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 1.40 2004/11/14 00:49:08 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, (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 tp->t_bw_rtttime = ticks; 696 /* 697 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 698 * because the socket may be bound to an IPv6 wildcard address, 699 * which may match an IPv4-mapped IPv6 address. 700 */ 701 inp->inp_ip_ttl = ip_defttl; 702 inp->inp_ppcb = (caddr_t)tp; 703 tcp_sack_tcpcb_init(tp); 704 return (tp); /* XXX */ 705 } 706 707 /* 708 * Drop a TCP connection, reporting the specified error. 709 * If connection is synchronized, then send a RST to peer. 710 */ 711 struct tcpcb * 712 tcp_drop(struct tcpcb *tp, int errno) 713 { 714 struct socket *so = tp->t_inpcb->inp_socket; 715 716 if (TCPS_HAVERCVDSYN(tp->t_state)) { 717 tp->t_state = TCPS_CLOSED; 718 (void) tcp_output(tp); 719 tcpstat.tcps_drops++; 720 } else 721 tcpstat.tcps_conndrops++; 722 if (errno == ETIMEDOUT && tp->t_softerror) 723 errno = tp->t_softerror; 724 so->so_error = errno; 725 return (tcp_close(tp)); 726 } 727 728 #ifdef SMP 729 730 struct netmsg_remwildcard { 731 struct lwkt_msg nm_lmsg; 732 struct inpcb *nm_inp; 733 struct inpcbinfo *nm_pcbinfo; 734 #if defined(INET6) 735 int nm_isinet6; 736 #else 737 int nm_unused01; 738 #endif 739 }; 740 741 /* 742 * Wildcard inpcb's on SMP boxes must be removed from all cpus before the 743 * inp can be detached. We do this by cycling through the cpus, ending up 744 * on the cpu controlling the inp last and then doing the disconnect. 745 */ 746 static int 747 in_pcbremwildcardhash_handler(struct lwkt_msg *msg0) 748 { 749 struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0; 750 int cpu; 751 752 cpu = msg->nm_pcbinfo->cpu; 753 754 if (cpu == msg->nm_inp->inp_pcbinfo->cpu) { 755 /* note: detach removes any wildcard hash entry */ 756 #ifdef INET6 757 if (msg->nm_isinet6) 758 in6_pcbdetach(msg->nm_inp); 759 else 760 #endif 761 in_pcbdetach(msg->nm_inp); 762 lwkt_replymsg(&msg->nm_lmsg, 0); 763 } else { 764 in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo); 765 cpu = (cpu + 1) % ncpus2; 766 msg->nm_pcbinfo = &tcbinfo[cpu]; 767 lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_lmsg); 768 } 769 return (EASYNC); 770 } 771 772 #endif 773 774 /* 775 * Close a TCP control block: 776 * discard all space held by the tcp 777 * discard internet protocol block 778 * wake up any sleepers 779 */ 780 struct tcpcb * 781 tcp_close(struct tcpcb *tp) 782 { 783 struct tseg_qent *q; 784 struct inpcb *inp = tp->t_inpcb; 785 struct socket *so = inp->inp_socket; 786 struct rtentry *rt; 787 boolean_t dosavessthresh; 788 #ifdef SMP 789 int cpu; 790 #endif 791 #ifdef INET6 792 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 793 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0); 794 #else 795 const boolean_t isipv6 = FALSE; 796 #endif 797 798 /* 799 * The tp is not instantly destroyed in the wildcard case. Setting 800 * the state to TCPS_TERMINATING will prevent the TCP stack from 801 * messing with it, though it should be noted that this change may 802 * not take effect on other cpus until we have chained the wildcard 803 * hash removal. 804 * 805 * XXX we currently depend on the BGL to synchronize the tp->t_state 806 * update and prevent other tcp protocol threads from accepting new 807 * connections on the listen socket we might be trying to close down. 808 */ 809 KKASSERT(tp->t_state != TCPS_TERMINATING); 810 tp->t_state = TCPS_TERMINATING; 811 812 /* 813 * Make sure that all of our timers are stopped before we 814 * delete the PCB. 815 */ 816 callout_stop(tp->tt_rexmt); 817 callout_stop(tp->tt_persist); 818 callout_stop(tp->tt_keep); 819 callout_stop(tp->tt_2msl); 820 callout_stop(tp->tt_delack); 821 822 if (tp->t_flags & TF_ONOUTPUTQ) { 823 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 824 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 825 tp->t_flags &= ~TF_ONOUTPUTQ; 826 } 827 828 /* 829 * If we got enough samples through the srtt filter, 830 * save the rtt and rttvar in the routing entry. 831 * 'Enough' is arbitrarily defined as the 16 samples. 832 * 16 samples is enough for the srtt filter to converge 833 * to within 5% of the correct value; fewer samples and 834 * we could save a very bogus rtt. 835 * 836 * Don't update the default route's characteristics and don't 837 * update anything that the user "locked". 838 */ 839 if (tp->t_rttupdated >= 16) { 840 u_long i = 0; 841 842 if (isipv6) { 843 struct sockaddr_in6 *sin6; 844 845 if ((rt = inp->in6p_route.ro_rt) == NULL) 846 goto no_valid_rt; 847 sin6 = (struct sockaddr_in6 *)rt_key(rt); 848 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 849 goto no_valid_rt; 850 } else 851 if ((rt = inp->inp_route.ro_rt) == NULL || 852 ((struct sockaddr_in *)rt_key(rt))-> 853 sin_addr.s_addr == INADDR_ANY) 854 goto no_valid_rt; 855 856 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 857 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 858 if (rt->rt_rmx.rmx_rtt && i) 859 /* 860 * filter this update to half the old & half 861 * the new values, converting scale. 