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 1.43 2004/12/21 02:54:15 hsu 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, &tcpstats_ary[cpu], 248 sizeof(struct tcp_stats)))) 249 break; 250 if ((error = SYSCTL_IN(req, &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, &tmp->tt_ipgen, &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 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 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL); 637 if ((ro == &sro) && (ro->ro_rt != NULL)) { 638 RTFREE(ro->ro_rt); 639 ro->ro_rt = NULL; 640 } 641 } 642 } 643 644 /* 645 * Create a new TCP control block, making an 646 * empty reassembly queue and hooking it to the argument 647 * protocol control block. The `inp' parameter must have 648 * come from the zone allocator set up in tcp_init(). 649 */ 650 struct tcpcb * 651 tcp_newtcpcb(struct inpcb *inp) 652 { 653 struct inp_tp *it; 654 struct tcpcb *tp; 655 #ifdef INET6 656 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 657 #else 658 const boolean_t isipv6 = FALSE; 659 #endif 660 661 it = (struct inp_tp *)inp; 662 tp = &it->tcb; 663 bzero(tp, sizeof(struct tcpcb)); 664 LIST_INIT(&tp->t_segq); 665 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 666 667 /* Set up our timeouts. */ 668 callout_init(tp->tt_rexmt = &it->inp_tp_rexmt); 669 callout_init(tp->tt_persist = &it->inp_tp_persist); 670 callout_init(tp->tt_keep = &it->inp_tp_keep); 671 callout_init(tp->tt_2msl = &it->inp_tp_2msl); 672 callout_init(tp->tt_delack = &it->inp_tp_delack); 673 674 if (tcp_do_rfc1323) 675 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP); 676 if (tcp_do_rfc1644) 677 tp->t_flags |= TF_REQ_CC; 678 tp->t_inpcb = inp; /* XXX */ 679 tp->t_state = TCPS_CLOSED; 680 /* 681 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 682 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 683 * reasonable initial retransmit time. 684 */ 685 tp->t_srtt = TCPTV_SRTTBASE; 686 tp->t_rttvar = 687 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 688 tp->t_rttmin = tcp_rexmit_min; 689 tp->t_rxtcur = TCPTV_RTOBASE; 690 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 691 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 692 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 693 tp->t_rcvtime = ticks; 694 /* 695 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 696 * because the socket may be bound to an IPv6 wildcard address, 697 * which may match an IPv4-mapped IPv6 address. 698 */ 699 inp->inp_ip_ttl = ip_defttl; 700 inp->inp_ppcb = tp; 701 tcp_sack_tcpcb_init(tp); 702 return (tp); /* XXX */ 703 } 704 705 /* 706 * Drop a TCP connection, reporting the specified error. 707 * If connection is synchronized, then send a RST to peer. 708 */ 709 struct tcpcb * 710 tcp_drop(struct tcpcb *tp, int errno) 711 { 712 struct socket *so = tp->t_inpcb->inp_socket; 713 714 if (TCPS_HAVERCVDSYN(tp->t_state)) { 715 tp->t_state = TCPS_CLOSED; 716 tcp_output(tp); 717 tcpstat.tcps_drops++; 718 } else 719 tcpstat.tcps_conndrops++; 720 if (errno == ETIMEDOUT && tp->t_softerror) 721 errno = tp->t_softerror; 722 so->so_error = errno; 723 return (tcp_close(tp)); 724 } 725 726 #ifdef SMP 727 728 struct netmsg_remwildcard { 729 struct lwkt_msg nm_lmsg; 730 struct inpcb *nm_inp; 731 struct inpcbinfo *nm_pcbinfo; 732 #if defined(INET6) 733 int nm_isinet6; 734 #else 735 int nm_unused01; 736 #endif 737 }; 738 739 /* 740 * Wildcard inpcb's on SMP boxes must be removed from all cpus before the 741 * inp can be detached. We do this by cycling through the cpus, ending up 742 * on the cpu controlling the inp last and then doing the disconnect. 743 */ 744 static int 745 in_pcbremwildcardhash_handler(struct lwkt_msg *msg0) 746 { 747 struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0; 748 int cpu; 749 750 cpu = msg->nm_pcbinfo->cpu; 751 752 if (cpu == msg->nm_inp->inp_pcbinfo->cpu) { 753 /* note: detach removes any wildcard hash entry */ 754 #ifdef INET6 755 if (msg->nm_isinet6) 756 in6_pcbdetach(msg->nm_inp); 757 else 758 #endif 759 in_pcbdetach(msg->nm_inp); 760 lwkt_replymsg(&msg->nm_lmsg, 0); 761 } else { 762 in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo); 763 cpu = (cpu + 1) % ncpus2; 764 msg->nm_pcbinfo = &tcbinfo[cpu]; 765 lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_lmsg); 766 } 767 return (EASYNC); 768 } 769 770 #endif 771 772 /* 773 * Close a TCP control block: 774 * discard all space held by the tcp 775 * discard internet protocol block 776 * wake up any sleepers 777 */ 778 struct tcpcb * 779 tcp_close(struct tcpcb *tp) 780 { 781 struct tseg_qent *q; 782 struct inpcb *inp = tp->t_inpcb; 783 struct socket *so = inp->inp_socket; 784 struct rtentry *rt; 785 boolean_t dosavessthresh; 786 #ifdef SMP 787 int cpu; 788 #endif 789 #ifdef INET6 790 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 791 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0); 792 #else 793 const boolean_t isipv6 = FALSE; 794 #endif 795 796 /* 797 * The tp is not instantly destroyed in the wildcard case. Setting 798 * the state to TCPS_TERMINATING will prevent the TCP stack from 799 * messing with it, though it should be noted that this change may 800 * not take effect on other cpus until we have chained the wildcard 801 * hash removal. 