1 /* 2 * Copyright (c) 2003, 2004 Jeffrey M. Hsu. All rights reserved. 3 * Copyright (c) 2003, 2004 The DragonFly Project. All rights reserved. 4 * 5 * This code is derived from software contributed to The DragonFly Project 6 * by Jeffrey M. Hsu. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. Neither the name of The DragonFly Project nor the names of its 17 * contributors may be used to endorse or promote products derived 18 * from this software without specific, prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 22 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 23 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 24 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 25 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 26 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 27 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 28 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 29 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 30 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 */ 33 34 /* 35 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 36 * The Regents of the University of California. All rights reserved. 37 * 38 * Redistribution and use in source and binary forms, with or without 39 * modification, are permitted provided that the following conditions 40 * are met: 41 * 1. Redistributions of source code must retain the above copyright 42 * notice, this list of conditions and the following disclaimer. 43 * 2. Redistributions in binary form must reproduce the above copyright 44 * notice, this list of conditions and the following disclaimer in the 45 * documentation and/or other materials provided with the distribution. 46 * 3. All advertising materials mentioning features or use of this software 47 * must display the following acknowledgement: 48 * This product includes software developed by the University of 49 * California, Berkeley and its contributors. 50 * 4. Neither the name of the University nor the names of its contributors 51 * may be used to endorse or promote products derived from this software 52 * without specific prior written permission. 53 * 54 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 55 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 56 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 57 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 58 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 59 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 60 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 61 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 62 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 63 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 64 * SUCH DAMAGE. 65 * 66 * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95 67 * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $ 68 */ 69 70 #include "opt_compat.h" 71 #include "opt_inet.h" 72 #include "opt_inet6.h" 73 #include "opt_ipsec.h" 74 #include "opt_tcpdebug.h" 75 76 #include <sys/param.h> 77 #include <sys/systm.h> 78 #include <sys/callout.h> 79 #include <sys/kernel.h> 80 #include <sys/sysctl.h> 81 #include <sys/malloc.h> 82 #include <sys/mpipe.h> 83 #include <sys/mbuf.h> 84 #ifdef INET6 85 #include <sys/domain.h> 86 #endif 87 #include <sys/proc.h> 88 #include <sys/priv.h> 89 #include <sys/socket.h> 90 #include <sys/socketvar.h> 91 #include <sys/protosw.h> 92 #include <sys/random.h> 93 #include <sys/in_cksum.h> 94 #include <sys/ktr.h> 95 96 #include <net/route.h> 97 #include <net/if.h> 98 #include <net/netisr.h> 99 100 #define _IP_VHL 101 #include <netinet/in.h> 102 #include <netinet/in_systm.h> 103 #include <netinet/ip.h> 104 #include <netinet/ip6.h> 105 #include <netinet/in_pcb.h> 106 #include <netinet6/in6_pcb.h> 107 #include <netinet/in_var.h> 108 #include <netinet/ip_var.h> 109 #include <netinet6/ip6_var.h> 110 #include <netinet/ip_icmp.h> 111 #ifdef INET6 112 #include <netinet/icmp6.h> 113 #endif 114 #include <netinet/tcp.h> 115 #include <netinet/tcp_fsm.h> 116 #include <netinet/tcp_seq.h> 117 #include <netinet/tcp_timer.h> 118 #include <netinet/tcp_timer2.h> 119 #include <netinet/tcp_var.h> 120 #include <netinet6/tcp6_var.h> 121 #include <netinet/tcpip.h> 122 #ifdef TCPDEBUG 123 #include <netinet/tcp_debug.h> 124 #endif 125 #include <netinet6/ip6protosw.h> 126 127 #ifdef IPSEC 128 #include <netinet6/ipsec.h> 129 #include <netproto/key/key.h> 130 #ifdef INET6 131 #include <netinet6/ipsec6.h> 132 #endif 133 #endif 134 135 #ifdef FAST_IPSEC 136 #include <netproto/ipsec/ipsec.h> 137 #ifdef INET6 138 #include <netproto/ipsec/ipsec6.h> 139 #endif 140 #define IPSEC 141 #endif 142 143 #include <sys/md5.h> 144 #include <machine/smp.h> 145 146 #include <sys/msgport2.h> 147 #include <sys/mplock2.h> 148 #include <net/netmsg2.h> 149 150 #if !defined(KTR_TCP) 151 #define KTR_TCP KTR_ALL 152 #endif 153 /* 154 KTR_INFO_MASTER(tcp); 155 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0); 156 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0); 157 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0); 158 #define logtcp(name) KTR_LOG(tcp_ ## name) 159 */ 160 161 #define TCP_IW_MAXSEGS_DFLT 4 162 #define TCP_IW_CAPSEGS_DFLT 3 163 164 struct inpcbinfo tcbinfo[MAXCPU]; 165 struct tcpcbackqhead tcpcbackq[MAXCPU]; 166 167 static struct lwkt_token tcp_port_token = 168 LWKT_TOKEN_INITIALIZER(tcp_port_token); 169 170 int tcp_mssdflt = TCP_MSS; 171 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 172 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size"); 173 174 #ifdef INET6 175 int tcp_v6mssdflt = TCP6_MSS; 176 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW, 177 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6"); 178 #endif 179 180 /* 181 * Minimum MSS we accept and use. This prevents DoS attacks where 182 * we are forced to a ridiculous low MSS like 20 and send hundreds 183 * of packets instead of one. The effect scales with the available 184 * bandwidth and quickly saturates the CPU and network interface 185 * with packet generation and sending. Set to zero to disable MINMSS 186 * checking. This setting prevents us from sending too small packets. 187 */ 188 int tcp_minmss = TCP_MINMSS; 189 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW, 190 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size"); 191 192 #if 0 193 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ; 194 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW, 195 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time"); 196 #endif 197 198 int tcp_do_rfc1323 = 1; 199 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW, 200 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions"); 201 202 static int tcp_tcbhashsize = 0; 203 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD, 204 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable"); 205 206 static int do_tcpdrain = 1; 207 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0, 208 "Enable tcp_drain routine for extra help when low on mbufs"); 209 210 static int icmp_may_rst = 1; 211 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0, 212 "Certain ICMP unreachable messages may abort connections in SYN_SENT"); 213 214 static int tcp_isn_reseed_interval = 0; 215 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW, 216 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret"); 217 218 /* 219 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on 220 * by default, but with generous values which should allow maximal 221 * bandwidth. In particular, the slop defaults to 50 (5 packets). 222 * 223 * The reason for doing this is that the limiter is the only mechanism we 224 * have which seems to do a really good job preventing receiver RX rings 225 * on network interfaces from getting blown out. Even though GigE/10GigE 226 * is supposed to flow control it looks like either it doesn't actually 227 * do it or Open Source drivers do not properly enable it. 228 * 229 * People using the limiter to reduce bottlenecks on slower WAN connections 230 * should set the slop to 20 (2 packets). 