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, "Fudge bw 1/10% (50=5%)"); 245 246 static int tcp_inflight_adjrtt = 2; 247 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_adjrtt, CTLFLAG_RW, 248 &tcp_inflight_adjrtt, 0, "Slop for rtt 1/(hz*32)"); 249 250 static int tcp_do_rfc3390 = 1; 251 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_RW, 252 &tcp_do_rfc3390, 0, 253 "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)"); 254 255 static u_long tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 256 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwmaxsegs, CTLFLAG_RW, 257 &tcp_iw_maxsegs, 0, "TCP IW segments max"); 258 259 static u_long tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 260 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwcapsegs, CTLFLAG_RW, 261 &tcp_iw_capsegs, 0, "TCP IW segments"); 262 263 int tcp_low_rtobase = 1; 264 SYSCTL_INT(_net_inet_tcp, OID_AUTO, low_rtobase, CTLFLAG_RW, 265 &tcp_low_rtobase, 0, "Lowering the Initial RTO (RFC 6298)"); 266 267 static int tcp_do_ncr = 1; 268 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr, CTLFLAG_RW, 269 &tcp_do_ncr, 0, "Non-Congestion Robustness (RFC 4653)"); 270 271 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives"); 272 static struct malloc_pipe tcptemp_mpipe; 273 274 static void tcp_willblock(void); 275 static void tcp_notify (struct inpcb *, int); 276 277 struct tcp_stats tcpstats_percpu[MAXCPU] __cachealign; 278 279 static int 280 sysctl_tcpstats(SYSCTL_HANDLER_ARGS) 281 { 282 int cpu, error = 0; 283 284 for (cpu = 0; cpu < ncpus2; ++cpu) { 285 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu], 286 sizeof(struct tcp_stats)))) 287 break; 288 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu], 289 sizeof(struct tcp_stats)))) 290 break; 291 } 292 293 return (error); 294 } 295 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW), 296 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics"); 297 298 /* 299 * Target size of TCP PCB hash tables. Must be a power of two. 300 * 301 * Note that this can be overridden by the kernel environment 302 * variable net.inet.tcp.tcbhashsize 303 */ 304 #ifndef TCBHASHSIZE 305 #define TCBHASHSIZE 512 306 #endif 307 308 /* 309 * This is the actual shape of what we allocate using the zone 310 * allocator. Doing it this way allows us to protect both structures 311 * using the same generation count, and also eliminates the overhead 312 * of allocating tcpcbs separately. By hiding the structure here, 313 * we avoid changing most of the rest of the code (although it needs 314 * to be changed, eventually, for greater efficiency). 315 */ 316 #define ALIGNMENT 32 317 #define ALIGNM1 (ALIGNMENT - 1) 318 struct inp_tp { 319 union { 320 struct inpcb inp; 321 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1]; 322 } inp_tp_u; 323 struct tcpcb tcb; 324 struct tcp_callout inp_tp_rexmt; 325 struct tcp_callout inp_tp_persist; 326 struct tcp_callout inp_tp_keep; 327 struct tcp_callout inp_tp_2msl; 328 struct tcp_callout inp_tp_delack; 329 struct netmsg_tcp_timer inp_tp_timermsg; 330 struct netmsg_base inp_tp_sndmore; 331 }; 332 #undef ALIGNMENT 333 #undef ALIGNM1 334 335 /* 336 * Tcp initialization 337 */ 338 void 339 tcp_init(void) 340 { 341 struct inpcbportinfo *portinfo; 342 struct inpcbinfo *ticb; 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; 368 369 portinfo = kmalloc_cachealign(sizeof(*portinfo) * ncpus2, M_PCB, 370 M_WAITOK); 371 372 for (cpu = 0; cpu < ncpus2; cpu++) { 373 ticb = &tcbinfo[cpu]; 374 in_pcbinfo_init(ticb); 375 ticb->cpu = cpu; 376 ticb->hashbase = hashinit(hashsize, M_PCB, 377 &ticb->hashmask); 378 in_pcbportinfo_init(&portinfo[cpu], hashsize, TRUE, cpu); 379 ticb->portinfo = portinfo; 380 ticb->portinfo_mask = ncpus2_mask; 381 ticb->wildcardhashbase = hashinit(hashsize, M_PCB, 382 &ticb->wildcardhashmask); 383 ticb->localgrphashbase = hashinit(hashsize, M_PCB, 384 &ticb->localgrphashmask); 385 ticb->ipi_size = sizeof(struct inp_tp); 386 TAILQ_INIT(&tcpcbackq[cpu]); 387 } 388 389 tcp_reass_maxseg = nmbclusters / 16; 390 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg); 391 392 #ifdef INET6 393 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) 394 #else 395 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) 396 #endif 397 if (max_protohdr < TCP_MINPROTOHDR) 398 max_protohdr = TCP_MINPROTOHDR; 399 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) 400 panic("tcp_init"); 401 #undef TCP_MINPROTOHDR 402 403 /* 404 * Initialize TCP statistics counters for each CPU. 405 */ 406 for (cpu = 0; cpu < ncpus2; ++cpu) 407 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats)); 408 409 syncache_init(); 410 netisr_register_rollup(tcp_willblock, NETISR_ROLLUP_PRIO_TCP); 411 } 412 413 static void 414 tcp_willblock(void) 415 { 416 struct tcpcb *tp; 417 int cpu = mycpu->gd_cpuid; 418 419 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) { 420 KKASSERT(tp->t_flags & TF_ONOUTPUTQ); 421 tp->t_flags &= ~TF_ONOUTPUTQ; 422 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq); 423 tcp_output(tp); 424 } 425 } 426 427 /* 428 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. 429 * tcp_template used to store this data in mbufs, but we now recopy it out 430 * of the tcpcb each time to conserve mbufs. 431 */ 432 void 433 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr, boolean_t tso) 434 { 435 struct inpcb *inp = tp->t_inpcb; 436 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr; 437 438 #ifdef INET6 439 if (inp->inp_vflag & INP_IPV6) { 440 struct ip6_hdr *ip6; 441 442 ip6 = (struct ip6_hdr *)ip_ptr; 443 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | 444 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK); 445 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | 446 (IPV6_VERSION & IPV6_VERSION_MASK); 447 ip6->ip6_nxt = IPPROTO_TCP; 448 ip6->ip6_plen = sizeof(struct tcphdr); 449 ip6->ip6_src = inp->in6p_laddr; 450 ip6->ip6_dst = inp->in6p_faddr; 451 tcp_hdr->th_sum = 0; 452 } else 453 #endif 454 { 455 struct ip *ip = (struct ip *) ip_ptr; 456 u_int plen; 457 458 ip->ip_vhl = IP_VHL_BORING; 459 ip->ip_tos = 0; 460 ip->ip_len = 0; 461 ip->ip_id = 0; 462 ip->ip_off = 0; 463 ip->ip_ttl = 0; 464 ip->ip_sum = 0; 465 ip->ip_p = IPPROTO_TCP; 466 ip->ip_src = inp->inp_laddr; 467 ip->ip_dst = inp->inp_faddr; 468 469 if (tso) 470 plen = htons(IPPROTO_TCP); 471 else 472 plen = htons(sizeof(struct tcphdr) + IPPROTO_TCP); 473 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr, 474 ip->ip_dst.s_addr, plen); 475 } 476 477 tcp_hdr->th_sport = inp->inp_lport; 478 tcp_hdr->th_dport = inp->inp_fport; 479 tcp_hdr->th_seq = 0; 480 tcp_hdr->th_ack = 0; 481 tcp_hdr->th_x2 = 0; 482 tcp_hdr->th_off = 5; 483 tcp_hdr->th_flags = 0; 484 tcp_hdr->th_win = 0; 485 tcp_hdr->th_urp = 0; 486 } 487 488 /* 489 * Create template to be used to send tcp packets on a connection. 490 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only 491 * use for this function is in keepalives, which use tcp_respond. 492 */ 493 struct tcptemp * 494 tcp_maketemplate(struct tcpcb *tp) 495 { 496 struct tcptemp *tmp; 497 498 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL) 499 return (NULL); 500 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t, FALSE); 501 return (tmp); 502 } 503 504 void 505 tcp_freetemplate(struct tcptemp *tmp) 506 { 507 mpipe_free(&tcptemp_mpipe, tmp); 508 } 509 510 /* 511 * Send a single message to the TCP at address specified by 512 * the given TCP/IP header. If m == NULL, then we make a copy 513 * of the tcpiphdr at ti and send directly to the addressed host. 514 * This is used to force keep alive messages out using the TCP 515 * template for a connection. If flags are given then we send 516 * a message back to the TCP which originated the * segment ti, 517 * and discard the mbuf containing it and any other attached mbufs. 