1 /* 2 * Copyright (c) 2003, 2004 Jeffrey M. Hsu. All rights reserved. 3 * Copyright (c) 2003, 2004 The DragonFly Project. All rights reserved. 4 * 5 * This code is derived from software contributed to The DragonFly Project 6 * by Jeffrey M. Hsu. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. Neither the name of The DragonFly Project nor the names of its 17 * contributors may be used to endorse or promote products derived 18 * from this software without specific, prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 22 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 23 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 24 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 25 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 26 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 27 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 28 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 29 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 30 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 */ 33 34 /* 35 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 36 * The Regents of the University of California. All rights reserved. 37 * 38 * Redistribution and use in source and binary forms, with or without 39 * modification, are permitted provided that the following conditions 40 * are met: 41 * 1. Redistributions of source code must retain the above copyright 42 * notice, this list of conditions and the following disclaimer. 43 * 2. Redistributions in binary form must reproduce the above copyright 44 * notice, this list of conditions and the following disclaimer in the 45 * documentation and/or other materials provided with the distribution. 46 * 3. Neither the name of the University nor the names of its contributors 47 * may be used to endorse or promote products derived from this software 48 * without specific prior written permission. 49 * 50 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 51 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 52 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 53 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 54 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 55 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 56 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 57 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 58 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 59 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 60 * SUCH DAMAGE. 61 * 62 * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95 63 * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $ 64 */ 65 66 #include "opt_compat.h" 67 #include "opt_inet.h" 68 #include "opt_inet6.h" 69 #include "opt_ipsec.h" 70 #include "opt_tcpdebug.h" 71 72 #include <sys/param.h> 73 #include <sys/systm.h> 74 #include <sys/callout.h> 75 #include <sys/kernel.h> 76 #include <sys/sysctl.h> 77 #include <sys/malloc.h> 78 #include <sys/mpipe.h> 79 #include <sys/mbuf.h> 80 #ifdef INET6 81 #include <sys/domain.h> 82 #endif 83 #include <sys/proc.h> 84 #include <sys/priv.h> 85 #include <sys/socket.h> 86 #include <sys/socketops.h> 87 #include <sys/socketvar.h> 88 #include <sys/protosw.h> 89 #include <sys/random.h> 90 #include <sys/in_cksum.h> 91 #include <sys/ktr.h> 92 93 #include <net/route.h> 94 #include <net/if.h> 95 #include <net/netisr2.h> 96 97 #define _IP_VHL 98 #include <netinet/in.h> 99 #include <netinet/in_systm.h> 100 #include <netinet/ip.h> 101 #include <netinet/ip6.h> 102 #include <netinet/in_pcb.h> 103 #include <netinet6/in6_pcb.h> 104 #include <netinet/in_var.h> 105 #include <netinet/ip_var.h> 106 #include <netinet6/ip6_var.h> 107 #include <netinet/ip_icmp.h> 108 #ifdef INET6 109 #include <netinet/icmp6.h> 110 #endif 111 #include <netinet/tcp.h> 112 #include <netinet/tcp_fsm.h> 113 #include <netinet/tcp_seq.h> 114 #include <netinet/tcp_timer.h> 115 #include <netinet/tcp_timer2.h> 116 #include <netinet/tcp_var.h> 117 #include <netinet6/tcp6_var.h> 118 #include <netinet/tcpip.h> 119 #ifdef TCPDEBUG 120 #include <netinet/tcp_debug.h> 121 #endif 122 #include <netinet6/ip6protosw.h> 123 124 #ifdef IPSEC 125 #include <netinet6/ipsec.h> 126 #include <netproto/key/key.h> 127 #ifdef INET6 128 #include <netinet6/ipsec6.h> 129 #endif 130 #endif 131 132 #ifdef FAST_IPSEC 133 #include <netproto/ipsec/ipsec.h> 134 #ifdef INET6 135 #include <netproto/ipsec/ipsec6.h> 136 #endif 137 #define IPSEC 138 #endif 139 140 #include <sys/md5.h> 141 #include <machine/smp.h> 142 143 #include <sys/msgport2.h> 144 #include <sys/mplock2.h> 145 #include <net/netmsg2.h> 146 147 #if !defined(KTR_TCP) 148 #define KTR_TCP KTR_ALL 149 #endif 150 /* 151 KTR_INFO_MASTER(tcp); 152 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0); 153 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0); 154 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0); 155 #define logtcp(name) KTR_LOG(tcp_ ## name) 156 */ 157 158 #define TCP_IW_MAXSEGS_DFLT 4 159 #define TCP_IW_CAPSEGS_DFLT 3 160 161 struct inpcbinfo tcbinfo[MAXCPU]; 162 struct tcpcbackqhead tcpcbackq[MAXCPU]; 163 164 static struct lwkt_token tcp_port_token = 165 LWKT_TOKEN_INITIALIZER(tcp_port_token); 166 167 int tcp_mssdflt = TCP_MSS; 168 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 169 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size"); 170 171 #ifdef INET6 172 int tcp_v6mssdflt = TCP6_MSS; 173 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW, 174 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6"); 175 #endif 176 177 /* 178 * Minimum MSS we accept and use. This prevents DoS attacks where 179 * we are forced to a ridiculous low MSS like 20 and send hundreds 180 * of packets instead of one. The effect scales with the available 181 * bandwidth and quickly saturates the CPU and network interface 182 * with packet generation and sending. Set to zero to disable MINMSS 183 * checking. This setting prevents us from sending too small packets. 184 */ 185 int tcp_minmss = TCP_MINMSS; 186 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW, 187 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size"); 188 189 #if 0 190 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ; 191 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW, 192 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time"); 193 #endif 194 195 int tcp_do_rfc1323 = 1; 196 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW, 197 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions"); 198 199 static int tcp_tcbhashsize = 0; 200 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD, 201 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable"); 202 203 static int do_tcpdrain = 1; 204 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0, 205 "Enable tcp_drain routine for extra help when low on mbufs"); 206 207 static int icmp_may_rst = 1; 208 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0, 209 "Certain ICMP unreachable messages may abort connections in SYN_SENT"); 210 211 static int tcp_isn_reseed_interval = 0; 212 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW, 213 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret"); 214 215 /* 216 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on 217 * by default, but with generous values which should allow maximal 218 * bandwidth. In particular, the slop defaults to 50 (5 packets). 219 * 220 * The reason for doing this is that the limiter is the only mechanism we 221 * have which seems to do a really good job preventing receiver RX rings 222 * on network interfaces from getting blown out. Even though GigE/10GigE 223 * is supposed to flow control it looks like either it doesn't actually 224 * do it or Open Source drivers do not properly enable it. 225 * 226 * People using the limiter to reduce bottlenecks on slower WAN connections 227 * should set the slop to 20 (2 packets). 