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. Neither the name of the University nor the names of its contributors 47 * may be used to endorse or promote products derived from this software 48 * without specific prior written permission. 49 * 50 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 51 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 52 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 53 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 54 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 55 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 56 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 57 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 58 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 59 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 60 * SUCH DAMAGE. 61 * 62 * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95 63 * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $ 64 */ 65 66 #include "opt_compat.h" 67 #include "opt_inet.h" 68 #include "opt_inet6.h" 69 #include "opt_ipsec.h" 70 #include "opt_tcpdebug.h" 71 72 #include <sys/param.h> 73 #include <sys/systm.h> 74 #include <sys/callout.h> 75 #include <sys/kernel.h> 76 #include <sys/sysctl.h> 77 #include <sys/malloc.h> 78 #include <sys/mpipe.h> 79 #include <sys/mbuf.h> 80 #ifdef INET6 81 #include <sys/domain.h> 82 #endif 83 #include <sys/proc.h> 84 #include <sys/priv.h> 85 #include <sys/socket.h> 86 #include <sys/socketops.h> 87 #include <sys/socketvar.h> 88 #include <sys/protosw.h> 89 #include <sys/random.h> 90 #include <sys/in_cksum.h> 91 #include <sys/ktr.h> 92 93 #include <net/route.h> 94 #include <net/if.h> 95 #include <net/netisr2.h> 96 97 #define _IP_VHL 98 #include <netinet/in.h> 99 #include <netinet/in_systm.h> 100 #include <netinet/ip.h> 101 #include <netinet/ip6.h> 102 #include <netinet/in_pcb.h> 103 #include <netinet6/in6_pcb.h> 104 #include <netinet/in_var.h> 105 #include <netinet/ip_var.h> 106 #include <netinet6/ip6_var.h> 107 #include <netinet/ip_icmp.h> 108 #ifdef INET6 109 #include <netinet/icmp6.h> 110 #endif 111 #include <netinet/tcp.h> 112 #include <netinet/tcp_fsm.h> 113 #include <netinet/tcp_seq.h> 114 #include <netinet/tcp_timer.h> 115 #include <netinet/tcp_timer2.h> 116 #include <netinet/tcp_var.h> 117 #include <netinet6/tcp6_var.h> 118 #include <netinet/tcpip.h> 119 #ifdef TCPDEBUG 120 #include <netinet/tcp_debug.h> 121 #endif 122 #include <netinet6/ip6protosw.h> 123 124 #ifdef IPSEC 125 #include <netinet6/ipsec.h> 126 #include <netproto/key/key.h> 127 #ifdef INET6 128 #include <netinet6/ipsec6.h> 129 #endif 130 #endif 131 132 #ifdef FAST_IPSEC 133 #include <netproto/ipsec/ipsec.h> 134 #ifdef INET6 135 #include <netproto/ipsec/ipsec6.h> 136 #endif 137 #define IPSEC 138 #endif 139 140 #include <sys/md5.h> 141 #include <machine/smp.h> 142 143 #include <sys/msgport2.h> 144 #include <sys/mplock2.h> 145 #include <net/netmsg2.h> 146 147 #if !defined(KTR_TCP) 148 #define KTR_TCP KTR_ALL 149 #endif 150 /* 151 KTR_INFO_MASTER(tcp); 152 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0); 153 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0); 154 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0); 155 #define logtcp(name) KTR_LOG(tcp_ ## name) 156 */ 157 158 #define TCP_IW_MAXSEGS_DFLT 4 159 #define TCP_IW_CAPSEGS_DFLT 3 160 161 struct inpcbinfo tcbinfo[MAXCPU]; 162 struct tcpcbackqhead tcpcbackq[MAXCPU]; 163 164 static struct lwkt_token tcp_port_token = 165 LWKT_TOKEN_INITIALIZER(tcp_port_token); 166 167 int tcp_mssdflt = TCP_MSS; 168 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 169 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size"); 170 171 #ifdef INET6 172 int tcp_v6mssdflt = TCP6_MSS; 173 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW, 174 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6"); 175 #endif 176 177 /* 178 * Minimum MSS we accept and use. This prevents DoS attacks where 179 * we are forced to a ridiculous low MSS like 20 and send hundreds 180 * of packets instead of one. The effect scales with the available 181 * bandwidth and quickly saturates the CPU and network interface 182 * with packet generation and sending. Set to zero to disable MINMSS 183 * checking. This setting prevents us from sending too small packets. 184 */ 185 int tcp_minmss = TCP_MINMSS; 186 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW, 187 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size"); 188 189 #if 0 190 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ; 191 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW, 192 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time"); 193 #endif 194 195 int tcp_do_rfc1323 = 1; 196 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW, 197 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions"); 198 199 static int tcp_tcbhashsize = 0; 200 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD, 201 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable"); 202 203 static int do_tcpdrain = 1; 204 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0, 205 "Enable tcp_drain routine for extra help when low on mbufs"); 206 207 static int icmp_may_rst = 1; 208 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0, 209 "Certain ICMP unreachable messages may abort connections in SYN_SENT"); 210 211 static int tcp_isn_reseed_interval = 0; 212 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW, 213 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret"); 214 215 /* 216 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on 217 * by default, but with generous values which should allow maximal 218 * bandwidth. In particular, the slop defaults to 50 (5 packets). 219 * 220 * The reason for doing this is that the limiter is the only mechanism we 221 * have which seems to do a really good job preventing receiver RX rings 222 * on network interfaces from getting blown out. Even though GigE/10GigE 223 * is supposed to flow control it looks like either it doesn't actually 224 * do it or Open Source drivers do not properly enable it. 225 * 226 * People using the limiter to reduce bottlenecks on slower WAN connections 227 * should set the slop to 20 (2 packets). 228 */ 229 static int tcp_inflight_enable = 1; 230 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW, 231 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting"); 232 233 static int tcp_inflight_debug = 0; 234 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW, 235 &tcp_inflight_debug, 0, "Debug TCP inflight calculations"); 236 237 static int tcp_inflight_min = 6144; 238 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW, 239 &tcp_inflight_min, 0, "Lower bound for TCP inflight window"); 240 241 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT; 242 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW, 243 &tcp_inflight_max, 0, "Upper bound for TCP inflight window"); 244 245 static int tcp_inflight_stab = 50; 246 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW, 247 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 3 packets)"); 248 249 static int tcp_do_rfc3390 = 1; 250 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_RW, 251 &tcp_do_rfc3390, 0, 252 "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)"); 253 254 static u_long tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 255 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwmaxsegs, CTLFLAG_RW, 256 &tcp_iw_maxsegs, 0, "TCP IW segments max"); 257 258 static u_long tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 259 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwcapsegs, CTLFLAG_RW, 260 &tcp_iw_capsegs, 0, "TCP IW segments"); 261 262 int tcp_low_rtobase = 1; 263 SYSCTL_INT(_net_inet_tcp, OID_AUTO, low_rtobase, CTLFLAG_RW, 264 &tcp_low_rtobase, 0, "Lowering the Initial RTO (RFC 6298)"); 265 266 static int tcp_do_ncr = 1; 267 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr, CTLFLAG_RW, 268 &tcp_do_ncr, 0, "Non-Congestion Robustness (RFC 4653)"); 269 270 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives"); 271 static struct malloc_pipe tcptemp_mpipe; 272 273 static void tcp_willblock(void); 274 static void tcp_notify (struct inpcb *, int); 275 276 struct tcp_stats tcpstats_percpu[MAXCPU] __cachealign; 277 278 static int 279 sysctl_tcpstats(SYSCTL_HANDLER_ARGS) 280 { 281 int cpu, error = 0; 282 283 for (cpu = 0; cpu < ncpus; ++cpu) { 284 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu], 285 sizeof(struct tcp_stats)))) 286 break; 287 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu], 288 sizeof(struct tcp_stats)))) 289 break; 290 } 291 292 return (error); 293 } 294 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW), 295 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics"); 296 297 /* 298 * Target size of TCP PCB hash tables. Must be a power of two. 299 * 300 * Note that this can be overridden by the kernel environment 301 * variable net.inet.tcp.tcbhashsize 302 */ 303 #ifndef TCBHASHSIZE 304 #define TCBHASHSIZE 512 305 #endif 306 307 /* 308 * This is the actual shape of what we allocate using the zone 309 * allocator. Doing it this way allows us to protect both structures 310 * using the same generation count, and also eliminates the overhead 311 * of allocating tcpcbs separately. By hiding the structure here, 312 * we avoid changing most of the rest of the code (although it needs 313 * to be changed, eventually, for greater efficiency). 