1 /* 2 * Copyright (c) 2003, 2004 Jeffrey M. Hsu. All rights reserved. 3 * Copyright (c) 2003, 2004 The DragonFly Project. All rights reserved. 4 * 5 * This code is derived from software contributed to The DragonFly Project 6 * by Jeffrey M. Hsu. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. Neither the name of The DragonFly Project nor the names of its 17 * contributors may be used to endorse or promote products derived 18 * from this software without specific, prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 22 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 23 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 24 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 25 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 26 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 27 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 28 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 29 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 30 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 */ 33 34 /* 35 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 36 * The Regents of the University of California. All rights reserved. 37 * 38 * Redistribution and use in source and binary forms, with or without 39 * modification, are permitted provided that the following conditions 40 * are met: 41 * 1. Redistributions of source code must retain the above copyright 42 * notice, this list of conditions and the following disclaimer. 43 * 2. Redistributions in binary form must reproduce the above copyright 44 * notice, this list of conditions and the following disclaimer in the 45 * documentation and/or other materials provided with the distribution. 46 * 3. Neither the name of the University nor the names of its contributors 47 * may be used to endorse or promote products derived from this software 48 * without specific prior written permission. 49 * 50 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 51 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 52 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 53 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 54 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 55 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 56 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 57 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 58 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 59 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 60 * SUCH DAMAGE. 61 * 62 * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95 63 * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $ 64 */ 65 66 #include "opt_compat.h" 67 #include "opt_inet.h" 68 #include "opt_inet6.h" 69 #include "opt_ipsec.h" 70 #include "opt_tcpdebug.h" 71 72 #include <sys/param.h> 73 #include <sys/systm.h> 74 #include <sys/callout.h> 75 #include <sys/kernel.h> 76 #include <sys/sysctl.h> 77 #include <sys/malloc.h> 78 #include <sys/mpipe.h> 79 #include <sys/mbuf.h> 80 #ifdef INET6 81 #include <sys/domain.h> 82 #endif 83 #include <sys/proc.h> 84 #include <sys/priv.h> 85 #include <sys/socket.h> 86 #include <sys/socketops.h> 87 #include <sys/socketvar.h> 88 #include <sys/protosw.h> 89 #include <sys/random.h> 90 #include <sys/in_cksum.h> 91 #include <sys/ktr.h> 92 93 #include <net/route.h> 94 #include <net/if.h> 95 #include <net/netisr2.h> 96 97 #define _IP_VHL 98 #include <netinet/in.h> 99 #include <netinet/in_systm.h> 100 #include <netinet/ip.h> 101 #include <netinet/ip6.h> 102 #include <netinet/in_pcb.h> 103 #include <netinet6/in6_pcb.h> 104 #include <netinet/in_var.h> 105 #include <netinet/ip_var.h> 106 #include <netinet6/ip6_var.h> 107 #include <netinet/ip_icmp.h> 108 #ifdef INET6 109 #include <netinet/icmp6.h> 110 #endif 111 #include <netinet/tcp.h> 112 #include <netinet/tcp_fsm.h> 113 #include <netinet/tcp_seq.h> 114 #include <netinet/tcp_timer.h> 115 #include <netinet/tcp_timer2.h> 116 #include <netinet/tcp_var.h> 117 #include <netinet6/tcp6_var.h> 118 #include <netinet/tcpip.h> 119 #ifdef TCPDEBUG 120 #include <netinet/tcp_debug.h> 121 #endif 122 #include <netinet6/ip6protosw.h> 123 124 #ifdef IPSEC 125 #include <netinet6/ipsec.h> 126 #include <netproto/key/key.h> 127 #ifdef INET6 128 #include <netinet6/ipsec6.h> 129 #endif 130 #endif 131 132 #ifdef FAST_IPSEC 133 #include <netproto/ipsec/ipsec.h> 134 #ifdef INET6 135 #include <netproto/ipsec/ipsec6.h> 136 #endif 137 #define IPSEC 138 #endif 139 140 #include <sys/md5.h> 141 #include <machine/smp.h> 142 143 #include <sys/msgport2.h> 144 #include <sys/mplock2.h> 145 #include <net/netmsg2.h> 146 147 #if !defined(KTR_TCP) 148 #define KTR_TCP KTR_ALL 149 #endif 150 /* 151 KTR_INFO_MASTER(tcp); 152 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0); 153 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0); 154 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0); 155 #define logtcp(name) KTR_LOG(tcp_ ## name) 156 */ 157 158 #define TCP_IW_MAXSEGS_DFLT 4 159 #define TCP_IW_CAPSEGS_DFLT 3 160 161 struct inpcbinfo tcbinfo[MAXCPU]; 162 struct tcpcbackqhead tcpcbackq[MAXCPU]; 163 164 int tcp_mssdflt = TCP_MSS; 165 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 166 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size"); 167 168 #ifdef INET6 169 int tcp_v6mssdflt = TCP6_MSS; 170 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW, 171 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6"); 172 #endif 173 174 /* 175 * Minimum MSS we accept and use. This prevents DoS attacks where 176 * we are forced to a ridiculous low MSS like 20 and send hundreds 177 * of packets instead of one. The effect scales with the available 178 * bandwidth and quickly saturates the CPU and network interface 179 * with packet generation and sending. Set to zero to disable MINMSS 180 * checking. This setting prevents us from sending too small packets. 