228 */ 229 static int tcp_inflight_enable = 1; 230 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW, 231 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting"); 232 233 static int tcp_inflight_debug = 0; 234 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW, 235 &tcp_inflight_debug, 0, "Debug TCP inflight calculations"); 236 237 static int tcp_inflight_min = 6144; 238 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW, 239 &tcp_inflight_min, 0, "Lower bound for TCP inflight window"); 240 241 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT; 242 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW, 243 &tcp_inflight_max, 0, "Upper bound for TCP inflight window"); 244 245 static int tcp_inflight_stab = 50; 246 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW, 247 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 3 packets)"); 248 249 static int tcp_do_rfc3390 = 1; 250 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_RW, 251 &tcp_do_rfc3390, 0, 252 "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)"); 253 254 static u_long tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 255 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwmaxsegs, CTLFLAG_RW, 256 &tcp_iw_maxsegs, 0, "TCP IW segments max"); 257 258 static u_long tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 259 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwcapsegs, CTLFLAG_RW, 260 &tcp_iw_capsegs, 0, "TCP IW segments"); 261 262 int tcp_low_rtobase = 1; 263 SYSCTL_INT(_net_inet_tcp, OID_AUTO, low_rtobase, CTLFLAG_RW, 264 &tcp_low_rtobase, 0, "Lowering the Initial RTO (RFC 6298)"); 265 266 static int tcp_do_ncr = 1; 267 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr, CTLFLAG_RW, 268 &tcp_do_ncr, 0, "Non-Congestion Robustness (RFC 4653)"); 269 270 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives"); 271 static struct malloc_pipe tcptemp_mpipe; 272 273 static void tcp_willblock(void); 274 static void tcp_notify (struct inpcb *, int); 275 276 struct tcp_stats tcpstats_percpu[MAXCPU] __cachealign; 277 278 static int 279 sysctl_tcpstats(SYSCTL_HANDLER_ARGS) 280 { 281 int cpu, error = 0; 282 283 for (cpu = 0; cpu < ncpus; ++cpu) { 284 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu], 285 sizeof(struct tcp_stats)))) 286 break; 287 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu], 288 sizeof(struct tcp_stats)))) 289 break; 290 } 291 292 return (error); 293 } 294 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW), 295 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics"); 296 297 /* 298 * Target size of TCP PCB hash tables. Must be a power of two. 299 * 300 * Note that this can be overridden by the kernel environment 301 * variable net.inet.tcp.tcbhashsize 302 */ 303 #ifndef TCBHASHSIZE 304 #define TCBHASHSIZE 512 305 #endif 306 307 /* 308 * This is the actual shape of what we allocate using the zone 309 * allocator. Doing it this way allows us to protect both structures 310 * using the same generation count, and also eliminates the overhead 311 * of allocating tcpcbs separately. By hiding the structure here, 312 * we avoid changing most of the rest of the code (although it needs 313 * to be changed, eventually, for greater efficiency). 314 */ 315 #define ALIGNMENT 32 316 #define ALIGNM1 (ALIGNMENT - 1) 317 struct inp_tp { 318 union { 319 struct inpcb inp; 320 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1]; 321 } inp_tp_u; 322 struct tcpcb tcb; 323 struct tcp_callout inp_tp_rexmt; 324 struct tcp_callout inp_tp_persist; 325 struct tcp_callout inp_tp_keep; 326 struct tcp_callout inp_tp_2msl; 327 struct tcp_callout inp_tp_delack; 328 struct netmsg_tcp_timer inp_tp_timermsg; 329 struct netmsg_base inp_tp_sndmore; 330 }; 331 #undef ALIGNMENT 332 #undef ALIGNM1 333 334 /* 335 * Tcp initialization 336 */ 337 void 338 tcp_init(void) 339 { 340 struct inpcbporthead *porthashbase; 341 struct inpcbinfo *ticb; 342 u_long porthashmask; 343 int hashsize = TCBHASHSIZE; 344 int cpu; 345 346 /* 347 * note: tcptemp is used for keepalives, and it is ok for an 348 * allocation to fail so do not specify MPF_INT. 349 */ 350 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp), 351 25, -1, 0, NULL, NULL, NULL); 352 353 tcp_delacktime = TCPTV_DELACK; 354 tcp_keepinit = TCPTV_KEEP_INIT; 355 tcp_keepidle = TCPTV_KEEP_IDLE; 356 tcp_keepintvl = TCPTV_KEEPINTVL; 357 tcp_maxpersistidle = TCPTV_KEEP_IDLE; 358 tcp_msl = TCPTV_MSL; 359 tcp_rexmit_min = TCPTV_MIN; 360 tcp_rexmit_slop = TCPTV_CPU_VAR; 361 362 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize); 363 if (!powerof2(hashsize)) { 364 kprintf("WARNING: TCB hash size not a power of 2\n"); 365 hashsize = 512; /* safe default */ 366 } 367 tcp_tcbhashsize = hashsize; 368 porthashbase = hashinit(hashsize, M_PCB, &porthashmask); 369 370 for (cpu = 0; cpu < ncpus2; cpu++) { 371 ticb = &tcbinfo[cpu]; 372 in_pcbinfo_init(ticb); 373 ticb->cpu = cpu; 374 ticb->hashbase = hashinit(hashsize, M_PCB, 375 &ticb->hashmask); 376 ticb->porthashbase = porthashbase; 377 ticb->porthashmask = porthashmask; 378 ticb->porttoken = &tcp_port_token; 379 #if 0 380 ticb->porthashbase = hashinit(hashsize, M_PCB, 381 &ticb->porthashmask); 382 #endif 383 ticb->wildcardhashbase = hashinit(hashsize, M_PCB, 384 &ticb->wildcardhashmask); 385 ticb->localgrphashbase = hashinit(hashsize, M_PCB, 386 &ticb->localgrphashmask); 387 ticb->ipi_size = sizeof(struct inp_tp); 388 TAILQ_INIT(&tcpcbackq[cpu]); 389 } 390 391 tcp_reass_maxseg = nmbclusters / 16; 392 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg); 393 394 #ifdef INET6 395 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) 396 #else 397 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) 398 #endif 399 if (max_protohdr < TCP_MINPROTOHDR) 400 max_protohdr = TCP_MINPROTOHDR; 401 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) 402 panic("tcp_init"); 403 #undef TCP_MINPROTOHDR 404 405 /* 406 * Initialize TCP statistics counters for each CPU. 407 */ 408 for (cpu = 0; cpu < ncpus; ++cpu) { 409 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats)); 410 } 411 412 syncache_init(); 413 netisr_register_rollup(tcp_willblock, NETISR_ROLLUP_PRIO_TCP); 414 } 415 416 static void 417 tcp_willblock(void) 418 { 419 struct tcpcb *tp; 420 int cpu = mycpu->gd_cpuid; 421 422 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) { 423 KKASSERT(tp->t_flags & TF_ONOUTPUTQ); 424 tp->t_flags &= ~TF_ONOUTPUTQ; 425 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq); 426 tcp_output(tp); 427 } 428 } 429 430 /* 431 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. 432 * tcp_template used to store this data in mbufs, but we now recopy it out 433 * of the tcpcb each time to conserve mbufs. 434 */ 435 void 436 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr, boolean_t tso) 437 { 438 struct inpcb *inp = tp->t_inpcb; 439 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr; 440 441 #ifdef INET6 442 if (inp->inp_vflag & INP_IPV6) { 443 struct ip6_hdr *ip6; 444 445 ip6 = (struct ip6_hdr *)ip_ptr; 446 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | 447 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK); 448 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | 449 (IPV6_VERSION & IPV6_VERSION_MASK); 450 ip6->ip6_nxt = IPPROTO_TCP; 451 ip6->ip6_plen = sizeof(struct tcphdr); 452 ip6->ip6_src = inp->in6p_laddr; 453 ip6->ip6_dst = inp->in6p_faddr; 454 tcp_hdr->th_sum = 0; 455 } else 456 #endif 457 { 458 struct ip *ip = (struct ip *) ip_ptr; 459 u_int plen; 460 461 ip->ip_vhl = IP_VHL_BORING; 462 ip->ip_tos = 0; 463 ip->ip_len = 0; 464 ip->ip_id = 0; 465 ip->ip_off = 0; 466 ip->ip_ttl = 0; 467 ip->ip_sum = 0; 468 ip->ip_p = IPPROTO_TCP; 469 ip->ip_src = inp->inp_laddr; 470 ip->ip_dst = inp->inp_faddr; 471 472 if (tso) 473 plen = htons(IPPROTO_TCP); 474 else 475 plen = htons(sizeof(struct tcphdr) + IPPROTO_TCP); 476 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr, 477 ip->ip_dst.s_addr, plen); 478 } 479 480 tcp_hdr->th_sport = inp->inp_lport; 481 tcp_hdr->th_dport = inp->inp_fport; 482 tcp_hdr->th_seq = 0; 483 tcp_hdr->th_ack = 0; 484 tcp_hdr->th_x2 = 0; 485 tcp_hdr->th_off = 5; 486 tcp_hdr->th_flags = 0; 487 tcp_hdr->th_win = 0; 488 tcp_hdr->th_urp = 0; 489 } 490 491 /* 492 * Create template to be used to send tcp packets on a connection. 493 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only 494 * use for this function is in keepalives, which use tcp_respond. 495 */ 496 struct tcptemp * 497 tcp_maketemplate(struct tcpcb *tp) 498 { 499 struct tcptemp *tmp; 500 501 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL) 502 return (NULL); 503 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t, FALSE); 504 return (tmp); 505 } 506 507 void 508 tcp_freetemplate(struct tcptemp *tmp) 509 { 510 mpipe_free(&tcptemp_mpipe, tmp); 511 } 512 513 /* 514 * Send a single message to the TCP at address specified by 515 * the given TCP/IP header. If m == NULL, then we make a copy 516 * of the tcpiphdr at ti and send directly to the addressed host. 517 * This is used to force keep alive messages out using the TCP 518 * template for a connection. If flags are given then we send 519 * a message back to the TCP which originated the * segment ti, 520 * and discard the mbuf containing it and any other attached mbufs. 521 * 522 * In any case the ack and sequence number of the transmitted 523 * segment are as specified by the parameters. 524 * 525 * NOTE: If m != NULL, then ti must point to *inside* the mbuf. 526 */ 527 void 528 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, 529 tcp_seq ack, tcp_seq seq, int flags) 530 { 531 int tlen; 532 int win = 0; 533 struct route *ro = NULL; 534 struct route sro; 535 struct ip *ip = ipgen; 536 struct tcphdr *nth; 537 int ipflags = 0; 538 struct route_in6 *ro6 = NULL; 539 struct route_in6 sro6; 540 struct ip6_hdr *ip6 = ipgen; 541 boolean_t use_tmpro = TRUE; 542 #ifdef INET6 543 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6); 544 #else 545 const boolean_t isipv6 = FALSE; 546 #endif 547 548 if (tp != NULL) { 549 if (!