518 * 519 * In any case the ack and sequence number of the transmitted 520 * segment are as specified by the parameters. 521 * 522 * NOTE: If m != NULL, then ti must point to *inside* the mbuf. 523 */ 524 void 525 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, 526 tcp_seq ack, tcp_seq seq, int flags) 527 { 528 int tlen; 529 int win = 0; 530 struct route *ro = NULL; 531 struct route sro; 532 struct ip *ip = ipgen; 533 struct tcphdr *nth; 534 int ipflags = 0; 535 struct route_in6 *ro6 = NULL; 536 struct route_in6 sro6; 537 struct ip6_hdr *ip6 = ipgen; 538 boolean_t use_tmpro = TRUE; 539 #ifdef INET6 540 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6); 541 #else 542 const boolean_t isipv6 = FALSE; 543 #endif 544 545 if (tp != NULL) { 546 if (!(flags & TH_RST)) { 547 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv); 548 if (win < 0) 549 win = 0; 550 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 551 win = (long)TCP_MAXWIN << tp->rcv_scale; 552 } 553 /* 554 * Don't use the route cache of a listen socket, 555 * it is not MPSAFE; use temporary route cache. 556 */ 557 if (tp->t_state != TCPS_LISTEN) { 558 if (isipv6) 559 ro6 = &tp->t_inpcb->in6p_route; 560 else 561 ro = &tp->t_inpcb->inp_route; 562 use_tmpro = FALSE; 563 } 564 } 565 if (use_tmpro) { 566 if (isipv6) { 567 ro6 = &sro6; 568 bzero(ro6, sizeof *ro6); 569 } else { 570 ro = &sro; 571 bzero(ro, sizeof *ro); 572 } 573 } 574 if (m == NULL) { 575 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 576 if (m == NULL) 577 return; 578 tlen = 0; 579 m->m_data += max_linkhdr; 580 if (isipv6) { 581 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); 582 ip6 = mtod(m, struct ip6_hdr *); 583 nth = (struct tcphdr *)(ip6 + 1); 584 } else { 585 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip)); 586 ip = mtod(m, struct ip *); 587 nth = (struct tcphdr *)(ip + 1); 588 } 589 bcopy(th, nth, sizeof(struct tcphdr)); 590 flags = TH_ACK; 591 } else { 592 m_freem(m->m_next); 593 m->m_next = NULL; 594 m->m_data = (caddr_t)ipgen; 595 /* m_len is set later */ 596 tlen = 0; 597 #define xchg(a, b, type) { type t; t = a; a = b; b = t; } 598 if (isipv6) { 599 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); 600 nth = (struct tcphdr *)(ip6 + 1); 601 } else { 602 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long); 603 nth = (struct tcphdr *)(ip + 1); 604 } 605 if (th != nth) { 606 /* 607 * this is usually a case when an extension header 608 * exists between the IPv6 header and the 609 * TCP header. 610 */ 611 nth->th_sport = th->th_sport; 612 nth->th_dport = th->th_dport; 613 } 614 xchg(nth->th_dport, nth->th_sport, n_short); 615 #undef xchg 616 } 617 if (isipv6) { 618 ip6->ip6_flow = 0; 619 ip6->ip6_vfc = IPV6_VERSION; 620 ip6->ip6_nxt = IPPROTO_TCP; 621 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen)); 622 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 623 } else { 624 tlen += sizeof(struct tcpiphdr); 625 ip->ip_len = tlen; 626 ip->ip_ttl = ip_defttl; 627 } 628 m->m_len = tlen; 629 m->m_pkthdr.len = tlen; 630 m->m_pkthdr.rcvif = NULL; 631 nth->th_seq = htonl(seq); 632 nth->th_ack = htonl(ack); 633 nth->th_x2 = 0; 634 nth->th_off = sizeof(struct tcphdr) >> 2; 635 nth->th_flags = flags; 636 if (tp != NULL) 637 nth->th_win = htons((u_short) (win >> tp->rcv_scale)); 638 else 639 nth->th_win = htons((u_short)win); 640 nth->th_urp = 0; 641 if (isipv6) { 642 nth->th_sum = 0; 643 nth->th_sum = in6_cksum(m, IPPROTO_TCP, 644 sizeof(struct ip6_hdr), 645 tlen - sizeof(struct ip6_hdr)); 646 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL, 647 (ro6 && ro6->ro_rt) ? 648 ro6->ro_rt->rt_ifp : NULL); 649 } else { 650 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 651 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); 652 m->m_pkthdr.csum_flags = CSUM_TCP; 653 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 654 m->m_pkthdr.csum_thlen = sizeof(struct tcphdr); 655 } 656 #ifdef TCPDEBUG 657 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) 658 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); 659 #endif 660 if (isipv6) { 661 ip6_output(m, NULL, ro6, ipflags, NULL, NULL, 662 tp ? tp->t_inpcb : NULL); 663 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) { 664 RTFREE(ro6->ro_rt); 665 ro6->ro_rt = NULL; 666 } 667 } else { 668 ipflags |= IP_DEBUGROUTE; 669 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL); 670 if ((ro == &sro) && (ro->ro_rt != NULL)) { 671 RTFREE(ro->ro_rt); 672 ro->ro_rt = NULL; 673 } 674 } 675 } 676 677 /* 678 * Create a new TCP control block, making an 679 * empty reassembly queue and hooking it to the argument 680 * protocol control block. The `inp' parameter must have 681 * come from the zone allocator set up in tcp_init(). 682 */ 683 struct tcpcb * 684 tcp_newtcpcb(struct inpcb *inp) 685 { 686 struct inp_tp *it; 687 struct tcpcb *tp; 688 #ifdef INET6 689 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 690 #else 691 const boolean_t isipv6 = FALSE; 692 #endif 693 694 it = (struct inp_tp *)inp; 695 tp = &it->tcb; 696 bzero(tp, sizeof(struct tcpcb)); 697 TAILQ_INIT(&tp->t_segq); 698 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 699 tp->t_rxtthresh = tcprexmtthresh; 700 701 /* Set up our timeouts. */ 702 tp->tt_rexmt = &it->inp_tp_rexmt; 703 tp->tt_persist = &it->inp_tp_persist; 704 tp->tt_keep = &it->inp_tp_keep; 705 tp->tt_2msl = &it->inp_tp_2msl; 706 tp->tt_delack = &it->inp_tp_delack; 707 tcp_inittimers(tp); 708 709 /* 710 * Zero out timer message. We don't create it here, 711 * since the current CPU may not be the owner of this 712 * inpcb. 713 */ 714 tp->tt_msg = &it->inp_tp_timermsg; 715 bzero(tp->tt_msg, sizeof(*tp->tt_msg)); 716 717 tp->t_keepinit = tcp_keepinit; 718 tp->t_keepidle = tcp_keepidle; 719 tp->t_keepintvl = tcp_keepintvl; 720 tp->t_keepcnt = tcp_keepcnt; 721 tp->t_maxidle = tp->t_keepintvl * tp->t_keepcnt; 722 723 if (tcp_do_ncr) 724 tp->t_flags |= TF_NCR; 725 if (tcp_do_rfc1323) 726 tp->t_flags |= (TF_REQ_SCALE | TF_REQ_TSTMP); 727 728 tp->t_inpcb = inp; /* XXX */ 729 tp->t_state = TCPS_CLOSED; 730 /* 731 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 732 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 733 * reasonable initial retransmit time. 734 */ 735 tp->t_srtt = TCPTV_SRTTBASE; 736 tp->t_rttvar = 737 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 738 tp->t_rttmin = tcp_rexmit_min; 739 tp->t_rxtcur = TCPTV_RTOBASE; 740 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 741 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 742 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 743 tp->snd_last = ticks; 744 tp->t_rcvtime = ticks; 745 /* 746 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 747 * because the socket may be bound to an IPv6 wildcard address, 748 * which may match an IPv4-mapped IPv6 address. 749 */ 750 inp->inp_ip_ttl = ip_defttl; 751 inp->inp_ppcb = tp; 752 tcp_sack_tcpcb_init(tp); 753 754 tp->tt_sndmore = &it->inp_tp_sndmore; 755 tcp_output_init(tp); 756 757 return (tp); /* XXX */ 758 } 759 760 /* 761 * Drop a TCP connection, reporting the specified error. 