802 * 803 * XXX we currently depend on the BGL to synchronize the tp->t_state 804 * update and prevent other tcp protocol threads from accepting new 805 * connections on the listen socket we might be trying to close down. 806 */ 807 KKASSERT(tp->t_state != TCPS_TERMINATING); 808 tp->t_state = TCPS_TERMINATING; 809 810 /* 811 * Make sure that all of our timers are stopped before we 812 * delete the PCB. 813 */ 814 callout_stop(tp->tt_rexmt); 815 callout_stop(tp->tt_persist); 816 callout_stop(tp->tt_keep); 817 callout_stop(tp->tt_2msl); 818 callout_stop(tp->tt_delack); 819 820 if (tp->t_flags & TF_ONOUTPUTQ) { 821 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 822 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 823 tp->t_flags &= ~TF_ONOUTPUTQ; 824 } 825 826 /* 827 * If we got enough samples through the srtt filter, 828 * save the rtt and rttvar in the routing entry. 829 * 'Enough' is arbitrarily defined as the 16 samples. 830 * 16 samples is enough for the srtt filter to converge 831 * to within 5% of the correct value; fewer samples and 832 * we could save a very bogus rtt. 833 * 834 * Don't update the default route's characteristics and don't 835 * update anything that the user "locked". 836 */ 837 if (tp->t_rttupdated >= 16) { 838 u_long i = 0; 839 840 if (isipv6) { 841 struct sockaddr_in6 *sin6; 842 843 if ((rt = inp->in6p_route.ro_rt) == NULL) 844 goto no_valid_rt; 845 sin6 = (struct sockaddr_in6 *)rt_key(rt); 846 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 847 goto no_valid_rt; 848 } else 849 if ((rt = inp->inp_route.ro_rt) == NULL || 850 ((struct sockaddr_in *)rt_key(rt))-> 851 sin_addr.s_addr == INADDR_ANY) 852 goto no_valid_rt; 853 854 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 855 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 856 if (rt->rt_rmx.rmx_rtt && i) 857 /* 858 * filter this update to half the old & half 859 * the new values, converting scale. 860 * See route.h and tcp_var.h for a 861 * description of the scaling constants. 862 */ 863 rt->rt_rmx.rmx_rtt = 864 (rt->rt_rmx.rmx_rtt + i) / 2; 865 else 866 rt->rt_rmx.rmx_rtt = i; 867 tcpstat.tcps_cachedrtt++; 868 } 869 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 870 i = tp->t_rttvar * 871 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 872 if (rt->rt_rmx.rmx_rttvar && i) 873 rt->rt_rmx.rmx_rttvar = 874 (rt->rt_rmx.rmx_rttvar + i) / 2; 875 else 876 rt->rt_rmx.rmx_rttvar = i; 877 tcpstat.tcps_cachedrttvar++; 878 } 879 /* 880 * The old comment here said: 881 * update the pipelimit (ssthresh) if it has been updated 882 * already or if a pipesize was specified & the threshhold 883 * got below half the pipesize. I.e., wait for bad news 884 * before we start updating, then update on both good 885 * and bad news. 886 * 887 * But we want to save the ssthresh even if no pipesize is 888 * specified explicitly in the route, because such 889 * connections still have an implicit pipesize specified 890 * by the global tcp_sendspace. In the absence of a reliable 891 * way to calculate the pipesize, it will have to do. 892 */ 893 i = tp->snd_ssthresh; 894 if (rt->rt_rmx.rmx_sendpipe != 0) 895 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 896 else 897 dosavessthresh = (i < so->so_snd.sb_hiwat/2); 898 if (dosavessthresh || 899 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 900 (rt->rt_rmx.rmx_ssthresh != 0))) { 901 /* 902 * convert the limit from user data bytes to 903 * packets then to packet data bytes. 904 */ 905 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 906 if (i < 2) 907 i = 2; 908 i *= tp->t_maxseg + 909 (isipv6 ? 910 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 911 sizeof(struct tcpiphdr)); 912 if (rt->rt_rmx.rmx_ssthresh) 913 rt->rt_rmx.rmx_ssthresh = 914 (rt->rt_rmx.rmx_ssthresh + i) / 2; 915 else 916 rt->rt_rmx.rmx_ssthresh = i; 917 tcpstat.tcps_cachedssthresh++; 918 } 919 } 920 921 no_valid_rt: 922 /* free the reassembly queue, if any */ 923 while((q = LIST_FIRST(&tp->t_segq)) != NULL) { 924 LIST_REMOVE(q, tqe_q); 925 m_freem(q->tqe_m); 926 FREE(q, M_TSEGQ); 927 tcp_reass_qsize--; 928 } 929 /* throw away SACK blocks in scoreboard*/ 930 if (TCP_DO_SACK(tp)) 931 tcp_sack_cleanup(&tp->scb); 932 933 inp->inp_ppcb = NULL; 934 soisdisconnected(so); 935 /* 936 * Discard the inp. In the SMP case a wildcard inp's hash (created 937 * by a listen socket or an INADDR_ANY udp socket) is replicated 938 * for each protocol thread and must be removed in the context of 939 * that thread. This is accomplished by chaining the message 940 * through the cpus. 941 * 942 * If the inp is not wildcarded we simply detach, which will remove 943 * the any hashes still present for this inp. 944 */ 945 #ifdef SMP 946 if (inp->inp_flags & INP_WILDCARD_MP) { 947 struct netmsg_remwildcard *msg; 948 949 cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2; 950 msg = malloc(sizeof(struct netmsg_remwildcard), 951 M_LWKTMSG, M_INTWAIT); 952 lwkt_initmsg(&msg->nm_lmsg, &netisr_afree_rport, 0, 953 lwkt_cmd_func(in_pcbremwildcardhash_handler), 954 lwkt_cmd_op_none); 955 #ifdef INET6 956 msg->nm_isinet6 = isafinet6; 957 #endif 958 msg->nm_inp = inp; 959 msg->nm_pcbinfo = &tcbinfo[cpu]; 960 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg); 961 } else 962 #endif 963 { 964 /* note: detach removes any wildcard hash entry */ 965 #ifdef INET6 966 if (isafinet6) 967 in6_pcbdetach(inp); 968 else 969 #endif 970 in_pcbdetach(inp); 971 } 972 tcpstat.tcps_closed++; 973 return (NULL); 974 } 975 976 static __inline void 977 tcp_drain_oncpu(struct inpcbhead *head) 978 { 979 struct inpcb *inpb; 980 struct tcpcb *tcpb; 981 struct tseg_qent *te; 982 983 LIST_FOREACH(inpb, head, inp_list) { 984 if (inpb->inp_flags & INP_PLACEMARKER) 985 continue; 986 if ((tcpb = intotcpcb(inpb))) { 987 while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) { 988 LIST_REMOVE(te, tqe_q); 989 m_freem(te->tqe_m); 990 FREE(te, M_TSEGQ); 991 tcp_reass_qsize--; 992 } 993 } 994 } 995 } 996 997 #ifdef SMP 998 struct netmsg_tcp_drain { 999 struct lwkt_msg nm_lmsg; 1000 struct inpcbhead *nm_head; 1001 }; 1002 1003 static int 1004 tcp_drain_handler(lwkt_msg_t lmsg) 1005 { 1006 struct netmsg_tcp_drain *nm = (void *)lmsg; 1007 1008 tcp_drain_oncpu(nm->nm_head); 1009 lwkt_replymsg(lmsg, 0); 1010 return(EASYNC); 1011 } 1012 #endif 1013 1014 void 1015 tcp_drain() 1016 { 1017 #ifdef SMP 1018 int cpu; 1019 #endif 1020 1021 if (!do_tcpdrain) 1022 return; 1023 1024 /* 1025 * Walk the tcpbs, if existing, and flush the reassembly queue, 1026 * if there is one... 1027 * XXX: The "Net/3" implementation doesn't imply that the TCP 1028 * reassembly queue should be flushed, but in a situation 1029 * where we're really low on mbufs, this is potentially 1030 * useful. 1031 */ 1032 #ifdef SMP 1033 for (cpu = 0; cpu < ncpus2; cpu++) { 1034 struct netmsg_tcp_drain *msg; 1035 1036 if (cpu == mycpu->gd_cpuid) { 1037 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead); 1038 } else { 1039 msg = malloc(sizeof(struct netmsg_tcp_drain), 1040 M_LWKTMSG, M_NOWAIT); 1041 if (msg == NULL) 1042 continue; 1043 lwkt_initmsg(&msg->nm_lmsg, &netisr_afree_rport, 0, 1044 lwkt_cmd_func(tcp_drain_handler), 1045 lwkt_cmd_op_none); 1046 msg->nm_head = &tcbinfo[cpu].pcblisthead; 1047 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg); 1048 } 1049 } 1050 #else 1051 tcp_drain_oncpu(&tcbinfo[0].pcblisthead); 1052 #endif 1053 } 1054 1055 /* 1056 * Notify a tcp user of an asynchronous error; 1057 * store error as soft error, but wake up user 1058 * (for now, won't do anything until can select for soft error). 1059 * 1060 * Do not wake up user since there currently is no mechanism for 1061 * reporting soft errors (yet - a kqueue filter may be added). 1062 */ 1063 static void 1064 tcp_notify(struct inpcb *inp, int error) 1065 { 1066 struct tcpcb *tp = intotcpcb(inp); 1067 1068 /* 1069 * Ignore some errors if we are hooked up. 1070 * If connection hasn't completed, has retransmitted several times, 1071 * and receives a second error, give up now. This is better 1072 * than waiting a long time to establish a connection that 1073 * can never complete. 1074 */ 1075 if (tp->t_state == TCPS_ESTABLISHED && 1076 (error == EHOSTUNREACH || error == ENETUNREACH || 1077 error == EHOSTDOWN)) { 1078 return; 1079 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1080 tp->t_softerror) 1081 tcp_drop(tp, error); 1082 else 1083 tp->t_softerror = error; 1084 #if 0 1085 wakeup(&so->so_timeo); 1086 sorwakeup(so); 1087 sowwakeup(so); 1088 #endif 1089 } 1090 1091 static int 1092 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1093 { 1094 int error, i, n; 1095 struct inpcb *marker; 1096 struct inpcb *inp; 1097 inp_gen_t gencnt; 1098 globaldata_t gd; 1099 int origcpu, ccpu; 1100 1101 error = 0; 1102 n = 0; 1103 1104 /* 1105 * The process of preparing the TCB list is too time-consuming and 1106 * resource-intensive to repeat twice on every request. 1107 */ 1108 if (req->oldptr == NULL) { 1109 for (ccpu = 0; ccpu < ncpus; ++ccpu) { 1110 gd = globaldata_find(ccpu); 1111 n += tcbinfo[gd->gd_cpuid].ipi_count; 1112 } 1113 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1114 return (0); 1115 } 1116 1117 if (req->newptr != NULL) 1118 return (EPERM); 1119 1120 marker = malloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1121 marker->inp_flags |= INP_PLACEMARKER; 1122 1123 /* 1124 * OK, now we're committed to doing something. Run the inpcb list 1125 * for each cpu in the system and construct the output. Use a 1126 * list placemarker to deal with list changes occuring during 1127 * copyout blockages (but otherwise depend on being on the correct 1128 * cpu to avoid races). 