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) 2003, 2004 Jeffrey M. Hsu. All rights reserved. 36 * 37 * License terms: all terms for the DragonFly license above plus the following: 38 * 39 * 4. All advertising materials mentioning features or use of this software 40 * must display the following acknowledgement: 41 * 42 * This product includes software developed by Jeffrey M. Hsu 43 * for the DragonFly Project. 44 * 45 * This requirement may be waived with permission from Jeffrey Hsu. 46 * This requirement will sunset and may be removed on July 8 2005, 47 * after which the standard DragonFly license (as shown above) will 48 * apply. 49 */ 50 51 /* 52 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 53 * The Regents of the University of California. All rights reserved. 54 * 55 * Redistribution and use in source and binary forms, with or without 56 * modification, are permitted provided that the following conditions 57 * are met: 58 * 1. Redistributions of source code must retain the above copyright 59 * notice, this list of conditions and the following disclaimer. 60 * 2. Redistributions in binary form must reproduce the above copyright 61 * notice, this list of conditions and the following disclaimer in the 62 * documentation and/or other materials provided with the distribution. 63 * 3. All advertising materials mentioning features or use of this software 64 * must display the following acknowledgement: 65 * This product includes software developed by the University of 66 * California, Berkeley and its contributors. 67 * 4. Neither the name of the University nor the names of its contributors 68 * may be used to endorse or promote products derived from this software 69 * without specific prior written permission. 70 * 71 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 72 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 73 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 74 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 75 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 76 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 77 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 78 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 79 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 80 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 81 * SUCH DAMAGE. 82 * 83 * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95 84 * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $ 85 * $DragonFly: src/sys/netinet/tcp_subr.c,v 1.42 2004/12/20 11:03:16 joerg Exp $ 86 */ 87 88 #include "opt_compat.h" 89 #include "opt_inet6.h" 90 #include "opt_ipsec.h" 91 #include "opt_tcpdebug.h" 92 93 #include <sys/param.h> 94 #include <sys/systm.h> 95 #include <sys/callout.h> 96 #include <sys/kernel.h> 97 #include <sys/sysctl.h> 98 #include <sys/malloc.h> 99 #include <sys/mpipe.h> 100 #include <sys/mbuf.h> 101 #ifdef INET6 102 #include <sys/domain.h> 103 #endif 104 #include <sys/proc.h> 105 #include <sys/socket.h> 106 #include <sys/socketvar.h> 107 #include <sys/protosw.h> 108 #include <sys/random.h> 109 #include <sys/in_cksum.h> 110 111 #include <vm/vm_zone.h> 112 113 #include <net/route.h> 114 #include <net/if.h> 115 #include <net/netisr.h> 116 117 #define _IP_VHL 118 #include <netinet/in.h> 119 #include <netinet/in_systm.h> 120 #include <netinet/ip.h> 121 #include <netinet/ip6.h> 122 #include <netinet/in_pcb.h> 123 #include <netinet6/in6_pcb.h> 124 #include <netinet/in_var.h> 125 #include <netinet/ip_var.h> 126 #include <netinet6/ip6_var.h> 127 #include <netinet/tcp.h> 128 #include <netinet/tcp_fsm.h> 129 #include <netinet/tcp_seq.h> 130 #include <netinet/tcp_timer.h> 131 #include <netinet/tcp_var.h> 132 #include <netinet6/tcp6_var.h> 133 #include <netinet/tcpip.h> 134 #ifdef TCPDEBUG 135 #include <netinet/tcp_debug.h> 136 #endif 137 #include <netinet6/ip6protosw.h> 138 139 #ifdef IPSEC 140 #include <netinet6/ipsec.h> 141 #ifdef INET6 142 #include <netinet6/ipsec6.h> 143 #endif 144 #endif 145 146 #ifdef FAST_IPSEC 147 #include <netproto/ipsec/ipsec.h> 148 #ifdef INET6 149 #include <netproto/ipsec/ipsec6.h> 150 #endif 151 #define IPSEC 152 #endif 153 154 #include <sys/md5.h> 155 156 #include <sys/msgport2.h> 157 158 #include <machine/smp.h> 159 160 struct inpcbinfo tcbinfo[MAXCPU]; 161 struct tcpcbackqhead tcpcbackq[MAXCPU]; 162 163 int tcp_mssdflt = TCP_MSS; 164 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 165 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size"); 166 167 #ifdef INET6 168 int tcp_v6mssdflt = TCP6_MSS; 169 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW, 170 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6"); 171 #endif 172 173 #if 0 174 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ; 175 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW, 176 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time"); 177 #endif 178 179 int tcp_do_rfc1323 = 1; 180 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW, 181 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions"); 182 183 int tcp_do_rfc1644 = 0; 184 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1644, rfc1644, CTLFLAG_RW, 185 &tcp_do_rfc1644, 0, "Enable rfc1644 (TTCP) extensions"); 186 187 static int tcp_tcbhashsize = 0; 188 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD, 189 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable"); 190 191 static int do_tcpdrain = 1; 192 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0, 193 "Enable tcp_drain routine for extra help when low on mbufs"); 194 195 /* XXX JH */ 196 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD, 197 &tcbinfo[0].ipi_count, 0, "Number of active PCBs"); 198 199 static int icmp_may_rst = 1; 200 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0, 201 "Certain ICMP unreachable messages may abort connections in SYN_SENT"); 202 203 static int tcp_isn_reseed_interval = 0; 204 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW, 205 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret"); 206 207 /* 208 * TCP bandwidth limiting sysctls. Note that the default lower bound of 209 * 1024 exists only for debugging. A good production default would be 210 * something like 6100. 211 */ 212 static int tcp_inflight_enable = 0; 213 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW, 214 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting"); 215 216 static int tcp_inflight_debug = 0; 217 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW, 218 &tcp_inflight_debug, 0, "Debug TCP inflight calculations"); 219 220 static int tcp_inflight_min = 6144; 221 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW, 222 &tcp_inflight_min, 0, "Lower bound for TCP inflight window"); 223 224 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT; 225 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW, 226 &tcp_inflight_max, 0, "Upper bound for TCP inflight window"); 227 228 static int tcp_inflight_stab = 20; 229 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW, 230 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 2 packets)"); 231 232 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives"); 233 static struct malloc_pipe tcptemp_mpipe; 234 235 static void tcp_willblock(void); 236 static void tcp_cleartaocache (void); 237 static void tcp_notify (struct inpcb *, int); 