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 1.61 2008/09/23 11:28:49 sephe Exp $ 69 */ 70 71 #include "opt_compat.h" 72 #include "opt_inet6.h" 73 #include "opt_ipsec.h" 74 #include "opt_tcpdebug.h" 75 76 #include <sys/param.h> 77 #include <sys/systm.h> 78 #include <sys/callout.h> 79 #include <sys/kernel.h> 80 #include <sys/sysctl.h> 81 #include <sys/malloc.h> 82 #include <sys/mpipe.h> 83 #include <sys/mbuf.h> 84 #ifdef INET6 85 #include <sys/domain.h> 86 #endif 87 #include <sys/proc.h> 88 #include <sys/socket.h> 89 #include <sys/socketvar.h> 90 #include <sys/protosw.h> 91 #include <sys/random.h> 92 #include <sys/in_cksum.h> 93 #include <sys/ktr.h> 94 95 #include <vm/vm_zone.h> 96 97 #include <net/route.h> 98 #include <net/if.h> 99 #include <net/netisr.h> 100 101 #define _IP_VHL 102 #include <netinet/in.h> 103 #include <netinet/in_systm.h> 104 #include <netinet/ip.h> 105 #include <netinet/ip6.h> 106 #include <netinet/in_pcb.h> 107 #include <netinet6/in6_pcb.h> 108 #include <netinet/in_var.h> 109 #include <netinet/ip_var.h> 110 #include <netinet6/ip6_var.h> 111 #include <netinet/ip_icmp.h> 112 #ifdef INET6 113 #include <netinet/icmp6.h> 114 #endif 115 #include <netinet/tcp.h> 116 #include <netinet/tcp_fsm.h> 117 #include <netinet/tcp_seq.h> 118 #include <netinet/tcp_timer.h> 119 #include <netinet/tcp_var.h> 120 #include <netinet6/tcp6_var.h> 121 #include <netinet/tcpip.h> 122 #ifdef TCPDEBUG 123 #include <netinet/tcp_debug.h> 124 #endif 125 #include <netinet6/ip6protosw.h> 126 127 #ifdef IPSEC 128 #include <netinet6/ipsec.h> 129 #ifdef INET6 130 #include <netinet6/ipsec6.h> 131 #endif 132 #endif 133 134 #ifdef FAST_IPSEC 135 #include <netproto/ipsec/ipsec.h> 136 #ifdef INET6 137 #include <netproto/ipsec/ipsec6.h> 138 #endif 139 #define IPSEC 140 #endif 141 142 #include <sys/md5.h> 143 #include <sys/msgport2.h> 144 #include <machine/smp.h> 145 146 #include <net/netmsg2.h> 147 148 #if !defined(KTR_TCP) 149 #define KTR_TCP KTR_ALL 150 #endif 151 KTR_INFO_MASTER(tcp); 152 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0); 153 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0); 154 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0); 155 #define logtcp(name) KTR_LOG(tcp_ ## name) 156 157 struct inpcbinfo tcbinfo[MAXCPU]; 158 struct tcpcbackqhead tcpcbackq[MAXCPU]; 159 160 int tcp_mpsafe_proto = 0; 161 TUNABLE_INT("net.inet.tcp.mpsafe_proto", &tcp_mpsafe_proto); 162 163 static int tcp_mpsafe_thread = 0; 164 TUNABLE_INT("net.inet.tcp.mpsafe_thread", &tcp_mpsafe_thread); 165 SYSCTL_INT(_net_inet_tcp, OID_AUTO, mpsafe_thread, CTLFLAG_RW, 166 &tcp_mpsafe_thread, 0, 167 "0:BGL, 1:Adaptive BGL, 2:No BGL(experimental)"); 168 169 int tcp_mssdflt = TCP_MSS; 170 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 171 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size"); 172 173 #ifdef INET6 174 int tcp_v6mssdflt = TCP6_MSS; 175 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW, 176 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6"); 177 #endif 178 179 #if 0 180 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ; 181 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW, 182 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time"); 183 #endif 184 185 int tcp_do_rfc1323 = 1; 186 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW, 187 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions"); 188 189 int tcp_do_rfc1644 = 0; 190 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1644, rfc1644, CTLFLAG_RW, 191 &tcp_do_rfc1644, 0, "Enable rfc1644 (TTCP) extensions"); 192 193 static int tcp_tcbhashsize = 0; 194 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD, 195 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable"); 196 197 static int do_tcpdrain = 1; 198 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0, 199 "Enable tcp_drain routine for extra help when low on mbufs"); 200 201 /* XXX JH */ 202 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD, 203 &tcbinfo[0].ipi_count, 0, "Number of active PCBs"); 204 205 static int icmp_may_rst = 1; 206 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0, 207 "Certain ICMP unreachable messages may abort connections in SYN_SENT"); 208 209 static int tcp_isn_reseed_interval = 0; 210 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW, 211 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret"); 212 213 /* 214 * TCP bandwidth limiting sysctls. Note that the default lower bound of 215 * 1024 exists only for debugging. A good production default would be 216 * something like 6100. 217 */ 218 static int tcp_inflight_enable = 0; 219 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW, 220 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting"); 221 222 static int tcp_inflight_debug = 0; 223 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW, 224 &tcp_inflight_debug, 0, "Debug TCP inflight calculations"); 225 226 static int tcp_inflight_min = 6144; 227 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW, 228 &tcp_inflight_min, 0, "Lower bound for TCP inflight window"); 229 230 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT; 231 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW, 232 &tcp_inflight_max, 0, "Upper bound for TCP inflight window"); 233 234 static int tcp_inflight_stab = 20; 235 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW, 236 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 2 packets)"); 237 238 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives"); 239 static struct malloc_pipe tcptemp_mpipe; 240 241 static void tcp_willblock(int); 242 static void tcp_cleartaocache (void); 243 static void tcp_notify (struct inpcb *, int); 244 245 struct tcp_stats tcpstats_percpu[MAXCPU]; 246 #ifdef SMP 247 static int 248 sysctl_tcpstats(SYSCTL_HANDLER_ARGS) 249 { 250 int cpu, error = 0; 251 252 for (cpu = 0; cpu < ncpus; ++cpu) { 253 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu], 254 sizeof(struct tcp_stats)))) 255 break; 256 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu], 257 sizeof(struct tcp_stats)))) 258 break; 259 } 260 261 return (error); 262 } 263 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW), 264 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics"); 265 #else 266 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW, 267 &tcpstat, tcp_stats, "TCP statistics"); 268 #endif 269 270 /* 271 * Target size of TCP PCB hash tables. Must be a power of two. 272 * 273 * Note that this can be overridden by the kernel environment 274 * variable net.inet.tcp.tcbhashsize 275 */ 276 #ifndef TCBHASHSIZE 277 #define TCBHASHSIZE 512 278 #endif 279 280 /* 281 * This is the actual shape of what we allocate using the zone 282 * allocator. Doing it this way allows us to protect both structures 283 * using the same generation count, and also eliminates the overhead 284 * of allocating tcpcbs separately. By hiding the structure here, 285 * we avoid changing most of the rest of the code (although it needs 286 * to be changed, eventually, for greater efficiency). 