231 */ 232 static int tcp_inflight_enable = 1; 233 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW, 234 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting"); 235 236 static int tcp_inflight_debug = 0; 237 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW, 238 &tcp_inflight_debug, 0, "Debug TCP inflight calculations"); 239 240 static int tcp_inflight_min = 6144; 241 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW, 242 &tcp_inflight_min, 0, "Lower bound for TCP inflight window"); 243 244 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT; 245 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW, 246 &tcp_inflight_max, 0, "Upper bound for TCP inflight window"); 247 248 static int tcp_inflight_stab = 50; 249 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW, 250 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 3 packets)"); 251 252 static int tcp_do_rfc3390 = 1; 253 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_RW, 254 &tcp_do_rfc3390, 0, 255 "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)"); 256 257 static u_long tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 258 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwmaxsegs, CTLFLAG_RW, 259 &tcp_iw_maxsegs, 0, "TCP IW segments max"); 260 261 static u_long tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 262 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwcapsegs, CTLFLAG_RW, 263 &tcp_iw_capsegs, 0, "TCP IW segments"); 264 265 int tcp_low_rtobase = 1; 266 SYSCTL_INT(_net_inet_tcp, OID_AUTO, low_rtobase, CTLFLAG_RW, 267 &tcp_low_rtobase, 0, "Lowering the Initial RTO (RFC 6298)"); 268 269 static int tcp_do_ncr = 1; 270 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr, CTLFLAG_RW, 271 &tcp_do_ncr, 0, "Non-Congestion Robustness (RFC 4653)"); 272 273 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives"); 274 static struct malloc_pipe tcptemp_mpipe; 275 276 static void tcp_willblock(void); 277 static void tcp_notify (struct inpcb *, int); 278 279 struct tcp_stats tcpstats_percpu[MAXCPU]; 280 #ifdef SMP 281 static int 282 sysctl_tcpstats(SYSCTL_HANDLER_ARGS) 283 { 284 int cpu, error = 0; 285 286 for (cpu = 0; cpu < ncpus; ++cpu) { 287 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu], 288 sizeof(struct tcp_stats)))) 289 break; 290 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu], 291 sizeof(struct tcp_stats)))) 292 break; 293 } 294 295 return (error); 296 } 297 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW), 298 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics"); 299 #else 300 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW, 301 &tcpstat, tcp_stats, "TCP statistics"); 302 #endif 303 304 /* 305 * Target size of TCP PCB hash tables. Must be a power of two. 306 * 307 * Note that this can be overridden by the kernel environment 308 * variable net.inet.tcp.tcbhashsize 309 */ 310 #ifndef TCBHASHSIZE 311 #define TCBHASHSIZE 512 312 #endif 313 314 /* 315 * This is the actual shape of what we allocate using the zone 316 * allocator. Doing it this way allows us to protect both structures 317 * using the same generation count, and also eliminates the overhead 318 * of allocating tcpcbs separately. By hiding the structure here, 319 * we avoid changing most of the rest of the code (although it needs 320 * to be changed, eventually, for greater efficiency). 321 */ 322 #define ALIGNMENT 32 323 #define ALIGNM1 (ALIGNMENT - 1) 324 struct inp_tp { 325 union { 326 struct inpcb inp; 327 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1]; 328 } inp_tp_u; 329 struct tcpcb tcb; 330 struct tcp_callout inp_tp_rexmt; 331 struct tcp_callout inp_tp_persist; 332 struct tcp_callout inp_tp_keep; 333 struct tcp_callout inp_tp_2msl; 334 struct tcp_callout inp_tp_delack; 335 struct netmsg_tcp_timer inp_tp_timermsg; 336 }; 337 #undef ALIGNMENT 338 #undef ALIGNM1 339 340 /* 341 * Tcp initialization 342 */ 343 void 344 tcp_init(void) 345 { 346 struct inpcbporthead *porthashbase; 347 struct inpcbinfo *ticb; 348 u_long porthashmask; 349 int hashsize = TCBHASHSIZE; 350 int cpu; 351 352 /* 353 * note: tcptemp is used for keepalives, and it is ok for an 354 * allocation to fail so do not specify MPF_INT. 355 */ 356 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp), 357 25, -1, 0, NULL, NULL, NULL); 358 359 tcp_delacktime = TCPTV_DELACK; 360 tcp_keepinit = TCPTV_KEEP_INIT; 361 tcp_keepidle = TCPTV_KEEP_IDLE; 362 tcp_keepintvl = TCPTV_KEEPINTVL; 363 tcp_maxpersistidle = TCPTV_KEEP_IDLE; 364 tcp_msl = TCPTV_MSL; 365 tcp_rexmit_min = TCPTV_MIN; 366 tcp_rexmit_slop = TCPTV_CPU_VAR; 367 368 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize); 369 if (!powerof2(hashsize)) { 370 kprintf("WARNING: TCB hash size not a power of 2\n"); 371 hashsize = 512; /* safe default */ 372 } 373 tcp_tcbhashsize = hashsize; 374 porthashbase = hashinit(hashsize, M_PCB, &porthashmask); 375 376 for (cpu = 0; cpu < ncpus2; cpu++) { 377 ticb = &tcbinfo[cpu]; 378 in_pcbinfo_init(ticb); 379 ticb->cpu = cpu; 380 ticb->hashbase = hashinit(hashsize, M_PCB, 381 &ticb->hashmask); 382 ticb->porthashbase = porthashbase; 383 ticb->porthashmask = porthashmask; 384 ticb->porttoken = &tcp_port_token; 385 #if 0 386 ticb->porthashbase = hashinit(hashsize, M_PCB, 387 &ticb->porthashmask); 388 #endif 389 ticb->wildcardhashbase = hashinit(hashsize, M_PCB, 390 &ticb->wildcardhashmask); 391 ticb->ipi_size = sizeof(struct inp_tp); 392 TAILQ_INIT(&tcpcbackq[cpu]); 393 } 394 395 tcp_reass_maxseg = nmbclusters / 16; 396 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg); 397 398 #ifdef INET6 399 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) 400 #else 401 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) 402 #endif 403 if (max_protohdr < TCP_MINPROTOHDR) 404 max_protohdr = TCP_MINPROTOHDR; 405 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) 406 panic("tcp_init"); 407 #undef TCP_MINPROTOHDR 408 409 /* 410 * Initialize TCP statistics counters for each CPU. 411 */ 412 #ifdef SMP 413 for (cpu = 0; cpu < ncpus; ++cpu) { 414 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats)); 415 } 416 #else 417 bzero(&tcpstat, sizeof(struct tcp_stats)); 418 #endif 419 420 syncache_init(); 421 netisr_register_rollup(tcp_willblock); 422 } 423 424 static void 425 tcp_willblock(void) 426 { 427 struct tcpcb *tp; 428 int cpu = mycpu->gd_cpuid; 429 430 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) { 431 KKASSERT(tp->t_flags & TF_ONOUTPUTQ); 432 tp->t_flags &= ~TF_ONOUTPUTQ; 433 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq); 434 tcp_output(tp); 435 } 436 } 437 438 /* 439 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. 440 * tcp_template used to store this data in mbufs, but we now recopy it out 441 * of the tcpcb each time to conserve mbufs. 442 */ 443 void 444 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr) 445 { 446 struct inpcb *inp = tp->t_inpcb; 447 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr; 448 449 #ifdef INET6 450 if (inp->inp_vflag & INP_IPV6) { 451 struct ip6_hdr *ip6; 452 453 ip6 = (struct ip6_hdr *)ip_ptr; 454 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | 455 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK); 456 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | 457 (IPV6_VERSION & IPV6_VERSION_MASK); 458 ip6->ip6_nxt = IPPROTO_TCP; 459 ip6->ip6_plen = sizeof(struct tcphdr); 460 ip6->ip6_src = inp->in6p_laddr; 461 ip6->ip6_dst = inp->in6p_faddr; 462 tcp_hdr->th_sum = 0; 463 } else 464 #endif 465 { 466 struct ip *ip = (struct ip *) ip_ptr; 467 468 ip->ip_vhl = IP_VHL_BORING; 469 ip->ip_tos = 0; 470 ip->ip_len = 0; 471 ip->ip_id = 0; 472 ip->ip_off = 0; 473 ip->ip_ttl = 0; 474 ip->ip_sum = 0; 475 ip->ip_p = IPPROTO_TCP; 476 ip->ip_src = inp->inp_laddr; 477 ip->ip_dst = inp->inp_faddr; 478 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr, 479 ip->ip_dst.s_addr, 480 htons(sizeof(struct tcphdr) + IPPROTO_TCP)); 481 } 482 483 tcp_hdr->th_sport = inp->inp_lport; 484 tcp_hdr->th_dport = inp->inp_fport; 485 tcp_hdr->th_seq = 0; 486 tcp_hdr->th_ack = 0; 487 tcp_hdr->th_x2 = 0; 488 tcp_hdr->th_off = 5; 489 tcp_hdr->th_flags = 0; 490 tcp_hdr->th_win = 0; 491 tcp_hdr->th_urp = 0; 492 } 493 494 /* 495 * Create template to be used to send tcp packets on a connection. 496 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only 497 * use for this function is in keepalives, which use tcp_respond. 498 */ 499 struct tcptemp * 500 tcp_maketemplate(struct tcpcb *tp) 501 { 502 struct tcptemp *tmp; 503 504 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL) 505 return (NULL); 506 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t); 507 return (tmp); 508 } 509 510 void 511 tcp_freetemplate(struct tcptemp *tmp) 512 { 513 mpipe_free(&tcptemp_mpipe, tmp); 514 } 515 516 /* 517 * Send a single message to the TCP at address specified by 518 * the given TCP/IP header. If m == NULL, then we make a copy 519 * of the tcpiphdr at ti and send directly to the addressed host. 520 * This is used to force keep alive messages out using the TCP 521 * template for a connection. If flags are given then we send 522 * a message back to the TCP which originated the * segment ti, 523 * and discard the mbuf containing it and any other attached mbufs. 524 * 525 * In any case the ack and sequence number of the transmitted 526 * segment are as specified by the parameters. 