1129 */ 1130 origcpu = mycpu->gd_cpuid; 1131 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) { 1132 globaldata_t rgd; 1133 caddr_t inp_ppcb; 1134 struct xtcpcb xt; 1135 int cpu_id; 1136 1137 cpu_id = (origcpu + ccpu) % ncpus; 1138 if ((smp_active_mask & (1 << cpu_id)) == 0) 1139 continue; 1140 rgd = globaldata_find(cpu_id); 1141 lwkt_setcpu_self(rgd); 1142 1143 /* indicate change of CPU */ 1144 cpu_mb1(); 1145 1146 gencnt = tcbinfo[cpu_id].ipi_gencnt; 1147 n = tcbinfo[cpu_id].ipi_count; 1148 1149 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list); 1150 i = 0; 1151 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1152 /* 1153 * process a snapshot of pcbs, ignoring placemarkers 1154 * and using our own to allow SYSCTL_OUT to block. 1155 */ 1156 LIST_REMOVE(marker, inp_list); 1157 LIST_INSERT_AFTER(inp, marker, inp_list); 1158 1159 if (inp->inp_flags & INP_PLACEMARKER) 1160 continue; 1161 if (inp->inp_gencnt > gencnt) 1162 continue; 1163 if (prison_xinpcb(req->td, inp)) 1164 continue; 1165 1166 xt.xt_len = sizeof xt; 1167 bcopy(inp, &xt.xt_inp, sizeof *inp); 1168 inp_ppcb = inp->inp_ppcb; 1169 if (inp_ppcb != NULL) 1170 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1171 else 1172 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1173 if (inp->inp_socket) 1174 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1175 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1176 break; 1177 ++i; 1178 } 1179 LIST_REMOVE(marker, inp_list); 1180 if (error == 0 && i < n) { 1181 bzero(&xt, sizeof(xt)); 1182 xt.xt_len = sizeof(xt); 1183 while (i < n) { 1184 error = SYSCTL_OUT(req, &xt, sizeof xt); 1185 if (error) 1186 break; 1187 ++i; 1188 } 1189 } 1190 } 1191 1192 /* 1193 * Make sure we are on the same cpu we were on originally, since 1194 * higher level callers expect this. Also don't pollute caches with 1195 * migrated userland data by (eventually) returning to userland 1196 * on a different cpu. 1197 */ 1198 lwkt_setcpu_self(globaldata_find(origcpu)); 1199 free(marker, M_TEMP); 1200 return (error); 1201 } 1202 1203 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1204 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1205 1206 static int 1207 tcp_getcred(SYSCTL_HANDLER_ARGS) 1208 { 1209 struct sockaddr_in addrs[2]; 1210 struct inpcb *inp; 1211 int cpu; 1212 int error, s; 1213 1214 error = suser(req->td); 1215 if (error != 0) 1216 return (error); 1217 error = SYSCTL_IN(req, addrs, sizeof addrs); 1218 if (error != 0) 1219 return (error); 1220 s = splnet(); 1221 1222 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1223 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1224 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1225 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1226 if (inp == NULL || inp->inp_socket == NULL) { 1227 error = ENOENT; 1228 goto out; 1229 } 1230 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1231 out: 1232 splx(s); 1233 return (error); 1234 } 1235 1236 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1237 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1238 1239 #ifdef INET6 1240 static int 1241 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1242 { 1243 struct sockaddr_in6 addrs[2]; 1244 struct inpcb *inp; 1245 int error, s; 1246 boolean_t mapped = FALSE; 1247 1248 error = suser(req->td); 1249 if (error != 0) 1250 return (error); 1251 error = SYSCTL_IN(req, addrs, sizeof addrs); 1252 if (error != 0) 1253 return (error); 1254 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { 1255 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) 1256 mapped = TRUE; 1257 else 1258 return (EINVAL); 1259 } 1260 s = splnet(); 1261 if (mapped) { 1262 inp = in_pcblookup_hash(&tcbinfo[0], 1263 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], 1264 addrs[1].sin6_port, 1265 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], 1266 addrs[0].sin6_port, 1267 0, NULL); 1268 } else { 1269 inp = in6_pcblookup_hash(&tcbinfo[0], 1270 &addrs[1].sin6_addr, addrs[1].sin6_port, 1271 &addrs[0].sin6_addr, addrs[0].sin6_port, 1272 0, NULL); 1273 } 1274 if (inp == NULL || inp->inp_socket == NULL) { 1275 error = ENOENT; 1276 goto out; 1277 } 1278 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1279 out: 1280 splx(s); 1281 return (error); 1282 } 1283 1284 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1285 0, 0, 1286 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1287 #endif 1288 1289 void 1290 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip) 1291 { 1292 struct ip *ip = vip; 1293 struct tcphdr *th; 1294 struct in_addr faddr; 1295 struct inpcb *inp; 1296 struct tcpcb *tp; 1297 void (*notify)(struct inpcb *, int) = tcp_notify; 1298 tcp_seq icmp_seq; 1299 int cpu; 1300 int s; 1301 1302 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1303 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1304 return; 1305 1306 if (cmd == PRC_QUENCH) 1307 notify = tcp_quench; 1308 else if (icmp_may_rst && 1309 (cmd == PRC_UNREACH_ADMIN_PROHIB || cmd == PRC_UNREACH_PORT || 1310 cmd == PRC_TIMXCEED_INTRANS) && 