1 /*- 2 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 3 * The Regents of the University of California. All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 4. Neither the name of the University nor the names of its contributors 14 * may be used to endorse or promote products derived from this software 15 * without specific prior written permission. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 * 29 * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95 30 */ 31 32 #include <sys/cdefs.h> 33 __FBSDID("$FreeBSD$"); 34 35 #include "opt_compat.h" 36 #include "opt_inet.h" 37 #include "opt_inet6.h" 38 #include "opt_ipsec.h" 39 #include "opt_mac.h" 40 #include "opt_tcpdebug.h" 41 42 #include <sys/param.h> 43 #include <sys/systm.h> 44 #include <sys/callout.h> 45 #include <sys/kernel.h> 46 #include <sys/sysctl.h> 47 #include <sys/malloc.h> 48 #include <sys/mbuf.h> 49 #ifdef INET6 50 #include <sys/domain.h> 51 #endif 52 #include <sys/priv.h> 53 #include <sys/proc.h> 54 #include <sys/socket.h> 55 #include <sys/socketvar.h> 56 #include <sys/protosw.h> 57 #include <sys/random.h> 58 59 #include <vm/uma.h> 60 61 #include <net/route.h> 62 #include <net/if.h> 63 64 #include <netinet/in.h> 65 #include <netinet/in_systm.h> 66 #include <netinet/ip.h> 67 #ifdef INET6 68 #include <netinet/ip6.h> 69 #endif 70 #include <netinet/in_pcb.h> 71 #ifdef INET6 72 #include <netinet6/in6_pcb.h> 73 #endif 74 #include <netinet/in_var.h> 75 #include <netinet/ip_var.h> 76 #ifdef INET6 77 #include <netinet6/ip6_var.h> 78 #include <netinet6/scope6_var.h> 79 #include <netinet6/nd6.h> 80 #endif 81 #include <netinet/ip_icmp.h> 82 #include <netinet/tcp.h> 83 #include <netinet/tcp_fsm.h> 84 #include <netinet/tcp_seq.h> 85 #include <netinet/tcp_timer.h> 86 #include <netinet/tcp_var.h> 87 #include <netinet/tcp_syncache.h> 88 #include <netinet/tcp_offload.h> 89 #ifdef INET6 90 #include <netinet6/tcp6_var.h> 91 #endif 92 #include <netinet/tcpip.h> 93 #ifdef TCPDEBUG 94 #include <netinet/tcp_debug.h> 95 #endif 96 #include <netinet6/ip6protosw.h> 97 98 #ifdef IPSEC 99 #include <netipsec/ipsec.h> 100 #include <netipsec/xform.h> 101 #ifdef INET6 102 #include <netipsec/ipsec6.h> 103 #endif 104 #include <netipsec/key.h> 105 #endif /*IPSEC*/ 106 107 #include <machine/in_cksum.h> 108 #include <sys/md5.h> 109 110 #include <security/mac/mac_framework.h> 111 112 int tcp_mssdflt = TCP_MSS; 113 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 114 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size"); 115 116 #ifdef INET6 117 int tcp_v6mssdflt = TCP6_MSS; 118 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, 119 CTLFLAG_RW, &tcp_v6mssdflt , 0, 120 "Default TCP Maximum Segment Size for IPv6"); 121 #endif 122 123 /* 124 * Minimum MSS we accept and use. This prevents DoS attacks where 125 * we are forced to a ridiculous low MSS like 20 and send hundreds 126 * of packets instead of one. The effect scales with the available 127 * bandwidth and quickly saturates the CPU and network interface 128 * with packet generation and sending. Set to zero to disable MINMSS 129 * checking. This setting prevents us from sending too small packets. 130 */ 131 int tcp_minmss = TCP_MINMSS; 132 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW, 133 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size"); 134 135 int tcp_do_rfc1323 = 1; 136 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW, 137 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions"); 138 139 static int tcp_log_debug = 0; 140 SYSCTL_INT(_net_inet_tcp, OID_AUTO, log_debug, CTLFLAG_RW, 141 &tcp_log_debug, 0, "Log errors caused by incoming TCP segments"); 142 143 static int tcp_tcbhashsize = 0; 144 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RDTUN, 145 &tcp_tcbhashsize, 0, "Size of TCP control-block hashtable"); 146 147 static int do_tcpdrain = 1; 148 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, 149 &do_tcpdrain, 0, 150 "Enable tcp_drain routine for extra help when low on mbufs"); 151 152 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD, 153 &tcbinfo.ipi_count, 0, "Number of active PCBs"); 154 155 static int icmp_may_rst = 1; 156 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, 157 &icmp_may_rst, 0, 158 "Certain ICMP unreachable messages may abort connections in SYN_SENT"); 159 160 static int tcp_isn_reseed_interval = 0; 161 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW, 162 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret"); 163 164 /* 165 * TCP bandwidth limiting sysctls. Note that the default lower bound of 166 * 1024 exists only for debugging. A good production default would be 167 * something like 6100. 168 */ 169 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, inflight, CTLFLAG_RW, 0, 170 "TCP inflight data limiting"); 171 172 static int tcp_inflight_enable = 1; 173 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, enable, CTLFLAG_RW, 174 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting"); 175 176 static int tcp_inflight_debug = 0; 177 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, debug, CTLFLAG_RW, 178 &tcp_inflight_debug, 0, "Debug TCP inflight calculations"); 179 180 static int tcp_inflight_rttthresh; 181 SYSCTL_PROC(_net_inet_tcp_inflight, OID_AUTO, rttthresh, CTLTYPE_INT|CTLFLAG_RW, 182 &tcp_inflight_rttthresh, 0, sysctl_msec_to_ticks, "I", 183 "RTT threshold below which inflight will deactivate itself"); 184 185 static int tcp_inflight_min = 6144; 186 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, min, CTLFLAG_RW, 187 &tcp_inflight_min, 0, "Lower-bound for TCP inflight window"); 188 189 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT; 190 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, max, CTLFLAG_RW, 191 &tcp_inflight_max, 0, "Upper-bound for TCP inflight window"); 192 193 static int tcp_inflight_stab = 20; 194 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, stab, CTLFLAG_RW, 195 &tcp_inflight_stab, 0, "Inflight Algorithm Stabilization 20 = 2 packets"); 196 197 uma_zone_t sack_hole_zone; 198 199 static struct inpcb *tcp_notify(struct inpcb *, int); 200 static void tcp_isn_tick(void *); 201 202 /* 203 * Target size of TCP PCB hash tables. Must be a power of two. 204 * 205 * Note that this can be overridden by the kernel environment 206 * variable net.inet.tcp.tcbhashsize 207 */ 208 #ifndef TCBHASHSIZE 209 #define TCBHASHSIZE 512 210 #endif 211 212 /* 213 * XXX 214 * Callouts should be moved into struct tcp directly. They are currently 215 * separate because the tcpcb structure is exported to userland for sysctl 216 * parsing purposes, which do not know about callouts. 217 */ 218 struct tcpcb_mem { 219 struct tcpcb tcb; 220 struct tcp_timer tt; 221 }; 222 223 static uma_zone_t tcpcb_zone; 224 MALLOC_DEFINE(M_TCPLOG, "tcplog", "TCP address and flags print buffers"); 225 struct callout isn_callout; 226 static struct mtx isn_mtx; 227 228 #define ISN_LOCK_INIT() mtx_init(&isn_mtx, "isn_mtx", NULL, MTX_DEF) 229 #define ISN_LOCK() mtx_lock(&isn_mtx) 230 #define ISN_UNLOCK() mtx_unlock(&isn_mtx) 231 232 /* 233 * TCP initialization. 234 */ 235 static void 236 tcp_zone_change(void *tag) 237 { 238 239 uma_zone_set_max(tcbinfo.ipi_zone, maxsockets); 240 uma_zone_set_max(tcpcb_zone, maxsockets); 241 tcp_tw_zone_change(); 242 } 243 244 static int 245 tcp_inpcb_init(void *mem, int size, int flags) 246 { 247 struct inpcb *inp = mem; 248 249 INP_LOCK_INIT(inp, "inp", "tcpinp"); 250 return (0); 251 } 252 253 void 254 tcp_init(void) 255 { 256 257 int hashsize = TCBHASHSIZE; 258 tcp_delacktime = TCPTV_DELACK; 259 tcp_keepinit = TCPTV_KEEP_INIT; 260 tcp_keepidle = TCPTV_KEEP_IDLE; 261 tcp_keepintvl = TCPTV_KEEPINTVL; 262 tcp_maxpersistidle = TCPTV_KEEP_IDLE; 263 tcp_msl = TCPTV_MSL; 264 tcp_rexmit_min = TCPTV_MIN; 265 if (tcp_rexmit_min < 1) 266 tcp_rexmit_min = 1; 267 tcp_rexmit_slop = TCPTV_CPU_VAR; 268 tcp_inflight_rttthresh = TCPTV_INFLIGHT_RTTTHRESH; 269 tcp_finwait2_timeout = TCPTV_FINWAIT2_TIMEOUT; 270 271 INP_INFO_LOCK_INIT(&tcbinfo, "tcp"); 272 LIST_INIT(&tcb); 273 tcbinfo.ipi_listhead = &tcb; 274 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize); 275 if (!powerof2(hashsize)) { 276 printf("WARNING: TCB hash size not a power of 2\n"); 277 hashsize = 512; /* safe default */ 278 } 279 tcp_tcbhashsize = hashsize; 280 tcbinfo.ipi_hashbase = hashinit(hashsize, M_PCB, 281 &tcbinfo.ipi_hashmask); 282 tcbinfo.ipi_porthashbase = hashinit(hashsize, M_PCB, 283 &tcbinfo.ipi_porthashmask); 284 tcbinfo.ipi_zone = uma_zcreate("inpcb", sizeof(struct inpcb), 285 NULL, NULL, tcp_inpcb_init, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 286 uma_zone_set_max(tcbinfo.ipi_zone, maxsockets); 287 #ifdef INET6 288 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) 289 #else /* INET6 */ 290 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) 291 #endif /* INET6 */ 292 if (max_protohdr < TCP_MINPROTOHDR) 293 max_protohdr = TCP_MINPROTOHDR; 294 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) 295 panic("tcp_init"); 296 #undef TCP_MINPROTOHDR 297 /* 298 * These have to be type stable for the benefit of the timers. 