762 * If connection is synchronized, then send a RST to peer. 763 */ 764 struct tcpcb * 765 tcp_drop(struct tcpcb *tp, int error) 766 { 767 struct socket *so = tp->t_inpcb->inp_socket; 768 769 if (TCPS_HAVERCVDSYN(tp->t_state)) { 770 tp->t_state = TCPS_CLOSED; 771 tcp_output(tp); 772 tcpstat.tcps_drops++; 773 } else 774 tcpstat.tcps_conndrops++; 775 if (error == ETIMEDOUT && tp->t_softerror) 776 error = tp->t_softerror; 777 so->so_error = error; 778 return (tcp_close(tp)); 779 } 780 781 struct netmsg_listen_detach { 782 struct netmsg_base base; 783 struct tcpcb *nm_tp; 784 struct tcpcb *nm_tp_inh; 785 }; 786 787 static void 788 tcp_listen_detach_handler(netmsg_t msg) 789 { 790 struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg; 791 struct tcpcb *tp = nmsg->nm_tp; 792 int cpu = mycpuid, nextcpu; 793 794 if (tp->t_flags & TF_LISTEN) 795 syncache_destroy(tp, nmsg->nm_tp_inh); 796 797 in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]); 798 799 nextcpu = cpu + 1; 800 if (nextcpu < ncpus2) 801 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nmsg->base.lmsg); 802 else 803 lwkt_replymsg(&nmsg->base.lmsg, 0); 804 } 805 806 /* 807 * Close a TCP control block: 808 * discard all space held by the tcp 809 * discard internet protocol block 810 * wake up any sleepers 811 */ 812 struct tcpcb * 813 tcp_close(struct tcpcb *tp) 814 { 815 struct tseg_qent *q; 816 struct inpcb *inp = tp->t_inpcb; 817 struct inpcb *inp_inh = NULL; 818 struct tcpcb *tp_inh = NULL; 819 struct socket *so = inp->inp_socket; 820 struct rtentry *rt; 821 boolean_t dosavessthresh; 822 #ifdef INET6 823 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 824 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0); 825 #else 826 const boolean_t isipv6 = FALSE; 827 #endif 828 829 if (tp->t_flags & TF_LISTEN) { 830 /* 831 * Pending socket/syncache inheritance 832 * 833 * If this is a listen(2) socket, find another listen(2) 834 * socket in the same local group, which could inherit 835 * the syncache and sockets pending on the completion 836 * and incompletion queues. 837 * 838 * NOTE: 839 * Currently the inheritance could only happen on the 840 * listen(2) sockets w/ SO_REUSEPORT set. 841 */ 842 KASSERT(&curthread->td_msgport == netisr_cpuport(0), 843 ("listen socket close not in netisr0")); 844 inp_inh = in_pcblocalgroup_last(&tcbinfo[0], inp); 845 if (inp_inh != NULL) 846 tp_inh = intotcpcb(inp_inh); 847 } 848 849 /* 850 * INP_WILDCARD_MP indicates that listen(2) has been called on 851 * this socket. This implies: 852 * - A wildcard inp's hash is replicated for each protocol thread. 853 * - Syncache for this inp grows independently in each protocol 854 * thread. 855 * - There is more than one cpu 856 * 857 * We have to chain a message to the rest of the protocol threads 858 * to cleanup the wildcard hash and the syncache. The cleanup 859 * in the current protocol thread is defered till the end of this 860 * function. 861 * 862 * NOTE: 863 * After cleanup the inp's hash and syncache entries, this inp will 864 * no longer be available to the rest of the protocol threads, so we 865 * are safe to whack the inp in the following code. 866 */ 867 if (inp->inp_flags & INP_WILDCARD_MP) { 868 struct netmsg_listen_detach nmsg; 869 870 KKASSERT(so->so_port == netisr_cpuport(0)); 871 KKASSERT(&curthread->td_msgport == netisr_cpuport(0)); 872 KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]); 873 874 netmsg_init(&nmsg.base, NULL, &curthread->td_msgport, 875 MSGF_PRIORITY, tcp_listen_detach_handler); 876 nmsg.nm_tp = tp; 877 nmsg.nm_tp_inh = tp_inh; 878 lwkt_domsg(netisr_cpuport(1), &nmsg.base.lmsg, 0); 879 880 inp->inp_flags &= ~INP_WILDCARD_MP; 881 } 882 883 KKASSERT(tp->t_state != TCPS_TERMINATING); 884 tp->t_state = TCPS_TERMINATING; 885 886 /* 887 * Make sure that all of our timers are stopped before we 888 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL), 889 * timers are never used. If timer message is never created 890 * (tp->tt_msg->tt_tcb == NULL), timers are never used too. 891 */ 892 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) { 893 tcp_callout_stop(tp, tp->tt_rexmt); 894 tcp_callout_stop(tp, tp->tt_persist); 895 tcp_callout_stop(tp, tp->tt_keep); 896 tcp_callout_stop(tp, tp->tt_2msl); 897 tcp_callout_stop(tp, tp->tt_delack); 898 } 899 900 if (tp->t_flags & TF_ONOUTPUTQ) { 901 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 902 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 903 tp->t_flags &= ~TF_ONOUTPUTQ; 904 } 905 906 /* 907 * If we got enough samples through the srtt filter, 908 * save the rtt and rttvar in the routing entry. 909 * 'Enough' is arbitrarily defined as the 16 samples. 910 * 16 samples is enough for the srtt filter to converge 911 * to within 5% of the correct value; fewer samples and 912 * we could save a very bogus rtt. 913 * 914 * Don't update the default route's characteristics and don't 915 * update anything that the user "locked". 916 */ 917 if (tp->t_rttupdated >= 16) { 918 u_long i = 0; 919 920 if (isipv6) { 921 struct sockaddr_in6 *sin6; 922 923 if ((rt = inp->in6p_route.ro_rt) == NULL) 924 goto no_valid_rt; 925 sin6 = (struct sockaddr_in6 *)rt_key(rt); 926 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 927 goto no_valid_rt; 928 } else 929 if ((rt = inp->inp_route.ro_rt) == NULL || 930 ((struct sockaddr_in *)rt_key(rt))-> 931 sin_addr.s_addr == INADDR_ANY) 932 goto no_valid_rt; 933 934 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 935 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 936 if (rt->rt_rmx.rmx_rtt && i) 937 /* 938 * filter this update to half the old & half 939 * the new values, converting scale. 940 * See route.h and tcp_var.h for a 941 * description of the scaling constants. 942 */ 943 rt->rt_rmx.rmx_rtt = 944 (rt->rt_rmx.rmx_rtt + i) / 2; 945 else 946 rt->rt_rmx.rmx_rtt = i; 947 tcpstat.tcps_cachedrtt++; 948 } 949 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 950 i = tp->t_rttvar * 951 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 952 if (rt->rt_rmx.rmx_rttvar && i) 953 rt->rt_rmx.rmx_rttvar = 954 (rt->rt_rmx.rmx_rttvar + i) / 2; 955 else 956 rt->rt_rmx.rmx_rttvar = i; 957 tcpstat.tcps_cachedrttvar++; 958 } 959 /* 960 * The old comment here said: 961 * update the pipelimit (ssthresh) if it has been updated 962 * already or if a pipesize was specified & the threshhold 963 * got below half the pipesize. I.e., wait for bad news 964 * before we start updating, then update on both good 965 * and bad news. 966 * 967 * But we want to save the ssthresh even if no pipesize is 968 * specified explicitly in the route, because such 969 * connections still have an implicit pipesize specified 970 * by the global tcp_sendspace. In the absence of a reliable 971 * way to calculate the pipesize, it will have to do. 972 */ 973 i = tp->snd_ssthresh; 974 if (rt->rt_rmx.rmx_sendpipe != 0) 975 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 976 else 977 dosavessthresh = (i < so->so_snd.ssb_hiwat/2); 978 if (dosavessthresh || 979 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 980 (rt->rt_rmx.rmx_ssthresh != 0))) { 981 /* 982 * convert the limit from user data bytes to 983 * packets then to packet data bytes. 984 */ 985 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 986 if (i < 2) 987 i = 2; 988 i *= tp->t_maxseg + 989 (isipv6 ? 990 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 991 sizeof(struct tcpiphdr)); 992 if (rt->rt_rmx.rmx_ssthresh) 993 rt->rt_rmx.rmx_ssthresh = 994 (rt->rt_rmx.rmx_ssthresh + i) / 2; 995 else 996 rt->rt_rmx.rmx_ssthresh = i; 997 tcpstat.tcps_cachedssthresh++; 998 } 999 } 1000 1001 no_valid_rt: 1002 /* free the reassembly queue, if any */ 1003 while((q = TAILQ_FIRST(&tp->t_segq)) != NULL) { 1004 TAILQ_REMOVE(&tp->t_segq, q, tqe_q); 1005 m_freem(q->tqe_m); 1006 kfree(q, M_TSEGQ); 1007 atomic_add_int(&tcp_reass_qsize, -1); 1008 } 1009 /* throw away SACK blocks in scoreboard*/ 1010 if (TCP_DO_SACK(tp)) 1011 tcp_sack_destroy(&tp->scb); 1012 1013 inp->inp_ppcb = NULL; 1014 soisdisconnected(so); 1015 /* note: pcb detached later on */ 1016 1017 tcp_destroy_timermsg(tp); 1018 tcp_output_cancel(tp); 1019 1020 if (tp->t_flags & TF_LISTEN) { 1021 syncache_destroy(tp, tp_inh); 1022 if (inp_inh != NULL && inp_inh->inp_socket != NULL) { 1023 /* 1024 * Pending sockets inheritance only needs 1025 * to be done once in the current thread, 1026 * i.e. netisr0. 