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