238 239 struct tcp_stats tcpstats_ary[MAXCPU]; 240 #ifdef SMP 241 static int 242 sysctl_tcpstats(SYSCTL_HANDLER_ARGS) 243 { 244 int cpu, error = 0; 245 246 for (cpu = 0; cpu < ncpus; ++cpu) { 247 if ((error = SYSCTL_OUT(req, (void *)&tcpstats_ary[cpu], 248 sizeof(struct tcp_stats)))) 249 break; 250 if ((error = SYSCTL_IN(req, (void *)&tcpstats_ary[cpu], 251 sizeof(struct tcp_stats)))) 252 break; 253 } 254 255 return (error); 256 } 257 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW), 258 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics"); 259 #else 260 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW, 261 &tcpstat, tcp_stats, "TCP statistics"); 262 #endif 263 264 /* 265 * Target size of TCP PCB hash tables. Must be a power of two. 266 * 267 * Note that this can be overridden by the kernel environment 268 * variable net.inet.tcp.tcbhashsize 269 */ 270 #ifndef TCBHASHSIZE 271 #define TCBHASHSIZE 512 272 #endif 273 274 /* 275 * This is the actual shape of what we allocate using the zone 276 * allocator. Doing it this way allows us to protect both structures 277 * using the same generation count, and also eliminates the overhead 278 * of allocating tcpcbs separately. By hiding the structure here, 279 * we avoid changing most of the rest of the code (although it needs 280 * to be changed, eventually, for greater efficiency). 281 */ 282 #define ALIGNMENT 32 283 #define ALIGNM1 (ALIGNMENT - 1) 284 struct inp_tp { 285 union { 286 struct inpcb inp; 287 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1]; 288 } inp_tp_u; 289 struct tcpcb tcb; 290 struct callout inp_tp_rexmt, inp_tp_persist, inp_tp_keep, inp_tp_2msl; 291 struct callout inp_tp_delack; 292 }; 293 #undef ALIGNMENT 294 #undef ALIGNM1 295 296 /* 297 * Tcp initialization 298 */ 299 void 300 tcp_init() 301 { 302 struct inpcbporthead *porthashbase; 303 u_long porthashmask; 304 struct vm_zone *ipi_zone; 305 int hashsize = TCBHASHSIZE; 306 int cpu; 307 308 /* 309 * note: tcptemp is used for keepalives, and it is ok for an 310 * allocation to fail so do not specify MPF_INT. 311 */ 312 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp), 313 25, -1, 0, NULL); 314 315 tcp_ccgen = 1; 316 tcp_cleartaocache(); 317 318 tcp_delacktime = TCPTV_DELACK; 319 tcp_keepinit = TCPTV_KEEP_INIT; 320 tcp_keepidle = TCPTV_KEEP_IDLE; 321 tcp_keepintvl = TCPTV_KEEPINTVL; 322 tcp_maxpersistidle = TCPTV_KEEP_IDLE; 323 tcp_msl = TCPTV_MSL; 324 tcp_rexmit_min = TCPTV_MIN; 325 tcp_rexmit_slop = TCPTV_CPU_VAR; 326 327 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize); 328 if (!powerof2(hashsize)) { 329 printf("WARNING: TCB hash size not a power of 2\n"); 330 hashsize = 512; /* safe default */ 331 } 332 tcp_tcbhashsize = hashsize; 333 porthashbase = hashinit(hashsize, M_PCB, &porthashmask); 334 ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets, 335 ZONE_INTERRUPT, 0); 336 337 for (cpu = 0; cpu < ncpus2; cpu++) { 338 in_pcbinfo_init(&tcbinfo[cpu]); 339 tcbinfo[cpu].cpu = cpu; 340 tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB, 341 &tcbinfo[cpu].hashmask); 342 tcbinfo[cpu].porthashbase = porthashbase; 343 tcbinfo[cpu].porthashmask = porthashmask; 344 tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB, 345 &tcbinfo[cpu].wildcardhashmask); 346 tcbinfo[cpu].ipi_zone = ipi_zone; 347 TAILQ_INIT(&tcpcbackq[cpu]); 348 } 349 350 tcp_reass_maxseg = nmbclusters / 16; 351 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg); 352 353 #ifdef INET6 354 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) 355 #else 356 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) 357 #endif 358 if (max_protohdr < TCP_MINPROTOHDR) 359 max_protohdr = TCP_MINPROTOHDR; 360 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) 361 panic("tcp_init"); 362 #undef TCP_MINPROTOHDR 363 364 /* 365 * Initialize TCP statistics. 366 * 367 * It is layed out as an array which is has one element for UP, 368 * and SMP_MAXCPU elements for SMP. This allows us to retain 369 * the access mechanism from userland for both UP and SMP. 370 */ 371 #ifdef SMP 372 for (cpu = 0; cpu < ncpus; ++cpu) { 373 bzero(&tcpstats_ary[cpu], sizeof(struct tcp_stats)); 374 } 375 #else 376 bzero(&tcpstat, sizeof(struct tcp_stats)); 377 #endif 378 379 syncache_init(); 380 tcp_sack_init(); 381 tcp_thread_init(); 382 } 383 384 void 385 tcpmsg_service_loop(void *dummy) 386 { 387 struct netmsg *msg; 388 389 while ((msg = lwkt_waitport(&curthread->td_msgport, NULL))) { 390 do { 391 msg->nm_lmsg.ms_cmd.cm_func(&msg->nm_lmsg); 392 } while ((msg = lwkt_getport(&curthread->td_msgport)) != NULL); 393 tcp_willblock(); 394 } 395 } 396 397 static void 398 tcp_willblock(void) 399 { 400 struct tcpcb *tp; 401 int cpu = mycpu->gd_cpuid; 402 403 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) { 404 KKASSERT(tp->t_flags & TF_ONOUTPUTQ); 405 tp->t_flags &= ~TF_ONOUTPUTQ; 406 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq); 407 tcp_output(tp); 408 } 409 } 410 411 412 /* 413 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. 414 * tcp_template used to store this data in mbufs, but we now recopy it out 415 * of the tcpcb each time to conserve mbufs. 416 */ 417 void 418 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr) 419 { 420 struct inpcb *inp = tp->t_inpcb; 421 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr; 422 423 #ifdef INET6 424 if (inp->inp_vflag & INP_IPV6) { 425 struct ip6_hdr *ip6; 426 427 ip6 = (struct ip6_hdr *)ip_ptr; 428 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | 429 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK); 430 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | 431 (IPV6_VERSION & IPV6_VERSION_MASK); 432 ip6->ip6_nxt = IPPROTO_TCP; 433 ip6->ip6_plen = sizeof(struct tcphdr); 434 ip6->ip6_src = inp->in6p_laddr; 435 ip6->ip6_dst = inp->in6p_faddr; 436 tcp_hdr->th_sum = 0; 437 } else 438 #endif 439 { 440 struct ip *ip = (struct ip *) ip_ptr; 441 442 ip->ip_vhl = IP_VHL_BORING; 443 ip->ip_tos = 0; 444 ip->ip_len = 0; 445 ip->ip_id = 0; 446 ip->ip_off = 0; 447 ip->ip_ttl = 0; 448 ip->ip_sum = 0; 449 ip->ip_p = IPPROTO_TCP; 450 ip->ip_src = inp->inp_laddr; 451 ip->ip_dst = inp->inp_faddr; 452 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr, 453 ip->ip_dst.s_addr, 454 htons(sizeof(struct tcphdr) + IPPROTO_TCP)); 455 } 456 457 tcp_hdr->th_sport = inp->inp_lport; 458 tcp_hdr->th_dport = inp->inp_fport; 459 tcp_hdr->th_seq = 0; 460 tcp_hdr->th_ack = 0; 461 tcp_hdr->th_x2 = 0; 462 tcp_hdr->th_off = 5; 463 tcp_hdr->th_flags = 0; 464 tcp_hdr->th_win = 0; 465 tcp_hdr->th_urp = 0; 466 } 467 468 /* 469 * Create template to be used to send tcp packets on a connection. 470 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only 471 * use for this function is in keepalives, which use tcp_respond. 472 */ 473 struct tcptemp * 474 tcp_maketemplate(struct tcpcb *tp) 475 { 476 struct tcptemp *tmp; 477 478 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL) 479 return (NULL); 480 tcp_fillheaders(tp, (void *)&tmp->tt_ipgen, (void *)&tmp->tt_t); 481 return (tmp); 482 } 483 484 void 485 tcp_freetemplate(struct tcptemp *tmp) 486 { 487 mpipe_free(&tcptemp_mpipe, tmp); 488 } 489 490 /* 491 * Send a single message to the TCP at address specified by 492 * the given TCP/IP header. If m == NULL, then we make a copy 493 * of the tcpiphdr at ti and send directly to the addressed host. 494 * This is used to force keep alive messages out using the TCP 495 * template for a connection. If flags are given then we send 496 * a message back to the TCP which originated the * segment ti, 497 * and discard the mbuf containing it and any other attached mbufs. 