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 int tcp_mssdflt = TCP_MSS; 165 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 166 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size"); 167 168 #ifdef INET6 169 int tcp_v6mssdflt = TCP6_MSS; 170 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW, 171 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6"); 172 #endif 173 174 /* 175 * Minimum MSS we accept and use. This prevents DoS attacks where 176 * we are forced to a ridiculous low MSS like 20 and send hundreds 177 * of packets instead of one. The effect scales with the available 178 * bandwidth and quickly saturates the CPU and network interface 179 * with packet generation and sending. Set to zero to disable MINMSS 180 * checking. This setting prevents us from sending too small packets. 181 */ 182 int tcp_minmss = TCP_MINMSS; 183 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW, 184 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size"); 185 186 #if 0 187 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ; 188 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW, 189 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time"); 190 #endif 191 192 int tcp_do_rfc1323 = 1; 193 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW, 194 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions"); 195 196 static int tcp_tcbhashsize = 0; 197 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD, 198 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable"); 199 200 static int do_tcpdrain = 1; 201 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0, 202 "Enable tcp_drain routine for extra help when low on mbufs"); 203 204 static int icmp_may_rst = 1; 205 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0, 206 "Certain ICMP unreachable messages may abort connections in SYN_SENT"); 207 208 static int tcp_isn_reseed_interval = 0; 209 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW, 210 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret"); 211 212 /* 213 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on 214 * by default, but with generous values which should allow maximal 215 * bandwidth. In particular, the slop defaults to 50 (5 packets). 216 * 217 * The reason for doing this is that the limiter is the only mechanism we 218 * have which seems to do a really good job preventing receiver RX rings 219 * on network interfaces from getting blown out. Even though GigE/10GigE 220 * is supposed to flow control it looks like either it doesn't actually 221 * do it or Open Source drivers do not properly enable it. 222 * 223 * People using the limiter to reduce bottlenecks on slower WAN connections 224 * should set the slop to 20 (2 packets). 225 */ 226 static int tcp_inflight_enable = 1; 227 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW, 228 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting"); 229 230 static int tcp_inflight_debug = 0; 231 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW, 232 &tcp_inflight_debug, 0, "Debug TCP inflight calculations"); 233 234 static int tcp_inflight_min = 6144; 235 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW, 236 &tcp_inflight_min, 0, "Lower bound for TCP inflight window"); 237 238 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT; 239 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW, 240 &tcp_inflight_max, 0, "Upper bound for TCP inflight window"); 241 242 static int tcp_inflight_stab = 50; 243 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW, 244 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 3 packets)"); 245 246 static int tcp_do_rfc3390 = 1; 247 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_RW, 248 &tcp_do_rfc3390, 0, 249 "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)"); 250 251 static u_long tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 252 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwmaxsegs, CTLFLAG_RW, 253 &tcp_iw_maxsegs, 0, "TCP IW segments max"); 254 255 static u_long tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 256 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwcapsegs, CTLFLAG_RW, 257 &tcp_iw_capsegs, 0, "TCP IW segments"); 258 259 int tcp_low_rtobase = 1; 260 SYSCTL_INT(_net_inet_tcp, OID_AUTO, low_rtobase, CTLFLAG_RW, 261 &tcp_low_rtobase, 0, "Lowering the Initial RTO (RFC 6298)"); 262 263 static int tcp_do_ncr = 1; 264 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr, CTLFLAG_RW, 265 &tcp_do_ncr, 0, "Non-Congestion Robustness (RFC 4653)"); 266 267 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives"); 268 static struct malloc_pipe tcptemp_mpipe; 269 270 static void tcp_willblock(void); 271 static void tcp_notify (struct inpcb *, int); 272 273 struct tcp_stats tcpstats_percpu[MAXCPU] __cachealign; 274 275 static int 276 sysctl_tcpstats(SYSCTL_HANDLER_ARGS) 277 { 278 int cpu, error = 0; 279 280 for (cpu = 0; cpu < ncpus; ++cpu) { 281 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu], 282 sizeof(struct tcp_stats)))) 283 break; 284 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu], 285 sizeof(struct tcp_stats)))) 286 break; 287 } 288 289 return (error); 290 } 291 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW), 292 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics"); 293 294 /* 295 * Target size of TCP PCB hash tables. Must be a power of two. 296 * 297 * Note that this can be overridden by the kernel environment 298 * variable net.inet.tcp.tcbhashsize 299 */ 300 #ifndef TCBHASHSIZE 301 #define TCBHASHSIZE 512 302 #endif 303 304 /* 305 * This is the actual shape of what we allocate using the zone 306 * allocator. Doing it this way allows us to protect both structures 307 * using the same generation count, and also eliminates the overhead 308 * of allocating tcpcbs separately. By hiding the structure here, 309 * we avoid changing most of the rest of the code (although it needs 310 * to be changed, eventually, for greater efficiency). 311 */ 312 #define ALIGNMENT 32 313 #define ALIGNM1 (ALIGNMENT - 1) 314 struct inp_tp { 315 union { 316 struct inpcb inp; 317 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1]; 318 } inp_tp_u; 319 struct tcpcb tcb; 320 struct tcp_callout inp_tp_rexmt; 321 struct tcp_callout inp_tp_persist; 322 struct tcp_callout inp_tp_keep; 323 struct tcp_callout inp_tp_2msl; 324 struct tcp_callout inp_tp_delack; 325 struct netmsg_tcp_timer inp_tp_timermsg; 326 struct netmsg_base inp_tp_sndmore; 327 }; 328 #undef ALIGNMENT 329 #undef ALIGNM1 330 331 /* 332 * Tcp initialization 333 */ 334 void 335 tcp_init(void) 336 { 337 struct inpcbportinfo *portinfo; 338 struct inpcbinfo *ticb; 339 int hashsize = TCBHASHSIZE; 340 int cpu; 341 342 /* 343 * note: tcptemp is used for keepalives, and it is ok for an 344 * allocation to fail so do not specify MPF_INT. 345 */ 346 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp), 347 25, -1, 0, NULL, NULL, NULL); 348 349 tcp_delacktime = TCPTV_DELACK; 350 tcp_keepinit = TCPTV_KEEP_INIT; 351 tcp_keepidle = TCPTV_KEEP_IDLE; 352 tcp_keepintvl = TCPTV_KEEPINTVL; 353 tcp_maxpersistidle = TCPTV_KEEP_IDLE; 354 tcp_msl = TCPTV_MSL; 355 tcp_rexmit_min = TCPTV_MIN; 356 tcp_rexmit_slop = TCPTV_CPU_VAR; 357 358 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize); 359 if (!powerof2(hashsize)) { 360 kprintf("WARNING: TCB hash size not a power of 2\n"); 361 hashsize = 512; /* safe default */ 362 } 363 tcp_tcbhashsize = hashsize; 364 365 portinfo = kmalloc_cachealign(sizeof(*portinfo) * ncpus2, M_PCB, 366 M_WAITOK); 367 368 for (cpu = 0; cpu < ncpus2; cpu++) { 369 ticb = &tcbinfo[cpu]; 370 in_pcbinfo_init(ticb); 371 ticb->cpu = cpu; 372 ticb->hashbase = hashinit(hashsize, M_PCB, 373 &ticb->hashmask); 374 in_pcbportinfo_init(&portinfo[cpu], hashsize, TRUE, cpu); 375 ticb->portinfo = portinfo; 376 ticb->portinfo_mask = ncpus2_mask; 377 ticb->wildcardhashbase = hashinit(hashsize, M_PCB, 378 &ticb->wildcardhashmask); 379 ticb->localgrphashbase = hashinit(hashsize, M_PCB, 380 &ticb->localgrphashmask); 381 ticb->ipi_size = sizeof(struct inp_tp); 382 TAILQ_INIT(&tcpcbackq[cpu]); 383 } 384 385 tcp_reass_maxseg = nmbclusters / 16; 386 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg); 387 388 #ifdef INET6 389 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) 390 #else 391 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) 392 #endif 393 if (max_protohdr < TCP_MINPROTOHDR) 394 max_protohdr = TCP_MINPROTOHDR; 395 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) 396 panic("tcp_init"); 397 #undef TCP_MINPROTOHDR 398 399 /* 400 * Initialize TCP statistics counters for each CPU. 401 */ 402 for (cpu = 0; cpu < ncpus; ++cpu) { 403 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats)); 404 } 405 406 syncache_init(); 407 netisr_register_rollup(tcp_willblock, NETISR_ROLLUP_PRIO_TCP); 408 } 409 410 static void 411 tcp_willblock(void) 412 { 413 struct tcpcb *tp; 414 int cpu = mycpu->gd_cpuid; 415 416 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) { 417 KKASSERT(tp->t_flags & TF_ONOUTPUTQ); 418 tp->t_flags &= ~TF_ONOUTPUTQ; 419 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq); 420 tcp_output(tp); 421 } 422 } 423 424 /* 425 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. 