862 * See route.h and tcp_var.h for a 863 * description of the scaling constants. 864 */ 865 rt->rt_rmx.rmx_rtt = 866 (rt->rt_rmx.rmx_rtt + i) / 2; 867 else 868 rt->rt_rmx.rmx_rtt = i; 869 tcpstat.tcps_cachedrtt++; 870 } 871 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 872 i = tp->t_rttvar * 873 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 874 if (rt->rt_rmx.rmx_rttvar && i) 875 rt->rt_rmx.rmx_rttvar = 876 (rt->rt_rmx.rmx_rttvar + i) / 2; 877 else 878 rt->rt_rmx.rmx_rttvar = i; 879 tcpstat.tcps_cachedrttvar++; 880 } 881 /* 882 * The old comment here said: 883 * update the pipelimit (ssthresh) if it has been updated 884 * already or if a pipesize was specified & the threshhold 885 * got below half the pipesize. I.e., wait for bad news 886 * before we start updating, then update on both good 887 * and bad news. 888 * 889 * But we want to save the ssthresh even if no pipesize is 890 * specified explicitly in the route, because such 891 * connections still have an implicit pipesize specified 892 * by the global tcp_sendspace. In the absence of a reliable 893 * way to calculate the pipesize, it will have to do. 894 */ 895 i = tp->snd_ssthresh; 896 if (rt->rt_rmx.rmx_sendpipe != 0) 897 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 898 else 899 dosavessthresh = (i < so->so_snd.sb_hiwat/2); 900 if (dosavessthresh || 901 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 902 (rt->rt_rmx.rmx_ssthresh != 0))) { 903 /* 904 * convert the limit from user data bytes to 905 * packets then to packet data bytes. 906 */ 907 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 908 if (i < 2) 909 i = 2; 910 i *= tp->t_maxseg + 911 (isipv6 ? 912 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 913 sizeof(struct tcpiphdr)); 914 if (rt->rt_rmx.rmx_ssthresh) 915 rt->rt_rmx.rmx_ssthresh = 916 (rt->rt_rmx.rmx_ssthresh + i) / 2; 917 else 918 rt->rt_rmx.rmx_ssthresh = i; 919 tcpstat.tcps_cachedssthresh++; 920 } 921 } 922 923 no_valid_rt: 924 /* free the reassembly queue, if any */ 925 while((q = LIST_FIRST(&tp->t_segq)) != NULL) { 926 LIST_REMOVE(q, tqe_q); 927 m_freem(q->tqe_m); 928 FREE(q, M_TSEGQ); 929 tcp_reass_qsize--; 930 } 931 /* throw away SACK blocks in scoreboard*/ 932 if (TCP_DO_SACK(tp)) 933 tcp_sack_cleanup(&tp->scb); 934 935 inp->inp_ppcb = NULL; 936 soisdisconnected(so); 937 /* 938 * Discard the inp. In the SMP case a wildcard inp's hash (created 939 * by a listen socket or an INADDR_ANY udp socket) is replicated 940 * for each protocol thread and must be removed in the context of 941 * that thread. This is accomplished by chaining the message 942 * through the cpus. 943 * 944 * If the inp is not wildcarded we simply detach, which will remove 945 * the any hashes still present for this inp. 946 */ 947 #ifdef SMP 948 if (inp->inp_flags & INP_WILDCARD_MP) { 949 struct netmsg_remwildcard *msg; 950 951 cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2; 952 msg = malloc(sizeof(struct netmsg_remwildcard), 953 M_LWKTMSG, M_INTWAIT); 954 lwkt_initmsg(&msg->nm_lmsg, &netisr_afree_rport, 0, 955 lwkt_cmd_func(in_pcbremwildcardhash_handler), 956 lwkt_cmd_op_none); 957 #ifdef INET6 958 msg->nm_isinet6 = isafinet6; 959 #endif 960 msg->nm_inp = inp; 961 msg->nm_pcbinfo = &tcbinfo[cpu]; 962 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg); 963 } else 964 #endif 965 { 966 /* note: detach removes any wildcard hash entry */ 967 #ifdef INET6 968 if (isafinet6) 969 in6_pcbdetach(inp); 970 else 971 #endif 972 in_pcbdetach(inp); 973 } 974 tcpstat.tcps_closed++; 975 return (NULL); 976 } 977 978 static __inline void 979 tcp_drain_oncpu(struct inpcbhead *head) 980 { 981 struct inpcb *inpb; 982 struct tcpcb *tcpb; 983 struct tseg_qent *te; 984 985 LIST_FOREACH(inpb, head, inp_list) { 986 if (inpb->inp_flags & INP_PLACEMARKER) 987 continue; 988 if ((tcpb = intotcpcb(inpb))) { 989 while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) { 990 LIST_REMOVE(te, tqe_q); 991 m_freem(te->tqe_m); 992 FREE(te, M_TSEGQ); 993 tcp_reass_qsize--; 994 } 995 } 996 } 997 } 998 999 #ifdef SMP 1000 struct netmsg_tcp_drain { 1001 struct lwkt_msg nm_lmsg; 1002 struct inpcbhead *nm_head; 1003 }; 1004 1005 static int 1006 tcp_drain_handler(lwkt_msg_t lmsg) 1007 { 1008 struct netmsg_tcp_drain *nm = (void *)lmsg; 1009 1010 tcp_drain_oncpu(nm->nm_head); 1011 lwkt_replymsg(lmsg, 0); 1012 return(EASYNC); 1013 } 1014 #endif 1015 1016 void 1017 tcp_drain() 1018 { 1019 #ifdef SMP 1020 int cpu; 1021 #endif 1022 1023 if (!do_tcpdrain) 1024 return; 1025 1026 /* 1027 * Walk the tcpbs, if existing, and flush the reassembly queue, 1028 * if there is one... 1029 * XXX: The "Net/3" implementation doesn't imply that the TCP 1030 * reassembly queue should be flushed, but in a situation 1031 * where we're really low on mbufs, this is potentially 1032 * useful. 