498 * 499 * In any case the ack and sequence number of the transmitted 500 * segment are as specified by the parameters. 501 * 502 * NOTE: If m != NULL, then ti must point to *inside* the mbuf. 503 */ 504 void 505 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, 506 tcp_seq ack, tcp_seq seq, int flags) 507 { 508 int tlen; 509 int win = 0; 510 struct route *ro = NULL; 511 struct route sro; 512 struct ip *ip = ipgen; 513 struct tcphdr *nth; 514 int ipflags = 0; 515 struct route_in6 *ro6 = NULL; 516 struct route_in6 sro6; 517 struct ip6_hdr *ip6 = ipgen; 518 #ifdef INET6 519 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6); 520 #else 521 const boolean_t isipv6 = FALSE; 522 #endif 523 524 if (tp != NULL) { 525 if (!(flags & TH_RST)) { 526 win = sbspace(&tp->t_inpcb->inp_socket->so_rcv); 527 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 528 win = (long)TCP_MAXWIN << tp->rcv_scale; 529 } 530 if (isipv6) 531 ro6 = &tp->t_inpcb->in6p_route; 532 else 533 ro = &tp->t_inpcb->inp_route; 534 } else { 535 if (isipv6) { 536 ro6 = &sro6; 537 bzero(ro6, sizeof *ro6); 538 } else { 539 ro = &sro; 540 bzero(ro, sizeof *ro); 541 } 542 } 543 if (m == NULL) { 544 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 545 if (m == NULL) 546 return; 547 tlen = 0; 548 m->m_data += max_linkhdr; 549 if (isipv6) { 550 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); 551 ip6 = mtod(m, struct ip6_hdr *); 552 nth = (struct tcphdr *)(ip6 + 1); 553 } else { 554 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip)); 555 ip = mtod(m, struct ip *); 556 nth = (struct tcphdr *)(ip + 1); 557 } 558 bcopy(th, nth, sizeof(struct tcphdr)); 559 flags = TH_ACK; 560 } else { 561 m_freem(m->m_next); 562 m->m_next = NULL; 563 m->m_data = (caddr_t)ipgen; 564 /* m_len is set later */ 565 tlen = 0; 566 #define xchg(a, b, type) { type t; t = a; a = b; b = t; } 567 if (isipv6) { 568 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); 569 nth = (struct tcphdr *)(ip6 + 1); 570 } else { 571 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long); 572 nth = (struct tcphdr *)(ip + 1); 573 } 574 if (th != nth) { 575 /* 576 * this is usually a case when an extension header 577 * exists between the IPv6 header and the 578 * TCP header. 579 */ 580 nth->th_sport = th->th_sport; 581 nth->th_dport = th->th_dport; 582 } 583 xchg(nth->th_dport, nth->th_sport, n_short); 584 #undef xchg 585 } 586 if (isipv6) { 587 ip6->ip6_flow = 0; 588 ip6->ip6_vfc = IPV6_VERSION; 589 ip6->ip6_nxt = IPPROTO_TCP; 590 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen)); 591 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 592 } else { 593 tlen += sizeof(struct tcpiphdr); 594 ip->ip_len = tlen; 595 ip->ip_ttl = ip_defttl; 596 } 597 m->m_len = tlen; 598 m->m_pkthdr.len = tlen; 599 m->m_pkthdr.rcvif = (struct ifnet *) NULL; 600 nth->th_seq = htonl(seq); 601 nth->th_ack = htonl(ack); 602 nth->th_x2 = 0; 603 nth->th_off = sizeof(struct tcphdr) >> 2; 604 nth->th_flags = flags; 605 if (tp != NULL) 606 nth->th_win = htons((u_short) (win >> tp->rcv_scale)); 607 else 608 nth->th_win = htons((u_short)win); 609 nth->th_urp = 0; 610 if (isipv6) { 611 nth->th_sum = 0; 612 nth->th_sum = in6_cksum(m, IPPROTO_TCP, 613 sizeof(struct ip6_hdr), 614 tlen - sizeof(struct ip6_hdr)); 615 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL, 616 (ro6 && ro6->ro_rt) ? 617 ro6->ro_rt->rt_ifp : NULL); 618 } else { 619 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 620 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); 621 m->m_pkthdr.csum_flags = CSUM_TCP; 622 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 623 } 624 #ifdef TCPDEBUG 625 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) 626 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); 627 #endif 628 if (isipv6) { 629 (void)ip6_output(m, NULL, ro6, ipflags, NULL, NULL, 630 tp ? tp->t_inpcb : NULL); 631 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) { 632 RTFREE(ro6->ro_rt); 633 ro6->ro_rt = NULL; 634 } 635 } else { 636 (void)ip_output(m, NULL, ro, ipflags, NULL, 637 tp ? tp->t_inpcb : NULL); 638 if ((ro == &sro) && (ro->ro_rt != NULL)) { 639 RTFREE(ro->ro_rt); 640 ro->ro_rt = NULL; 641 } 642 } 643 } 644 645 /* 646 * Create a new TCP control block, making an 647 * empty reassembly queue and hooking it to the argument 648 * protocol control block. The `inp' parameter must have 649 * come from the zone allocator set up in tcp_init(). 650 */ 651 struct tcpcb * 652 tcp_newtcpcb(struct inpcb *inp) 653 { 654 struct inp_tp *it; 655 struct tcpcb *tp; 656 #ifdef INET6 657 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 658 #else 659 const boolean_t isipv6 = FALSE; 660 #endif 661 662 it = (struct inp_tp *)inp; 663 tp = &it->tcb; 664 bzero(tp, sizeof(struct tcpcb)); 665 LIST_INIT(&tp->t_segq); 666 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 667 668 /* Set up our timeouts. */ 669 callout_init(tp->tt_rexmt = &it->inp_tp_rexmt); 670 callout_init(tp->tt_persist = &it->inp_tp_persist); 671 callout_init(tp->tt_keep = &it->inp_tp_keep); 672 callout_init(tp->tt_2msl = &it->inp_tp_2msl); 673 callout_init(tp->tt_delack = &it->inp_tp_delack); 674 675 if (tcp_do_rfc1323) 676 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP); 677 if (tcp_do_rfc1644) 678 tp->t_flags |= TF_REQ_CC; 679 tp->t_inpcb = inp; /* XXX */ 680 tp->t_state = TCPS_CLOSED; 681 /* 682 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 683 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 684 * reasonable initial retransmit time. 685 */ 686 tp->t_srtt = TCPTV_SRTTBASE; 687 tp->t_rttvar = 688 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 689 tp->t_rttmin = tcp_rexmit_min; 690 tp->t_rxtcur = TCPTV_RTOBASE; 691 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 692 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 693 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 694 tp->t_rcvtime = ticks; 695 /* 696 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 697 * because the socket may be bound to an IPv6 wildcard address, 698 * which may match an IPv4-mapped IPv6 address. 699 */ 700 inp->inp_ip_ttl = ip_defttl; 701 inp->inp_ppcb = (caddr_t)tp; 702 tcp_sack_tcpcb_init(tp); 703 return (tp); /* XXX */ 704 } 705 706 /* 707 * Drop a TCP connection, reporting the specified error. 708 * If connection is synchronized, then send a RST to peer. 709 */ 710 struct tcpcb * 711 tcp_drop(struct tcpcb *tp, int errno) 712 { 713 struct socket *so = tp->t_inpcb->inp_socket; 714 715 if (TCPS_HAVERCVDSYN(tp->t_state)) { 716 tp->t_state = TCPS_CLOSED; 717 (void) tcp_output(tp); 718 tcpstat.tcps_drops++; 719 } else 720 tcpstat.tcps_conndrops++; 721 if (errno == ETIMEDOUT && tp->t_softerror) 722 errno = tp->t_softerror; 723 so->so_error = errno; 724 return (tcp_close(tp)); 725 } 726 727 #ifdef SMP 728 729 struct netmsg_remwildcard { 730 struct lwkt_msg nm_lmsg; 731 struct inpcb *nm_inp; 732 struct inpcbinfo *nm_pcbinfo; 733 #if defined(INET6) 734 int nm_isinet6; 735 #else 736 int nm_unused01; 737 #endif 738 }; 739 740 /* 741 * Wildcard inpcb's on SMP boxes must be removed from all cpus before the 742 * inp can be detached. We do this by cycling through the cpus, ending up 743 * on the cpu controlling the inp last and then doing the disconnect. 