426 * tcp_template used to store this data in mbufs, but we now recopy it out 427 * of the tcpcb each time to conserve mbufs. 428 */ 429 void 430 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr, boolean_t tso) 431 { 432 struct inpcb *inp = tp->t_inpcb; 433 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr; 434 435 #ifdef INET6 436 if (inp->inp_vflag & INP_IPV6) { 437 struct ip6_hdr *ip6; 438 439 ip6 = (struct ip6_hdr *)ip_ptr; 440 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | 441 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK); 442 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | 443 (IPV6_VERSION & IPV6_VERSION_MASK); 444 ip6->ip6_nxt = IPPROTO_TCP; 445 ip6->ip6_plen = sizeof(struct tcphdr); 446 ip6->ip6_src = inp->in6p_laddr; 447 ip6->ip6_dst = inp->in6p_faddr; 448 tcp_hdr->th_sum = 0; 449 } else 450 #endif 451 { 452 struct ip *ip = (struct ip *) ip_ptr; 453 u_int plen; 454 455 ip->ip_vhl = IP_VHL_BORING; 456 ip->ip_tos = 0; 457 ip->ip_len = 0; 458 ip->ip_id = 0; 459 ip->ip_off = 0; 460 ip->ip_ttl = 0; 461 ip->ip_sum = 0; 462 ip->ip_p = IPPROTO_TCP; 463 ip->ip_src = inp->inp_laddr; 464 ip->ip_dst = inp->inp_faddr; 465 466 if (tso) 467 plen = htons(IPPROTO_TCP); 468 else 469 plen = htons(sizeof(struct tcphdr) + IPPROTO_TCP); 470 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr, 471 ip->ip_dst.s_addr, plen); 472 } 473 474 tcp_hdr->th_sport = inp->inp_lport; 475 tcp_hdr->th_dport = inp->inp_fport; 476 tcp_hdr->th_seq = 0; 477 tcp_hdr->th_ack = 0; 478 tcp_hdr->th_x2 = 0; 479 tcp_hdr->th_off = 5; 480 tcp_hdr->th_flags = 0; 481 tcp_hdr->th_win = 0; 482 tcp_hdr->th_urp = 0; 483 } 484 485 /* 486 * Create template to be used to send tcp packets on a connection. 487 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only 488 * use for this function is in keepalives, which use tcp_respond. 489 */ 490 struct tcptemp * 491 tcp_maketemplate(struct tcpcb *tp) 492 { 493 struct tcptemp *tmp; 494 495 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL) 496 return (NULL); 497 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t, FALSE); 498 return (tmp); 499 } 500 501 void 502 tcp_freetemplate(struct tcptemp *tmp) 503 { 504 mpipe_free(&tcptemp_mpipe, tmp); 505 } 506 507 /* 508 * Send a single message to the TCP at address specified by 509 * the given TCP/IP header. If m == NULL, then we make a copy 510 * of the tcpiphdr at ti and send directly to the addressed host. 511 * This is used to force keep alive messages out using the TCP 512 * template for a connection. If flags are given then we send 513 * a message back to the TCP which originated the * segment ti, 514 * and discard the mbuf containing it and any other attached mbufs. 515 * 516 * In any case the ack and sequence number of the transmitted 517 * segment are as specified by the parameters. 518 * 519 * NOTE: If m != NULL, then ti must point to *inside* the mbuf. 520 */ 521 void 522 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, 523 tcp_seq ack, tcp_seq seq, int flags) 524 { 525 int tlen; 526 int win = 0; 527 struct route *ro = NULL; 528 struct route sro; 529 struct ip *ip = ipgen; 530 struct tcphdr *nth; 531 int ipflags = 0; 532 struct route_in6 *ro6 = NULL; 533 struct route_in6 sro6; 534 struct ip6_hdr *ip6 = ipgen; 535 boolean_t use_tmpro = TRUE; 536 #ifdef INET6 537 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6); 538 #else 539 const boolean_t isipv6 = FALSE; 540 #endif 541 542 if (tp != NULL) { 543 if (!(flags & TH_RST)) { 544 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv); 545 if (win < 0) 546 win = 0; 547 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 548 win = (long)TCP_MAXWIN << tp->rcv_scale; 549 } 550 /* 551 * Don't use the route cache of a listen socket, 552 * it is not MPSAFE; use temporary route cache. 553 */ 554 if (tp->t_state != TCPS_LISTEN) { 555 if (isipv6) 556 ro6 = &tp->t_inpcb->in6p_route; 557 else 558 ro = &tp->t_inpcb->inp_route; 559 use_tmpro = FALSE; 560 } 561 } 562 if (use_tmpro) { 563 if (isipv6) { 564 ro6 = &sro6; 565 bzero(ro6, sizeof *ro6); 566 } else { 567 ro = &sro; 568 bzero(ro, sizeof *ro); 569 } 570 } 571 if (m == NULL) { 572 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 573 if (m == NULL) 574 return; 575 tlen = 0; 576 m->m_data += max_linkhdr; 577 if (isipv6) { 578 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); 579 ip6 = mtod(m, struct ip6_hdr *); 580 nth = (struct tcphdr *)(ip6 + 1); 581 } else { 582 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip)); 583 ip = mtod(m, struct ip *); 584 nth = (struct tcphdr *)(ip + 1); 585 } 586 bcopy(th, nth, sizeof(struct tcphdr)); 587 flags = TH_ACK; 588 } else { 589 m_freem(m->m_next); 590 m->m_next = NULL; 591 m->m_data = (caddr_t)ipgen; 592 /* m_len is set later */ 593 tlen = 0; 594 #define xchg(a, b, type) { type t; t = a; a = b; b = t; } 595 if (isipv6) { 596 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); 597 nth = (struct tcphdr *)(ip6 + 1); 598 } else { 599 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long); 600 nth = (struct tcphdr *)(ip + 1); 601 } 602 if (th != nth) { 603 /* 604 * this is usually a case when an extension header 605 * exists between the IPv6 header and the 606 * TCP header. 607 */ 608 nth->th_sport = th->th_sport; 609 nth->th_dport = th->th_dport; 610 } 611 xchg(nth->th_dport, nth->th_sport, n_short); 612 #undef xchg 613 } 614 if (isipv6) { 615 ip6->ip6_flow = 0; 616 ip6->ip6_vfc = IPV6_VERSION; 617 ip6->ip6_nxt = IPPROTO_TCP; 618 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen)); 619 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 620 } else { 621 tlen += sizeof(struct tcpiphdr); 622 ip->ip_len = tlen; 623 ip->ip_ttl = ip_defttl; 624 } 625 m->m_len = tlen; 626 m->m_pkthdr.len = tlen; 627 m->m_pkthdr.rcvif = NULL; 628 nth->th_seq = htonl(seq); 629 nth->th_ack = htonl(ack); 630 nth->th_x2 = 0; 631 nth->th_off = sizeof(struct tcphdr) >> 2; 632 nth->th_flags = flags; 633 if (tp != NULL) 634 nth->th_win = htons((u_short) (win >> tp->rcv_scale)); 635 else 636 nth->th_win = htons((u_short)win); 637 nth->th_urp = 0; 638 if (isipv6) { 639 nth->th_sum = 0; 640 nth->th_sum = in6_cksum(m, IPPROTO_TCP, 641 sizeof(struct ip6_hdr), 642 tlen - sizeof(struct ip6_hdr)); 643 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL, 644 (ro6 && ro6->ro_rt) ? 645 ro6->ro_rt->rt_ifp : NULL); 646 } else { 647 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 648 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); 649 m->m_pkthdr.csum_flags = CSUM_TCP; 650 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 651 m->m_pkthdr.csum_thlen = sizeof(struct tcphdr); 652 } 653 #ifdef TCPDEBUG 654 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) 655 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); 656 #endif 657 if (isipv6) { 658 ip6_output(m, NULL, ro6, ipflags, NULL, NULL, 659 tp ? tp->t_inpcb : NULL); 660 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) { 661 RTFREE(ro6->ro_rt); 662 ro6->ro_rt = NULL; 663 } 664 } else { 665 ipflags |= IP_DEBUGROUTE; 666 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL); 667 if ((ro == &sro) && (ro->ro_rt != NULL)) { 668 RTFREE(ro->ro_rt); 669 ro->ro_rt = NULL; 670 } 671 } 672 } 673 674 /* 675 * Create a new TCP control block, making an 676 * empty reassembly queue and hooking it to the argument 677 * protocol control block. The `inp' parameter must have 678 * come from the zone allocator set up in tcp_init(). 