314 */ 315 #define ALIGNMENT 32 316 #define ALIGNM1 (ALIGNMENT - 1) 317 struct inp_tp { 318 union { 319 struct inpcb inp; 320 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1]; 321 } inp_tp_u; 322 struct tcpcb tcb; 323 struct tcp_callout inp_tp_rexmt; 324 struct tcp_callout inp_tp_persist; 325 struct tcp_callout inp_tp_keep; 326 struct tcp_callout inp_tp_2msl; 327 struct tcp_callout inp_tp_delack; 328 struct netmsg_tcp_timer inp_tp_timermsg; 329 struct netmsg_base inp_tp_sndmore; 330 }; 331 #undef ALIGNMENT 332 #undef ALIGNM1 333 334 /* 335 * Tcp initialization 336 */ 337 void 338 tcp_init(void) 339 { 340 struct inpcbporthead *porthashbase; 341 struct inpcbinfo *ticb; 342 u_long porthashmask; 343 int hashsize = TCBHASHSIZE; 344 int cpu; 345 346 /* 347 * note: tcptemp is used for keepalives, and it is ok for an 348 * allocation to fail so do not specify MPF_INT. 349 */ 350 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp), 351 25, -1, 0, NULL, NULL, NULL); 352 353 tcp_delacktime = TCPTV_DELACK; 354 tcp_keepinit = TCPTV_KEEP_INIT; 355 tcp_keepidle = TCPTV_KEEP_IDLE; 356 tcp_keepintvl = TCPTV_KEEPINTVL; 357 tcp_maxpersistidle = TCPTV_KEEP_IDLE; 358 tcp_msl = TCPTV_MSL; 359 tcp_rexmit_min = TCPTV_MIN; 360 tcp_rexmit_slop = TCPTV_CPU_VAR; 361 362 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize); 363 if (!powerof2(hashsize)) { 364 kprintf("WARNING: TCB hash size not a power of 2\n"); 365 hashsize = 512; /* safe default */ 366 } 367 tcp_tcbhashsize = hashsize; 368 porthashbase = hashinit(hashsize, M_PCB, &porthashmask); 369 370 for (cpu = 0; cpu < ncpus2; cpu++) { 371 ticb = &tcbinfo[cpu]; 372 in_pcbinfo_init(ticb); 373 ticb->cpu = cpu; 374 ticb->hashbase = hashinit(hashsize, M_PCB, 375 &ticb->hashmask); 376 ticb->porthashbase = porthashbase; 377 ticb->porthashmask = porthashmask; 378 ticb->porttoken = &tcp_port_token; 379 ticb->wildcardhashbase = hashinit(hashsize, M_PCB, 380 &ticb->wildcardhashmask); 381 ticb->localgrphashbase = hashinit(hashsize, M_PCB, 382 &ticb->localgrphashmask); 383 ticb->ipi_size = sizeof(struct inp_tp); 384 TAILQ_INIT(&tcpcbackq[cpu]); 385 } 386 387 tcp_reass_maxseg = nmbclusters / 16; 388 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg); 389 390 #ifdef INET6 391 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) 392 #else 393 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) 394 #endif 395 if (max_protohdr < TCP_MINPROTOHDR) 396 max_protohdr = TCP_MINPROTOHDR; 397 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) 398 panic("tcp_init"); 399 #undef TCP_MINPROTOHDR 400 401 /* 402 * Initialize TCP statistics counters for each CPU. 403 */ 404 for (cpu = 0; cpu < ncpus; ++cpu) { 405 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats)); 406 } 407 408 syncache_init(); 409 netisr_register_rollup(tcp_willblock, NETISR_ROLLUP_PRIO_TCP); 410 } 411 412 static void 413 tcp_willblock(void) 414 { 415 struct tcpcb *tp; 416 int cpu = mycpu->gd_cpuid; 417 418 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) { 419 KKASSERT(tp->t_flags & TF_ONOUTPUTQ); 420 tp->t_flags &= ~TF_ONOUTPUTQ; 421 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq); 422 tcp_output(tp); 423 } 424 } 425 426 /* 427 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. 428 * tcp_template used to store this data in mbufs, but we now recopy it out 429 * of the tcpcb each time to conserve mbufs. 430 */ 431 void 432 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr, boolean_t tso) 433 { 434 struct inpcb *inp = tp->t_inpcb; 435 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr; 436 437 #ifdef INET6 438 if (inp->inp_vflag & INP_IPV6) { 439 struct ip6_hdr *ip6; 440 441 ip6 = (struct ip6_hdr *)ip_ptr; 442 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | 443 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK); 444 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | 445 (IPV6_VERSION & IPV6_VERSION_MASK); 446 ip6->ip6_nxt = IPPROTO_TCP; 447 ip6->ip6_plen = sizeof(struct tcphdr); 448 ip6->ip6_src = inp->in6p_laddr; 449 ip6->ip6_dst = inp->in6p_faddr; 450 tcp_hdr->th_sum = 0; 451 } else 452 #endif 453 { 454 struct ip *ip = (struct ip *) ip_ptr; 455 u_int plen; 456 457 ip->ip_vhl = IP_VHL_BORING; 458 ip->ip_tos = 0; 459 ip->ip_len = 0; 460 ip->ip_id = 0; 461 ip->ip_off = 0; 462 ip->ip_ttl = 0; 463 ip->ip_sum = 0; 464 ip->ip_p = IPPROTO_TCP; 465 ip->ip_src = inp->inp_laddr; 466 ip->ip_dst = inp->inp_faddr; 467 468 if (tso) 469 plen = htons(IPPROTO_TCP); 470 else 471 plen = htons(sizeof(struct tcphdr) + IPPROTO_TCP); 472 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr, 473 ip->ip_dst.s_addr, plen); 474 } 475 476 tcp_hdr->th_sport = inp->inp_lport; 477 tcp_hdr->th_dport = inp->inp_fport; 478 tcp_hdr->th_seq = 0; 479 tcp_hdr->th_ack = 0; 480 tcp_hdr->th_x2 = 0; 481 tcp_hdr->th_off = 5; 482 tcp_hdr->th_flags = 0; 483 tcp_hdr->th_win = 0; 484 tcp_hdr->th_urp = 0; 485 } 486 487 /* 488 * Create template to be used to send tcp packets on a connection. 489 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only 490 * use for this function is in keepalives, which use tcp_respond. 491 */ 492 struct tcptemp * 493 tcp_maketemplate(struct tcpcb *tp) 494 { 495 struct tcptemp *tmp; 496 497 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL) 498 return (NULL); 499 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t, FALSE); 500 return (tmp); 501 } 502 503 void 504 tcp_freetemplate(struct tcptemp *tmp) 505 { 506 mpipe_free(&tcptemp_mpipe, tmp); 507 } 508 509 /* 510 * Send a single message to the TCP at address specified by 511 * the given TCP/IP header. If m == NULL, then we make a copy 512 * of the tcpiphdr at ti and send directly to the addressed host. 513 * This is used to force keep alive messages out using the TCP 514 * template for a connection. If flags are given then we send 515 * a message back to the TCP which originated the * segment ti, 516 * and discard the mbuf containing it and any other attached mbufs. 517 * 518 * In any case the ack and sequence number of the transmitted 519 * segment are as specified by the parameters. 520 * 521 * NOTE: If m != NULL, then ti must point to *inside* the mbuf. 522 */ 523 void 524 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, 525 tcp_seq ack, tcp_seq seq, int flags) 526 { 527 int tlen; 528 int win = 0; 529 struct route *ro = NULL; 530 struct route sro; 531 struct ip *ip = ipgen; 532 struct tcphdr *nth; 533 int ipflags = 0; 534 struct route_in6 *ro6 = NULL; 535 struct route_in6 sro6; 536 struct ip6_hdr *ip6 = ipgen; 537 boolean_t use_tmpro = TRUE; 538 #ifdef INET6 539 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6); 540 #else 541 const boolean_t isipv6 = FALSE; 542 #endif 543 544 if (tp != NULL) { 545 if (!(flags & TH_RST)) { 546 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv); 547 if (win < 0) 548 win = 0; 549 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 550 win = (long)TCP_MAXWIN << tp->rcv_scale; 551 } 552 /* 553 * Don't use the route cache of a listen socket, 554 * it is not MPSAFE; use temporary route cache. 555 */ 556 if (tp->t_state != TCPS_LISTEN) { 557 if (isipv6) 558 ro6 = &tp->t_inpcb->in6p_route; 559 else 560 ro = &tp->t_inpcb->inp_route; 561 use_tmpro = FALSE; 562 } 563 } 564 if (use_tmpro) { 565 if (isipv6) { 566 ro6 = &sro6; 567 bzero(ro6, sizeof *ro6); 568 } else { 569 ro = &sro; 570 bzero(ro, sizeof *ro); 571 } 572 } 573 if (m == NULL) { 574 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 575 if (m == NULL) 576 return; 577 tlen = 0; 578 m->m_data += max_linkhdr; 579 if (isipv6) { 580 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); 581 ip6 = mtod(m, struct ip6_hdr *); 582 nth = (struct tcphdr *)(ip6 + 1); 583 } else { 584 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip)); 585 ip = mtod(m, struct ip *); 586 nth = (struct tcphdr *)(ip + 1); 587 } 588 bcopy(th, nth, sizeof(struct tcphdr)); 589 flags = TH_ACK; 590 } else { 591 m_freem(m->m_next); 592 m->m_next = NULL; 593 m->m_data = (caddr_t)ipgen; 594 /* m_len is set later */ 595 tlen = 0; 596 #define xchg(a, b, type) { type t; t = a; a = b; b = t; } 597 if (isipv6) { 598 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); 599 nth = (struct tcphdr *)(ip6 + 1); 600 } else { 601 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long); 602 nth = (struct tcphdr *)(ip + 1); 603 } 604 if (th != nth) { 605 /* 606 * this is usually a case when an extension header 607 * exists between the IPv6 header and the 608 * TCP header. 