1033 */ 1034 #ifdef SMP 1035 for (cpu = 0; cpu < ncpus2; cpu++) { 1036 struct netmsg_tcp_drain *msg; 1037 1038 if (cpu == mycpu->gd_cpuid) { 1039 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead); 1040 } else { 1041 msg = malloc(sizeof(struct netmsg_tcp_drain), 1042 M_LWKTMSG, M_NOWAIT); 1043 if (msg == NULL) 1044 continue; 1045 lwkt_initmsg(&msg->nm_lmsg, &netisr_afree_rport, 0, 1046 lwkt_cmd_func(tcp_drain_handler), 1047 lwkt_cmd_op_none); 1048 msg->nm_head = &tcbinfo[cpu].pcblisthead; 1049 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg); 1050 } 1051 } 1052 #else 1053 tcp_drain_oncpu(&tcbinfo[0].pcblisthead); 1054 #endif 1055 } 1056 1057 /* 1058 * Notify a tcp user of an asynchronous error; 1059 * store error as soft error, but wake up user 1060 * (for now, won't do anything until can select for soft error). 1061 * 1062 * Do not wake up user since there currently is no mechanism for 1063 * reporting soft errors (yet - a kqueue filter may be added). 1064 */ 1065 static void 1066 tcp_notify(struct inpcb *inp, int error) 1067 { 1068 struct tcpcb *tp = intotcpcb(inp); 1069 1070 /* 1071 * Ignore some errors if we are hooked up. 1072 * If connection hasn't completed, has retransmitted several times, 1073 * and receives a second error, give up now. This is better 1074 * than waiting a long time to establish a connection that 1075 * can never complete. 1076 */ 1077 if (tp->t_state == TCPS_ESTABLISHED && 1078 (error == EHOSTUNREACH || error == ENETUNREACH || 1079 error == EHOSTDOWN)) { 1080 return; 1081 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1082 tp->t_softerror) 1083 tcp_drop(tp, error); 1084 else 1085 tp->t_softerror = error; 1086 #if 0 1087 wakeup((caddr_t) &so->so_timeo); 1088 sorwakeup(so); 1089 sowwakeup(so); 1090 #endif 1091 } 1092 1093 static int 1094 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1095 { 1096 int error, i, n; 1097 struct inpcb *marker; 1098 struct inpcb *inp; 1099 inp_gen_t gencnt; 1100 struct xinpgen xig; 1101 globaldata_t gd; 1102 int origcpu, ccpu; 1103 1104 error = 0; 1105 n = 0; 1106 1107 /* 1108 * The process of preparing the TCB list is too time-consuming and 1109 * resource-intensive to repeat twice on every request. 1110 */ 1111 if (req->oldptr == NULL) { 1112 for (ccpu = 0; ccpu < ncpus; ++ccpu) { 1113 gd = globaldata_find(ccpu); 1114 n += tcbinfo[gd->gd_cpuid].ipi_count; 1115 } 1116 req->oldidx = 2 * ncpus * (sizeof xig) + 1117 (n + n/8) * sizeof(struct xtcpcb); 1118 return (0); 1119 } 1120 1121 if (req->newptr != NULL) 1122 return (EPERM); 1123 1124 marker = malloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1125 marker->inp_flags |= INP_PLACEMARKER; 1126 1127 /* 1128 * OK, now we're committed to doing something. Run the inpcb list 1129 * for each cpu in the system and construct the output. Use a 1130 * list placemarker to deal with list changes occuring during 1131 * copyout blockages (but otherwise depend on being on the correct 1132 * cpu to avoid races). 1133 */ 1134 origcpu = mycpu->gd_cpuid; 1135 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) { 1136 globaldata_t rgd; 1137 caddr_t inp_ppcb; 1138 struct xtcpcb xt; 1139 int cpu_id; 1140 1141 cpu_id = (origcpu + ccpu) % ncpus; 1142 if ((smp_active_mask & (1 << cpu_id)) == 0) 1143 continue; 1144 rgd = globaldata_find(cpu_id); 1145 lwkt_setcpu_self(rgd); 1146 1147 /* indicate change of CPU */ 1148 cpu_mb1(); 1149 1150 gencnt = tcbinfo[cpu_id].ipi_gencnt; 1151 n = tcbinfo[cpu_id].ipi_count; 1152 1153 xig.xig_len = sizeof xig; 1154 xig.xig_count = n; 1155 xig.xig_gen = gencnt; 1156 xig.xig_sogen = so_gencnt; 1157 xig.xig_cpu = cpu_id; 1158 error = SYSCTL_OUT(req, &xig, sizeof xig); 1159 if (error != 0) 1160 break; 1161 1162 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list); 1163 i = 0; 1164 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1165 /* 1166 * process a snapshot of pcbs, ignoring placemarkers 1167 * and using our own to allow SYSCTL_OUT to block. 