(flags & TH_RST)) { 550 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv); 551 if (win < 0) 552 win = 0; 553 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 554 win = (long)TCP_MAXWIN << tp->rcv_scale; 555 } 556 /* 557 * Don't use the route cache of a listen socket, 558 * it is not MPSAFE; use temporary route cache. 559 */ 560 if (tp->t_state != TCPS_LISTEN) { 561 if (isipv6) 562 ro6 = &tp->t_inpcb->in6p_route; 563 else 564 ro = &tp->t_inpcb->inp_route; 565 use_tmpro = FALSE; 566 } 567 } 568 if (use_tmpro) { 569 if (isipv6) { 570 ro6 = &sro6; 571 bzero(ro6, sizeof *ro6); 572 } else { 573 ro = &sro; 574 bzero(ro, sizeof *ro); 575 } 576 } 577 if (m == NULL) { 578 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 579 if (m == NULL) 580 return; 581 tlen = 0; 582 m->m_data += max_linkhdr; 583 if (isipv6) { 584 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); 585 ip6 = mtod(m, struct ip6_hdr *); 586 nth = (struct tcphdr *)(ip6 + 1); 587 } else { 588 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip)); 589 ip = mtod(m, struct ip *); 590 nth = (struct tcphdr *)(ip + 1); 591 } 592 bcopy(th, nth, sizeof(struct tcphdr)); 593 flags = TH_ACK; 594 } else { 595 m_freem(m->m_next); 596 m->m_next = NULL; 597 m->m_data = (caddr_t)ipgen; 598 /* m_len is set later */ 599 tlen = 0; 600 #define xchg(a, b, type) { type t; t = a; a = b; b = t; } 601 if (isipv6) { 602 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); 603 nth = (struct tcphdr *)(ip6 + 1); 604 } else { 605 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long); 606 nth = (struct tcphdr *)(ip + 1); 607 } 608 if (th != nth) { 609 /* 610 * this is usually a case when an extension header 611 * exists between the IPv6 header and the 612 * TCP header. 613 */ 614 nth->th_sport = th->th_sport; 615 nth->th_dport = th->th_dport; 616 } 617 xchg(nth->th_dport, nth->th_sport, n_short); 618 #undef xchg 619 } 620 if (isipv6) { 621 ip6->ip6_flow = 0; 622 ip6->ip6_vfc = IPV6_VERSION; 623 ip6->ip6_nxt = IPPROTO_TCP; 624 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen)); 625 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 626 } else { 627 tlen += sizeof(struct tcpiphdr); 628 ip->ip_len = tlen; 629 ip->ip_ttl = ip_defttl; 630 } 631 m->m_len = tlen; 632 m->m_pkthdr.len = tlen; 633 m->m_pkthdr.rcvif = NULL; 634 nth->th_seq = htonl(seq); 635 nth->th_ack = htonl(ack); 636 nth->th_x2 = 0; 637 nth->th_off = sizeof(struct tcphdr) >> 2; 638 nth->th_flags = flags; 639 if (tp != NULL) 640 nth->th_win = htons((u_short) (win >> tp->rcv_scale)); 641 else 642 nth->th_win = htons((u_short)win); 643 nth->th_urp = 0; 644 if (isipv6) { 645 nth->th_sum = 0; 646 nth->th_sum = in6_cksum(m, IPPROTO_TCP, 647 sizeof(struct ip6_hdr), 648 tlen - sizeof(struct ip6_hdr)); 649 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL, 650 (ro6 && ro6->ro_rt) ? 651 ro6->ro_rt->rt_ifp : NULL); 652 } else { 653 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 654 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); 655 m->m_pkthdr.csum_flags = CSUM_TCP; 656 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 657 m->m_pkthdr.csum_thlen = sizeof(struct tcphdr); 658 } 659 #ifdef TCPDEBUG 660 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) 661 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); 662 #endif 663 if (isipv6) { 664 ip6_output(m, NULL, ro6, ipflags, NULL, NULL, 665 tp ? tp->t_inpcb : NULL); 666 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) { 667 RTFREE(ro6->ro_rt); 668 ro6->ro_rt = NULL; 669 } 670 } else { 671 ipflags |= IP_DEBUGROUTE; 672 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL); 673 if ((ro == &sro) && (ro->ro_rt != NULL)) { 674 RTFREE(ro->ro_rt); 675 ro->ro_rt = NULL; 676 } 677 } 678 } 679 680 /* 681 * Create a new TCP control block, making an 682 * empty reassembly queue and hooking it to the argument 683 * protocol control block. The `inp' parameter must have 684 * come from the zone allocator set up in tcp_init(). 685 */ 686 struct tcpcb * 687 tcp_newtcpcb(struct inpcb *inp) 688 { 689 struct inp_tp *it; 690 struct tcpcb *tp; 691 #ifdef INET6 692 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 693 #else 694 const boolean_t isipv6 = FALSE; 695 #endif 696 697 it = (struct inp_tp *)inp; 698 tp = &it->tcb; 699 bzero(tp, sizeof(struct tcpcb)); 700 TAILQ_INIT(&tp->t_segq); 701 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 702 tp->t_rxtthresh = tcprexmtthresh; 703 704 /* Set up our timeouts. */ 705 tp->tt_rexmt = &it->inp_tp_rexmt; 706 tp->tt_persist = &it->inp_tp_persist; 707 tp->tt_keep = &it->inp_tp_keep; 708 tp->tt_2msl = &it->inp_tp_2msl; 709 tp->tt_delack = &it->inp_tp_delack; 710 tcp_inittimers(tp); 711 712 /* 713 * Zero out timer message. We don't create it here, 714 * since the current CPU may not be the owner of this 715 * inpcb. 716 */ 717 tp->tt_msg = &it->inp_tp_timermsg; 718 bzero(tp->tt_msg, sizeof(*tp->tt_msg)); 719 720 tp->t_keepinit = tcp_keepinit; 721 tp->t_keepidle = tcp_keepidle; 722 tp->t_keepintvl = tcp_keepintvl; 723 tp->t_keepcnt = tcp_keepcnt; 724 tp->t_maxidle = tp->t_keepintvl * tp->t_keepcnt; 725 726 if (tcp_do_ncr) 727 tp->t_flags |= TF_NCR; 728 if (tcp_do_rfc1323) 729 tp->t_flags |= (TF_REQ_SCALE | TF_REQ_TSTMP); 730 731 tp->t_inpcb = inp; /* XXX */ 732 tp->t_state = TCPS_CLOSED; 733 /* 734 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 735 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 736 * reasonable initial retransmit time. 737 */ 738 tp->t_srtt = TCPTV_SRTTBASE; 739 tp->t_rttvar = 740 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 741 tp->t_rttmin = tcp_rexmit_min; 742 tp->t_rxtcur = TCPTV_RTOBASE; 743 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 744 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 745 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 746 tp->snd_last = ticks; 747 tp->t_rcvtime = ticks; 748 /* 749 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 750 * because the socket may be bound to an IPv6 wildcard address, 751 * which may match an IPv4-mapped IPv6 address. 752 */ 753 inp->inp_ip_ttl = ip_defttl; 754 inp->inp_ppcb = tp; 755 tcp_sack_tcpcb_init(tp); 756 757 tp->tt_sndmore = &it->inp_tp_sndmore; 758 tcp_output_init(tp); 759 760 return (tp); /* XXX */ 761 } 762 763 /* 764 * Drop a TCP connection, reporting the specified error. 765 * If connection is synchronized, then send a RST to peer. 766 */ 767 struct tcpcb * 768 tcp_drop(struct tcpcb *tp, int error) 769 { 770 struct socket *so = tp->t_inpcb->inp_socket; 771 772 if (TCPS_HAVERCVDSYN(tp->t_state)) { 773 tp->t_state = TCPS_CLOSED; 774 tcp_output(tp); 775 tcpstat.tcps_drops++; 776 } else 777 tcpstat.tcps_conndrops++; 778 if (error == ETIMEDOUT && tp->t_softerror) 779 error = tp->t_softerror; 780 so->so_error = error; 781 return (tcp_close(tp)); 782 } 783 784 struct netmsg_listen_detach { 785 struct netmsg_base base; 786 struct tcpcb *nm_tp; 787 struct tcpcb *nm_tp_inh; 788 }; 789 790 static void 791 tcp_listen_detach_handler(netmsg_t msg) 792 { 793 struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg; 794 struct tcpcb *tp = nmsg->nm_tp; 795 int cpu = mycpuid, nextcpu; 796 797 if (tp->t_flags & TF_LISTEN) 798 syncache_destroy(tp, nmsg->nm_tp_inh); 799 800 in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]); 801 802 nextcpu = cpu + 1; 803 if (nextcpu < ncpus2) 804 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nmsg->base.lmsg); 805 else 806 lwkt_replymsg(&nmsg->base.lmsg, 0); 807 } 808 809 /* 810 * Close a TCP control block: 811 * discard all space held by the tcp 812 * discard internet protocol block 813 * wake up any sleepers 814 */ 815 struct tcpcb * 816 tcp_close(struct tcpcb *tp) 817 { 818 struct tseg_qent *q; 819 struct inpcb *inp = tp->t_inpcb; 820 struct inpcb *inp_inh = NULL; 821 struct tcpcb *tp_inh = NULL; 822 struct socket *so = inp->inp_socket; 823 struct rtentry *rt; 824 boolean_t dosavessthresh; 825 #ifdef INET6 826 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 827 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0); 828 #else 829 const boolean_t isipv6 = FALSE; 830 #endif 831 832 if (tp->t_flags & TF_LISTEN) { 833 /* 834 * Pending socket/syncache inheritance 835 * 836 * If this is a listen(2) socket, find another listen(2) 837 * socket in the same local group, which could inherit 838 * the syncache and sockets pending on the completion 839 * and incompletion queues. 840 * 841 * NOTE: 842 * Currently the inheritance could only happen on the 843 * listen(2) sockets w/ SO_REUSEPORT set. 844 */ 845 KASSERT(&curthread->td_msgport == netisr_cpuport(0), 846 ("listen socket close not in netisr0")); 847 inp_inh = in_pcblocalgroup_last(&tcbinfo[0], inp); 848 if (inp_inh != NULL) 849 tp_inh = intotcpcb(inp_inh); 850 } 851 852 /* 853 * INP_WILDCARD_MP indicates that listen(2) has been called on 854 * this socket. This implies: 855 * - A wildcard inp's hash is replicated for each protocol thread. 