609 */ 610 nth->th_sport = th->th_sport; 611 nth->th_dport = th->th_dport; 612 } 613 xchg(nth->th_dport, nth->th_sport, n_short); 614 #undef xchg 615 } 616 if (isipv6) { 617 ip6->ip6_flow = 0; 618 ip6->ip6_vfc = IPV6_VERSION; 619 ip6->ip6_nxt = IPPROTO_TCP; 620 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen)); 621 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 622 } else { 623 tlen += sizeof(struct tcpiphdr); 624 ip->ip_len = tlen; 625 ip->ip_ttl = ip_defttl; 626 } 627 m->m_len = tlen; 628 m->m_pkthdr.len = tlen; 629 m->m_pkthdr.rcvif = NULL; 630 nth->th_seq = htonl(seq); 631 nth->th_ack = htonl(ack); 632 nth->th_x2 = 0; 633 nth->th_off = sizeof(struct tcphdr) >> 2; 634 nth->th_flags = flags; 635 if (tp != NULL) 636 nth->th_win = htons((u_short) (win >> tp->rcv_scale)); 637 else 638 nth->th_win = htons((u_short)win); 639 nth->th_urp = 0; 640 if (isipv6) { 641 nth->th_sum = 0; 642 nth->th_sum = in6_cksum(m, IPPROTO_TCP, 643 sizeof(struct ip6_hdr), 644 tlen - sizeof(struct ip6_hdr)); 645 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL, 646 (ro6 && ro6->ro_rt) ? 647 ro6->ro_rt->rt_ifp : NULL); 648 } else { 649 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 650 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); 651 m->m_pkthdr.csum_flags = CSUM_TCP; 652 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 653 m->m_pkthdr.csum_thlen = sizeof(struct tcphdr); 654 } 655 #ifdef TCPDEBUG 656 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) 657 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); 658 #endif 659 if (isipv6) { 660 ip6_output(m, NULL, ro6, ipflags, NULL, NULL, 661 tp ? tp->t_inpcb : NULL); 662 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) { 663 RTFREE(ro6->ro_rt); 664 ro6->ro_rt = NULL; 665 } 666 } else { 667 ipflags |= IP_DEBUGROUTE; 668 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL); 669 if ((ro == &sro) && (ro->ro_rt != NULL)) { 670 RTFREE(ro->ro_rt); 671 ro->ro_rt = NULL; 672 } 673 } 674 } 675 676 /* 677 * Create a new TCP control block, making an 678 * empty reassembly queue and hooking it to the argument 679 * protocol control block. The `inp' parameter must have 680 * come from the zone allocator set up in tcp_init(). 681 */ 682 struct tcpcb * 683 tcp_newtcpcb(struct inpcb *inp) 684 { 685 struct inp_tp *it; 686 struct tcpcb *tp; 687 #ifdef INET6 688 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 689 #else 690 const boolean_t isipv6 = FALSE; 691 #endif 692 693 it = (struct inp_tp *)inp; 694 tp = &it->tcb; 695 bzero(tp, sizeof(struct tcpcb)); 696 TAILQ_INIT(&tp->t_segq); 697 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 698 tp->t_rxtthresh = tcprexmtthresh; 699 700 /* Set up our timeouts. */ 701 tp->tt_rexmt = &it->inp_tp_rexmt; 702 tp->tt_persist = &it->inp_tp_persist; 703 tp->tt_keep = &it->inp_tp_keep; 704 tp->tt_2msl = &it->inp_tp_2msl; 705 tp->tt_delack = &it->inp_tp_delack; 706 tcp_inittimers(tp); 707 708 /* 709 * Zero out timer message. We don't create it here, 710 * since the current CPU may not be the owner of this 711 * inpcb. 712 */ 713 tp->tt_msg = &it->inp_tp_timermsg; 714 bzero(tp->tt_msg, sizeof(*tp->tt_msg)); 715 716 tp->t_keepinit = tcp_keepinit; 717 tp->t_keepidle = tcp_keepidle; 718 tp->t_keepintvl = tcp_keepintvl; 719 tp->t_keepcnt = tcp_keepcnt; 720 tp->t_maxidle = tp->t_keepintvl * tp->t_keepcnt; 721 722 if (tcp_do_ncr) 723 tp->t_flags |= TF_NCR; 724 if (tcp_do_rfc1323) 725 tp->t_flags |= (TF_REQ_SCALE | TF_REQ_TSTMP); 726 727 tp->t_inpcb = inp; /* XXX */ 728 tp->t_state = TCPS_CLOSED; 729 /* 730 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 731 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 732 * reasonable initial retransmit time. 733 */ 734 tp->t_srtt = TCPTV_SRTTBASE; 735 tp->t_rttvar = 736 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 737 tp->t_rttmin = tcp_rexmit_min; 738 tp->t_rxtcur = TCPTV_RTOBASE; 739 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 740 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 741 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 742 tp->snd_last = ticks; 743 tp->t_rcvtime = ticks; 744 /* 745 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 746 * because the socket may be bound to an IPv6 wildcard address, 747 * which may match an IPv4-mapped IPv6 address. 748 */ 749 inp->inp_ip_ttl = ip_defttl; 750 inp->inp_ppcb = tp; 751 tcp_sack_tcpcb_init(tp); 752 753 tp->tt_sndmore = &it->inp_tp_sndmore; 754 tcp_output_init(tp); 755 756 return (tp); /* XXX */ 757 } 758 759 /* 760 * Drop a TCP connection, reporting the specified error. 761 * If connection is synchronized, then send a RST to peer. 762 */ 763 struct tcpcb * 764 tcp_drop(struct tcpcb *tp, int error) 765 { 766 struct socket *so = tp->t_inpcb->inp_socket; 767 768 if (TCPS_HAVERCVDSYN(tp->t_state)) { 769 tp->t_state = TCPS_CLOSED; 770 tcp_output(tp); 771 tcpstat.tcps_drops++; 772 } else 773 tcpstat.tcps_conndrops++; 774 if (error == ETIMEDOUT && tp->t_softerror) 775 error = tp->t_softerror; 776 so->so_error = error; 777 return (tcp_close(tp)); 778 } 779 780 struct netmsg_listen_detach { 781 struct netmsg_base base; 782 struct tcpcb *nm_tp; 783 struct tcpcb *nm_tp_inh; 784 }; 785 786 static void 787 tcp_listen_detach_handler(netmsg_t msg) 788 { 789 struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg; 790 struct tcpcb *tp = nmsg->nm_tp; 791 int cpu = mycpuid, nextcpu; 792 793 if (tp->t_flags & TF_LISTEN) 794 syncache_destroy(tp, nmsg->nm_tp_inh); 795 796 in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]); 797 798 nextcpu = cpu + 1; 799 if (nextcpu < ncpus2) 800 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nmsg->base.lmsg); 801 else 802 lwkt_replymsg(&nmsg->base.lmsg, 0); 803 } 804 805 /* 806 * Close a TCP control block: 807 * discard all space held by the tcp 808 * discard internet protocol block 809 * wake up any sleepers 810 */ 811 struct tcpcb * 812 tcp_close(struct tcpcb *tp) 813 { 814 struct tseg_qent *q; 815 struct inpcb *inp = tp->t_inpcb; 816 struct inpcb *inp_inh = NULL; 817 struct tcpcb *tp_inh = NULL; 818 struct socket *so = inp->inp_socket; 819 struct rtentry *rt; 820 boolean_t dosavessthresh; 821 #ifdef INET6 822 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 823 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0); 824 #else 825 const boolean_t isipv6 = FALSE; 826 #endif 827 828 if (tp->t_flags & TF_LISTEN) { 829 /* 830 * Pending socket/syncache inheritance 831 * 832 * If this is a listen(2) socket, find another listen(2) 833 * socket in the same local group, which could inherit 834 * the syncache and sockets pending on the completion 835 * and incompletion queues. 836 * 837 * NOTE: 838 * Currently the inheritance could only happen on the 839 * listen(2) sockets w/ SO_REUSEPORT set. 840 */ 841 KASSERT(&curthread->td_msgport == netisr_cpuport(0), 842 ("listen socket close not in netisr0")); 843 inp_inh = in_pcblocalgroup_last(&tcbinfo[0], inp); 844 if (inp_inh != NULL) 845 tp_inh = intotcpcb(inp_inh); 846 } 847 848 /* 849 * INP_WILDCARD_MP indicates that listen(2) has been called on 850 * this socket. This implies: 851 * - A wildcard inp's hash is replicated for each protocol thread. 852 * - Syncache for this inp grows independently in each protocol 853 * thread. 854 * - There is more than one cpu 855 * 856 * We have to chain a message to the rest of the protocol threads 857 * to cleanup the wildcard hash and the syncache. The cleanup 858 * in the current protocol thread is defered till the end of this 859 * function. 