744 */ 745 static int 746 in_pcbremwildcardhash_handler(struct lwkt_msg *msg0) 747 { 748 struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0; 749 int cpu; 750 751 cpu = msg->nm_pcbinfo->cpu; 752 753 if (cpu == msg->nm_inp->inp_pcbinfo->cpu) { 754 /* note: detach removes any wildcard hash entry */ 755 #ifdef INET6 756 if (msg->nm_isinet6) 757 in6_pcbdetach(msg->nm_inp); 758 else 759 #endif 760 in_pcbdetach(msg->nm_inp); 761 lwkt_replymsg(&msg->nm_lmsg, 0); 762 } else { 763 in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo); 764 cpu = (cpu + 1) % ncpus2; 765 msg->nm_pcbinfo = &tcbinfo[cpu]; 766 lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_lmsg); 767 } 768 return (EASYNC); 769 } 770 771 #endif 772 773 /* 774 * Close a TCP control block: 775 * discard all space held by the tcp 776 * discard internet protocol block 777 * wake up any sleepers 778 */ 779 struct tcpcb * 780 tcp_close(struct tcpcb *tp) 781 { 782 struct tseg_qent *q; 783 struct inpcb *inp = tp->t_inpcb; 784 struct socket *so = inp->inp_socket; 785 struct rtentry *rt; 786 boolean_t dosavessthresh; 787 #ifdef SMP 788 int cpu; 789 #endif 790 #ifdef INET6 791 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 792 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0); 793 #else 794 const boolean_t isipv6 = FALSE; 795 #endif 796 797 /* 798 * The tp is not instantly destroyed in the wildcard case. Setting 799 * the state to TCPS_TERMINATING will prevent the TCP stack from 800 * messing with it, though it should be noted that this change may 801 * not take effect on other cpus until we have chained the wildcard 802 * hash removal. 803 * 804 * XXX we currently depend on the BGL to synchronize the tp->t_state 805 * update and prevent other tcp protocol threads from accepting new 806 * connections on the listen socket we might be trying to close down. 807 */ 808 KKASSERT(tp->t_state != TCPS_TERMINATING); 809 tp->t_state = TCPS_TERMINATING; 810 811 /* 812 * Make sure that all of our timers are stopped before we 813 * delete the PCB. 814 */ 815 callout_stop(tp->tt_rexmt); 816 callout_stop(tp->tt_persist); 817 callout_stop(tp->tt_keep); 818 callout_stop(tp->tt_2msl); 819 callout_stop(tp->tt_delack); 820 821 if (tp->t_flags & TF_ONOUTPUTQ) { 822 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 823 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 824 tp->t_flags &= ~TF_ONOUTPUTQ; 825 } 826 827 /* 828 * If we got enough samples through the srtt filter, 829 * save the rtt and rttvar in the routing entry. 830 * 'Enough' is arbitrarily defined as the 16 samples. 831 * 16 samples is enough for the srtt filter to converge 832 * to within 5% of the correct value; fewer samples and 833 * we could save a very bogus rtt. 834 * 835 * Don't update the default route's characteristics and don't 836 * update anything that the user "locked". 837 */ 838 if (tp->t_rttupdated >= 16) { 839 u_long i = 0; 840 841 if (isipv6) { 842 struct sockaddr_in6 *sin6; 843 844 if ((rt = inp->in6p_route.ro_rt) == NULL) 845 goto no_valid_rt; 846 sin6 = (struct sockaddr_in6 *)rt_key(rt); 847 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 848 goto no_valid_rt; 849 } else 850 if ((rt = inp->inp_route.ro_rt) == NULL || 851 ((struct sockaddr_in *)rt_key(rt))-> 852 sin_addr.s_addr == INADDR_ANY) 853 goto no_valid_rt; 854 855 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 856 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 857 if (rt->rt_rmx.rmx_rtt && i) 858 /* 859 * filter this update to half the old & half 860 * the new values, converting scale. 861 * See route.h and tcp_var.h for a 862 * description of the scaling constants. 863 */ 864 rt->rt_rmx.rmx_rtt = 865 (rt->rt_rmx.rmx_rtt + i) / 2; 866 else 867 rt->rt_rmx.rmx_rtt = i; 868 tcpstat.tcps_cachedrtt++; 869 } 870 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 871 i = tp->t_rttvar * 872 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 873 if (rt->rt_rmx.rmx_rttvar && i) 874 rt->rt_rmx.rmx_rttvar = 875 (rt->rt_rmx.rmx_rttvar + i) / 2; 876 else 877 rt->rt_rmx.rmx_rttvar = i; 878 tcpstat.tcps_cachedrttvar++; 879 } 880 /* 881 * The old comment here said: 882 * update the pipelimit (ssthresh) if it has been updated 883 * already or if a pipesize was specified & the threshhold 884 * got below half the pipesize. I.e., wait for bad news 885 * before we start updating, then update on both good 886 * and bad news. 887 * 888 * But we want to save the ssthresh even if no pipesize is 889 * specified explicitly in the route, because such 890 * connections still have an implicit pipesize specified 891 * by the global tcp_sendspace. In the absence of a reliable 892 * way to calculate the pipesize, it will have to do. 893 */ 894 i = tp->snd_ssthresh; 895 if (rt->rt_rmx.rmx_sendpipe != 0) 896 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 897 else 898 dosavessthresh = (i < so->so_snd.sb_hiwat/2); 899 if (dosavessthresh || 900 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 901 (rt->rt_rmx.rmx_ssthresh != 0))) { 902 /* 903 * convert the limit from user data bytes to 904 * packets then to packet data bytes. 905 */ 906 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 907 if (i < 2) 908 i = 2; 909 i *= tp->t_maxseg + 910 (isipv6 ? 911 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 912 sizeof(struct tcpiphdr)); 913 if (rt->rt_rmx.rmx_ssthresh) 914 rt->rt_rmx.rmx_ssthresh = 915 (rt->rt_rmx.rmx_ssthresh + i) / 2; 916 else 917 rt->rt_rmx.rmx_ssthresh = i; 918 tcpstat.tcps_cachedssthresh++; 919 } 920 } 921 922 no_valid_rt: 923 /* free the reassembly queue, if any */ 924 while((q = LIST_FIRST(&tp->t_segq)) != NULL) { 925 LIST_REMOVE(q, tqe_q); 926 m_freem(q->tqe_m); 927 FREE(q, M_TSEGQ); 928 tcp_reass_qsize--; 929 } 930 /* throw away SACK blocks in scoreboard*/ 931 if (TCP_DO_SACK(tp)) 932 tcp_sack_cleanup(&tp->scb); 933 934 inp->inp_ppcb = NULL; 935 soisdisconnected(so); 936 /* 937 * Discard the inp. In the SMP case a wildcard inp's hash (created 938 * by a listen socket or an INADDR_ANY udp socket) is replicated 939 * for each protocol thread and must be removed in the context of 940 * that thread. This is accomplished by chaining the message 941 * through the cpus. 942 * 943 * If the inp is not wildcarded we simply detach, which will remove 944 * the any hashes still present for this inp. 945 */ 946 #ifdef SMP 947 if (inp->inp_flags & INP_WILDCARD_MP) { 948 struct netmsg_remwildcard *msg; 949 950 cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2; 951 msg = malloc(sizeof(struct netmsg_remwildcard), 952 M_LWKTMSG, M_INTWAIT); 953 lwkt_initmsg(&msg->nm_lmsg, &netisr_afree_rport, 0, 954 lwkt_cmd_func(in_pcbremwildcardhash_handler), 955 lwkt_cmd_op_none); 956 #ifdef INET6 957 msg->nm_isinet6 = isafinet6; 958 #endif 959 msg->nm_inp = inp; 960 msg->nm_pcbinfo = &tcbinfo[cpu]; 961 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg); 962 } else 963 #endif 964 { 965 /* note: detach removes any wildcard hash entry */ 966 #ifdef INET6 967 if (isafinet6) 968 in6_pcbdetach(inp); 969 else 970 #endif 971 in_pcbdetach(inp); 972 } 973 tcpstat.tcps_closed++; 974 return (NULL); 