679 */ 680 struct tcpcb * 681 tcp_newtcpcb(struct inpcb *inp) 682 { 683 struct inp_tp *it; 684 struct tcpcb *tp; 685 #ifdef INET6 686 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 687 #else 688 const boolean_t isipv6 = FALSE; 689 #endif 690 691 it = (struct inp_tp *)inp; 692 tp = &it->tcb; 693 bzero(tp, sizeof(struct tcpcb)); 694 TAILQ_INIT(&tp->t_segq); 695 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 696 tp->t_rxtthresh = tcprexmtthresh; 697 698 /* Set up our timeouts. */ 699 tp->tt_rexmt = &it->inp_tp_rexmt; 700 tp->tt_persist = &it->inp_tp_persist; 701 tp->tt_keep = &it->inp_tp_keep; 702 tp->tt_2msl = &it->inp_tp_2msl; 703 tp->tt_delack = &it->inp_tp_delack; 704 tcp_inittimers(tp); 705 706 /* 707 * Zero out timer message. We don't create it here, 708 * since the current CPU may not be the owner of this 709 * inpcb. 710 */ 711 tp->tt_msg = &it->inp_tp_timermsg; 712 bzero(tp->tt_msg, sizeof(*tp->tt_msg)); 713 714 tp->t_keepinit = tcp_keepinit; 715 tp->t_keepidle = tcp_keepidle; 716 tp->t_keepintvl = tcp_keepintvl; 717 tp->t_keepcnt = tcp_keepcnt; 718 tp->t_maxidle = tp->t_keepintvl * tp->t_keepcnt; 719 720 if (tcp_do_ncr) 721 tp->t_flags |= TF_NCR; 722 if (tcp_do_rfc1323) 723 tp->t_flags |= (TF_REQ_SCALE | TF_REQ_TSTMP); 724 725 tp->t_inpcb = inp; /* XXX */ 726 tp->t_state = TCPS_CLOSED; 727 /* 728 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 729 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 730 * reasonable initial retransmit time. 731 */ 732 tp->t_srtt = TCPTV_SRTTBASE; 733 tp->t_rttvar = 734 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 735 tp->t_rttmin = tcp_rexmit_min; 736 tp->t_rxtcur = TCPTV_RTOBASE; 737 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 738 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 739 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 740 tp->snd_last = ticks; 741 tp->t_rcvtime = ticks; 742 /* 743 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 744 * because the socket may be bound to an IPv6 wildcard address, 745 * which may match an IPv4-mapped IPv6 address. 746 */ 747 inp->inp_ip_ttl = ip_defttl; 748 inp->inp_ppcb = tp; 749 tcp_sack_tcpcb_init(tp); 750 751 tp->tt_sndmore = &it->inp_tp_sndmore; 752 tcp_output_init(tp); 753 754 return (tp); /* XXX */ 755 } 756 757 /* 758 * Drop a TCP connection, reporting the specified error. 759 * If connection is synchronized, then send a RST to peer. 760 */ 761 struct tcpcb * 762 tcp_drop(struct tcpcb *tp, int error) 763 { 764 struct socket *so = tp->t_inpcb->inp_socket; 765 766 if (TCPS_HAVERCVDSYN(tp->t_state)) { 767 tp->t_state = TCPS_CLOSED; 768 tcp_output(tp); 769 tcpstat.tcps_drops++; 770 } else 771 tcpstat.tcps_conndrops++; 772 if (error == ETIMEDOUT && tp->t_softerror) 773 error = tp->t_softerror; 774 so->so_error = error; 775 return (tcp_close(tp)); 776 } 777 778 struct netmsg_listen_detach { 779 struct netmsg_base base; 780 struct tcpcb *nm_tp; 781 struct tcpcb *nm_tp_inh; 782 }; 783 784 static void 785 tcp_listen_detach_handler(netmsg_t msg) 786 { 787 struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg; 788 struct tcpcb *tp = nmsg->nm_tp; 789 int cpu = mycpuid, nextcpu; 790 791 if (tp->t_flags & TF_LISTEN) 792 syncache_destroy(tp, nmsg->nm_tp_inh); 793 794 in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]); 795 796 nextcpu = cpu + 1; 797 if (nextcpu < ncpus2) 798 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nmsg->base.lmsg); 799 else 800 lwkt_replymsg(&nmsg->base.lmsg, 0); 801 } 802 803 /* 804 * Close a TCP control block: 805 * discard all space held by the tcp 806 * discard internet protocol block 807 * wake up any sleepers 808 */ 809 struct tcpcb * 810 tcp_close(struct tcpcb *tp) 811 { 812 struct tseg_qent *q; 813 struct inpcb *inp = tp->t_inpcb; 814 struct inpcb *inp_inh = NULL; 815 struct tcpcb *tp_inh = NULL; 816 struct socket *so = inp->inp_socket; 817 struct rtentry *rt; 818 boolean_t dosavessthresh; 819 #ifdef INET6 820 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 821 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0); 822 #else 823 const boolean_t isipv6 = FALSE; 824 #endif 825 826 if (tp->t_flags & TF_LISTEN) { 827 /* 828 * Pending socket/syncache inheritance 829 * 830 * If this is a listen(2) socket, find another listen(2) 831 * socket in the same local group, which could inherit 832 * the syncache and sockets pending on the completion 833 * and incompletion queues. 834 * 835 * NOTE: 836 * Currently the inheritance could only happen on the 837 * listen(2) sockets w/ SO_REUSEPORT set. 838 */ 839 KASSERT(&curthread->td_msgport == netisr_cpuport(0), 840 ("listen socket close not in netisr0")); 841 inp_inh = in_pcblocalgroup_last(&tcbinfo[0], inp); 842 if (inp_inh != NULL) 843 tp_inh = intotcpcb(inp_inh); 844 } 845 846 /* 847 * INP_WILDCARD_MP indicates that listen(2) has been called on 848 * this socket. This implies: 849 * - A wildcard inp's hash is replicated for each protocol thread. 850 * - Syncache for this inp grows independently in each protocol 851 * thread. 852 * - There is more than one cpu 853 * 854 * We have to chain a message to the rest of the protocol threads 855 * to cleanup the wildcard hash and the syncache. The cleanup 856 * in the current protocol thread is defered till the end of this 857 * function. 858 * 859 * NOTE: 860 * After cleanup the inp's hash and syncache entries, this inp will 861 * no longer be available to the rest of the protocol threads, so we 862 * are safe to whack the inp in the following code. 863 */ 864 if (inp->inp_flags & INP_WILDCARD_MP) { 865 struct netmsg_listen_detach nmsg; 866 867 KKASSERT(so->so_port == netisr_cpuport(0)); 868 KKASSERT(&curthread->td_msgport == netisr_cpuport(0)); 869 KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]); 870 871 netmsg_init(&nmsg.base, NULL, &curthread->td_msgport, 872 MSGF_PRIORITY, tcp_listen_detach_handler); 873 nmsg.nm_tp = tp; 874 nmsg.nm_tp_inh = tp_inh; 875 lwkt_domsg(netisr_cpuport(1), &nmsg.base.lmsg, 0); 876 877 inp->inp_flags &= ~INP_WILDCARD_MP; 878 } 879 880 KKASSERT(tp->t_state != TCPS_TERMINATING); 881 tp->t_state = TCPS_TERMINATING; 882 883 /* 884 * Make sure that all of our timers are stopped before we 885 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL), 886 * timers are never used. If timer message is never created 887 * (tp->tt_msg->tt_tcb == NULL), timers are never used too. 888 */ 889 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) { 890 tcp_callout_stop(tp, tp->tt_rexmt); 891 tcp_callout_stop(tp, tp->tt_persist); 892 tcp_callout_stop(tp, tp->tt_keep); 893 tcp_callout_stop(tp, tp->tt_2msl); 894 tcp_callout_stop(tp, tp->tt_delack); 895 } 896 897 if (tp->t_flags & TF_ONOUTPUTQ) { 898 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 899 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 900 tp->t_flags &= ~TF_ONOUTPUTQ; 901 } 902 903 /* 904 * If we got enough samples through the srtt filter, 905 * save the rtt and rttvar in the routing entry. 906 * 'Enough' is arbitrarily defined as the 16 samples. 907 * 16 samples is enough for the srtt filter to converge 908 * to within 5% of the correct value; fewer samples and 909 * we could save a very bogus rtt. 910 * 911 * Don't update the default route's characteristics and don't 912 * update anything that the user "locked". 913 */ 914 if (tp->t_rttupdated >= 16) { 915 u_long i = 0; 916 917 if (isipv6) { 918 struct sockaddr_in6 *sin6; 919 920 if ((rt = inp->in6p_route.ro_rt) == NULL) 921 goto no_valid_rt; 922 sin6 = (struct sockaddr_in6 *)rt_key(rt); 923 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 924 goto no_valid_rt; 925 } else 926 if ((rt = inp->inp_route.ro_rt) == NULL || 927 ((struct sockaddr_in *)rt_key(rt))-> 928 sin_addr.s_addr == INADDR_ANY) 929 goto no_valid_rt; 930 931 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 932 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 933 if (rt->rt_rmx.rmx_rtt && i) 934 /* 935 * filter this update to half the old & half 936 * the new values, converting scale. 937 * See route.h and tcp_var.h for a 938 * description of the scaling constants. 