860 * 861 * NOTE: 862 * After cleanup the inp's hash and syncache entries, this inp will 863 * no longer be available to the rest of the protocol threads, so we 864 * are safe to whack the inp in the following code. 865 */ 866 if (inp->inp_flags & INP_WILDCARD_MP) { 867 struct netmsg_listen_detach nmsg; 868 869 KKASSERT(so->so_port == netisr_cpuport(0)); 870 KKASSERT(&curthread->td_msgport == netisr_cpuport(0)); 871 KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]); 872 873 netmsg_init(&nmsg.base, NULL, &curthread->td_msgport, 874 MSGF_PRIORITY, tcp_listen_detach_handler); 875 nmsg.nm_tp = tp; 876 nmsg.nm_tp_inh = tp_inh; 877 lwkt_domsg(netisr_cpuport(1), &nmsg.base.lmsg, 0); 878 879 inp->inp_flags &= ~INP_WILDCARD_MP; 880 } 881 882 KKASSERT(tp->t_state != TCPS_TERMINATING); 883 tp->t_state = TCPS_TERMINATING; 884 885 /* 886 * Make sure that all of our timers are stopped before we 887 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL), 888 * timers are never used. If timer message is never created 889 * (tp->tt_msg->tt_tcb == NULL), timers are never used too. 890 */ 891 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) { 892 tcp_callout_stop(tp, tp->tt_rexmt); 893 tcp_callout_stop(tp, tp->tt_persist); 894 tcp_callout_stop(tp, tp->tt_keep); 895 tcp_callout_stop(tp, tp->tt_2msl); 896 tcp_callout_stop(tp, tp->tt_delack); 897 } 898 899 if (tp->t_flags & TF_ONOUTPUTQ) { 900 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 901 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 902 tp->t_flags &= ~TF_ONOUTPUTQ; 903 } 904 905 /* 906 * If we got enough samples through the srtt filter, 907 * save the rtt and rttvar in the routing entry. 908 * 'Enough' is arbitrarily defined as the 16 samples. 909 * 16 samples is enough for the srtt filter to converge 910 * to within 5% of the correct value; fewer samples and 911 * we could save a very bogus rtt. 912 * 913 * Don't update the default route's characteristics and don't 914 * update anything that the user "locked". 915 */ 916 if (tp->t_rttupdated >= 16) { 917 u_long i = 0; 918 919 if (isipv6) { 920 struct sockaddr_in6 *sin6; 921 922 if ((rt = inp->in6p_route.ro_rt) == NULL) 923 goto no_valid_rt; 924 sin6 = (struct sockaddr_in6 *)rt_key(rt); 925 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 926 goto no_valid_rt; 927 } else 928 if ((rt = inp->inp_route.ro_rt) == NULL || 929 ((struct sockaddr_in *)rt_key(rt))-> 930 sin_addr.s_addr == INADDR_ANY) 931 goto no_valid_rt; 932 933 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 934 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 935 if (rt->rt_rmx.rmx_rtt && i) 936 /* 937 * filter this update to half the old & half 938 * the new values, converting scale. 939 * See route.h and tcp_var.h for a 940 * description of the scaling constants. 941 */ 942 rt->rt_rmx.rmx_rtt = 943 (rt->rt_rmx.rmx_rtt + i) / 2; 944 else 945 rt->rt_rmx.rmx_rtt = i; 946 tcpstat.tcps_cachedrtt++; 947 } 948 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 949 i = tp->t_rttvar * 950 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 951 if (rt->rt_rmx.rmx_rttvar && i) 952 rt->rt_rmx.rmx_rttvar = 953 (rt->rt_rmx.rmx_rttvar + i) / 2; 954 else 955 rt->rt_rmx.rmx_rttvar = i; 956 tcpstat.tcps_cachedrttvar++; 957 } 958 /* 959 * The old comment here said: 960 * update the pipelimit (ssthresh) if it has been updated 961 * already or if a pipesize was specified & the threshhold 962 * got below half the pipesize. I.e., wait for bad news 963 * before we start updating, then update on both good 964 * and bad news. 965 * 966 * But we want to save the ssthresh even if no pipesize is 967 * specified explicitly in the route, because such 968 * connections still have an implicit pipesize specified 969 * by the global tcp_sendspace. In the absence of a reliable 970 * way to calculate the pipesize, it will have to do. 971 */ 972 i = tp->snd_ssthresh; 973 if (rt->rt_rmx.rmx_sendpipe != 0) 974 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 975 else 976 dosavessthresh = (i < so->so_snd.ssb_hiwat/2); 977 if (dosavessthresh || 978 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 979 (rt->rt_rmx.rmx_ssthresh != 0))) { 980 /* 981 * convert the limit from user data bytes to 982 * packets then to packet data bytes. 983 */ 984 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 985 if (i < 2) 986 i = 2; 987 i *= tp->t_maxseg + 988 (isipv6 ? 989 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 990 sizeof(struct tcpiphdr)); 991 if (rt->rt_rmx.rmx_ssthresh) 992 rt->rt_rmx.rmx_ssthresh = 993 (rt->rt_rmx.rmx_ssthresh + i) / 2; 994 else 995 rt->rt_rmx.rmx_ssthresh = i; 996 tcpstat.tcps_cachedssthresh++; 997 } 998 } 999 1000 no_valid_rt: 1001 /* free the reassembly queue, if any */ 1002 while((q = TAILQ_FIRST(&tp->t_segq)) != NULL) { 1003 TAILQ_REMOVE(&tp->t_segq, q, tqe_q); 1004 m_freem(q->tqe_m); 1005 kfree(q, M_TSEGQ); 1006 atomic_add_int(&tcp_reass_qsize, -1); 1007 } 1008 /* throw away SACK blocks in scoreboard*/ 1009 if (TCP_DO_SACK(tp)) 1010 tcp_sack_destroy(&tp->scb); 1011 1012 inp->inp_ppcb = NULL; 1013 soisdisconnected(so); 1014 /* note: pcb detached later on */ 1015 1016 tcp_destroy_timermsg(tp); 1017 tcp_output_cancel(tp); 1018 1019 if (tp->t_flags & TF_LISTEN) { 1020 syncache_destroy(tp, tp_inh); 1021 if (inp_inh != NULL && inp_inh->inp_socket != NULL) { 1022 /* 1023 * Pending sockets inheritance only needs 1024 * to be done once in the current thread, 1025 * i.e. netisr0. 1026 */ 1027 soinherit(so, inp_inh->inp_socket); 1028 } 1029 } 1030 1031 so_async_rcvd_drop(so); 1032 /* Drop the reference for the asynchronized pru_rcvd */ 1033 sofree(so); 1034 1035 /* 1036 * NOTE: 1037 * pcbdetach removes any wildcard hash entry on the current CPU. 1038 */ 1039 #ifdef INET6 1040 if (isafinet6) 1041 in6_pcbdetach(inp); 1042 else 1043 #endif 1044 in_pcbdetach(inp); 1045 1046 tcpstat.tcps_closed++; 1047 return (NULL); 1048 } 1049 1050 static __inline void 1051 tcp_drain_oncpu(struct inpcbhead *head) 1052 { 1053 struct inpcb *marker; 1054 struct inpcb *inpb; 1055 struct tcpcb *tcpb; 1056 struct tseg_qent *te; 1057 1058 /* 1059 * Allows us to block while running the list 1060 */ 1061 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1062 marker->inp_flags |= INP_PLACEMARKER; 1063 LIST_INSERT_HEAD(head, marker, inp_list); 1064 1065 while ((inpb = LIST_NEXT(marker, inp_list)) != NULL) { 1066 if ((inpb->inp_flags & INP_PLACEMARKER) == 0 && 1067 (tcpb = intotcpcb(inpb)) != NULL && 1068 (te = TAILQ_FIRST(&tcpb->t_segq)) != NULL) { 1069 TAILQ_REMOVE(&tcpb->t_segq, te, tqe_q); 1070 if (te->tqe_th->th_flags & TH_FIN) 1071 tcpb->t_flags &= ~TF_QUEDFIN; 1072 m_freem(te->tqe_m); 1073 kfree(te, M_TSEGQ); 1074 atomic_add_int(&tcp_reass_qsize, -1); 1075 /* retry */ 1076 } else { 1077 LIST_REMOVE(marker, inp_list); 1078 LIST_INSERT_AFTER(inpb, marker, inp_list); 1079 } 1080 } 1081 LIST_REMOVE(marker, inp_list); 1082 kfree(marker, M_TEMP); 1083 } 1084 1085 struct netmsg_tcp_drain { 1086 struct netmsg_base base; 1087 struct inpcbhead *nm_head; 1088 }; 1089 1090 static void 1091 tcp_drain_handler(netmsg_t msg) 1092 { 1093 struct netmsg_tcp_drain *nm = (void *)msg; 1094 1095 tcp_drain_oncpu(nm->nm_head); 1096 lwkt_replymsg(&nm->base.lmsg, 0); 1097 } 1098 1099 void 1100 tcp_drain(void) 1101 { 1102 int cpu; 1103 1104 if (!do_tcpdrain) 1105 return; 1106 1107 /* 1108 * Walk the tcpbs, if existing, and flush the reassembly queue, 1109 * if there is one... 1110 * XXX: The "Net/3" implementation doesn't imply that the TCP 1111 * reassembly queue should be flushed, but in a situation 1112 * where we're really low on mbufs, this is potentially 1113 * useful. 1114 */ 1115 for (cpu = 0; cpu < ncpus2; cpu++) { 1116 struct netmsg_tcp_drain *nm; 1117 1118 if (cpu == mycpu->gd_cpuid) { 1119 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead); 1120 } else { 1121 nm = kmalloc(sizeof(struct netmsg_tcp_drain), 1122 M_LWKTMSG, M_NOWAIT); 1123 if (nm == NULL) 1124 continue; 1125 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1126 0, tcp_drain_handler); 1127 nm->nm_head = &tcbinfo[cpu].pcblisthead; 1128 lwkt_sendmsg(netisr_cpuport(cpu), &nm->base.lmsg); 1129 } 1130 } 1131 } 1132 1133 /* 1134 * Notify a tcp user of an asynchronous error; 1135 * store error as soft error, but wake up user 1136 * (for now, won't do anything until can select for soft error). 1137 * 1138 * Do not wake up user since there currently is no mechanism for 1139 * reporting soft errors (yet - a kqueue filter may be added). 1140 */ 1141 static void 1142 tcp_notify(struct inpcb *inp, int error) 1143 { 1144 struct tcpcb *tp = intotcpcb(inp); 1145 1146 /* 1147 * Ignore some errors if we are hooked up. 1148 * If connection hasn't completed, has retransmitted several times, 1149 * and receives a second error, give up now. This is better 1150 * than waiting a long time to establish a connection that 1151 * can never complete. 1152 */ 1153 if (tp->t_state == TCPS_ESTABLISHED && 1154 (error == EHOSTUNREACH || error == ENETUNREACH || 1155 error == EHOSTDOWN)) { 1156 return; 1157 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1158 tp->t_softerror) 1159 tcp_drop(tp, error); 1160 else 1161 tp->t_softerror = error; 1162 #if 0 1163 wakeup(&so->so_timeo); 1164 sorwakeup(so); 1165 sowwakeup(so); 1166 #endif 1167 } 1168 1169 static int 1170 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1171 { 1172 int error, i, n; 1173 struct inpcb *marker; 1174 struct inpcb *inp; 1175 globaldata_t gd; 1176 int origcpu, ccpu; 1177 1178 error = 0; 1179 n = 0; 1180 1181 /* 1182 * The process of preparing the TCB list is too time-consuming and 1183 * resource-intensive to repeat twice on every request. 