287 */ 288 #define ALIGNMENT 32 289 #define ALIGNM1 (ALIGNMENT - 1) 290 struct inp_tp { 291 union { 292 struct inpcb inp; 293 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1]; 294 } inp_tp_u; 295 struct tcpcb tcb; 296 struct callout inp_tp_rexmt, inp_tp_persist, inp_tp_keep, inp_tp_2msl; 297 struct callout inp_tp_delack; 298 }; 299 #undef ALIGNMENT 300 #undef ALIGNM1 301 302 /* 303 * Tcp initialization 304 */ 305 void 306 tcp_init(void) 307 { 308 struct inpcbporthead *porthashbase; 309 u_long porthashmask; 310 struct vm_zone *ipi_zone; 311 int hashsize = TCBHASHSIZE; 312 int cpu; 313 314 /* 315 * note: tcptemp is used for keepalives, and it is ok for an 316 * allocation to fail so do not specify MPF_INT. 317 */ 318 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp), 319 25, -1, 0, NULL); 320 321 tcp_ccgen = 1; 322 tcp_cleartaocache(); 323 324 tcp_delacktime = TCPTV_DELACK; 325 tcp_keepinit = TCPTV_KEEP_INIT; 326 tcp_keepidle = TCPTV_KEEP_IDLE; 327 tcp_keepintvl = TCPTV_KEEPINTVL; 328 tcp_maxpersistidle = TCPTV_KEEP_IDLE; 329 tcp_msl = TCPTV_MSL; 330 tcp_rexmit_min = TCPTV_MIN; 331 tcp_rexmit_slop = TCPTV_CPU_VAR; 332 333 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize); 334 if (!powerof2(hashsize)) { 335 kprintf("WARNING: TCB hash size not a power of 2\n"); 336 hashsize = 512; /* safe default */ 337 } 338 tcp_tcbhashsize = hashsize; 339 porthashbase = hashinit(hashsize, M_PCB, &porthashmask); 340 ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets, 341 ZONE_INTERRUPT, 0); 342 343 for (cpu = 0; cpu < ncpus2; cpu++) { 344 in_pcbinfo_init(&tcbinfo[cpu]); 345 tcbinfo[cpu].cpu = cpu; 346 tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB, 347 &tcbinfo[cpu].hashmask); 348 tcbinfo[cpu].porthashbase = porthashbase; 349 tcbinfo[cpu].porthashmask = porthashmask; 350 tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB, 351 &tcbinfo[cpu].wildcardhashmask); 352 tcbinfo[cpu].ipi_zone = ipi_zone; 353 TAILQ_INIT(&tcpcbackq[cpu]); 354 } 355 356 tcp_reass_maxseg = nmbclusters / 16; 357 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg); 358 359 #ifdef INET6 360 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) 361 #else 362 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) 363 #endif 364 if (max_protohdr < TCP_MINPROTOHDR) 365 max_protohdr = TCP_MINPROTOHDR; 366 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) 367 panic("tcp_init"); 368 #undef TCP_MINPROTOHDR 369 370 /* 371 * Initialize TCP statistics counters for each CPU. 372 */ 373 #ifdef SMP 374 for (cpu = 0; cpu < ncpus; ++cpu) { 375 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats)); 376 } 377 #else 378 bzero(&tcpstat, sizeof(struct tcp_stats)); 379 #endif 380 381 syncache_init(); 382 tcp_thread_init(); 383 } 384 385 void 386 tcpmsg_service_loop(void *dummy) 387 { 388 struct netmsg *msg; 389 int mplocked; 390 391 /* 392 * Thread was started with TDF_MPSAFE 393 */ 394 mplocked = 0; 395 396 while ((msg = lwkt_waitport(&curthread->td_msgport, 0))) { 397 do { 398 logtcp(rxmsg); 399 mplocked = netmsg_service(msg, tcp_mpsafe_thread, 400 mplocked); 401 } while ((msg = lwkt_getport(&curthread->td_msgport)) != NULL); 402 403 logtcp(delayed); 404 tcp_willblock(mplocked); 405 logtcp(wait); 406 } 407 } 408 409 static void 410 tcp_willblock(int mplocked) 411 { 412 struct tcpcb *tp; 413 int cpu = mycpu->gd_cpuid; 414 int unlock = 0; 415 416 if (!mplocked && !tcp_mpsafe_proto) { 417 if (TAILQ_EMPTY(&tcpcbackq[cpu])) 418 return; 419 420 get_mplock(); 421 mplocked = 1; 422 unlock = 1; 423 } 424 425 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) { 426 KKASSERT(tp->t_flags & TF_ONOUTPUTQ); 427 tp->t_flags &= ~TF_ONOUTPUTQ; 428 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq); 429 tcp_output(tp); 430 } 431 432 if (unlock) 433 rel_mplock(); 434 } 435 436 437 /* 438 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. 439 * tcp_template used to store this data in mbufs, but we now recopy it out 440 * of the tcpcb each time to conserve mbufs. 441 */ 442 void 443 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr) 444 { 445 struct inpcb *inp = tp->t_inpcb; 446 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr; 447 448 #ifdef INET6 449 if (inp->inp_vflag & INP_IPV6) { 450 struct ip6_hdr *ip6; 451 452 ip6 = (struct ip6_hdr *)ip_ptr; 453 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | 454 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK); 455 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | 456 (IPV6_VERSION & IPV6_VERSION_MASK); 457 ip6->ip6_nxt = IPPROTO_TCP; 458 ip6->ip6_plen = sizeof(struct tcphdr); 459 ip6->ip6_src = inp->in6p_laddr; 460 ip6->ip6_dst = inp->in6p_faddr; 461 tcp_hdr->th_sum = 0; 462 } else 463 #endif 464 { 465 struct ip *ip = (struct ip *) ip_ptr; 466 467 ip->ip_vhl = IP_VHL_BORING; 468 ip->ip_tos = 0; 469 ip->ip_len = 0; 470 ip->ip_id = 0; 471 ip->ip_off = 0; 472 ip->ip_ttl = 0; 473 ip->ip_sum = 0; 474 ip->ip_p = IPPROTO_TCP; 475 ip->ip_src = inp->inp_laddr; 476 ip->ip_dst = inp->inp_faddr; 477 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr, 478 ip->ip_dst.s_addr, 479 htons(sizeof(struct tcphdr) + IPPROTO_TCP)); 480 } 481 482 tcp_hdr->th_sport = inp->inp_lport; 483 tcp_hdr->th_dport = inp->inp_fport; 484 tcp_hdr->th_seq = 0; 485 tcp_hdr->th_ack = 0; 486 tcp_hdr->th_x2 = 0; 487 tcp_hdr->th_off = 5; 488 tcp_hdr->th_flags = 0; 489 tcp_hdr->th_win = 0; 490 tcp_hdr->th_urp = 0; 491 } 492 493 /* 494 * Create template to be used to send tcp packets on a connection. 495 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only 496 * use for this function is in keepalives, which use tcp_respond. 497 */ 498 struct tcptemp * 499 tcp_maketemplate(struct tcpcb *tp) 500 { 501 struct tcptemp *tmp; 502 503 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL) 504 return (NULL); 505 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t); 506 return (tmp); 507 } 508 509 void 510 tcp_freetemplate(struct tcptemp *tmp) 511 { 512 mpipe_free(&tcptemp_mpipe, tmp); 513 } 514 515 /* 516 * Send a single message to the TCP at address specified by 517 * the given TCP/IP header. If m == NULL, then we make a copy 518 * of the tcpiphdr at ti and send directly to the addressed host. 519 * This is used to force keep alive messages out using the TCP 520 * template for a connection. If flags are given then we send 521 * a message back to the TCP which originated the * segment ti, 522 * and discard the mbuf containing it and any other attached mbufs. 523 * 524 * In any case the ack and sequence number of the transmitted 525 * segment are as specified by the parameters. 526 * 527 * NOTE: If m != NULL, then ti must point to *inside* the mbuf. 528 */ 529 void 530 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, 531 tcp_seq ack, tcp_seq seq, int flags) 532 { 533 int tlen; 534 int win = 0; 535 struct route *ro = NULL; 536 struct route sro; 537 struct ip *ip = ipgen; 538 struct tcphdr *nth; 539 int ipflags = 0; 540 struct route_in6 *ro6 = NULL; 541 struct route_in6 sro6; 542 struct ip6_hdr *ip6 = ipgen; 543 #ifdef INET6 544 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6); 545 #else 546 const boolean_t isipv6 = FALSE; 547 #endif 548 549 if (tp != NULL) { 550 if (!