975 } 976 977 static __inline void 978 tcp_drain_oncpu(struct inpcbhead *head) 979 { 980 struct inpcb *inpb; 981 struct tcpcb *tcpb; 982 struct tseg_qent *te; 983 984 LIST_FOREACH(inpb, head, inp_list) { 985 if (inpb->inp_flags & INP_PLACEMARKER) 986 continue; 987 if ((tcpb = intotcpcb(inpb))) { 988 while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) { 989 LIST_REMOVE(te, tqe_q); 990 m_freem(te->tqe_m); 991 FREE(te, M_TSEGQ); 992 tcp_reass_qsize--; 993 } 994 } 995 } 996 } 997 998 #ifdef SMP 999 struct netmsg_tcp_drain { 1000 struct lwkt_msg nm_lmsg; 1001 struct inpcbhead *nm_head; 1002 }; 1003 1004 static int 1005 tcp_drain_handler(lwkt_msg_t lmsg) 1006 { 1007 struct netmsg_tcp_drain *nm = (void *)lmsg; 1008 1009 tcp_drain_oncpu(nm->nm_head); 1010 lwkt_replymsg(lmsg, 0); 1011 return(EASYNC); 1012 } 1013 #endif 1014 1015 void 1016 tcp_drain() 1017 { 1018 #ifdef SMP 1019 int cpu; 1020 #endif 1021 1022 if (!do_tcpdrain) 1023 return; 1024 1025 /* 1026 * Walk the tcpbs, if existing, and flush the reassembly queue, 1027 * if there is one... 1028 * XXX: The "Net/3" implementation doesn't imply that the TCP 1029 * reassembly queue should be flushed, but in a situation 1030 * where we're really low on mbufs, this is potentially 1031 * useful. 1032 */ 1033 #ifdef SMP 1034 for (cpu = 0; cpu < ncpus2; cpu++) { 1035 struct netmsg_tcp_drain *msg; 1036 1037 if (cpu == mycpu->gd_cpuid) { 1038 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead); 1039 } else { 1040 msg = malloc(sizeof(struct netmsg_tcp_drain), 1041 M_LWKTMSG, M_NOWAIT); 1042 if (msg == NULL) 1043 continue; 1044 lwkt_initmsg(&msg->nm_lmsg, &netisr_afree_rport, 0, 1045 lwkt_cmd_func(tcp_drain_handler), 1046 lwkt_cmd_op_none); 1047 msg->nm_head = &tcbinfo[cpu].pcblisthead; 1048 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg); 1049 } 1050 } 1051 #else 1052 tcp_drain_oncpu(&tcbinfo[0].pcblisthead); 1053 #endif 1054 } 1055 1056 /* 1057 * Notify a tcp user of an asynchronous error; 1058 * store error as soft error, but wake up user 1059 * (for now, won't do anything until can select for soft error). 1060 * 1061 * Do not wake up user since there currently is no mechanism for 1062 * reporting soft errors (yet - a kqueue filter may be added). 1063 */ 1064 static void 1065 tcp_notify(struct inpcb *inp, int error) 1066 { 1067 struct tcpcb *tp = intotcpcb(inp); 1068 1069 /* 1070 * Ignore some errors if we are hooked up. 1071 * If connection hasn't completed, has retransmitted several times, 1072 * and receives a second error, give up now. This is better 1073 * than waiting a long time to establish a connection that 1074 * can never complete. 1075 */ 1076 if (tp->t_state == TCPS_ESTABLISHED && 1077 (error == EHOSTUNREACH || error == ENETUNREACH || 1078 error == EHOSTDOWN)) { 1079 return; 1080 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1081 tp->t_softerror) 1082 tcp_drop(tp, error); 1083 else 1084 tp->t_softerror = error; 1085 #if 0 1086 wakeup((caddr_t) &so->so_timeo); 1087 sorwakeup(so); 1088 sowwakeup(so); 1089 #endif 1090 } 1091 1092 static int 1093 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1094 { 1095 int error, i, n; 1096 struct inpcb *marker; 1097 struct inpcb *inp; 1098 inp_gen_t gencnt; 1099 globaldata_t gd; 1100 int origcpu, ccpu; 1101 1102 error = 0; 1103 n = 0; 1104 1105 /* 1106 * The process of preparing the TCB list is too time-consuming and 1107 * resource-intensive to repeat twice on every request. 1108 */ 1109 if (req->oldptr == NULL) { 1110 for (ccpu = 0; ccpu < ncpus; ++ccpu) { 1111 gd = globaldata_find(ccpu); 1112 n += tcbinfo[gd->gd_cpuid].ipi_count; 1113 } 1114 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1115 return (0); 1116 } 1117 1118 if (req->newptr != NULL) 1119 return (EPERM); 1120 1121 marker = malloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1122 marker->inp_flags |= INP_PLACEMARKER; 1123 1124 /* 1125 * OK, now we're committed to doing something. Run the inpcb list 1126 * for each cpu in the system and construct the output. Use a 1127 * list placemarker to deal with list changes occuring during 1128 * copyout blockages (but otherwise depend on being on the correct 1129 * cpu to avoid races). 1130 */ 1131 origcpu = mycpu->gd_cpuid; 1132 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) { 1133 globaldata_t rgd; 1134 caddr_t inp_ppcb; 1135 struct xtcpcb xt; 1136 int cpu_id; 1137 1138 cpu_id = (origcpu + ccpu) % ncpus; 1139 if ((smp_active_mask & (1 << cpu_id)) == 0) 1140 continue; 1141 rgd = globaldata_find(cpu_id); 1142 lwkt_setcpu_self(rgd); 1143 1144 /* indicate change of CPU */ 1145 cpu_mb1(); 1146 1147 gencnt = tcbinfo[cpu_id].ipi_gencnt; 1148 n = tcbinfo[cpu_id].ipi_count; 1149 1150 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list); 1151 i = 0; 1152 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1153 /* 1154 * process a snapshot of pcbs, ignoring placemarkers 1155 * and using our own to allow SYSCTL_OUT to block. 1156 */ 1157 LIST_REMOVE(marker, inp_list); 1158 LIST_INSERT_AFTER(inp, marker, inp_list); 1159 1160 if (inp->inp_flags & INP_PLACEMARKER) 1161 continue; 1162 if (inp->inp_gencnt > gencnt) 1163 continue; 1164 if (prison_xinpcb(req->td, inp)) 1165 continue; 1166 1167 xt.xt_len = sizeof xt; 1168 bcopy(inp, &xt.xt_inp, sizeof *inp); 1169 inp_ppcb = inp->inp_ppcb; 1170 if (inp_ppcb != NULL) 1171 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1172 else 1173 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1174 if (inp->inp_socket) 1175 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1176 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1177 break; 1178 ++i; 1179 } 1180 LIST_REMOVE(marker, inp_list); 1181 if (error == 0 && i < n) { 1182 bzero(&xt, sizeof(xt)); 1183 xt.xt_len = sizeof(xt); 1184 while (i < n) { 1185 error = SYSCTL_OUT(req, &xt, sizeof (xt)); 1186 if (error) 1187 break; 1188 ++i; 1189 } 1190 } 1191 } 1192 1193 /* 1194 * Make sure we are on the same cpu we were on originally, since 1195 * higher level callers expect this. Also don't pollute caches with 1196 * migrated userland data by (eventually) returning to userland 1197 * on a different cpu. 1198 */ 1199 lwkt_setcpu_self(globaldata_find(origcpu)); 1200 free(marker, M_TEMP); 1201 return (error); 1202 } 1203 1204 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1205 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1206 1207 static int 1208 tcp_getcred(SYSCTL_HANDLER_ARGS) 1209 { 1210 struct sockaddr_in addrs[2]; 1211 struct inpcb *inp; 1212 int cpu; 1213 int error, s; 1214 1215 error = suser(req->td); 1216 if (error != 0) 1217 return (error); 1218 error = SYSCTL_IN(req, addrs, sizeof addrs); 1219 if (error != 0) 1220 return (error); 1221 s = splnet(); 1222 1223 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1224 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1225 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1226 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1227 if (inp == NULL || inp->inp_socket == NULL) { 1228 error = ENOENT; 1229 goto out; 1230 } 1231 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1232 out: 1233 splx(s); 1234 return (error); 1235 } 1236 1237 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1238 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1239 1240 #ifdef INET6 1241 static int 1242 