299 */ 300 tcpcb_zone = uma_zcreate("tcpcb", sizeof(struct tcpcb_mem), 301 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 302 uma_zone_set_max(tcpcb_zone, maxsockets); 303 tcp_tw_init(); 304 syncache_init(); 305 tcp_hc_init(); 306 tcp_reass_init(); 307 ISN_LOCK_INIT(); 308 callout_init(&isn_callout, CALLOUT_MPSAFE); 309 tcp_isn_tick(NULL); 310 EVENTHANDLER_REGISTER(shutdown_pre_sync, tcp_fini, NULL, 311 SHUTDOWN_PRI_DEFAULT); 312 sack_hole_zone = uma_zcreate("sackhole", sizeof(struct sackhole), 313 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 314 EVENTHANDLER_REGISTER(maxsockets_change, tcp_zone_change, NULL, 315 EVENTHANDLER_PRI_ANY); 316 } 317 318 void 319 tcp_fini(void *xtp) 320 { 321 322 callout_stop(&isn_callout); 323 } 324 325 /* 326 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. 327 * tcp_template used to store this data in mbufs, but we now recopy it out 328 * of the tcpcb each time to conserve mbufs. 329 */ 330 void 331 tcpip_fillheaders(struct inpcb *inp, void *ip_ptr, void *tcp_ptr) 332 { 333 struct tcphdr *th = (struct tcphdr *)tcp_ptr; 334 335 INP_WLOCK_ASSERT(inp); 336 337 #ifdef INET6 338 if ((inp->inp_vflag & INP_IPV6) != 0) { 339 struct ip6_hdr *ip6; 340 341 ip6 = (struct ip6_hdr *)ip_ptr; 342 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | 343 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK); 344 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | 345 (IPV6_VERSION & IPV6_VERSION_MASK); 346 ip6->ip6_nxt = IPPROTO_TCP; 347 ip6->ip6_plen = sizeof(struct tcphdr); 348 ip6->ip6_src = inp->in6p_laddr; 349 ip6->ip6_dst = inp->in6p_faddr; 350 } else 351 #endif 352 { 353 struct ip *ip; 354 355 ip = (struct ip *)ip_ptr; 356 ip->ip_v = IPVERSION; 357 ip->ip_hl = 5; 358 ip->ip_tos = inp->inp_ip_tos; 359 ip->ip_len = 0; 360 ip->ip_id = 0; 361 ip->ip_off = 0; 362 ip->ip_ttl = inp->inp_ip_ttl; 363 ip->ip_sum = 0; 364 ip->ip_p = IPPROTO_TCP; 365 ip->ip_src = inp->inp_laddr; 366 ip->ip_dst = inp->inp_faddr; 367 } 368 th->th_sport = inp->inp_lport; 369 th->th_dport = inp->inp_fport; 370 th->th_seq = 0; 371 th->th_ack = 0; 372 th->th_x2 = 0; 373 th->th_off = 5; 374 th->th_flags = 0; 375 th->th_win = 0; 376 th->th_urp = 0; 377 th->th_sum = 0; /* in_pseudo() is called later for ipv4 */ 378 } 379 380 /* 381 * Create template to be used to send tcp packets on a connection. 382 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only 383 * use for this function is in keepalives, which use tcp_respond. 384 */ 385 struct tcptemp * 386 tcpip_maketemplate(struct inpcb *inp) 387 { 388 struct mbuf *m; 389 struct tcptemp *n; 390 391 m = m_get(M_DONTWAIT, MT_DATA); 392 if (m == NULL) 393 return (0); 394 m->m_len = sizeof(struct tcptemp); 395 n = mtod(m, struct tcptemp *); 396 397 tcpip_fillheaders(inp, (void *)&n->tt_ipgen, (void *)&n->tt_t); 398 return (n); 399 } 400 401 /* 402 * Send a single message to the TCP at address specified by 403 * the given TCP/IP header. If m == NULL, then we make a copy 404 * of the tcpiphdr at ti and send directly to the addressed host. 405 * This is used to force keep alive messages out using the TCP 406 * template for a connection. If flags are given then we send 407 * a message back to the TCP which originated the * segment ti, 408 * and discard the mbuf containing it and any other attached mbufs. 409 * 410 * In any case the ack and sequence number of the transmitted 411 * segment are as specified by the parameters. 412 * 413 * NOTE: If m != NULL, then ti must point to *inside* the mbuf. 414 */ 415 void 416 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, 417 tcp_seq ack, tcp_seq seq, int flags) 418 { 419 int tlen; 420 int win = 0; 421 struct ip *ip; 422 struct tcphdr *nth; 423 #ifdef INET6 424 struct ip6_hdr *ip6; 425 int isipv6; 426 #endif /* INET6 */ 427 int ipflags = 0; 428 struct inpcb *inp; 429 430 KASSERT(tp != NULL || m != NULL, ("tcp_respond: tp and m both NULL")); 431 432 #ifdef INET6 433 isipv6 = ((struct ip *)ipgen)->ip_v == 6; 434 ip6 = ipgen; 435 #endif /* INET6 */ 436 ip = ipgen; 437 438 if (tp != NULL) { 439 inp = tp->t_inpcb; 440 KASSERT(inp != NULL, ("tcp control block w/o inpcb")); 441 INP_WLOCK_ASSERT(inp); 442 } else 443 inp = NULL; 444 445 if (tp != NULL) { 446 if (!(flags & TH_RST)) { 447 win = sbspace(&inp->inp_socket->so_rcv); 448 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 449 win = (long)TCP_MAXWIN << tp->rcv_scale; 450 } 451 } 452 if (m == NULL) { 453 m = m_gethdr(M_DONTWAIT, MT_DATA); 454 if (m == NULL) 455 return; 456 tlen = 0; 457 m->m_data += max_linkhdr; 458 #ifdef INET6 459 if (isipv6) { 460 bcopy((caddr_t)ip6, mtod(m, caddr_t), 461 sizeof(struct ip6_hdr)); 462 ip6 = mtod(m, struct ip6_hdr *); 463 nth = (struct tcphdr *)(ip6 + 1); 464 } else 465 #endif /* INET6 */ 466 { 467 bcopy((caddr_t)ip, mtod(m, caddr_t), sizeof(struct ip)); 468 ip = mtod(m, struct ip *); 469 nth = (struct tcphdr *)(ip + 1); 470 } 471 bcopy((caddr_t)th, (caddr_t)nth, sizeof(struct tcphdr)); 472 flags = TH_ACK; 473 } else { 474 /* 475 * reuse the mbuf. 476 * XXX MRT We inherrit the FIB, which is lucky. 477 */ 478 m_freem(m->m_next); 479 m->m_next = NULL; 480 m->m_data = (caddr_t)ipgen; 481 /* m_len is set later */ 482 tlen = 0; 483 #define xchg(a,b,type) { type t; t=a; a=b; b=t; } 484 #ifdef INET6 485 if (isipv6) { 486 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); 487 nth = (struct tcphdr *)(ip6 + 1); 488 } else 489 #endif /* INET6 */ 490 { 491 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long); 492 nth = (struct tcphdr *)(ip + 1); 493 } 494 if (th != nth) { 495 /* 496 * this is usually a case when an extension header 497 * exists between the IPv6 header and the 498 * TCP header. 499 */ 500 nth->th_sport = th->th_sport; 501 nth->th_dport = th->th_dport; 502 } 503 xchg(nth->th_dport, nth->th_sport, n_short); 504 #undef xchg 505 } 506 #ifdef INET6 507 if (isipv6) { 508 ip6->ip6_flow = 0; 509 ip6->ip6_vfc = IPV6_VERSION; 510 ip6->ip6_nxt = IPPROTO_TCP; 511 ip6->ip6_plen = htons((u_short)(sizeof (struct tcphdr) + 512 tlen)); 513 tlen += sizeof (struct ip6_hdr) + sizeof (struct tcphdr); 514 } else 515 #endif 516 { 517 tlen += sizeof (struct tcpiphdr); 518 ip->ip_len = tlen; 519 ip->ip_ttl = ip_defttl; 520 if (path_mtu_discovery) 521 ip->ip_off |= IP_DF; 522 } 523 m->m_len = tlen; 524 m->m_pkthdr.len = tlen; 525 m->m_pkthdr.rcvif = NULL; 526 #ifdef MAC 527 if (inp != NULL) { 528 /* 529 * Packet is associated with a socket, so allow the 530 * label of the response to reflect the socket label. 531 */ 532 INP_WLOCK_ASSERT(inp); 533 mac_inpcb_create_mbuf(inp, m); 534 } else { 535 /* 536 * Packet is not associated with a socket, so possibly 537 * update the label in place. 538 */ 539 mac_netinet_tcp_reply(m); 540 } 541 #endif 542 nth->th_seq = htonl(seq); 543 nth->th_ack = htonl(ack); 544 nth->th_x2 = 0; 545 nth->th_off = sizeof (struct tcphdr) >> 2; 546 nth->th_flags = flags; 547 if (tp != NULL) 548 nth->th_win = htons((u_short) (win >> tp->rcv_scale)); 549 else 550 nth->th_win = htons((u_short)win); 551 nth->th_urp = 0; 552 #ifdef INET6 553 if (isipv6) { 554 nth->th_sum = 0; 555 nth->th_sum = in6_cksum(m, IPPROTO_TCP, 556 sizeof(struct ip6_hdr), 557 tlen - sizeof(struct ip6_hdr)); 558 ip6->ip6_hlim = in6_selecthlim(tp != NULL ? tp->t_inpcb : 559 NULL, NULL); 560 } else 561 #endif /* INET6 */ 562 { 563 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 564 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); 565 m->m_pkthdr.csum_flags = CSUM_TCP; 566 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 567 } 568 #ifdef TCPDEBUG 569 if (tp == NULL || (inp->inp_socket->so_options & SO_DEBUG)) 570 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); 571 #endif 572 #ifdef INET6 573 if (isipv6) 574 (void) ip6_output(m, NULL, NULL, ipflags, NULL, NULL, inp); 575 else 576 #endif /* INET6 */ 577 (void) ip_output(m, NULL, NULL, ipflags, NULL, inp); 578 } 579 580 /* 581 * Create a new TCP control block, making an 582 * empty reassembly queue and hooking it to the argument 583 * protocol control block. The `inp' parameter must have 584 * come from the zone allocator set up in tcp_init(). 