181 */ 182 int tcp_minmss = TCP_MINMSS; 183 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW, 184 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size"); 185 186 #if 0 187 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ; 188 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW, 189 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time"); 190 #endif 191 192 int tcp_do_rfc1323 = 1; 193 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW, 194 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions"); 195 196 static int tcp_tcbhashsize = 0; 197 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD, 198 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable"); 199 200 static int do_tcpdrain = 1; 201 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0, 202 "Enable tcp_drain routine for extra help when low on mbufs"); 203 204 static int icmp_may_rst = 1; 205 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0, 206 "Certain ICMP unreachable messages may abort connections in SYN_SENT"); 207 208 static int tcp_isn_reseed_interval = 0; 209 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW, 210 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret"); 211 212 /* 213 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on 214 * by default, but with generous values which should allow maximal 215 * bandwidth. In particular, the slop defaults to 50 (5 packets). 216 * 217 * The reason for doing this is that the limiter is the only mechanism we 218 * have which seems to do a really good job preventing receiver RX rings 219 * on network interfaces from getting blown out. Even though GigE/10GigE 220 * is supposed to flow control it looks like either it doesn't actually 221 * do it or Open Source drivers do not properly enable it. 222 * 223 * People using the limiter to reduce bottlenecks on slower WAN connections 224 * should set the slop to 20 (2 packets). 225 */ 226 static int tcp_inflight_enable = 1; 227 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW, 228 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting"); 229 230 static int tcp_inflight_debug = 0; 231 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW, 232 &tcp_inflight_debug, 0, "Debug TCP inflight calculations"); 233 234 static int tcp_inflight_min = 6144; 235 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW, 236 &tcp_inflight_min, 0, "Lower bound for TCP inflight window"); 237 238 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT; 239 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW, 240 &tcp_inflight_max, 0, "Upper bound for TCP inflight window"); 241 242 static int tcp_inflight_stab = 50; 243 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW, 244 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 3 packets)"); 245 246 static int tcp_do_rfc3390 = 1; 247 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_RW, 248 &tcp_do_rfc3390, 0, 249 "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)"); 250 251 static u_long tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 252 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwmaxsegs, CTLFLAG_RW, 253 &tcp_iw_maxsegs, 0, "TCP IW segments max"); 254 255 static u_long tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 256 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwcapsegs, CTLFLAG_RW, 257 &tcp_iw_capsegs, 0, "TCP IW segments"); 258 259 int tcp_low_rtobase = 1; 260 SYSCTL_INT(_net_inet_tcp, OID_AUTO, low_rtobase, CTLFLAG_RW, 261 &tcp_low_rtobase, 0, "Lowering the Initial RTO (RFC 6298)"); 262 263 static int tcp_do_ncr = 1; 264 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr, CTLFLAG_RW, 265 &tcp_do_ncr, 0, "Non-Congestion Robustness (RFC 4653)"); 266 267 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives"); 268 static struct malloc_pipe tcptemp_mpipe; 269 270 static void tcp_willblock(void); 271 static void tcp_notify (struct inpcb *, int); 272 273 struct tcp_stats tcpstats_percpu[MAXCPU] __cachealign; 274 275 static int 276 sysctl_tcpstats(SYSCTL_HANDLER_ARGS) 277 { 278 int cpu, error = 0; 279 280 for (cpu = 0; cpu < ncpus2; ++cpu) { 281 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu], 282 sizeof(struct tcp_stats)))) 283 break; 284 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu], 285 sizeof(struct tcp_stats)))) 286 break; 287 } 288 289 return (error); 290 } 291 