1184 */ 1185 if (req->oldptr == NULL) { 1186 for (ccpu = 0; ccpu < ncpus; ++ccpu) { 1187 gd = globaldata_find(ccpu); 1188 n += tcbinfo[gd->gd_cpuid].ipi_count; 1189 } 1190 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1191 return (0); 1192 } 1193 1194 if (req->newptr != NULL) 1195 return (EPERM); 1196 1197 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1198 marker->inp_flags |= INP_PLACEMARKER; 1199 1200 /* 1201 * OK, now we're committed to doing something. Run the inpcb list 1202 * for each cpu in the system and construct the output. Use a 1203 * list placemarker to deal with list changes occuring during 1204 * copyout blockages (but otherwise depend on being on the correct 1205 * cpu to avoid races). 1206 */ 1207 origcpu = mycpu->gd_cpuid; 1208 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) { 1209 globaldata_t rgd; 1210 caddr_t inp_ppcb; 1211 struct xtcpcb xt; 1212 int cpu_id; 1213 1214 cpu_id = (origcpu + ccpu) % ncpus; 1215 if ((smp_active_mask & CPUMASK(cpu_id)) == 0) 1216 continue; 1217 rgd = globaldata_find(cpu_id); 1218 lwkt_setcpu_self(rgd); 1219 1220 n = tcbinfo[cpu_id].ipi_count; 1221 1222 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list); 1223 i = 0; 1224 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1225 /* 1226 * process a snapshot of pcbs, ignoring placemarkers 1227 * and using our own to allow SYSCTL_OUT to block. 1228 */ 1229 LIST_REMOVE(marker, inp_list); 1230 LIST_INSERT_AFTER(inp, marker, inp_list); 1231 1232 if (inp->inp_flags & INP_PLACEMARKER) 1233 continue; 1234 if (prison_xinpcb(req->td, inp)) 1235 continue; 1236 1237 xt.xt_len = sizeof xt; 1238 bcopy(inp, &xt.xt_inp, sizeof *inp); 1239 inp_ppcb = inp->inp_ppcb; 1240 if (inp_ppcb != NULL) 1241 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1242 else 1243 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1244 if (inp->inp_socket) 1245 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1246 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1247 break; 1248 ++i; 1249 } 1250 LIST_REMOVE(marker, inp_list); 1251 if (error == 0 && i < n) { 1252 bzero(&xt, sizeof xt); 1253 xt.xt_len = sizeof xt; 1254 while (i < n) { 1255 error = SYSCTL_OUT(req, &xt, sizeof xt); 1256 if (error) 1257 break; 1258 ++i; 1259 } 1260 } 1261 } 1262 1263 /* 1264 * Make sure we are on the same cpu we were on originally, since 1265 * higher level callers expect this. Also don't pollute caches with 1266 * migrated userland data by (eventually) returning to userland 1267 * on a different cpu. 1268 */ 1269 lwkt_setcpu_self(globaldata_find(origcpu)); 1270 kfree(marker, M_TEMP); 1271 return (error); 1272 } 1273 1274 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1275 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1276 1277 static int 1278 tcp_getcred(SYSCTL_HANDLER_ARGS) 1279 { 1280 struct sockaddr_in addrs[2]; 1281 struct inpcb *inp; 1282 int cpu; 1283 int error; 1284 1285 error = priv_check(req->td, PRIV_ROOT); 1286 if (error != 0) 1287 return (error); 1288 error = SYSCTL_IN(req, addrs, sizeof addrs); 1289 if (error != 0) 1290 return (error); 1291 crit_enter(); 1292 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1293 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1294 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1295 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1296 if (inp == NULL || inp->inp_socket == NULL) { 1297 error = ENOENT; 1298 goto out; 1299 } 1300 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1301 out: 1302 crit_exit(); 1303 return (error); 1304 } 1305 1306 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1307 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1308 1309 #ifdef INET6 1310 static int 1311 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1312 { 1313 struct sockaddr_in6 addrs[2]; 1314 struct inpcb *inp; 1315 int error; 1316 boolean_t mapped = FALSE; 1317 1318 error = priv_check(req->td, PRIV_ROOT); 1319 if (error != 0) 1320 return (error); 1321 error = SYSCTL_IN(req, addrs, sizeof addrs); 1322 if (error != 0) 1323 return (error); 1324 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { 1325 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) 1326 mapped = TRUE; 1327 else 1328 return (EINVAL); 1329 } 1330 crit_enter(); 1331 if (mapped) { 1332 inp = in_pcblookup_hash(&tcbinfo[0], 1333 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], 1334 addrs[1].sin6_port, 1335 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], 1336 addrs[0].sin6_port, 1337 0, NULL); 1338 } else { 1339 inp = in6_pcblookup_hash(&tcbinfo[0], 1340 &addrs[1].sin6_addr, addrs[1].sin6_port, 1341 &addrs[0].sin6_addr, addrs[0].sin6_port, 1342 0, NULL); 1343 } 1344 if (inp == NULL || inp->inp_socket == NULL) { 1345 error = ENOENT; 1346 goto out; 1347 } 1348 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1349 out: 1350 crit_exit(); 1351 return (error); 1352 } 1353 1354 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1355 0, 0, 1356 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1357 #endif 1358 1359 struct netmsg_tcp_notify { 1360 struct netmsg_base base; 1361 void (*nm_notify)(struct inpcb *, int); 1362 struct in_addr nm_faddr; 1363 int nm_arg; 1364 }; 1365 1366 static void 1367 tcp_notifyall_oncpu(netmsg_t msg) 1368 { 1369 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg; 1370 int nextcpu; 1371 1372 in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nm->nm_faddr, 1373 nm->nm_arg, nm->nm_notify); 1374 1375 nextcpu = mycpuid + 1; 1376 if (nextcpu < ncpus2) 1377 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nm->base.lmsg); 1378 else 1379 lwkt_replymsg(&nm->base.lmsg, 0); 1380 } 1381 1382 void 1383 tcp_ctlinput(netmsg_t msg) 1384 { 1385 int cmd = msg->ctlinput.nm_cmd; 1386 struct sockaddr *sa = msg->ctlinput.nm_arg; 1387 struct ip *ip = msg->ctlinput.nm_extra; 1388 struct tcphdr *th; 1389 struct in_addr faddr; 1390 struct inpcb *inp; 1391 struct tcpcb *tp; 1392 void (*notify)(struct inpcb *, int) = tcp_notify; 1393 tcp_seq icmpseq; 1394 int arg, cpu; 1395 1396 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) { 1397 goto done; 1398 } 1399 1400 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1401 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1402 goto done; 1403 1404 arg = inetctlerrmap[cmd]; 1405 if (cmd == PRC_QUENCH) { 1406 notify = tcp_quench; 1407 } else if (icmp_may_rst && 1408 (cmd == PRC_UNREACH_ADMIN_PROHIB || 1409 cmd == PRC_UNREACH_PORT || 1410 cmd == PRC_TIMXCEED_INTRANS) && 1411 ip != NULL) { 1412 notify = tcp_drop_syn_sent; 1413 } else if (cmd == PRC_MSGSIZE) { 1414 struct icmp *icmp = (struct icmp *) 1415 ((caddr_t)ip - offsetof(struct icmp, icmp_ip)); 1416 1417 arg = ntohs(icmp->icmp_nextmtu); 1418 notify = tcp_mtudisc; 1419 } else if (PRC_IS_REDIRECT(cmd)) { 1420 ip = NULL; 1421 notify = in_rtchange; 1422 } else if (cmd == PRC_HOSTDEAD) { 1423 ip = NULL; 1424 } 1425 1426 if (ip != NULL) { 1427 crit_enter(); 1428 th = (struct tcphdr *)((caddr_t)ip + 1429 (IP_VHL_HL(ip->ip_vhl) << 2)); 1430 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport, 1431 ip->ip_src.s_addr, th->th_sport); 1432 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport, 1433 ip->ip_src, th->th_sport, 0, NULL); 1434 if ((inp != NULL) && (inp->inp_socket != NULL)) { 1435 icmpseq = htonl(th->th_seq); 1436 tp = intotcpcb(inp); 1437 if (SEQ_GEQ(icmpseq, tp->snd_una) && 1438 SEQ_LT(icmpseq, tp->snd_max)) 1439 (*notify)(inp, arg); 1440 } else { 1441 struct in_conninfo inc; 1442 1443 inc.inc_fport = th->th_dport; 1444 inc.inc_lport = th->th_sport; 1445 inc.inc_faddr = faddr; 1446 inc.inc_laddr = ip->ip_src; 1447 #ifdef INET6 1448 inc.inc_isipv6 = 0; 1449 #endif 1450 syncache_unreach(&inc, th); 1451 } 1452 crit_exit(); 1453 } else { 1454 struct netmsg_tcp_notify *nm; 1455 1456 KKASSERT(&curthread->td_msgport == netisr_cpuport(0)); 1457 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT); 1458 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1459 0, tcp_notifyall_oncpu); 1460 nm->nm_faddr = faddr; 1461 nm->nm_arg = arg; 1462 nm->nm_notify = notify; 1463 1464 lwkt_sendmsg(netisr_cpuport(0), &nm->base.lmsg); 1465 } 1466 done: 1467 lwkt_replymsg(&msg->lmsg, 0); 1468 } 1469 1470 #ifdef INET6 1471 1472 void 1473 tcp6_ctlinput(netmsg_t msg) 1474 { 1475 int cmd = msg->ctlinput.nm_cmd; 1476 struct sockaddr *sa = msg->ctlinput.nm_arg; 1477 void *d = msg->ctlinput.nm_extra; 1478 struct tcphdr th; 1479 void (*notify) (struct inpcb *, int) = tcp_notify; 1480 struct ip6_hdr *ip6; 1481 struct mbuf *m; 1482 struct ip6ctlparam *ip6cp = NULL; 1483 const struct sockaddr_in6 *sa6_src = NULL; 1484 int off; 1485 struct tcp_portonly { 1486 u_int16_t th_sport; 1487 u_int16_t th_dport; 1488 } *thp; 1489 int arg; 1490 1491 if (sa->sa_family != AF_INET6 || 1492 sa->sa_len != sizeof(struct sockaddr_in6)) { 1493 goto out; 1494 } 1495 1496 arg = 0; 1497 if (cmd == PRC_QUENCH) 1498 notify = tcp_quench; 1499 else if (cmd == PRC_MSGSIZE) { 1500 struct ip6ctlparam *ip6cp = d; 1501 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6; 1502 1503 arg = ntohl(icmp6->icmp6_mtu); 1504 notify = tcp_mtudisc; 1505 } else if (!PRC_IS_REDIRECT(cmd) && 1506 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) { 1507 goto out; 1508 } 1509 1510 /* if the parameter is from icmp6, decode it. */ 1511 if (d != NULL) { 1512 ip6cp = (struct ip6ctlparam *)d; 1513 m = ip6cp->ip6c_m; 1514 ip6 = ip6cp->ip6c_ip6; 1515 off = ip6cp->ip6c_off; 1516 sa6_src = ip6cp->ip6c_src; 1517 } else { 1518 m = NULL; 1519 ip6 = NULL; 1520 off = 0; /* fool gcc */ 1521 sa6_src = &sa6_any; 1522 } 1523 1524 if (ip6 != NULL) { 1525 struct in_conninfo inc; 1526 /* 1527 * XXX: We assume that when IPV6 is non NULL, 1528 * M and OFF are valid. 