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