1311 ip != NULL) 1312 notify = tcp_drop_syn_sent; 1313 else if (cmd == PRC_MSGSIZE) 1314 notify = tcp_mtudisc; 1315 else if (PRC_IS_REDIRECT(cmd)) { 1316 ip = NULL; 1317 notify = in_rtchange; 1318 } else if (cmd == PRC_HOSTDEAD) 1319 ip = NULL; 1320 else if ((unsigned)cmd > PRC_NCMDS || inetctlerrmap[cmd] == 0) 1321 return; 1322 if (ip != NULL) { 1323 s = splnet(); 1324 th = (struct tcphdr *)((caddr_t)ip + 1325 (IP_VHL_HL(ip->ip_vhl) << 2)); 1326 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport, 1327 ip->ip_src.s_addr, th->th_sport); 1328 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport, 1329 ip->ip_src, th->th_sport, 0, NULL); 1330 if ((inp != NULL) && (inp->inp_socket != NULL)) { 1331 icmp_seq = htonl(th->th_seq); 1332 tp = intotcpcb(inp); 1333 if (SEQ_GEQ(icmp_seq, tp->snd_una) && 1334 SEQ_LT(icmp_seq, tp->snd_max)) 1335 (*notify)(inp, inetctlerrmap[cmd]); 1336 } else { 1337 struct in_conninfo inc; 1338 1339 inc.inc_fport = th->th_dport; 1340 inc.inc_lport = th->th_sport; 1341 inc.inc_faddr = faddr; 1342 inc.inc_laddr = ip->ip_src; 1343 #ifdef INET6 1344 inc.inc_isipv6 = 0; 1345 #endif 1346 syncache_unreach(&inc, th); 1347 } 1348 splx(s); 1349 } else { 1350 for (cpu = 0; cpu < ncpus2; cpu++) { 1351 in_pcbnotifyall(&tcbinfo[cpu].pcblisthead, faddr, 1352 inetctlerrmap[cmd], notify); 1353 } 1354 } 1355 } 1356 1357 #ifdef INET6 1358 void 1359 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d) 1360 { 1361 struct tcphdr th; 1362 void (*notify) (struct inpcb *, int) = tcp_notify; 1363 struct ip6_hdr *ip6; 1364 struct mbuf *m; 1365 struct ip6ctlparam *ip6cp = NULL; 1366 const struct sockaddr_in6 *sa6_src = NULL; 1367 int off; 1368 struct tcp_portonly { 1369 u_int16_t th_sport; 1370 u_int16_t th_dport; 1371 } *thp; 1372 1373 if (sa->sa_family != AF_INET6 || 1374 sa->sa_len != sizeof(struct sockaddr_in6)) 1375 return; 1376 1377 if (cmd == PRC_QUENCH) 1378 notify = tcp_quench; 1379 else if (cmd == PRC_MSGSIZE) 1380 notify = tcp_mtudisc; 1381 else if (!PRC_IS_REDIRECT(cmd) && 1382 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) 1383 return; 1384 1385 /* if the parameter is from icmp6, decode it. */ 1386 if (d != NULL) { 1387 ip6cp = (struct ip6ctlparam *)d; 1388 m = ip6cp->ip6c_m; 1389 ip6 = ip6cp->ip6c_ip6; 1390 off = ip6cp->ip6c_off; 1391 sa6_src = ip6cp->ip6c_src; 1392 } else { 1393 m = NULL; 1394 ip6 = NULL; 1395 off = 0; /* fool gcc */ 1396 sa6_src = &sa6_any; 1397 } 1398 1399 if (ip6 != NULL) { 1400 struct in_conninfo inc; 1401 /* 1402 * XXX: We assume that when IPV6 is non NULL, 1403 * M and OFF are valid. 1404 */ 1405 1406 /* check if we can safely examine src and dst ports */ 1407 if (m->m_pkthdr.len < off + sizeof *thp) 1408 return; 1409 1410 bzero(&th, sizeof th); 1411 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1412 1413 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport, 1414 (struct sockaddr *)ip6cp->ip6c_src, 1415 th.th_sport, cmd, notify); 1416 1417 inc.inc_fport = th.th_dport; 1418 inc.inc_lport = th.th_sport; 1419 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1420 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1421 inc.inc_isipv6 = 1; 1422 syncache_unreach(&inc, &th); 1423 } else 1424 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0, 1425 (const struct sockaddr *)sa6_src, 0, cmd, notify); 1426 } 1427 #endif 1428 1429 /* 1430 * Following is where TCP initial sequence number generation occurs. 1431 * 1432 * There are two places where we must use initial sequence numbers: 1433 * 1. In SYN-ACK packets. 1434 * 2. In SYN packets. 1435 * 1436 * All ISNs for SYN-ACK packets are generated by the syncache. See 1437 * tcp_syncache.c for details. 1438 * 1439 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1440 * depends on this property. In addition, these ISNs should be 1441 * unguessable so as to prevent connection hijacking. To satisfy 1442 * the requirements of this situation, the algorithm outlined in 1443 * RFC 1948 is used to generate sequence numbers. 1444 * 1445 * Implementation details: 1446 * 1447 * Time is based off the system timer, and is corrected so that it 1448 * increases by one megabyte per second. This allows for proper 1449 * recycling on high speed LANs while still leaving over an hour 1450 * before rollover. 1451 * 1452 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1453 * between seeding of isn_secret. This is normally set to zero, 1454 * as reseeding should not be necessary. 