939 */ 940 rt->rt_rmx.rmx_rtt = 941 (rt->rt_rmx.rmx_rtt + i) / 2; 942 else 943 rt->rt_rmx.rmx_rtt = i; 944 tcpstat.tcps_cachedrtt++; 945 } 946 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 947 i = tp->t_rttvar * 948 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 949 if (rt->rt_rmx.rmx_rttvar && i) 950 rt->rt_rmx.rmx_rttvar = 951 (rt->rt_rmx.rmx_rttvar + i) / 2; 952 else 953 rt->rt_rmx.rmx_rttvar = i; 954 tcpstat.tcps_cachedrttvar++; 955 } 956 /* 957 * The old comment here said: 958 * update the pipelimit (ssthresh) if it has been updated 959 * already or if a pipesize was specified & the threshhold 960 * got below half the pipesize. I.e., wait for bad news 961 * before we start updating, then update on both good 962 * and bad news. 963 * 964 * But we want to save the ssthresh even if no pipesize is 965 * specified explicitly in the route, because such 966 * connections still have an implicit pipesize specified 967 * by the global tcp_sendspace. In the absence of a reliable 968 * way to calculate the pipesize, it will have to do. 969 */ 970 i = tp->snd_ssthresh; 971 if (rt->rt_rmx.rmx_sendpipe != 0) 972 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 973 else 974 dosavessthresh = (i < so->so_snd.ssb_hiwat/2); 975 if (dosavessthresh || 976 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 977 (rt->rt_rmx.rmx_ssthresh != 0))) { 978 /* 979 * convert the limit from user data bytes to 980 * packets then to packet data bytes. 981 */ 982 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 983 if (i < 2) 984 i = 2; 985 i *= tp->t_maxseg + 986 (isipv6 ? 987 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 988 sizeof(struct tcpiphdr)); 989 if (rt->rt_rmx.rmx_ssthresh) 990 rt->rt_rmx.rmx_ssthresh = 991 (rt->rt_rmx.rmx_ssthresh + i) / 2; 992 else 993 rt->rt_rmx.rmx_ssthresh = i; 994 tcpstat.tcps_cachedssthresh++; 995 } 996 } 997 998 no_valid_rt: 999 /* free the reassembly queue, if any */ 1000 while((q = TAILQ_FIRST(&tp->t_segq)) != NULL) { 1001 TAILQ_REMOVE(&tp->t_segq, q, tqe_q); 1002 m_freem(q->tqe_m); 1003 kfree(q, M_TSEGQ); 1004 atomic_add_int(&tcp_reass_qsize, -1); 1005 } 1006 /* throw away SACK blocks in scoreboard*/ 1007 if (TCP_DO_SACK(tp)) 1008 tcp_sack_destroy(&tp->scb); 1009 1010 inp->inp_ppcb = NULL; 1011 soisdisconnected(so); 1012 /* note: pcb detached later on */ 1013 1014 tcp_destroy_timermsg(tp); 1015 tcp_output_cancel(tp); 1016 1017 if (tp->t_flags & TF_LISTEN) { 1018 syncache_destroy(tp, tp_inh); 1019 if (inp_inh != NULL && inp_inh->inp_socket != NULL) { 1020 /* 1021 * Pending sockets inheritance only needs 1022 * to be done once in the current thread, 1023 * i.e. netisr0. 1024 */ 1025 soinherit(so, inp_inh->inp_socket); 1026 } 1027 } 1028 1029 so_async_rcvd_drop(so); 1030 /* Drop the reference for the asynchronized pru_rcvd */ 1031 sofree(so); 1032 1033 /* 1034 * NOTE: 1035 * pcbdetach removes any wildcard hash entry on the current CPU. 1036 */ 1037 #ifdef INET6 1038 if (isafinet6) 1039 in6_pcbdetach(inp); 1040 else 1041 #endif 1042 in_pcbdetach(inp); 1043 1044 tcpstat.tcps_closed++; 1045 return (NULL); 1046 } 1047 1048 static __inline void 1049 tcp_drain_oncpu(struct inpcbhead *head) 1050 { 1051 struct inpcb *marker; 1052 struct inpcb *inpb; 1053 struct tcpcb *tcpb; 1054 struct tseg_qent *te; 1055 1056 /* 1057 * Allows us to block while running the list 1058 */ 1059 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1060 marker->inp_flags |= INP_PLACEMARKER; 1061 LIST_INSERT_HEAD(head, marker, inp_list); 1062 1063 while ((inpb = LIST_NEXT(marker, inp_list)) != NULL) { 1064 if ((inpb->inp_flags & INP_PLACEMARKER) == 0 && 1065 (tcpb = intotcpcb(inpb)) != NULL && 1066 (te = TAILQ_FIRST(&tcpb->t_segq)) != NULL) { 1067 TAILQ_REMOVE(&tcpb->t_segq, te, tqe_q); 1068 if (te->tqe_th->th_flags & TH_FIN) 1069 tcpb->t_flags &= ~TF_QUEDFIN; 1070 m_freem(te->tqe_m); 1071 kfree(te, M_TSEGQ); 1072 atomic_add_int(&tcp_reass_qsize, -1); 1073 /* retry */ 1074 } else { 1075 LIST_REMOVE(marker, inp_list); 1076 LIST_INSERT_AFTER(inpb, marker, inp_list); 1077 } 1078 } 1079 LIST_REMOVE(marker, inp_list); 1080 kfree(marker, M_TEMP); 1081 } 1082 1083 struct netmsg_tcp_drain { 1084 struct netmsg_base base; 1085 struct inpcbhead *nm_head; 1086 }; 1087 1088 static void 1089 tcp_drain_handler(netmsg_t msg) 1090 { 1091 struct netmsg_tcp_drain *nm = (void *)msg; 1092 1093 tcp_drain_oncpu(nm->nm_head); 1094 lwkt_replymsg(&nm->base.lmsg, 0); 1095 } 1096 1097 void 1098 tcp_drain(void) 1099 { 1100 int cpu; 1101 1102 if (!do_tcpdrain) 1103 return; 1104 1105 /* 1106 * Walk the tcpbs, if existing, and flush the reassembly queue, 1107 * if there is one... 1108 * XXX: The "Net/3" implementation doesn't imply that the TCP 1109 * reassembly queue should be flushed, but in a situation 1110 * where we're really low on mbufs, this is potentially 1111 * useful. 1112 */ 1113 for (cpu = 0; cpu < ncpus2; cpu++) { 1114 struct netmsg_tcp_drain *nm; 1115 1116 if (cpu == mycpu->gd_cpuid) { 1117 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead); 1118 } else { 1119 nm = kmalloc(sizeof(struct netmsg_tcp_drain), 1120 M_LWKTMSG, M_NOWAIT); 1121 if (nm == NULL) 1122 continue; 1123 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1124 0, tcp_drain_handler); 1125 nm->nm_head = &tcbinfo[cpu].pcblisthead; 1126 lwkt_sendmsg(netisr_cpuport(cpu), &nm->base.lmsg); 1127 } 1128 } 1129 } 1130 1131 /* 1132 * Notify a tcp user of an asynchronous error; 1133 * store error as soft error, but wake up user 1134 * (for now, won't do anything until can select for soft error). 1135 * 1136 * Do not wake up user since there currently is no mechanism for 1137 * reporting soft errors (yet - a kqueue filter may be added). 1138 */ 1139 static void 1140 tcp_notify(struct inpcb *inp, int error) 1141 { 1142 struct tcpcb *tp = intotcpcb(inp); 1143 1144 /* 1145 * Ignore some errors if we are hooked up. 1146 * If connection hasn't completed, has retransmitted several times, 1147 * and receives a second error, give up now. This is better 1148 * than waiting a long time to establish a connection that 1149 * can never complete. 1150 */ 1151 if (tp->t_state == TCPS_ESTABLISHED && 1152 (error == EHOSTUNREACH || error == ENETUNREACH || 1153 error == EHOSTDOWN)) { 1154 return; 1155 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1156 tp->t_softerror) 1157 tcp_drop(tp, error); 1158 else 1159 tp->t_softerror = error; 1160 #if 0 1161 wakeup(&so->so_timeo); 1162 sorwakeup(so); 1163 sowwakeup(so); 1164 #endif 1165 } 1166 1167 static int 1168 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1169 { 1170 int error, i, n; 1171 struct inpcb *marker; 1172 struct inpcb *inp; 1173 globaldata_t gd; 1174 int origcpu, ccpu; 1175 1176 error = 0; 1177 n = 0; 1178 1179 /* 1180 * The process of preparing the TCB list is too time-consuming and 1181 * resource-intensive to repeat twice on every request. 1182 */ 1183 if (req->oldptr == NULL) { 1184 for (ccpu = 0; ccpu < ncpus; ++ccpu) { 1185 gd = globaldata_find(ccpu); 1186 n += tcbinfo[gd->gd_cpuid].ipi_count; 1187 } 1188 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1189 return (0); 1190 } 1191 1192 if (req->newptr != NULL) 1193 return (EPERM); 1194 1195 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1196 marker->inp_flags |= INP_PLACEMARKER; 1197 1198 /* 1199 * OK, now we're committed to doing something. Run the inpcb list 1200 * for each cpu in the system and construct the output. Use a 1201 * list placemarker to deal with list changes occuring during 1202 * copyout blockages (but otherwise depend on being on the correct 1203 * cpu to avoid races). 1204 */ 1205 origcpu = mycpu->gd_cpuid; 1206 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) { 1207 globaldata_t rgd; 1208 caddr_t inp_ppcb; 1209 struct xtcpcb xt; 1210 int cpu_id; 1211 1212 cpu_id = (origcpu + ccpu) % ncpus; 1213 if ((smp_active_mask & CPUMASK(cpu_id)) == 0) 1214 continue; 1215 rgd = globaldata_find(cpu_id); 1216 lwkt_setcpu_self(rgd); 1217 1218 n = tcbinfo[cpu_id].ipi_count; 1219 1220 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list); 1221 i = 0; 1222 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1223 /* 1224 * process a snapshot of pcbs, ignoring placemarkers 1225 * and using our own to allow SYSCTL_OUT to block. 