527 * 528 * NOTE: If m != NULL, then ti must point to *inside* the mbuf. 529 */ 530 void 531 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, 532 tcp_seq ack, tcp_seq seq, int flags) 533 { 534 int tlen; 535 int win = 0; 536 struct route *ro = NULL; 537 struct route sro; 538 struct ip *ip = ipgen; 539 struct tcphdr *nth; 540 int ipflags = 0; 541 struct route_in6 *ro6 = NULL; 542 struct route_in6 sro6; 543 struct ip6_hdr *ip6 = ipgen; 544 boolean_t use_tmpro = TRUE; 545 #ifdef INET6 546 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6); 547 #else 548 const boolean_t isipv6 = FALSE; 549 #endif 550 551 if (tp != NULL) { 552 if (!(flags & TH_RST)) { 553 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv); 554 if (win < 0) 555 win = 0; 556 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 557 win = (long)TCP_MAXWIN << tp->rcv_scale; 558 } 559 /* 560 * Don't use the route cache of a listen socket, 561 * it is not MPSAFE; use temporary route cache. 562 */ 563 if (tp->t_state != TCPS_LISTEN) { 564 if (isipv6) 565 ro6 = &tp->t_inpcb->in6p_route; 566 else 567 ro = &tp->t_inpcb->inp_route; 568 use_tmpro = FALSE; 569 } 570 } 571 if (use_tmpro) { 572 if (isipv6) { 573 ro6 = &sro6; 574 bzero(ro6, sizeof *ro6); 575 } else { 576 ro = &sro; 577 bzero(ro, sizeof *ro); 578 } 579 } 580 if (m == NULL) { 581 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 582 if (m == NULL) 583 return; 584 tlen = 0; 585 m->m_data += max_linkhdr; 586 if (isipv6) { 587 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); 588 ip6 = mtod(m, struct ip6_hdr *); 589 nth = (struct tcphdr *)(ip6 + 1); 590 } else { 591 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip)); 592 ip = mtod(m, struct ip *); 593 nth = (struct tcphdr *)(ip + 1); 594 } 595 bcopy(th, nth, sizeof(struct tcphdr)); 596 flags = TH_ACK; 597 } else { 598 m_freem(m->m_next); 599 m->m_next = NULL; 600 m->m_data = (caddr_t)ipgen; 601 /* m_len is set later */ 602 tlen = 0; 603 #define xchg(a, b, type) { type t; t = a; a = b; b = t; } 604 if (isipv6) { 605 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); 606 nth = (struct tcphdr *)(ip6 + 1); 607 } else { 608 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long); 609 nth = (struct tcphdr *)(ip + 1); 610 } 611 if (th != nth) { 612 /* 613 * this is usually a case when an extension header 614 * exists between the IPv6 header and the 615 * TCP header. 616 */ 617 nth->th_sport = th->th_sport; 618 nth->th_dport = th->th_dport; 619 } 620 xchg(nth->th_dport, nth->th_sport, n_short); 621 #undef xchg 622 } 623 if (isipv6) { 624 ip6->ip6_flow = 0; 625 ip6->ip6_vfc = IPV6_VERSION; 626 ip6->ip6_nxt = IPPROTO_TCP; 627 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen)); 628 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 629 } else { 630 tlen += sizeof(struct tcpiphdr); 631 ip->ip_len = tlen; 632 ip->ip_ttl = ip_defttl; 633 } 634 m->m_len = tlen; 635 m->m_pkthdr.len = tlen; 636 m->m_pkthdr.rcvif = NULL; 637 nth->th_seq = htonl(seq); 638 nth->th_ack = htonl(ack); 639 nth->th_x2 = 0; 640 nth->th_off = sizeof(struct tcphdr) >> 2; 641 nth->th_flags = flags; 642 if (tp != NULL) 643 nth->th_win = htons((u_short) (win >> tp->rcv_scale)); 644 else 645 nth->th_win = htons((u_short)win); 646 nth->th_urp = 0; 647 if (isipv6) { 648 nth->th_sum = 0; 649 nth->th_sum = in6_cksum(m, IPPROTO_TCP, 650 sizeof(struct ip6_hdr), 651 tlen - sizeof(struct ip6_hdr)); 652 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL, 653 (ro6 && ro6->ro_rt) ? 654 ro6->ro_rt->rt_ifp : NULL); 655 } else { 656 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 657 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); 658 m->m_pkthdr.csum_flags = CSUM_TCP; 659 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 660 } 661 #ifdef TCPDEBUG 662 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) 663 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); 664 #endif 665 if (isipv6) { 666 ip6_output(m, NULL, ro6, ipflags, NULL, NULL, 667 tp ? tp->t_inpcb : NULL); 668 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) { 669 RTFREE(ro6->ro_rt); 670 ro6->ro_rt = NULL; 671 } 672 } else { 673 ipflags |= IP_DEBUGROUTE; 674 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL); 675 if ((ro == &sro) && (ro->ro_rt != NULL)) { 676 RTFREE(ro->ro_rt); 677 ro->ro_rt = NULL; 678 } 679 } 680 } 681 682 /* 683 * Create a new TCP control block, making an 684 * empty reassembly queue and hooking it to the argument 685 * protocol control block. The `inp' parameter must have 686 * come from the zone allocator set up in tcp_init(). 687 */ 688 struct tcpcb * 689 tcp_newtcpcb(struct inpcb *inp) 690 { 691 struct inp_tp *it; 692 struct tcpcb *tp; 693 #ifdef INET6 694 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 695 #else 696 const boolean_t isipv6 = FALSE; 697 #endif 698 699 it = (struct inp_tp *)inp; 700 tp = &it->tcb; 701 bzero(tp, sizeof(struct tcpcb)); 702 TAILQ_INIT(&tp->t_segq); 703 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 704 tp->t_rxtthresh = tcprexmtthresh; 705 706 /* Set up our timeouts. */ 707 tp->tt_rexmt = &it->inp_tp_rexmt; 708 tp->tt_persist = &it->inp_tp_persist; 709 tp->tt_keep = &it->inp_tp_keep; 710 tp->tt_2msl = &it->inp_tp_2msl; 711 tp->tt_delack = &it->inp_tp_delack; 712 tcp_inittimers(tp); 713 714 /* 715 * Zero out timer message. We don't create it here, 716 * since the current CPU may not be the owner of this 717 * inpcb. 718 */ 719 tp->tt_msg = &it->inp_tp_timermsg; 720 bzero(tp->tt_msg, sizeof(*tp->tt_msg)); 721 722 tp->t_keepinit = tcp_keepinit; 723 tp->t_keepidle = tcp_keepidle; 724 tp->t_keepintvl = tcp_keepintvl; 725 tp->t_keepcnt = tcp_keepcnt; 726 tp->t_maxidle = tp->t_keepintvl * tp->t_keepcnt; 727 728 if (tcp_do_ncr) 729 tp->t_flags |= TF_NCR; 730 if (tcp_do_rfc1323) 731 tp->t_flags |= (TF_REQ_SCALE | TF_REQ_TSTMP); 732 733 tp->t_inpcb = inp; /* XXX */ 734 tp->t_state = TCPS_CLOSED; 735 /* 736 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 737 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 738 * reasonable initial retransmit time. 739 */ 740 tp->t_srtt = TCPTV_SRTTBASE; 741 tp->t_rttvar = 742 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 743 tp->t_rttmin = tcp_rexmit_min; 744 tp->t_rxtcur = TCPTV_RTOBASE; 745 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 746 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 747 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 748 tp->snd_last = ticks; 749 tp->t_rcvtime = ticks; 750 /* 751 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 752 * because the socket may be bound to an IPv6 wildcard address, 753 * which may match an IPv4-mapped IPv6 address. 754 */ 755 inp->inp_ip_ttl = ip_defttl; 756 inp->inp_ppcb = tp; 757 tcp_sack_tcpcb_init(tp); 758 return (tp); /* XXX */ 759 } 760 761 /* 762 * Drop a TCP connection, reporting the specified error. 763 * If connection is synchronized, then send a RST to peer. 764 */ 765 struct tcpcb * 766 tcp_drop(struct tcpcb *tp, int error) 767 { 768 struct socket *so = tp->t_inpcb->inp_socket; 769 770 if (TCPS_HAVERCVDSYN(tp->t_state)) { 771 tp->t_state = TCPS_CLOSED; 772 tcp_output(tp); 773 tcpstat.tcps_drops++; 774 } else 775 tcpstat.tcps_conndrops++; 776 if (error == ETIMEDOUT && tp->t_softerror) 777 error = tp->t_softerror; 778 so->so_error = error; 779 return (tcp_close(tp)); 780 } 781 782 #ifdef SMP 783 784 struct netmsg_listen_detach { 785 struct netmsg_base base; 786 struct tcpcb *nm_tp; 787 }; 788 789 static void 790 tcp_listen_detach_handler(netmsg_t msg) 791 { 792 struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg; 793 struct tcpcb *tp = nmsg->nm_tp; 794 int cpu = mycpuid, nextcpu; 795 796 if (tp->t_flags & TF_LISTEN) 797 syncache_destroy(tp); 798 799 in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]); 800 801 nextcpu = cpu + 1; 802 if (nextcpu < ncpus2) 803 lwkt_forwardmsg(cpu_portfn(nextcpu), &nmsg->base.lmsg); 804 else 805 lwkt_replymsg(&nmsg->base.lmsg, 0); 806 } 807 808 #endif 809 810 /* 811 * Close a TCP control block: 812 * discard all space held by the tcp 813 * discard internet protocol block 814 * wake up any sleepers 815 */ 816 struct tcpcb * 817 tcp_close(struct tcpcb *tp) 818 { 819 struct tseg_qent *q; 820 struct inpcb *inp = tp->t_inpcb; 821 struct socket *so = inp->inp_socket; 822 struct rtentry *rt; 823 boolean_t dosavessthresh; 824 #ifdef INET6 825 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 826 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0); 827 #else 828 const boolean_t isipv6 = FALSE; 829 #endif 830 831 #ifdef SMP 832 /* 833 * INP_WILDCARD_MP indicates that listen(2) has been called on 834 * this socket. This implies: 835 * - A wildcard inp's hash is replicated for each protocol thread. 836 * - Syncache for this inp grows independently in each protocol 837 * thread. 