1027 */ 1028 soinherit(so, inp_inh->inp_socket); 1029 } 1030 } 1031 1032 so_async_rcvd_drop(so); 1033 /* Drop the reference for the asynchronized pru_rcvd */ 1034 sofree(so); 1035 1036 /* 1037 * NOTE: 1038 * pcbdetach removes any wildcard hash entry on the current CPU. 1039 */ 1040 #ifdef INET6 1041 if (isafinet6) 1042 in6_pcbdetach(inp); 1043 else 1044 #endif 1045 in_pcbdetach(inp); 1046 1047 tcpstat.tcps_closed++; 1048 return (NULL); 1049 } 1050 1051 static __inline void 1052 tcp_drain_oncpu(struct inpcbhead *head) 1053 { 1054 struct inpcb *marker; 1055 struct inpcb *inpb; 1056 struct tcpcb *tcpb; 1057 struct tseg_qent *te; 1058 1059 /* 1060 * Allows us to block while running the list 1061 */ 1062 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1063 marker->inp_flags |= INP_PLACEMARKER; 1064 LIST_INSERT_HEAD(head, marker, inp_list); 1065 1066 while ((inpb = LIST_NEXT(marker, inp_list)) != NULL) { 1067 if ((inpb->inp_flags & INP_PLACEMARKER) == 0 && 1068 (tcpb = intotcpcb(inpb)) != NULL && 1069 (te = TAILQ_FIRST(&tcpb->t_segq)) != NULL) { 1070 TAILQ_REMOVE(&tcpb->t_segq, te, tqe_q); 1071 if (te->tqe_th->th_flags & TH_FIN) 1072 tcpb->t_flags &= ~TF_QUEDFIN; 1073 m_freem(te->tqe_m); 1074 kfree(te, M_TSEGQ); 1075 atomic_add_int(&tcp_reass_qsize, -1); 1076 /* retry */ 1077 } else { 1078 LIST_REMOVE(marker, inp_list); 1079 LIST_INSERT_AFTER(inpb, marker, inp_list); 1080 } 1081 } 1082 LIST_REMOVE(marker, inp_list); 1083 kfree(marker, M_TEMP); 1084 } 1085 1086 struct netmsg_tcp_drain { 1087 struct netmsg_base base; 1088 struct inpcbhead *nm_head; 1089 }; 1090 1091 static void 1092 tcp_drain_handler(netmsg_t msg) 1093 { 1094 struct netmsg_tcp_drain *nm = (void *)msg; 1095 1096 tcp_drain_oncpu(nm->nm_head); 1097 lwkt_replymsg(&nm->base.lmsg, 0); 1098 } 1099 1100 void 1101 tcp_drain(void) 1102 { 1103 int cpu; 1104 1105 if (!do_tcpdrain) 1106 return; 1107 1108 /* 1109 * Walk the tcpbs, if existing, and flush the reassembly queue, 1110 * if there is one... 1111 * XXX: The "Net/3" implementation doesn't imply that the TCP 1112 * reassembly queue should be flushed, but in a situation 1113 * where we're really low on mbufs, this is potentially 1114 * useful. 1115 */ 1116 for (cpu = 0; cpu < ncpus2; cpu++) { 1117 struct netmsg_tcp_drain *nm; 1118 1119 if (cpu == mycpu->gd_cpuid) { 1120 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead); 1121 } else { 1122 nm = kmalloc(sizeof(struct netmsg_tcp_drain), 1123 M_LWKTMSG, M_NOWAIT); 1124 if (nm == NULL) 1125 continue; 1126 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1127 0, tcp_drain_handler); 1128 nm->nm_head = &tcbinfo[cpu].pcblisthead; 1129 lwkt_sendmsg(netisr_cpuport(cpu), &nm->base.lmsg); 1130 } 1131 } 1132 } 1133 1134 /* 1135 * Notify a tcp user of an asynchronous error; 1136 * store error as soft error, but wake up user 1137 * (for now, won't do anything until can select for soft error). 1138 * 1139 * Do not wake up user since there currently is no mechanism for 1140 * reporting soft errors (yet - a kqueue filter may be added). 1141 */ 1142 static void 1143 tcp_notify(struct inpcb *inp, int error) 1144 { 1145 struct tcpcb *tp = intotcpcb(inp); 1146 1147 /* 1148 * Ignore some errors if we are hooked up. 1149 * If connection hasn't completed, has retransmitted several times, 1150 * and receives a second error, give up now. This is better 1151 * than waiting a long time to establish a connection that 1152 * can never complete. 1153 */ 1154 if (tp->t_state == TCPS_ESTABLISHED && 1155 (error == EHOSTUNREACH || error == ENETUNREACH || 1156 error == EHOSTDOWN)) { 1157 return; 1158 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1159 tp->t_softerror) 1160 tcp_drop(tp, error); 1161 else 1162 tp->t_softerror = error; 1163 #if 0 1164 wakeup(&so->so_timeo); 1165 sorwakeup(so); 1166 sowwakeup(so); 1167 #endif 1168 } 1169 1170 static int 1171 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1172 { 1173 int error, i, n; 1174 struct inpcb *marker; 1175 struct inpcb *inp; 1176 globaldata_t gd; 1177 int origcpu, ccpu; 1178 1179 error = 0; 1180 n = 0; 1181 1182 /* 1183 * The process of preparing the TCB list is too time-consuming and 1184 * resource-intensive to repeat twice on every request. 1185 */ 1186 if (req->oldptr == NULL) { 1187 for (ccpu = 0; ccpu < ncpus2; ++ccpu) { 1188 gd = globaldata_find(ccpu); 1189 n += tcbinfo[gd->gd_cpuid].ipi_count; 1190 } 1191 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1192 return (0); 1193 } 1194 1195 if (req->newptr != NULL) 1196 return (EPERM); 1197 1198 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1199 marker->inp_flags |= INP_PLACEMARKER; 1200 1201 /* 1202 * OK, now we're committed to doing something. Run the inpcb list 1203 * for each cpu in the system and construct the output. Use a 1204 * list placemarker to deal with list changes occuring during 1205 * copyout blockages (but otherwise depend on being on the correct 1206 * cpu to avoid races). 1207 */ 1208 origcpu = mycpu->gd_cpuid; 1209 for (ccpu = 0; ccpu < ncpus2 && error == 0; ++ccpu) { 1210 caddr_t inp_ppcb; 1211 struct xtcpcb xt; 1212 1213 lwkt_migratecpu(ccpu); 1214 1215 n = tcbinfo[ccpu].ipi_count; 1216 1217 LIST_INSERT_HEAD(&tcbinfo[ccpu].pcblisthead, marker, inp_list); 1218 i = 0; 1219 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1220 /* 1221 * process a snapshot of pcbs, ignoring placemarkers 1222 * and using our own to allow SYSCTL_OUT to block. 1223 */ 1224 LIST_REMOVE(marker, inp_list); 1225 LIST_INSERT_AFTER(inp, marker, inp_list); 1226 1227 if (inp->inp_flags & INP_PLACEMARKER) 1228 continue; 1229 if (prison_xinpcb(req->td, inp)) 1230 continue; 1231 1232 xt.xt_len = sizeof xt; 1233 bcopy(inp, &xt.xt_inp, sizeof *inp); 1234 inp_ppcb = inp->inp_ppcb; 1235 if (inp_ppcb != NULL) 1236 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1237 else 1238 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1239 if (inp->inp_socket) 1240 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1241 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1242 break; 1243 ++i; 1244 } 1245 LIST_REMOVE(marker, inp_list); 1246 if (error == 0 && i < n) { 1247 bzero(&xt, sizeof xt); 1248 xt.xt_len = sizeof xt; 1249 while (i < n) { 1250 error = SYSCTL_OUT(req, &xt, sizeof xt); 1251 if (error) 1252 break; 1253 ++i; 1254 } 1255 } 1256 } 1257 1258 /* 1259 * Make sure we are on the same cpu we were on originally, since 1260 * higher level callers expect this. Also don't pollute caches with 1261 * migrated userland data by (eventually) returning to userland 1262 * on a different cpu. 1263 */ 1264 lwkt_migratecpu(origcpu); 1265 kfree(marker, M_TEMP); 1266 return (error); 1267 } 1268 1269 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1270 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1271 1272 static int 1273 tcp_getcred(SYSCTL_HANDLER_ARGS) 1274 { 1275 struct sockaddr_in addrs[2]; 1276 struct inpcb *inp; 1277 int cpu; 1278 int error; 1279 1280 error = priv_check(req->td, PRIV_ROOT); 1281 if (error != 0) 1282 return (error); 1283 error = SYSCTL_IN(req, addrs, sizeof addrs); 1284 if (error != 0) 1285 return (error); 1286 crit_enter(); 1287 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1288 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1289 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1290 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1291 if (inp == NULL || inp->inp_socket == NULL) { 1292 error = ENOENT; 1293 goto out; 1294 } 1295 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1296 out: 1297 crit_exit(); 1298 return (error); 1299 } 1300 1301 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1302 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1303 1304 #ifdef INET6 1305 static int 1306 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1307 { 1308 