1226 */ 1227 LIST_REMOVE(marker, inp_list); 1228 LIST_INSERT_AFTER(inp, marker, inp_list); 1229 1230 if (inp->inp_flags & INP_PLACEMARKER) 1231 continue; 1232 if (prison_xinpcb(req->td, inp)) 1233 continue; 1234 1235 xt.xt_len = sizeof xt; 1236 bcopy(inp, &xt.xt_inp, sizeof *inp); 1237 inp_ppcb = inp->inp_ppcb; 1238 if (inp_ppcb != NULL) 1239 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1240 else 1241 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1242 if (inp->inp_socket) 1243 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1244 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1245 break; 1246 ++i; 1247 } 1248 LIST_REMOVE(marker, inp_list); 1249 if (error == 0 && i < n) { 1250 bzero(&xt, sizeof xt); 1251 xt.xt_len = sizeof xt; 1252 while (i < n) { 1253 error = SYSCTL_OUT(req, &xt, sizeof xt); 1254 if (error) 1255 break; 1256 ++i; 1257 } 1258 } 1259 } 1260 1261 /* 1262 * Make sure we are on the same cpu we were on originally, since 1263 * higher level callers expect this. Also don't pollute caches with 1264 * migrated userland data by (eventually) returning to userland 1265 * on a different cpu. 1266 */ 1267 lwkt_setcpu_self(globaldata_find(origcpu)); 1268 kfree(marker, M_TEMP); 1269 return (error); 1270 } 1271 1272 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1273 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1274 1275 static int 1276 tcp_getcred(SYSCTL_HANDLER_ARGS) 1277 { 1278 struct sockaddr_in addrs[2]; 1279 struct inpcb *inp; 1280 int cpu; 1281 int error; 1282 1283 error = priv_check(req->td, PRIV_ROOT); 1284 if (error != 0) 1285 return (error); 1286 error = SYSCTL_IN(req, addrs, sizeof addrs); 1287 if (error != 0) 1288 return (error); 1289 crit_enter(); 1290 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1291 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1292 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1293 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1294 if (inp == NULL || inp->inp_socket == NULL) { 1295 error = ENOENT; 1296 goto out; 1297 } 1298 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1299 out: 1300 crit_exit(); 1301 return (error); 1302 } 1303 1304 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1305 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1306 1307 #ifdef INET6 1308 static int 1309 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1310 { 1311 struct sockaddr_in6 addrs[2]; 1312 struct inpcb *inp; 1313 int error; 1314 boolean_t mapped = FALSE; 1315 1316 error = priv_check(req->td, PRIV_ROOT); 1317 if (error != 0) 1318 return (error); 1319 error = SYSCTL_IN(req, addrs, sizeof addrs); 1320 if (error != 0) 1321 return (error); 1322 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { 1323 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) 1324 mapped = TRUE; 1325 else 1326 return (EINVAL); 1327 } 1328 crit_enter(); 1329 if (mapped) { 1330 inp = in_pcblookup_hash(&tcbinfo[0], 1331 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], 1332 addrs[1].sin6_port, 1333 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], 1334 addrs[0].sin6_port, 1335 0, NULL); 1336 } else { 1337 inp = in6_pcblookup_hash(&tcbinfo[0], 1338 &addrs[1].sin6_addr, addrs[1].sin6_port, 1339 &addrs[0].sin6_addr, addrs[0].sin6_port, 1340 0, NULL); 1341 } 1342 if (inp == NULL || inp->inp_socket == NULL) { 1343 error = ENOENT; 1344 goto out; 1345 } 1346 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1347 out: 1348 crit_exit(); 1349 return (error); 1350 } 1351 1352 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1353 0, 0, 1354 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1355 #endif 1356 1357 struct netmsg_tcp_notify { 1358 struct netmsg_base base; 1359 void (*nm_notify)(struct inpcb *, int); 1360 struct in_addr nm_faddr; 1361 int nm_arg; 1362 }; 1363 1364 static void 1365 tcp_notifyall_oncpu(netmsg_t msg) 1366 { 1367 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg; 1368 int nextcpu; 1369 1370 in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nm->nm_faddr, 1371 nm->nm_arg, nm->nm_notify); 1372 1373 nextcpu = mycpuid + 1; 1374 if (nextcpu < ncpus2) 1375 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nm->base.lmsg); 1376 else 1377 lwkt_replymsg(&nm->base.lmsg, 0); 1378 } 1379 1380 void 1381 tcp_ctlinput(netmsg_t msg) 1382 { 1383 int cmd = msg->ctlinput.nm_cmd; 1384 struct sockaddr *sa = msg->ctlinput.nm_arg; 1385 struct ip *ip = msg->ctlinput.nm_extra; 1386 struct tcphdr *th; 1387 struct in_addr faddr; 1388 struct inpcb *inp; 1389 struct tcpcb *tp; 1390 void (*notify)(struct inpcb *, int) = tcp_notify; 1391 tcp_seq icmpseq; 1392 int arg, cpu; 1393 1394 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) { 1395 goto done; 1396 } 1397 1398 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1399 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1400 goto done; 1401 1402 arg = inetctlerrmap[cmd]; 1403 if (cmd == PRC_QUENCH) { 1404 notify = tcp_quench; 1405 } else if (icmp_may_rst && 1406 (cmd == PRC_UNREACH_ADMIN_PROHIB || 1407 cmd == PRC_UNREACH_PORT || 1408 cmd == PRC_TIMXCEED_INTRANS) && 1409 ip != NULL) { 1410 notify = tcp_drop_syn_sent; 1411 } else if (cmd == PRC_MSGSIZE) { 1412 struct icmp *icmp = (struct icmp *) 1413 ((caddr_t)ip - offsetof(struct icmp, icmp_ip)); 1414 1415 arg = ntohs(icmp->icmp_nextmtu); 1416 notify = tcp_mtudisc; 1417 } else if (PRC_IS_REDIRECT(cmd)) { 1418 ip = NULL; 1419 notify = in_rtchange; 1420 } else if (cmd == PRC_HOSTDEAD) { 1421 ip = NULL; 1422 } 1423 1424 if (ip != NULL) { 1425 crit_enter(); 1426 th = (struct tcphdr *)((caddr_t)ip + 1427 (IP_VHL_HL(ip->ip_vhl) << 2)); 1428 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport, 1429 ip->ip_src.s_addr, th->th_sport); 1430 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport, 1431 ip->ip_src, th->th_sport, 0, NULL); 1432 if ((inp != NULL) && (inp->inp_socket != NULL)) { 1433 icmpseq = htonl(th->th_seq); 1434 tp = intotcpcb(inp); 1435 if (SEQ_GEQ(icmpseq, tp->snd_una) && 1436 SEQ_LT(icmpseq, tp->snd_max)) 1437 (*notify)(inp, arg); 1438 } else { 1439 struct in_conninfo inc; 1440 1441 inc.inc_fport = th->th_dport; 1442 inc.inc_lport = th->th_sport; 1443 inc.inc_faddr = faddr; 1444 inc.inc_laddr = ip->ip_src; 1445 #ifdef INET6 1446 inc.inc_isipv6 = 0; 1447 #endif 1448 syncache_unreach(&inc, th); 1449 } 1450 crit_exit(); 1451 } else { 1452 struct netmsg_tcp_notify *nm; 1453 1454 KKASSERT(&curthread->td_msgport == netisr_cpuport(0)); 1455 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT); 1456 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1457 0, tcp_notifyall_oncpu); 1458 nm->nm_faddr = faddr; 1459 nm->nm_arg = arg; 1460 nm->nm_notify = notify; 1461 1462 lwkt_sendmsg(netisr_cpuport(0), &nm->base.lmsg); 1463 } 1464 done: 1465 lwkt_replymsg(&msg->lmsg, 0); 1466 } 1467 1468 #ifdef INET6 1469 1470 void 1471 tcp6_ctlinput(netmsg_t msg) 1472 { 1473 int cmd = msg->ctlinput.nm_cmd; 1474 struct sockaddr *sa = msg->ctlinput.nm_arg; 1475 void *d = msg->ctlinput.nm_extra; 1476 struct tcphdr th; 1477 void (*notify) (struct inpcb *, int) = tcp_notify; 1478 struct ip6_hdr *ip6; 1479 struct mbuf *m; 1480 struct ip6ctlparam *ip6cp = NULL; 1481 const struct sockaddr_in6 *sa6_src = NULL; 1482 int off; 1483 struct tcp_portonly { 1484 u_int16_t th_sport; 1485 u_int16_t th_dport; 1486 } *thp; 1487 int arg; 1488 1489 if (sa->sa_family != AF_INET6 || 1490 sa->sa_len != sizeof(struct sockaddr_in6)) { 1491 goto out; 1492 } 1493 1494 arg = 0; 1495 if (cmd == PRC_QUENCH) 1496 notify = tcp_quench; 1497 else if (cmd == PRC_MSGSIZE) { 1498 struct ip6ctlparam *ip6cp = d; 1499 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6; 1500 1501 arg = ntohl(icmp6->icmp6_mtu); 1502 notify = tcp_mtudisc; 1503 } else if (!PRC_IS_REDIRECT(cmd) && 1504 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) { 1505 goto out; 1506 } 1507 1508 /* if the parameter is from icmp6, decode it. */ 1509 if (d != NULL) { 1510 ip6cp = (struct ip6ctlparam *)d; 1511 m = ip6cp->ip6c_m; 1512 ip6 = ip6cp->ip6c_ip6; 1513 off = ip6cp->ip6c_off; 1514 sa6_src = ip6cp->ip6c_src; 1515 } else { 1516 m = NULL; 1517 ip6 = NULL; 1518 off = 0; /* fool gcc */ 1519 sa6_src = &sa6_any; 1520 } 1521 1522 if (ip6 != NULL) { 1523 struct in_conninfo inc; 1524 /* 1525 * XXX: We assume that when IPV6 is non NULL, 1526 * M and OFF are valid. 