585 */ 586 struct tcpcb * 587 tcp_newtcpcb(struct inpcb *inp) 588 { 589 struct tcpcb_mem *tm; 590 struct tcpcb *tp; 591 #ifdef INET6 592 int isipv6 = (inp->inp_vflag & INP_IPV6) != 0; 593 #endif /* INET6 */ 594 595 tm = uma_zalloc(tcpcb_zone, M_NOWAIT | M_ZERO); 596 if (tm == NULL) 597 return (NULL); 598 tp = &tm->tcb; 599 tp->t_timers = &tm->tt; 600 /* LIST_INIT(&tp->t_segq); */ /* XXX covered by M_ZERO */ 601 tp->t_maxseg = tp->t_maxopd = 602 #ifdef INET6 603 isipv6 ? tcp_v6mssdflt : 604 #endif /* INET6 */ 605 tcp_mssdflt; 606 607 /* Set up our timeouts. */ 608 callout_init(&tp->t_timers->tt_rexmt, CALLOUT_MPSAFE); 609 callout_init(&tp->t_timers->tt_persist, CALLOUT_MPSAFE); 610 callout_init(&tp->t_timers->tt_keep, CALLOUT_MPSAFE); 611 callout_init(&tp->t_timers->tt_2msl, CALLOUT_MPSAFE); 612 callout_init(&tp->t_timers->tt_delack, CALLOUT_MPSAFE); 613 614 if (tcp_do_rfc1323) 615 tp->t_flags = (TF_REQ_SCALE|TF_REQ_TSTMP); 616 if (tcp_do_sack) 617 tp->t_flags |= TF_SACK_PERMIT; 618 TAILQ_INIT(&tp->snd_holes); 619 tp->t_inpcb = inp; /* XXX */ 620 /* 621 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 622 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 623 * reasonable initial retransmit time. 624 */ 625 tp->t_srtt = TCPTV_SRTTBASE; 626 tp->t_rttvar = ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 627 tp->t_rttmin = tcp_rexmit_min; 628 tp->t_rxtcur = TCPTV_RTOBASE; 629 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 630 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 631 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 632 tp->t_rcvtime = ticks; 633 tp->t_bw_rtttime = ticks; 634 /* 635 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 636 * because the socket may be bound to an IPv6 wildcard address, 637 * which may match an IPv4-mapped IPv6 address. 638 */ 639 inp->inp_ip_ttl = ip_defttl; 640 inp->inp_ppcb = tp; 641 return (tp); /* XXX */ 642 } 643 644 /* 645 * Drop a TCP connection, reporting 646 * the specified error. If connection is synchronized, 647 * then send a RST to peer. 648 */ 649 struct tcpcb * 650 tcp_drop(struct tcpcb *tp, int errno) 651 { 652 struct socket *so = tp->t_inpcb->inp_socket; 653 654 INP_INFO_WLOCK_ASSERT(&tcbinfo); 655 INP_WLOCK_ASSERT(tp->t_inpcb); 656 657 if (TCPS_HAVERCVDSYN(tp->t_state)) { 658 tp->t_state = TCPS_CLOSED; 659 (void) tcp_output_reset(tp); 660 tcpstat.tcps_drops++; 661 } else 662 tcpstat.tcps_conndrops++; 663 if (errno == ETIMEDOUT && tp->t_softerror) 664 errno = tp->t_softerror; 665 so->so_error = errno; 666 return (tcp_close(tp)); 667 } 668 669 void 670 tcp_discardcb(struct tcpcb *tp) 671 { 672 struct tseg_qent *q; 673 struct inpcb *inp = tp->t_inpcb; 674 struct socket *so = inp->inp_socket; 675 #ifdef INET6 676 int isipv6 = (inp->inp_vflag & INP_IPV6) != 0; 677 #endif /* INET6 */ 678 679 INP_WLOCK_ASSERT(inp); 680 681 /* 682 * Make sure that all of our timers are stopped before we 683 * delete the PCB. 684 */ 685 callout_stop(&tp->t_timers->tt_rexmt); 686 callout_stop(&tp->t_timers->tt_persist); 687 callout_stop(&tp->t_timers->tt_keep); 688 callout_stop(&tp->t_timers->tt_2msl); 689 callout_stop(&tp->t_timers->tt_delack); 690 691 /* 692 * If we got enough samples through the srtt filter, 693 * save the rtt and rttvar in the routing entry. 694 * 'Enough' is arbitrarily defined as 4 rtt samples. 695 * 4 samples is enough for the srtt filter to converge 696 * to within enough % of the correct value; fewer samples 697 * and we could save a bogus rtt. The danger is not high 698 * as tcp quickly recovers from everything. 699 * XXX: Works very well but needs some more statistics! 700 */ 701 if (tp->t_rttupdated >= 4) { 702 struct hc_metrics_lite metrics; 703 u_long ssthresh; 704 705 bzero(&metrics, sizeof(metrics)); 706 /* 707 * Update the ssthresh always when the conditions below 708 * are satisfied. This gives us better new start value 709 * for the congestion avoidance for new connections. 710 * ssthresh is only set if packet loss occured on a session. 711 * 712 * XXXRW: 'so' may be NULL here, and/or socket buffer may be 713 * being torn down. Ideally this code would not use 'so'. 714 */ 715 ssthresh = tp->snd_ssthresh; 716 if (ssthresh != 0 && ssthresh < so->so_snd.sb_hiwat / 2) { 717 /* 718 * convert the limit from user data bytes to 719 * packets then to packet data bytes. 720 */ 721 ssthresh = (ssthresh + tp->t_maxseg / 2) / tp->t_maxseg; 722 if (ssthresh < 2) 723 ssthresh = 2; 724 ssthresh *= (u_long)(tp->t_maxseg + 725 #ifdef INET6 726 (isipv6 ? sizeof (struct ip6_hdr) + 727 sizeof (struct tcphdr) : 728 #endif 729 sizeof (struct tcpiphdr) 730 #ifdef INET6 731 ) 732 #endif 733 ); 734 } else 735 ssthresh = 0; 736 metrics.rmx_ssthresh = ssthresh; 737 738 metrics.rmx_rtt = tp->t_srtt; 739 metrics.rmx_rttvar = tp->t_rttvar; 740 /* XXX: This wraps if the pipe is more than 4 Gbit per second */ 741 metrics.rmx_bandwidth = tp->snd_bandwidth; 742 metrics.rmx_cwnd = tp->snd_cwnd; 743 metrics.rmx_sendpipe = 0; 744 metrics.rmx_recvpipe = 0; 745 746 tcp_hc_update(&inp->inp_inc, &metrics); 747 } 748 749 /* free the reassembly queue, if any */ 750 while ((q = LIST_FIRST(&tp->t_segq)) != NULL) { 751 LIST_REMOVE(q, tqe_q); 752 m_freem(q->tqe_m); 753 uma_zfree(tcp_reass_zone, q); 754 tp->t_segqlen--; 755 tcp_reass_qsize--; 756 } 757 /* Disconnect offload device, if any. */ 758 tcp_offload_detach(tp); 759 760 tcp_free_sackholes(tp); 761 inp->inp_ppcb = NULL; 762 tp->t_inpcb = NULL; 763 uma_zfree(tcpcb_zone, tp); 764 } 765 766 /* 767 * Attempt to close a TCP control block, marking it as dropped, and freeing 768 * the socket if we hold the only reference. 769 */ 770 struct tcpcb * 771 tcp_close(struct tcpcb *tp) 772 { 773 struct inpcb *inp = tp->t_inpcb; 774 struct socket *so; 775 776 INP_INFO_WLOCK_ASSERT(&tcbinfo); 777 INP_WLOCK_ASSERT(inp); 778 779 /* Notify any offload devices of listener close */ 780 if (tp->t_state == TCPS_LISTEN) 781 tcp_offload_listen_close(tp); 782 in_pcbdrop(inp); 783 tcpstat.tcps_closed++; 784 KASSERT(inp->inp_socket != NULL, ("tcp_close: inp_socket NULL")); 785 so = inp->inp_socket; 786 soisdisconnected(so); 787 if (inp->inp_vflag & INP_SOCKREF) { 788 KASSERT(so->so_state & SS_PROTOREF, 789 ("tcp_close: !SS_PROTOREF")); 790 inp->inp_vflag &= ~INP_SOCKREF; 791 INP_WUNLOCK(inp); 792 ACCEPT_LOCK(); 793 SOCK_LOCK(so); 794 so->so_state &= ~SS_PROTOREF; 795 sofree(so); 796 return (NULL); 797 } 798 return (tp); 799 } 800 801 void 802 tcp_drain(void) 803 { 804 805 if (do_tcpdrain) { 806 struct inpcb *inpb; 807 struct tcpcb *tcpb; 808 struct tseg_qent *te; 809 810 /* 811 * Walk the tcpbs, if existing, and flush the reassembly queue, 812 * if there is one... 813 * XXX: The "Net/3" implementation doesn't imply that the TCP 814 * reassembly queue should be flushed, but in a situation 815 * where we're really low on mbufs, this is potentially 816 * usefull. 817 */ 818 INP_INFO_RLOCK(&tcbinfo); 819 LIST_FOREACH(inpb, tcbinfo.ipi_listhead, inp_list) { 820 if (inpb->inp_vflag & INP_TIMEWAIT) 821 continue; 822 INP_WLOCK(inpb); 823 if ((tcpb = intotcpcb(inpb)) != NULL) { 824 while ((te = LIST_FIRST(&tcpb->t_segq)) 825 != NULL) { 826 LIST_REMOVE(te, tqe_q); 827 m_freem(te->tqe_m); 828 uma_zfree(tcp_reass_zone, te); 829 tcpb->t_segqlen--; 830 tcp_reass_qsize--; 831 } 832 tcp_clean_sackreport(tcpb); 833 } 834 INP_WUNLOCK(inpb); 835 } 836 INP_INFO_RUNLOCK(&tcbinfo); 837 } 838 } 839 840 /* 841 * Notify a tcp user of an asynchronous error; 842 * store error as soft error, but wake up user 843 * (for now, won't do anything until can select for soft error). 844 * 845 * Do not wake up user since there currently is no mechanism for 846 * reporting soft errors (yet - a kqueue filter may be added). 