856 * - Syncache for this inp grows independently in each protocol 857 * thread. 858 * - There is more than one cpu 859 * 860 * We have to chain a message to the rest of the protocol threads 861 * to cleanup the wildcard hash and the syncache. The cleanup 862 * in the current protocol thread is defered till the end of this 863 * function. 864 * 865 * NOTE: 866 * After cleanup the inp's hash and syncache entries, this inp will 867 * no longer be available to the rest of the protocol threads, so we 868 * are safe to whack the inp in the following code. 869 */ 870 if (inp->inp_flags & INP_WILDCARD_MP) { 871 struct netmsg_listen_detach nmsg; 872 873 KKASSERT(so->so_port == netisr_cpuport(0)); 874 KKASSERT(&curthread->td_msgport == netisr_cpuport(0)); 875 KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]); 876 877 netmsg_init(&nmsg.base, NULL, &curthread->td_msgport, 878 MSGF_PRIORITY, tcp_listen_detach_handler); 879 nmsg.nm_tp = tp; 880 nmsg.nm_tp_inh = tp_inh; 881 lwkt_domsg(netisr_cpuport(1), &nmsg.base.lmsg, 0); 882 883 inp->inp_flags &= ~INP_WILDCARD_MP; 884 } 885 886 KKASSERT(tp->t_state != TCPS_TERMINATING); 887 tp->t_state = TCPS_TERMINATING; 888 889 /* 890 * Make sure that all of our timers are stopped before we 891 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL), 892 * timers are never used. If timer message is never created 893 * (tp->tt_msg->tt_tcb == NULL), timers are never used too. 894 */ 895 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) { 896 tcp_callout_stop(tp, tp->tt_rexmt); 897 tcp_callout_stop(tp, tp->tt_persist); 898 tcp_callout_stop(tp, tp->tt_keep); 899 tcp_callout_stop(tp, tp->tt_2msl); 900 tcp_callout_stop(tp, tp->tt_delack); 901 } 902 903 if (tp->t_flags & TF_ONOUTPUTQ) { 904 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 905 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 906 tp->t_flags &= ~TF_ONOUTPUTQ; 907 } 908 909 /* 910 * If we got enough samples through the srtt filter, 911 * save the rtt and rttvar in the routing entry. 912 * 'Enough' is arbitrarily defined as the 16 samples. 913 * 16 samples is enough for the srtt filter to converge 914 * to within 5% of the correct value; fewer samples and 915 * we could save a very bogus rtt. 916 * 917 * Don't update the default route's characteristics and don't 918 * update anything that the user "locked". 919 */ 920 if (tp->t_rttupdated >= 16) { 921 u_long i = 0; 922 923 if (isipv6) { 924 struct sockaddr_in6 *sin6; 925 926 if ((rt = inp->in6p_route.ro_rt) == NULL) 927 goto no_valid_rt; 928 sin6 = (struct sockaddr_in6 *)rt_key(rt); 929 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 930 goto no_valid_rt; 931 } else 932 if ((rt = inp->inp_route.ro_rt) == NULL || 933 ((struct sockaddr_in *)rt_key(rt))-> 934 sin_addr.s_addr == INADDR_ANY) 935 goto no_valid_rt; 936 937 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 938 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 939 if (rt->rt_rmx.rmx_rtt && i) 940 /* 941 * filter this update to half the old & half 942 * the new values, converting scale. 943 * See route.h and tcp_var.h for a 944 * description of the scaling constants. 945 */ 946 rt->rt_rmx.rmx_rtt = 947 (rt->rt_rmx.rmx_rtt + i) / 2; 948 else 949 rt->rt_rmx.rmx_rtt = i; 950 tcpstat.tcps_cachedrtt++; 951 } 952 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 953 i = tp->t_rttvar * 954 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 955 if (rt->rt_rmx.rmx_rttvar && i) 956 rt->rt_rmx.rmx_rttvar = 957 (rt->rt_rmx.rmx_rttvar + i) / 2; 958 else 959 rt->rt_rmx.rmx_rttvar = i; 960 tcpstat.tcps_cachedrttvar++; 961 } 962 /* 963 * The old comment here said: 964 * update the pipelimit (ssthresh) if it has been updated 965 * already or if a pipesize was specified & the threshhold 966 * got below half the pipesize. I.e., wait for bad news 967 * before we start updating, then update on both good 968 * and bad news. 969 * 970 * But we want to save the ssthresh even if no pipesize is 971 * specified explicitly in the route, because such 972 * connections still have an implicit pipesize specified 973 * by the global tcp_sendspace. In the absence of a reliable 974 * way to calculate the pipesize, it will have to do. 975 */ 976 i = tp->snd_ssthresh; 977 if (rt->rt_rmx.rmx_sendpipe != 0) 978 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 979 else 980 dosavessthresh = (i < so->so_snd.ssb_hiwat/2); 981 if (dosavessthresh || 982 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 983 (rt->rt_rmx.rmx_ssthresh != 0))) { 984 /* 985 * convert the limit from user data bytes to 986 * packets then to packet data bytes. 987 */ 988 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 989 if (i < 2) 990 i = 2; 991 i *= tp->t_maxseg + 992 (isipv6 ? 993 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 994 sizeof(struct tcpiphdr)); 995 if (rt->rt_rmx.rmx_ssthresh) 996 rt->rt_rmx.rmx_ssthresh = 997 (rt->rt_rmx.rmx_ssthresh + i) / 2; 998 else 999 rt->rt_rmx.rmx_ssthresh = i; 1000 tcpstat.tcps_cachedssthresh++; 1001 } 1002 } 1003 1004 no_valid_rt: 1005 /* free the reassembly queue, if any */ 1006 while((q = TAILQ_FIRST(&tp->t_segq)) != NULL) { 1007 TAILQ_REMOVE(&tp->t_segq, q, tqe_q); 1008 m_freem(q->tqe_m); 1009 kfree(q, M_TSEGQ); 1010 atomic_add_int(&tcp_reass_qsize, -1); 1011 } 1012 /* throw away SACK blocks in scoreboard*/ 1013 if (TCP_DO_SACK(tp)) 1014 tcp_sack_destroy(&tp->scb); 1015 1016 inp->inp_ppcb = NULL; 1017 soisdisconnected(so); 1018 /* note: pcb detached later on */ 1019 1020 tcp_destroy_timermsg(tp); 1021 tcp_output_cancel(tp); 1022 1023 if (tp->t_flags & TF_LISTEN) { 1024 syncache_destroy(tp, tp_inh); 1025 if (inp_inh != NULL && inp_inh->inp_socket != NULL) { 1026 /* 1027 * Pending sockets inheritance only needs 1028 * to be done once in the current thread, 1029 * i.e. netisr0. 1030 */ 1031 soinherit(so, inp_inh->inp_socket); 1032 } 1033 } 1034 1035 so_async_rcvd_drop(so); 1036 /* Drop the reference for the asynchronized pru_rcvd */ 1037 sofree(so); 1038 1039 /* 1040 * NOTE: 1041 * pcbdetach removes any wildcard hash entry on the current CPU. 1042 */ 1043 #ifdef INET6 1044 if (isafinet6) 1045 in6_pcbdetach(inp); 1046 else 1047 #endif 1048 in_pcbdetach(inp); 1049 1050 tcpstat.tcps_closed++; 1051 return (NULL); 1052 } 1053 1054 static __inline void 1055 tcp_drain_oncpu(struct inpcbhead *head) 1056 { 1057 struct inpcb *marker; 1058 struct inpcb *inpb; 1059 struct tcpcb *tcpb; 1060 struct tseg_qent *te; 1061 1062 /* 1063 * Allows us to block while running the list 1064 */ 1065 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1066 marker->inp_flags |= INP_PLACEMARKER; 1067 LIST_INSERT_HEAD(head, marker, inp_list); 1068 1069 while ((inpb = LIST_NEXT(marker, inp_list)) != NULL) { 1070 if ((inpb->inp_flags & INP_PLACEMARKER) == 0 && 1071 (tcpb = intotcpcb(inpb)) != NULL && 1072 (te = TAILQ_FIRST(&tcpb->t_segq)) != NULL) { 1073 TAILQ_REMOVE(&tcpb->t_segq, te, tqe_q); 1074 if (te->tqe_th->th_flags & TH_FIN) 1075 tcpb->t_flags &= ~TF_QUEDFIN; 1076 m_freem(te->tqe_m); 1077 kfree(te, M_TSEGQ); 1078 atomic_add_int(&tcp_reass_qsize, -1); 1079 /* retry */ 1080 } else { 1081 LIST_REMOVE(marker, inp_list); 1082 LIST_INSERT_AFTER(inpb, marker, inp_list); 1083 } 1084 } 1085 LIST_REMOVE(marker, inp_list); 1086 kfree(marker, M_TEMP); 1087 } 1088 1089 struct netmsg_tcp_drain { 1090 struct netmsg_base base; 1091 struct inpcbhead *nm_head; 1092 }; 1093 1094 static void 1095 tcp_drain_handler(netmsg_t msg) 1096 { 1097 struct netmsg_tcp_drain *nm = (void *)msg; 1098 1099 tcp_drain_oncpu(nm->nm_head); 1100 lwkt_replymsg(&nm->base.lmsg, 0); 1101 } 1102 1103 void 1104 tcp_drain(void) 1105 { 1106 int cpu; 1107 1108 if (!do_tcpdrain) 1109 return; 1110 1111 /* 1112 * Walk the tcpbs, if existing, and flush the reassembly queue, 1113 * if there is one... 1114 * XXX: The "Net/3" implementation doesn't imply that the TCP 1115 * reassembly queue should be flushed, but in a situation 1116 * where we're really low on mbufs, this is potentially 1117 * useful. 1118 */ 1119 for (cpu = 0; cpu < ncpus2; cpu++) { 1120 struct netmsg_tcp_drain *nm; 1121 1122 if (cpu == mycpu->gd_cpuid) { 1123 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead); 1124 } else { 1125 nm = kmalloc(sizeof(struct netmsg_tcp_drain), 1126 M_LWKTMSG, M_NOWAIT); 1127 if (nm == NULL) 1128 continue; 1129 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1130 0, tcp_drain_handler); 1131 nm->nm_head = &tcbinfo[cpu].pcblisthead; 1132 lwkt_sendmsg(netisr_cpuport(cpu), &nm->base.lmsg); 1133 } 1134 } 1135 } 1136 1137 /* 1138 * Notify a tcp user of an asynchronous error; 1139 * store error as soft error, but wake up user 1140 * (for now, won't do anything until can select for soft error). 1141 * 1142 * Do not wake up user since there currently is no mechanism for 1143 * reporting soft errors (yet - a kqueue filter may be added). 