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW), 292 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics"); 293 294 /* 295 * Target size of TCP PCB hash tables. Must be a power of two. 296 * 297 * Note that this can be overridden by the kernel environment 298 * variable net.inet.tcp.tcbhashsize 299 */ 300 #ifndef TCBHASHSIZE 301 #define TCBHASHSIZE 512 302 #endif 303 304 /* 305 * This is the actual shape of what we allocate using the zone 306 * allocator. Doing it this way allows us to protect both structures 307 * using the same generation count, and also eliminates the overhead 308 * of allocating tcpcbs separately. By hiding the structure here, 309 * we avoid changing most of the rest of the code (although it needs 310 * to be changed, eventually, for greater efficiency). 311 */ 312 #define ALIGNMENT 32 313 #define ALIGNM1 (ALIGNMENT - 1) 314 struct inp_tp { 315 union { 316 struct inpcb inp; 317 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1]; 318 } inp_tp_u; 319 struct tcpcb tcb; 320 struct tcp_callout inp_tp_rexmt; 321 struct tcp_callout inp_tp_persist; 322 struct tcp_callout inp_tp_keep; 323 struct tcp_callout inp_tp_2msl; 324 struct tcp_callout inp_tp_delack; 325 struct netmsg_tcp_timer inp_tp_timermsg; 326 struct netmsg_base inp_tp_sndmore; 327 }; 328 #undef ALIGNMENT 329 #undef ALIGNM1 330 331 /* 332 * Tcp initialization 333 */ 334 void 335 tcp_init(void) 336 { 337 struct inpcbportinfo *portinfo; 338 struct inpcbinfo *ticb; 339 int hashsize = TCBHASHSIZE; 340 int cpu; 341 342 /* 343 * note: tcptemp is used for keepalives, and it is ok for an 344 * allocation to fail so do not specify MPF_INT. 345 */ 346 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp), 347 25, -1, 0, NULL, NULL, NULL); 348 349 tcp_delacktime = TCPTV_DELACK; 350 tcp_keepinit = TCPTV_KEEP_INIT; 351 tcp_keepidle = TCPTV_KEEP_IDLE; 352 tcp_keepintvl = TCPTV_KEEPINTVL; 353 tcp_maxpersistidle = TCPTV_KEEP_IDLE; 354 tcp_msl = TCPTV_MSL; 355 tcp_rexmit_min = TCPTV_MIN; 356 tcp_rexmit_slop = TCPTV_CPU_VAR; 357 358 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize); 359 if (!powerof2(hashsize)) { 360 kprintf("WARNING: TCB hash size not a power of 2\n"); 361 hashsize = 512; /* safe default */ 362 } 363 tcp_tcbhashsize = hashsize; 364 365 portinfo = kmalloc_cachealign(sizeof(*portinfo) * ncpus2, M_PCB, 366 M_WAITOK); 367 368 for (cpu = 0; cpu < ncpus2; cpu++) { 369 ticb = &tcbinfo[cpu]; 370 in_pcbinfo_init(ticb); 371 ticb->cpu = cpu; 372 ticb->hashbase = hashinit(hashsize, M_PCB, 373 &ticb->hashmask); 374 in_pcbportinfo_init(&portinfo[cpu], hashsize, TRUE, cpu); 375 ticb->portinfo = portinfo; 376 ticb->portinfo_mask = ncpus2_mask; 377 ticb->wildcardhashbase = hashinit(hashsize, M_PCB, 378 &ticb->wildcardhashmask); 379 ticb->localgrphashbase = hashinit(hashsize, M_PCB, 380 &ticb->localgrphashmask); 381 ticb->ipi_size = sizeof(struct inp_tp); 382 TAILQ_INIT(&tcpcbackq[cpu]); 383 } 384 385 tcp_reass_maxseg = nmbclusters / 16; 386 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg); 387 388 #ifdef INET6 389 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) 390 #else 391 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) 392 #endif 393 if (max_protohdr < TCP_MINPROTOHDR) 394 max_protohdr = TCP_MINPROTOHDR; 395 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) 396 panic("tcp_init"); 397 #undef TCP_MINPROTOHDR 398 399 /* 400 * Initialize TCP statistics counters for each CPU. 401 */ 402 for (cpu = 0; cpu < ncpus2; ++cpu) 403 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats)); 404 405 syncache_init(); 406 netisr_register_rollup(tcp_willblock, NETISR_ROLLUP_PRIO_TCP); 407 } 408 409 static void 410 tcp_willblock(void) 411 { 412 struct tcpcb *tp; 413 int cpu = mycpu->gd_cpuid; 414 415 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) { 416 KKASSERT(tp->t_flags & TF_ONOUTPUTQ); 417 tp->t_flags &= ~TF_ONOUTPUTQ; 418 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq); 419 tcp_output(tp); 420 } 421 } 422 423 /* 424 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. 425 * tcp_template used to store this data in mbufs, but we now recopy it out 426 * of the tcpcb each time to conserve mbufs. 427 */ 428 void 429 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr, boolean_t tso) 430 { 431 struct inpcb *inp = tp->t_inpcb; 432 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr; 433 434 #ifdef INET6 435 if (inp->inp_vflag & INP_IPV6) { 436 struct ip6_hdr *ip6; 437 438 ip6 = (struct ip6_hdr *)ip_ptr; 439 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | 440 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK); 441 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | 442 (IPV6_VERSION & IPV6_VERSION_MASK); 443 ip6->ip6_nxt = IPPROTO_TCP; 444 ip6->ip6_plen = sizeof(struct tcphdr); 445 ip6->ip6_src = inp->in6p_laddr; 446 ip6->ip6_dst = inp->in6p_faddr; 447 tcp_hdr->th_sum = 0; 448 } else 449 #endif 450 { 451 struct ip *ip = (struct ip *) ip_ptr; 452 u_int plen; 453 454 ip->ip_vhl = IP_VHL_BORING; 455 ip->ip_tos = 0; 456 ip->ip_len = 0; 457 ip->ip_id = 0; 458 ip->ip_off = 0; 459 ip->ip_ttl = 0; 460 ip->ip_sum = 0; 461 ip->ip_p = IPPROTO_TCP; 462 ip->ip_src = inp->inp_laddr; 463 ip->ip_dst = inp->inp_faddr; 464 465 if (tso) 466 plen = htons(IPPROTO_TCP); 467 else 468 plen = htons(sizeof(struct tcphdr) + IPPROTO_TCP); 469 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr, 470 ip->ip_dst.s_addr, plen); 471 } 472 473 tcp_hdr->th_sport = inp->inp_lport; 474 tcp_hdr->th_dport = inp->inp_fport; 475 tcp_hdr->th_seq = 0; 476 tcp_hdr->th_ack = 0; 477 tcp_hdr->th_x2 = 0; 478 tcp_hdr->th_off = 5; 479 tcp_hdr->th_flags = 0; 480 tcp_hdr->th_win = 0; 481 tcp_hdr->th_urp = 0; 482 } 483 484 /* 485 * Create template to be used to send tcp packets on a connection. 