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1243 { 1244 struct sockaddr_in6 addrs[2]; 1245 struct inpcb *inp; 1246 int error, s; 1247 boolean_t mapped = FALSE; 1248 1249 error = suser(req->td); 1250 if (error != 0) 1251 return (error); 1252 error = SYSCTL_IN(req, addrs, sizeof addrs); 1253 if (error != 0) 1254 return (error); 1255 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { 1256 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) 1257 mapped = TRUE; 1258 else 1259 return (EINVAL); 1260 } 1261 s = splnet(); 1262 if (mapped) { 1263 inp = in_pcblookup_hash(&tcbinfo[0], 1264 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], 1265 addrs[1].sin6_port, 1266 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], 1267 addrs[0].sin6_port, 1268 0, NULL); 1269 } else { 1270 inp = in6_pcblookup_hash(&tcbinfo[0], 1271 &addrs[1].sin6_addr, addrs[1].sin6_port, 1272 &addrs[0].sin6_addr, addrs[0].sin6_port, 1273 0, NULL); 1274 } 1275 if (inp == NULL || inp->inp_socket == NULL) { 1276 error = ENOENT; 1277 goto out; 1278 } 1279 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1280 out: 1281 splx(s); 1282 return (error); 1283 } 1284 1285 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1286 0, 0, 1287 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1288 #endif 1289 1290 void 1291 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip) 1292 { 1293 struct ip *ip = vip; 1294 struct tcphdr *th; 1295 struct in_addr faddr; 1296 struct inpcb *inp; 1297 struct tcpcb *tp; 1298 void (*notify)(struct inpcb *, int) = tcp_notify; 1299 tcp_seq icmp_seq; 1300 int cpu; 1301 int s; 1302 1303 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1304 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1305 return; 1306 1307 if (cmd == PRC_QUENCH) 1308 notify = tcp_quench; 1309 else if (icmp_may_rst && 1310 (cmd == PRC_UNREACH_ADMIN_PROHIB || cmd == PRC_UNREACH_PORT || 1311 cmd == PRC_TIMXCEED_INTRANS) && 1312 ip != NULL) 1313 notify = tcp_drop_syn_sent; 1314 else if (cmd == PRC_MSGSIZE) 1315 notify = tcp_mtudisc; 1316 else if (PRC_IS_REDIRECT(cmd)) { 1317 ip = NULL; 1318 notify = in_rtchange; 1319 } else if (cmd == PRC_HOSTDEAD) 1320 ip = NULL; 1321 else if ((unsigned)cmd > PRC_NCMDS || inetctlerrmap[cmd] == 0) 1322 return; 1323 if (ip != NULL) { 1324 s = splnet(); 1325 th = (struct tcphdr *)((caddr_t)ip + 1326 (IP_VHL_HL(ip->ip_vhl) << 2)); 1327 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport, 1328 ip->ip_src.s_addr, th->th_sport); 1329 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport, 1330 ip->ip_src, th->th_sport, 0, NULL); 1331 if ((inp != NULL) && (inp->inp_socket != NULL)) { 1332 icmp_seq = htonl(th->th_seq); 1333 tp = intotcpcb(inp); 1334 if (SEQ_GEQ(icmp_seq, tp->snd_una) && 1335 SEQ_LT(icmp_seq, tp->snd_max)) 1336 (*notify)(inp, inetctlerrmap[cmd]); 1337 } else { 1338 struct in_conninfo inc; 1339 1340 inc.inc_fport = th->th_dport; 1341 inc.inc_lport = th->th_sport; 1342 inc.inc_faddr = faddr; 1343 inc.inc_laddr = ip->ip_src; 1344 #ifdef INET6 1345 inc.inc_isipv6 = 0; 1346 #endif 1347 syncache_unreach(&inc, th); 1348 } 1349 splx(s); 1350 } else { 1351 for (cpu = 0; cpu < ncpus2; cpu++) { 1352 in_pcbnotifyall(&tcbinfo[cpu].pcblisthead, faddr, 1353 inetctlerrmap[cmd], notify); 1354 } 1355 } 1356 } 1357 1358 #ifdef INET6 1359 void 1360 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d) 1361 { 1362 struct tcphdr th; 1363 void (*notify) (struct inpcb *, int) = tcp_notify; 1364 struct ip6_hdr *ip6; 1365 struct mbuf *m; 1366 struct ip6ctlparam *ip6cp = NULL; 1367 const struct sockaddr_in6 *sa6_src = NULL; 1368 int off; 1369 struct tcp_portonly { 1370 u_int16_t th_sport; 1371 u_int16_t th_dport; 1372 } *thp; 1373 1374 if (sa->sa_family != AF_INET6 || 1375 sa->sa_len != sizeof(struct sockaddr_in6)) 1376 return; 1377 1378 if (cmd == PRC_QUENCH) 1379 notify = tcp_quench; 1380 else if (cmd == PRC_MSGSIZE) 1381 notify = tcp_mtudisc; 1382 else if (!PRC_IS_REDIRECT(cmd) && 1383 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) 1384 return; 1385 1386 /* if the parameter is from icmp6, decode it. */ 1387 if (d != NULL) { 1388 ip6cp = (struct ip6ctlparam *)d; 1389 m = ip6cp->ip6c_m; 1390 ip6 = ip6cp->ip6c_ip6; 1391 off = ip6cp->ip6c_off; 1392 sa6_src = ip6cp->ip6c_src; 1393 } else { 1394 m = NULL; 1395 ip6 = NULL; 1396 off = 0; /* fool gcc */ 1397 sa6_src = &sa6_any; 1398 } 1399 1400 if (ip6 != NULL) { 1401 struct in_conninfo inc; 1402 /* 1403 * XXX: We assume that when IPV6 is non NULL, 1404 * M and OFF are valid. 1405 */ 1406 1407 /* check if we can safely examine src and dst ports */ 1408 if (m->m_pkthdr.len < off + sizeof *thp) 1409 return; 1410 1411 bzero(&th, sizeof th); 1412 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1413 1414 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport, 1415 (struct sockaddr *)ip6cp->ip6c_src, 1416 th.th_sport, cmd, notify); 1417 1418 inc.inc_fport = th.th_dport; 1419 inc.inc_lport = th.th_sport; 1420 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1421 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1422 inc.inc_isipv6 = 1; 1423 syncache_unreach(&inc, &th); 1424 } else 1425 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0, 1426 (const struct sockaddr *)sa6_src, 0, cmd, notify); 1427 } 1428 #endif 1429 1430 /* 1431 * Following is where TCP initial sequence number generation occurs. 1432 * 1433 * There are two places where we must use initial sequence numbers: 1434 * 1. In SYN-ACK packets. 1435 * 2. In SYN packets. 1436 * 1437 * All ISNs for SYN-ACK packets are generated by the syncache. See 1438 * tcp_syncache.c for details. 1439 * 1440 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1441 * depends on this property. In addition, these ISNs should be 1442 * unguessable so as to prevent connection hijacking. To satisfy 1443 * the requirements of this situation, the algorithm outlined in 1444 * RFC 1948 is used to generate sequence numbers. 1445 * 1446 * Implementation details: 1447 * 1448 * Time is based off the system timer, and is corrected so that it 1449 * increases by one megabyte per second. This allows for proper 1450 * recycling on high speed LANs while still leaving over an hour 1451 * before rollover. 1452 * 1453 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1454 * between seeding of isn_secret. This is normally set to zero, 1455 * as reseeding should not be necessary. 1456 * 1457 */ 1458 1459 #define ISN_BYTES_PER_SECOND 1048576 1460 1461 u_char isn_secret[32]; 1462 int isn_last_reseed; 1463 MD5_CTX isn_ctx; 1464 1465 tcp_seq 1466 tcp_new_isn(struct tcpcb *tp) 1467 { 1468 u_int32_t md5_buffer[4]; 1469 tcp_seq new_isn; 1470 1471 /* Seed if this is the first use, reseed if requested. */ 1472 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1473 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1474 < (u_int)ticks))) { 1475 read_random_unlimited(&isn_secret, sizeof isn_secret); 1476 isn_last_reseed = ticks; 1477 } 1478 1479 /* Compute the md5 hash and return the ISN. */ 1480 MD5Init(&isn_ctx); 1481 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1482 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1483 #ifdef INET6 1484 if (tp->t_inpcb->inp_vflag & INP_IPV6) { 1485 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1486 sizeof(struct in6_addr)); 1487 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1488 sizeof(struct in6_addr)); 1489 } else 1490 #endif 1491 { 1492 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1493 sizeof(struct in_addr)); 1494 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1495 sizeof(struct in_addr)); 1496 } 1497 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1498 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1499 new_isn = (tcp_seq) md5_buffer[0]; 1500 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1501 return (new_isn); 1502 } 1503 1504 /* 1505 * When a source quench is received, close congestion window 1506 * to one segment. We will gradually open it again as we proceed. 