1527 */ 1528 1529 /* check if we can safely examine src and dst ports */ 1530 if (m->m_pkthdr.len < off + sizeof *thp) 1531 goto out; 1532 1533 bzero(&th, sizeof th); 1534 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1535 1536 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport, 1537 (struct sockaddr *)ip6cp->ip6c_src, 1538 th.th_sport, cmd, arg, notify); 1539 1540 inc.inc_fport = th.th_dport; 1541 inc.inc_lport = th.th_sport; 1542 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1543 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1544 inc.inc_isipv6 = 1; 1545 syncache_unreach(&inc, &th); 1546 } else { 1547 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0, 1548 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify); 1549 } 1550 out: 1551 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0); 1552 } 1553 1554 #endif 1555 1556 /* 1557 * Following is where TCP initial sequence number generation occurs. 1558 * 1559 * There are two places where we must use initial sequence numbers: 1560 * 1. In SYN-ACK packets. 1561 * 2. In SYN packets. 1562 * 1563 * All ISNs for SYN-ACK packets are generated by the syncache. See 1564 * tcp_syncache.c for details. 1565 * 1566 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1567 * depends on this property. In addition, these ISNs should be 1568 * unguessable so as to prevent connection hijacking. To satisfy 1569 * the requirements of this situation, the algorithm outlined in 1570 * RFC 1948 is used to generate sequence numbers. 1571 * 1572 * Implementation details: 1573 * 1574 * Time is based off the system timer, and is corrected so that it 1575 * increases by one megabyte per second. This allows for proper 1576 * recycling on high speed LANs while still leaving over an hour 1577 * before rollover. 1578 * 1579 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1580 * between seeding of isn_secret. This is normally set to zero, 1581 * as reseeding should not be necessary. 1582 * 1583 */ 1584 1585 #define ISN_BYTES_PER_SECOND 1048576 1586 1587 u_char isn_secret[32]; 1588 int isn_last_reseed; 1589 MD5_CTX isn_ctx; 1590 1591 tcp_seq 1592 tcp_new_isn(struct tcpcb *tp) 1593 { 1594 u_int32_t md5_buffer[4]; 1595 tcp_seq new_isn; 1596 1597 /* Seed if this is the first use, reseed if requested. */ 1598 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1599 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1600 < (u_int)ticks))) { 1601 read_random_unlimited(&isn_secret, sizeof isn_secret); 1602 isn_last_reseed = ticks; 1603 } 1604 1605 /* Compute the md5 hash and return the ISN. */ 1606 MD5Init(&isn_ctx); 1607 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1608 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1609 #ifdef INET6 1610 if (tp->t_inpcb->inp_vflag & INP_IPV6) { 1611 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1612 sizeof(struct in6_addr)); 1613 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1614 sizeof(struct in6_addr)); 1615 } else 1616 #endif 1617 { 1618 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1619 sizeof(struct in_addr)); 1620 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1621 sizeof(struct in_addr)); 1622 } 1623 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1624 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1625 new_isn = (tcp_seq) md5_buffer[0]; 1626 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1627 return (new_isn); 1628 } 1629 1630 /* 1631 * When a source quench is received, close congestion window 1632 * to one segment. We will gradually open it again as we proceed. 1633 */ 1634 void 1635 tcp_quench(struct inpcb *inp, int error) 1636 { 1637 struct tcpcb *tp = intotcpcb(inp); 1638 1639 if (tp != NULL) { 1640 tp->snd_cwnd = tp->t_maxseg; 1641 tp->snd_wacked = 0; 1642 } 1643 } 1644 1645 /* 1646 * When a specific ICMP unreachable message is received and the 1647 * connection state is SYN-SENT, drop the connection. This behavior 1648 * is controlled by the icmp_may_rst sysctl. 1649 */ 1650 void 1651 tcp_drop_syn_sent(struct inpcb *inp, int error) 1652 { 1653 struct tcpcb *tp = intotcpcb(inp); 1654 1655 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT)) 1656 tcp_drop(tp, error); 1657 } 1658 1659 /* 1660 * When a `need fragmentation' ICMP is received, update our idea of the MSS 1661 * based on the new value in the route. Also nudge TCP to send something, 1662 * since we know the packet we just sent was dropped. 1663 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1664 */ 1665 void 1666 tcp_mtudisc(struct inpcb *inp, int mtu) 1667 { 1668 struct tcpcb *tp = intotcpcb(inp); 1669 struct rtentry *rt; 1670 struct socket *so = inp->inp_socket; 1671 int maxopd, mss; 1672 #ifdef INET6 1673 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0); 1674 #else 1675 const boolean_t isipv6 = FALSE; 1676 #endif 1677 1678 if (tp == NULL) 1679 return; 1680 1681 /* 1682 * If no MTU is provided in the ICMP message, use the 1683 * next lower likely value, as specified in RFC 1191. 1684 */ 1685 if (mtu == 0) { 1686 int oldmtu; 1687 1688 oldmtu = tp->t_maxopd + 1689 (isipv6 ? 1690 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1691 sizeof(struct tcpiphdr)); 1692 mtu = ip_next_mtu(oldmtu, 0); 1693 } 1694 1695 if (isipv6) 1696 rt = tcp_rtlookup6(&inp->inp_inc); 1697 else 1698 rt = tcp_rtlookup(&inp->inp_inc); 1699 if (rt != NULL) { 1700 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu) 1701 mtu = rt->rt_rmx.rmx_mtu; 1702 1703 maxopd = mtu - 1704 (isipv6 ? 1705 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1706 sizeof(struct tcpiphdr)); 1707 1708 /* 1709 * XXX - The following conditional probably violates the TCP 1710 * spec. The problem is that, since we don't know the 1711 * other end's MSS, we are supposed to use a conservative 1712 * default. But, if we do that, then MTU discovery will 1713 * never actually take place, because the conservative 1714 * default is much less than the MTUs typically seen 1715 * on the Internet today. For the moment, we'll sweep 1716 * this under the carpet. 1717 * 1718 * The conservative default might not actually be a problem 1719 * if the only case this occurs is when sending an initial 1720 * SYN with options and data to a host we've never talked 1721 * to before. Then, they will reply with an MSS value which 1722 * will get recorded and the new parameters should get 1723 * recomputed. For Further Study. 1724 */ 1725 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd) 1726 maxopd = rt->rt_rmx.rmx_mssopt; 1727 } else 1728 maxopd = mtu - 1729 (isipv6 ? 1730 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1731 sizeof(struct tcpiphdr)); 1732 1733 if (tp->t_maxopd <= maxopd) 1734 return; 1735 tp->t_maxopd = maxopd; 1736 1737 mss = maxopd; 1738 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) == 1739 (TF_REQ_TSTMP | TF_RCVD_TSTMP)) 1740 mss -= TCPOLEN_TSTAMP_APPA; 1741 1742 /* round down to multiple of MCLBYTES */ 1743 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */ 1744 if (mss > MCLBYTES) 1745 mss &= ~(MCLBYTES - 1); 1746 #else 1747 if (mss > MCLBYTES) 1748 mss = (mss / MCLBYTES) * MCLBYTES; 1749 #endif 1750 1751 if (so->so_snd.ssb_hiwat < mss) 1752 mss = so->so_snd.ssb_hiwat; 1753 1754 tp->t_maxseg = mss; 1755 tp->t_rtttime = 0; 1756 tp->snd_nxt = tp->snd_una; 1757 tcp_output(tp); 1758 tcpstat.tcps_mturesent++; 1759 } 1760 1761 /* 1762 * Look-up the routing entry to the peer of this inpcb. If no route 1763 * is found and it cannot be allocated the return NULL. This routine 1764 * is called by TCP routines that access the rmx structure and by tcp_mss 1765 * to get the interface MTU. 1766 */ 1767 struct rtentry * 1768 tcp_rtlookup(struct in_conninfo *inc) 1769 { 1770 struct route *ro = &inc->inc_route; 1771 1772 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1773 /* No route yet, so try to acquire one */ 1774 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1775 /* 1776 * unused portions of the structure MUST be zero'd 1777 * out because rtalloc() treats it as opaque data 1778 */ 1779 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1780 ro->ro_dst.sa_family = AF_INET; 1781 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1782 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1783 inc->inc_faddr; 1784 rtalloc(ro); 1785 } 1786 } 