1529 */ 1530 1531 /* check if we can safely examine src and dst ports */ 1532 if (m->m_pkthdr.len < off + sizeof *thp) 1533 goto out; 1534 1535 bzero(&th, sizeof th); 1536 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1537 1538 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport, 1539 (struct sockaddr *)ip6cp->ip6c_src, 1540 th.th_sport, cmd, arg, notify); 1541 1542 inc.inc_fport = th.th_dport; 1543 inc.inc_lport = th.th_sport; 1544 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1545 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1546 inc.inc_isipv6 = 1; 1547 syncache_unreach(&inc, &th); 1548 } else { 1549 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0, 1550 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify); 1551 } 1552 out: 1553 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0); 1554 } 1555 1556 #endif 1557 1558 /* 1559 * Following is where TCP initial sequence number generation occurs. 1560 * 1561 * There are two places where we must use initial sequence numbers: 1562 * 1. In SYN-ACK packets. 1563 * 2. In SYN packets. 1564 * 1565 * All ISNs for SYN-ACK packets are generated by the syncache. See 1566 * tcp_syncache.c for details. 1567 * 1568 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1569 * depends on this property. In addition, these ISNs should be 1570 * unguessable so as to prevent connection hijacking. To satisfy 1571 * the requirements of this situation, the algorithm outlined in 1572 * RFC 1948 is used to generate sequence numbers. 1573 * 1574 * Implementation details: 1575 * 1576 * Time is based off the system timer, and is corrected so that it 1577 * increases by one megabyte per second. This allows for proper 1578 * recycling on high speed LANs while still leaving over an hour 1579 * before rollover. 1580 * 1581 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1582 * between seeding of isn_secret. This is normally set to zero, 1583 * as reseeding should not be necessary. 1584 * 1585 */ 1586 1587 #define ISN_BYTES_PER_SECOND 1048576 1588 1589 u_char isn_secret[32]; 1590 int isn_last_reseed; 1591 MD5_CTX isn_ctx; 1592 1593 tcp_seq 1594 tcp_new_isn(struct tcpcb *tp) 1595 { 1596 u_int32_t md5_buffer[4]; 1597 tcp_seq new_isn; 1598 1599 /* Seed if this is the first use, reseed if requested. */ 1600 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1601 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1602 < (u_int)ticks))) { 1603 read_random_unlimited(&isn_secret, sizeof isn_secret); 1604 isn_last_reseed = ticks; 1605 } 1606 1607 /* Compute the md5 hash and return the ISN. */ 1608 MD5Init(&isn_ctx); 1609 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1610 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1611 #ifdef INET6 1612 if (tp->t_inpcb->inp_vflag & INP_IPV6) { 1613 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1614 sizeof(struct in6_addr)); 1615 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1616 sizeof(struct in6_addr)); 1617 } else 1618 #endif 1619 { 1620 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1621 sizeof(struct in_addr)); 1622 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1623 sizeof(struct in_addr)); 1624 } 1625 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1626 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1627 new_isn = (tcp_seq) md5_buffer[0]; 1628 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1629 return (new_isn); 1630 } 1631 1632 /* 1633 * When a source quench is received, close congestion window 1634 * to one segment. We will gradually open it again as we proceed. 1635 */ 1636 void 1637 tcp_quench(struct inpcb *inp, int error) 1638 { 1639 struct tcpcb *tp = intotcpcb(inp); 1640 1641 if (tp != NULL) { 1642 tp->snd_cwnd = tp->t_maxseg; 1643 tp->snd_wacked = 0; 1644 } 1645 } 1646 1647 /* 1648 * When a specific ICMP unreachable message is received and the 1649 * connection state is SYN-SENT, drop the connection. This behavior 1650 * is controlled by the icmp_may_rst sysctl. 1651 */ 1652 void 1653 tcp_drop_syn_sent(struct inpcb *inp, int error) 1654 { 1655 struct tcpcb *tp = intotcpcb(inp); 1656 1657 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT)) 1658 tcp_drop(tp, error); 1659 } 1660 1661 /* 1662 * When a `need fragmentation' ICMP is received, update our idea of the MSS 1663 * based on the new value in the route. Also nudge TCP to send something, 1664 * since we know the packet we just sent was dropped. 1665 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1666 */ 1667 void 1668 tcp_mtudisc(struct inpcb *inp, int mtu) 1669 { 1670 struct tcpcb *tp = intotcpcb(inp); 1671 struct rtentry *rt; 1672 struct socket *so = inp->inp_socket; 1673 int maxopd, mss; 1674 #ifdef INET6 1675 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0); 1676 #else 1677 const boolean_t isipv6 = FALSE; 1678 #endif 1679 1680 if (tp == NULL) 1681 return; 1682 1683 /* 1684 * If no MTU is provided in the ICMP message, use the 1685 * next lower likely value, as specified in RFC 1191. 1686 */ 1687 if (mtu == 0) { 1688 int oldmtu; 1689 1690 oldmtu = tp->t_maxopd + 1691 (isipv6 ? 1692 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1693 sizeof(struct tcpiphdr)); 1694 mtu = ip_next_mtu(oldmtu, 0); 1695 } 1696 1697 if (isipv6) 1698 rt = tcp_rtlookup6(&inp->inp_inc); 1699 else 1700 rt = tcp_rtlookup(&inp->inp_inc); 1701 if (rt != NULL) { 1702 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu) 1703 mtu = rt->rt_rmx.rmx_mtu; 1704 1705 maxopd = mtu - 1706 (isipv6 ? 1707 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1708 sizeof(struct tcpiphdr)); 1709 1710 /* 1711 * XXX - The following conditional probably violates the TCP 1712 * spec. The problem is that, since we don't know the 1713 * other end's MSS, we are supposed to use a conservative 1714 * default. But, if we do that, then MTU discovery will 1715 * never actually take place, because the conservative 1716 * default is much less than the MTUs typically seen 1717 * on the Internet today. For the moment, we'll sweep 1718 * this under the carpet. 1719 * 1720 * The conservative default might not actually be a problem 1721 * if the only case this occurs is when sending an initial 1722 * SYN with options and data to a host we've never talked 1723 * to before. Then, they will reply with an MSS value which 1724 * will get recorded and the new parameters should get 1725 * recomputed. For Further Study. 1726 */ 1727 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd) 1728 maxopd = rt->rt_rmx.rmx_mssopt; 1729 } else 1730 maxopd = mtu - 1731 (isipv6 ? 