1455 * 1456 */ 1457 1458 #define ISN_BYTES_PER_SECOND 1048576 1459 1460 u_char isn_secret[32]; 1461 int isn_last_reseed; 1462 MD5_CTX isn_ctx; 1463 1464 tcp_seq 1465 tcp_new_isn(struct tcpcb *tp) 1466 { 1467 u_int32_t md5_buffer[4]; 1468 tcp_seq new_isn; 1469 1470 /* Seed if this is the first use, reseed if requested. */ 1471 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1472 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1473 < (u_int)ticks))) { 1474 read_random_unlimited(&isn_secret, sizeof isn_secret); 1475 isn_last_reseed = ticks; 1476 } 1477 1478 /* Compute the md5 hash and return the ISN. */ 1479 MD5Init(&isn_ctx); 1480 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1481 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1482 #ifdef INET6 1483 if (tp->t_inpcb->inp_vflag & INP_IPV6) { 1484 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1485 sizeof(struct in6_addr)); 1486 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1487 sizeof(struct in6_addr)); 1488 } else 1489 #endif 1490 { 1491 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1492 sizeof(struct in_addr)); 1493 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1494 sizeof(struct in_addr)); 1495 } 1496 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1497 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1498 new_isn = (tcp_seq) md5_buffer[0]; 1499 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1500 return (new_isn); 1501 } 1502 1503 /* 1504 * When a source quench is received, close congestion window 1505 * to one segment. We will gradually open it again as we proceed. 1506 */ 1507 void 1508 tcp_quench(struct inpcb *inp, int errno) 1509 { 1510 struct tcpcb *tp = intotcpcb(inp); 1511 1512 if (tp != NULL) 1513 tp->snd_cwnd = tp->t_maxseg; 1514 } 1515 1516 /* 1517 * When a specific ICMP unreachable message is received and the 1518 * connection state is SYN-SENT, drop the connection. This behavior 1519 * is controlled by the icmp_may_rst sysctl. 1520 */ 1521 void 1522 tcp_drop_syn_sent(struct inpcb *inp, int errno) 1523 { 1524 struct tcpcb *tp = intotcpcb(inp); 1525 1526 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT)) 1527 tcp_drop(tp, errno); 1528 } 1529 1530 /* 1531 * When `need fragmentation' ICMP is received, update our idea of the MSS 1532 * based on the new value in the route. Also nudge TCP to send something, 1533 * since we know the packet we just sent was dropped. 1534 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1535 */ 1536 void 1537 tcp_mtudisc(struct inpcb *inp, int errno) 1538 { 1539 struct tcpcb *tp = intotcpcb(inp); 1540 struct rtentry *rt; 1541 struct rmxp_tao *taop; 1542 struct socket *so = inp->inp_socket; 1543 int offered; 1544 int mss; 1545 #ifdef INET6 1546 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0); 1547 #else 1548 const boolean_t isipv6 = FALSE; 1549 #endif 1550 1551 if (tp != NULL) { 1552 if (isipv6) 1553 rt = tcp_rtlookup6(&inp->inp_inc); 1554 else 1555 rt = tcp_rtlookup(&inp->inp_inc); 1556 if (rt == NULL || rt->rt_rmx.rmx_mtu == 0) { 1557 tp->t_maxopd = tp->t_maxseg = 1558 isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 1559 return; 1560 } 1561 taop = rmx_taop(rt->rt_rmx); 1562 offered = taop->tao_mssopt; 1563 mss = rt->rt_rmx.rmx_mtu - 1564 (isipv6 ? 1565 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1566 sizeof(struct tcpiphdr)); 1567 1568 if (offered != 0) 1569 mss = min(mss, offered); 1570 /* 1571 * XXX - The above conditional probably violates the TCP 1572 * spec. The problem is that, since we don't know the 1573 * other end's MSS, we are supposed to use a conservative 1574 * default. But, if we do that, then MTU discovery will 1575 * never actually take place, because the conservative 1576 * default is much less than the MTUs typically seen 1577 * on the Internet today. For the moment, we'll sweep 1578 * this under the carpet. 1579 * 1580 * The conservative default might not actually be a problem 1581 * if the only case this occurs is when sending an initial 1582 * SYN with options and data to a host we've never talked 1583 * to before. Then, they will reply with an MSS value which 1584 * will get recorded and the new parameters should get 1585 * recomputed. For Further Study. 1586 */ 1587 if (tp->t_maxopd <= mss) 1588 return; 1589 tp->t_maxopd = mss; 1590 1591 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP && 1592 (tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP) 1593 mss -= TCPOLEN_TSTAMP_APPA; 1594 if ((tp->t_flags & (TF_REQ_CC | TF_NOOPT)) == TF_REQ_CC && 1595 (tp->t_flags & TF_RCVD_CC) == TF_RCVD_CC) 1596 mss -= TCPOLEN_CC_APPA; 1597 #if (MCLBYTES & (MCLBYTES - 1)) == 0 1598 if (mss > MCLBYTES) 1599 mss &= ~(MCLBYTES - 1); 1600 #else 1601 if (mss > MCLBYTES) 1602 mss = mss / MCLBYTES * MCLBYTES; 1603 #endif 1604 if (so->so_snd.sb_hiwat < mss) 1605 mss = so->so_snd.sb_hiwat; 1606 1607 tp->t_maxseg = mss; 1608 1609 tcpstat.tcps_mturesent++; 1610 tp->t_rtttime = 0; 1611 tp->snd_nxt = tp->snd_una; 1612 tcp_output(tp); 1613 } 1614 } 1615 1616 /* 1617 * Look-up the routing entry to the peer of this inpcb. If no route 1618 * is found and it cannot be allocated the return NULL. This routine 1619 * is called by TCP routines that access the rmx structure and by tcp_mss 1620 * to get the interface MTU. 1621 */ 1622 struct rtentry * 1623 tcp_rtlookup(struct in_conninfo *inc) 1624 { 1625 struct route *ro = &inc->inc_route; 1626 1627 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1628 /* No route yet, so try to acquire one */ 1629 