1168 */ 1169 LIST_REMOVE(marker, inp_list); 1170 LIST_INSERT_AFTER(inp, marker, inp_list); 1171 1172 if (inp->inp_flags & INP_PLACEMARKER) 1173 continue; 1174 if (inp->inp_gencnt > gencnt) 1175 continue; 1176 if (prison_xinpcb(req->td, inp)) 1177 continue; 1178 1179 xt.xt_len = sizeof xt; 1180 bcopy(inp, &xt.xt_inp, sizeof *inp); 1181 inp_ppcb = inp->inp_ppcb; 1182 if (inp_ppcb != NULL) 1183 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1184 else 1185 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1186 if (inp->inp_socket) 1187 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1188 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1189 break; 1190 ++i; 1191 } 1192 LIST_REMOVE(marker, inp_list); 1193 if (error == 0 && i < n) { 1194 bzero(&xt, sizeof(xt)); 1195 xt.xt_len = sizeof(xt); 1196 while (i < n) { 1197 error = SYSCTL_OUT(req, &xt, sizeof (xt)); 1198 if (error) 1199 break; 1200 ++i; 1201 } 1202 } 1203 if (error == 0) { 1204 /* 1205 * Give the user an updated idea of our state. 1206 * If the generation differs from what we told 1207 * her before, she knows that something happened 1208 * while we were processing this request, and it 1209 * might be necessary to retry. 1210 */ 1211 xig.xig_gen = tcbinfo[cpu_id].ipi_gencnt; 1212 xig.xig_sogen = so_gencnt; 1213 xig.xig_count = tcbinfo[cpu_id].ipi_count; 1214 error = SYSCTL_OUT(req, &xig, sizeof xig); 1215 } 1216 } 1217 1218 /* 1219 * Make sure we are on the same cpu we were on originally, since 1220 * higher level callers expect this. Also don't pollute caches with 1221 * migrated userland data by (eventually) returning to userland 1222 * on a different cpu. 1223 */ 1224 lwkt_setcpu_self(globaldata_find(origcpu)); 1225 free(marker, M_TEMP); 1226 return (error); 1227 } 1228 1229 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1230 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1231 1232 static int 1233 tcp_getcred(SYSCTL_HANDLER_ARGS) 1234 { 1235 struct sockaddr_in addrs[2]; 1236 struct inpcb *inp; 1237 int cpu; 1238 int error, s; 1239 1240 error = suser(req->td); 1241 if (error != 0) 1242 return (error); 1243 error = SYSCTL_IN(req, addrs, sizeof addrs); 1244 if (error != 0) 1245 return (error); 1246 s = splnet(); 1247 1248 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1249 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1250 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1251 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1252 if (inp == NULL || inp->inp_socket == NULL) { 1253 error = ENOENT; 1254 goto out; 1255 } 1256 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1257 out: 1258 splx(s); 1259 return (error); 1260 } 1261 1262 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1263 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1264 1265 #ifdef INET6 1266 static int 1267 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1268 { 1269 struct sockaddr_in6 addrs[2]; 1270 struct inpcb *inp; 1271 int error, s; 1272 boolean_t mapped = FALSE; 1273 1274 error = suser(req->td); 1275 if (error != 0) 1276 return (error); 1277 error = SYSCTL_IN(req, addrs, sizeof addrs); 1278 if (error != 0) 1279 return (error); 1280 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { 1281 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) 1282 mapped = TRUE; 1283 else 1284 return (EINVAL); 1285 } 1286 s = splnet(); 1287 if (mapped) { 1288 inp = in_pcblookup_hash(&tcbinfo[0], 1289 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], 1290 addrs[1].sin6_port, 1291 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], 1292 addrs[0].sin6_port, 1293 0, NULL); 1294 } else { 1295 inp = in6_pcblookup_hash(&tcbinfo[0], 1296 &addrs[1].sin6_addr, addrs[1].sin6_port, 1297 &addrs[0].sin6_addr, addrs[0].sin6_port, 1298 0, NULL); 1299 } 1300 if (inp == NULL || inp->inp_socket == NULL) { 1301 error = ENOENT; 1302 goto out; 1303 } 1304 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1305 out: 1306 splx(s); 1307 return (error); 1308 } 1309 1310 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1311 0, 0, 1312 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1313 #endif 1314 1315 void 1316 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip) 1317 { 1318 struct ip *ip = vip; 1319 struct tcphdr *th; 1320 struct in_addr faddr; 1321 struct inpcb *inp; 1322 struct tcpcb *tp; 1323 void (*notify)(struct inpcb *, int) = tcp_notify; 1324 tcp_seq icmp_seq; 1325 int cpu; 1326 int s; 1327 1328 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1329 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1330 return; 1331 1332 if (cmd == PRC_QUENCH) 1333 notify = tcp_quench; 1334 else if (icmp_may_rst && 1335 (cmd == PRC_UNREACH_ADMIN_PROHIB || cmd == PRC_UNREACH_PORT || 1336 cmd == PRC_TIMXCEED_INTRANS) && 1337 ip != NULL) 1338 notify = tcp_drop_syn_sent; 1339 else if (cmd == PRC_MSGSIZE) 1340 notify = tcp_mtudisc; 1341 else if (PRC_IS_REDIRECT(cmd)) { 1342 ip = NULL; 1343 notify = in_rtchange; 1344 } else if (cmd == PRC_HOSTDEAD) 1345 ip = NULL; 1346 else if ((unsigned)cmd > PRC_NCMDS || inetctlerrmap[cmd] == 0) 1347 return; 1348 if (ip != NULL) { 1349 s = splnet(); 1350 th = (struct tcphdr *)((caddr_t)ip + 1351 (IP_VHL_HL(ip->ip_vhl) << 2)); 1352 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport, 1353 ip->ip_src.s_addr, th->th_sport); 