838 * - There is more than one cpu 839 * 840 * We have to chain a message to the rest of the protocol threads 841 * to cleanup the wildcard hash and the syncache. The cleanup 842 * in the current protocol thread is defered till the end of this 843 * function. 844 * 845 * NOTE: 846 * After cleanup the inp's hash and syncache entries, this inp will 847 * no longer be available to the rest of the protocol threads, so we 848 * are safe to whack the inp in the following code. 849 */ 850 if (inp->inp_flags & INP_WILDCARD_MP) { 851 struct netmsg_listen_detach nmsg; 852 853 KKASSERT(so->so_port == cpu_portfn(0)); 854 KKASSERT(&curthread->td_msgport == cpu_portfn(0)); 855 KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]); 856 857 netmsg_init(&nmsg.base, NULL, &curthread->td_msgport, 858 MSGF_PRIORITY, tcp_listen_detach_handler); 859 nmsg.nm_tp = tp; 860 lwkt_domsg(cpu_portfn(1), &nmsg.base.lmsg, 0); 861 862 inp->inp_flags &= ~INP_WILDCARD_MP; 863 } 864 #endif 865 866 KKASSERT(tp->t_state != TCPS_TERMINATING); 867 tp->t_state = TCPS_TERMINATING; 868 869 /* 870 * Make sure that all of our timers are stopped before we 871 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL), 872 * timers are never used. If timer message is never created 873 * (tp->tt_msg->tt_tcb == NULL), timers are never used too. 874 */ 875 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) { 876 tcp_callout_stop(tp, tp->tt_rexmt); 877 tcp_callout_stop(tp, tp->tt_persist); 878 tcp_callout_stop(tp, tp->tt_keep); 879 tcp_callout_stop(tp, tp->tt_2msl); 880 tcp_callout_stop(tp, tp->tt_delack); 881 } 882 883 if (tp->t_flags & TF_ONOUTPUTQ) { 884 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 885 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 886 tp->t_flags &= ~TF_ONOUTPUTQ; 887 } 888 889 /* 890 * If we got enough samples through the srtt filter, 891 * save the rtt and rttvar in the routing entry. 892 * 'Enough' is arbitrarily defined as the 16 samples. 893 * 16 samples is enough for the srtt filter to converge 894 * to within 5% of the correct value; fewer samples and 895 * we could save a very bogus rtt. 896 * 897 * Don't update the default route's characteristics and don't 898 * update anything that the user "locked". 899 */ 900 if (tp->t_rttupdated >= 16) { 901 u_long i = 0; 902 903 if (isipv6) { 904 struct sockaddr_in6 *sin6; 905 906 if ((rt = inp->in6p_route.ro_rt) == NULL) 907 goto no_valid_rt; 908 sin6 = (struct sockaddr_in6 *)rt_key(rt); 909 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 910 goto no_valid_rt; 911 } else 912 if ((rt = inp->inp_route.ro_rt) == NULL || 913 ((struct sockaddr_in *)rt_key(rt))-> 914 sin_addr.s_addr == INADDR_ANY) 915 goto no_valid_rt; 916 917 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 918 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 919 if (rt->rt_rmx.rmx_rtt && i) 920 /* 921 * filter this update to half the old & half 922 * the new values, converting scale. 923 * See route.h and tcp_var.h for a 924 * description of the scaling constants. 925 */ 926 rt->rt_rmx.rmx_rtt = 927 (rt->rt_rmx.rmx_rtt + i) / 2; 928 else 929 rt->rt_rmx.rmx_rtt = i; 930 tcpstat.tcps_cachedrtt++; 931 } 932 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 933 i = tp->t_rttvar * 934 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 935 if (rt->rt_rmx.rmx_rttvar && i) 936 rt->rt_rmx.rmx_rttvar = 937 (rt->rt_rmx.rmx_rttvar + i) / 2; 938 else 939 rt->rt_rmx.rmx_rttvar = i; 940 tcpstat.tcps_cachedrttvar++; 941 } 942 /* 943 * The old comment here said: 944 * update the pipelimit (ssthresh) if it has been updated 945 * already or if a pipesize was specified & the threshhold 946 * got below half the pipesize. I.e., wait for bad news 947 * before we start updating, then update on both good 948 * and bad news. 949 * 950 * But we want to save the ssthresh even if no pipesize is 951 * specified explicitly in the route, because such 952 * connections still have an implicit pipesize specified 953 * by the global tcp_sendspace. In the absence of a reliable 954 * way to calculate the pipesize, it will have to do. 955 */ 956 i = tp->snd_ssthresh; 957 if (rt->rt_rmx.rmx_sendpipe != 0) 958 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 959 else 960 dosavessthresh = (i < so->so_snd.ssb_hiwat/2); 961 if (dosavessthresh || 962 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 963 (rt->rt_rmx.rmx_ssthresh != 0))) { 964 /* 965 * convert the limit from user data bytes to 966 * packets then to packet data bytes. 967 */ 968 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 969 if (i < 2) 970 i = 2; 971 i *= tp->t_maxseg + 972 (isipv6 ? 973 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 974 sizeof(struct tcpiphdr)); 975 if (rt->rt_rmx.rmx_ssthresh) 976 rt->rt_rmx.rmx_ssthresh = 977 (rt->rt_rmx.rmx_ssthresh + i) / 2; 978 else 979 rt->rt_rmx.rmx_ssthresh = i; 980 tcpstat.tcps_cachedssthresh++; 981 } 982 } 983 984 no_valid_rt: 985 /* free the reassembly queue, if any */ 986 while((q = TAILQ_FIRST(&tp->t_segq)) != NULL) { 987 TAILQ_REMOVE(&tp->t_segq, q, tqe_q); 988 m_freem(q->tqe_m); 989 kfree(q, M_TSEGQ); 990 atomic_add_int(&tcp_reass_qsize, -1); 991 } 992 /* throw away SACK blocks in scoreboard*/ 993 if (TCP_DO_SACK(tp)) 994 tcp_sack_destroy(&tp->scb); 995 996 inp->inp_ppcb = NULL; 997 soisdisconnected(so); 998 /* note: pcb detached later on */ 999 1000 tcp_destroy_timermsg(tp); 1001 1002 if (tp->t_flags & TF_LISTEN) 1003 syncache_destroy(tp); 1004 1005 /* 1006 * NOTE: 1007 * pcbdetach removes any wildcard hash entry on the current CPU. 1008 */ 1009 #ifdef INET6 1010 if (isafinet6) 1011 in6_pcbdetach(inp); 1012 else 1013 #endif 1014 in_pcbdetach(inp); 1015 1016 tcpstat.tcps_closed++; 1017 return (NULL); 1018 } 1019 1020 static __inline void 1021 tcp_drain_oncpu(struct inpcbhead *head) 1022 { 1023 struct inpcb *marker; 1024 struct inpcb *inpb; 1025 struct tcpcb *tcpb; 1026 struct tseg_qent *te; 1027 1028 /* 1029 * Allows us to block while running the list 1030 */ 1031 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1032 marker->inp_flags |= INP_PLACEMARKER; 1033 LIST_INSERT_HEAD(head, marker, inp_list); 1034 1035 while ((inpb = LIST_NEXT(marker, inp_list)) != NULL) { 1036 if ((inpb->inp_flags & INP_PLACEMARKER) == 0 && 1037 (tcpb = intotcpcb(inpb)) != NULL && 1038 (te = TAILQ_FIRST(&tcpb->t_segq)) != NULL) { 1039 TAILQ_REMOVE(&tcpb->t_segq, te, tqe_q); 1040 m_freem(te->tqe_m); 1041 kfree(te, M_TSEGQ); 1042 atomic_add_int(&tcp_reass_qsize, -1); 1043 /* retry */ 1044 } else { 1045 LIST_REMOVE(marker, inp_list); 1046 LIST_INSERT_AFTER(inpb, marker, inp_list); 1047 } 1048 } 1049 LIST_REMOVE(marker, inp_list); 1050 kfree(marker, M_TEMP); 1051 } 1052 1053 #ifdef SMP 1054 struct netmsg_tcp_drain { 1055 struct netmsg_base base; 1056 struct inpcbhead *nm_head; 1057 }; 1058 1059 static void 1060 tcp_drain_handler(netmsg_t msg) 1061 { 1062 struct netmsg_tcp_drain *nm = (void *)msg; 1063 1064 tcp_drain_oncpu(nm->nm_head); 1065 lwkt_replymsg(&nm->base.lmsg, 0); 1066 } 1067 #endif 1068 1069 void 1070 tcp_drain(void) 1071 { 1072 #ifdef SMP 1073 int cpu; 1074 #endif 1075 1076 if (!do_tcpdrain) 1077 return; 1078 1079 /* 1080 * Walk the tcpbs, if existing, and flush the reassembly queue, 1081 * if there is one... 1082 * XXX: The "Net/3" implementation doesn't imply that the TCP 1083 * reassembly queue should be flushed, but in a situation 1084 * where we're really low on mbufs, this is potentially 1085 * useful. 1086 */ 1087 #ifdef SMP 1088 for (cpu = 0; cpu < ncpus2; cpu++) { 1089 struct netmsg_tcp_drain *nm; 1090 1091 if (cpu == mycpu->gd_cpuid) { 1092 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead); 1093 } else { 1094 nm = kmalloc(sizeof(struct netmsg_tcp_drain), 1095 M_LWKTMSG, M_NOWAIT); 1096 if (nm == NULL) 1097 continue; 1098 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1099 0, tcp_drain_handler); 1100 nm->nm_head = &tcbinfo[cpu].pcblisthead; 1101 lwkt_sendmsg(cpu_portfn(cpu), &nm->base.lmsg); 1102 } 1103 } 1104 #else 1105 tcp_drain_oncpu(&tcbinfo[0].pcblisthead); 1106 #endif 1107 } 1108 1109 /* 1110 * Notify a tcp user of an asynchronous error; 1111 * store error as soft error, but wake up user 1112 * (for now, won't do anything until can select for soft error). 1113 * 1114 * Do not wake up user since there currently is no mechanism for 1115 * reporting soft errors (yet - a kqueue filter may be added). 1116 */ 1117 static void 1118 tcp_notify(struct inpcb *inp, int error) 1119 { 1120 struct tcpcb *tp = intotcpcb(inp); 1121 1122 /* 1123 * Ignore some errors if we are hooked up. 1124 * If connection hasn't completed, has retransmitted several times, 1125 * and receives a second error, give up now. This is better 1126 * than waiting a long time to establish a connection that 1127 * can never complete. 