthdr.len < off + sizeof *thp) 1506 return; 1507 1508 bzero(&th, sizeof th); 1509 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1510 1511 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport, 1512 (struct sockaddr *)ip6cp->ip6c_src, 1513 th.th_sport, cmd, arg, notify); 1514 1515 inc.inc_fport = th.th_dport; 1516 inc.inc_lport = th.th_sport; 1517 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1518 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1519 inc.inc_isipv6 = 1; 1520 syncache_unreach(&inc, &th); 1521 } else 1522 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0, 1523 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify); 1524 } 1525 #endif 1526 1527 /* 1528 * Following is where TCP initial sequence number generation occurs. 1529 * 1530 * There are two places where we must use initial sequence numbers: 1531 * 1. In SYN-ACK packets. 1532 * 2. In SYN packets. 1533 * 1534 * All ISNs for SYN-ACK packets are generated by the syncache. See 1535 * tcp_syncache.c for details. 1536 * 1537 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1538 * depends on this property. In addition, these ISNs should be 1539 * unguessable so as to prevent connection hijacking. To satisfy 1540 * the requirements of this situation, the algorithm outlined in 1541 * RFC 1948 is used to generate sequence numbers. 1542 * 1543 * Implementation details: 1544 * 1545 * Time is based off the system timer, and is corrected so that it 1546 * increases by one megabyte per second. This allows for proper 1547 * recycling on high speed LANs while still leaving over an hour 1548 * before rollover. 1549 * 1550 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1551 * between seeding of isn_secret. This is normally set to zero, 1552 * as reseeding should not be necessary. 1553 * 1554 */ 1555 1556 #define ISN_BYTES_PER_SECOND 1048576 1557 1558 u_char isn_secret[32]; 1559 int isn_last_reseed; 1560 MD5_CTX isn_ctx; 1561 1562 tcp_seq 1563 tcp_new_isn(struct tcpcb *tp) 1564 { 1565 u_int32_t md5_buffer[4]; 1566 tcp_seq new_isn; 1567 1568 /* Seed if this is the first use, reseed if requested. */ 1569 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1570 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1571 < (u_int)ticks))) { 1572 read_random_unlimited(&isn_secret, sizeof isn_secret); 1573 isn_last_reseed = ticks; 1574 } 1575 1576 /* Compute the md5 hash and return the ISN. */ 1577 MD5Init(&isn_ctx); 1578 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1579 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1580 #ifdef INET6 1581 if (tp->t_inpcb->inp_vflag & INP_IPV6) { 1582 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1583 sizeof(struct in6_addr)); 1584 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1585 sizeof(struct in6_addr)); 1586 } else 1587 #endif 1588 { 1589 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1590 sizeof(struct in_addr)); 1591 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1592 sizeof(struct in_addr)); 1593 } 1594 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1595 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1596 new_isn = (tcp_seq) md5_buffer[0]; 1597 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1598 return (new_isn); 1599 } 1600 1601 /* 1602 * When a source quench is received, close congestion window 1603 * to one segment. We will gradually open it again as we proceed. 1604 */ 1605 void 1606 tcp_quench(struct inpcb *inp, int error) 1607 { 1608 struct tcpcb *tp = intotcpcb(inp); 1609 1610 if (tp != NULL) { 1611 tp->snd_cwnd = tp->t_maxseg; 1612 tp->snd_wacked = 0; 1613 } 1614 } 1615 1616 /* 1617 * When a specific ICMP unreachable message is received and the 1618 * connection state is SYN-SENT, drop the connection. This behavior 1619 * is controlled by the icmp_may_rst sysctl. 1620 */ 1621 void 1622 tcp_drop_syn_sent(struct inpcb *inp, int error) 1623 { 1624 struct tcpcb *tp = intotcpcb(inp); 1625 1626 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT)) 1627 tcp_drop(tp, error); 1628 } 1629 1630 /* 1631 * When a `need fragmentation' ICMP is received, update our idea of the MSS 1632 * based on the new value in the route. Also nudge TCP to send something, 1633 * since we know the packet we just sent was dropped. 