struct sockaddr_in6 addrs[2]; 1309 struct inpcb *inp; 1310 int error; 1311 boolean_t mapped = FALSE; 1312 1313 error = priv_check(req->td, PRIV_ROOT); 1314 if (error != 0) 1315 return (error); 1316 error = SYSCTL_IN(req, addrs, sizeof addrs); 1317 if (error != 0) 1318 return (error); 1319 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { 1320 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) 1321 mapped = TRUE; 1322 else 1323 return (EINVAL); 1324 } 1325 crit_enter(); 1326 if (mapped) { 1327 inp = in_pcblookup_hash(&tcbinfo[0], 1328 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], 1329 addrs[1].sin6_port, 1330 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], 1331 addrs[0].sin6_port, 1332 0, NULL); 1333 } else { 1334 inp = in6_pcblookup_hash(&tcbinfo[0], 1335 &addrs[1].sin6_addr, addrs[1].sin6_port, 1336 &addrs[0].sin6_addr, addrs[0].sin6_port, 1337 0, NULL); 1338 } 1339 if (inp == NULL || inp->inp_socket == NULL) { 1340 error = ENOENT; 1341 goto out; 1342 } 1343 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1344 out: 1345 crit_exit(); 1346 return (error); 1347 } 1348 1349 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1350 0, 0, 1351 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1352 #endif 1353 1354 struct netmsg_tcp_notify { 1355 struct netmsg_base base; 1356 void (*nm_notify)(struct inpcb *, int); 1357 struct in_addr nm_faddr; 1358 int nm_arg; 1359 }; 1360 1361 static void 1362 tcp_notifyall_oncpu(netmsg_t msg) 1363 { 1364 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg; 1365 int nextcpu; 1366 1367 in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nm->nm_faddr, 1368 nm->nm_arg, nm->nm_notify); 1369 1370 nextcpu = mycpuid + 1; 1371 if (nextcpu < ncpus2) 1372 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nm->base.lmsg); 1373 else 1374 lwkt_replymsg(&nm->base.lmsg, 0); 1375 } 1376 1377 void 1378 tcp_ctlinput(netmsg_t msg) 1379 { 1380 int cmd = msg->ctlinput.nm_cmd; 1381 struct sockaddr *sa = msg->ctlinput.nm_arg; 1382 struct ip *ip = msg->ctlinput.nm_extra; 1383 struct tcphdr *th; 1384 struct in_addr faddr; 1385 struct inpcb *inp; 1386 struct tcpcb *tp; 1387 void (*notify)(struct inpcb *, int) = tcp_notify; 1388 tcp_seq icmpseq; 1389 int arg, cpu; 1390 1391 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) { 1392 goto done; 1393 } 1394 1395 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1396 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1397 goto done; 1398 1399 arg = inetctlerrmap[cmd]; 1400 if (cmd == PRC_QUENCH) { 1401 notify = tcp_quench; 1402 } else if (icmp_may_rst && 1403 (cmd == PRC_UNREACH_ADMIN_PROHIB || 1404 cmd == PRC_UNREACH_PORT || 1405 cmd == PRC_TIMXCEED_INTRANS) && 1406 ip != NULL) { 1407 notify = tcp_drop_syn_sent; 1408 } else if (cmd == PRC_MSGSIZE) { 1409 struct icmp *icmp = (struct icmp *) 1410 ((caddr_t)ip - offsetof(struct icmp, icmp_ip)); 1411 1412 arg = ntohs(icmp->icmp_nextmtu); 1413 notify = tcp_mtudisc; 1414 } else if (PRC_IS_REDIRECT(cmd)) { 1415 ip = NULL; 1416 notify = in_rtchange; 1417 } else if (cmd == PRC_HOSTDEAD) { 1418 ip = NULL; 1419 } 1420 1421 if (ip != NULL) { 1422 crit_enter(); 1423 th = (struct tcphdr *)((caddr_t)ip + 1424 (IP_VHL_HL(ip->ip_vhl) << 2)); 1425 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport, 1426 ip->ip_src.s_addr, th->th_sport); 1427 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport, 1428 ip->ip_src, th->th_sport, 0, NULL); 1429 if ((inp != NULL) && (inp->inp_socket != NULL)) { 1430 icmpseq = htonl(th->th_seq); 1431 tp = intotcpcb(inp); 1432 if (SEQ_GEQ(icmpseq, tp->snd_una) && 1433 SEQ_LT(icmpseq, tp->snd_max)) 1434 (*notify)(inp, arg); 1435 } else { 1436 struct in_conninfo inc; 1437 1438 inc.inc_fport = th->th_dport; 1439 inc.inc_lport = th->th_sport; 1440 inc.inc_faddr = faddr; 1441 inc.inc_laddr = ip->ip_src; 1442 #ifdef INET6 1443 inc.inc_isipv6 = 0; 1444 #endif 1445 syncache_unreach(&inc, th); 1446 } 1447 crit_exit(); 1448 } else { 1449 struct netmsg_tcp_notify *nm; 1450 1451 KKASSERT(&curthread->td_msgport == netisr_cpuport(0)); 1452 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT); 1453 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1454 0, tcp_notifyall_oncpu); 1455 nm->nm_faddr = faddr; 1456 nm->nm_arg = arg; 1457 nm->nm_notify = notify; 1458 1459 lwkt_sendmsg(netisr_cpuport(0), &nm->base.lmsg); 1460 } 1461 done: 1462 lwkt_replymsg(&msg->lmsg, 0); 1463 } 1464 1465 #ifdef INET6 1466 1467 void 1468 tcp6_ctlinput(netmsg_t msg) 1469 { 1470 int cmd = msg->ctlinput.nm_cmd; 1471 struct sockaddr *sa = msg->ctlinput.nm_arg; 1472 void *d = msg->ctlinput.nm_extra; 1473 struct tcphdr th; 1474 void (*notify) (struct inpcb *, int) = tcp_notify; 1475 struct ip6_hdr *ip6; 1476 struct mbuf *m; 1477 struct ip6ctlparam *ip6cp = NULL; 1478 const struct sockaddr_in6 *sa6_src = NULL; 1479 int off; 1480 struct tcp_portonly { 1481 u_int16_t th_sport; 1482 u_int16_t th_dport; 1483 } *thp; 1484 int arg; 1485 1486 if (sa->sa_family != AF_INET6 || 1487 sa->sa_len != sizeof(struct sockaddr_in6)) { 1488 goto out; 1489 } 1490 1491 arg = 0; 1492 if (cmd == PRC_QUENCH) 1493 notify = tcp_quench; 1494 else if (cmd == PRC_MSGSIZE) { 1495 struct ip6ctlparam *ip6cp = d; 1496 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6; 1497 1498 arg = ntohl(icmp6->icmp6_mtu); 1499 notify = tcp_mtudisc; 1500 } else if (!PRC_IS_REDIRECT(cmd) && 1501 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) { 1502 goto out; 1503 } 1504 1505 /* if the parameter is from icmp6, decode it. */ 1506 if (d != NULL) { 1507 ip6cp = (struct ip6ctlparam *)d; 1508 m = ip6cp->ip6c_m; 1509 ip6 = ip6cp->ip6c_ip6; 1510 off = ip6cp->ip6c_off; 1511 sa6_src = ip6cp->ip6c_src; 1512 } else { 1513 m = NULL; 1514 ip6 = NULL; 1515 off = 0; /* fool gcc */ 1516 sa6_src = &sa6_any; 1517 } 1518 1519 if (ip6 != NULL) { 1520 struct in_conninfo inc; 1521 /* 1522 * XXX: We assume that when IPV6 is non NULL, 1523 * M and OFF are valid. 1524 */ 1525 1526 /* check if we can safely examine src and dst ports */ 1527 if (m->m_pkthdr.len < off + sizeof *thp) 1528 goto out; 1529 1530 bzero(&th, sizeof th); 1531 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1532 1533 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport, 1534 (struct sockaddr *)ip6cp->ip6c_src, 1535 th.th_sport, cmd, arg, notify); 1536 1537 inc.inc_fport = th.th_dport; 1538 inc.inc_lport = th.th_sport; 1539 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1540 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1541 inc.inc_isipv6 = 1; 1542 syncache_unreach(&inc, &th); 1543 } else { 1544 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0, 1545 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify); 1546 } 1547 out: 1548 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0); 1549 } 1550 1551 #endif 1552 1553 /* 1554 * Following is where TCP initial sequence number generation occurs. 1555 * 1556 * There are two places where we must use initial sequence numbers: 1557 * 1. In SYN-ACK packets. 1558 * 2. In SYN packets. 1559 * 1560 * All ISNs for SYN-ACK packets are generated by the syncache. See 1561 * tcp_syncache.c for details. 1562 * 1563 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1564 * depends on this property. In addition, these ISNs should be 1565 * unguessable so as to prevent connection hijacking. To satisfy 1566 * the requirements of this situation, the algorithm outlined in 1567 * RFC 1948 is used to generate sequence numbers. 1568 * 1569 * Implementation details: 1570 * 1571 * Time is based off the system timer, and is corrected so that it 1572 * increases by one megabyte per second. This allows for proper 1573 * recycling on high speed LANs while still leaving over an hour 1574 * before rollover. 