(flags & TH_RST)) { 551 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv); 552 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 553 win = (long)TCP_MAXWIN << tp->rcv_scale; 554 } 555 if (isipv6) 556 ro6 = &tp->t_inpcb->in6p_route; 557 else 558 ro = &tp->t_inpcb->inp_route; 559 } else { 560 if (isipv6) { 561 ro6 = &sro6; 562 bzero(ro6, sizeof *ro6); 563 } else { 564 ro = &sro; 565 bzero(ro, sizeof *ro); 566 } 567 } 568 if (m == NULL) { 569 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 570 if (m == NULL) 571 return; 572 tlen = 0; 573 m->m_data += max_linkhdr; 574 if (isipv6) { 575 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); 576 ip6 = mtod(m, struct ip6_hdr *); 577 nth = (struct tcphdr *)(ip6 + 1); 578 } else { 579 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip)); 580 ip = mtod(m, struct ip *); 581 nth = (struct tcphdr *)(ip + 1); 582 } 583 bcopy(th, nth, sizeof(struct tcphdr)); 584 flags = TH_ACK; 585 } else { 586 m_freem(m->m_next); 587 m->m_next = NULL; 588 m->m_data = (caddr_t)ipgen; 589 /* m_len is set later */ 590 tlen = 0; 591 #define xchg(a, b, type) { type t; t = a; a = b; b = t; } 592 if (isipv6) { 593 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); 594 nth = (struct tcphdr *)(ip6 + 1); 595 } else { 596 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long); 597 nth = (struct tcphdr *)(ip + 1); 598 } 599 if (th != nth) { 600 /* 601 * this is usually a case when an extension header 602 * exists between the IPv6 header and the 603 * TCP header. 604 */ 605 nth->th_sport = th->th_sport; 606 nth->th_dport = th->th_dport; 607 } 608 xchg(nth->th_dport, nth->th_sport, n_short); 609 #undef xchg 610 } 611 if (isipv6) { 612 ip6->ip6_flow = 0; 613 ip6->ip6_vfc = IPV6_VERSION; 614 ip6->ip6_nxt = IPPROTO_TCP; 615 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen)); 616 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 617 } else { 618 tlen += sizeof(struct tcpiphdr); 619 ip->ip_len = tlen; 620 ip->ip_ttl = ip_defttl; 621 } 622 m->m_len = tlen; 623 m->m_pkthdr.len = tlen; 624 m->m_pkthdr.rcvif = (struct ifnet *) NULL; 625 nth->th_seq = htonl(seq); 626 nth->th_ack = htonl(ack); 627 nth->th_x2 = 0; 628 nth->th_off = sizeof(struct tcphdr) >> 2; 629 nth->th_flags = flags; 630 if (tp != NULL) 631 nth->th_win = htons((u_short) (win >> tp->rcv_scale)); 632 else 633 nth->th_win = htons((u_short)win); 634 nth->th_urp = 0; 635 if (isipv6) { 636 nth->th_sum = 0; 637 nth->th_sum = in6_cksum(m, IPPROTO_TCP, 638 sizeof(struct ip6_hdr), 639 tlen - sizeof(struct ip6_hdr)); 640 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL, 641 (ro6 && ro6->ro_rt) ? 642 ro6->ro_rt->rt_ifp : NULL); 643 } else { 644 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 645 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); 646 m->m_pkthdr.csum_flags = CSUM_TCP; 647 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 648 } 649 #ifdef TCPDEBUG 650 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) 651 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); 652 #endif 653 if (isipv6) { 654 ip6_output(m, NULL, ro6, ipflags, NULL, NULL, 655 tp ? tp->t_inpcb : NULL); 656 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) { 657 RTFREE(ro6->ro_rt); 658 ro6->ro_rt = NULL; 659 } 660 } else { 661 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL); 662 if ((ro == &sro) && (ro->ro_rt != NULL)) { 663 RTFREE(ro->ro_rt); 664 ro->ro_rt = NULL; 665 } 666 } 667 } 668 669 /* 670 * Create a new TCP control block, making an 671 * empty reassembly queue and hooking it to the argument 672 * protocol control block. The `inp' parameter must have 673 * come from the zone allocator set up in tcp_init(). 674 */ 675 struct tcpcb * 676 tcp_newtcpcb(struct inpcb *inp) 677 { 678 struct inp_tp *it; 679 struct tcpcb *tp; 680 #ifdef INET6 681 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 682 #else 683 const boolean_t isipv6 = FALSE; 684 #endif 685 686 it = (struct inp_tp *)inp; 687 tp = &it->tcb; 688 bzero(tp, sizeof(struct tcpcb)); 689 LIST_INIT(&tp->t_segq); 690 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 691 692 /* Set up our timeouts. */ 693 callout_init(tp->tt_rexmt = &it->inp_tp_rexmt); 694 callout_init(tp->tt_persist = &it->inp_tp_persist); 695 callout_init(tp->tt_keep = &it->inp_tp_keep); 696 callout_init(tp->tt_2msl = &it->inp_tp_2msl); 697 callout_init(tp->tt_delack = &it->inp_tp_delack); 698 699 if (tcp_do_rfc1323) 700 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP); 701 if (tcp_do_rfc1644) 702 tp->t_flags |= TF_REQ_CC; 703 tp->t_inpcb = inp; /* XXX */ 704 tp->t_state = TCPS_CLOSED; 705 /* 706 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 707 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 708 * reasonable initial retransmit time. 709 */ 710 tp->t_srtt = TCPTV_SRTTBASE; 711 tp->t_rttvar = 712 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 713 tp->t_rttmin = tcp_rexmit_min; 714 tp->t_rxtcur = TCPTV_RTOBASE; 715 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 716 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 717 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 718 tp->t_rcvtime = ticks; 719 /* 720 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 721 * because the socket may be bound to an IPv6 wildcard address, 722 * which may match an IPv4-mapped IPv6 address. 723 */ 724 inp->inp_ip_ttl = ip_defttl; 725 inp->inp_ppcb = tp; 726 tcp_sack_tcpcb_init(tp); 727 return (tp); /* XXX */ 728 } 729 730 /* 731 * Drop a TCP connection, reporting the specified error. 732 * If connection is synchronized, then send a RST to peer. 733 */ 734 struct tcpcb * 735 tcp_drop(struct tcpcb *tp, int error) 736 { 737 struct socket *so = tp->t_inpcb->inp_socket; 738 739 if (TCPS_HAVERCVDSYN(tp->t_state)) { 740 tp->t_state = TCPS_CLOSED; 741 tcp_output(tp); 742 tcpstat.tcps_drops++; 743 } else 744 tcpstat.tcps_conndrops++; 745 if (error == ETIMEDOUT && tp->t_softerror) 746 error = tp->t_softerror; 747 so->so_error = error; 748 return (tcp_close(tp)); 749 } 750 751 #ifdef SMP 752 753 struct netmsg_remwildcard { 754 struct netmsg nm_netmsg; 755 struct inpcb *nm_inp; 756 struct inpcbinfo *nm_pcbinfo; 757 #if defined(INET6) 758 int nm_isinet6; 759 #else 760 int nm_unused01; 761 #endif 762 }; 763 764 /* 765 * Wildcard inpcb's on SMP boxes must be removed from all cpus before the 766 * inp can be detached. We do this by cycling through the cpus, ending up 767 * on the cpu controlling the inp last and then doing the disconnect. 768 */ 769 static void 770 in_pcbremwildcardhash_handler(struct netmsg *msg0) 771 { 772 struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0; 773 int cpu; 774 775 cpu = msg->nm_pcbinfo->cpu; 776 777 if (cpu == msg->nm_inp->inp_pcbinfo->cpu) { 778 /* note: detach removes any wildcard hash entry */ 779 #ifdef INET6 780 if (msg->nm_isinet6) 781 in6_pcbdetach(msg->nm_inp); 782 else 783 #endif 784 in_pcbdetach(msg->nm_inp); 785 lwkt_replymsg(&msg->nm_netmsg.nm_lmsg, 0); 786 } else { 787 in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo); 788 cpu = (cpu + 1) % ncpus2; 789 msg->nm_pcbinfo = &tcbinfo[cpu]; 790 lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg); 791 } 792 } 793 794 #endif 795 796 /* 797 * Close a TCP control block: 798 * discard all space held by the tcp 799 * discard internet protocol block 800 * wake up any sleepers 801 */ 802 struct tcpcb * 803 tcp_close(struct tcpcb *tp) 804 { 805 struct tseg_qent *q; 806 struct inpcb *inp = tp->t_inpcb; 807 struct socket *so = inp->inp_socket; 808 struct rtentry *rt; 809 boolean_t dosavessthresh; 810 #ifdef SMP 811 int cpu; 812 #endif 813 #ifdef INET6 814 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 815 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0); 816 #else 817 const boolean_t isipv6 = FALSE; 818 #endif 819 820 /* 821 * The tp is not instantly destroyed in the wildcard case. Setting 822 * the state to TCPS_TERMINATING will prevent the TCP stack from 823 * messing with it, though it should be noted that this change may 824 * not take effect on other cpus until we have chained the wildcard 825 * hash removal. 