847 */ 848 static struct inpcb * 849 tcp_notify(struct inpcb *inp, int error) 850 { 851 struct tcpcb *tp; 852 853 INP_INFO_WLOCK_ASSERT(&tcbinfo); 854 INP_WLOCK_ASSERT(inp); 855 856 if ((inp->inp_vflag & INP_TIMEWAIT) || 857 (inp->inp_vflag & INP_DROPPED)) 858 return (inp); 859 860 tp = intotcpcb(inp); 861 KASSERT(tp != NULL, ("tcp_notify: tp == NULL")); 862 863 /* 864 * Ignore some errors if we are hooked up. 865 * If connection hasn't completed, has retransmitted several times, 866 * and receives a second error, give up now. This is better 867 * than waiting a long time to establish a connection that 868 * can never complete. 869 */ 870 if (tp->t_state == TCPS_ESTABLISHED && 871 (error == EHOSTUNREACH || error == ENETUNREACH || 872 error == EHOSTDOWN)) { 873 return (inp); 874 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 875 tp->t_softerror) { 876 tp = tcp_drop(tp, error); 877 if (tp != NULL) 878 return (inp); 879 else 880 return (NULL); 881 } else { 882 tp->t_softerror = error; 883 return (inp); 884 } 885 #if 0 886 wakeup( &so->so_timeo); 887 sorwakeup(so); 888 sowwakeup(so); 889 #endif 890 } 891 892 static int 893 tcp_pcblist(SYSCTL_HANDLER_ARGS) 894 { 895 int error, i, m, n, pcb_count; 896 struct inpcb *inp, **inp_list; 897 inp_gen_t gencnt; 898 struct xinpgen xig; 899 900 /* 901 * The process of preparing the TCB list is too time-consuming and 902 * resource-intensive to repeat twice on every request. 903 */ 904 if (req->oldptr == NULL) { 905 m = syncache_pcbcount(); 906 n = tcbinfo.ipi_count; 907 req->oldidx = 2 * (sizeof xig) 908 + ((m + n) + n/8) * sizeof(struct xtcpcb); 909 return (0); 910 } 911 912 if (req->newptr != NULL) 913 return (EPERM); 914 915 /* 916 * OK, now we're committed to doing something. 917 */ 918 INP_INFO_RLOCK(&tcbinfo); 919 gencnt = tcbinfo.ipi_gencnt; 920 n = tcbinfo.ipi_count; 921 INP_INFO_RUNLOCK(&tcbinfo); 922 923 m = syncache_pcbcount(); 924 925 error = sysctl_wire_old_buffer(req, 2 * (sizeof xig) 926 + (n + m) * sizeof(struct xtcpcb)); 927 if (error != 0) 928 return (error); 929 930 xig.xig_len = sizeof xig; 931 xig.xig_count = n + m; 932 xig.xig_gen = gencnt; 933 xig.xig_sogen = so_gencnt; 934 error = SYSCTL_OUT(req, &xig, sizeof xig); 935 if (error) 936 return (error); 937 938 error = syncache_pcblist(req, m, &pcb_count); 939 if (error) 940 return (error); 941 942 inp_list = malloc(n * sizeof *inp_list, M_TEMP, M_WAITOK); 943 if (inp_list == NULL) 944 return (ENOMEM); 945 946 INP_INFO_RLOCK(&tcbinfo); 947 for (inp = LIST_FIRST(tcbinfo.ipi_listhead), i = 0; inp != NULL && i 948 < n; inp = LIST_NEXT(inp, inp_list)) { 949 INP_RLOCK(inp); 950 if (inp->inp_gencnt <= gencnt) { 951 /* 952 * XXX: This use of cr_cansee(), introduced with 953 * TCP state changes, is not quite right, but for 954 * now, better than nothing. 955 */ 956 if (inp->inp_vflag & INP_TIMEWAIT) { 957 if (intotw(inp) != NULL) 958 error = cr_cansee(req->td->td_ucred, 959 intotw(inp)->tw_cred); 960 else 961 error = EINVAL; /* Skip this inp. */ 962 } else 963 error = cr_canseesocket(req->td->td_ucred, 964 inp->inp_socket); 965 if (error == 0) 966 inp_list[i++] = inp; 967 } 968 INP_RUNLOCK(inp); 969 } 970 INP_INFO_RUNLOCK(&tcbinfo); 971 n = i; 972 973 error = 0; 974 for (i = 0; i < n; i++) { 975 inp = inp_list[i]; 976 INP_RLOCK(inp); 977 if (inp->inp_gencnt <= gencnt) { 978 struct xtcpcb xt; 979 void *inp_ppcb; 980 981 bzero(&xt, sizeof(xt)); 982 xt.xt_len = sizeof xt; 983 /* XXX should avoid extra copy */ 984 bcopy(inp, &xt.xt_inp, sizeof *inp); 985 inp_ppcb = inp->inp_ppcb; 986 if (inp_ppcb == NULL) 987 bzero((char *) &xt.xt_tp, sizeof xt.xt_tp); 988 else if (inp->inp_vflag & INP_TIMEWAIT) { 989 bzero((char *) &xt.xt_tp, sizeof xt.xt_tp); 990 xt.xt_tp.t_state = TCPS_TIME_WAIT; 991 } else 992 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 993 if (inp->inp_socket != NULL) 994 sotoxsocket(inp->inp_socket, &xt.xt_socket); 995 else { 996 bzero(&xt.xt_socket, sizeof xt.xt_socket); 997 xt.xt_socket.xso_protocol = IPPROTO_TCP; 998 } 999 xt.xt_inp.inp_gencnt = inp->inp_gencnt; 1000 INP_RUNLOCK(inp); 1001 error = SYSCTL_OUT(req, &xt, sizeof xt); 1002 } else 1003 INP_RUNLOCK(inp); 1004 1005 } 1006 if (!error) { 1007 /* 1008 * Give the user an updated idea of our state. 1009 * If the generation differs from what we told 1010 * her before, she knows that something happened 1011 * while we were processing this request, and it 1012 * might be necessary to retry. 1013 */ 1014 INP_INFO_RLOCK(&tcbinfo); 1015 xig.xig_gen = tcbinfo.ipi_gencnt; 1016 xig.xig_sogen = so_gencnt; 1017 xig.xig_count = tcbinfo.ipi_count + pcb_count; 1018 INP_INFO_RUNLOCK(&tcbinfo); 1019 error = SYSCTL_OUT(req, &xig, sizeof xig); 1020 } 1021 free(inp_list, M_TEMP); 1022 return (error); 1023 } 1024 1025 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1026 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1027 1028 static int 1029 tcp_getcred(SYSCTL_HANDLER_ARGS) 1030 { 1031 struct xucred xuc; 1032 struct sockaddr_in addrs[2]; 1033 struct inpcb *inp; 1034 int error; 1035 1036 error = priv_check(req->td, PRIV_NETINET_GETCRED); 1037 if (error) 1038 return (error); 1039 error = SYSCTL_IN(req, addrs, sizeof(addrs)); 1040 if (error) 1041 return (error); 1042 INP_INFO_RLOCK(&tcbinfo); 1043 inp = in_pcblookup_hash(&tcbinfo, addrs[1].sin_addr, addrs[1].sin_port, 1044 addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1045 if (inp != NULL) { 1046 INP_RLOCK(inp); 1047 INP_INFO_RUNLOCK(&tcbinfo); 1048 if (inp->inp_socket == NULL) 1049 error = ENOENT; 1050 if (error == 0) 1051 error = cr_canseesocket(req->td->td_ucred, 1052 inp->inp_socket); 1053 if (error == 0) 1054 cru2x(inp->inp_socket->so_cred, &xuc); 1055 INP_RUNLOCK(inp); 1056 } else { 1057 INP_INFO_RUNLOCK(&tcbinfo); 1058 error = ENOENT; 1059 } 1060 if (error == 0) 1061 error = SYSCTL_OUT(req, &xuc, sizeof(struct xucred)); 1062 return (error); 1063 } 1064 1065 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, 1066 CTLTYPE_OPAQUE|CTLFLAG_RW|CTLFLAG_PRISON, 0, 0, 1067 tcp_getcred, "S,xucred", "Get the xucred of a TCP connection"); 1068 1069 #ifdef INET6 1070 static int 1071 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1072 { 1073 struct xucred xuc; 1074 struct sockaddr_in6 addrs[2]; 1075 struct inpcb *inp; 1076 int error, mapped = 0; 1077 1078 error = priv_check(req->td, PRIV_NETINET_GETCRED); 1079 if (error) 1080 return (error); 1081 error = SYSCTL_IN(req, addrs, sizeof(addrs)); 1082 if (error) 1083 return (error); 1084 if ((error = sa6_embedscope(&addrs[0], ip6_use_defzone)) != 0 || 1085 (error = sa6_embedscope(&addrs[1], ip6_use_defzone)) != 0) { 1086 return (error); 1087 } 1088 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { 1089 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) 1090 mapped = 1; 1091 else 1092 return (EINVAL); 1093 } 1094 1095 INP_INFO_RLOCK(&tcbinfo); 1096 if (mapped == 1) 1097 inp = in_pcblookup_hash(&tcbinfo, 1098 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], 1099 addrs[1].sin6_port, 1100 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], 1101 addrs[0].sin6_port, 1102 0, NULL); 1103 else 1104 inp = in6_pcblookup_hash(&tcbinfo, 1105 &addrs[1].sin6_addr, addrs[1].sin6_port, 1106 &addrs[0].sin6_addr, addrs[0].sin6_port, 0, NULL); 1107 if (inp != NULL) { 1108 INP_RLOCK(inp); 1109 INP_INFO_RUNLOCK(&tcbinfo); 1110 if (inp->inp_socket == NULL) 1111 error = ENOENT; 1112 if (error == 0) 1113 error = cr_canseesocket(req->td->td_ucred, 1114 inp->inp_socket); 1115 if (error == 0) 1116 cru2x(inp->inp_socket->so_cred, &xuc); 1117 INP_RUNLOCK(inp); 1118 } else { 1119 INP_INFO_RUNLOCK(&tcbinfo); 1120 error = ENOENT; 1121 } 1122 if (error == 0) 1123 error = SYSCTL_OUT(req, &xuc, sizeof(struct xucred)); 1124 return (error); 1125 } 1126 1127 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, 1128 CTLTYPE_OPAQUE|CTLFLAG_RW|CTLFLAG_PRISON, 0, 0, 1129 tcp6_getcred, "S,xucred", "Get the xucred of a TCP6 connection"); 1130 #endif 1131 1132 1133 void 1134 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip) 1135 { 1136 struct ip *ip = vip; 1137 struct tcphdr *th; 1138 struct in_addr faddr; 1139 struct inpcb *inp; 1140 struct tcpcb *tp; 1141 struct inpcb *(*notify)(struct inpcb *, int) = tcp_notify; 1142 struct icmp *icp; 1143 struct in_conninfo inc; 1144 tcp_seq icmp_tcp_seq; 1145 int mtu; 1146 1147 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1148 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1149 return; 1150 1151 if (cmd == PRC_MSGSIZE) 1152 notify = tcp_mtudisc; 1153 else if (icmp_may_rst && (cmd == PRC_UNREACH_ADMIN_PROHIB || 1154 cmd == PRC_UNREACH_PORT || cmd == PRC_TIMXCEED_INTRANS) && ip) 1155 notify = tcp_drop_syn_sent; 1156 /* 1157 * Redirects don't need to be handled up here. 1158 */ 1159 else if (PRC_IS_REDIRECT(cmd)) 1160 return; 1161 /* 1162 * Source quench is depreciated. 