1575 * 1576 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1577 * between seeding of isn_secret. This is normally set to zero, 1578 * as reseeding should not be necessary. 1579 * 1580 */ 1581 1582 #define ISN_BYTES_PER_SECOND 1048576 1583 1584 u_char isn_secret[32]; 1585 int isn_last_reseed; 1586 MD5_CTX isn_ctx; 1587 1588 tcp_seq 1589 tcp_new_isn(struct tcpcb *tp) 1590 { 1591 u_int32_t md5_buffer[4]; 1592 tcp_seq new_isn; 1593 1594 /* Seed if this is the first use, reseed if requested. */ 1595 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1596 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1597 < (u_int)ticks))) { 1598 read_random_unlimited(&isn_secret, sizeof isn_secret); 1599 isn_last_reseed = ticks; 1600 } 1601 1602 /* Compute the md5 hash and return the ISN. */ 1603 MD5Init(&isn_ctx); 1604 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1605 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1606 #ifdef INET6 1607 if (tp->t_inpcb->inp_vflag & INP_IPV6) { 1608 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1609 sizeof(struct in6_addr)); 1610 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1611 sizeof(struct in6_addr)); 1612 } else 1613 #endif 1614 { 1615 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1616 sizeof(struct in_addr)); 1617 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1618 sizeof(struct in_addr)); 1619 } 1620 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1621 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1622 new_isn = (tcp_seq) md5_buffer[0]; 1623 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1624 return (new_isn); 1625 } 1626 1627 /* 1628 * When a source quench is received, close congestion window 1629 * to one segment. We will gradually open it again as we proceed. 1630 */ 1631 void 1632 tcp_quench(struct inpcb *inp, int error) 1633 { 1634 struct tcpcb *tp = intotcpcb(inp); 1635 1636 if (tp != NULL) { 1637 tp->snd_cwnd = tp->t_maxseg; 1638 tp->snd_wacked = 0; 1639 } 1640 } 1641 1642 /* 1643 * When a specific ICMP unreachable message is received and the 1644 * connection state is SYN-SENT, drop the connection. This behavior 1645 * is controlled by the icmp_may_rst sysctl. 1646 */ 1647 void 1648 tcp_drop_syn_sent(struct inpcb *inp, int error) 1649 { 1650 struct tcpcb *tp = intotcpcb(inp); 1651 1652 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT)) 1653 tcp_drop(tp, error); 1654 } 1655 1656 /* 1657 * When a `need fragmentation' ICMP is received, update our idea of the MSS 1658 * based on the new value in the route. Also nudge TCP to send something, 1659 * since we know the packet we just sent was dropped. 1660 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1661 */ 1662 void 1663 tcp_mtudisc(struct inpcb *inp, int mtu) 1664 { 1665 struct tcpcb *tp = intotcpcb(inp); 1666 struct rtentry *rt; 1667 struct socket *so = inp->inp_socket; 1668 int maxopd, mss; 1669 #ifdef INET6 1670 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0); 1671 #else 1672 const boolean_t isipv6 = FALSE; 1673 #endif 1674 1675 if (tp == NULL) 1676 return; 1677 1678 /* 1679 * If no MTU is provided in the ICMP message, use the 1680 * next lower likely value, as specified in RFC 1191. 1681 */ 1682 if (mtu == 0) { 1683 int oldmtu; 1684 1685 oldmtu = tp->t_maxopd + 1686 (isipv6 ? 1687 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1688 sizeof(struct tcpiphdr)); 1689 mtu = ip_next_mtu(oldmtu, 0); 1690 } 1691 1692 if (isipv6) 1693 rt = tcp_rtlookup6(&inp->inp_inc); 1694 else 1695 rt = tcp_rtlookup(&inp->inp_inc); 1696 if (rt != NULL) { 1697 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu) 1698 mtu = rt->rt_rmx.rmx_mtu; 1699 1700 maxopd = mtu - 1701 (isipv6 ? 1702 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1703 sizeof(struct tcpiphdr)); 1704 1705 /* 1706 * XXX - The following conditional probably violates the TCP 1707 * spec. The problem is that, since we don't know the 1708 * other end's MSS, we are supposed to use a conservative 1709 * default. But, if we do that, then MTU discovery will 1710 * never actually take place, because the conservative 1711 * default is much less than the MTUs typically seen 1712 * on the Internet today. For the moment, we'll sweep 1713 * this under the carpet. 1714 * 1715 * The conservative default might not actually be a problem 1716 * if the only case this occurs is when sending an initial 1717 * SYN with options and data to a host we've never talked 1718 * to before. Then, they will reply with an MSS value which 1719 * will get recorded and the new parameters should get 1720 * recomputed. For Further Study. 1721 */ 1722 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd) 1723 maxopd = rt->rt_rmx.rmx_mssopt; 1724 } else 1725 maxopd = mtu - 1726 (isipv6 ? 1727 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1728 sizeof(struct tcpiphdr)); 1729 1730 if (tp->t_maxopd <= maxopd) 1731 return; 1732 tp->t_maxopd = maxopd; 1733 1734 mss = maxopd; 1735 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) == 1736 (TF_REQ_TSTMP | TF_RCVD_TSTMP)) 1737 mss -= TCPOLEN_TSTAMP_APPA; 1738 1739 /* round down to multiple of MCLBYTES */ 1740 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */ 1741 if (mss > MCLBYTES) 1742 mss &= ~(MCLBYTES - 1); 1743 #else 1744 if (mss > MCLBYTES) 1745 mss = (mss / MCLBYTES) * MCLBYTES; 1746 #endif 1747 1748 if (so->so_snd.ssb_hiwat < mss) 1749 mss = so->so_snd.ssb_hiwat; 1750 1751 tp->t_maxseg = mss; 1752 tp->t_rtttime = 0; 1753 tp->snd_nxt = tp->snd_una; 1754 tcp_output(tp); 1755 tcpstat.tcps_mturesent++; 1756 } 1757 1758 /* 1759 * Look-up the routing entry to the peer of this inpcb. If no route 1760 * is found and it cannot be allocated the return NULL. This routine 1761 * is called by TCP routines that access the rmx structure and by tcp_mss 1762 * to get the interface MTU. 1763 */ 1764 struct rtentry * 1765 tcp_rtlookup(struct in_conninfo *inc) 1766 { 1767 struct route *ro = &inc->inc_route; 1768 1769 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1770 /* No route yet, so try to acquire one */ 1771 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1772 /* 1773 * unused portions of the structure MUST be zero'd 1774 * out because rtalloc() treats it as opaque data 1775 */ 1776 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1777 ro->ro_dst.sa_family = AF_INET; 1778 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1779 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1780 inc->inc_faddr; 1781 rtalloc(ro); 1782 } 1783 } 1784 return (ro->ro_rt); 1785 } 1786 1787 #ifdef INET6 1788 struct rtentry * 1789 tcp_rtlookup6(struct in_conninfo *inc) 1790 { 1791 struct route_in6 *ro6 = &inc->inc6_route; 1792 1793 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1794 /* No route yet, so try to acquire one */ 1795 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1796 /* 1797 * unused portions of the structure MUST be zero'd 1798 * out because rtalloc() treats it as opaque data 1799 */ 1800 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1801 ro6->ro_dst.sin6_family = AF_INET6; 1802 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1803 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1804 rtalloc((struct route *)ro6); 1805 } 1806 } 1807 return (ro6->ro_rt); 1808 } 1809 #endif 1810 1811 #ifdef IPSEC 1812 /* compute ESP/AH header size for TCP, including outer IP header. */ 1813 size_t 1814 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1815 { 1816 struct inpcb *inp; 1817 struct mbuf *m; 1818 size_t hdrsiz; 1819 struct ip *ip; 1820 struct tcphdr *th; 1821 1822 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1823 return (0); 1824 MGETHDR(m, MB_DONTWAIT, MT_DATA); 1825 if (!m) 1826 return (0); 1827 1828 #ifdef INET6 1829 if (inp->inp_vflag & INP_IPV6) { 1830 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1831 1832 th = (struct tcphdr *)(ip6 + 1); 1833 m->m_pkthdr.len = m->m_len = 1834 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1835 tcp_fillheaders(tp, ip6, th, FALSE); 1836 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1837 } else 1838 #endif 1839 { 1840 ip = mtod(m, struct ip *); 1841 th = (struct tcphdr *)(ip + 1); 1842 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1843 tcp_fillheaders(tp, ip, th, FALSE); 1844 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1845 } 1846 1847 m_free(m); 1848 return (hdrsiz); 1849 } 1850 #endif 1851 1852 /* 1853 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1854 * 1855 * This code attempts to calculate the bandwidth-delay product as a 1856 * means of determining the optimal window size to maximize bandwidth, 1857 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1858 * routers. This code also does a fairly good job keeping RTTs in check 1859 * across slow links like modems. We implement an algorithm which is very 1860 * similar (but not meant to be) TCP/Vegas. The code operates on the 1861 * transmitter side of a TCP connection and so only effects the transmit 1862 * side of the connection. 