1354 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport, 1355 ip->ip_src, th->th_sport, 0, NULL); 1356 if ((inp != NULL) && (inp->inp_socket != NULL)) { 1357 icmp_seq = htonl(th->th_seq); 1358 tp = intotcpcb(inp); 1359 if (SEQ_GEQ(icmp_seq, tp->snd_una) && 1360 SEQ_LT(icmp_seq, tp->snd_max)) 1361 (*notify)(inp, inetctlerrmap[cmd]); 1362 } else { 1363 struct in_conninfo inc; 1364 1365 inc.inc_fport = th->th_dport; 1366 inc.inc_lport = th->th_sport; 1367 inc.inc_faddr = faddr; 1368 inc.inc_laddr = ip->ip_src; 1369 #ifdef INET6 1370 inc.inc_isipv6 = 0; 1371 #endif 1372 syncache_unreach(&inc, th); 1373 } 1374 splx(s); 1375 } else { 1376 for (cpu = 0; cpu < ncpus2; cpu++) { 1377 in_pcbnotifyall(&tcbinfo[cpu].pcblisthead, faddr, 1378 inetctlerrmap[cmd], notify); 1379 } 1380 } 1381 } 1382 1383 #ifdef INET6 1384 void 1385 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d) 1386 { 1387 struct tcphdr th; 1388 void (*notify) (struct inpcb *, int) = tcp_notify; 1389 struct ip6_hdr *ip6; 1390 struct mbuf *m; 1391 struct ip6ctlparam *ip6cp = NULL; 1392 const struct sockaddr_in6 *sa6_src = NULL; 1393 int off; 1394 struct tcp_portonly { 1395 u_int16_t th_sport; 1396 u_int16_t th_dport; 1397 } *thp; 1398 1399 if (sa->sa_family != AF_INET6 || 1400 sa->sa_len != sizeof(struct sockaddr_in6)) 1401 return; 1402 1403 if (cmd == PRC_QUENCH) 1404 notify = tcp_quench; 1405 else if (cmd == PRC_MSGSIZE) 1406 notify = tcp_mtudisc; 1407 else if (!PRC_IS_REDIRECT(cmd) && 1408 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) 1409 return; 1410 1411 /* if the parameter is from icmp6, decode it. */ 1412 if (d != NULL) { 1413 ip6cp = (struct ip6ctlparam *)d; 1414 m = ip6cp->ip6c_m; 1415 ip6 = ip6cp->ip6c_ip6; 1416 off = ip6cp->ip6c_off; 1417 sa6_src = ip6cp->ip6c_src; 1418 } else { 1419 m = NULL; 1420 ip6 = NULL; 1421 off = 0; /* fool gcc */ 1422 sa6_src = &sa6_any; 1423 } 1424 1425 if (ip6 != NULL) { 1426 struct in_conninfo inc; 1427 /* 1428 * XXX: We assume that when IPV6 is non NULL, 1429 * M and OFF are valid. 1430 */ 1431 1432 /* check if we can safely examine src and dst ports */ 1433 if (m->m_pkthdr.len < off + sizeof *thp) 1434 return; 1435 1436 bzero(&th, sizeof th); 1437 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1438 1439 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport, 1440 (struct sockaddr *)ip6cp->ip6c_src, 1441 th.th_sport, cmd, notify); 1442 1443 inc.inc_fport = th.th_dport; 1444 inc.inc_lport = th.th_sport; 1445 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1446 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1447 inc.inc_isipv6 = 1; 1448 syncache_unreach(&inc, &th); 1449 } else 1450 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0, 1451 (const struct sockaddr *)sa6_src, 0, cmd, notify); 1452 } 1453 #endif 1454 1455 /* 1456 * Following is where TCP initial sequence number generation occurs. 1457 * 1458 * There are two places where we must use initial sequence numbers: 1459 * 1. In SYN-ACK packets. 1460 * 2. In SYN packets. 1461 * 1462 * All ISNs for SYN-ACK packets are generated by the syncache. See 1463 * tcp_syncache.c for details. 1464 * 1465 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1466 * depends on this property. In addition, these ISNs should be 1467 * unguessable so as to prevent connection hijacking. To satisfy 1468 * the requirements of this situation, the algorithm outlined in 1469 * RFC 1948 is used to generate sequence numbers. 1470 * 1471 * Implementation details: 1472 * 1473 * Time is based off the system timer, and is corrected so that it 1474 * increases by one megabyte per second. This allows for proper 1475 * recycling on high speed LANs while still leaving over an hour 1476 * before rollover. 1477 * 1478 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1479 * between seeding of isn_secret. This is normally set to zero, 1480 * as reseeding should not be necessary. 1481 * 1482 */ 1483 1484 #define ISN_BYTES_PER_SECOND 1048576 1485 1486 u_char isn_secret[32]; 1487 int isn_last_reseed; 1488 MD5_CTX isn_ctx; 1489 1490 tcp_seq 1491 tcp_new_isn(struct tcpcb *tp) 1492 { 1493 u_int32_t md5_buffer[4]; 1494 tcp_seq new_isn; 1495 1496 /* Seed if this is the first use, reseed if requested. */ 1497 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1498 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1499 < (u_int)ticks))) { 1500 read_random_unlimited(&isn_secret, sizeof isn_secret); 1501 isn_last_reseed = ticks; 1502 } 1503 1504 /* Compute the md5 hash and return the ISN. */ 1505 MD5Init(&isn_ctx); 1506 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1507 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1508 #ifdef INET6 1509 if (tp->t_inpcb->inp_vflag & INP_IPV6) { 1510 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1511 sizeof(struct in6_addr)); 1512 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1513 sizeof(struct in6_addr)); 1514 } else 1515 #endif 1516 { 1517 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1518 sizeof(struct in_addr)); 1519 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1520 sizeof(struct in_addr)); 1521 } 1522 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1523 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1524 new_isn = (tcp_seq) md5_buffer[0]; 1525 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1526 return (new_isn); 1527 } 1528 1529 /* 1530 * When a source quench is received, close congestion window 1531 * to one segment. We will gradually open it again as we proceed. 