826 * 827 * XXX we currently depend on the BGL to synchronize the tp->t_state 828 * update and prevent other tcp protocol threads from accepting new 829 * connections on the listen socket we might be trying to close down. 830 */ 831 KKASSERT(tp->t_state != TCPS_TERMINATING); 832 tp->t_state = TCPS_TERMINATING; 833 834 /* 835 * Make sure that all of our timers are stopped before we 836 * delete the PCB. 837 */ 838 callout_stop(tp->tt_rexmt); 839 callout_stop(tp->tt_persist); 840 callout_stop(tp->tt_keep); 841 callout_stop(tp->tt_2msl); 842 callout_stop(tp->tt_delack); 843 844 if (tp->t_flags & TF_ONOUTPUTQ) { 845 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 846 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 847 tp->t_flags &= ~TF_ONOUTPUTQ; 848 } 849 850 /* 851 * If we got enough samples through the srtt filter, 852 * save the rtt and rttvar in the routing entry. 853 * 'Enough' is arbitrarily defined as the 16 samples. 854 * 16 samples is enough for the srtt filter to converge 855 * to within 5% of the correct value; fewer samples and 856 * we could save a very bogus rtt. 857 * 858 * Don't update the default route's characteristics and don't 859 * update anything that the user "locked". 860 */ 861 if (tp->t_rttupdated >= 16) { 862 u_long i = 0; 863 864 if (isipv6) { 865 struct sockaddr_in6 *sin6; 866 867 if ((rt = inp->in6p_route.ro_rt) == NULL) 868 goto no_valid_rt; 869 sin6 = (struct sockaddr_in6 *)rt_key(rt); 870 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 871 goto no_valid_rt; 872 } else 873 if ((rt = inp->inp_route.ro_rt) == NULL || 874 ((struct sockaddr_in *)rt_key(rt))-> 875 sin_addr.s_addr == INADDR_ANY) 876 goto no_valid_rt; 877 878 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 879 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 880 if (rt->rt_rmx.rmx_rtt && i) 881 /* 882 * filter this update to half the old & half 883 * the new values, converting scale. 884 * See route.h and tcp_var.h for a 885 * description of the scaling constants. 886 */ 887 rt->rt_rmx.rmx_rtt = 888 (rt->rt_rmx.rmx_rtt + i) / 2; 889 else 890 rt->rt_rmx.rmx_rtt = i; 891 tcpstat.tcps_cachedrtt++; 892 } 893 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 894 i = tp->t_rttvar * 895 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 896 if (rt->rt_rmx.rmx_rttvar && i) 897 rt->rt_rmx.rmx_rttvar = 898 (rt->rt_rmx.rmx_rttvar + i) / 2; 899 else 900 rt->rt_rmx.rmx_rttvar = i; 901 tcpstat.tcps_cachedrttvar++; 902 } 903 /* 904 * The old comment here said: 905 * update the pipelimit (ssthresh) if it has been updated 906 * already or if a pipesize was specified & the threshhold 907 * got below half the pipesize. I.e., wait for bad news 908 * before we start updating, then update on both good 909 * and bad news. 910 * 911 * But we want to save the ssthresh even if no pipesize is 912 * specified explicitly in the route, because such 913 * connections still have an implicit pipesize specified 914 * by the global tcp_sendspace. In the absence of a reliable 915 * way to calculate the pipesize, it will have to do. 916 */ 917 i = tp->snd_ssthresh; 918 if (rt->rt_rmx.rmx_sendpipe != 0) 919 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 920 else 921 dosavessthresh = (i < so->so_snd.ssb_hiwat/2); 922 if (dosavessthresh || 923 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 924 (rt->rt_rmx.rmx_ssthresh != 0))) { 925 /* 926 * convert the limit from user data bytes to 927 * packets then to packet data bytes. 928 */ 929 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 930 if (i < 2) 931 i = 2; 932 i *= tp->t_maxseg + 933 (isipv6 ? 934 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 935 sizeof(struct tcpiphdr)); 936 if (rt->rt_rmx.rmx_ssthresh) 937 rt->rt_rmx.rmx_ssthresh = 938 (rt->rt_rmx.rmx_ssthresh + i) / 2; 939 else 940 rt->rt_rmx.rmx_ssthresh = i; 941 tcpstat.tcps_cachedssthresh++; 942 } 943 } 944 945 no_valid_rt: 946 /* free the reassembly queue, if any */ 947 while((q = LIST_FIRST(&tp->t_segq)) != NULL) { 948 LIST_REMOVE(q, tqe_q); 949 m_freem(q->tqe_m); 950 FREE(q, M_TSEGQ); 951 tcp_reass_qsize--; 952 } 953 /* throw away SACK blocks in scoreboard*/ 954 if (TCP_DO_SACK(tp)) 955 tcp_sack_cleanup(&tp->scb); 956 957 inp->inp_ppcb = NULL; 958 soisdisconnected(so); 959 /* 960 * Discard the inp. In the SMP case a wildcard inp's hash (created 961 * by a listen socket or an INADDR_ANY udp socket) is replicated 962 * for each protocol thread and must be removed in the context of 963 * that thread. This is accomplished by chaining the message 964 * through the cpus. 965 * 966 * If the inp is not wildcarded we simply detach, which will remove 967 * the any hashes still present for this inp. 968 */ 969 #ifdef SMP 970 if (inp->inp_flags & INP_WILDCARD_MP) { 971 struct netmsg_remwildcard *msg; 972 973 cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2; 974 msg = kmalloc(sizeof(struct netmsg_remwildcard), 975 M_LWKTMSG, M_INTWAIT); 976 netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0, 977 in_pcbremwildcardhash_handler); 978 #ifdef INET6 979 msg->nm_isinet6 = isafinet6; 980 #endif 981 msg->nm_inp = inp; 982 msg->nm_pcbinfo = &tcbinfo[cpu]; 983 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg); 984 } else 985 #endif 986 { 987 /* note: detach removes any wildcard hash entry */ 988 #ifdef INET6 989 if (isafinet6) 990 in6_pcbdetach(inp); 991 else 992 #endif 993 in_pcbdetach(inp); 994 } 995 tcpstat.tcps_closed++; 996 return (NULL); 997 } 998 999 static __inline void 1000 tcp_drain_oncpu(struct inpcbhead *head) 1001 { 1002 struct inpcb *inpb; 1003 struct tcpcb *tcpb; 1004 struct tseg_qent *te; 1005 1006 LIST_FOREACH(inpb, head, inp_list) { 1007 if (inpb->inp_flags & INP_PLACEMARKER) 1008 continue; 1009 if ((tcpb = intotcpcb(inpb))) { 1010 while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) { 1011 LIST_REMOVE(te, tqe_q); 1012 m_freem(te->tqe_m); 1013 FREE(te, M_TSEGQ); 1014 tcp_reass_qsize--; 1015 } 1016 } 1017 } 1018 } 1019 1020 #ifdef SMP 1021 struct netmsg_tcp_drain { 1022 struct netmsg nm_netmsg; 1023 struct inpcbhead *nm_head; 1024 }; 1025 1026 static void 1027 tcp_drain_handler(netmsg_t netmsg) 1028 { 1029 struct netmsg_tcp_drain *nm = (void *)netmsg; 1030 1031 tcp_drain_oncpu(nm->nm_head); 1032 lwkt_replymsg(&nm->nm_netmsg.nm_lmsg, 0); 1033 } 1034 #endif 1035 1036 void 1037 tcp_drain(void) 1038 { 1039 #ifdef SMP 1040 int cpu; 1041 #endif 1042 1043 if (!do_tcpdrain) 1044 return; 1045 1046 /* 1047 * Walk the tcpbs, if existing, and flush the reassembly queue, 1048 * if there is one... 1049 * XXX: The "Net/3" implementation doesn't imply that the TCP 1050 * reassembly queue should be flushed, but in a situation 1051 * where we're really low on mbufs, this is potentially 1052 * useful. 