1863 * 1864 * BACKGROUND: TCP makes no provision for the management of buffer space 1865 * at the end points or at the intermediate routers and switches. A TCP 1866 * stream, whether using NewReno or not, will eventually buffer as 1867 * many packets as it is able and the only reason this typically works is 1868 * due to the fairly small default buffers made available for a connection 1869 * (typicaly 16K or 32K). As machines use larger windows and/or window 1870 * scaling it is now fairly easy for even a single TCP connection to blow-out 1871 * all available buffer space not only on the local interface, but on 1872 * intermediate routers and switches as well. NewReno makes a misguided 1873 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1874 * then backing off, then steadily increasing the window again until another 1875 * failure occurs, ad-infinitum. This results in terrible oscillation that 1876 * is only made worse as network loads increase and the idea of intentionally 1877 * blowing out network buffers is, frankly, a terrible way to manage network 1878 * resources. 1879 * 1880 * It is far better to limit the transmit window prior to the failure 1881 * condition being achieved. There are two general ways to do this: First 1882 * you can 'scan' through different transmit window sizes and locate the 1883 * point where the RTT stops increasing, indicating that you have filled the 1884 * pipe, then scan backwards until you note that RTT stops decreasing, then 1885 * repeat ad-infinitum. This method works in principle but has severe 1886 * implementation issues due to RTT variances, timer granularity, and 1887 * instability in the algorithm which can lead to many false positives and 1888 * create oscillations as well as interact badly with other TCP streams 1889 * implementing the same algorithm. 1890 * 1891 * The second method is to limit the window to the bandwidth delay product 1892 * of the link. This is the method we implement. RTT variances and our 1893 * own manipulation of the congestion window, bwnd, can potentially 1894 * destabilize the algorithm. For this reason we have to stabilize the 1895 * elements used to calculate the window. We do this by using the minimum 1896 * observed RTT, the long term average of the observed bandwidth, and 1897 * by adding two segments worth of slop. It isn't perfect but it is able 1898 * to react to changing conditions and gives us a very stable basis on 1899 * which to extend the algorithm. 1900 */ 1901 void 1902 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1903 { 1904 u_long bw; 1905 u_long bwnd; 1906 int save_ticks; 1907 int delta_ticks; 1908 1909 /* 1910 * If inflight_enable is disabled in the middle of a tcp connection, 1911 * make sure snd_bwnd is effectively disabled. 1912 */ 1913 if (!tcp_inflight_enable) { 1914 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1915 tp->snd_bandwidth = 0; 1916 return; 1917 } 1918 1919 /* 1920 * Validate the delta time. If a connection is new or has been idle 1921 * a long time we have to reset the bandwidth calculator. 1922 */ 1923 save_ticks = ticks; 1924 cpu_ccfence(); 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 1934 /* 1935 * A delta of at least 1 tick is required. Waiting 2 ticks will 1936 * result in better (bw) accuracy. More than that and the ramp-up 1937 * will be too slow. 1938 */ 1939 if (delta_ticks == 0 || delta_ticks == 1) 1940 return; 1941 1942 /* 1943 * Sanity check, plus ignore pure window update acks. 1944 */ 1945 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1946 return; 1947 1948 /* 1949 * Figure out the bandwidth. Due to the tick granularity this 1950 * is a very rough number and it MUST be averaged over a fairly 1951 * long period of time. XXX we need to take into account a link 1952 * that is not using all available bandwidth, but for now our 1953 * slop will ramp us up if this case occurs and the bandwidth later 1954 * increases. 1955 */ 1956 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 1957 tp->t_bw_rtttime = save_ticks; 1958 tp->t_bw_rtseq = ack_seq; 1959 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4; 1960 1961 tp->snd_bandwidth = bw; 1962 1963 /* 1964 * Calculate the semi-static bandwidth delay product, plus two maximal 1965 * segments. The additional slop puts us squarely in the sweet 1966 * spot and also handles the bandwidth run-up case. Without the 1967 * slop we could be locking ourselves into a lower bandwidth. 1968 * 1969 * At very high speeds the bw calculation can become overly sensitive 1970 * and error prone when delta_ticks is low (e.g. usually 1). To deal 1971 * with the problem the stab must be scaled to the bw. A stab of 50 1972 * (the default) increases the bw for the purposes of the bwnd 1973 * calculation by 5%. 1974 * 1975 * Situations Handled: 1976 * (1) Prevents over-queueing of packets on LANs, especially on 1977 * high speed LANs, allowing larger TCP buffers to be 1978 * specified, and also does a good job preventing 1979 * over-queueing of packets over choke points like modems 1980 * (at least for the transmit side). 1981 * 1982 * (2) Is able to handle changing network loads (bandwidth 1983 * drops so bwnd drops, bandwidth increases so bwnd 1984 * increases). 1985 * 1986 * (3) Theoretically should stabilize in the face of multiple 1987 * connections implementing the same algorithm (this may need 1988 * a little work). 1989 * 1990 * (4) Stability value (defaults to 20 = 2 maximal packets) can 1991 * be adjusted with a sysctl but typically only needs to be on 1992 * very slow connections. A value no smaller then 5 should 1993 * be used, but only reduce this default if you have no other 1994 * choice. 