486 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only 487 * use for this function is in keepalives, which use tcp_respond. 488 */ 489 struct tcptemp * 490 tcp_maketemplate(struct tcpcb *tp) 491 { 492 struct tcptemp *tmp; 493 494 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL) 495 return (NULL); 496 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t, FALSE); 497 return (tmp); 498 } 499 500 void 501 tcp_freetemplate(struct tcptemp *tmp) 502 { 503 mpipe_free(&tcptemp_mpipe, tmp); 504 } 505 506 /* 507 * Send a single message to the TCP at address specified by 508 * the given TCP/IP header. If m == NULL, then we make a copy 509 * of the tcpiphdr at ti and send directly to the addressed host. 510 * This is used to force keep alive messages out using the TCP 511 * template for a connection. If flags are given then we send 512 * a message back to the TCP which originated the * segment ti, 513 * and discard the mbuf containing it and any other attached mbufs. 514 * 515 * In any case the ack and sequence number of the transmitted 516 * segment are as specified by the parameters. 517 * 518 * NOTE: If m != NULL, then ti must point to *inside* the mbuf. 519 */ 520 void 521 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, 522 tcp_seq ack, tcp_seq seq, int flags) 523 { 524 int tlen; 525 int win = 0; 526 struct route *ro = NULL; 527 struct route sro; 528 struct ip *ip = ipgen; 529 struct tcphdr *nth; 530 int ipflags = 0; 531 struct route_in6 *ro6 = NULL; 532 struct route_in6 sro6; 533 struct ip6_hdr *ip6 = ipgen; 534 boolean_t use_tmpro = TRUE; 535 #ifdef INET6 536 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6); 537 #else 538 const boolean_t isipv6 = FALSE; 539 #endif 540 541 if (tp != NULL) { 542 if (!(flags & TH_RST)) { 543 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv); 544 if (win < 0) 545 win = 0; 546 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 547 win = (long)TCP_MAXWIN << tp->rcv_scale; 548 } 549 /* 550 * Don't use the route cache of a listen socket, 551 * it is not MPSAFE; use temporary route cache. 552 */ 553 if (tp->t_state != TCPS_LISTEN) { 554 if (isipv6) 555 ro6 = &tp->t_inpcb->in6p_route; 556 else 557 ro = &tp->t_inpcb->inp_route; 558 use_tmpro = FALSE; 559 } 560 } 561 if (use_tmpro) { 562 if (isipv6) { 563 ro6 = &sro6; 564 bzero(ro6, sizeof *ro6); 565 } else { 566 ro = &sro; 567 bzero(ro, sizeof *ro); 568 } 569 } 570 if (m == NULL) { 571 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 572 if (m == NULL) 573 return; 574 tlen = 0; 575 m->m_data += max_linkhdr; 576 if (isipv6) { 577 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); 578 ip6 = mtod(m, struct ip6_hdr *); 579 nth = (struct tcphdr *)(ip6 + 1); 580 } else { 581 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip)); 582 ip = mtod(m, struct ip *); 583 nth = (struct tcphdr *)(ip + 1); 584 } 585 bcopy(th, nth, sizeof(struct tcphdr)); 586 flags = TH_ACK; 587 } else { 588 m_freem(m->m_next); 589 m->m_next = NULL; 590 m->m_data = (caddr_t)ipgen; 591 /* m_len is set later */ 592 tlen = 0; 593 #define xchg(a, b, type) { type t; t = a; a = b; b = t; } 594 if (isipv6) { 595 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); 596 nth = (struct tcphdr *)(ip6 + 1); 597 } else { 598 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long); 599 nth = (struct tcphdr *)(ip + 1); 600 } 601 if (th != nth) { 602 /* 603 * this is usually a case when an extension header 604 * exists between the IPv6 header and the 605 * TCP header. 606 */ 607 nth->th_sport = th->th_sport; 608 nth->th_dport = th->th_dport; 609 } 610 xchg(nth->th_dport, nth->th_sport, n_short); 611 #undef xchg 612 } 613 if (isipv6) { 614 ip6->ip6_flow = 0; 615 ip6->ip6_vfc = IPV6_VERSION; 616 ip6->ip6_nxt = IPPROTO_TCP; 617 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen)); 618 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 619 } else { 620 tlen += sizeof(struct tcpiphdr); 621 ip->ip_len = tlen; 622 ip->ip_ttl = ip_defttl; 623 } 624 m->m_len = tlen; 625 m->m_pkthdr.len = tlen; 626 m->m_pkthdr.rcvif = NULL; 627 nth->th_seq = htonl(seq); 628 nth->th_ack = htonl(ack); 629 nth->th_x2 = 0; 630 nth->th_off = sizeof(struct tcphdr) >> 2; 631 nth->th_flags = flags; 632 if (tp != NULL) 633 nth->th_win = htons((u_short) (win >> tp->rcv_scale)); 634 else 635 nth->th_win = htons((u_short)win); 636 nth->th_urp = 0; 637 if (isipv6) { 638 nth->th_sum = 0; 639 nth->th_sum = in6_cksum(m, IPPROTO_TCP, 640 sizeof(struct ip6_hdr), 641 tlen - sizeof(struct ip6_hdr)); 642 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL, 643 (ro6 && ro6->ro_rt) ? 