1507 */ 1508 void 1509 tcp_quench(struct inpcb *inp, int errno) 1510 { 1511 struct tcpcb *tp = intotcpcb(inp); 1512 1513 if (tp != NULL) 1514 tp->snd_cwnd = tp->t_maxseg; 1515 } 1516 1517 /* 1518 * When a specific ICMP unreachable message is received and the 1519 * connection state is SYN-SENT, drop the connection. This behavior 1520 * is controlled by the icmp_may_rst sysctl. 1521 */ 1522 void 1523 tcp_drop_syn_sent(struct inpcb *inp, int errno) 1524 { 1525 struct tcpcb *tp = intotcpcb(inp); 1526 1527 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT)) 1528 tcp_drop(tp, errno); 1529 } 1530 1531 /* 1532 * When `need fragmentation' ICMP is received, update our idea of the MSS 1533 * based on the new value in the route. Also nudge TCP to send something, 1534 * since we know the packet we just sent was dropped. 1535 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1536 */ 1537 void 1538 tcp_mtudisc(struct inpcb *inp, int errno) 1539 { 1540 struct tcpcb *tp = intotcpcb(inp); 1541 struct rtentry *rt; 1542 struct rmxp_tao *taop; 1543 struct socket *so = inp->inp_socket; 1544 int offered; 1545 int mss; 1546 #ifdef INET6 1547 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0); 1548 #else 1549 const boolean_t isipv6 = FALSE; 1550 #endif 1551 1552 if (tp != NULL) { 1553 if (isipv6) 1554 rt = tcp_rtlookup6(&inp->inp_inc); 1555 else 1556 rt = tcp_rtlookup(&inp->inp_inc); 1557 if (rt == NULL || rt->rt_rmx.rmx_mtu == 0) { 1558 tp->t_maxopd = tp->t_maxseg = 1559 isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 1560 return; 1561 } 1562 taop = rmx_taop(rt->rt_rmx); 1563 offered = taop->tao_mssopt; 1564 mss = rt->rt_rmx.rmx_mtu - 1565 (isipv6 ? 1566 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1567 sizeof(struct tcpiphdr)); 1568 1569 if (offered != 0) 1570 mss = min(mss, offered); 1571 /* 1572 * XXX - The above conditional probably violates the TCP 1573 * spec. The problem is that, since we don't know the 1574 * other end's MSS, we are supposed to use a conservative 1575 * default. But, if we do that, then MTU discovery will 1576 * never actually take place, because the conservative 1577 * default is much less than the MTUs typically seen 1578 * on the Internet today. For the moment, we'll sweep 1579 * this under the carpet. 1580 * 1581 * The conservative default might not actually be a problem 1582 * if the only case this occurs is when sending an initial 1583 * SYN with options and data to a host we've never talked 1584 * to before. Then, they will reply with an MSS value which 1585 * will get recorded and the new parameters should get 1586 * recomputed. For Further Study. 1587 */ 1588 if (tp->t_maxopd <= mss) 1589 return; 1590 tp->t_maxopd = mss; 1591 1592 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP && 1593 (tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP) 1594 mss -= TCPOLEN_TSTAMP_APPA; 1595 if ((tp->t_flags & (TF_REQ_CC | TF_NOOPT)) == TF_REQ_CC && 1596 (tp->t_flags & TF_RCVD_CC) == TF_RCVD_CC) 1597 mss -= TCPOLEN_CC_APPA; 1598 #if (MCLBYTES & (MCLBYTES - 1)) == 0 1599 if (mss > MCLBYTES) 1600 mss &= ~(MCLBYTES - 1); 1601 #else 1602 if (mss > MCLBYTES) 1603 mss = mss / MCLBYTES * MCLBYTES; 1604 #endif 1605 if (so->so_snd.sb_hiwat < mss) 1606 mss = so->so_snd.sb_hiwat; 1607 1608 tp->t_maxseg = mss; 1609 1610 tcpstat.tcps_mturesent++; 1611 tp->t_rtttime = 0; 1612 tp->snd_nxt = tp->snd_una; 1613 tcp_output(tp); 1614 } 1615 } 1616 1617 /* 1618 * Look-up the routing entry to the peer of this inpcb. If no route 1619 * is found and it cannot be allocated the return NULL. This routine 1620 * is called by TCP routines that access the rmx structure and by tcp_mss 1621 * to get the interface MTU. 1622 */ 1623 struct rtentry * 1624 tcp_rtlookup(struct in_conninfo *inc) 1625 { 1626 struct route *ro; 1627 struct rtentry *rt; 1628 1629 ro = &inc->inc_route; 1630 rt = ro->ro_rt; 1631 if (rt == NULL || !(rt->rt_flags & RTF_UP)) { 1632 /* No route yet, so try to acquire one */ 1633 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1634 /* 1635 * unused portions of the structure MUST be zero'd 1636 * out because rtalloc() treats it as opaque data 1637 */ 1638 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1639 ro->ro_dst.sa_family = AF_INET; 1640 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1641 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1642 inc->inc_faddr; 1643 rtalloc(ro); 1644 rt = ro->ro_rt; 1645 } 1646 } 1647 return (rt); 1648 } 1649 1650 #ifdef INET6 1651 struct rtentry * 1652 tcp_rtlookup6(struct in_conninfo *inc) 1653 { 1654 struct route_in6 *ro6; 1655 struct rtentry *rt; 1656 1657 ro6 = &inc->inc6_route; 1658 rt = ro6->ro_rt; 1659 if (rt == NULL || !(rt->rt_flags & RTF_UP)) { 1660 /* No route yet, so try to acquire one */ 1661 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1662 /* 1663 * unused portions of the structure MUST be zero'd 1664 * out because rtalloc() treats it as opaque data 1665 */ 1666 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1667 ro6->ro_dst.sin6_family = AF_INET6; 1668 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1669 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1670 rtalloc((struct route *)ro6); 1671 rt = ro6->ro_rt; 1672 } 1673 } 1674 return (rt); 1675 } 1676 #endif 1677 1678 #ifdef IPSEC 1679 /* compute ESP/AH header size for TCP, including outer IP header. */ 1680 size_t 1681 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1682 { 1683 struct inpcb *inp; 1684 struct mbuf *m; 1685 size_t hdrsiz; 1686 struct ip *ip; 1687 struct tcphdr *th; 1688 1689 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1690 return (0); 1691 MGETHDR(m, MB_DONTWAIT, MT_DATA); 1692 if (!m) 1693 return (0); 1694 1695 #ifdef INET6 1696 if (inp->inp_vflag & INP_IPV6) { 1697 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1698 1699 th = (struct tcphdr *)(ip6 + 1); 1700 m->m_pkthdr.len = m->m_len = 1701 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1702 tcp_fillheaders(tp, ip6, th); 1703 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1704 } else 1705 #endif 1706 { 1707 ip = mtod(m, struct ip *); 1708 th = (struct tcphdr *)(ip + 1); 1709 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1710 tcp_fillheaders(tp, ip, th); 1711 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1712 } 1713 1714 m_free(m); 1715 return (hdrsiz); 1716 } 1717 #endif 1718 1719 /* 1720 * Return a pointer to the cached information about the remote host. 1721 * The cached information is stored in the protocol specific part of 1722 * the route metrics. 