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1630 /* 1631 * unused portions of the structure MUST be zero'd 1632 * out because rtalloc() treats it as opaque data 1633 */ 1634 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1635 ro->ro_dst.sa_family = AF_INET; 1636 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1637 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1638 inc->inc_faddr; 1639 rtalloc(ro); 1640 } 1641 } 1642 return (ro->ro_rt); 1643 } 1644 1645 #ifdef INET6 1646 struct rtentry * 1647 tcp_rtlookup6(struct in_conninfo *inc) 1648 { 1649 struct route_in6 *ro6 = &inc->inc6_route; 1650 1651 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1652 /* No route yet, so try to acquire one */ 1653 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1654 /* 1655 * unused portions of the structure MUST be zero'd 1656 * out because rtalloc() treats it as opaque data 1657 */ 1658 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1659 ro6->ro_dst.sin6_family = AF_INET6; 1660 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1661 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1662 rtalloc((struct route *)ro6); 1663 } 1664 } 1665 return (ro6->ro_rt); 1666 } 1667 #endif 1668 1669 #ifdef IPSEC 1670 /* compute ESP/AH header size for TCP, including outer IP header. */ 1671 size_t 1672 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1673 { 1674 struct inpcb *inp; 1675 struct mbuf *m; 1676 size_t hdrsiz; 1677 struct ip *ip; 1678 struct tcphdr *th; 1679 1680 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1681 return (0); 1682 MGETHDR(m, MB_DONTWAIT, MT_DATA); 1683 if (!m) 1684 return (0); 1685 1686 #ifdef INET6 1687 if (inp->inp_vflag & INP_IPV6) { 1688 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1689 1690 th = (struct tcphdr *)(ip6 + 1); 1691 m->m_pkthdr.len = m->m_len = 1692 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1693 tcp_fillheaders(tp, ip6, th); 1694 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1695 } else 1696 #endif 1697 { 1698 ip = mtod(m, struct ip *); 1699 th = (struct tcphdr *)(ip + 1); 1700 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1701 tcp_fillheaders(tp, ip, th); 1702 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1703 } 1704 1705 m_free(m); 1706 return (hdrsiz); 1707 } 1708 #endif 1709 1710 /* 1711 * Return a pointer to the cached information about the remote host. 1712 * The cached information is stored in the protocol specific part of 1713 * the route metrics. 1714 */ 1715 struct rmxp_tao * 1716 tcp_gettaocache(struct in_conninfo *inc) 1717 { 1718 struct rtentry *rt; 1719 1720 #ifdef INET6 1721 if (inc->inc_isipv6) 1722 rt = tcp_rtlookup6(inc); 1723 else 1724 #endif 1725 rt = tcp_rtlookup(inc); 1726 1727 /* Make sure this is a host route and is up. */ 1728 if (rt == NULL || 1729 (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST)) 1730 return (NULL); 1731 1732 return (rmx_taop(rt->rt_rmx)); 1733 } 1734 1735 /* 1736 * Clear all the TAO cache entries, called from tcp_init. 1737 * 1738 * XXX 1739 * This routine is just an empty one, because we assume that the routing 1740 * routing tables are initialized at the same time when TCP, so there is 1741 * nothing in the cache left over. 1742 */ 1743 static void 1744 tcp_cleartaocache() 1745 { 1746 } 1747 1748 /* 1749 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1750 * 1751 * This code attempts to calculate the bandwidth-delay product as a 1752 * means of determining the optimal window size to maximize bandwidth, 1753 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1754 * routers. This code also does a fairly good job keeping RTTs in check 1755 * across slow links like modems. We implement an algorithm which is very 1756 * similar (but not meant to be) TCP/Vegas. The code operates on the 1757 * transmitter side of a TCP connection and so only effects the transmit 1758 * side of the connection. 1759 * 1760 * BACKGROUND: TCP makes no provision for the management of buffer space 1761 * at the end points or at the intermediate routers and switches. A TCP 1762 * stream, whether using NewReno or not, will eventually buffer as 1763 * many packets as it is able and the only reason this typically works is 1764 * due to the fairly small default buffers made available for a connection 1765 * (typicaly 16K or 32K). As machines use larger windows and/or window 1766 * scaling it is now fairly easy for even a single TCP connection to blow-out 1767 * all available buffer space not only on the local interface, but on 1768 * intermediate routers and switches as well. NewReno makes a misguided 1769 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1770 * then backing off, then steadily increasing the window again until another 1771 * failure occurs, ad-infinitum. This results in terrible oscillation that 1772 * is only made worse as network loads increase and the idea of intentionally 1773 * blowing out network buffers is, frankly, a terrible way to manage network 1774 * resources. 