644 ro6->ro_rt->rt_ifp : NULL); 645 } else { 646 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 647 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); 648 m->m_pkthdr.csum_flags = CSUM_TCP; 649 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 650 m->m_pkthdr.csum_thlen = sizeof(struct tcphdr); 651 } 652 #ifdef TCPDEBUG 653 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) 654 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); 655 #endif 656 if (isipv6) { 657 ip6_output(m, NULL, ro6, ipflags, NULL, NULL, 658 tp ? tp->t_inpcb : NULL); 659 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) { 660 RTFREE(ro6->ro_rt); 661 ro6->ro_rt = NULL; 662 } 663 } else { 664 ipflags |= IP_DEBUGROUTE; 665 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL); 666 if ((ro == &sro) && (ro->ro_rt != NULL)) { 667 RTFREE(ro->ro_rt); 668 ro->ro_rt = NULL; 669 } 670 } 671 } 672 673 /* 674 * Create a new TCP control block, making an 675 * empty reassembly queue and hooking it to the argument 676 * protocol control block. The `inp' parameter must have 677 * come from the zone allocator set up in tcp_init(). 678 */ 679 struct tcpcb * 680 tcp_newtcpcb(struct inpcb *inp) 681 { 682 struct inp_tp *it; 683 struct tcpcb *tp; 684 #ifdef INET6 685 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 686 #else 687 const boolean_t isipv6 = FALSE; 688 #endif 689 690 it = (struct inp_tp *)inp; 691 tp = &it->tcb; 692 bzero(tp, sizeof(struct tcpcb)); 693 TAILQ_INIT(&tp->t_segq); 694 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 695 tp->t_rxtthresh = tcprexmtthresh; 696 697 /* Set up our timeouts. */ 698 tp->tt_rexmt = &it->inp_tp_rexmt; 699 tp->tt_persist = &it->inp_tp_persist; 700 tp->tt_keep = &it->inp_tp_keep; 701 tp->tt_2msl = &it->inp_tp_2msl; 702 tp->tt_delack = &it->inp_tp_delack; 703 tcp_inittimers(tp); 704 705 /* 706 * Zero out timer message. We don't create it here, 707 * since the current CPU may not be the owner of this 708 * inpcb. 709 */ 710 tp->tt_msg = &it->inp_tp_timermsg; 711 bzero(tp->tt_msg, sizeof(*tp->tt_msg)); 712 713 tp->t_keepinit = tcp_keepinit; 714 tp->t_keepidle = tcp_keepidle; 715 tp->t_keepintvl = tcp_keepintvl; 716 tp->t_keepcnt = tcp_keepcnt; 717 tp->t_maxidle = tp->t_keepintvl * tp->t_keepcnt; 718 719 if (tcp_do_ncr) 720 tp->t_flags |= TF_NCR; 721 if (tcp_do_rfc1323) 722 tp->t_flags |= (TF_REQ_SCALE | TF_REQ_TSTMP); 723 724 tp->t_inpcb = inp; /* XXX */ 725 tp->t_state = TCPS_CLOSED; 726 /* 727 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 728 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 729 * reasonable initial retransmit time. 730 */ 731 tp->t_srtt = TCPTV_SRTTBASE; 732 tp->t_rttvar = 733 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 734 tp->t_rttmin = tcp_rexmit_min; 735 tp->t_rxtcur = TCPTV_RTOBASE; 736 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 737 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 738 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 739 tp->snd_last = ticks; 740 tp->t_rcvtime = ticks; 741 /* 742 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 743 * because the socket may be bound to an IPv6 wildcard address, 744 * which may match an IPv4-mapped IPv6 address. 745 */ 746 inp->inp_ip_ttl = ip_defttl; 747 inp->inp_ppcb = tp; 748 tcp_sack_tcpcb_init(tp); 749 750 tp->tt_sndmore = &it->inp_tp_sndmore; 751 tcp_output_init(tp); 752 753 return (tp); /* XXX */ 754 } 755 756 /* 757 * Drop a TCP connection, reporting the specified error. 758 * If connection is synchronized, then send a RST to peer. 759 */ 760 struct tcpcb * 761 tcp_drop(struct tcpcb *tp, int error) 762 { 763 struct socket *so = tp->t_inpcb->inp_socket; 764 765 if (TCPS_HAVERCVDSYN(tp->t_state)) { 766 tp->t_state = TCPS_CLOSED; 767 tcp_output(tp); 768 tcpstat.tcps_drops++; 769 } else 770 tcpstat.tcps_conndrops++; 771 if (error == ETIMEDOUT && tp->t_softerror) 772 error = tp->t_softerror; 773 so->so_error = error; 774 return (tcp_close(tp)); 775 } 776 777 struct netmsg_listen_detach { 778 struct netmsg_base base; 779 struct tcpcb *nm_tp; 780 struct tcpcb *nm_tp_inh; 781 }; 782 783 static void 784 tcp_listen_detach_handler(netmsg_t msg) 785 { 786 struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg; 787 struct tcpcb *tp = nmsg->nm_tp; 788 int cpu = mycpuid, nextcpu; 789 790 if (tp->t_flags & TF_LISTEN) 791 syncache_destroy(tp, nmsg->nm_tp_inh); 792 793 in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]); 794 795 nextcpu = cpu + 1; 796 if (nextcpu < ncpus2) 797 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nmsg->base.lmsg); 798 else 799 lwkt_replymsg(&nmsg->base.lmsg, 0); 800 } 801 802 /* 803 * Close a TCP control block: 804 * discard all space held by the tcp 805 * discard internet protocol block 806 * wake up any sleepers 807 */ 808 struct tcpcb * 809 tcp_close(struct tcpcb *tp) 810 { 811 struct tseg_qent *q; 812 struct inpcb *inp = tp->t_inpcb; 813 struct inpcb *inp_inh = NULL; 814 struct tcpcb *tp_inh = NULL; 815 struct socket *so = inp->inp_socket; 816 struct rtentry *rt; 817 boolean_t dosavessthresh; 818 #ifdef INET6 819 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 820 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0); 821 #else 822 const boolean_t isipv6 = FALSE; 823 #endif 824 825 if (tp->t_flags & TF_LISTEN) { 826 /* 827 * Pending socket/syncache inheritance 828 * 829 * If this is a listen(2) socket, find another listen(2) 830 * socket in the same local group, which could inherit 831 * the syncache and sockets pending on the completion 832 * and incompletion queues. 