1532 */ 1533 void 1534 tcp_quench(struct inpcb *inp, int errno) 1535 { 1536 struct tcpcb *tp = intotcpcb(inp); 1537 1538 if (tp != NULL) 1539 tp->snd_cwnd = tp->t_maxseg; 1540 } 1541 1542 /* 1543 * When a specific ICMP unreachable message is received and the 1544 * connection state is SYN-SENT, drop the connection. This behavior 1545 * is controlled by the icmp_may_rst sysctl. 1546 */ 1547 void 1548 tcp_drop_syn_sent(struct inpcb *inp, int errno) 1549 { 1550 struct tcpcb *tp = intotcpcb(inp); 1551 1552 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT)) 1553 tcp_drop(tp, errno); 1554 } 1555 1556 /* 1557 * When `need fragmentation' ICMP is received, update our idea of the MSS 1558 * based on the new value in the route. Also nudge TCP to send something, 1559 * since we know the packet we just sent was dropped. 1560 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1561 */ 1562 void 1563 tcp_mtudisc(struct inpcb *inp, int errno) 1564 { 1565 struct tcpcb *tp = intotcpcb(inp); 1566 struct rtentry *rt; 1567 struct rmxp_tao *taop; 1568 struct socket *so = inp->inp_socket; 1569 int offered; 1570 int mss; 1571 #ifdef INET6 1572 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0); 1573 #else 1574 const boolean_t isipv6 = FALSE; 1575 #endif 1576 1577 if (tp != NULL) { 1578 if (isipv6) 1579 rt = tcp_rtlookup6(&inp->inp_inc); 1580 else 1581 rt = tcp_rtlookup(&inp->inp_inc); 1582 if (rt == NULL || rt->rt_rmx.rmx_mtu == 0) { 1583 tp->t_maxopd = tp->t_maxseg = 1584 isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 1585 return; 1586 } 1587 taop = rmx_taop(rt->rt_rmx); 1588 offered = taop->tao_mssopt; 1589 mss = rt->rt_rmx.rmx_mtu - 1590 (isipv6 ? 1591 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1592 sizeof(struct tcpiphdr)); 1593 1594 if (offered != 0) 1595 mss = min(mss, offered); 1596 /* 1597 * XXX - The above conditional probably violates the TCP 1598 * spec. The problem is that, since we don't know the 1599 * other end's MSS, we are supposed to use a conservative 1600 * default. But, if we do that, then MTU discovery will 1601 * never actually take place, because the conservative 1602 * default is much less than the MTUs typically seen 1603 * on the Internet today. For the moment, we'll sweep 1604 * this under the carpet. 1605 * 1606 * The conservative default might not actually be a problem 1607 * if the only case this occurs is when sending an initial 1608 * SYN with options and data to a host we've never talked 1609 * to before. Then, they will reply with an MSS value which 1610 * will get recorded and the new parameters should get 1611 * recomputed. For Further Study. 1612 */ 1613 if (tp->t_maxopd <= mss) 1614 return; 1615 tp->t_maxopd = mss; 1616 1617 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP && 1618 (tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP) 1619 mss -= TCPOLEN_TSTAMP_APPA; 1620 if ((tp->t_flags & (TF_REQ_CC | TF_NOOPT)) == TF_REQ_CC && 1621 (tp->t_flags & TF_RCVD_CC) == TF_RCVD_CC) 1622 mss -= TCPOLEN_CC_APPA; 1623 #if (MCLBYTES & (MCLBYTES - 1)) == 0 1624 if (mss > MCLBYTES) 1625 mss &= ~(MCLBYTES - 1); 1626 #else 1627 if (mss > MCLBYTES) 1628 mss = mss / MCLBYTES * MCLBYTES; 1629 #endif 1630 if (so->so_snd.sb_hiwat < mss) 1631 mss = so->so_snd.sb_hiwat; 1632 1633 tp->t_maxseg = mss; 1634 1635 tcpstat.tcps_mturesent++; 1636 tp->t_rtttime = 0; 1637 tp->snd_nxt = tp->snd_una; 1638 tcp_output(tp); 1639 } 1640 } 1641 1642 /* 1643 * Look-up the routing entry to the peer of this inpcb. If no route 1644 * is found and it cannot be allocated the return NULL. This routine 1645 * is called by TCP routines that access the rmx structure and by tcp_mss 1646 * to get the interface MTU. 1647 */ 1648 struct rtentry * 1649 tcp_rtlookup(struct in_conninfo *inc) 1650 { 1651 struct route *ro; 1652 struct rtentry *rt; 1653 1654 ro = &inc->inc_route; 1655 rt = ro->ro_rt; 1656 if (rt == NULL || !