1128 */ 1129 if (tp->t_state == TCPS_ESTABLISHED && 1130 (error == EHOSTUNREACH || error == ENETUNREACH || 1131 error == EHOSTDOWN)) { 1132 return; 1133 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1134 tp->t_softerror) 1135 tcp_drop(tp, error); 1136 else 1137 tp->t_softerror = error; 1138 #if 0 1139 wakeup(&so->so_timeo); 1140 sorwakeup(so); 1141 sowwakeup(so); 1142 #endif 1143 } 1144 1145 static int 1146 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1147 { 1148 int error, i, n; 1149 struct inpcb *marker; 1150 struct inpcb *inp; 1151 globaldata_t gd; 1152 int origcpu, ccpu; 1153 1154 error = 0; 1155 n = 0; 1156 1157 /* 1158 * The process of preparing the TCB list is too time-consuming and 1159 * resource-intensive to repeat twice on every request. 1160 */ 1161 if (req->oldptr == NULL) { 1162 for (ccpu = 0; ccpu < ncpus; ++ccpu) { 1163 gd = globaldata_find(ccpu); 1164 n += tcbinfo[gd->gd_cpuid].ipi_count; 1165 } 1166 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1167 return (0); 1168 } 1169 1170 if (req->newptr != NULL) 1171 return (EPERM); 1172 1173 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1174 marker->inp_flags |= INP_PLACEMARKER; 1175 1176 /* 1177 * OK, now we're committed to doing something. Run the inpcb list 1178 * for each cpu in the system and construct the output. Use a 1179 * list placemarker to deal with list changes occuring during 1180 * copyout blockages (but otherwise depend on being on the correct 1181 * cpu to avoid races). 1182 */ 1183 origcpu = mycpu->gd_cpuid; 1184 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) { 1185 globaldata_t rgd; 1186 caddr_t inp_ppcb; 1187 struct xtcpcb xt; 1188 int cpu_id; 1189 1190 cpu_id = (origcpu + ccpu) % ncpus; 1191 if ((smp_active_mask & CPUMASK(cpu_id)) == 0) 1192 continue; 1193 rgd = globaldata_find(cpu_id); 1194 lwkt_setcpu_self(rgd); 1195 1196 n = tcbinfo[cpu_id].ipi_count; 1197 1198 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list); 1199 i = 0; 1200 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1201 /* 1202 * process a snapshot of pcbs, ignoring placemarkers 1203 * and using our own to allow SYSCTL_OUT to block. 1204 */ 1205 LIST_REMOVE(marker, inp_list); 1206 LIST_INSERT_AFTER(inp, marker, inp_list); 1207 1208 if (inp->inp_flags & INP_PLACEMARKER) 1209 continue; 1210 if (prison_xinpcb(req->td, inp)) 1211 continue; 1212 1213 xt.xt_len = sizeof xt; 1214 bcopy(inp, &xt.xt_inp, sizeof *inp); 1215 inp_ppcb = inp->inp_ppcb; 1216 if (inp_ppcb != NULL) 1217 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1218 else 1219 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1220 if (inp->inp_socket) 1221 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1222 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1223 break; 1224 ++i; 1225 } 1226 LIST_REMOVE(marker, inp_list); 1227 if (error == 0 && i < n) { 1228 bzero(&xt, sizeof xt); 1229 xt.xt_len = sizeof xt; 1230 while (i < n) { 1231 error = SYSCTL_OUT(req, &xt, sizeof xt); 1232 if (error) 1233 break; 1234 ++i; 1235 } 1236 } 1237 } 1238 1239 /* 1240 * Make sure we are on the same cpu we were on originally, since 1241 * higher level callers expect this. Also don't pollute caches with 1242 * migrated userland data by (eventually) returning to userland 1243 * on a different cpu. 1244 */ 1245 lwkt_setcpu_self(globaldata_find(origcpu)); 1246 kfree(marker, M_TEMP); 1247 return (error); 1248 } 1249 1250 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1251 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1252 1253 static int 1254 tcp_getcred(SYSCTL_HANDLER_ARGS) 1255 { 1256 struct sockaddr_in addrs[2]; 1257 struct inpcb *inp; 1258 int cpu; 1259 int error; 1260 1261 error = priv_check(req->td, PRIV_ROOT); 1262 if (error != 0) 1263 return (error); 1264 error = SYSCTL_IN(req, addrs, sizeof addrs); 1265 if (error != 0) 1266 return (error); 1267 crit_enter(); 1268 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1269 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1270 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1271 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1272 if (inp == NULL || inp->inp_socket == NULL) { 1273 error = ENOENT; 1274 goto out; 1275 } 1276 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1277 out: 1278 crit_exit(); 1279 return (error); 1280 } 1281 1282 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1283 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1284 1285 #ifdef INET6 1286 static int 1287 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1288 { 1289 struct sockaddr_in6 addrs[2]; 1290 struct inpcb *inp; 1291 int error; 1292 boolean_t mapped = FALSE; 1293 1294 error = priv_check(req->td, PRIV_ROOT); 1295 if (error != 0) 1296 return (error); 1297 error = SYSCTL_IN(req, addrs, sizeof addrs); 1298 if (error != 0) 1299 return (error); 1300 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { 1301 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) 1302 mapped = TRUE; 1303 else 1304 return (EINVAL); 1305 } 1306 crit_enter(); 1307 if (mapped) { 1308 inp = in_pcblookup_hash(&tcbinfo[0], 1309 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], 1310 addrs[1].sin6_port, 1311 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], 1312 addrs[0].sin6_port, 1313 0, NULL); 1314 } else { 1315 inp = in6_pcblookup_hash(&tcbinfo[0], 1316 &addrs[1].sin6_addr, addrs[1].sin6_port, 1317 &addrs[0].sin6_addr, addrs[0].sin6_port, 1318 0, NULL); 1319 } 1320 if (inp == NULL || inp->inp_socket == NULL) { 1321 error = ENOENT; 1322 goto out; 1323 } 1324 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1325 out: 1326 crit_exit(); 1327 return (error); 1328 } 1329 1330 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1331 0, 0, 1332 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1333 #endif 1334 1335 struct netmsg_tcp_notify { 1336 struct netmsg_base base; 1337 void (*nm_notify)(struct inpcb *, int); 1338 struct in_addr nm_faddr; 1339 int nm_arg; 1340 }; 1341 1342 static void 1343 tcp_notifyall_oncpu(netmsg_t msg) 1344 { 1345 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg; 1346 int nextcpu; 1347 1348 in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nm->nm_faddr, 1349 nm->nm_arg, nm->nm_notify); 1350 1351 nextcpu = mycpuid + 1; 1352 if (nextcpu < ncpus2) 1353 lwkt_forwardmsg(cpu_portfn(nextcpu), &nm->base.lmsg); 1354 else 1355 lwkt_replymsg(&nm->base.lmsg, 0); 1356 } 1357 1358 void 1359 tcp_ctlinput(netmsg_t msg) 1360 { 1361 int cmd = msg->ctlinput.nm_cmd; 1362 struct sockaddr *sa = msg->ctlinput.nm_arg; 1363 struct ip *ip = msg->ctlinput.nm_extra; 1364 struct tcphdr *th; 1365 struct in_addr faddr; 1366 struct inpcb *inp; 1367 struct tcpcb *tp; 1368 void (*notify)(struct inpcb *, int) = tcp_notify; 1369 tcp_seq icmpseq; 1370 int arg, cpu; 1371 1372 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) { 1373 goto done; 1374 } 1375 1376 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1377 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1378 goto done; 1379 1380 arg = inetctlerrmap[cmd]; 1381 if (cmd == PRC_QUENCH) { 1382 notify = tcp_quench; 1383 } else if (icmp_may_rst && 1384 (cmd == PRC_UNREACH_ADMIN_PROHIB || 1385 cmd == PRC_UNREACH_PORT || 1386 cmd == PRC_TIMXCEED_INTRANS) && 1387 ip != NULL) { 1388 notify = tcp_drop_syn_sent; 1389 } else if (cmd == PRC_MSGSIZE) { 1390 struct icmp *icmp = (struct icmp *) 1391 ((caddr_t)ip - offsetof(struct icmp, icmp_ip)); 1392 1393 arg = ntohs(icmp->icmp_nextmtu); 1394 notify = tcp_mtudisc; 1395 } else if (PRC_IS_REDIRECT(cmd)) { 1396 ip = NULL; 1397 notify = in_rtchange; 1398 } else if (cmd == PRC_HOSTDEAD) { 1399 ip = NULL; 1400 } 1401 1402 if (ip != NULL) { 1403 crit_enter(); 1404 th = (struct tcphdr *)((caddr_t)ip + 1405 (IP_VHL_HL(ip->ip_vhl) << 2)); 1406 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport, 1407 ip->ip_src.s_addr, th->th_sport); 1408 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport, 1409 ip->ip_src, th->th_sport, 0, NULL); 1410 if ((inp != NULL) && (inp->inp_socket != NULL)) { 1411 icmpseq = htonl(th->th_seq); 1412 tp = intotcpcb(inp); 1413 if (SEQ_GEQ(icmpseq, tp->snd_una) && 1414 SEQ_LT(icmpseq, tp->snd_max)) 1415 (*notify)(inp, arg); 1416 } else { 1417 struct in_conninfo inc; 1418 1419 inc.inc_fport = th->th_dport; 1420 inc.inc_lport = th->th_sport; 1421 inc.inc_faddr = faddr; 1422 inc.inc_laddr = ip->ip_src; 1423 #ifdef INET6 1424 inc.inc_isipv6 = 0; 1425 #endif 1426 syncache_unreach(&inc, th); 1427 } 1428 crit_exit(); 1429 } else { 1430 struct netmsg_tcp_notify *nm; 1431 1432 KKASSERT(&curthread->td_msgport == cpu_portfn(0)); 1433 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT); 1434 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1435 0, tcp_notifyall_oncpu); 1436 nm->nm_faddr = faddr; 1437 nm->nm_arg = arg; 1438 nm->nm_notify = notify; 1439 1440 lwkt_sendmsg(cpu_portfn(0), &nm->base.lmsg); 1441 } 1442 done: 1443 lwkt_replymsg(&msg->lmsg, 