1995 */ 1996 1997 #define USERTT (((tp->t_srtt + tp->t_rttbest) / 2) + tcp_inflight_adjrtt) 1998 bw += bw * tcp_inflight_stab / 1000; 1999 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 2000 (int)tp->t_maxseg * 2; 2001 #undef USERTT 2002 2003 if (tcp_inflight_debug > 0) { 2004 static int ltime; 2005 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) { 2006 ltime = ticks; 2007 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n", 2008 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd); 2009 } 2010 } 2011 if ((long)bwnd < tcp_inflight_min) 2012 bwnd = tcp_inflight_min; 2013 if (bwnd > tcp_inflight_max) 2014 bwnd = tcp_inflight_max; 2015 if ((long)bwnd < tp->t_maxseg * 2) 2016 bwnd = tp->t_maxseg * 2; 2017 tp->snd_bwnd = bwnd; 2018 } 2019 2020 static void 2021 tcp_rmx_iwsegs(struct tcpcb *tp, u_long *maxsegs, u_long *capsegs) 2022 { 2023 struct rtentry *rt; 2024 struct inpcb *inp = tp->t_inpcb; 2025 #ifdef INET6 2026 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE); 2027 #else 2028 const boolean_t isipv6 = FALSE; 2029 #endif 2030 2031 /* XXX */ 2032 if (tcp_iw_maxsegs < TCP_IW_MAXSEGS_DFLT) 2033 tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 2034 if (tcp_iw_capsegs < TCP_IW_CAPSEGS_DFLT) 2035 tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 2036 2037 if (isipv6) 2038 rt = tcp_rtlookup6(&inp->inp_inc); 2039 else 2040 rt = tcp_rtlookup(&inp->inp_inc); 2041 if (rt == NULL || 2042 rt->rt_rmx.rmx_iwmaxsegs < TCP_IW_MAXSEGS_DFLT || 2043 rt->rt_rmx.rmx_iwcapsegs < TCP_IW_CAPSEGS_DFLT) { 2044 *maxsegs = tcp_iw_maxsegs; 2045 *capsegs = tcp_iw_capsegs; 2046 return; 2047 } 2048 *maxsegs = rt->rt_rmx.rmx_iwmaxsegs; 2049 *capsegs = rt->rt_rmx.rmx_iwcapsegs; 2050 } 2051 2052 u_long 2053 tcp_initial_window(struct tcpcb *tp) 2054 { 2055 if (tcp_do_rfc3390) { 2056 /* 2057 * RFC3390: 2058 * "If the SYN or SYN/ACK is lost, the initial window 2059 * used by a sender after a correctly transmitted SYN 2060 * MUST be one segment consisting of MSS bytes." 2061 * 2062 * However, we do something a little bit more aggressive 2063 * then RFC3390 here: 2064 * - Only if time spent in the SYN or SYN|ACK retransmition 2065 * >= 3 seconds, the IW is reduced. We do this mainly 2066 * because when RFC3390 is published, the initial RTO is 2067 * still 3 seconds (the threshold we test here), while 2068 * after RFC6298, the initial RTO is 1 second. This 2069 * behaviour probably still falls within the spirit of 2070 * RFC3390. 2071 * - When IW is reduced, 2*MSS is used instead of 1*MSS. 2072 * Mainly to avoid sender and receiver deadlock until 2073 * delayed ACK timer expires. And even RFC2581 does not 2074 * try to reduce IW upon SYN or SYN|ACK retransmition 2075 * timeout. 2076 * 2077 * See also: 2078 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03 2079 */ 2080 if (tp->t_rxtsyn >= TCPTV_RTOBASE3) { 2081 return (2 * tp->t_maxseg); 2082 } else { 2083 u_long maxsegs, capsegs; 2084 2085 tcp_rmx_iwsegs(tp, &maxsegs, &capsegs); 2086 return min(maxsegs * tp->t_maxseg, 2087 max(2 * tp->t_maxseg, capsegs * 1460)); 2088 } 2089 } else { 2090 /* 2091 * Even RFC2581 (back to 1999) allows 2*SMSS IW. 2092 * 2093 * Mainly to avoid sender and receiver deadlock 2094 * until delayed ACK timer expires. 2095 */ 2096 return (2 * tp->t_maxseg); 2097 } 2098 } 2099 2100 #ifdef TCP_SIGNATURE 2101 /* 2102 * Compute TCP-MD5 hash of a TCP segment. (RFC2385) 2103 * 2104 * We do this over ip, tcphdr, segment data, and the key in the SADB. 2105 * When called from tcp_input(), we can be sure that th_sum has been 2106 * zeroed out and verified already. 2107 * 2108 * Return 0 if successful, otherwise return -1. 2109 * 2110 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a 2111 * search with the destination IP address, and a 'magic SPI' to be 2112 * determined by the application. This is hardcoded elsewhere to 1179 2113 * right now. Another branch of this code exists which uses the SPD to 2114 * specify per-application flows but it is unstable. 2115 */ 2116 int 2117 tcpsignature_compute( 2118 struct mbuf *m, /* mbuf chain */ 2119 int len, /* length of TCP data */ 2120 int optlen, /* length of TCP options */ 2121 u_char *buf, /* storage for MD5 digest */ 2122 u_int direction) /* direction of flow */ 2123 { 2124 struct ippseudo ippseudo; 2125 MD5_CTX ctx; 2126 int doff; 2127 struct ip *ip; 2128 struct ipovly *ipovly; 2129 struct secasvar *sav; 2130 struct tcphdr *th; 2131 #ifdef INET6 2132 struct ip6_hdr *ip6; 2133 struct in6_addr in6; 2134 uint32_t plen; 2135 uint16_t nhdr; 2136 #endif /* INET6 */ 2137 u_short savecsum; 2138 2139 KASSERT(m != NULL, ("passed NULL mbuf. Game over.")); 2140 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature")); 2141 /* 2142 * Extract the destination from the IP header in the mbuf. 2143 */ 2144 ip = mtod(m, struct ip *); 2145 #ifdef INET6 2146 ip6 = NULL; /* Make the compiler happy. */ 2147 #endif /* INET6 */ 2148 /* 2149 * Look up an SADB entry which matches the address found in 2150 * the segment. 2151 */ 2152 switch (IP_VHL_V(ip->ip_vhl)) { 2153 case IPVERSION: 2154 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst, 2155 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2156 break; 2157 #ifdef INET6 2158 case (IPV6_VERSION >> 4): 2159 ip6 = mtod(m, struct ip6_hdr *); 2160 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst, 2161 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2162 break; 2163 #endif /* INET6 */ 2164 default: 2165 return (EINVAL); 2166 /* NOTREACHED */ 2167 break; 2168 } 2169 if (sav == NULL) { 2170 kprintf("%s: SADB lookup failed\n", __func__); 2171 return (EINVAL); 2172 } 2173 MD5Init(&ctx); 2174 2175 /* 2176 * Step 1: Update MD5 hash with IP pseudo-header. 2177 * 2178 * XXX The ippseudo header MUST be digested in network byte order, 2179 * or else we'll fail the regression test. Assume all fields we've 2180 * been doing arithmetic on have been in host byte order. 2181 * XXX One cannot depend on ipovly->ih_len here. When called from 2182 * tcp_output(), the underlying ip_len member has not yet been set. 2183 */ 2184 switch (IP_VHL_V(ip->ip_vhl)) { 2185 case IPVERSION: 2186 ipovly = (struct ipovly *)ip; 2187 ippseudo.ippseudo_src = ipovly->ih_src; 2188 ippseudo.ippseudo_dst = ipovly->ih_dst; 2189 ippseudo.ippseudo_pad = 0; 2190 ippseudo.ippseudo_p = IPPROTO_TCP; 2191 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen); 2192 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo)); 2193 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip)); 2194 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen; 2195 break; 2196 #ifdef INET6 2197 /* 2198 * RFC 2385, 2.0 Proposal 2199 * For IPv6, the pseudo-header is as described in RFC 2460, namely the 2200 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero- 2201 * extended next header value (to form 32 bits), and 32-bit segment 2202 * length. 2203 * Note: Upper-Layer Packet Length comes before Next Header. 2204 */ 2205 case (IPV6_VERSION >> 4): 2206 in6 = ip6->ip6_src; 2207 in6_clearscope(&in6); 2208 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2209 in6 = ip6->ip6_dst; 2210 in6_clearscope(&in6); 2211 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2212 plen = htonl(len + sizeof(struct tcphdr) + optlen); 2213 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t)); 2214 nhdr = 0; 2215 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2216 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2217 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2218 nhdr = IPPROTO_TCP; 2219 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2220 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr)); 2221 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen; 2222 break; 2223 #endif /* INET6 */ 2224 default: 2225 return (EINVAL); 2226 /* NOTREACHED */ 2227 break; 2228 } 2229 /* 2230 * Step 2: Update MD5 hash with TCP header, excluding options. 2231 * The TCP checksum must be set to zero. 2232 */ 2233 savecsum = th->th_sum; 2234 th->th_sum = 0; 2235 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr)); 2236 th->th_sum = savecsum; 2237 /* 2238 * Step 3: Update MD5 hash with TCP segment data. 2239 * Use m_apply() to avoid an early m_pullup(). 2240 */ 2241 if (len > 0) 2242 m_apply(m, doff, len, tcpsignature_apply, &ctx); 2243 /* 2244 * Step 4: Update MD5 hash with shared secret. 2245 */ 2246 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth)); 2247 MD5Final(buf, &ctx); 2248 key_sa_recordxfer(sav, m); 2249 key_freesav(sav); 2250 return (0); 2251 } 2252 2253 int 2254 tcpsignature_apply(void *fstate, void *data, unsigned int len) 2255 { 2256 2257 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len); 2258 return (0); 2259 } 2260 #endif /* TCP_SIGNATURE */ 2261