1144 */ 1145 static void 1146 tcp_notify(struct inpcb *inp, int error) 1147 { 1148 struct tcpcb *tp = intotcpcb(inp); 1149 1150 /* 1151 * Ignore some errors if we are hooked up. 1152 * If connection hasn't completed, has retransmitted several times, 1153 * and receives a second error, give up now. This is better 1154 * than waiting a long time to establish a connection that 1155 * can never complete. 1156 */ 1157 if (tp->t_state == TCPS_ESTABLISHED && 1158 (error == EHOSTUNREACH || error == ENETUNREACH || 1159 error == EHOSTDOWN)) { 1160 return; 1161 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1162 tp->t_softerror) 1163 tcp_drop(tp, error); 1164 else 1165 tp->t_softerror = error; 1166 #if 0 1167 wakeup(&so->so_timeo); 1168 sorwakeup(so); 1169 sowwakeup(so); 1170 #endif 1171 } 1172 1173 static int 1174 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1175 { 1176 int error, i, n; 1177 struct inpcb *marker; 1178 struct inpcb *inp; 1179 globaldata_t gd; 1180 int origcpu, ccpu; 1181 1182 error = 0; 1183 n = 0; 1184 1185 /* 1186 * The process of preparing the TCB list is too time-consuming and 1187 * resource-intensive to repeat twice on every request. 1188 */ 1189 if (req->oldptr == NULL) { 1190 for (ccpu = 0; ccpu < ncpus; ++ccpu) { 1191 gd = globaldata_find(ccpu); 1192 n += tcbinfo[gd->gd_cpuid].ipi_count; 1193 } 1194 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1195 return (0); 1196 } 1197 1198 if (req->newptr != NULL) 1199 return (EPERM); 1200 1201 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1202 marker->inp_flags |= INP_PLACEMARKER; 1203 1204 /* 1205 * OK, now we're committed to doing something. Run the inpcb list 1206 * for each cpu in the system and construct the output. Use a 1207 * list placemarker to deal with list changes occuring during 1208 * copyout blockages (but otherwise depend on being on the correct 1209 * cpu to avoid races). 1210 */ 1211 origcpu = mycpu->gd_cpuid; 1212 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) { 1213 globaldata_t rgd; 1214 caddr_t inp_ppcb; 1215 struct xtcpcb xt; 1216 int cpu_id; 1217 1218 cpu_id = (origcpu + ccpu) % ncpus; 1219 if ((smp_active_mask & CPUMASK(cpu_id)) == 0) 1220 continue; 1221 rgd = globaldata_find(cpu_id); 1222 lwkt_setcpu_self(rgd); 1223 1224 n = tcbinfo[cpu_id].ipi_count; 1225 1226 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list); 1227 i = 0; 1228 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1229 /* 1230 * process a snapshot of pcbs, ignoring placemarkers 1231 * and using our own to allow SYSCTL_OUT to block. 1232 */ 1233 LIST_REMOVE(marker, inp_list); 1234 LIST_INSERT_AFTER(inp, marker, inp_list); 1235 1236 if (inp->inp_flags & INP_PLACEMARKER) 1237 continue; 1238 if (prison_xinpcb(req->td, inp)) 1239 continue; 1240 1241 xt.xt_len = sizeof xt; 1242 bcopy(inp, &xt.xt_inp, sizeof *inp); 1243 inp_ppcb = inp->inp_ppcb; 1244 if (inp_ppcb != NULL) 1245 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1246 else 1247 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1248 if (inp->inp_socket) 1249 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1250 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1251 break; 1252 ++i; 1253 } 1254 LIST_REMOVE(marker, inp_list); 1255 if (error == 0 && i < n) { 1256 bzero(&xt, sizeof xt); 1257 xt.xt_len = sizeof xt; 1258 while (i < n) { 1259 error = SYSCTL_OUT(req, &xt, sizeof xt); 1260 if (error) 1261 break; 1262 ++i; 1263 } 1264 } 1265 } 1266 1267 /* 1268 * Make sure we are on the same cpu we were on originally, since 1269 * higher level callers expect this. Also don't pollute caches with 1270 * migrated userland data by (eventually) returning to userland 1271 * on a different cpu. 1272 */ 1273 lwkt_setcpu_self(globaldata_find(origcpu)); 1274 kfree(marker, M_TEMP); 1275 return (error); 1276 } 1277 1278 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1279 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1280 1281 static int 1282 tcp_getcred(SYSCTL_HANDLER_ARGS) 1283 { 1284 struct sockaddr_in addrs[2]; 1285 struct inpcb *inp; 1286 int cpu; 1287 int error; 1288 1289 error = priv_check(req->td, PRIV_ROOT); 1290 if (error != 0) 1291 return (error); 1292 error = SYSCTL_IN(req, addrs, sizeof addrs); 1293 if (error != 0) 1294 return (error); 1295 crit_enter(); 1296 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1297 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1298 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1299 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1300 if (inp == NULL || inp->inp_socket == NULL) { 1301 error = ENOENT; 1302 goto out; 1303 } 1304 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1305 out: 1306 crit_exit(); 1307 return (error); 1308 } 1309 1310 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1311 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1312 1313 #ifdef INET6 1314 static int 1315 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1316 { 1317 struct sockaddr_in6 addrs[2]; 1318 struct inpcb *inp; 1319 int error; 1320 boolean_t mapped = FALSE; 1321 1322 error = priv_check(req->td, PRIV_ROOT); 1323 if (error != 0) 1324 return (error); 1325 error = SYSCTL_IN(req, addrs, sizeof addrs); 1326 if (error != 0) 1327 return (error); 1328 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { 1329 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) 1330 mapped = TRUE; 1331 else 1332 return (EINVAL); 1333 } 1334 crit_enter(); 1335 if (mapped) { 1336 inp = in_pcblookup_hash(&tcbinfo[0], 1337 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], 1338 addrs[1].sin6_port, 1339 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], 1340 addrs[0].sin6_port, 1341 0, NULL); 1342 } else { 1343 inp = in6_pcblookup_hash(&tcbinfo[0], 1344 &addrs[1].sin6_addr, addrs[1].sin6_port, 1345 &addrs[0].sin6_addr, addrs[0].sin6_port, 1346 0, NULL); 1347 } 1348 if (inp == NULL || inp->inp_socket == NULL) { 1349 error = ENOENT; 1350 goto out; 1351 } 1352 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1353 out: 1354 crit_exit(); 1355 return (error); 1356 } 1357 1358 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1359 0, 0, 1360 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1361 #endif 1362 1363 struct netmsg_tcp_notify { 1364 struct netmsg_base base; 1365 void (*nm_notify)(struct inpcb *, int); 1366 struct in_addr nm_faddr; 1367 int nm_arg; 1368 }; 1369 1370 static void 1371 tcp_notifyall_oncpu(netmsg_t msg) 1372 { 1373 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg; 1374 int nextcpu; 1375 1376 in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nm->nm_faddr, 1377 nm->nm_arg, nm->nm_notify); 1378 1379 nextcpu = mycpuid + 1; 1380 if (nextcpu < ncpus2) 1381 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nm->base.lmsg); 1382 else 1383 lwkt_replymsg(&nm->base.lmsg, 0); 1384 } 1385 1386 void 1387 tcp_ctlinput(netmsg_t msg) 1388 { 1389 int cmd = msg->ctlinput.nm_cmd; 1390 struct sockaddr *sa = msg->ctlinput.nm_arg; 1391 struct ip *ip = msg->ctlinput.nm_extra; 1392 struct tcphdr *th; 1393 struct in_addr faddr; 1394 struct inpcb *inp; 1395 struct tcpcb *tp; 1396 void (*notify)(struct inpcb *, int) = tcp_notify; 1397 tcp_seq icmpseq; 1398 int arg, cpu; 1399 1400 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) { 1401 goto done; 1402 } 1403 1404 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1405 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1406 goto done; 1407 1408 arg = inetctlerrmap[cmd]; 1409 if (cmd == PRC_QUENCH) { 1410 notify = tcp_quench; 1411 } else if (icmp_may_rst && 1412 (cmd == PRC_UNREACH_ADMIN_PROHIB || 1413 cmd == PRC_UNREACH_PORT || 1414 cmd == PRC_TIMXCEED_INTRANS) && 1415 ip != NULL) { 1416 notify = tcp_drop_syn_sent; 1417 } else if (cmd == PRC_MSGSIZE) { 1418 struct icmp *icmp = (struct icmp *) 1419 ((caddr_t)ip - offsetof(struct icmp, icmp_ip)); 1420 1421 arg = ntohs(icmp->icmp_nextmtu); 1422 notify = tcp_mtudisc; 1423 } else if (PRC_IS_REDIRECT(cmd)) { 1424 ip = NULL; 1425 notify = in_rtchange; 1426 } else if (cmd == PRC_HOSTDEAD) { 1427 ip = NULL; 1428 } 1429 1430 if (ip != NULL) { 1431 crit_enter(); 1432 th = (struct tcphdr *)((caddr_t)ip + 1433 (IP_VHL_HL(ip->ip_vhl) << 2)); 1434 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport, 1435 ip->ip_src.s_addr, th->th_sport); 1436 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport, 1437 ip->ip_src, th->th_sport, 0, NULL); 1438 if ((inp != NULL) && (inp->inp_socket != NULL)) { 1439 icmpseq = htonl(th->th_seq); 1440 tp = intotcpcb(inp); 1441 if (SEQ_GEQ(icmpseq, tp->snd_una) && 1442 SEQ_LT(icmpseq, tp->snd_max)) 1443 (*notify)(inp, arg); 1444 } else { 1445 struct in_conninfo inc; 1446 1447 inc.inc_fport = th->th_dport; 1448 inc.inc_lport = th->th_sport; 1449 inc.inc_faddr = faddr; 1450 inc.inc_laddr = ip->ip_src; 1451 #ifdef INET6 1452 inc.inc_isipv6 = 0; 1453 #endif 1454 syncache_unreach(&inc, th); 1455 } 1456 crit_exit(); 1457 } else { 1458 struct netmsg_tcp_notify *nm; 