0); 1444 } 1445 1446 #ifdef INET6 1447 1448 void 1449 tcp6_ctlinput(netmsg_t msg) 1450 { 1451 int cmd = msg->ctlinput.nm_cmd; 1452 struct sockaddr *sa = msg->ctlinput.nm_arg; 1453 void *d = msg->ctlinput.nm_extra; 1454 struct tcphdr th; 1455 void (*notify) (struct inpcb *, int) = tcp_notify; 1456 struct ip6_hdr *ip6; 1457 struct mbuf *m; 1458 struct ip6ctlparam *ip6cp = NULL; 1459 const struct sockaddr_in6 *sa6_src = NULL; 1460 int off; 1461 struct tcp_portonly { 1462 u_int16_t th_sport; 1463 u_int16_t th_dport; 1464 } *thp; 1465 int arg; 1466 1467 if (sa->sa_family != AF_INET6 || 1468 sa->sa_len != sizeof(struct sockaddr_in6)) { 1469 goto out; 1470 } 1471 1472 arg = 0; 1473 if (cmd == PRC_QUENCH) 1474 notify = tcp_quench; 1475 else if (cmd == PRC_MSGSIZE) { 1476 struct ip6ctlparam *ip6cp = d; 1477 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6; 1478 1479 arg = ntohl(icmp6->icmp6_mtu); 1480 notify = tcp_mtudisc; 1481 } else if (!PRC_IS_REDIRECT(cmd) && 1482 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) { 1483 goto out; 1484 } 1485 1486 /* if the parameter is from icmp6, decode it. */ 1487 if (d != NULL) { 1488 ip6cp = (struct ip6ctlparam *)d; 1489 m = ip6cp->ip6c_m; 1490 ip6 = ip6cp->ip6c_ip6; 1491 off = ip6cp->ip6c_off; 1492 sa6_src = ip6cp->ip6c_src; 1493 } else { 1494 m = NULL; 1495 ip6 = NULL; 1496 off = 0; /* fool gcc */ 1497 sa6_src = &sa6_any; 1498 } 1499 1500 if (ip6 != NULL) { 1501 struct in_conninfo inc; 1502 /* 1503 * XXX: We assume that when IPV6 is non NULL, 1504 * M and OFF are valid. 1505 */ 1506 1507 /* check if we can safely examine src and dst ports */ 1508 if (m->m_pkthdr.len < off + sizeof *thp) 1509 goto out; 1510 1511 bzero(&th, sizeof th); 1512 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1513 1514 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport, 1515 (struct sockaddr *)ip6cp->ip6c_src, 1516 th.th_sport, cmd, arg, notify); 1517 1518 inc.inc_fport = th.th_dport; 1519 inc.inc_lport = th.th_sport; 1520 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1521 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1522 inc.inc_isipv6 = 1; 1523 syncache_unreach(&inc, &th); 1524 } else { 1525 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0, 1526 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify); 1527 } 1528 out: 1529 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0); 1530 } 1531 1532 #endif 1533 1534 /* 1535 * Following is where TCP initial sequence number generation occurs. 1536 * 1537 * There are two places where we must use initial sequence numbers: 1538 * 1. In SYN-ACK packets. 1539 * 2. In SYN packets. 1540 * 1541 * All ISNs for SYN-ACK packets are generated by the syncache. See 1542 * tcp_syncache.c for details. 1543 * 1544 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1545 * depends on this property. In addition, these ISNs should be 1546 * unguessable so as to prevent connection hijacking. To satisfy 1547 * the requirements of this situation, the algorithm outlined in 1548 * RFC 1948 is used to generate sequence numbers. 1549 * 1550 * Implementation details: 1551 * 1552 * Time is based off the system timer, and is corrected so that it 1553 * increases by one megabyte per second. This allows for proper 1554 * recycling on high speed LANs while still leaving over an hour 1555 * before rollover. 1556 * 1557 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1558 * between seeding of isn_secret. This is normally set to zero, 1559 * as reseeding should not be necessary. 1560 * 1561 */ 1562 1563 #define ISN_BYTES_PER_SECOND 1048576 1564 1565 u_char isn_secret[32]; 1566 int isn_last_reseed; 1567 MD5_CTX isn_ctx; 1568 1569 tcp_seq 1570 tcp_new_isn(struct tcpcb *tp) 1571 { 1572 u_int32_t md5_buffer[4]; 1573 tcp_seq new_isn; 1574 1575 /* Seed if this is the first use, reseed if requested. */ 1576 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1577 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1578 < (u_int)ticks))) { 1579 read_random_unlimited(&isn_secret, sizeof isn_secret); 1580 isn_last_reseed = ticks; 1581 } 1582 1583 /* Compute the md5 hash and return the ISN. */ 1584 MD5Init(&isn_ctx); 1585 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1586 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1587 #ifdef INET6 1588 if (tp->t_inpcb->inp_vflag & INP_IPV6) { 1589 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1590 sizeof(struct in6_addr)); 1591 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1592 sizeof(struct in6_addr)); 1593 } else 1594 #endif 1595 { 1596 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1597 sizeof(struct in_addr)); 1598 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1599 sizeof(struct in_addr)); 1600 } 1601 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1602 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1603 new_isn = (tcp_seq) md5_buffer[0]; 1604 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1605 return (new_isn); 1606 } 1607 1608 /* 1609 * When a source quench is received, close congestion window 1610 * to one segment. We will gradually open it again as we proceed. 1611 */ 1612 void 1613 tcp_quench(struct inpcb *inp, int error) 1614 { 1615 struct tcpcb *tp = intotcpcb(inp); 1616 1617 if (tp != NULL) { 1618 tp->snd_cwnd = tp->t_maxseg; 1619 tp->snd_wacked = 0; 1620 } 1621 } 1622 1623 /* 1624 * When a specific ICMP unreachable message is received and the 1625 * connection state is SYN-SENT, drop the connection. This behavior 1626 * is controlled by the icmp_may_rst sysctl. 1627 */ 1628 void 1629 tcp_drop_syn_sent(struct inpcb *inp, int error) 1630 { 1631 struct tcpcb *tp = intotcpcb(inp); 1632 1633 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT)) 1634 tcp_drop(tp, error); 1635 } 1636 1637 /* 1638 * When a `need fragmentation' ICMP is received, update our idea of the MSS 1639 * based on the new value in the route. Also nudge TCP to send something, 1640 * since we know the packet we just sent was dropped. 1641 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1642 */ 1643 void 1644 tcp_mtudisc(struct inpcb *inp, int mtu) 1645 { 1646 struct tcpcb *tp = intotcpcb(inp); 1647 struct rtentry *rt; 1648 struct socket *so = inp->inp_socket; 1649 int maxopd, mss; 1650 #ifdef INET6 1651 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0); 1652 #else 1653 const boolean_t isipv6 = FALSE; 1654 #endif 1655 1656 if (tp == NULL) 1657 return; 1658 1659 /* 1660 * If no MTU is provided in the ICMP message, use the 1661 * next lower likely value, as specified in RFC 1191. 1662 */ 1663 if (mtu == 0) { 1664 int oldmtu; 1665 1666 oldmtu = tp->t_maxopd + 1667 (isipv6 ? 1668 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1669 sizeof(struct tcpiphdr)); 1670 mtu = ip_next_mtu(oldmtu, 0); 1671 } 1672 1673 if (isipv6) 1674 rt = tcp_rtlookup6(&inp->inp_inc); 1675 else 1676 rt = tcp_rtlookup(&inp->inp_inc); 1677 if (rt != NULL) { 1678 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu) 1679 mtu = rt->rt_rmx.rmx_mtu; 1680 1681 maxopd = mtu - 1682 (isipv6 ? 1683 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1684 sizeof(struct tcpiphdr)); 1685 1686 /* 1687 * XXX - The following conditional probably violates the TCP 1688 * spec. The problem is that, since we don't know the 1689 * other end's MSS, we are supposed to use a conservative 1690 * default. But, if we do that, then MTU discovery will 1691 * never actually take place, because the conservative 1692 * default is much less than the MTUs typically seen 1693 * on the Internet today. For the moment, we'll sweep 1694 * this under the carpet. 1695 * 1696 * The conservative default might not actually be a problem 1697 * if the only case this occurs is when sending an initial 1698 * SYN with options and data to a host we've never talked 1699 * to before. Then, they will reply with an MSS value which 1700 * will get recorded and the new parameters should get 1701 * recomputed. For Further Study. 1702 */ 1703 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd) 1704 maxopd = rt->rt_rmx.rmx_mssopt; 1705 } else 1706 maxopd = mtu - 1707 (isipv6 ? 