1163 */ 1164 else if (cmd == PRC_QUENCH) 1165 return; 1166 /* 1167 * Hostdead is ugly because it goes linearly through all PCBs. 1168 * XXX: We never get this from ICMP, otherwise it makes an 1169 * excellent DoS attack on machines with many connections. 1170 */ 1171 else if (cmd == PRC_HOSTDEAD) 1172 ip = NULL; 1173 else if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) 1174 return; 1175 if (ip != NULL) { 1176 icp = (struct icmp *)((caddr_t)ip 1177 - offsetof(struct icmp, icmp_ip)); 1178 th = (struct tcphdr *)((caddr_t)ip 1179 + (ip->ip_hl << 2)); 1180 INP_INFO_WLOCK(&tcbinfo); 1181 inp = in_pcblookup_hash(&tcbinfo, faddr, th->th_dport, 1182 ip->ip_src, th->th_sport, 0, NULL); 1183 if (inp != NULL) { 1184 INP_WLOCK(inp); 1185 if (!(inp->inp_vflag & INP_TIMEWAIT) && 1186 !(inp->inp_vflag & INP_DROPPED) && 1187 !(inp->inp_socket == NULL)) { 1188 icmp_tcp_seq = htonl(th->th_seq); 1189 tp = intotcpcb(inp); 1190 if (SEQ_GEQ(icmp_tcp_seq, tp->snd_una) && 1191 SEQ_LT(icmp_tcp_seq, tp->snd_max)) { 1192 if (cmd == PRC_MSGSIZE) { 1193 /* 1194 * MTU discovery: 1195 * If we got a needfrag set the MTU 1196 * in the route to the suggested new 1197 * value (if given) and then notify. 1198 */ 1199 bzero(&inc, sizeof(inc)); 1200 inc.inc_flags = 0; /* IPv4 */ 1201 inc.inc_faddr = faddr; 1202 inc.inc_fibnum = 1203 inp->inp_inc.inc_fibnum; 1204 1205 mtu = ntohs(icp->icmp_nextmtu); 1206 /* 1207 * If no alternative MTU was 1208 * proposed, try the next smaller 1209 * one. ip->ip_len has already 1210 * been swapped in icmp_input(). 1211 */ 1212 if (!mtu) 1213 mtu = ip_next_mtu(ip->ip_len, 1214 1); 1215 if (mtu < max(296, (tcp_minmss) 1216 + sizeof(struct tcpiphdr))) 1217 mtu = 0; 1218 if (!mtu) 1219 mtu = tcp_mssdflt 1220 + sizeof(struct tcpiphdr); 1221 /* 1222 * Only cache the the MTU if it 1223 * is smaller than the interface 1224 * or route MTU. tcp_mtudisc() 1225 * will do right thing by itself. 1226 */ 1227 if (mtu <= tcp_maxmtu(&inc, NULL)) 1228 tcp_hc_updatemtu(&inc, mtu); 1229 } 1230 1231 inp = (*notify)(inp, inetctlerrmap[cmd]); 1232 } 1233 } 1234 if (inp != NULL) 1235 INP_WUNLOCK(inp); 1236 } else { 1237 inc.inc_fport = th->th_dport; 1238 inc.inc_lport = th->th_sport; 1239 inc.inc_faddr = faddr; 1240 inc.inc_laddr = ip->ip_src; 1241 #ifdef INET6 1242 inc.inc_isipv6 = 0; 1243 #endif 1244 syncache_unreach(&inc, th); 1245 } 1246 INP_INFO_WUNLOCK(&tcbinfo); 1247 } else 1248 in_pcbnotifyall(&tcbinfo, faddr, inetctlerrmap[cmd], notify); 1249 } 1250 1251 #ifdef INET6 1252 void 1253 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d) 1254 { 1255 struct tcphdr th; 1256 struct inpcb *(*notify)(struct inpcb *, int) = tcp_notify; 1257 struct ip6_hdr *ip6; 1258 struct mbuf *m; 1259 struct ip6ctlparam *ip6cp = NULL; 1260 const struct sockaddr_in6 *sa6_src = NULL; 1261 int off; 1262 struct tcp_portonly { 1263 u_int16_t th_sport; 1264 u_int16_t th_dport; 1265 } *thp; 1266 1267 if (sa->sa_family != AF_INET6 || 1268 sa->sa_len != sizeof(struct sockaddr_in6)) 1269 return; 1270 1271 if (cmd == PRC_MSGSIZE) 1272 notify = tcp_mtudisc; 1273 else if (!PRC_IS_REDIRECT(cmd) && 1274 ((unsigned)cmd >= PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) 1275 return; 1276 /* Source quench is depreciated. */ 1277 else if (cmd == PRC_QUENCH) 1278 return; 1279 1280 /* if the parameter is from icmp6, decode it. */ 1281 if (d != NULL) { 1282 ip6cp = (struct ip6ctlparam *)d; 1283 m = ip6cp->ip6c_m; 1284 ip6 = ip6cp->ip6c_ip6; 1285 off = ip6cp->ip6c_off; 1286 sa6_src = ip6cp->ip6c_src; 1287 } else { 1288 m = NULL; 1289 ip6 = NULL; 1290 off = 0; /* fool gcc */ 1291 sa6_src = &sa6_any; 1292 } 1293 1294 if (ip6 != NULL) { 1295 struct in_conninfo inc; 1296 /* 1297 * XXX: We assume that when IPV6 is non NULL, 1298 * M and OFF are valid. 1299 */ 1300 1301 /* check if we can safely examine src and dst ports */ 1302 if (m->m_pkthdr.len < off + sizeof(*thp)) 1303 return; 1304 1305 bzero(&th, sizeof(th)); 1306 m_copydata(m, off, sizeof(*thp), (caddr_t)&th); 1307 1308 in6_pcbnotify(&tcbinfo, sa, th.th_dport, 1309 (struct sockaddr *)ip6cp->ip6c_src, 1310 th.th_sport, cmd, NULL, notify); 1311 1312 inc.inc_fport = th.th_dport; 1313 inc.inc_lport = th.th_sport; 1314 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1315 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1316 inc.inc_isipv6 = 1; 1317 INP_INFO_WLOCK(&tcbinfo); 1318 syncache_unreach(&inc, &th); 1319 INP_INFO_WUNLOCK(&tcbinfo); 1320 } else 1321 in6_pcbnotify(&tcbinfo, sa, 0, (const struct sockaddr *)sa6_src, 1322 0, cmd, NULL, notify); 1323 } 1324 #endif /* INET6 */ 1325 1326 1327 /* 1328 * Following is where TCP initial sequence number generation occurs. 1329 * 1330 * There are two places where we must use initial sequence numbers: 1331 * 1. In SYN-ACK packets. 1332 * 2. In SYN packets. 1333 * 1334 * All ISNs for SYN-ACK packets are generated by the syncache. See 1335 * tcp_syncache.c for details. 1336 * 1337 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1338 * depends on this property. In addition, these ISNs should be 1339 * unguessable so as to prevent connection hijacking. To satisfy 1340 * the requirements of this situation, the algorithm outlined in 1341 * RFC 1948 is used, with only small modifications. 1342 * 1343 * Implementation details: 1344 * 1345 * Time is based off the system timer, and is corrected so that it 1346 * increases by one megabyte per second. This allows for proper 1347 * recycling on high speed LANs while still leaving over an hour 1348 * before rollover. 1349 * 1350 * As reading the *exact* system time is too expensive to be done 1351 * whenever setting up a TCP connection, we increment the time 1352 * offset in two ways. First, a small random positive increment 1353 * is added to isn_offset for each connection that is set up. 1354 * Second, the function tcp_isn_tick fires once per clock tick 1355 * and increments isn_offset as necessary so that sequence numbers 1356 * are incremented at approximately ISN_BYTES_PER_SECOND. The 1357 * random positive increments serve only to ensure that the same 1358 * exact sequence number is never sent out twice (as could otherwise 1359 * happen when a port is recycled in less than the system tick 1360 * interval.) 1361 * 1362 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1363 * between seeding of isn_secret. This is normally set to zero, 1364 * as reseeding should not be necessary. 1365 * 1366 * Locking of the global variables isn_secret, isn_last_reseed, isn_offset, 1367 * isn_offset_old, and isn_ctx is performed using the TCP pcbinfo lock. In 1368 * general, this means holding an exclusive (write) lock. 1369 */ 1370 1371 #define ISN_BYTES_PER_SECOND 1048576 1372 #define ISN_STATIC_INCREMENT 4096 1373 #define ISN_RANDOM_INCREMENT (4096 - 1) 1374 1375 static u_char isn_secret[32]; 1376 static int isn_last_reseed; 1377 static u_int32_t isn_offset, isn_offset_old; 1378 static MD5_CTX isn_ctx; 1379 1380 tcp_seq 1381 tcp_new_isn(struct tcpcb *tp) 1382 { 1383 u_int32_t md5_buffer[4]; 1384 tcp_seq new_isn; 1385 1386 INP_WLOCK_ASSERT(tp->t_inpcb); 1387 1388 ISN_LOCK(); 1389 /* Seed if this is the first use, reseed if requested. */ 1390 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1391 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1392 < (u_int)ticks))) { 1393 read_random(&isn_secret, sizeof(isn_secret)); 1394 isn_last_reseed = ticks; 1395 } 1396 1397 /* Compute the md5 hash and return the ISN. */ 1398 MD5Init(&isn_ctx); 1399 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_fport, sizeof(u_short)); 1400 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_lport, sizeof(u_short)); 1401 #ifdef INET6 1402 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) { 1403 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1404 sizeof(struct in6_addr)); 1405 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1406 sizeof(struct in6_addr)); 1407 } else 1408 #endif 1409 { 1410 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1411 sizeof(struct in_addr)); 1412 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1413 sizeof(struct in_addr)); 1414 } 1415 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1416 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1417 new_isn = (tcp_seq) md5_buffer[0]; 1418 isn_offset += ISN_STATIC_INCREMENT + 1419 (arc4random() & ISN_RANDOM_INCREMENT); 1420 new_isn += isn_offset; 1421 ISN_UNLOCK(); 1422 return (new_isn); 1423 } 1424 1425 /* 1426 * Increment the offset to the next ISN_BYTES_PER_SECOND / 100 boundary 1427 * to keep time flowing at a relatively constant rate. If the random 1428 * increments have already pushed us past the projected offset, do nothing. 