833 * 834 * NOTE: 835 * Currently the inheritance could only happen on the 836 * listen(2) sockets w/ SO_REUSEPORT set. 837 */ 838 KASSERT(&curthread->td_msgport == netisr_cpuport(0), 839 ("listen socket close not in netisr0")); 840 inp_inh = in_pcblocalgroup_last(&tcbinfo[0], inp); 841 if (inp_inh != NULL) 842 tp_inh = intotcpcb(inp_inh); 843 } 844 845 /* 846 * INP_WILDCARD_MP indicates that listen(2) has been called on 847 * this socket. This implies: 848 * - A wildcard inp's hash is replicated for each protocol thread. 849 * - Syncache for this inp grows independently in each protocol 850 * thread. 851 * - There is more than one cpu 852 * 853 * We have to chain a message to the rest of the protocol threads 854 * to cleanup the wildcard hash and the syncache. The cleanup 855 * in the current protocol thread is defered till the end of this 856 * function. 857 * 858 * NOTE: 859 * After cleanup the inp's hash and syncache entries, this inp will 860 * no longer be available to the rest of the protocol threads, so we 861 * are safe to whack the inp in the following code. 862 */ 863 if (inp->inp_flags & INP_WILDCARD_MP) { 864 struct netmsg_listen_detach nmsg; 865 866 KKASSERT(so->so_port == netisr_cpuport(0)); 867 KKASSERT(&curthread->td_msgport == netisr_cpuport(0)); 868 KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]); 869 870 netmsg_init(&nmsg.base, NULL, &curthread->td_msgport, 871 MSGF_PRIORITY, tcp_listen_detach_handler); 872 nmsg.nm_tp = tp; 873 nmsg.nm_tp_inh = tp_inh; 874 lwkt_domsg(netisr_cpuport(1), &nmsg.base.lmsg, 0); 875 876 inp->inp_flags &= ~INP_WILDCARD_MP; 877 } 878 879 KKASSERT(tp->t_state != TCPS_TERMINATING); 880 tp->t_state = TCPS_TERMINATING; 881 882 /* 883 * Make sure that all of our timers are stopped before we 884 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL), 885 * timers are never used. If timer message is never created 886 * (tp->tt_msg->tt_tcb == NULL), timers are never used too. 887 */ 888 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) { 889 tcp_callout_stop(tp, tp->tt_rexmt); 890 tcp_callout_stop(tp, tp->tt_persist); 891 tcp_callout_stop(tp, tp->tt_keep); 892 tcp_callout_stop(tp, tp->tt_2msl); 893 tcp_callout_stop(tp, tp->tt_delack); 894 } 895 896 if (tp->t_flags & TF_ONOUTPUTQ) { 897 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 898 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 899 tp->t_flags &= ~TF_ONOUTPUTQ; 900 } 901 902 /* 903 * If we got enough samples through the srtt filter, 904 * save the rtt and rttvar in the routing entry. 905 * 'Enough' is arbitrarily defined as the 16 samples. 906 * 16 samples is enough for the srtt filter to converge 907 * to within 5% of the correct value; fewer samples and 908 * we could save a very bogus rtt. 909 * 910 * Don't update the default route's characteristics and don't 911 * update anything that the user "locked". 912 */ 913 if (tp->t_rttupdated >= 16) { 914 u_long i = 0; 915 916 if (isipv6) { 917 struct sockaddr_in6 *sin6; 918 919 if ((rt = inp->in6p_route.ro_rt) == NULL) 920 goto no_valid_rt; 921 sin6 = (struct sockaddr_in6 *)rt_key(rt); 922 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 923 goto no_valid_rt; 924 } else 925 if ((rt = inp->inp_route.ro_rt) == NULL || 926 ((struct sockaddr_in *)rt_key(rt))-> 927 sin_addr.s_addr == INADDR_ANY) 928 goto no_valid_rt; 929 930 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 931 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 932 if (rt->rt_rmx.rmx_rtt && i) 933 /* 934 * filter this update to half the old & half 935 * the new values, converting scale. 936 * See route.h and tcp_var.h for a 937 * description of the scaling constants. 938 */ 939 rt->rt_rmx.rmx_rtt = 940 (rt->rt_rmx.rmx_rtt + i) / 2; 941 else 942 rt->rt_rmx.rmx_rtt = i; 943 tcpstat.tcps_cachedrtt++; 944 } 945 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 946 i = tp->t_rttvar * 947 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 948 if (rt->rt_rmx.rmx_rttvar && i) 949 rt->rt_rmx.rmx_rttvar = 950 (rt->rt_rmx.rmx_rttvar + i) / 2; 951 else 952 rt->rt_rmx.rmx_rttvar = i; 953 tcpstat.tcps_cachedrttvar++; 954 } 955 /* 956 * The old comment here said: 957 * update the pipelimit (ssthresh) if it has been updated 958 * already or if a pipesize was specified & the threshhold 959 * got below half the pipesize. I.e., wait for bad news 960 * before we start updating, then update on both good 961 * and bad news. 962 * 963 * But we want to save the ssthresh even if no pipesize is 964 * specified explicitly in the route, because such 965 * connections still have an implicit pipesize specified 966 * by the global tcp_sendspace. In the absence of a reliable 967 * way to calculate the pipesize, it will have to do. 968 */ 969 i = tp->snd_ssthresh; 970 if (rt->rt_rmx.rmx_sendpipe != 0) 971 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 972 else 973 dosavessthresh = (i < so->so_snd.ssb_hiwat/2); 974 if (dosavessthresh || 975 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 976 (rt->rt_rmx.rmx_ssthresh != 0))) { 977 /* 978 * convert the limit from user data bytes to 979 * packets then to packet data bytes. 