1708 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1709 sizeof(struct tcpiphdr)); 1710 1711 if (tp->t_maxopd <= maxopd) 1712 return; 1713 tp->t_maxopd = maxopd; 1714 1715 mss = maxopd; 1716 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) == 1717 (TF_REQ_TSTMP | TF_RCVD_TSTMP)) 1718 mss -= TCPOLEN_TSTAMP_APPA; 1719 1720 /* round down to multiple of MCLBYTES */ 1721 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */ 1722 if (mss > MCLBYTES) 1723 mss &= ~(MCLBYTES - 1); 1724 #else 1725 if (mss > MCLBYTES) 1726 mss = (mss / MCLBYTES) * MCLBYTES; 1727 #endif 1728 1729 if (so->so_snd.ssb_hiwat < mss) 1730 mss = so->so_snd.ssb_hiwat; 1731 1732 tp->t_maxseg = mss; 1733 tp->t_rtttime = 0; 1734 tp->snd_nxt = tp->snd_una; 1735 tcp_output(tp); 1736 tcpstat.tcps_mturesent++; 1737 } 1738 1739 /* 1740 * Look-up the routing entry to the peer of this inpcb. If no route 1741 * is found and it cannot be allocated the return NULL. This routine 1742 * is called by TCP routines that access the rmx structure and by tcp_mss 1743 * to get the interface MTU. 1744 */ 1745 struct rtentry * 1746 tcp_rtlookup(struct in_conninfo *inc) 1747 { 1748 struct route *ro = &inc->inc_route; 1749 1750 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1751 /* No route yet, so try to acquire one */ 1752 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1753 /* 1754 * unused portions of the structure MUST be zero'd 1755 * out because rtalloc() treats it as opaque data 1756 */ 1757 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1758 ro->ro_dst.sa_family = AF_INET; 1759 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1760 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1761 inc->inc_faddr; 1762 rtalloc(ro); 1763 } 1764 } 1765 return (ro->ro_rt); 1766 } 1767 1768 #ifdef INET6 1769 struct rtentry * 1770 tcp_rtlookup6(struct in_conninfo *inc) 1771 { 1772 struct route_in6 *ro6 = &inc->inc6_route; 1773 1774 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1775 /* No route yet, so try to acquire one */ 1776 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1777 /* 1778 * unused portions of the structure MUST be zero'd 1779 * out because rtalloc() treats it as opaque data 1780 */ 1781 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1782 ro6->ro_dst.sin6_family = AF_INET6; 1783 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1784 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1785 rtalloc((struct route *)ro6); 1786 } 1787 } 1788 return (ro6->ro_rt); 1789 } 1790 #endif 1791 1792 #ifdef IPSEC 1793 /* compute ESP/AH header size for TCP, including outer IP header. */ 1794 size_t 1795 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1796 { 1797 struct inpcb *inp; 1798 struct mbuf *m; 1799 size_t hdrsiz; 1800 struct ip *ip; 1801 struct tcphdr *th; 1802 1803 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1804 return (0); 1805 MGETHDR(m, MB_DONTWAIT, MT_DATA); 1806 if (!m) 1807 return (0); 1808 1809 #ifdef INET6 1810 if (inp->inp_vflag & INP_IPV6) { 1811 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1812 1813 th = (struct tcphdr *)(ip6 + 1); 1814 m->m_pkthdr.len = m->m_len = 1815 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1816 tcp_fillheaders(tp, ip6, th); 1817 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1818 } else 1819 #endif 1820 { 1821 ip = mtod(m, struct ip *); 1822 th = (struct tcphdr *)(ip + 1); 1823 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1824 tcp_fillheaders(tp, ip, th); 1825 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1826 } 1827 1828 m_free(m); 1829 return (hdrsiz); 1830 } 1831 #endif 1832 1833 /* 1834 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1835 * 1836 * This code attempts to calculate the bandwidth-delay product as a 1837 * means of determining the optimal window size to maximize bandwidth, 1838 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1839 * routers. This code also does a fairly good job keeping RTTs in check 1840 * across slow links like modems. We implement an algorithm which is very 1841 * similar (but not meant to be) TCP/Vegas. The code operates on the 1842 * transmitter side of a TCP connection and so only effects the transmit 1843 * side of the connection. 1844 * 1845 * BACKGROUND: TCP makes no provision for the management of buffer space 1846 * at the end points or at the intermediate routers and switches. A TCP 1847 * stream, whether using NewReno or not, will eventually buffer as 1848 * many packets as it is able and the only reason this typically works is 1849 * due to the fairly small default buffers made available for a connection 1850 * (typicaly 16K or 32K). As machines use larger windows and/or window 1851 * scaling it is now fairly easy for even a single TCP connection to blow-out 1852 * all available buffer space not only on the local interface, but on 1853 * intermediate routers and switches as well. NewReno makes a misguided 1854 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1855 * then backing off, then steadily increasing the window again until another 1856 * failure occurs, ad-infinitum. This results in terrible oscillation that 1857 * is only made worse as network loads increase and the idea of intentionally 1858 * blowing out network buffers is, frankly, a terrible way to manage network 1859 * resources. 1860 * 1861 * It is far better to limit the transmit window prior to the failure 1862 * condition being achieved. There are two general ways to do this: First 1863 * you can 'scan' through different transmit window sizes and locate the 1864 * point where the RTT stops increasing, indicating that you have filled the 1865 * pipe, then scan backwards until you note that RTT stops decreasing, then 1866 * repeat ad-infinitum. This method works in principle but has severe 1867 * implementation issues due to RTT variances, timer granularity, and 1868 * instability in the algorithm which can lead to many false positives and 1869 * create oscillations as well as interact badly with other TCP streams 1870 * implementing the same algorithm. 1871 * 1872 * The second method is to limit the window to the bandwidth delay product 1873 * of the link. This is the method we implement. RTT variances and our 1874 * own manipulation of the congestion window, bwnd, can potentially 1875 * destabilize the algorithm. For this reason we have to stabilize the 1876 * elements used to calculate the window. We do this by using the minimum 1877 * observed RTT, the long term average of the observed bandwidth, and 1878 * by adding two segments worth of slop. It isn't perfect but it is able 1879 * to react to changing conditions and gives us a very stable basis on 1880 * which to extend the algorithm. 1881 */ 1882 void 1883 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1884 { 1885 u_long bw; 1886 u_long bwnd; 1887 int save_ticks; 1888 int delta_ticks; 1889 1890 /* 1891 * If inflight_enable is disabled in the middle of a tcp connection, 1892 * make sure snd_bwnd is effectively disabled. 1893 */ 1894 if (!tcp_inflight_enable) { 1895 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1896 tp->snd_bandwidth = 0; 1897 return; 1898 } 1899 1900 /* 1901 * Validate the delta time. If a connection is new or has been idle 1902 * a long time we have to reset the bandwidth calculator. 1903 */ 1904 save_ticks = ticks; 1905 delta_ticks = save_ticks - tp->t_bw_rtttime; 1906 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1907 tp->t_bw_rtttime = ticks; 1908 tp->t_bw_rtseq = ack_seq; 1909 if (tp->snd_bandwidth == 0) 1910 tp->snd_bandwidth = tcp_inflight_min; 1911 return; 1912 } 1913 if (delta_ticks == 0) 1914 return; 1915 1916 /* 1917 * Sanity check, plus ignore pure window update acks. 1918 */ 1919 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1920 return; 1921 1922 /* 1923 * Figure out the bandwidth. Due to the tick granularity this 1924 * is a very rough number and it MUST be averaged over a fairly 1925 * long period of time. XXX we need to take into account a link 1926 * that is not using all available bandwidth, but for now our 1927 * slop will ramp us up if this case occurs and the bandwidth later 1928 * increases. 1929 */ 1930 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 1931 tp->t_bw_rtttime = save_ticks; 1932 tp->t_bw_rtseq = ack_seq; 1933 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4; 1934 1935 tp->snd_bandwidth = bw; 1936 1937 /* 1938 * Calculate the semi-static bandwidth delay product, plus two maximal 1939 * segments. The additional slop puts us squarely in the sweet 1940 * spot and also handles the bandwidth run-up case. Without the 1941 * slop we could be locking ourselves into a lower bandwidth. 1942 * 1943 * Situations Handled: 1944 * (1) Prevents over-queueing of packets on LANs, especially on 1945 * high speed LANs, allowing larger TCP buffers to be 1946 * specified, and also does a good job preventing 1947 * over-queueing of packets over choke points like modems 1948 * (at least for the transmit side). 1949 * 1950 * (2) Is able to handle changing network loads (bandwidth 1951 * drops so bwnd drops, bandwidth increases so bwnd 1952 * increases). 1953 * 1954 * (3) Theoretically should stabilize in the face of multiple 1955 * connections implementing the same algorithm (this may need 1956 * a little work). 1957 * 1958 * (4) Stability value (defaults to 20 = 2 maximal packets) can 1959 * be adjusted with a sysctl but typically only needs to be on 1960 * very slow connections. A value no smaller then 5 should 1961 * be used, but only reduce this default if you have no other 1962 * choice. 