1634 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1635 */ 1636 void 1637 tcp_mtudisc(struct inpcb *inp, int mtu) 1638 { 1639 struct tcpcb *tp = intotcpcb(inp); 1640 struct rtentry *rt; 1641 struct socket *so = inp->inp_socket; 1642 int maxopd, mss; 1643 #ifdef INET6 1644 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0); 1645 #else 1646 const boolean_t isipv6 = FALSE; 1647 #endif 1648 1649 if (tp == NULL) 1650 return; 1651 1652 /* 1653 * If no MTU is provided in the ICMP message, use the 1654 * next lower likely value, as specified in RFC 1191. 1655 */ 1656 if (mtu == 0) { 1657 int oldmtu; 1658 1659 oldmtu = tp->t_maxopd + 1660 (isipv6 ? 1661 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1662 sizeof(struct tcpiphdr)); 1663 mtu = ip_next_mtu(oldmtu, 0); 1664 } 1665 1666 if (isipv6) 1667 rt = tcp_rtlookup6(&inp->inp_inc); 1668 else 1669 rt = tcp_rtlookup(&inp->inp_inc); 1670 if (rt != NULL) { 1671 struct rmxp_tao *taop = rmx_taop(rt->rt_rmx); 1672 1673 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu) 1674 mtu = rt->rt_rmx.rmx_mtu; 1675 1676 maxopd = mtu - 1677 (isipv6 ? 1678 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1679 sizeof(struct tcpiphdr)); 1680 1681 /* 1682 * XXX - The following conditional probably violates the TCP 1683 * spec. The problem is that, since we don't know the 1684 * other end's MSS, we are supposed to use a conservative 1685 * default. But, if we do that, then MTU discovery will 1686 * never actually take place, because the conservative 1687 * default is much less than the MTUs typically seen 1688 * on the Internet today. For the moment, we'll sweep 1689 * this under the carpet. 1690 * 1691 * The conservative default might not actually be a problem 1692 * if the only case this occurs is when sending an initial 1693 * SYN with options and data to a host we've never talked 1694 * to before. Then, they will reply with an MSS value which 1695 * will get recorded and the new parameters should get 1696 * recomputed. For Further Study. 1697 */ 1698 if (taop->tao_mssopt != 0 && taop->tao_mssopt < maxopd) 1699 maxopd = taop->tao_mssopt; 1700 } else 1701 maxopd = mtu - 1702 (isipv6 ? 1703 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1704 sizeof(struct tcpiphdr)); 1705 1706 if (tp->t_maxopd <= maxopd) 1707 return; 1708 tp->t_maxopd = maxopd; 1709 1710 mss = maxopd; 1711 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) == 1712 (TF_REQ_TSTMP | TF_RCVD_TSTMP)) 1713 mss -= TCPOLEN_TSTAMP_APPA; 1714 1715 if ((tp->t_flags & (TF_REQ_CC | TF_RCVD_CC | TF_NOOPT)) == 1716 (TF_REQ_CC | TF_RCVD_CC)) 1717 mss -= TCPOLEN_CC_APPA; 1718 1719 /* round down to multiple of MCLBYTES */ 1720 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */ 1721 if (mss > MCLBYTES) 1722 mss &= ~(MCLBYTES - 1); 1723 #else 1724 if (mss > MCLBYTES) 1725 mss = (mss / MCLBYTES) * MCLBYTES; 1726 #endif 1727 1728 if (so->so_snd.ssb_hiwat < mss) 1729 mss = so->so_snd.ssb_hiwat; 1730 1731 tp->t_maxseg = mss; 1732 tp->t_rtttime = 0; 1733 tp->snd_nxt = tp->snd_una; 1734 tcp_output(tp); 1735 tcpstat.tcps_mturesent++; 1736 } 1737 1738 /* 1739 * Look-up the routing entry to the peer of this inpcb. If no route 1740 * is found and it cannot be allocated the return NULL. This routine 1741 * is called by TCP routines that access the rmx structure and by tcp_mss 1742 * to get the interface MTU. 1743 */ 1744 struct rtentry * 1745 tcp_rtlookup(struct in_conninfo *inc) 1746 { 1747 struct route *ro = &inc->inc_route; 1748 1749 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1750 /* No route yet, so try to acquire one */ 1751 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1752 /* 1753 * unused portions of the structure MUST be zero'd 1754 * out because rtalloc() treats it as opaque data 1755 */ 1756 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1757 ro->ro_dst.sa_family = AF_INET; 1758 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1759 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1760 inc->inc_faddr; 1761 rtalloc(ro); 1762 } 1763 } 1764 return (ro->ro_rt); 1765 } 1766 1767 #ifdef INET6 1768 struct rtentry * 1769 tcp_rtlookup6(struct in_conninfo *inc) 1770 { 1771 struct route_in6 *ro6 = &inc->inc6_route; 1772 1773 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1774 /* No route yet, so try to acquire one */ 1775 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1776 /* 1777 * unused portions of the structure MUST be zero'd 1778 * out because rtalloc() treats it as opaque data 1779 */ 1780 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1781 ro6->ro_dst.sin6_family = AF_INET6; 1782 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1783 