980 */ 981 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 982 if (i < 2) 983 i = 2; 984 i *= tp->t_maxseg + 985 (isipv6 ? 986 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 987 sizeof(struct tcpiphdr)); 988 if (rt->rt_rmx.rmx_ssthresh) 989 rt->rt_rmx.rmx_ssthresh = 990 (rt->rt_rmx.rmx_ssthresh + i) / 2; 991 else 992 rt->rt_rmx.rmx_ssthresh = i; 993 tcpstat.tcps_cachedssthresh++; 994 } 995 } 996 997 no_valid_rt: 998 /* free the reassembly queue, if any */ 999 while((q = TAILQ_FIRST(&tp->t_segq)) != NULL) { 1000 TAILQ_REMOVE(&tp->t_segq, q, tqe_q); 1001 m_freem(q->tqe_m); 1002 kfree(q, M_TSEGQ); 1003 atomic_add_int(&tcp_reass_qsize, -1); 1004 } 1005 /* throw away SACK blocks in scoreboard*/ 1006 if (TCP_DO_SACK(tp)) 1007 tcp_sack_destroy(&tp->scb); 1008 1009 inp->inp_ppcb = NULL; 1010 soisdisconnected(so); 1011 /* note: pcb detached later on */ 1012 1013 tcp_destroy_timermsg(tp); 1014 tcp_output_cancel(tp); 1015 1016 if (tp->t_flags & TF_LISTEN) { 1017 syncache_destroy(tp, tp_inh); 1018 if (inp_inh != NULL && inp_inh->inp_socket != NULL) { 1019 /* 1020 * Pending sockets inheritance only needs 1021 * to be done once in the current thread, 1022 * i.e. netisr0. 1023 */ 1024 soinherit(so, inp_inh->inp_socket); 1025 } 1026 } 1027 1028 so_async_rcvd_drop(so); 1029 /* Drop the reference for the asynchronized pru_rcvd */ 1030 sofree(so); 1031 1032 /* 1033 * NOTE: 1034 * pcbdetach removes any wildcard hash entry on the current CPU. 1035 */ 1036 #ifdef INET6 1037 if (isafinet6) 1038 in6_pcbdetach(inp); 1039 else 1040 #endif 1041 in_pcbdetach(inp); 1042 1043 tcpstat.tcps_closed++; 1044 return (NULL); 1045 } 1046 1047 static __inline void 1048 tcp_drain_oncpu(struct inpcbhead *head) 1049 { 1050 struct inpcb *marker; 1051 struct inpcb *inpb; 1052 struct tcpcb *tcpb; 1053 struct tseg_qent *te; 1054 1055 /* 1056 * Allows us to block while running the list 1057 */ 1058 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1059 marker->inp_flags |= INP_PLACEMARKER; 1060 LIST_INSERT_HEAD(head, marker, inp_list); 1061 1062 while ((inpb = LIST_NEXT(marker, inp_list)) != NULL) { 1063 if ((inpb->inp_flags & INP_PLACEMARKER) == 0 && 1064 (tcpb = intotcpcb(inpb)) != NULL && 1065 (te = TAILQ_FIRST(&tcpb->t_segq)) != NULL) { 1066 TAILQ_REMOVE(&tcpb->t_segq, te, tqe_q); 1067 if (te->tqe_th->th_flags & TH_FIN) 1068 tcpb->t_flags &= ~TF_QUEDFIN; 1069 m_freem(te->tqe_m); 1070 kfree(te, M_TSEGQ); 1071 atomic_add_int(&tcp_reass_qsize, -1); 1072 /* retry */ 1073 } else { 1074 LIST_REMOVE(marker, inp_list); 1075 LIST_INSERT_AFTER(inpb, marker, inp_list); 1076 } 1077 } 1078 LIST_REMOVE(marker, inp_list); 1079 kfree(marker, M_TEMP); 1080 } 1081 1082 struct netmsg_tcp_drain { 1083 struct netmsg_base base; 1084 struct inpcbhead *nm_head; 1085 }; 1086 1087 static void 1088 tcp_drain_handler(netmsg_t msg) 1089 { 1090 struct netmsg_tcp_drain *nm = (void *)msg; 1091 1092 tcp_drain_oncpu(nm->nm_head); 1093 lwkt_replymsg(&nm->base.lmsg, 0); 1094 } 1095 1096 void 1097 tcp_drain(void) 1098 { 1099 int cpu; 1100 1101 if (!do_tcpdrain) 1102 return; 1103 1104 /* 1105 * Walk the tcpbs, if existing, and flush the reassembly queue, 1106 * if there is one... 1107 * XXX: The "Net/3" implementation doesn't imply that the TCP 1108 * reassembly queue should be flushed, but in a situation 1109 * where we're really low on mbufs, this is potentially 1110 * useful. 1111 */ 1112 for (cpu = 0; cpu < ncpus2; cpu++) { 1113 struct netmsg_tcp_drain *nm; 1114 1115 if (cpu == mycpu->gd_cpuid) { 1116 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead); 1117 } else { 1118 nm = kmalloc(sizeof(struct netmsg_tcp_drain), 1119 M_LWKTMSG, M_NOWAIT); 1120 if (nm == NULL) 1121 continue; 1122 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1123 0, tcp_drain_handler); 1124 nm->nm_head = &tcbinfo[cpu].pcblisthead; 1125 lwkt_sendmsg(netisr_cpuport(cpu), &nm->base.lmsg); 1126 } 1127 } 1128 } 1129 1130 /* 1131 * Notify a tcp user of an asynchronous error; 1132 * store error as soft error, but wake up user 1133 * (for now, won't do anything until can select for soft error). 1134 * 1135 * Do not wake up user since there currently is no mechanism for 1136 * reporting soft errors (yet - a kqueue filter may be added). 1137 */ 1138 static void 1139 tcp_notify(struct inpcb *inp, int error) 1140 { 1141 struct tcpcb *tp = intotcpcb(inp); 1142 1143 /* 1144 * Ignore some errors if we are hooked up. 1145 * If connection hasn't completed, has retransmitted several times, 1146 * and receives a second error, give up now. This is better 1147 * than waiting a long time to establish a connection that 1148 * can never complete. 