1723 */ 1724 struct rmxp_tao * 1725 tcp_gettaocache(struct in_conninfo *inc) 1726 { 1727 struct rtentry *rt; 1728 1729 #ifdef INET6 1730 if (inc->inc_isipv6) 1731 rt = tcp_rtlookup6(inc); 1732 else 1733 #endif 1734 rt = tcp_rtlookup(inc); 1735 1736 /* Make sure this is a host route and is up. */ 1737 if (rt == NULL || 1738 (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST)) 1739 return (NULL); 1740 1741 return (rmx_taop(rt->rt_rmx)); 1742 } 1743 1744 /* 1745 * Clear all the TAO cache entries, called from tcp_init. 1746 * 1747 * XXX 1748 * This routine is just an empty one, because we assume that the routing 1749 * routing tables are initialized at the same time when TCP, so there is 1750 * nothing in the cache left over. 1751 */ 1752 static void 1753 tcp_cleartaocache() 1754 { 1755 } 1756 1757 /* 1758 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1759 * 1760 * This code attempts to calculate the bandwidth-delay product as a 1761 * means of determining the optimal window size to maximize bandwidth, 1762 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1763 * routers. This code also does a fairly good job keeping RTTs in check 1764 * across slow links like modems. We implement an algorithm which is very 1765 * similar (but not meant to be) TCP/Vegas. The code operates on the 1766 * transmitter side of a TCP connection and so only effects the transmit 1767 * side of the connection. 1768 * 1769 * BACKGROUND: TCP makes no provision for the management of buffer space 1770 * at the end points or at the intermediate routers and switches. A TCP 1771 * stream, whether using NewReno or not, will eventually buffer as 1772 * many packets as it is able and the only reason this typically works is 1773 * due to the fairly small default buffers made available for a connection 1774 * (typicaly 16K or 32K). As machines use larger windows and/or window 1775 * scaling it is now fairly easy for even a single TCP connection to blow-out 1776 * all available buffer space not only on the local interface, but on 1777 * intermediate routers and switches as well. NewReno makes a misguided 1778 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1779 * then backing off, then steadily increasing the window again until another 1780 * failure occurs, ad-infinitum. This results in terrible oscillation that 1781 * is only made worse as network loads increase and the idea of intentionally 1782 * blowing out network buffers is, frankly, a terrible way to manage network 1783 * resources. 1784 * 1785 * It is far better to limit the transmit window prior to the failure 1786 * condition being achieved. There are two general ways to do this: First 1787 * you can 'scan' through different transmit window sizes and locate the 1788 * point where the RTT stops increasing, indicating that you have filled the 1789 * pipe, then scan backwards until you note that RTT stops decreasing, then 1790 * repeat ad-infinitum. This method works in principle but has severe 1791 * implementation issues due to RTT variances, timer granularity, and 1792 * instability in the algorithm which can lead to many false positives and 1793 * create oscillations as well as interact badly with other TCP streams 1794 * implementing the same algorithm. 1795 * 1796 * The second method is to limit the window to the bandwidth delay product 1797 * of the link. This is the method we implement. RTT variances and our 1798 * own manipulation of the congestion window, bwnd, can potentially 1799 * destabilize the algorithm. For this reason we have to stabilize the 1800 * elements used to calculate the window. We do this by using the minimum 1801 * observed RTT, the long term average of the observed bandwidth, and 1802 * by adding two segments worth of slop. It isn't perfect but it is able 1803 * to react to changing conditions and gives us a very stable basis on 1804 * which to extend the algorithm. 1805 */ 1806 void 1807 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1808 { 1809 u_long bw; 1810 u_long bwnd; 1811 int save_ticks; 1812 int delta_ticks; 1813 1814 /* 1815 * If inflight_enable is disabled in the middle of a tcp connection, 1816 * make sure snd_bwnd is effectively disabled. 1817 */ 1818 if (!tcp_inflight_enable) { 1819 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1820 tp->snd_bandwidth = 0; 1821 return; 1822 } 1823 1824 /* 1825 * Validate the delta time. If a connection is new or has been idle 1826 * a long time we have to reset the bandwidth calculator. 1827 */ 1828 save_ticks = ticks; 1829 delta_ticks = save_ticks - tp->t_bw_rtttime; 1830 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1831 tp->t_bw_rtttime = ticks; 1832 tp->t_bw_rtseq = ack_seq; 1833 if (tp->snd_bandwidth == 0) 1834 tp->snd_bandwidth = tcp_inflight_min; 1835 return; 1836 } 1837 if (delta_ticks == 0) 1838 return; 1839 1840 /* 1841 * Sanity check, plus ignore pure window update acks. 1842 */ 1843 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1844 return; 1845 1846 /* 1847 * Figure out the bandwidth. Due to the tick granularity this 1848 * is a very rough number and it MUST be averaged over a fairly 1849 * long period of time. XXX we need to take into account a link 1850 * that is not using all available bandwidth, but for now our 1851 * slop will ramp us up if this case occurs and the bandwidth later 1852 * increases. 1853 */ 1854 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 1855 tp->t_bw_rtttime = save_ticks; 1856 tp->t_bw_rtseq = ack_seq; 1857 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4; 1858 1859 tp->snd_bandwidth = bw; 1860 1861 /* 1862 * Calculate the semi-static bandwidth delay product, plus two maximal 1863 * segments. The additional slop puts us squarely in the sweet 1864 * spot and also handles the bandwidth run-up case. Without the 1865 * slop we could be locking ourselves into a lower bandwidth. 1866 * 1867 * Situations Handled: 1868 * (1) Prevents over-queueing of packets on LANs, especially on 1869 * high speed LANs, allowing larger TCP buffers to be 1870 * specified, and also does a good job preventing 1871 * over-queueing of packets over choke points like modems 1872 * (at least for the transmit side). 1873 * 1874 * (2) Is able to handle changing network loads (bandwidth 1875 * drops so bwnd drops, bandwidth increases so bwnd 1876 * increases). 1877 * 1878 * (3) Theoretically should stabilize in the face of multiple 1879 * connections implementing the same algorithm (this may need 1880 * a little work). 1881 * 1882 * (4) Stability value (defaults to 20 = 2 maximal packets) can 1883 * be adjusted with a sysctl but typically only needs to be on 1884 * very slow connections. A value no smaller then 5 should 1885 * be used, but only reduce this default if you have no other 1886 * choice. 1887 */ 1888 1889 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2) 1890 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 1891 tcp_inflight_stab * (int)tp->t_maxseg / 10; 1892 #undef USERTT 1893 1894 if (tcp_inflight_debug > 0) { 1895 static int ltime; 1896 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) { 1897 ltime = ticks; 1898 printf("%p bw %ld rttbest %d srtt %d bwnd %ld\n", 1899 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd); 1900 } 1901 } 1902 if ((long)bwnd < tcp_inflight_min) 1903 bwnd = tcp_inflight_min; 1904 if (bwnd > tcp_inflight_max) 1905 bwnd = tcp_inflight_max; 1906 if ((long)bwnd < tp->t_maxseg * 2) 1907 bwnd = tp->t_maxseg * 2; 1908 tp->snd_bwnd = bwnd; 1909 } 1910