1787 return (ro->ro_rt); 1788 } 1789 1790 #ifdef INET6 1791 struct rtentry * 1792 tcp_rtlookup6(struct in_conninfo *inc) 1793 { 1794 struct route_in6 *ro6 = &inc->inc6_route; 1795 1796 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1797 /* No route yet, so try to acquire one */ 1798 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1799 /* 1800 * unused portions of the structure MUST be zero'd 1801 * out because rtalloc() treats it as opaque data 1802 */ 1803 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1804 ro6->ro_dst.sin6_family = AF_INET6; 1805 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1806 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1807 rtalloc((struct route *)ro6); 1808 } 1809 } 1810 return (ro6->ro_rt); 1811 } 1812 #endif 1813 1814 #ifdef IPSEC 1815 /* compute ESP/AH header size for TCP, including outer IP header. */ 1816 size_t 1817 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1818 { 1819 struct inpcb *inp; 1820 struct mbuf *m; 1821 size_t hdrsiz; 1822 struct ip *ip; 1823 struct tcphdr *th; 1824 1825 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1826 return (0); 1827 MGETHDR(m, MB_DONTWAIT, MT_DATA); 1828 if (!m) 1829 return (0); 1830 1831 #ifdef INET6 1832 if (inp->inp_vflag & INP_IPV6) { 1833 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1834 1835 th = (struct tcphdr *)(ip6 + 1); 1836 m->m_pkthdr.len = m->m_len = 1837 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1838 tcp_fillheaders(tp, ip6, th, FALSE); 1839 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1840 } else 1841 #endif 1842 { 1843 ip = mtod(m, struct ip *); 1844 th = (struct tcphdr *)(ip + 1); 1845 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1846 tcp_fillheaders(tp, ip, th, FALSE); 1847 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1848 } 1849 1850 m_free(m); 1851 return (hdrsiz); 1852 } 1853 #endif 1854 1855 /* 1856 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1857 * 1858 * This code attempts to calculate the bandwidth-delay product as a 1859 * means of determining the optimal window size to maximize bandwidth, 1860 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1861 * routers. This code also does a fairly good job keeping RTTs in check 1862 * across slow links like modems. We implement an algorithm which is very 1863 * similar (but not meant to be) TCP/Vegas. The code operates on the 1864 * transmitter side of a TCP connection and so only effects the transmit 1865 * side of the connection. 1866 * 1867 * BACKGROUND: TCP makes no provision for the management of buffer space 1868 * at the end points or at the intermediate routers and switches. A TCP 1869 * stream, whether using NewReno or not, will eventually buffer as 1870 * many packets as it is able and the only reason this typically works is 1871 * due to the fairly small default buffers made available for a connection 1872 * (typicaly 16K or 32K). As machines use larger windows and/or window 1873 * scaling it is now fairly easy for even a single TCP connection to blow-out 1874 * all available buffer space not only on the local interface, but on 1875 * intermediate routers and switches as well. NewReno makes a misguided 1876 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1877 * then backing off, then steadily increasing the window again until another 1878 * failure occurs, ad-infinitum. This results in terrible oscillation that 1879 * is only made worse as network loads increase and the idea of intentionally 1880 * blowing out network buffers is, frankly, a terrible way to manage network 1881 * resources. 1882 * 1883 * It is far better to limit the transmit window prior to the failure 1884 * condition being achieved. There are two general ways to do this: First 1885 * you can 'scan' through different transmit window sizes and locate the 1886 * point where the RTT stops increasing, indicating that you have filled the 1887 * pipe, then scan backwards until you note that RTT stops decreasing, then 1888 * repeat ad-infinitum. This method works in principle but has severe 1889 * implementation issues due to RTT variances, timer granularity, and 1890 * instability in the algorithm which can lead to many false positives and 1891 * create oscillations as well as interact badly with other TCP streams 1892 * implementing the same algorithm. 1893 * 1894 * The second method is to limit the window to the bandwidth delay product 1895 * of the link. This is the method we implement. RTT variances and our 1896 * own manipulation of the congestion window, bwnd, can potentially 1897 * destabilize the algorithm. For this reason we have to stabilize the 1898 * elements used to calculate the window. We do this by using the minimum 1899 * observed RTT, the long term average of the observed bandwidth, and 1900 * by adding two segments worth of slop. It isn't perfect but it is able 1901 * to react to changing conditions and gives us a very stable basis on 1902 * which to extend the algorithm. 1903 */ 1904 void 1905 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1906 { 1907 u_long bw; 1908 u_long bwnd; 1909 int save_ticks; 1910 int delta_ticks; 1911 1912 /* 1913 * If inflight_enable is disabled in the middle of a tcp connection, 1914 * make sure snd_bwnd is effectively disabled. 1915 */ 1916 if (!tcp_inflight_enable) { 1917 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1918 tp->snd_bandwidth = 0; 1919 return; 1920 } 1921 1922 /* 1923 * Validate the delta time. If a connection is new or has been idle 1924 * a long time we have to reset the bandwidth calculator. 1925 */ 1926 save_ticks = ticks; 1927 delta_ticks = save_ticks - tp->t_bw_rtttime; 1928 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1929 tp->t_bw_rtttime = ticks; 1930 tp->t_bw_rtseq = ack_seq; 1931 if (tp->snd_bandwidth == 0) 1932 tp->snd_bandwidth = tcp_inflight_min; 1933 return; 1934 } 1935 if (delta_ticks == 0) 1936 return; 1937 1938 /* 1939 * Sanity check, plus ignore pure window update acks. 1940 */ 1941 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1942 return; 1943 1944 /* 1945 * Figure out the bandwidth. Due to the tick granularity this 1946 * is a very rough number and it MUST be averaged over a fairly 1947 * long period of time. XXX we need to take into account a link 1948 * that is not using all available bandwidth, but for now our 1949 * slop will ramp us up if this case occurs and the bandwidth later 1950 * increases. 1951 */ 1952 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 1953 tp->t_bw_rtttime = save_ticks; 1954 tp->t_bw_rtseq = ack_seq; 1955 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4; 1956 1957 tp->snd_bandwidth = bw; 1958 1959 /* 1960 * Calculate the semi-static bandwidth delay product, plus two maximal 1961 * segments. The additional slop puts us squarely in the sweet 1962 * spot and also handles the bandwidth run-up case. Without the 1963 * slop we could be locking ourselves into a lower bandwidth. 1964 * 1965 * Situations Handled: 1966 * (1) Prevents over-queueing of packets on LANs, especially on 1967 * high speed LANs, allowing larger TCP buffers to be 1968 * specified, and also does a good job preventing 1969 * over-queueing of packets over choke points like modems 1970 * (at least for the transmit side). 1971 * 1972 * (2) Is able to handle changing network loads (bandwidth 1973 * drops so bwnd drops, bandwidth increases so bwnd 1974 * increases). 1975 * 1976 * (3) Theoretically should stabilize in the face of multiple 1977 * connections implementing the same algorithm (this may need 1978 * a little work). 1979 * 1980 * (4) Stability value (defaults to 20 = 2 maximal packets) can 1981 * be adjusted with a sysctl but typically only needs to be on 1982 * very slow connections. A value no smaller then 5 should 1983 * be used, but only reduce this default if you have no other 1984 * choice. 