1429 */ 1430 static void 1431 tcp_isn_tick(void *xtp) 1432 { 1433 u_int32_t projected_offset; 1434 1435 ISN_LOCK(); 1436 projected_offset = isn_offset_old + ISN_BYTES_PER_SECOND / 100; 1437 1438 if (SEQ_GT(projected_offset, isn_offset)) 1439 isn_offset = projected_offset; 1440 1441 isn_offset_old = isn_offset; 1442 callout_reset(&isn_callout, hz/100, tcp_isn_tick, NULL); 1443 ISN_UNLOCK(); 1444 } 1445 1446 /* 1447 * When a specific ICMP unreachable message is received and the 1448 * connection state is SYN-SENT, drop the connection. This behavior 1449 * is controlled by the icmp_may_rst sysctl. 1450 */ 1451 struct inpcb * 1452 tcp_drop_syn_sent(struct inpcb *inp, int errno) 1453 { 1454 struct tcpcb *tp; 1455 1456 INP_INFO_WLOCK_ASSERT(&tcbinfo); 1457 INP_WLOCK_ASSERT(inp); 1458 1459 if ((inp->inp_vflag & INP_TIMEWAIT) || 1460 (inp->inp_vflag & INP_DROPPED)) 1461 return (inp); 1462 1463 tp = intotcpcb(inp); 1464 if (tp->t_state != TCPS_SYN_SENT) 1465 return (inp); 1466 1467 tp = tcp_drop(tp, errno); 1468 if (tp != NULL) 1469 return (inp); 1470 else 1471 return (NULL); 1472 } 1473 1474 /* 1475 * When `need fragmentation' ICMP is received, update our idea of the MSS 1476 * based on the new value in the route. Also nudge TCP to send something, 1477 * since we know the packet we just sent was dropped. 1478 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1479 */ 1480 struct inpcb * 1481 tcp_mtudisc(struct inpcb *inp, int errno) 1482 { 1483 struct tcpcb *tp; 1484 struct socket *so = inp->inp_socket; 1485 u_int maxmtu; 1486 u_int romtu; 1487 int mss; 1488 #ifdef INET6 1489 int isipv6; 1490 #endif /* INET6 */ 1491 1492 INP_WLOCK_ASSERT(inp); 1493 if ((inp->inp_vflag & INP_TIMEWAIT) || 1494 (inp->inp_vflag & INP_DROPPED)) 1495 return (inp); 1496 1497 tp = intotcpcb(inp); 1498 KASSERT(tp != NULL, ("tcp_mtudisc: tp == NULL")); 1499 1500 #ifdef INET6 1501 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 1502 #endif 1503 maxmtu = tcp_hc_getmtu(&inp->inp_inc); /* IPv4 and IPv6 */ 1504 romtu = 1505 #ifdef INET6 1506 isipv6 ? tcp_maxmtu6(&inp->inp_inc, NULL) : 1507 #endif /* INET6 */ 1508 tcp_maxmtu(&inp->inp_inc, NULL); 1509 if (!maxmtu) 1510 maxmtu = romtu; 1511 else 1512 maxmtu = min(maxmtu, romtu); 1513 if (!maxmtu) { 1514 tp->t_maxopd = tp->t_maxseg = 1515 #ifdef INET6 1516 isipv6 ? tcp_v6mssdflt : 1517 #endif /* INET6 */ 1518 tcp_mssdflt; 1519 return (inp); 1520 } 1521 mss = maxmtu - 1522 #ifdef INET6 1523 (isipv6 ? sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1524 #endif /* INET6 */ 1525 sizeof(struct tcpiphdr) 1526 #ifdef INET6 1527 ) 1528 #endif /* INET6 */ 1529 ; 1530 1531 /* 1532 * XXX - The above conditional probably violates the TCP 1533 * spec. The problem is that, since we don't know the 1534 * other end's MSS, we are supposed to use a conservative 1535 * default. But, if we do that, then MTU discovery will 1536 * never actually take place, because the conservative 1537 * default is much less than the MTUs typically seen 1538 * on the Internet today. For the moment, we'll sweep 1539 * this under the carpet. 1540 * 1541 * The conservative default might not actually be a problem 1542 * if the only case this occurs is when sending an initial 1543 * SYN with options and data to a host we've never talked 1544 * to before. Then, they will reply with an MSS value which 1545 * will get recorded and the new parameters should get 1546 * recomputed. For Further Study. 1547 */ 1548 if (tp->t_maxopd <= mss) 1549 return (inp); 1550 tp->t_maxopd = mss; 1551 1552 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP && 1553 (tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP) 1554 mss -= TCPOLEN_TSTAMP_APPA; 1555 #if (MCLBYTES & (MCLBYTES - 1)) == 0 1556 if (mss > MCLBYTES) 1557 mss &= ~(MCLBYTES-1); 1558 #else 1559 if (mss > MCLBYTES) 1560 mss = mss / MCLBYTES * MCLBYTES; 1561 #endif 1562 if (so->so_snd.sb_hiwat < mss) 1563 mss = so->so_snd.sb_hiwat; 1564 1565 tp->t_maxseg = mss; 1566 1567 tcpstat.tcps_mturesent++; 1568 tp->t_rtttime = 0; 1569 tp->snd_nxt = tp->snd_una; 1570 tcp_free_sackholes(tp); 1571 tp->snd_recover = tp->snd_max; 1572 if (tp->t_flags & TF_SACK_PERMIT) 1573 EXIT_FASTRECOVERY(tp); 1574 tcp_output_send(tp); 1575 return (inp); 1576 } 1577 1578 /* 1579 * Look-up the routing entry to the peer of this inpcb. If no route 1580 * is found and it cannot be allocated, then return NULL. This routine 1581 * is called by TCP routines that access the rmx structure and by tcp_mss 1582 * to get the interface MTU. 1583 */ 1584 u_long 1585 tcp_maxmtu(struct in_conninfo *inc, int *flags) 1586 { 1587 struct route sro; 1588 struct sockaddr_in *dst; 1589 struct ifnet *ifp; 1590 u_long maxmtu = 0; 1591 1592 KASSERT(inc != NULL, ("tcp_maxmtu with NULL in_conninfo pointer")); 1593 1594 bzero(&sro, sizeof(sro)); 1595 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1596 dst = (struct sockaddr_in *)&sro.ro_dst; 1597 dst->sin_family = AF_INET; 1598 dst->sin_len = sizeof(*dst); 1599 dst->sin_addr = inc->inc_faddr; 1600 in_rtalloc_ign(&sro, RTF_CLONING, inc->inc_fibnum); 1601 } 1602 if (sro.ro_rt != NULL) { 1603 ifp = sro.ro_rt->rt_ifp; 1604 if (sro.ro_rt->rt_rmx.rmx_mtu == 0) 1605 maxmtu = ifp->if_mtu; 1606 else 1607 maxmtu = min(sro.ro_rt->rt_rmx.rmx_mtu, ifp->if_mtu); 1608 1609 /* Report additional interface capabilities. */ 1610 if (flags != NULL) { 1611 if (ifp->if_capenable & IFCAP_TSO4 && 1612 ifp->if_hwassist & CSUM_TSO) 1613 *flags |= CSUM_TSO; 1614 } 1615 RTFREE(sro.ro_rt); 1616 } 1617 return (maxmtu); 1618 } 1619 1620 #ifdef INET6 1621 u_long 1622 tcp_maxmtu6(struct in_conninfo *inc, int *flags) 1623 { 1624 struct route_in6 sro6; 1625 struct ifnet *ifp; 1626 u_long maxmtu = 0; 1627 1628 KASSERT(inc != NULL, ("tcp_maxmtu6 with NULL in_conninfo pointer")); 1629 1630 bzero(&sro6, sizeof(sro6)); 1631 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1632 sro6.ro_dst.sin6_family = AF_INET6; 1633 sro6.ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1634 sro6.ro_dst.sin6_addr = inc->inc6_faddr; 1635 rtalloc_ign((struct route *)&sro6, RTF_CLONING); 1636 } 1637 if (sro6.ro_rt != NULL) { 1638 ifp = sro6.ro_rt->rt_ifp; 1639 if (sro6.ro_rt->rt_rmx.rmx_mtu == 0) 1640 maxmtu = IN6_LINKMTU(sro6.ro_rt->rt_ifp); 1641 else 1642 maxmtu = min(sro6.ro_rt->rt_rmx.rmx_mtu, 1643 IN6_LINKMTU(sro6.ro_rt->rt_ifp)); 1644 1645 /* Report additional interface capabilities. */ 1646 if (flags != NULL) { 1647 if (ifp->if_capenable & IFCAP_TSO6 && 1648 ifp->if_hwassist & CSUM_TSO) 1649 *flags |= CSUM_TSO; 1650 } 1651 RTFREE(sro6.ro_rt); 1652 } 1653 1654 return (maxmtu); 1655 } 1656 #endif /* INET6 */ 1657 1658 #ifdef IPSEC 1659 /* compute ESP/AH header size for TCP, including outer IP header. */ 1660 size_t 1661 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1662 { 1663 struct inpcb *inp; 1664 struct mbuf *m; 1665 size_t hdrsiz; 1666 struct ip *ip; 1667 #ifdef INET6 1668 struct ip6_hdr *ip6; 1669 #endif 1670 struct tcphdr *th; 1671 1672 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1673 return (0); 1674 MGETHDR(m, M_DONTWAIT, MT_DATA); 1675 if (!m) 1676 return (0); 1677 1678 #ifdef INET6 1679 if ((inp->inp_vflag & INP_IPV6) != 0) { 1680 ip6 = mtod(m, struct ip6_hdr *); 1681 th = (struct tcphdr *)(ip6 + 1); 1682 m->m_pkthdr.len = m->m_len = 1683 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1684 tcpip_fillheaders(inp, ip6, th); 1685 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1686 } else 1687 #endif /* INET6 */ 1688 { 1689 ip = mtod(m, struct ip *); 1690 th = (struct tcphdr *)(ip + 1); 1691 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1692 tcpip_fillheaders(inp, ip, th); 1693 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1694 } 1695 1696 m_free(m); 1697 return (hdrsiz); 1698 } 1699 #endif /* IPSEC */ 1700 1701 /* 1702 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1703 * 1704 * This code attempts to calculate the bandwidth-delay product as a 1705 * means of determining the optimal window size to maximize bandwidth, 1706 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1707 * routers. This code also does a fairly good job keeping RTTs in check 1708 * across slow links like modems. We implement an algorithm which is very 1709 * similar (but not meant to be) TCP/Vegas. The code operates on the 1710 * transmitter side of a TCP connection and so only effects the transmit 1711 * side of the connection. 