1732 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1733 sizeof(struct tcpiphdr)); 1734 1735 if (tp->t_maxopd <= maxopd) 1736 return; 1737 tp->t_maxopd = maxopd; 1738 1739 mss = maxopd; 1740 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) == 1741 (TF_REQ_TSTMP | TF_RCVD_TSTMP)) 1742 mss -= TCPOLEN_TSTAMP_APPA; 1743 1744 /* round down to multiple of MCLBYTES */ 1745 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */ 1746 if (mss > MCLBYTES) 1747 mss &= ~(MCLBYTES - 1); 1748 #else 1749 if (mss > MCLBYTES) 1750 mss = (mss / MCLBYTES) * MCLBYTES; 1751 #endif 1752 1753 if (so->so_snd.ssb_hiwat < mss) 1754 mss = so->so_snd.ssb_hiwat; 1755 1756 tp->t_maxseg = mss; 1757 tp->t_rtttime = 0; 1758 tp->snd_nxt = tp->snd_una; 1759 tcp_output(tp); 1760 tcpstat.tcps_mturesent++; 1761 } 1762 1763 /* 1764 * Look-up the routing entry to the peer of this inpcb. If no route 1765 * is found and it cannot be allocated the return NULL. This routine 1766 * is called by TCP routines that access the rmx structure and by tcp_mss 1767 * to get the interface MTU. 1768 */ 1769 struct rtentry * 1770 tcp_rtlookup(struct in_conninfo *inc) 1771 { 1772 struct route *ro = &inc->inc_route; 1773 1774 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1775 /* No route yet, so try to acquire one */ 1776 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1777 /* 1778 * unused portions of the structure MUST be zero'd 1779 * out because rtalloc() treats it as opaque data 1780 */ 1781 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1782 ro->ro_dst.sa_family = AF_INET; 1783 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1784 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1785 inc->inc_faddr; 1786 rtalloc(ro); 1787 } 1788 } 1789 return (ro->ro_rt); 1790 } 1791 1792 #ifdef INET6 1793 struct rtentry * 1794 tcp_rtlookup6(struct in_conninfo *inc) 1795 { 1796 struct route_in6 *ro6 = &inc->inc6_route; 1797 1798 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1799 /* No route yet, so try to acquire one */ 1800 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1801 /* 1802 * unused portions of the structure MUST be zero'd 1803 * out because rtalloc() treats it as opaque data 1804 */ 1805 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1806 ro6->ro_dst.sin6_family = AF_INET6; 1807 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1808 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1809 rtalloc((struct route *)ro6); 1810 } 1811 } 1812 return (ro6->ro_rt); 1813 } 1814 #endif 1815 1816 #ifdef IPSEC 1817 /* compute ESP/AH header size for TCP, including outer IP header. */ 1818 size_t 1819 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1820 { 1821 struct inpcb *inp; 1822 struct mbuf *m; 1823 size_t hdrsiz; 1824 struct ip *ip; 1825 struct tcphdr *th; 1826 1827 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1828 return (0); 1829 MGETHDR(m, MB_DONTWAIT, MT_DATA); 1830 if (!m) 1831 return (0); 1832 1833 #ifdef INET6 1834 if (inp->inp_vflag & INP_IPV6) { 1835 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1836 1837 th = (struct tcphdr *)(ip6 + 1); 1838 m->m_pkthdr.len = m->m_len = 1839 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1840 tcp_fillheaders(tp, ip6, th, FALSE); 1841 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1842 } else 1843 #endif 1844 { 1845 ip = mtod(m, struct ip *); 1846 th = (struct tcphdr *)(ip + 1); 1847 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1848 tcp_fillheaders(tp, ip, th, FALSE); 1849 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1850 } 1851 1852 m_free(m); 1853 return (hdrsiz); 1854 } 1855 #endif 1856 1857 /* 1858 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1859 * 1860 * This code attempts to calculate the bandwidth-delay product as a 1861 * means of determining the optimal window size to maximize bandwidth, 1862 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1863 * routers. This code also does a fairly good job keeping RTTs in check 1864 * across slow links like modems. We implement an algorithm which is very 1865 * similar (but not meant to be) TCP/Vegas. The code operates on the 1866 * transmitter side of a TCP connection and so only effects the transmit 1867 * side of the connection. 1868 * 1869 * BACKGROUND: TCP makes no provision for the management of buffer space 1870 * at the end points or at the intermediate routers and switches. A TCP 1871 * stream, whether using NewReno or not, will eventually buffer as 1872 * many packets as it is able and the only reason this typically works is 1873 * due to the fairly small default buffers made available for a connection 1874 * (typicaly 16K or 32K). As machines use larger windows and/or window 1875 * scaling it is now fairly easy for even a single TCP connection to blow-out 1876 * all available buffer space not only on the local interface, but on 1877 * intermediate routers and switches as well. NewReno makes a misguided 1878 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1879 * then backing off, then steadily increasing the window again until another 1880 * failure occurs, ad-infinitum. This results in terrible oscillation that 1881 * is only made worse as network loads increase and the idea of intentionally 1882 * blowing out network buffers is, frankly, a terrible way to manage network 1883 * resources. 1884 * 1885 * It is far better to limit the transmit window prior to the failure 1886 * condition being achieved. There are two general ways to do this: First 1887 * you can 'scan' through different transmit window sizes and locate the 1888 * point where the RTT stops increasing, indicating that you have filled the 1889 * pipe, then scan backwards until you note that RTT stops decreasing, then 1890 * repeat ad-infinitum. This method works in principle but has severe 1891 * implementation issues due to RTT variances, timer granularity, and 1892 * instability in the algorithm which can lead to many false positives and 1893 * create oscillations as well as interact badly with other TCP streams 1894 * implementing the same algorithm. 1895 * 1896 * The second method is to limit the window to the bandwidth delay product 1897 * of the link. This is the method we implement. RTT variances and our 1898 * own manipulation of the congestion window, bwnd, can potentially 1899 * destabilize the algorithm. For this reason we have to stabilize the 1900 * elements used to calculate the window. We do this by using the minimum 1901 * observed RTT, the long term average of the observed bandwidth, and 1902 * by adding two segments worth of slop. It isn't perfect but it is able 1903 * to react to changing conditions and gives us a very stable basis on 1904 * which to extend the algorithm. 1905 */ 1906 void 1907 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1908 { 1909 u_long bw; 1910 u_long bwnd; 1911 int save_ticks; 1912 int delta_ticks; 1913 1914 /* 1915 * If inflight_enable is disabled in the middle of a tcp connection, 1916 * make sure snd_bwnd is effectively disabled. 1917 */ 1918 if (!tcp_inflight_enable) { 1919 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1920 tp->snd_bandwidth = 0; 1921 return; 1922 } 1923 1924 /* 1925 * Validate the delta time. If a connection is new or has been idle 1926 * a long time we have to reset the bandwidth calculator. 1927 */ 1928 save_ticks = ticks; 1929 delta_ticks = save_ticks - tp->t_bw_rtttime; 1930 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1931 tp->t_bw_rtttime = ticks; 1932 tp->t_bw_rtseq = ack_seq; 1933 if (tp->snd_bandwidth == 0) 1934 tp->snd_bandwidth = tcp_inflight_min; 1935 return; 1936 } 1937 if (delta_ticks == 0) 1938 return; 1939 1940 /* 1941 * Sanity check, plus ignore pure window update acks. 1942 */ 1943 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1944 return; 1945 1946 /* 1947 * Figure out the bandwidth. Due to the tick granularity this 1948 * is a very rough number and it MUST be averaged over a fairly 1949 * long period of time. XXX we need to take into account a link 1950 * that is not using all available bandwidth, but for now our 1951 * slop will ramp us up if this case occurs and the bandwidth later 1952 * increases. 1953 */ 1954 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 1955 tp->t_bw_rtttime = save_ticks; 1956 tp->t_bw_rtseq = ack_seq; 1957 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4; 1958 1959 tp->snd_bandwidth = bw; 1960 1961 /* 1962 * Calculate the semi-static bandwidth delay product, plus two maximal 1963 * segments. The additional slop puts us squarely in the sweet 1964 * spot and also handles the bandwidth run-up case. Without the 1965 * slop we could be locking ourselves into a lower bandwidth. 1966 * 1967 * Situations Handled: 1968 * (1) Prevents over-queueing of packets on LANs, especially on 1969 * high speed LANs, allowing larger TCP buffers to be 1970 * specified, and also does a good job preventing 1971 * over-queueing of packets over choke points like modems 1972 * (at least for the transmit side). 1973 * 1974 * (2) Is able to handle changing network loads (bandwidth 1975 * drops so bwnd drops, bandwidth increases so bwnd 1976 * increases). 1977 * 1978 * (3) Theoretically should stabilize in the face of multiple 1979 * connections implementing the same algorithm (this may need 1980 * a little work). 1981 * 1982 * (4) Stability value (defaults to 20 = 2 maximal packets) can 1983 * be adjusted with a sysctl but typically only needs to be on 1984 * very slow connections. A value no smaller then 5 should 1985 * be used, but only reduce this default if you have no other 1986 * choice. 