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1784 rtalloc((struct route *)ro6); 1785 } 1786 } 1787 return (ro6->ro_rt); 1788 } 1789 #endif 1790 1791 #ifdef IPSEC 1792 /* compute ESP/AH header size for TCP, including outer IP header. */ 1793 size_t 1794 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1795 { 1796 struct inpcb *inp; 1797 struct mbuf *m; 1798 size_t hdrsiz; 1799 struct ip *ip; 1800 struct tcphdr *th; 1801 1802 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1803 return (0); 1804 MGETHDR(m, MB_DONTWAIT, MT_DATA); 1805 if (!m) 1806 return (0); 1807 1808 #ifdef INET6 1809 if (inp->inp_vflag & INP_IPV6) { 1810 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1811 1812 th = (struct tcphdr *)(ip6 + 1); 1813 m->m_pkthdr.len = m->m_len = 1814 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1815 tcp_fillheaders(tp, ip6, th); 1816 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1817 } else 1818 #endif 1819 { 1820 ip = mtod(m, struct ip *); 1821 th = (struct tcphdr *)(ip + 1); 1822 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1823 tcp_fillheaders(tp, ip, th); 1824 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1825 } 1826 1827 m_free(m); 1828 return (hdrsiz); 1829 } 1830 #endif 1831 1832 /* 1833 * Return a pointer to the cached information about the remote host. 1834 * The cached information is stored in the protocol specific part of 1835 * the route metrics. 1836 */ 1837 struct rmxp_tao * 1838 tcp_gettaocache(struct in_conninfo *inc) 1839 { 1840 struct rtentry *rt; 1841 1842 #ifdef INET6 1843 if (inc->inc_isipv6) 1844 rt = tcp_rtlookup6(inc); 1845 else 1846 #endif 1847 rt = tcp_rtlookup(inc); 1848 1849 /* Make sure this is a host route and is up. */ 1850 if (rt == NULL || 1851 (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST)) 1852 return (NULL); 1853 1854 return (rmx_taop(rt->rt_rmx)); 1855 } 1856 1857 /* 1858 * Clear all the TAO cache entries, called from tcp_init. 1859 * 1860 * XXX 1861 * This routine is just an empty one, because we assume that the routing 1862 * routing tables are initialized at the same time when TCP, so there is 1863 * nothing in the cache left over. 1864 */ 1865 static void 1866 tcp_cleartaocache(void) 1867 { 1868 } 1869 1870 /* 1871 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1872 * 1873 * This code attempts to calculate the bandwidth-delay product as a 1874 * means of determining the optimal window size to maximize bandwidth, 1875 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1876 * routers. This code also does a fairly good job keeping RTTs in check 1877 * across slow links like modems. We implement an algorithm which is very 1878 * similar (but not meant to be) TCP/Vegas. The code operates on the 1879 * transmitter side of a TCP connection and so only effects the transmit 1880 * side of the connection. 1881 * 1882 * BACKGROUND: TCP makes no provision for the management of buffer space 1883 * at the end points or at the intermediate routers and switches. A TCP 1884 * stream, whether using NewReno or not, will eventually buffer as 1885 * many packets as it is able and the only reason this typically works is 1886 * due to the fairly small default buffers made available for a connection 1887 * (typicaly 16K or 32K). As machines use larger windows and/or window 1888 * scaling it is now fairly easy for even a single TCP connection to blow-out 1889 * all available buffer space not only on the local interface, but on 1890 * intermediate routers and switches as well. NewReno makes a misguided 1891 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1892 * then backing off, then steadily increasing the window again until another 1893 * failure occurs, ad-infinitum. This results in terrible oscillation that 1894 * is only made worse as network loads increase and the idea of intentionally 1895 * blowing out network buffers is, frankly, a terrible way to manage network 1896 * resources. 1897 * 1898 * It is far better to limit the transmit window prior to the failure 1899 * condition being achieved. There are two general ways to do this: First 1900 * you can 'scan' through different transmit window sizes and locate the 1901 * point where the RTT stops increasing, indicating that you have filled the 1902 * pipe, then scan backwards until you note that RTT stops decreasing, then 1903 * repeat ad-infinitum. This method works in principle but has severe 1904 * implementation issues due to RTT variances, timer granularity, and 1905 * instability in the algorithm which can lead to many false positives and 1906 * create oscillations as well as interact badly with other TCP streams 1907 * implementing the same algorithm. 