1712 * 1713 * BACKGROUND: TCP makes no provision for the management of buffer space 1714 * at the end points or at the intermediate routers and switches. A TCP 1715 * stream, whether using NewReno or not, will eventually buffer as 1716 * many packets as it is able and the only reason this typically works is 1717 * due to the fairly small default buffers made available for a connection 1718 * (typicaly 16K or 32K). As machines use larger windows and/or window 1719 * scaling it is now fairly easy for even a single TCP connection to blow-out 1720 * all available buffer space not only on the local interface, but on 1721 * intermediate routers and switches as well. NewReno makes a misguided 1722 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1723 * then backing off, then steadily increasing the window again until another 1724 * failure occurs, ad-infinitum. This results in terrible oscillation that 1725 * is only made worse as network loads increase and the idea of intentionally 1726 * blowing out network buffers is, frankly, a terrible way to manage network 1727 * resources. 1728 * 1729 * It is far better to limit the transmit window prior to the failure 1730 * condition being achieved. There are two general ways to do this: First 1731 * you can 'scan' through different transmit window sizes and locate the 1732 * point where the RTT stops increasing, indicating that you have filled the 1733 * pipe, then scan backwards until you note that RTT stops decreasing, then 1734 * repeat ad-infinitum. This method works in principle but has severe 1735 * implementation issues due to RTT variances, timer granularity, and 1736 * instability in the algorithm which can lead to many false positives and 1737 * create oscillations as well as interact badly with other TCP streams 1738 * implementing the same algorithm. 1739 * 1740 * The second method is to limit the window to the bandwidth delay product 1741 * of the link. This is the method we implement. RTT variances and our 1742 * own manipulation of the congestion window, bwnd, can potentially 1743 * destabilize the algorithm. For this reason we have to stabilize the 1744 * elements used to calculate the window. We do this by using the minimum 1745 * observed RTT, the long term average of the observed bandwidth, and 1746 * by adding two segments worth of slop. It isn't perfect but it is able 1747 * to react to changing conditions and gives us a very stable basis on 1748 * which to extend the algorithm. 1749 */ 1750 void 1751 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1752 { 1753 u_long bw; 1754 u_long bwnd; 1755 int save_ticks; 1756 1757 INP_WLOCK_ASSERT(tp->t_inpcb); 1758 1759 /* 1760 * If inflight_enable is disabled in the middle of a tcp connection, 1761 * make sure snd_bwnd is effectively disabled. 1762 */ 1763 if (tcp_inflight_enable == 0 || tp->t_rttlow < tcp_inflight_rttthresh) { 1764 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1765 tp->snd_bandwidth = 0; 1766 return; 1767 } 1768 1769 /* 1770 * Figure out the bandwidth. Due to the tick granularity this 1771 * is a very rough number and it MUST be averaged over a fairly 1772 * long period of time. XXX we need to take into account a link 1773 * that is not using all available bandwidth, but for now our 1774 * slop will ramp us up if this case occurs and the bandwidth later 1775 * increases. 1776 * 1777 * Note: if ticks rollover 'bw' may wind up negative. We must 1778 * effectively reset t_bw_rtttime for this case. 1779 */ 1780 save_ticks = ticks; 1781 if ((u_int)(save_ticks - tp->t_bw_rtttime) < 1) 1782 return; 1783 1784 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / 1785 (save_ticks - tp->t_bw_rtttime); 1786 tp->t_bw_rtttime = save_ticks; 1787 tp->t_bw_rtseq = ack_seq; 1788 if (tp->t_bw_rtttime == 0 || (int)bw < 0) 1789 return; 1790 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4; 1791 1792 tp->snd_bandwidth = bw; 1793 1794 /* 1795 * Calculate the semi-static bandwidth delay product, plus two maximal 1796 * segments. The additional slop puts us squarely in the sweet 1797 * spot and also handles the bandwidth run-up case and stabilization. 1798 * Without the slop we could be locking ourselves into a lower 1799 * bandwidth. 1800 * 1801 * Situations Handled: 1802 * (1) Prevents over-queueing of packets on LANs, especially on 1803 * high speed LANs, allowing larger TCP buffers to be 1804 * specified, and also does a good job preventing 1805 * over-queueing of packets over choke points like modems 1806 * (at least for the transmit side). 1807 * 1808 * (2) Is able to handle changing network loads (bandwidth 1809 * drops so bwnd drops, bandwidth increases so bwnd 1810 * increases). 1811 * 1812 * (3) Theoretically should stabilize in the face of multiple 1813 * connections implementing the same algorithm (this may need 1814 * a little work). 1815 * 1816 * (4) Stability value (defaults to 20 = 2 maximal packets) can 1817 * be adjusted with a sysctl but typically only needs to be 1818 * on very slow connections. A value no smaller then 5 1819 * should be used, but only reduce this default if you have 1820 * no other choice. 1821 */ 1822 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2) 1823 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + tcp_inflight_stab * tp->t_maxseg / 10; 1824 #undef USERTT 1825 1826 if (tcp_inflight_debug > 0) { 1827 static int ltime; 1828 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) { 1829 ltime = ticks; 1830 printf("%p bw %ld rttbest %d srtt %d bwnd %ld\n", 1831 tp, 1832 bw, 1833 tp->t_rttbest, 1834 tp->t_srtt, 1835 bwnd 1836 ); 1837 } 1838 } 1839 if ((long)bwnd < tcp_inflight_min) 1840 bwnd = tcp_inflight_min; 1841 if (bwnd > tcp_inflight_max) 1842 bwnd = tcp_inflight_max; 1843 if ((long)bwnd < tp->t_maxseg * 2) 1844 bwnd = tp->t_maxseg * 2; 1845 tp->snd_bwnd = bwnd; 1846 } 1847 1848 #ifdef TCP_SIGNATURE 1849 /* 1850 * Callback function invoked by m_apply() to digest TCP segment data 1851 * contained within an mbuf chain. 1852 */ 1853 static int 1854 tcp_signature_apply(void *fstate, void *data, u_int len) 1855 { 1856 1857 MD5Update(fstate, (u_char *)data, len); 1858 return (0); 1859 } 1860 1861 /* 1862 * Compute TCP-MD5 hash of a TCPv4 segment. (RFC2385) 1863 * 1864 * Parameters: 1865 * m pointer to head of mbuf chain 1866 * off0 offset to TCP header within the mbuf chain 1867 * len length of TCP segment data, excluding options 1868 * optlen length of TCP segment options 1869 * buf pointer to storage for computed MD5 digest 1870 * direction direction of flow (IPSEC_DIR_INBOUND or OUTBOUND) 1871 * 1872 * We do this over ip, tcphdr, segment data, and the key in the SADB. 1873 * When called from tcp_input(), we can be sure that th_sum has been 1874 * zeroed out and verified already. 1875 * 1876 * This function is for IPv4 use only. Calling this function with an 1877 * IPv6 packet in the mbuf chain will yield undefined results. 1878 * 1879 * Return 0 if successful, otherwise return -1. 1880 * 1881 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a 1882 * search with the destination IP address, and a 'magic SPI' to be 1883 * determined by the application. This is hardcoded elsewhere to 1179 1884 * right now. Another branch of this code exists which uses the SPD to 1885 * specify per-application flows but it is unstable. 1886 */ 1887 int 1888 tcp_signature_compute(struct mbuf *m, int off0, int len, int optlen, 1889 u_char *buf, u_int direction) 1890 { 1891 union sockaddr_union dst; 1892 struct ippseudo ippseudo; 1893 MD5_CTX ctx; 1894 int doff; 1895 struct ip *ip; 1896 struct ipovly *ipovly; 1897 struct secasvar *sav; 1898 struct tcphdr *th; 1899 u_short savecsum; 1900 1901 KASSERT(m != NULL, ("NULL mbuf chain")); 1902 KASSERT(buf != NULL, ("NULL signature pointer")); 1903 1904 /* Extract the destination from the IP header in the mbuf. */ 1905 ip = mtod(m, struct ip *); 1906 bzero(&dst, sizeof(union sockaddr_union)); 1907 dst.sa.sa_len = sizeof(struct sockaddr_in); 1908 dst.sa.sa_family = AF_INET; 1909 dst.sin.sin_addr = (direction == IPSEC_DIR_INBOUND) ? 