1053 */ 1054 #ifdef SMP 1055 for (cpu = 0; cpu < ncpus2; cpu++) { 1056 struct netmsg_tcp_drain *msg; 1057 1058 if (cpu == mycpu->gd_cpuid) { 1059 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead); 1060 } else { 1061 msg = kmalloc(sizeof(struct netmsg_tcp_drain), 1062 M_LWKTMSG, M_NOWAIT); 1063 if (msg == NULL) 1064 continue; 1065 netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0, 1066 tcp_drain_handler); 1067 msg->nm_head = &tcbinfo[cpu].pcblisthead; 1068 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg); 1069 } 1070 } 1071 #else 1072 tcp_drain_oncpu(&tcbinfo[0].pcblisthead); 1073 #endif 1074 } 1075 1076 /* 1077 * Notify a tcp user of an asynchronous error; 1078 * store error as soft error, but wake up user 1079 * (for now, won't do anything until can select for soft error). 1080 * 1081 * Do not wake up user since there currently is no mechanism for 1082 * reporting soft errors (yet - a kqueue filter may be added). 1083 */ 1084 static void 1085 tcp_notify(struct inpcb *inp, int error) 1086 { 1087 struct tcpcb *tp = intotcpcb(inp); 1088 1089 /* 1090 * Ignore some errors if we are hooked up. 1091 * If connection hasn't completed, has retransmitted several times, 1092 * and receives a second error, give up now. This is better 1093 * than waiting a long time to establish a connection that 1094 * can never complete. 1095 */ 1096 if (tp->t_state == TCPS_ESTABLISHED && 1097 (error == EHOSTUNREACH || error == ENETUNREACH || 1098 error == EHOSTDOWN)) { 1099 return; 1100 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1101 tp->t_softerror) 1102 tcp_drop(tp, error); 1103 else 1104 tp->t_softerror = error; 1105 #if 0 1106 wakeup(&so->so_timeo); 1107 sorwakeup(so); 1108 sowwakeup(so); 1109 #endif 1110 } 1111 1112 static int 1113 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1114 { 1115 int error, i, n; 1116 struct inpcb *marker; 1117 struct inpcb *inp; 1118 inp_gen_t gencnt; 1119 globaldata_t gd; 1120 int origcpu, ccpu; 1121 1122 error = 0; 1123 n = 0; 1124 1125 /* 1126 * The process of preparing the TCB list is too time-consuming and 1127 * resource-intensive to repeat twice on every request. 1128 */ 1129 if (req->oldptr == NULL) { 1130 for (ccpu = 0; ccpu < ncpus; ++ccpu) { 1131 gd = globaldata_find(ccpu); 1132 n += tcbinfo[gd->gd_cpuid].ipi_count; 1133 } 1134 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1135 return (0); 1136 } 1137 1138 if (req->newptr != NULL) 1139 return (EPERM); 1140 1141 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1142 marker->inp_flags |= INP_PLACEMARKER; 1143 1144 /* 1145 * OK, now we're committed to doing something. Run the inpcb list 1146 * for each cpu in the system and construct the output. Use a 1147 * list placemarker to deal with list changes occuring during 1148 * copyout blockages (but otherwise depend on being on the correct 1149 * cpu to avoid races). 1150 */ 1151 origcpu = mycpu->gd_cpuid; 1152 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) { 1153 globaldata_t rgd; 1154 caddr_t inp_ppcb; 1155 struct xtcpcb xt; 1156 int cpu_id; 1157 1158 cpu_id = (origcpu + ccpu) % ncpus; 1159 if ((smp_active_mask & (1 << cpu_id)) == 0) 1160 continue; 1161 rgd = globaldata_find(cpu_id); 1162 lwkt_setcpu_self(rgd); 1163 1164 gencnt = tcbinfo[cpu_id].ipi_gencnt; 1165 n = tcbinfo[cpu_id].ipi_count; 1166 1167 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list); 1168 i = 0; 1169 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1170 /* 1171 * process a snapshot of pcbs, ignoring placemarkers 1172 * and using our own to allow SYSCTL_OUT to block. 1173 */ 1174 LIST_REMOVE(marker, inp_list); 1175 LIST_INSERT_AFTER(inp, marker, inp_list); 1176 1177 if (inp->inp_flags & INP_PLACEMARKER) 1178 continue; 1179 if (inp->inp_gencnt > gencnt) 1180 continue; 1181 if (prison_xinpcb(req->td, inp)) 1182 continue; 1183 1184 xt.xt_len = sizeof xt; 1185 bcopy(inp, &xt.xt_inp, sizeof *inp); 1186 inp_ppcb = inp->inp_ppcb; 1187 if (inp_ppcb != NULL) 1188 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1189 else 1190 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1191 if (inp->inp_socket) 1192 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1193 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1194 break; 1195 ++i; 1196 } 1197 LIST_REMOVE(marker, inp_list); 1198 if (error == 0 && i < n) { 1199 bzero(&xt, sizeof xt); 1200 xt.xt_len = sizeof xt; 1201 while (i < n) { 1202 error = SYSCTL_OUT(req, &xt, sizeof xt); 1203 if (error) 1204 break; 1205 ++i; 1206 } 1207 } 1208 } 1209 1210 /* 1211 * Make sure we are on the same cpu we were on originally, since 1212 * higher level callers expect this. Also don't pollute caches with 1213 * migrated userland data by (eventually) returning to userland 1214 * on a different cpu. 1215 */ 1216 lwkt_setcpu_self(globaldata_find(origcpu)); 1217 kfree(marker, M_TEMP); 1218 return (error); 1219 } 1220 1221 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1222 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1223 1224 static int 1225 tcp_getcred(SYSCTL_HANDLER_ARGS) 1226 { 1227 struct sockaddr_in addrs[2]; 1228 struct inpcb *inp; 1229 int cpu; 1230 int error; 1231 1232 error = suser(req->td); 1233 if (error != 0) 1234 return (error); 1235 error = SYSCTL_IN(req, addrs, sizeof addrs); 1236 if (error != 0) 1237 return (error); 1238 crit_enter(); 1239 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1240 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1241 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1242 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1243 if (inp == NULL || inp->inp_socket == NULL) { 1244 error = ENOENT; 1245 goto out; 1246 } 1247 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1248 out: 1249 crit_exit(); 1250 return (error); 1251 } 1252 1253 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1254 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1255 1256 #ifdef INET6 1257 static int 1258 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1259 { 1260 struct sockaddr_in6 addrs[2]; 1261 struct inpcb *inp; 1262 int error; 1263 boolean_t mapped = FALSE; 1264 1265 error = suser(req->td); 1266 if (error != 0) 1267 return (error); 1268 error = SYSCTL_IN(req, addrs, sizeof addrs); 1269 if (error != 0) 1270 return (error); 1271 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { 1272 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) 1273 mapped = TRUE; 1274 else 1275 return (EINVAL); 1276 } 1277 crit_enter(); 1278 if (mapped) { 1279 inp = in_pcblookup_hash(&tcbinfo[0], 1280 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], 1281 addrs[1].sin6_port, 1282 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], 1283 addrs[0].sin6_port, 1284 0, NULL); 1285 } else { 1286 inp = in6_pcblookup_hash(&tcbinfo[0], 1287 &addrs[1].sin6_addr, addrs[1].sin6_port, 1288 &addrs[0].sin6_addr, addrs[0].sin6_port, 1289 0, NULL); 1290 } 1291 if (inp == NULL || inp->inp_socket == NULL) { 1292 error = ENOENT; 1293 goto out; 1294 } 1295 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1296 out: 