(rt->rt_flags & RTF_UP)) { 1657 /* No route yet, so try to acquire one */ 1658 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1659 /* 1660 * unused portions of the structure MUST be zero'd 1661 * out because rtalloc() treats it as opaque data 1662 */ 1663 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1664 ro->ro_dst.sa_family = AF_INET; 1665 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1666 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1667 inc->inc_faddr; 1668 rtalloc(ro); 1669 rt = ro->ro_rt; 1670 } 1671 } 1672 return (rt); 1673 } 1674 1675 #ifdef INET6 1676 struct rtentry * 1677 tcp_rtlookup6(struct in_conninfo *inc) 1678 { 1679 struct route_in6 *ro6; 1680 struct rtentry *rt; 1681 1682 ro6 = &inc->inc6_route; 1683 rt = ro6->ro_rt; 1684 if (rt == NULL || !(rt->rt_flags & RTF_UP)) { 1685 /* No route yet, so try to acquire one */ 1686 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1687 /* 1688 * unused portions of the structure MUST be zero'd 1689 * out because rtalloc() treats it as opaque data 1690 */ 1691 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1692 ro6->ro_dst.sin6_family = AF_INET6; 1693 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1694 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1695 rtalloc((struct route *)ro6); 1696 rt = ro6->ro_rt; 1697 } 1698 } 1699 return (rt); 1700 } 1701 #endif 1702 1703 #ifdef IPSEC 1704 /* compute ESP/AH header size for TCP, including outer IP header. */ 1705 size_t 1706 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1707 { 1708 struct inpcb *inp; 1709 struct mbuf *m; 1710 size_t hdrsiz; 1711 struct ip *ip; 1712 struct tcphdr *th; 1713 1714 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1715 return (0); 1716 MGETHDR(m, MB_DONTWAIT, MT_DATA); 1717 if (!m) 1718 return (0); 1719 1720 #ifdef INET6 1721 if (inp->inp_vflag & INP_IPV6) { 1722 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1723 1724 th = (struct tcphdr *)(ip6 + 1); 1725 m->m_pkthdr.len = m->m_len = 1726 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1727 tcp_fillheaders(tp, ip6, th); 1728 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1729 } else 1730 #endif 1731 { 1732 ip = mtod(m, struct ip *); 1733 th = (struct tcphdr *)(ip + 1); 1734 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1735 tcp_fillheaders(tp, ip, th); 1736 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1737 } 1738 1739 m_free(m); 1740 return (hdrsiz); 1741 } 1742 #endif 1743 1744 /* 1745 * Return a pointer to the cached information about the remote host. 1746 * The cached information is stored in the protocol specific part of 1747 * the route metrics. 1748 */ 1749 struct rmxp_tao * 1750 tcp_gettaocache(struct in_conninfo *inc) 1751 { 1752 struct rtentry *rt; 1753 1754 #ifdef INET6 1755 if (inc->inc_isipv6) 1756 rt = tcp_rtlookup6(inc); 1757 else 1758 #endif 1759 rt = tcp_rtlookup(inc); 1760 1761 /* Make sure this is a host route and is up. */ 1762 if (rt == NULL || 1763 (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST)) 1764 return (NULL); 1765 1766 return (rmx_taop(rt->rt_rmx)); 1767 } 1768 1769 /* 1770 * Clear all the TAO cache entries, called from tcp_init. 1771 * 1772 * XXX 1773 * This routine is just an empty one, because we assume that the routing 1774 * routing tables are initialized at the same time when TCP, so there is 1775 * nothing in the cache left over. 1776 */ 1777 static void 1778 tcp_cleartaocache() 1779 { 1780 } 1781 1782 /* 1783 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1784 * 1785 * This code attempts to calculate the bandwidth-delay product as a 1786 * means of determining the optimal window size to maximize bandwidth, 1787 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1788 * routers. This code also does a fairly good job keeping RTTs in check 1789 * across slow links like modems. We implement an algorithm which is very 1790 * similar (but not meant to be) TCP/Vegas. The code operates on the 1791 * transmitter side of a TCP connection and so only effects the transmit 1792 * side of the connection. 1793 * 1794 * BACKGROUND: TCP makes no provision for the management of buffer space 1795 * at the end points or at the intermediate routers and switches. A TCP 1796 * stream, whether using NewReno or not, will eventually buffer as 1797 * many packets as it is able and the only reason this typically works is 1798 * due to the fairly small default buffers made available for a connection 1799 * (typicaly 16K or 32K). As machines use larger windows and/or window 1800 * scaling it is now fairly easy for even a single TCP connection to blow-out 1801 * all available buffer space not only on the local interface, but on 1802 * intermediate routers and switches as well. NewReno makes a misguided 1803 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1804 * then backing off, then steadily increasing the window again until another 1805 * failure occurs, ad-infinitum. This results in terrible oscillation that 1806 * is only made worse as network loads increase and the idea of intentionally 1807 * blowing out network buffers is, frankly, a terrible way to manage network 1808 * resources. 