1987 */ 1988 1989 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2) 1990 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 1991 tcp_inflight_stab * (int)tp->t_maxseg / 10; 1992 #undef USERTT 1993 1994 if (tcp_inflight_debug > 0) { 1995 static int ltime; 1996 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) { 1997 ltime = ticks; 1998 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n", 1999 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd); 2000 } 2001 } 2002 if ((long)bwnd < tcp_inflight_min) 2003 bwnd = tcp_inflight_min; 2004 if (bwnd > tcp_inflight_max) 2005 bwnd = tcp_inflight_max; 2006 if ((long)bwnd < tp->t_maxseg * 2) 2007 bwnd = tp->t_maxseg * 2; 2008 tp->snd_bwnd = bwnd; 2009 } 2010 2011 static void 2012 tcp_rmx_iwsegs(struct tcpcb *tp, u_long *maxsegs, u_long *capsegs) 2013 { 2014 struct rtentry *rt; 2015 struct inpcb *inp = tp->t_inpcb; 2016 #ifdef INET6 2017 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE); 2018 #else 2019 const boolean_t isipv6 = FALSE; 2020 #endif 2021 2022 /* XXX */ 2023 if (tcp_iw_maxsegs < TCP_IW_MAXSEGS_DFLT) 2024 tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 2025 if (tcp_iw_capsegs < TCP_IW_CAPSEGS_DFLT) 2026 tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 2027 2028 if (isipv6) 2029 rt = tcp_rtlookup6(&inp->inp_inc); 2030 else 2031 rt = tcp_rtlookup(&inp->inp_inc); 2032 if (rt == NULL || 2033 rt->rt_rmx.rmx_iwmaxsegs < TCP_IW_MAXSEGS_DFLT || 2034 rt->rt_rmx.rmx_iwcapsegs < TCP_IW_CAPSEGS_DFLT) { 2035 *maxsegs = tcp_iw_maxsegs; 2036 *capsegs = tcp_iw_capsegs; 2037 return; 2038 } 2039 *maxsegs = rt->rt_rmx.rmx_iwmaxsegs; 2040 *capsegs = rt->rt_rmx.rmx_iwcapsegs; 2041 } 2042 2043 u_long 2044 tcp_initial_window(struct tcpcb *tp) 2045 { 2046 if (tcp_do_rfc3390) { 2047 /* 2048 * RFC3390: 2049 * "If the SYN or SYN/ACK is lost, the initial window 2050 * used by a sender after a correctly transmitted SYN 2051 * MUST be one segment consisting of MSS bytes." 2052 * 2053 * However, we do something a little bit more aggressive 2054 * then RFC3390 here: 2055 * - Only if time spent in the SYN or SYN|ACK retransmition 2056 * >= 3 seconds, the IW is reduced. We do this mainly 2057 * because when RFC3390 is published, the initial RTO is 2058 * still 3 seconds (the threshold we test here), while 2059 * after RFC6298, the initial RTO is 1 second. This 2060 * behaviour probably still falls within the spirit of 2061 * RFC3390. 2062 * - When IW is reduced, 2*MSS is used instead of 1*MSS. 2063 * Mainly to avoid sender and receiver deadlock until 2064 * delayed ACK timer expires. And even RFC2581 does not 2065 * try to reduce IW upon SYN or SYN|ACK retransmition 2066 * timeout. 2067 * 2068 * See also: 2069 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03 2070 */ 2071 if (tp->t_rxtsyn >= TCPTV_RTOBASE3) { 2072 return (2 * tp->t_maxseg); 2073 } else { 2074 u_long maxsegs, capsegs; 2075 2076 tcp_rmx_iwsegs(tp, &maxsegs, &capsegs); 2077 return min(maxsegs * tp->t_maxseg, 2078 max(2 * tp->t_maxseg, capsegs * 1460)); 2079 } 2080 } else { 2081 /* 2082 * Even RFC2581 (back to 1999) allows 2*SMSS IW. 2083 * 2084 * Mainly to avoid sender and receiver deadlock 2085 * until delayed ACK timer expires. 2086 */ 2087 return (2 * tp->t_maxseg); 2088 } 2089 } 2090 2091 #ifdef TCP_SIGNATURE 2092 /* 2093 * Compute TCP-MD5 hash of a TCP segment. (RFC2385) 2094 * 2095 * We do this over ip, tcphdr, segment data, and the key in the SADB. 2096 * When called from tcp_input(), we can be sure that th_sum has been 2097 * zeroed out and verified already. 2098 * 2099 * Return 0 if successful, otherwise return -1. 2100 * 2101 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a 2102 * search with the destination IP address, and a 'magic SPI' to be 2103 * determined by the application. This is hardcoded elsewhere to 1179 2104 * right now. Another branch of this code exists which uses the SPD to 2105 * specify per-application flows but it is unstable. 2106 */ 2107 int 2108 tcpsignature_compute( 2109 struct mbuf *m, /* mbuf chain */ 2110 int len, /* length of TCP data */ 2111 int optlen, /* length of TCP options */ 2112 u_char *buf, /* storage for MD5 digest */ 2113 u_int direction) /* direction of flow */ 2114 { 2115 struct ippseudo ippseudo; 2116 MD5_CTX ctx; 2117 int doff; 2118 struct ip *ip; 2119 struct ipovly *ipovly; 2120 struct secasvar *sav; 2121 struct tcphdr *th; 2122 #ifdef INET6 2123 struct ip6_hdr *ip6; 2124 struct in6_addr in6; 2125 uint32_t plen; 2126 uint16_t nhdr; 2127 #endif /* INET6 */ 2128 u_short savecsum; 2129 2130 KASSERT(m != NULL, ("passed NULL mbuf. Game over.")); 2131 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature")); 2132 /* 2133 * Extract the destination from the IP header in the mbuf. 2134 */ 2135 ip = mtod(m, struct ip *); 2136 #ifdef INET6 2137 ip6 = NULL; /* Make the compiler happy. */ 2138 #endif /* INET6 */ 2139 /* 2140 * Look up an SADB entry which matches the address found in 2141 * the segment. 2142 */ 2143 switch (IP_VHL_V(ip->ip_vhl)) { 2144 case IPVERSION: 2145 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst, 2146 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2147 break; 2148 #ifdef INET6 2149 case (IPV6_VERSION >> 4): 2150 ip6 = mtod(m, struct ip6_hdr *); 2151 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst, 2152 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2153 break; 2154 #endif /* INET6 */ 2155 default: 2156 return (EINVAL); 2157 /* NOTREACHED */ 2158 break; 2159 } 2160 if (sav == NULL) { 2161 kprintf("%s: SADB lookup failed\n", __func__); 2162 return (EINVAL); 2163 } 2164 MD5Init(&ctx); 2165 2166 /* 2167 * Step 1: Update MD5 hash with IP pseudo-header. 2168 * 2169 * XXX The ippseudo header MUST be digested in network byte order, 2170 * or else we'll fail the regression test. Assume all fields we've 2171 * been doing arithmetic on have been in host byte order. 2172 * XXX One cannot depend on ipovly->ih_len here. When called from 2173 * tcp_output(), the underlying ip_len member has not yet been set. 2174 */ 2175 switch (IP_VHL_V(ip->ip_vhl)) { 2176 case IPVERSION: 2177 ipovly = (struct ipovly *)ip; 2178 ippseudo.ippseudo_src = ipovly->ih_src; 2179 ippseudo.ippseudo_dst = ipovly->ih_dst; 2180 ippseudo.ippseudo_pad = 0; 2181 ippseudo.ippseudo_p = IPPROTO_TCP; 2182 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen); 2183 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo)); 2184 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip)); 2185 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen; 2186 break; 2187 #ifdef INET6 2188 /* 2189 * RFC 2385, 2.0 Proposal 2190 * For IPv6, the pseudo-header is as described in RFC 2460, namely the 2191 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero- 2192 * extended next header value (to form 32 bits), and 32-bit segment 2193 * length. 2194 * Note: Upper-Layer Packet Length comes before Next Header. 2195 */ 2196 case (IPV6_VERSION >> 4): 2197 in6 = ip6->ip6_src; 2198 in6_clearscope(&in6); 2199 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2200 in6 = ip6->ip6_dst; 2201 in6_clearscope(&in6); 2202 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2203 plen = htonl(len + sizeof(struct tcphdr) + optlen); 2204 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t)); 2205 nhdr = 0; 2206 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2207 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2208 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2209 nhdr = IPPROTO_TCP; 2210 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2211 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr)); 2212 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen; 2213 break; 2214 #endif /* INET6 */ 2215 default: 2216 return (EINVAL); 2217 /* NOTREACHED */ 2218 break; 2219 } 2220 /* 2221 * Step 2: Update MD5 hash with TCP header, excluding options. 2222 * The TCP checksum must be set to zero. 2223 */ 2224 savecsum = th->th_sum; 2225 th->th_sum = 0; 2226 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr)); 2227 th->th_sum = savecsum; 2228 /* 2229 * Step 3: Update MD5 hash with TCP segment data. 2230 * Use m_apply() to avoid an early m_pullup(). 2231 */ 2232 if (len > 0) 2233 m_apply(m, doff, len, tcpsignature_apply, &ctx); 2234 /* 2235 * Step 4: Update MD5 hash with shared secret. 2236 */ 2237 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth)); 2238 MD5Final(buf, &ctx); 2239 key_sa_recordxfer(sav, m); 2240 key_freesav(sav); 2241 return (0); 2242 } 2243 2244 int 2245 tcpsignature_apply(void *fstate, void *data, unsigned int len) 2246 { 2247 2248 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len); 2249 return (0); 2250 } 2251 #endif /* TCP_SIGNATURE */ 2252