1908 * 1909 * The second method is to limit the window to the bandwidth delay product 1910 * of the link. This is the method we implement. RTT variances and our 1911 * own manipulation of the congestion window, bwnd, can potentially 1912 * destabilize the algorithm. For this reason we have to stabilize the 1913 * elements used to calculate the window. We do this by using the minimum 1914 * observed RTT, the long term average of the observed bandwidth, and 1915 * by adding two segments worth of slop. It isn't perfect but it is able 1916 * to react to changing conditions and gives us a very stable basis on 1917 * which to extend the algorithm. 1918 */ 1919 void 1920 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1921 { 1922 u_long bw; 1923 u_long bwnd; 1924 int save_ticks; 1925 int delta_ticks; 1926 1927 /* 1928 * If inflight_enable is disabled in the middle of a tcp connection, 1929 * make sure snd_bwnd is effectively disabled. 1930 */ 1931 if (!tcp_inflight_enable) { 1932 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1933 tp->snd_bandwidth = 0; 1934 return; 1935 } 1936 1937 /* 1938 * Validate the delta time. If a connection is new or has been idle 1939 * a long time we have to reset the bandwidth calculator. 1940 */ 1941 save_ticks = ticks; 1942 delta_ticks = save_ticks - tp->t_bw_rtttime; 1943 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1944 tp->t_bw_rtttime = ticks; 1945 tp->t_bw_rtseq = ack_seq; 1946 if (tp->snd_bandwidth == 0) 1947 tp->snd_bandwidth = tcp_inflight_min; 1948 return; 1949 } 1950 if (delta_ticks == 0) 1951 return; 1952 1953 /* 1954 * Sanity check, plus ignore pure window update acks. 1955 */ 1956 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1957 return; 1958 1959 /* 1960 * Figure out the bandwidth. Due to the tick granularity this 1961 * is a very rough number and it MUST be averaged over a fairly 1962 * long period of time. XXX we need to take into account a link 1963 * that is not using all available bandwidth, but for now our 1964 * slop will ramp us up if this case occurs and the bandwidth later 1965 * increases. 1966 */ 1967 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 1968 tp->t_bw_rtttime = save_ticks; 1969 tp->t_bw_rtseq = ack_seq; 1970 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4; 1971 1972 tp->snd_bandwidth = bw; 1973 1974 /* 1975 * Calculate the semi-static bandwidth delay product, plus two maximal 1976 * segments. The additional slop puts us squarely in the sweet 1977 * spot and also handles the bandwidth run-up case. Without the 1978 * slop we could be locking ourselves into a lower bandwidth. 1979 * 1980 * Situations Handled: 1981 * (1) Prevents over-queueing of packets on LANs, especially on 1982 * high speed LANs, allowing larger TCP buffers to be 1983 * specified, and also does a good job preventing 1984 * over-queueing of packets over choke points like modems 1985 * (at least for the transmit side). 1986 * 1987 * (2) Is able to handle changing network loads (bandwidth 1988 * drops so bwnd drops, bandwidth increases so bwnd 1989 * increases). 1990 * 1991 * (3) Theoretically should stabilize in the face of multiple 1992 * connections implementing the same algorithm (this may need 1993 * a little work). 1994 * 1995 * (4) Stability value (defaults to 20 = 2 maximal packets) can 1996 * be adjusted with a sysctl but typically only needs to be on 1997 * very slow connections. A value no smaller then 5 should 1998 * be used, but only reduce this default if you have no other 1999 * choice. 2000 */ 2001 2002 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2) 2003 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 2004 tcp_inflight_stab * (int)tp->t_maxseg / 10; 2005 #undef USERTT 2006 2007 if (tcp_inflight_debug > 0) { 2008 static int ltime; 2009 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) { 2010 ltime = ticks; 2011 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n", 2012 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd); 2013 } 2014 } 2015 if ((long)bwnd < tcp_inflight_min) 2016 bwnd = tcp_inflight_min; 2017 if (bwnd > tcp_inflight_max) 2018 bwnd = tcp_inflight_max; 2019 if ((long)bwnd < tp->t_maxseg * 2) 2020 bwnd = tp->t_maxseg * 2; 2021 tp->snd_bwnd = bwnd; 2022 } 2023