1809 * 1810 * It is far better to limit the transmit window prior to the failure 1811 * condition being achieved. There are two general ways to do this: First 1812 * you can 'scan' through different transmit window sizes and locate the 1813 * point where the RTT stops increasing, indicating that you have filled the 1814 * pipe, then scan backwards until you note that RTT stops decreasing, then 1815 * repeat ad-infinitum. This method works in principle but has severe 1816 * implementation issues due to RTT variances, timer granularity, and 1817 * instability in the algorithm which can lead to many false positives and 1818 * create oscillations as well as interact badly with other TCP streams 1819 * implementing the same algorithm. 1820 * 1821 * The second method is to limit the window to the bandwidth delay product 1822 * of the link. This is the method we implement. RTT variances and our 1823 * own manipulation of the congestion window, bwnd, can potentially 1824 * destabilize the algorithm. For this reason we have to stabilize the 1825 * elements used to calculate the window. We do this by using the minimum 1826 * observed RTT, the long term average of the observed bandwidth, and 1827 * by adding two segments worth of slop. It isn't perfect but it is able 1828 * to react to changing conditions and gives us a very stable basis on 1829 * which to extend the algorithm. 1830 */ 1831 void 1832 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1833 { 1834 u_long bw; 1835 u_long bwnd; 1836 int save_ticks; 1837 1838 /* 1839 * If inflight_enable is disabled in the middle of a tcp connection, 1840 * make sure snd_bwnd is effectively disabled. 1841 */ 1842 if (!tcp_inflight_enable) { 1843 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1844 tp->snd_bandwidth = 0; 1845 return; 1846 } 1847 1848 /* 1849 * Figure out the bandwidth. Due to the tick granularity this 1850 * is a very rough number and it MUST be averaged over a fairly 1851 * long period of time. XXX we need to take into account a link 1852 * that is not using all available bandwidth, but for now our 1853 * slop will ramp us up if this case occurs and the bandwidth later 1854 * increases. 1855 * 1856 * Note: if ticks rollover 'bw' may wind up negative. We must 1857 * effectively reset t_bw_rtttime for this case. 1858 */ 1859 save_ticks = ticks; 1860 if ((u_int)(save_ticks - tp->t_bw_rtttime) < 1) 1861 return; 1862 1863 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / 1864 (save_ticks - tp->t_bw_rtttime); 1865 tp->t_bw_rtttime = save_ticks; 1866 tp->t_bw_rtseq = ack_seq; 1867 if (tp->t_bw_rtttime == 0 || (int)bw < 0) 1868 return; 1869 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4; 1870 1871 tp->snd_bandwidth = bw; 1872 1873 /* 1874 * Calculate the semi-static bandwidth delay product, plus two maximal 1875 * segments. The additional slop puts us squarely in the sweet 1876 * spot and also handles the bandwidth run-up case. Without the 1877 * slop we could be locking ourselves into a lower bandwidth. 1878 * 1879 * Situations Handled: 1880 * (1) Prevents over-queueing of packets on LANs, especially on 1881 * high speed LANs, allowing larger TCP buffers to be 1882 * specified, and also does a good job preventing 1883 * over-queueing of packets over choke points like modems 1884 * (at least for the transmit side). 1885 * 1886 * (2) Is able to handle changing network loads (bandwidth 1887 * drops so bwnd drops, bandwidth increases so bwnd 1888 * increases). 1889 * 1890 * (3) Theoretically should stabilize in the face of multiple 1891 * connections implementing the same algorithm (this may need 1892 * a little work). 1893 * 1894 * (4) Stability value (defaults to 20 = 2 maximal packets) can 1895 * be adjusted with a sysctl but typically only needs to be on 1896 * very slow connections. A value no smaller then 5 should 1897 * be used, but only reduce this default if you have no other 1898 * choice. 1899 */ 1900 1901 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2) 1902 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 1903 tcp_inflight_stab * (int)tp->t_maxseg / 10; 1904 #undef USERTT 1905 1906 if (tcp_inflight_debug > 0) { 1907 static int ltime; 1908 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) { 1909 ltime = ticks; 1910 printf("%p bw %ld rttbest %d srtt %d bwnd %ld\n", 1911 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd); 1912 } 1913 } 1914 if ((long)bwnd < tcp_inflight_min) 1915 bwnd = tcp_inflight_min; 1916 if (bwnd > tcp_inflight_max) 1917 bwnd = tcp_inflight_max; 1918 if ((long)bwnd < tp->t_maxseg * 2) 1919 bwnd = tp->t_maxseg * 2; 1920 tp->snd_bwnd = bwnd; 1921 } 1922