1775 * 1776 * It is far better to limit the transmit window prior to the failure 1777 * condition being achieved. There are two general ways to do this: First 1778 * you can 'scan' through different transmit window sizes and locate the 1779 * point where the RTT stops increasing, indicating that you have filled the 1780 * pipe, then scan backwards until you note that RTT stops decreasing, then 1781 * repeat ad-infinitum. This method works in principle but has severe 1782 * implementation issues due to RTT variances, timer granularity, and 1783 * instability in the algorithm which can lead to many false positives and 1784 * create oscillations as well as interact badly with other TCP streams 1785 * implementing the same algorithm. 1786 * 1787 * The second method is to limit the window to the bandwidth delay product 1788 * of the link. This is the method we implement. RTT variances and our 1789 * own manipulation of the congestion window, bwnd, can potentially 1790 * destabilize the algorithm. For this reason we have to stabilize the 1791 * elements used to calculate the window. We do this by using the minimum 1792 * observed RTT, the long term average of the observed bandwidth, and 1793 * by adding two segments worth of slop. It isn't perfect but it is able 1794 * to react to changing conditions and gives us a very stable basis on 1795 * which to extend the algorithm. 1796 */ 1797 void 1798 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1799 { 1800 u_long bw; 1801 u_long bwnd; 1802 int save_ticks; 1803 int delta_ticks; 1804 1805 /* 1806 * If inflight_enable is disabled in the middle of a tcp connection, 1807 * make sure snd_bwnd is effectively disabled. 1808 */ 1809 if (!tcp_inflight_enable) { 1810 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1811 tp->snd_bandwidth = 0; 1812 return; 1813 } 1814 1815 /* 1816 * Validate the delta time. If a connection is new or has been idle 1817 * a long time we have to reset the bandwidth calculator. 1818 */ 1819 save_ticks = ticks; 1820 delta_ticks = save_ticks - tp->t_bw_rtttime; 1821 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1822 tp->t_bw_rtttime = ticks; 1823 tp->t_bw_rtseq = ack_seq; 1824 if (tp->snd_bandwidth == 0) 1825 tp->snd_bandwidth = tcp_inflight_min; 1826 return; 1827 } 1828 if (delta_ticks == 0) 1829 return; 1830 1831 /* 1832 * Sanity check, plus ignore pure window update acks. 1833 */ 1834 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1835 return; 1836 1837 /* 1838 * Figure out the bandwidth. Due to the tick granularity this 1839 * is a very rough number and it MUST be averaged over a fairly 1840 * long period of time. XXX we need to take into account a link 1841 * that is not using all available bandwidth, but for now our 1842 * slop will ramp us up if this case occurs and the bandwidth later 1843 * increases. 1844 */ 1845 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 1846 tp->t_bw_rtttime = save_ticks; 1847 tp->t_bw_rtseq = ack_seq; 1848 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4; 1849 1850 tp->snd_bandwidth = bw; 1851 1852 /* 1853 * Calculate the semi-static bandwidth delay product, plus two maximal 1854 * segments. The additional slop puts us squarely in the sweet 1855 * spot and also handles the bandwidth run-up case. Without the 1856 * slop we could be locking ourselves into a lower bandwidth. 1857 * 1858 * Situations Handled: 1859 * (1) Prevents over-queueing of packets on LANs, especially on 1860 * high speed LANs, allowing larger TCP buffers to be 1861 * specified, and also does a good job preventing 1862 * over-queueing of packets over choke points like modems 1863 * (at least for the transmit side). 1864 * 1865 * (2) Is able to handle changing network loads (bandwidth 1866 * drops so bwnd drops, bandwidth increases so bwnd 1867 * increases). 1868 * 1869 * (3) Theoretically should stabilize in the face of multiple 1870 * connections implementing the same algorithm (this may need 1871 * a little work). 1872 * 1873 * (4) Stability value (defaults to 20 = 2 maximal packets) can 1874 * be adjusted with a sysctl but typically only needs to be on 1875 * very slow connections. A value no smaller then 5 should 1876 * be used, but only reduce this default if you have no other 1877 * choice. 1878 */ 1879 1880 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2) 1881 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 1882 tcp_inflight_stab * (int)tp->t_maxseg / 10; 1883 #undef USERTT 1884 1885 if (tcp_inflight_debug > 0) { 1886 static int ltime; 1887 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) { 1888 ltime = ticks; 1889 printf("%p bw %ld rttbest %d srtt %d bwnd %ld\n", 1890 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd); 1891 } 1892 } 1893 if ((long)bwnd < tcp_inflight_min) 1894 bwnd = tcp_inflight_min; 1895 if (bwnd > tcp_inflight_max) 1896 bwnd = tcp_inflight_max; 1897 if ((long)bwnd < tp->t_maxseg * 2) 1898 bwnd = tp->t_maxseg * 2; 1899 tp->snd_bwnd = bwnd; 1900 } 1901