1910 ip->ip_src : ip->ip_dst; 1911 1912 /* Look up an SADB entry which matches the address of the peer. */ 1913 sav = KEY_ALLOCSA(&dst, IPPROTO_TCP, htonl(TCP_SIG_SPI)); 1914 if (sav == NULL) { 1915 printf("%s: SADB lookup failed for %s\n", __func__, 1916 inet_ntoa(dst.sin.sin_addr)); 1917 return (EINVAL); 1918 } 1919 1920 MD5Init(&ctx); 1921 ipovly = (struct ipovly *)ip; 1922 th = (struct tcphdr *)((u_char *)ip + off0); 1923 doff = off0 + sizeof(struct tcphdr) + optlen; 1924 1925 /* 1926 * Step 1: Update MD5 hash with IP pseudo-header. 1927 * 1928 * XXX The ippseudo header MUST be digested in network byte order, 1929 * or else we'll fail the regression test. Assume all fields we've 1930 * been doing arithmetic on have been in host byte order. 1931 * XXX One cannot depend on ipovly->ih_len here. When called from 1932 * tcp_output(), the underlying ip_len member has not yet been set. 1933 */ 1934 ippseudo.ippseudo_src = ipovly->ih_src; 1935 ippseudo.ippseudo_dst = ipovly->ih_dst; 1936 ippseudo.ippseudo_pad = 0; 1937 ippseudo.ippseudo_p = IPPROTO_TCP; 1938 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen); 1939 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo)); 1940 1941 /* 1942 * Step 2: Update MD5 hash with TCP header, excluding options. 1943 * The TCP checksum must be set to zero. 1944 */ 1945 savecsum = th->th_sum; 1946 th->th_sum = 0; 1947 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr)); 1948 th->th_sum = savecsum; 1949 1950 /* 1951 * Step 3: Update MD5 hash with TCP segment data. 1952 * Use m_apply() to avoid an early m_pullup(). 1953 */ 1954 if (len > 0) 1955 m_apply(m, doff, len, tcp_signature_apply, &ctx); 1956 1957 /* 1958 * Step 4: Update MD5 hash with shared secret. 1959 */ 1960 MD5Update(&ctx, sav->key_auth->key_data, _KEYLEN(sav->key_auth)); 1961 MD5Final(buf, &ctx); 1962 1963 key_sa_recordxfer(sav, m); 1964 KEY_FREESAV(&sav); 1965 return (0); 1966 } 1967 #endif /* TCP_SIGNATURE */ 1968 1969 static int 1970 sysctl_drop(SYSCTL_HANDLER_ARGS) 1971 { 1972 /* addrs[0] is a foreign socket, addrs[1] is a local one. */ 1973 struct sockaddr_storage addrs[2]; 1974 struct inpcb *inp; 1975 struct tcpcb *tp; 1976 struct tcptw *tw; 1977 struct sockaddr_in *fin, *lin; 1978 #ifdef INET6 1979 struct sockaddr_in6 *fin6, *lin6; 1980 struct in6_addr f6, l6; 1981 #endif 1982 int error; 1983 1984 inp = NULL; 1985 fin = lin = NULL; 1986 #ifdef INET6 1987 fin6 = lin6 = NULL; 1988 #endif 1989 error = 0; 1990 1991 if (req->oldptr != NULL || req->oldlen != 0) 1992 return (EINVAL); 1993 if (req->newptr == NULL) 1994 return (EPERM); 1995 if (req->newlen < sizeof(addrs)) 1996 return (ENOMEM); 1997 error = SYSCTL_IN(req, &addrs, sizeof(addrs)); 1998 if (error) 1999 return (error); 2000 2001 switch (addrs[0].ss_family) { 2002 #ifdef INET6 2003 case AF_INET6: 2004 fin6 = (struct sockaddr_in6 *)&addrs[0]; 2005 lin6 = (struct sockaddr_in6 *)&addrs[1]; 2006 if (fin6->sin6_len != sizeof(struct sockaddr_in6) || 2007 lin6->sin6_len != sizeof(struct sockaddr_in6)) 2008 return (EINVAL); 2009 if (IN6_IS_ADDR_V4MAPPED(&fin6->sin6_addr)) { 2010 if (!IN6_IS_ADDR_V4MAPPED(&lin6->sin6_addr)) 2011 return (EINVAL); 2012 in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[0]); 2013 in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[1]); 2014 fin = (struct sockaddr_in *)&addrs[0]; 2015 lin = (struct sockaddr_in *)&addrs[1]; 2016 break; 2017 } 2018 error = sa6_embedscope(fin6, ip6_use_defzone); 2019 if (error) 2020 return (error); 2021 error = sa6_embedscope(lin6, ip6_use_defzone); 2022 if (error) 2023 return (error); 2024 break; 2025 #endif 2026 case AF_INET: 2027 fin = (struct sockaddr_in *)&addrs[0]; 2028 lin = (struct sockaddr_in *)&addrs[1]; 2029 if (fin->sin_len != sizeof(struct sockaddr_in) || 2030 lin->sin_len != sizeof(struct sockaddr_in)) 2031 return (EINVAL); 2032 break; 2033 default: 2034 return (EINVAL); 2035 } 2036 INP_INFO_WLOCK(&tcbinfo); 2037 switch (addrs[0].ss_family) { 2038 #ifdef INET6 2039 case AF_INET6: 2040 inp = in6_pcblookup_hash(&tcbinfo, &f6, fin6->sin6_port, 2041 &l6, lin6->sin6_port, 0, NULL); 2042 break; 2043 #endif 2044 case AF_INET: 2045 inp = in_pcblookup_hash(&tcbinfo, fin->sin_addr, fin->sin_port, 2046 lin->sin_addr, lin->sin_port, 0, NULL); 2047 break; 2048 } 2049 if (inp != NULL) { 2050 INP_WLOCK(inp); 2051 if (inp->inp_vflag & INP_TIMEWAIT) { 2052 /* 2053 * XXXRW: There currently exists a state where an 2054 * inpcb is present, but its timewait state has been 2055 * discarded. For now, don't allow dropping of this 2056 * type of inpcb. 2057 */ 2058 tw = intotw(inp); 2059 if (tw != NULL) 2060 tcp_twclose(tw, 0); 2061 else 2062 INP_WUNLOCK(inp); 2063 } else if (!(inp->inp_vflag & INP_DROPPED) && 2064 !(inp->inp_socket->so_options & SO_ACCEPTCONN)) { 2065 tp = intotcpcb(inp); 2066 tp = tcp_drop(tp, ECONNABORTED); 2067 if (tp != NULL) 2068 INP_WUNLOCK(inp); 2069 } else 2070 INP_WUNLOCK(inp); 2071 } else 2072 error = ESRCH; 2073 INP_INFO_WUNLOCK(&tcbinfo); 2074 return (error); 2075 } 2076 2077 SYSCTL_PROC(_net_inet_tcp, TCPCTL_DROP, drop, 2078 CTLTYPE_STRUCT|CTLFLAG_WR|CTLFLAG_SKIP, NULL, 2079 0, sysctl_drop, "", "Drop TCP connection"); 2080 2081 /* 2082 * Generate a standardized TCP log line for use throughout the 2083 * tcp subsystem. Memory allocation is done with M_NOWAIT to 2084 * allow use in the interrupt context. 2085 * 2086 * NB: The caller MUST free(s, M_TCPLOG) the returned string. 2087 * NB: The function may return NULL if memory allocation failed. 2088 * 2089 * Due to header inclusion and ordering limitations the struct ip 2090 * and ip6_hdr pointers have to be passed as void pointers. 2091 */ 2092 char * 2093 tcp_log_addrs(struct in_conninfo *inc, struct tcphdr *th, void *ip4hdr, 2094 const void *ip6hdr) 2095 { 2096 char *s, *sp; 2097 size_t size; 2098 struct ip *ip; 2099 #ifdef INET6 2100 const struct ip6_hdr *ip6; 2101 2102 ip6 = (const struct ip6_hdr *)ip6hdr; 2103 #endif /* INET6 */ 2104 ip = (struct ip *)ip4hdr; 2105 2106 /* 2107 * The log line looks like this: 2108 * "TCP: [1.2.3.4]:50332 to [1.2.3.4]:80 tcpflags 0x2<SYN>" 2109 */ 2110 size = sizeof("TCP: []:12345 to []:12345 tcpflags 0x2<>") + 2111 sizeof(PRINT_TH_FLAGS) + 1 + 2112 #ifdef INET6 2113 2 * INET6_ADDRSTRLEN; 2114 #else 2115 2 * INET_ADDRSTRLEN; 2116 #endif /* INET6 */ 2117 2118 /* Is logging enabled? */ 2119 if (tcp_log_debug == 0 && tcp_log_in_vain == 0) 2120 return (NULL); 2121 2122 s = malloc(size, M_TCPLOG, M_ZERO|M_NOWAIT); 2123 if (s == NULL) 2124 return (NULL); 2125 2126 strcat(s, "TCP: ["); 2127 sp = s + strlen(s); 2128 2129 if (inc && inc->inc_isipv6 == 0) { 2130 inet_ntoa_r(inc->inc_faddr, sp); 2131 sp = s + strlen(s); 2132 sprintf(sp, "]:%i to [", ntohs(inc->inc_fport)); 2133 sp = s + strlen(s); 2134 inet_ntoa_r(inc->inc_laddr, sp); 2135 sp = s + strlen(s); 2136 sprintf(sp, "]:%i", ntohs(inc->inc_lport)); 2137 #ifdef INET6 2138 } else if (inc) { 2139 ip6_sprintf(sp, &inc->inc6_faddr); 2140 sp = s + strlen(s); 2141 sprintf(sp, "]:%i to [", ntohs(inc->inc_fport)); 2142 sp = s + strlen(s); 2143 ip6_sprintf(sp, &inc->inc6_laddr); 2144 sp = s + strlen(s); 2145 sprintf(sp, "]:%i", ntohs(inc->inc_lport)); 2146 } else if (ip6 && th) { 2147 ip6_sprintf(sp, &ip6->ip6_src); 2148 sp = s + strlen(s); 2149 sprintf(sp, "]:%i to [", ntohs(th->th_sport)); 2150 sp = s + strlen(s); 2151 ip6_sprintf(sp, &ip6->ip6_dst); 2152 sp = s + strlen(s); 2153 sprintf(sp, "]:%i", ntohs(th->th_dport)); 2154 #endif /* INET6 */ 2155 } else if (ip && th) { 2156 inet_ntoa_r(ip->ip_src, sp); 2157 sp = s + strlen(s); 2158 sprintf(sp, "]:%i to [", ntohs(th->th_sport)); 2159 sp = s + strlen(s); 2160 inet_ntoa_r(ip->ip_dst, sp); 2161 sp = s + strlen(s); 2162 sprintf(sp, "]:%i", ntohs(th->th_dport)); 2163 } else { 2164 free(s, M_TCPLOG); 2165 return (NULL); 2166 } 2167 sp = s + strlen(s); 2168 if (th) 2169 sprintf(sp, " tcpflags 0x%b", th->th_flags, PRINT_TH_FLAGS); 2170 if (*(s + size - 1) != '\0') 2171 panic("%s: string too long", __func__); 2172 return (s); 2173 } 2174