1297 crit_exit(); 1298 return (error); 1299 } 1300 1301 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1302 0, 0, 1303 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1304 #endif 1305 1306 void 1307 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip) 1308 { 1309 struct ip *ip = vip; 1310 struct tcphdr *th; 1311 struct in_addr faddr; 1312 struct inpcb *inp; 1313 struct tcpcb *tp; 1314 void (*notify)(struct inpcb *, int) = tcp_notify; 1315 tcp_seq icmpseq; 1316 int arg, cpu; 1317 1318 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) { 1319 return; 1320 } 1321 1322 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1323 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1324 return; 1325 1326 arg = inetctlerrmap[cmd]; 1327 if (cmd == PRC_QUENCH) { 1328 notify = tcp_quench; 1329 } else if (icmp_may_rst && 1330 (cmd == PRC_UNREACH_ADMIN_PROHIB || 1331 cmd == PRC_UNREACH_PORT || 1332 cmd == PRC_TIMXCEED_INTRANS) && 1333 ip != NULL) { 1334 notify = tcp_drop_syn_sent; 1335 } else if (cmd == PRC_MSGSIZE) { 1336 struct icmp *icmp = (struct icmp *) 1337 ((caddr_t)ip - offsetof(struct icmp, icmp_ip)); 1338 1339 arg = ntohs(icmp->icmp_nextmtu); 1340 notify = tcp_mtudisc; 1341 } else if (PRC_IS_REDIRECT(cmd)) { 1342 ip = NULL; 1343 notify = in_rtchange; 1344 } else if (cmd == PRC_HOSTDEAD) { 1345 ip = NULL; 1346 } 1347 1348 if (ip != NULL) { 1349 crit_enter(); 1350 th = (struct tcphdr *)((caddr_t)ip + 1351 (IP_VHL_HL(ip->ip_vhl) << 2)); 1352 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport, 1353 ip->ip_src.s_addr, th->th_sport); 1354 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport, 1355 ip->ip_src, th->th_sport, 0, NULL); 1356 if ((inp != NULL) && (inp->inp_socket != NULL)) { 1357 icmpseq = htonl(th->th_seq); 1358 tp = intotcpcb(inp); 1359 if (SEQ_GEQ(icmpseq, tp->snd_una) && 1360 SEQ_LT(icmpseq, tp->snd_max)) 1361 (*notify)(inp, arg); 1362 } else { 1363 struct in_conninfo inc; 1364 1365 inc.inc_fport = th->th_dport; 1366 inc.inc_lport = th->th_sport; 1367 inc.inc_faddr = faddr; 1368 inc.inc_laddr = ip->ip_src; 1369 #ifdef INET6 1370 inc.inc_isipv6 = 0; 1371 #endif 1372 syncache_unreach(&inc, th); 1373 } 1374 crit_exit(); 1375 } else { 1376 for (cpu = 0; cpu < ncpus2; cpu++) { 1377 in_pcbnotifyall(&tcbinfo[cpu].pcblisthead, faddr, arg, 1378 notify); 1379 } 1380 } 1381 } 1382 1383 #ifdef INET6 1384 void 1385 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d) 1386 { 1387 struct tcphdr th; 1388 void (*notify) (struct inpcb *, int) = tcp_notify; 1389 struct ip6_hdr *ip6; 1390 struct mbuf *m; 1391 struct ip6ctlparam *ip6cp = NULL; 1392 const struct sockaddr_in6 *sa6_src = NULL; 1393 int off; 1394 struct tcp_portonly { 1395 u_int16_t th_sport; 1396 u_int16_t th_dport; 1397 } *thp; 1398 int arg; 1399 1400 if (sa->sa_family != AF_INET6 || 1401 sa->sa_len != sizeof(struct sockaddr_in6)) 1402 return; 1403 1404 arg = 0; 1405 if (cmd == PRC_QUENCH) 1406 notify = tcp_quench; 1407 else if (cmd == PRC_MSGSIZE) { 1408 struct ip6ctlparam *ip6cp = d; 1409 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6; 1410 1411 arg = ntohl(icmp6->icmp6_mtu); 1412 notify = tcp_mtudisc; 1413 } else if (!PRC_IS_REDIRECT(cmd) && 1414 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) { 1415 return; 1416 } 1417 1418 /* if the parameter is from icmp6, decode it. */ 1419 if (d != NULL) { 1420 ip6cp = (struct ip6ctlparam *)d; 1421 m = ip6cp->ip6c_m; 1422 ip6 = ip6cp->ip6c_ip6; 1423 off = ip6cp->ip6c_off; 1424 sa6_src = ip6cp->ip6c_src; 1425 } else { 1426 m = NULL; 1427 ip6 = NULL; 1428 off = 0; /* fool gcc */ 1429 sa6_src = &sa6_any; 1430 } 1431 1432 if (ip6 != NULL) { 1433 struct in_conninfo inc; 1434 /* 1435 * XXX: We assume that when IPV6 is non NULL, 1436 * M and OFF are valid. 1437 */ 1438 1439 /* check if we can safely examine src and dst ports */ 1440 if (m->m_pkthdr.len < off + sizeof *thp) 1441 return; 1442 1443 bzero(&th, sizeof th); 1444 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1445 1446 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport, 1447 (struct sockaddr *)ip6cp->ip6c_src, 1448 th.th_sport, cmd, arg, notify); 1449 1450 inc.inc_fport = th.th_dport; 1451 inc.inc_lport = th.th_sport; 1452 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1453 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1454 inc.inc_isipv6 = 1; 1455 syncache_unreach(&inc, &th); 1456 } else 1457 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0, 1458 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify); 1459 } 1460 #endif 1461 1462 /* 1463 * Following is where TCP initial sequence number generation occurs. 1464 * 1465 * There are two places where we must use initial sequence numbers: 1466 * 1. In SYN-ACK packets. 1467 * 2. In SYN packets. 1468 * 1469 * All ISNs for SYN-ACK packets are generated by the syncache. See 1470 * tcp_syncache.c for details. 1471 * 1472 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1473 * depends on this property. In addition, these ISNs should be 1474 * unguessable so as to prevent connection hijacking. To satisfy 1475 * the requirements of this situation, the algorithm outlined in 1476 * RFC 1948 is used to generate sequence numbers. 1477 * 1478 * Implementation details: 1479 * 1480 * Time is based off the system timer, and is corrected so that it 1481 * increases by one megabyte per second. This allows for proper 1482 * recycling on high speed LANs while still leaving over an hour 1483 * before rollover. 1484 * 1485 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1486 * between seeding of isn_secret. This is normally set to zero, 1487 * as reseeding should not be necessary. 1488 * 1489 */ 1490 1491 #define ISN_BYTES_PER_SECOND 1048576 1492 1493 u_char isn_secret[32]; 1494 int isn_last_reseed; 1495 MD5_CTX isn_ctx; 1496 1497 tcp_seq 1498 tcp_new_isn(struct tcpcb *tp) 1499 { 1500 u_int32_t md5_buffer[4]; 1501 tcp_seq new_isn; 1502 1503 /* Seed if this is the first use, reseed if requested. */ 1504 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1505 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1506 < (u_int)ticks))) { 1507 read_random_unlimited(&isn_secret, sizeof isn_secret); 1508 isn_last_reseed = ticks; 1509 } 1510 1511 /* Compute the md5 hash and return the ISN. */ 1512 MD5Init(&isn_ctx); 1513 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1514 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1515 #ifdef INET6 1516 if (tp->t_inpcb->inp_vflag & INP_IPV6) { 1517 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1518 sizeof(struct in6_addr)); 1519 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1520 sizeof(struct in6_addr)); 1521 } else 1522 #endif 1523 { 1524 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1525 sizeof(struct in_addr)); 1526 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1527 sizeof(struct in_addr)); 1528 } 1529 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1530 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1531 new_isn = (tcp_seq) md5_buffer[0]; 1532 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1533 return (new_isn); 1534 } 1535 1536 /* 1537 * When a source quench is received, close congestion window 1538 * to one segment. We will gradually open it again as we proceed. 