1963 */ 1964 1965 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2) 1966 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 1967 tcp_inflight_stab * (int)tp->t_maxseg / 10; 1968 #undef USERTT 1969 1970 if (tcp_inflight_debug > 0) { 1971 static int ltime; 1972 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) { 1973 ltime = ticks; 1974 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n", 1975 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd); 1976 } 1977 } 1978 if ((long)bwnd < tcp_inflight_min) 1979 bwnd = tcp_inflight_min; 1980 if (bwnd > tcp_inflight_max) 1981 bwnd = tcp_inflight_max; 1982 if ((long)bwnd < tp->t_maxseg * 2) 1983 bwnd = tp->t_maxseg * 2; 1984 tp->snd_bwnd = bwnd; 1985 } 1986 1987 static void 1988 tcp_rmx_iwsegs(struct tcpcb *tp, u_long *maxsegs, u_long *capsegs) 1989 { 1990 struct rtentry *rt; 1991 struct inpcb *inp = tp->t_inpcb; 1992 #ifdef INET6 1993 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE); 1994 #else 1995 const boolean_t isipv6 = FALSE; 1996 #endif 1997 1998 /* XXX */ 1999 if (tcp_iw_maxsegs < TCP_IW_MAXSEGS_DFLT) 2000 tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 2001 if (tcp_iw_capsegs < TCP_IW_CAPSEGS_DFLT) 2002 tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 2003 2004 if (isipv6) 2005 rt = tcp_rtlookup6(&inp->inp_inc); 2006 else 2007 rt = tcp_rtlookup(&inp->inp_inc); 2008 if (rt == NULL || 2009 rt->rt_rmx.rmx_iwmaxsegs < TCP_IW_MAXSEGS_DFLT || 2010 rt->rt_rmx.rmx_iwcapsegs < TCP_IW_CAPSEGS_DFLT) { 2011 *maxsegs = tcp_iw_maxsegs; 2012 *capsegs = tcp_iw_capsegs; 2013 return; 2014 } 2015 *maxsegs = rt->rt_rmx.rmx_iwmaxsegs; 2016 *capsegs = rt->rt_rmx.rmx_iwcapsegs; 2017 } 2018 2019 u_long 2020 tcp_initial_window(struct tcpcb *tp) 2021 { 2022 if (tcp_do_rfc3390) { 2023 /* 2024 * RFC3390: 2025 * "If the SYN or SYN/ACK is lost, the initial window 2026 * used by a sender after a correctly transmitted SYN 2027 * MUST be one segment consisting of MSS bytes." 2028 * 2029 * However, we do something a little bit more aggressive 2030 * then RFC3390 here: 2031 * - Only if time spent in the SYN or SYN|ACK retransmition 2032 * >= 3 seconds, the IW is reduced. We do this mainly 2033 * because when RFC3390 is published, the initial RTO is 2034 * still 3 seconds (the threshold we test here), while 2035 * after RFC6298, the initial RTO is 1 second. This 2036 * behaviour probably still falls within the spirit of 2037 * RFC3390. 2038 * - When IW is reduced, 2*MSS is used instead of 1*MSS. 2039 * Mainly to avoid sender and receiver deadlock until 2040 * delayed ACK timer expires. And even RFC2581 does not 2041 * try to reduce IW upon SYN or SYN|ACK retransmition 2042 * timeout. 2043 * 2044 * See also: 2045 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03 2046 */ 2047 if (tp->t_rxtsyn >= TCPTV_RTOBASE3) { 2048 return (2 * tp->t_maxseg); 2049 } else { 2050 u_long maxsegs, capsegs; 2051 2052 tcp_rmx_iwsegs(tp, &maxsegs, &capsegs); 2053 return min(maxsegs * tp->t_maxseg, 2054 max(2 * tp->t_maxseg, capsegs * 1460)); 2055 } 2056 } else { 2057 /* 2058 * Even RFC2581 (back to 1999) allows 2*SMSS IW. 2059 * 2060 * Mainly to avoid sender and receiver deadlock 2061 * until delayed ACK timer expires. 2062 */ 2063 return (2 * tp->t_maxseg); 2064 } 2065 } 2066 2067 #ifdef TCP_SIGNATURE 2068 /* 2069 * Compute TCP-MD5 hash of a TCP segment. (RFC2385) 2070 * 2071 * We do this over ip, tcphdr, segment data, and the key in the SADB. 2072 * When called from tcp_input(), we can be sure that th_sum has been 2073 * zeroed out and verified already. 2074 * 2075 * Return 0 if successful, otherwise return -1. 2076 * 2077 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a 2078 * search with the destination IP address, and a 'magic SPI' to be 2079 * determined by the application. This is hardcoded elsewhere to 1179 2080 * right now. Another branch of this code exists which uses the SPD to 2081 * specify per-application flows but it is unstable. 2082 */ 2083 int 2084 tcpsignature_compute( 2085 struct mbuf *m, /* mbuf chain */ 2086 int len, /* length of TCP data */ 2087 int optlen, /* length of TCP options */ 2088 u_char *buf, /* storage for MD5 digest */ 2089 u_int direction) /* direction of flow */ 2090 { 2091 struct ippseudo ippseudo; 2092 MD5_CTX ctx; 2093 int doff; 2094 struct ip *ip; 2095 struct ipovly *ipovly; 2096 struct secasvar *sav; 2097 struct tcphdr *th; 2098 #ifdef INET6 2099 struct ip6_hdr *ip6; 2100 struct in6_addr in6; 2101 uint32_t plen; 2102 uint16_t nhdr; 2103 #endif /* INET6 */ 2104 u_short savecsum; 2105 2106 KASSERT(m != NULL, ("passed NULL mbuf. Game over.")); 2107 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature")); 2108 /* 2109 * Extract the destination from the IP header in the mbuf. 2110 */ 2111 ip = mtod(m, struct ip *); 2112 #ifdef INET6 2113 ip6 = NULL; /* Make the compiler happy. */ 2114 #endif /* INET6 */ 2115 /* 2116 * Look up an SADB entry which matches the address found in 2117 * the segment. 2118 */ 2119 switch (IP_VHL_V(ip->ip_vhl)) { 2120 case IPVERSION: 2121 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst, 2122 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2123 break; 2124 #ifdef INET6 2125 case (IPV6_VERSION >> 4): 2126 ip6 = mtod(m, struct ip6_hdr *); 2127 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst, 2128 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2129 break; 2130 #endif /* INET6 */ 2131 default: 2132 return (EINVAL); 2133 /* NOTREACHED */ 2134 break; 2135 } 2136 if (sav == NULL) { 2137 kprintf("%s: SADB lookup failed\n", __func__); 2138 return (EINVAL); 2139 } 2140 MD5Init(&ctx); 2141 2142 /* 2143 * Step 1: Update MD5 hash with IP pseudo-header. 2144 * 2145 * XXX The ippseudo header MUST be digested in network byte order, 2146 * or else we'll fail the regression test. Assume all fields we've 2147 * been doing arithmetic on have been in host byte order. 2148 * XXX One cannot depend on ipovly->ih_len here. When called from 2149 * tcp_output(), the underlying ip_len member has not yet been set. 2150 */ 2151 switch (IP_VHL_V(ip->ip_vhl)) { 2152 case IPVERSION: 2153 ipovly = (struct ipovly *)ip; 2154 ippseudo.ippseudo_src = ipovly->ih_src; 2155 ippseudo.ippseudo_dst = ipovly->ih_dst; 2156 ippseudo.ippseudo_pad = 0; 2157 ippseudo.ippseudo_p = IPPROTO_TCP; 2158 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen); 2159 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo)); 2160 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip)); 2161 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen; 2162 break; 2163 #ifdef INET6 2164 /* 2165 * RFC 2385, 2.0 Proposal 2166 * For IPv6, the pseudo-header is as described in RFC 2460, namely the 2167 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero- 2168 * extended next header value (to form 32 bits), and 32-bit segment 2169 * length. 2170 * Note: Upper-Layer Packet Length comes before Next Header. 2171 */ 2172 case (IPV6_VERSION >> 4): 2173 in6 = ip6->ip6_src; 2174 in6_clearscope(&in6); 2175 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2176 in6 = ip6->ip6_dst; 2177 in6_clearscope(&in6); 2178 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2179 plen = htonl(len + sizeof(struct tcphdr) + optlen); 2180 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t)); 2181 nhdr = 0; 2182 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2183 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2184 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2185 nhdr = IPPROTO_TCP; 2186 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2187 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr)); 2188 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen; 2189 break; 2190 #endif /* INET6 */ 2191 default: 2192 return (EINVAL); 2193 /* NOTREACHED */ 2194 break; 2195 } 2196 /* 2197 * Step 2: Update MD5 hash with TCP header, excluding options. 2198 * The TCP checksum must be set to zero. 2199 */ 2200 savecsum = th->th_sum; 2201 th->th_sum = 0; 2202 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr)); 2203 th->th_sum = savecsum; 2204 /* 2205 * Step 3: Update MD5 hash with TCP segment data. 2206 * Use m_apply() to avoid an early m_pullup(). 2207 */ 2208 if (len > 0) 2209 m_apply(m, doff, len, tcpsignature_apply, &ctx); 2210 /* 2211 * Step 4: Update MD5 hash with shared secret. 2212 */ 2213 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth)); 2214 MD5Final(buf, &ctx); 2215 key_sa_recordxfer(sav, m); 2216 key_freesav(sav); 2217 return (0); 2218 } 2219 2220 int 2221 tcpsignature_apply(void *fstate, void *data, unsigned int len) 2222 { 2223 2224 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len); 2225 return (0); 2226 } 2227 #endif /* TCP_SIGNATURE */ 2228