1459 1460 KKASSERT(&curthread->td_msgport == netisr_cpuport(0)); 1461 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT); 1462 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1463 0, tcp_notifyall_oncpu); 1464 nm->nm_faddr = faddr; 1465 nm->nm_arg = arg; 1466 nm->nm_notify = notify; 1467 1468 lwkt_sendmsg(netisr_cpuport(0), &nm->base.lmsg); 1469 } 1470 done: 1471 lwkt_replymsg(&msg->lmsg, 0); 1472 } 1473 1474 #ifdef INET6 1475 1476 void 1477 tcp6_ctlinput(netmsg_t msg) 1478 { 1479 int cmd = msg->ctlinput.nm_cmd; 1480 struct sockaddr *sa = msg->ctlinput.nm_arg; 1481 void *d = msg->ctlinput.nm_extra; 1482 struct tcphdr th; 1483 void (*notify) (struct inpcb *, int) = tcp_notify; 1484 struct ip6_hdr *ip6; 1485 struct mbuf *m; 1486 struct ip6ctlparam *ip6cp = NULL; 1487 const struct sockaddr_in6 *sa6_src = NULL; 1488 int off; 1489 struct tcp_portonly { 1490 u_int16_t th_sport; 1491 u_int16_t th_dport; 1492 } *thp; 1493 int arg; 1494 1495 if (sa->sa_family != AF_INET6 || 1496 sa->sa_len != sizeof(struct sockaddr_in6)) { 1497 goto out; 1498 } 1499 1500 arg = 0; 1501 if (cmd == PRC_QUENCH) 1502 notify = tcp_quench; 1503 else if (cmd == PRC_MSGSIZE) { 1504 struct ip6ctlparam *ip6cp = d; 1505 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6; 1506 1507 arg = ntohl(icmp6->icmp6_mtu); 1508 notify = tcp_mtudisc; 1509 } else if (!PRC_IS_REDIRECT(cmd) && 1510 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) { 1511 goto out; 1512 } 1513 1514 /* if the parameter is from icmp6, decode it. */ 1515 if (d != NULL) { 1516 ip6cp = (struct ip6ctlparam *)d; 1517 m = ip6cp->ip6c_m; 1518 ip6 = ip6cp->ip6c_ip6; 1519 off = ip6cp->ip6c_off; 1520 sa6_src = ip6cp->ip6c_src; 1521 } else { 1522 m = NULL; 1523 ip6 = NULL; 1524 off = 0; /* fool gcc */ 1525 sa6_src = &sa6_any; 1526 } 1527 1528 if (ip6 != NULL) { 1529 struct in_conninfo inc; 1530 /* 1531 * XXX: We assume that when IPV6 is non NULL, 1532 * M and OFF are valid. 1533 */ 1534 1535 /* check if we can safely examine src and dst ports */ 1536 if (m->m_pkthdr.len < off + sizeof *thp) 1537 goto out; 1538 1539 bzero(&th, sizeof th); 1540 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1541 1542 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport, 1543 (struct sockaddr *)ip6cp->ip6c_src, 1544 th.th_sport, cmd, arg, notify); 1545 1546 inc.inc_fport = th.th_dport; 1547 inc.inc_lport = th.th_sport; 1548 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1549 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1550 inc.inc_isipv6 = 1; 1551 syncache_unreach(&inc, &th); 1552 } else { 1553 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0, 1554 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify); 1555 } 1556 out: 1557 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0); 1558 } 1559 1560 #endif 1561 1562 /* 1563 * Following is where TCP initial sequence number generation occurs. 1564 * 1565 * There are two places where we must use initial sequence numbers: 1566 * 1. In SYN-ACK packets. 1567 * 2. In SYN packets. 1568 * 1569 * All ISNs for SYN-ACK packets are generated by the syncache. See 1570 * tcp_syncache.c for details. 1571 * 1572 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1573 * depends on this property. In addition, these ISNs should be 1574 * unguessable so as to prevent connection hijacking. To satisfy 1575 * the requirements of this situation, the algorithm outlined in 1576 * RFC 1948 is used to generate sequence numbers. 1577 * 1578 * Implementation details: 1579 * 1580 * Time is based off the system timer, and is corrected so that it 1581 * increases by one megabyte per second. This allows for proper 1582 * recycling on high speed LANs while still leaving over an hour 1583 * before rollover. 1584 * 1585 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1586 * between seeding of isn_secret. This is normally set to zero, 1587 * as reseeding should not be necessary. 1588 * 1589 */ 1590 1591 #define ISN_BYTES_PER_SECOND 1048576 1592 1593 u_char isn_secret[32]; 1594 int isn_last_reseed; 1595 MD5_CTX isn_ctx; 1596 1597 tcp_seq 1598 tcp_new_isn(struct tcpcb *tp) 1599 { 1600 u_int32_t md5_buffer[4]; 1601 tcp_seq new_isn; 1602 1603 /* Seed if this is the first use, reseed if requested. */ 1604 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1605 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1606 < (u_int)ticks))) { 1607 read_random_unlimited(&isn_secret, sizeof isn_secret); 1608 isn_last_reseed = ticks; 1609 } 1610 1611 /* Compute the md5 hash and return the ISN. */ 1612 MD5Init(&isn_ctx); 1613 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1614 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1615 #ifdef INET6 1616 if (tp->t_inpcb->inp_vflag & INP_IPV6) { 1617 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1618 sizeof(struct in6_addr)); 1619 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1620 sizeof(struct in6_addr)); 1621 } else 1622 #endif 1623 { 1624 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1625 sizeof(struct in_addr)); 1626 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1627 sizeof(struct in_addr)); 1628 } 1629 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1630 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1631 new_isn = (tcp_seq) md5_buffer[0]; 1632 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1633 return (new_isn); 1634 } 1635 1636 /* 1637 * When a source quench is received, close congestion window 1638 * to one segment. We will gradually open it again as we proceed. 1639 */ 1640 void 1641 tcp_quench(struct inpcb *inp, int error) 1642 { 1643 struct tcpcb *tp = intotcpcb(inp); 1644 1645 if (tp != NULL) { 1646 tp->snd_cwnd = tp->t_maxseg; 1647 tp->snd_wacked = 0; 1648 } 1649 } 1650 1651 /* 1652 * When a specific ICMP unreachable message is received and the 1653 * connection state is SYN-SENT, drop the connection. This behavior 1654 * is controlled by the icmp_may_rst sysctl. 1655 */ 1656 void 1657 tcp_drop_syn_sent(struct inpcb *inp, int error) 1658 { 1659 struct tcpcb *tp = intotcpcb(inp); 1660 1661 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT)) 1662 tcp_drop(tp, error); 1663 } 1664 1665 /* 1666 * When a `need fragmentation' ICMP is received, update our idea of the MSS 1667 * based on the new value in the route. Also nudge TCP to send something, 1668 * since we know the packet we just sent was dropped. 1669 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1670 */ 1671 void 1672 tcp_mtudisc(struct inpcb *inp, int mtu) 1673 { 1674 struct tcpcb *tp = intotcpcb(inp); 1675 struct rtentry *rt; 1676 struct socket *so = inp->inp_socket; 1677 int maxopd, mss; 1678 #ifdef INET6 1679 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0); 1680 #else 1681 const boolean_t isipv6 = FALSE; 1682 #endif 1683 1684 if (tp == NULL) 1685 return; 1686 1687 /* 1688 * If no MTU is provided in the ICMP message, use the 1689 * next lower likely value, as specified in RFC 1191. 1690 */ 1691 if (mtu == 0) { 1692 int oldmtu; 1693 1694 oldmtu = tp->t_maxopd + 1695 (isipv6 ? 1696 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1697 sizeof(struct tcpiphdr)); 1698 mtu = ip_next_mtu(oldmtu, 0); 1699 } 1700 1701 if (isipv6) 1702 rt = tcp_rtlookup6(&inp->inp_inc); 1703 else 1704 rt = tcp_rtlookup(&inp->inp_inc); 1705 if (rt != NULL) { 1706 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu) 1707 mtu = rt->rt_rmx.rmx_mtu; 1708 1709 maxopd = mtu - 1710 (isipv6 ? 1711 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1712 sizeof(struct tcpiphdr)); 1713 1714 /* 1715 * XXX - The following conditional probably violates the TCP 1716 * spec. The problem is that, since we don't know the 1717 * other end's MSS, we are supposed to use a conservative 1718 * default. But, if we do that, then MTU discovery will 1719 * never actually take place, because the conservative 1720 * default is much less than the MTUs typically seen 1721 * on the Internet today. For the moment, we'll sweep 1722 * this under the carpet. 1723 * 1724 * The conservative default might not actually be a problem 1725 * if the only case this occurs is when sending an initial 1726 * SYN with options and data to a host we've never talked 1727 * to before. Then, they will reply with an MSS value which 1728 * will get recorded and the new parameters should get 1729 * recomputed. For Further Study. 1730 */ 1731 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd) 1732 maxopd = rt->rt_rmx.rmx_mssopt; 1733 } else 1734 maxopd = mtu - 1735 (isipv6 ? 