1539 */ 1540 void 1541 tcp_quench(struct inpcb *inp, int error) 1542 { 1543 struct tcpcb *tp = intotcpcb(inp); 1544 1545 if (tp != NULL) { 1546 tp->snd_cwnd = tp->t_maxseg; 1547 tp->snd_wacked = 0; 1548 } 1549 } 1550 1551 /* 1552 * When a specific ICMP unreachable message is received and the 1553 * connection state is SYN-SENT, drop the connection. This behavior 1554 * is controlled by the icmp_may_rst sysctl. 1555 */ 1556 void 1557 tcp_drop_syn_sent(struct inpcb *inp, int error) 1558 { 1559 struct tcpcb *tp = intotcpcb(inp); 1560 1561 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT)) 1562 tcp_drop(tp, error); 1563 } 1564 1565 /* 1566 * When a `need fragmentation' ICMP is received, update our idea of the MSS 1567 * based on the new value in the route. Also nudge TCP to send something, 1568 * since we know the packet we just sent was dropped. 1569 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1570 */ 1571 void 1572 tcp_mtudisc(struct inpcb *inp, int mtu) 1573 { 1574 struct tcpcb *tp = intotcpcb(inp); 1575 struct rtentry *rt; 1576 struct socket *so = inp->inp_socket; 1577 int maxopd, mss; 1578 #ifdef INET6 1579 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0); 1580 #else 1581 const boolean_t isipv6 = FALSE; 1582 #endif 1583 1584 if (tp == NULL) 1585 return; 1586 1587 /* 1588 * If no MTU is provided in the ICMP message, use the 1589 * next lower likely value, as specified in RFC 1191. 1590 */ 1591 if (mtu == 0) { 1592 int oldmtu; 1593 1594 oldmtu = tp->t_maxopd + 1595 (isipv6 ? 1596 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1597 sizeof(struct tcpiphdr)); 1598 mtu = ip_next_mtu(oldmtu, 0); 1599 } 1600 1601 if (isipv6) 1602 rt = tcp_rtlookup6(&inp->inp_inc); 1603 else 1604 rt = tcp_rtlookup(&inp->inp_inc); 1605 if (rt != NULL) { 1606 struct rmxp_tao *taop = rmx_taop(rt->rt_rmx); 1607 1608 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu) 1609 mtu = rt->rt_rmx.rmx_mtu; 1610 1611 maxopd = mtu - 1612 (isipv6 ? 1613 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1614 sizeof(struct tcpiphdr)); 1615 1616 /* 1617 * XXX - The following conditional probably violates the TCP 1618 * spec. The problem is that, since we don't know the 1619 * other end's MSS, we are supposed to use a conservative 1620 * default. But, if we do that, then MTU discovery will 1621 * never actually take place, because the conservative 1622 * default is much less than the MTUs typically seen 1623 * on the Internet today. For the moment, we'll sweep 1624 * this under the carpet. 1625 * 1626 * The conservative default might not actually be a problem 1627 * if the only case this occurs is when sending an initial 1628 * SYN with options and data to a host we've never talked 1629 * to before. Then, they will reply with an MSS value which 1630 * will get recorded and the new parameters should get 1631 * recomputed. For Further Study. 1632 */ 1633 if (taop->tao_mssopt != 0 && taop->tao_mssopt < maxopd) 1634 maxopd = taop->tao_mssopt; 1635 } else 1636 maxopd = mtu - 1637 (isipv6 ? 1638 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1639 sizeof(struct tcpiphdr)); 1640 1641 if (tp->t_maxopd <= maxopd) 1642 return; 1643 tp->t_maxopd = maxopd; 1644 1645 mss = maxopd; 1646 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) == 1647 (TF_REQ_TSTMP | TF_RCVD_TSTMP)) 1648 mss -= TCPOLEN_TSTAMP_APPA; 1649 1650 if ((tp->t_flags & (TF_REQ_CC | TF_RCVD_CC | TF_NOOPT)) == 1651 (TF_REQ_CC | TF_RCVD_CC)) 1652 mss -= TCPOLEN_CC_APPA; 1653 1654 /* round down to multiple of MCLBYTES */ 1655 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */ 1656 if (mss > MCLBYTES) 1657 mss &= ~(MCLBYTES - 1); 1658 #else 1659 if (mss > MCLBYTES) 1660 mss = (mss / MCLBYTES) * MCLBYTES; 1661 #endif 1662 1663 if (so->so_snd.ssb_hiwat < mss) 1664 mss = so->so_snd.ssb_hiwat; 1665 1666 tp->t_maxseg = mss; 1667 tp->t_rtttime = 0; 1668 tp->snd_nxt = tp->snd_una; 1669 tcp_output(tp); 1670 tcpstat.tcps_mturesent++; 1671 } 1672 1673 /* 1674 * Look-up the routing entry to the peer of this inpcb. If no route 1675 * is found and it cannot be allocated the return NULL. This routine 1676 * is called by TCP routines that access the rmx structure and by tcp_mss 1677 * to get the interface MTU. 1678 */ 1679 struct rtentry * 1680 tcp_rtlookup(struct in_conninfo *inc) 1681 { 1682 struct route *ro = &inc->inc_route; 1683 1684 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1685 /* No route yet, so try to acquire one */ 1686 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1687 /* 1688 * unused portions of the structure MUST be zero'd 1689 * out because rtalloc() treats it as opaque data 1690 */ 1691 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1692 ro->ro_dst.sa_family = AF_INET; 1693 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1694 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1695 inc->inc_faddr; 1696 rtalloc(ro); 1697 } 1698 } 1699 return (ro->ro_rt); 1700 } 1701 1702 #ifdef INET6 1703 struct rtentry * 1704 tcp_rtlookup6(struct in_conninfo *inc) 1705 { 1706 struct route_in6 *ro6 = &inc->inc6_route; 1707 1708 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1709 /* No route yet, so try to acquire one */ 1710 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1711 /* 1712 * unused portions of the structure MUST be zero'd 1713 * out because rtalloc() treats it as opaque data 1714 */ 1715 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1716 ro6->ro_dst.sin6_family = AF_INET6; 1717 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1718 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1719 rtalloc((struct route *)ro6); 1720 } 1721 } 1722 return (ro6->ro_rt); 1723 } 1724 #endif 1725 1726 #ifdef IPSEC 1727 /* compute ESP/AH header size for TCP, including outer IP header. */ 1728 size_t 1729 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1730 { 1731 struct inpcb *inp; 1732 struct mbuf *m; 1733 size_t hdrsiz; 1734 struct ip *ip; 1735 struct tcphdr *th; 1736 1737 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1738 return (0); 1739 MGETHDR(m, MB_DONTWAIT, MT_DATA); 1740 if (!m) 1741 return (0); 1742 1743 #ifdef INET6 1744 if (inp->inp_vflag & INP_IPV6) { 1745 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1746 1747 th = (struct tcphdr *)(ip6 + 1); 1748 m->m_pkthdr.len = m->m_len = 1749 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1750 tcp_fillheaders(tp, ip6, th); 1751 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1752 } else 1753 #endif 1754 { 1755 ip = mtod(m, struct ip *); 1756 th = (struct tcphdr *)(ip + 1); 1757 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1758 tcp_fillheaders(tp, ip, th); 1759 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1760 } 1761 1762 m_free(m); 1763 return (hdrsiz); 1764 } 1765 #endif 1766 1767 /* 1768 * Return a pointer to the cached information about the remote host. 1769 * The cached information is stored in the protocol specific part of 1770 * the route metrics. 