1985 */ 1986 1987 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2) 1988 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 1989 tcp_inflight_stab * (int)tp->t_maxseg / 10; 1990 #undef USERTT 1991 1992 if (tcp_inflight_debug > 0) { 1993 static int ltime; 1994 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) { 1995 ltime = ticks; 1996 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n", 1997 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd); 1998 } 1999 } 2000 if ((long)bwnd < tcp_inflight_min) 2001 bwnd = tcp_inflight_min; 2002 if (bwnd > tcp_inflight_max) 2003 bwnd = tcp_inflight_max; 2004 if ((long)bwnd < tp->t_maxseg * 2) 2005 bwnd = tp->t_maxseg * 2; 2006 tp->snd_bwnd = bwnd; 2007 } 2008 2009 static void 2010 tcp_rmx_iwsegs(struct tcpcb *tp, u_long *maxsegs, u_long *capsegs) 2011 { 2012 struct rtentry *rt; 2013 struct inpcb *inp = tp->t_inpcb; 2014 #ifdef INET6 2015 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE); 2016 #else 2017 const boolean_t isipv6 = FALSE; 2018 #endif 2019 2020 /* XXX */ 2021 if (tcp_iw_maxsegs < TCP_IW_MAXSEGS_DFLT) 2022 tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 2023 if (tcp_iw_capsegs < TCP_IW_CAPSEGS_DFLT) 2024 tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 2025 2026 if (isipv6) 2027 rt = tcp_rtlookup6(&inp->inp_inc); 2028 else 2029 rt = tcp_rtlookup(&inp->inp_inc); 2030 if (rt == NULL || 2031 rt->rt_rmx.rmx_iwmaxsegs < TCP_IW_MAXSEGS_DFLT || 2032 rt->rt_rmx.rmx_iwcapsegs < TCP_IW_CAPSEGS_DFLT) { 2033 *maxsegs = tcp_iw_maxsegs; 2034 *capsegs = tcp_iw_capsegs; 2035 return; 2036 } 2037 *maxsegs = rt->rt_rmx.rmx_iwmaxsegs; 2038 *capsegs = rt->rt_rmx.rmx_iwcapsegs; 2039 } 2040 2041 u_long 2042 tcp_initial_window(struct tcpcb *tp) 2043 { 2044 if (tcp_do_rfc3390) { 2045 /* 2046 * RFC3390: 2047 * "If the SYN or SYN/ACK is lost, the initial window 2048 * used by a sender after a correctly transmitted SYN 2049 * MUST be one segment consisting of MSS bytes." 2050 * 2051 * However, we do something a little bit more aggressive 2052 * then RFC3390 here: 2053 * - Only if time spent in the SYN or SYN|ACK retransmition 2054 * >= 3 seconds, the IW is reduced. We do this mainly 2055 * because when RFC3390 is published, the initial RTO is 2056 * still 3 seconds (the threshold we test here), while 2057 * after RFC6298, the initial RTO is 1 second. This 2058 * behaviour probably still falls within the spirit of 2059 * RFC3390. 2060 * - When IW is reduced, 2*MSS is used instead of 1*MSS. 2061 * Mainly to avoid sender and receiver deadlock until 2062 * delayed ACK timer expires. And even RFC2581 does not 2063 * try to reduce IW upon SYN or SYN|ACK retransmition 2064 * timeout. 2065 * 2066 * See also: 2067 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03 2068 */ 2069 if (tp->t_rxtsyn >= TCPTV_RTOBASE3) { 2070 return (2 * tp->t_maxseg); 2071 } else { 2072 u_long maxsegs, capsegs; 2073 2074 tcp_rmx_iwsegs(tp, &maxsegs, &capsegs); 2075 return min(maxsegs * tp->t_maxseg, 2076 max(2 * tp->t_maxseg, capsegs * 1460)); 2077 } 2078 } else { 2079 /* 2080 * Even RFC2581 (back to 1999) allows 2*SMSS IW. 2081 * 2082 * Mainly to avoid sender and receiver deadlock 2083 * until delayed ACK timer expires. 2084 */ 2085 return (2 * tp->t_maxseg); 2086 } 2087 } 2088 2089 #ifdef TCP_SIGNATURE 2090 /* 2091 * Compute TCP-MD5 hash of a TCP segment. (RFC2385) 2092 * 2093 * We do this over ip, tcphdr, segment data, and the key in the SADB. 2094 * When called from tcp_input(), we can be sure that th_sum has been 2095 * zeroed out and verified already. 2096 * 2097 * Return 0 if successful, otherwise return -1. 2098 * 2099 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a 2100 * search with the destination IP address, and a 'magic SPI' to be 2101 * determined by the application. This is hardcoded elsewhere to 1179 2102 * right now. Another branch of this code exists which uses the SPD to 2103 * specify per-application flows but it is unstable. 2104 */ 2105 int 2106 tcpsignature_compute( 2107 struct mbuf *m, /* mbuf chain */ 2108 int len, /* length of TCP data */ 2109 int optlen, /* length of TCP options */ 2110 u_char *buf, /* storage for MD5 digest */ 2111 u_int direction) /* direction of flow */ 2112 { 2113 struct ippseudo ippseudo; 2114 MD5_CTX ctx; 2115 int doff; 2116 struct ip *ip; 2117 struct ipovly *ipovly; 2118 struct secasvar *sav; 2119 struct tcphdr *th; 2120 #ifdef INET6 2121 struct ip6_hdr *ip6; 2122 struct in6_addr in6; 2123 uint32_t plen; 2124 uint16_t nhdr; 2125 #endif /* INET6 */ 2126 u_short savecsum; 2127 2128 KASSERT(m != NULL, ("passed NULL mbuf. Game over.")); 2129 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature")); 2130 /* 2131 * Extract the destination from the IP header in the mbuf. 2132 */ 2133 ip = mtod(m, struct ip *); 2134 #ifdef INET6 2135 ip6 = NULL; /* Make the compiler happy. */ 2136 #endif /* INET6 */ 2137 /* 2138 * Look up an SADB entry which matches the address found in 2139 * the segment. 2140 */ 2141 switch (IP_VHL_V(ip->ip_vhl)) { 2142 case IPVERSION: 2143 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst, 2144 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2145 break; 2146 #ifdef INET6 2147 case (IPV6_VERSION >> 4): 2148 ip6 = mtod(m, struct ip6_hdr *); 2149 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst, 2150 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2151 break; 2152 #endif /* INET6 */ 2153 default: 2154 return (EINVAL); 2155 /* NOTREACHED */ 2156 break; 2157 } 2158 if (sav == NULL) { 2159 kprintf("%s: SADB lookup failed\n", __func__); 2160 return (EINVAL); 2161 } 2162 MD5Init(&ctx); 2163 2164 /* 2165 * Step 1: Update MD5 hash with IP pseudo-header. 2166 * 2167 * XXX The ippseudo header MUST be digested in network byte order, 2168 * or else we'll fail the regression test. Assume all fields we've 2169 * been doing arithmetic on have been in host byte order. 2170 * XXX One cannot depend on ipovly->ih_len here. When called from 2171 * tcp_output(), the underlying ip_len member has not yet been set. 2172 */ 2173 switch (IP_VHL_V(ip->ip_vhl)) { 2174 case IPVERSION: 2175 ipovly = (struct ipovly *)ip; 2176 ippseudo.ippseudo_src = ipovly->ih_src; 2177 ippseudo.ippseudo_dst = ipovly->ih_dst; 2178 ippseudo.ippseudo_pad = 0; 2179 ippseudo.ippseudo_p = IPPROTO_TCP; 2180 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen); 2181 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo)); 2182 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip)); 2183 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen; 2184 break; 2185 #ifdef INET6 2186 /* 2187 * RFC 2385, 2.0 Proposal 2188 * For IPv6, the pseudo-header is as described in RFC 2460, namely the 2189 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero- 2190 * extended next header value (to form 32 bits), and 32-bit segment 2191 * length. 2192 * Note: Upper-Layer Packet Length comes before Next Header. 2193 */ 2194 case (IPV6_VERSION >> 4): 2195 in6 = ip6->ip6_src; 2196 in6_clearscope(&in6); 2197 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2198 in6 = ip6->ip6_dst; 2199 in6_clearscope(&in6); 2200 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2201 plen = htonl(len + sizeof(struct tcphdr) + optlen); 2202 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t)); 2203 nhdr = 0; 2204 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2205 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2206 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2207 nhdr = IPPROTO_TCP; 2208 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2209 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr)); 2210 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen; 2211 break; 2212 #endif /* INET6 */ 2213 default: 2214 return (EINVAL); 2215 /* NOTREACHED */ 2216 break; 2217 } 2218 /* 2219 * Step 2: Update MD5 hash with TCP header, excluding options. 2220 * The TCP checksum must be set to zero. 2221 */ 2222 savecsum = th->th_sum; 2223 th->th_sum = 0; 2224 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr)); 2225 th->th_sum = savecsum; 2226 /* 2227 * Step 3: Update MD5 hash with TCP segment data. 2228 * Use m_apply() to avoid an early m_pullup(). 2229 */ 2230 if (len > 0) 2231 m_apply(m, doff, len, tcpsignature_apply, &ctx); 2232 /* 2233 * Step 4: Update MD5 hash with shared secret. 2234 */ 2235 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth)); 2236 MD5Final(buf, &ctx); 2237 key_sa_recordxfer(sav, m); 2238 key_freesav(sav); 2239 return (0); 2240 } 2241 2242 int 2243 tcpsignature_apply(void *fstate, void *data, unsigned int len) 2244 { 2245 2246 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len); 2247 return (0); 2248 } 2249 #endif /* TCP_SIGNATURE */ 2250