1736 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1737 sizeof(struct tcpiphdr)); 1738 1739 if (tp->t_maxopd <= maxopd) 1740 return; 1741 tp->t_maxopd = maxopd; 1742 1743 mss = maxopd; 1744 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) == 1745 (TF_REQ_TSTMP | TF_RCVD_TSTMP)) 1746 mss -= TCPOLEN_TSTAMP_APPA; 1747 1748 /* round down to multiple of MCLBYTES */ 1749 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */ 1750 if (mss > MCLBYTES) 1751 mss &= ~(MCLBYTES - 1); 1752 #else 1753 if (mss > MCLBYTES) 1754 mss = (mss / MCLBYTES) * MCLBYTES; 1755 #endif 1756 1757 if (so->so_snd.ssb_hiwat < mss) 1758 mss = so->so_snd.ssb_hiwat; 1759 1760 tp->t_maxseg = mss; 1761 tp->t_rtttime = 0; 1762 tp->snd_nxt = tp->snd_una; 1763 tcp_output(tp); 1764 tcpstat.tcps_mturesent++; 1765 } 1766 1767 /* 1768 * Look-up the routing entry to the peer of this inpcb. If no route 1769 * is found and it cannot be allocated the return NULL. This routine 1770 * is called by TCP routines that access the rmx structure and by tcp_mss 1771 * to get the interface MTU. 1772 */ 1773 struct rtentry * 1774 tcp_rtlookup(struct in_conninfo *inc) 1775 { 1776 struct route *ro = &inc->inc_route; 1777 1778 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1779 /* No route yet, so try to acquire one */ 1780 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1781 /* 1782 * unused portions of the structure MUST be zero'd 1783 * out because rtalloc() treats it as opaque data 1784 */ 1785 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1786 ro->ro_dst.sa_family = AF_INET; 1787 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1788 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1789 inc->inc_faddr; 1790 rtalloc(ro); 1791 } 1792 } 1793 return (ro->ro_rt); 1794 } 1795 1796 #ifdef INET6 1797 struct rtentry * 1798 tcp_rtlookup6(struct in_conninfo *inc) 1799 { 1800 struct route_in6 *ro6 = &inc->inc6_route; 1801 1802 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1803 /* No route yet, so try to acquire one */ 1804 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1805 /* 1806 * unused portions of the structure MUST be zero'd 1807 * out because rtalloc() treats it as opaque data 1808 */ 1809 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1810 ro6->ro_dst.sin6_family = AF_INET6; 1811 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1812 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1813 rtalloc((struct route *)ro6); 1814 } 1815 } 1816 return (ro6->ro_rt); 1817 } 1818 #endif 1819 1820 #ifdef IPSEC 1821 /* compute ESP/AH header size for TCP, including outer IP header. */ 1822 size_t 1823 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1824 { 1825 struct inpcb *inp; 1826 struct mbuf *m; 1827 size_t hdrsiz; 1828 struct ip *ip; 1829 struct tcphdr *th; 1830 1831 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1832 return (0); 1833 MGETHDR(m, MB_DONTWAIT, MT_DATA); 1834 if (!m) 1835 return (0); 1836 1837 #ifdef INET6 1838 if (inp->inp_vflag & INP_IPV6) { 1839 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1840 1841 th = (struct tcphdr *)(ip6 + 1); 1842 m->m_pkthdr.len = m->m_len = 1843 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1844 tcp_fillheaders(tp, ip6, th, FALSE); 1845 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1846 } else 1847 #endif 1848 { 1849 ip = mtod(m, struct ip *); 1850 th = (struct tcphdr *)(ip + 1); 1851 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1852 tcp_fillheaders(tp, ip, th, FALSE); 1853 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1854 } 1855 1856 m_free(m); 1857 return (hdrsiz); 1858 } 1859 #endif 1860 1861 /* 1862 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1863 * 1864 * This code attempts to calculate the bandwidth-delay product as a 1865 * means of determining the optimal window size to maximize bandwidth, 1866 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1867 * routers. This code also does a fairly good job keeping RTTs in check 1868 * across slow links like modems. We implement an algorithm which is very 1869 * similar (but not meant to be) TCP/Vegas. The code operates on the 1870 * transmitter side of a TCP connection and so only effects the transmit 1871 * side of the connection. 1872 * 1873 * BACKGROUND: TCP makes no provision for the management of buffer space 1874 * at the end points or at the intermediate routers and switches. A TCP 1875 * stream, whether using NewReno or not, will eventually buffer as 1876 * many packets as it is able and the only reason this typically works is 1877 * due to the fairly small default buffers made available for a connection 1878 * (typicaly 16K or 32K). As machines use larger windows and/or window 1879 * scaling it is now fairly easy for even a single TCP connection to blow-out 1880 * all available buffer space not only on the local interface, but on 1881 * intermediate routers and switches as well. NewReno makes a misguided 1882 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1883 * then backing off, then steadily increasing the window again until another 1884 * failure occurs, ad-infinitum. This results in terrible oscillation that 1885 * is only made worse as network loads increase and the idea of intentionally 1886 * blowing out network buffers is, frankly, a terrible way to manage network 1887 * resources. 1888 * 1889 * It is far better to limit the transmit window prior to the failure 1890 * condition being achieved. There are two general ways to do this: First 1891 * you can 'scan' through different transmit window sizes and locate the 1892 * point where the RTT stops increasing, indicating that you have filled the 1893 * pipe, then scan backwards until you note that RTT stops decreasing, then 1894 * repeat ad-infinitum. This method works in principle but has severe 1895 * implementation issues due to RTT variances, timer granularity, and 1896 * instability in the algorithm which can lead to many false positives and 1897 * create oscillations as well as interact badly with other TCP streams 1898 * implementing the same algorithm. 1899 * 1900 * The second method is to limit the window to the bandwidth delay product 1901 * of the link. This is the method we implement. RTT variances and our 1902 * own manipulation of the congestion window, bwnd, can potentially 1903 * destabilize the algorithm. For this reason we have to stabilize the 1904 * elements used to calculate the window. We do this by using the minimum 1905 * observed RTT, the long term average of the observed bandwidth, and 1906 * by adding two segments worth of slop. It isn't perfect but it is able 1907 * to react to changing conditions and gives us a very stable basis on 1908 * which to extend the algorithm. 1909 */ 1910 void 1911 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1912 { 1913 u_long bw; 1914 u_long bwnd; 1915 int save_ticks; 1916 int delta_ticks; 1917 1918 /* 1919 * If inflight_enable is disabled in the middle of a tcp connection, 1920 * make sure snd_bwnd is effectively disabled. 1921 */ 1922 if (!tcp_inflight_enable) { 1923 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1924 tp->snd_bandwidth = 0; 1925 return; 1926 } 1927 1928 /* 1929 * Validate the delta time. If a connection is new or has been idle 1930 * a long time we have to reset the bandwidth calculator. 1931 */ 1932 save_ticks = ticks; 1933 delta_ticks = save_ticks - tp->t_bw_rtttime; 1934 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1935 tp->t_bw_rtttime = ticks; 1936 tp->t_bw_rtseq = ack_seq; 1937 if (tp->snd_bandwidth == 0) 1938 tp->snd_bandwidth = tcp_inflight_min; 1939 return; 1940 } 1941 if (delta_ticks == 0) 1942 return; 1943 1944 /* 1945 * Sanity check, plus ignore pure window update acks. 1946 */ 1947 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1948 return; 1949 1950 /* 1951 * Figure out the bandwidth. Due to the tick granularity this 1952 * is a very rough number and it MUST be averaged over a fairly 1953 * long period of time. XXX we need to take into account a link 1954 * that is not using all available bandwidth, but for now our 1955 * slop will ramp us up if this case occurs and the bandwidth later 1956 * increases. 1957 */ 1958 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 1959 tp->t_bw_rtttime = save_ticks; 1960 tp->t_bw_rtseq = ack_seq; 1961 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4; 1962 1963 tp->snd_bandwidth = bw; 1964 1965 /* 1966 * Calculate the semi-static bandwidth delay product, plus two maximal 1967 * segments. The additional slop puts us squarely in the sweet 1968 * spot and also handles the bandwidth run-up case. Without the 1969 * slop we could be locking ourselves into a lower bandwidth. 1970 * 1971 * Situations Handled: 1972 * (1) Prevents over-queueing of packets on LANs, especially on 1973 * high speed LANs, allowing larger TCP buffers to be 1974 * specified, and also does a good job preventing 1975 * over-queueing of packets over choke points like modems 1976 * (at least for the transmit side). 1977 * 1978 * (2) Is able to handle changing network loads (bandwidth 1979 * drops so bwnd drops, bandwidth increases so bwnd 1980 * increases). 1981 * 1982 * (3) Theoretically should stabilize in the face of multiple 1983 * connections implementing the same algorithm (this may need 1984 * a little work). 