1771 */ 1772 struct rmxp_tao * 1773 tcp_gettaocache(struct in_conninfo *inc) 1774 { 1775 struct rtentry *rt; 1776 1777 #ifdef INET6 1778 if (inc->inc_isipv6) 1779 rt = tcp_rtlookup6(inc); 1780 else 1781 #endif 1782 rt = tcp_rtlookup(inc); 1783 1784 /* Make sure this is a host route and is up. */ 1785 if (rt == NULL || 1786 (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST)) 1787 return (NULL); 1788 1789 return (rmx_taop(rt->rt_rmx)); 1790 } 1791 1792 /* 1793 * Clear all the TAO cache entries, called from tcp_init. 1794 * 1795 * XXX 1796 * This routine is just an empty one, because we assume that the routing 1797 * routing tables are initialized at the same time when TCP, so there is 1798 * nothing in the cache left over. 1799 */ 1800 static void 1801 tcp_cleartaocache(void) 1802 { 1803 } 1804 1805 /* 1806 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1807 * 1808 * This code attempts to calculate the bandwidth-delay product as a 1809 * means of determining the optimal window size to maximize bandwidth, 1810 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1811 * routers. This code also does a fairly good job keeping RTTs in check 1812 * across slow links like modems. We implement an algorithm which is very 1813 * similar (but not meant to be) TCP/Vegas. The code operates on the 1814 * transmitter side of a TCP connection and so only effects the transmit 1815 * side of the connection. 1816 * 1817 * BACKGROUND: TCP makes no provision for the management of buffer space 1818 * at the end points or at the intermediate routers and switches. A TCP 1819 * stream, whether using NewReno or not, will eventually buffer as 1820 * many packets as it is able and the only reason this typically works is 1821 * due to the fairly small default buffers made available for a connection 1822 * (typicaly 16K or 32K). As machines use larger windows and/or window 1823 * scaling it is now fairly easy for even a single TCP connection to blow-out 1824 * all available buffer space not only on the local interface, but on 1825 * intermediate routers and switches as well. NewReno makes a misguided 1826 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1827 * then backing off, then steadily increasing the window again until another 1828 * failure occurs, ad-infinitum. This results in terrible oscillation that 1829 * is only made worse as network loads increase and the idea of intentionally 1830 * blowing out network buffers is, frankly, a terrible way to manage network 1831 * resources. 1832 * 1833 * It is far better to limit the transmit window prior to the failure 1834 * condition being achieved. There are two general ways to do this: First 1835 * you can 'scan' through different transmit window sizes and locate the 1836 * point where the RTT stops increasing, indicating that you have filled the 1837 * pipe, then scan backwards until you note that RTT stops decreasing, then 1838 * repeat ad-infinitum. This method works in principle but has severe 1839 * implementation issues due to RTT variances, timer granularity, and 1840 * instability in the algorithm which can lead to many false positives and 1841 * create oscillations as well as interact badly with other TCP streams 1842 * implementing the same algorithm. 1843 * 1844 * The second method is to limit the window to the bandwidth delay product 1845 * of the link. This is the method we implement. RTT variances and our 1846 * own manipulation of the congestion window, bwnd, can potentially 1847 * destabilize the algorithm. For this reason we have to stabilize the 1848 * elements used to calculate the window. We do this by using the minimum 1849 * observed RTT, the long term average of the observed bandwidth, and 1850 * by adding two segments worth of slop. It isn't perfect but it is able 1851 * to react to changing conditions and gives us a very stable basis on 1852 * which to extend the algorithm. 1853 */ 1854 void 1855 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1856 { 1857 u_long bw; 1858 u_long bwnd; 1859 int save_ticks; 1860 int delta_ticks; 1861 1862 /* 1863 * If inflight_enable is disabled in the middle of a tcp connection, 1864 * make sure snd_bwnd is effectively disabled. 1865 */ 1866 if (!tcp_inflight_enable) { 1867 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1868 tp->snd_bandwidth = 0; 1869 return; 1870 } 1871 1872 /* 1873 * Validate the delta time. If a connection is new or has been idle 1874 * a long time we have to reset the bandwidth calculator. 1875 */ 1876 save_ticks = ticks; 1877 delta_ticks = save_ticks - tp->t_bw_rtttime; 1878 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1879 tp->t_bw_rtttime = ticks; 1880 tp->t_bw_rtseq = ack_seq; 1881 if (tp->snd_bandwidth == 0) 1882 tp->snd_bandwidth = tcp_inflight_min; 1883 return; 1884 } 1885 if (delta_ticks == 0) 1886 return; 1887 1888 /* 1889 * Sanity check, plus ignore pure window update acks. 1890 */ 1891 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1892 return; 1893 1894 /* 1895 * Figure out the bandwidth. Due to the tick granularity this 1896 * is a very rough number and it MUST be averaged over a fairly 1897 * long period of time. XXX we need to take into account a link 1898 * that is not using all available bandwidth, but for now our 1899 * slop will ramp us up if this case occurs and the bandwidth later 1900 * increases. 1901 */ 1902 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 1903 tp->t_bw_rtttime = save_ticks; 1904 tp->t_bw_rtseq = ack_seq; 1905 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4; 1906 1907 tp->snd_bandwidth = bw; 1908 1909 /* 1910 * Calculate the semi-static bandwidth delay product, plus two maximal 1911 * segments. The additional slop puts us squarely in the sweet 1912 * spot and also handles the bandwidth run-up case. Without the 1913 * slop we could be locking ourselves into a lower bandwidth. 1914 * 1915 * Situations Handled: 1916 * (1) Prevents over-queueing of packets on LANs, especially on 1917 * high speed LANs, allowing larger TCP buffers to be 1918 * specified, and also does a good job preventing 1919 * over-queueing of packets over choke points like modems 1920 * (at least for the transmit side). 1921 * 1922 * (2) Is able to handle changing network loads (bandwidth 1923 * drops so bwnd drops, bandwidth increases so bwnd 1924 * increases). 1925 * 1926 * (3) Theoretically should stabilize in the face of multiple 1927 * connections implementing the same algorithm (this may need 1928 * a little work). 1929 * 1930 * (4) Stability value (defaults to 20 = 2 maximal packets) can 1931 * be adjusted with a sysctl but typically only needs to be on 1932 * very slow connections. A value no smaller then 5 should 1933 * be used, but only reduce this default if you have no other 1934 * choice. 1935 */ 1936 1937 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2) 1938 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 1939 tcp_inflight_stab * (int)tp->t_maxseg / 10; 1940 #undef USERTT 1941 1942 if (tcp_inflight_debug > 0) { 1943 static int ltime; 1944 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) { 1945 ltime = ticks; 1946 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n", 1947 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd); 1948 } 1949 } 1950 if ((long)bwnd < tcp_inflight_min) 1951 bwnd = tcp_inflight_min; 1952 if (bwnd > tcp_inflight_max) 1953 bwnd = tcp_inflight_max; 1954 if ((long)bwnd < tp->t_maxseg * 2) 1955 bwnd = tp->t_maxseg * 2; 1956 tp->snd_bwnd = bwnd; 1957 } 1958