1985 * 1986 * (4) Stability value (defaults to 20 = 2 maximal packets) can 1987 * be adjusted with a sysctl but typically only needs to be on 1988 * very slow connections. A value no smaller then 5 should 1989 * be used, but only reduce this default if you have no other 1990 * choice. 1991 */ 1992 1993 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2) 1994 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 1995 tcp_inflight_stab * (int)tp->t_maxseg / 10; 1996 #undef USERTT 1997 1998 if (tcp_inflight_debug > 0) { 1999 static int ltime; 2000 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) { 2001 ltime = ticks; 2002 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n", 2003 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd); 2004 } 2005 } 2006 if ((long)bwnd < tcp_inflight_min) 2007 bwnd = tcp_inflight_min; 2008 if (bwnd > tcp_inflight_max) 2009 bwnd = tcp_inflight_max; 2010 if ((long)bwnd < tp->t_maxseg * 2) 2011 bwnd = tp->t_maxseg * 2; 2012 tp->snd_bwnd = bwnd; 2013 } 2014 2015 static void 2016 tcp_rmx_iwsegs(struct tcpcb *tp, u_long *maxsegs, u_long *capsegs) 2017 { 2018 struct rtentry *rt; 2019 struct inpcb *inp = tp->t_inpcb; 2020 #ifdef INET6 2021 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE); 2022 #else 2023 const boolean_t isipv6 = FALSE; 2024 #endif 2025 2026 /* XXX */ 2027 if (tcp_iw_maxsegs < TCP_IW_MAXSEGS_DFLT) 2028 tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 2029 if (tcp_iw_capsegs < TCP_IW_CAPSEGS_DFLT) 2030 tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 2031 2032 if (isipv6) 2033 rt = tcp_rtlookup6(&inp->inp_inc); 2034 else 2035 rt = tcp_rtlookup(&inp->inp_inc); 2036 if (rt == NULL || 2037 rt->rt_rmx.rmx_iwmaxsegs < TCP_IW_MAXSEGS_DFLT || 2038 rt->rt_rmx.rmx_iwcapsegs < TCP_IW_CAPSEGS_DFLT) { 2039 *maxsegs = tcp_iw_maxsegs; 2040 *capsegs = tcp_iw_capsegs; 2041 return; 2042 } 2043 *maxsegs = rt->rt_rmx.rmx_iwmaxsegs; 2044 *capsegs = rt->rt_rmx.rmx_iwcapsegs; 2045 } 2046 2047 u_long 2048 tcp_initial_window(struct tcpcb *tp) 2049 { 2050 if (tcp_do_rfc3390) { 2051 /* 2052 * RFC3390: 2053 * "If the SYN or SYN/ACK is lost, the initial window 2054 * used by a sender after a correctly transmitted SYN 2055 * MUST be one segment consisting of MSS bytes." 2056 * 2057 * However, we do something a little bit more aggressive 2058 * then RFC3390 here: 2059 * - Only if time spent in the SYN or SYN|ACK retransmition 2060 * >= 3 seconds, the IW is reduced. We do this mainly 2061 * because when RFC3390 is published, the initial RTO is 2062 * still 3 seconds (the threshold we test here), while 2063 * after RFC6298, the initial RTO is 1 second. This 2064 * behaviour probably still falls within the spirit of 2065 * RFC3390. 2066 * - When IW is reduced, 2*MSS is used instead of 1*MSS. 2067 * Mainly to avoid sender and receiver deadlock until 2068 * delayed ACK timer expires. And even RFC2581 does not 2069 * try to reduce IW upon SYN or SYN|ACK retransmition 2070 * timeout. 2071 * 2072 * See also: 2073 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03 2074 */ 2075 if (tp->t_rxtsyn >= TCPTV_RTOBASE3) { 2076 return (2 * tp->t_maxseg); 2077 } else { 2078 u_long maxsegs, capsegs; 2079 2080 tcp_rmx_iwsegs(tp, &maxsegs, &capsegs); 2081 return min(maxsegs * tp->t_maxseg, 2082 max(2 * tp->t_maxseg, capsegs * 1460)); 2083 } 2084 } else { 2085 /* 2086 * Even RFC2581 (back to 1999) allows 2*SMSS IW. 2087 * 2088 * Mainly to avoid sender and receiver deadlock 2089 * until delayed ACK timer expires. 2090 */ 2091 return (2 * tp->t_maxseg); 2092 } 2093 } 2094 2095 #ifdef TCP_SIGNATURE 2096 /* 2097 * Compute TCP-MD5 hash of a TCP segment. (RFC2385) 2098 * 2099 * We do this over ip, tcphdr, segment data, and the key in the SADB. 2100 * When called from tcp_input(), we can be sure that th_sum has been 2101 * zeroed out and verified already. 2102 * 2103 * Return 0 if successful, otherwise return -1. 2104 * 2105 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a 2106 * search with the destination IP address, and a 'magic SPI' to be 2107 * determined by the application. This is hardcoded elsewhere to 1179 2108 * right now. Another branch of this code exists which uses the SPD to 2109 * specify per-application flows but it is unstable. 2110 */ 2111 int 2112 tcpsignature_compute( 2113 struct mbuf *m, /* mbuf chain */ 2114 int len, /* length of TCP data */ 2115 int optlen, /* length of TCP options */ 2116 u_char *buf, /* storage for MD5 digest */ 2117 u_int direction) /* direction of flow */ 2118 { 2119 struct ippseudo ippseudo; 2120 MD5_CTX ctx; 2121 int doff; 2122 struct ip *ip; 2123 struct ipovly *ipovly; 2124 struct secasvar *sav; 2125 struct tcphdr *th; 2126 #ifdef INET6 2127 struct ip6_hdr *ip6; 2128 struct in6_addr in6; 2129 uint32_t plen; 2130 uint16_t nhdr; 2131 #endif /* INET6 */ 2132 u_short savecsum; 2133 2134 KASSERT(m != NULL, ("passed NULL mbuf. Game over.")); 2135 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature")); 2136 /* 2137 * Extract the destination from the IP header in the mbuf. 2138 */ 2139 ip = mtod(m, struct ip *); 2140 #ifdef INET6 2141 ip6 = NULL; /* Make the compiler happy. */ 2142 #endif /* INET6 */ 2143 /* 2144 * Look up an SADB entry which matches the address found in 2145 * the segment. 2146 */ 2147 switch (IP_VHL_V(ip->ip_vhl)) { 2148 case IPVERSION: 2149 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst, 2150 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2151 break; 2152 #ifdef INET6 2153 case (IPV6_VERSION >> 4): 2154 ip6 = mtod(m, struct ip6_hdr *); 2155 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst, 2156 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2157 break; 2158 #endif /* INET6 */ 2159 default: 2160 return (EINVAL); 2161 /* NOTREACHED */ 2162 break; 2163 } 2164 if (sav == NULL) { 2165 kprintf("%s: SADB lookup failed\n", __func__); 2166 return (EINVAL); 2167 } 2168 MD5Init(&ctx); 2169 2170 /* 2171 * Step 1: Update MD5 hash with IP pseudo-header. 2172 * 2173 * XXX The ippseudo header MUST be digested in network byte order, 2174 * or else we'll fail the regression test. Assume all fields we've 2175 * been doing arithmetic on have been in host byte order. 2176 * XXX One cannot depend on ipovly->ih_len here. When called from 2177 * tcp_output(), the underlying ip_len member has not yet been set. 2178 */ 2179 switch (IP_VHL_V(ip->ip_vhl)) { 2180 case IPVERSION: 2181 ipovly = (struct ipovly *)ip; 2182 ippseudo.ippseudo_src = ipovly->ih_src; 2183 ippseudo.ippseudo_dst = ipovly->ih_dst; 2184 ippseudo.ippseudo_pad = 0; 2185 ippseudo.ippseudo_p = IPPROTO_TCP; 2186 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen); 2187 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo)); 2188 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip)); 2189 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen; 2190 break; 2191 #ifdef INET6 2192 /* 2193 * RFC 2385, 2.0 Proposal 2194 * For IPv6, the pseudo-header is as described in RFC 2460, namely the 2195 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero- 2196 * extended next header value (to form 32 bits), and 32-bit segment 2197 * length. 2198 * Note: Upper-Layer Packet Length comes before Next Header. 2199 */ 2200 case (IPV6_VERSION >> 4): 2201 in6 = ip6->ip6_src; 2202 in6_clearscope(&in6); 2203 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2204 in6 = ip6->ip6_dst; 2205 in6_clearscope(&in6); 2206 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2207 plen = htonl(len + sizeof(struct tcphdr) + optlen); 2208 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t)); 2209 nhdr = 0; 2210 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2211 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2212 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2213 nhdr = IPPROTO_TCP; 2214 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2215 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr)); 2216 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen; 2217 break; 2218 #endif /* INET6 */ 2219 default: 2220 return (EINVAL); 2221 /* NOTREACHED */ 2222 break; 2223 } 2224 /* 2225 * Step 2: Update MD5 hash with TCP header, excluding options. 2226 * The TCP checksum must be set to zero. 2227 */ 2228 savecsum = th->th_sum; 2229 th->th_sum = 0; 2230 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr)); 2231 th->th_sum = savecsum; 2232 /* 2233 * Step 3: Update MD5 hash with TCP segment data. 2234 * Use m_apply() to avoid an early m_pullup(). 2235 */ 2236 if (len > 0) 2237 m_apply(m, doff, len, tcpsignature_apply, &ctx); 2238 /* 2239 * Step 4: Update MD5 hash with shared secret. 2240 */ 2241 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth)); 2242 MD5Final(buf, &ctx); 2243 key_sa_recordxfer(sav, m); 2244 key_freesav(sav); 2245 return (0); 2246 } 2247 2248 int 2249 tcpsignature_apply(void *fstate, void *data, unsigned int len) 2250 { 2251 2252 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len); 2253 return (0); 2254 } 2255 #endif /* TCP_SIGNATURE */ 2256