1149 */ 1150 if (tp->t_state == TCPS_ESTABLISHED && 1151 (error == EHOSTUNREACH || error == ENETUNREACH || 1152 error == EHOSTDOWN)) { 1153 return; 1154 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1155 tp->t_softerror) 1156 tcp_drop(tp, error); 1157 else 1158 tp->t_softerror = error; 1159 #if 0 1160 wakeup(&so->so_timeo); 1161 sorwakeup(so); 1162 sowwakeup(so); 1163 #endif 1164 } 1165 1166 static int 1167 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1168 { 1169 int error, i, n; 1170 struct inpcb *marker; 1171 struct inpcb *inp; 1172 globaldata_t gd; 1173 int origcpu, ccpu; 1174 1175 error = 0; 1176 n = 0; 1177 1178 /* 1179 * The process of preparing the TCB list is too time-consuming and 1180 * resource-intensive to repeat twice on every request. 1181 */ 1182 if (req->oldptr == NULL) { 1183 for (ccpu = 0; ccpu < ncpus2; ++ccpu) { 1184 gd = globaldata_find(ccpu); 1185 n += tcbinfo[gd->gd_cpuid].ipi_count; 1186 } 1187 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1188 return (0); 1189 } 1190 1191 if (req->newptr != NULL) 1192 return (EPERM); 1193 1194 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1195 marker->inp_flags |= INP_PLACEMARKER; 1196 1197 /* 1198 * OK, now we're committed to doing something. Run the inpcb list 1199 * for each cpu in the system and construct the output. Use a 1200 * list placemarker to deal with list changes occuring during 1201 * copyout blockages (but otherwise depend on being on the correct 1202 * cpu to avoid races). 1203 */ 1204 origcpu = mycpu->gd_cpuid; 1205 for (ccpu = 0; ccpu < ncpus2 && error == 0; ++ccpu) { 1206 caddr_t inp_ppcb; 1207 struct xtcpcb xt; 1208 1209 lwkt_migratecpu(ccpu); 1210 1211 n = tcbinfo[ccpu].ipi_count; 1212 1213 LIST_INSERT_HEAD(&tcbinfo[ccpu].pcblisthead, marker, inp_list); 1214 i = 0; 1215 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1216 /* 1217 * process a snapshot of pcbs, ignoring placemarkers 1218 * and using our own to allow SYSCTL_OUT to block. 1219 */ 1220 LIST_REMOVE(marker, inp_list); 1221 LIST_INSERT_AFTER(inp, marker, inp_list); 1222 1223 if (inp->inp_flags & INP_PLACEMARKER) 1224 continue; 1225 if (prison_xinpcb(req->td, inp)) 1226 continue; 1227 1228 xt.xt_len = sizeof xt; 1229 bcopy(inp, &xt.xt_inp, sizeof *inp); 1230 inp_ppcb = inp->inp_ppcb; 1231 if (inp_ppcb != NULL) 1232 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1233 else 1234 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1235 if (inp->inp_socket) 1236 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1237 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1238 break; 1239 ++i; 1240 } 1241 LIST_REMOVE(marker, inp_list); 1242 if (error == 0 && i < n) { 1243 bzero(&xt, sizeof xt); 1244 xt.xt_len = sizeof xt; 1245 while (i < n) { 1246 error = SYSCTL_OUT(req, &xt, sizeof xt); 1247 if (error) 1248 break; 1249 ++i; 1250 } 1251 } 1252 } 1253 1254 /* 1255 * Make sure we are on the same cpu we were on originally, since 1256 * higher level callers expect this. Also don't pollute caches with 1257 * migrated userland data by (eventually) returning to userland 1258 * on a different cpu. 1259 */ 1260 lwkt_migratecpu(origcpu); 1261 kfree(marker, M_TEMP); 1262 return (error); 1263 } 1264 1265 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1266 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1267 1268 static int 1269 tcp_getcred(SYSCTL_HANDLER_ARGS) 1270 { 1271 struct sockaddr_in addrs[2]; 1272 struct inpcb *inp; 1273 int cpu; 1274 int error; 1275 1276 error = priv_check(req->td, PRIV_ROOT); 1277 if (error != 0) 1278 return (error); 1279 error = SYSCTL_IN(req, addrs, sizeof addrs); 1280 if (error != 0) 1281 return (error); 1282 crit_enter(); 1283 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1284 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1285 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1286 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1287 if (inp == NULL || inp->inp_socket == NULL) { 1288 error = ENOENT; 1289 goto out; 1290 } 1291 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1292 out: 1293 crit_exit(); 1294 return (error); 1295 } 1296 1297 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1298 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1299 1300 #ifdef INET6 1301 static int 1302 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1303 { 1304 struct sockaddr_in6 addrs[2]; 1305 struct inpcb *inp; 1306 int error; 1307 boolean_t mapped = FALSE; 1308 1309 error = priv_check(req->td, PRIV_ROOT); 1310 if (error != 0) 1311 return (error); 1312 error = SYSCTL_IN(req, addrs, sizeof addrs); 1313 if (error != 0) 1314 return (error); 1315 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { 1316 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) 1317 mapped = TRUE; 1318 else 1319 return (EINVAL); 1320 } 1321 crit_enter(); 1322 if (mapped) { 1323 inp = in_pcblookup_hash(&tcbinfo[0], 1324 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], 1325 addrs[1].sin6_port, 1326 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], 1327 addrs[0].sin6_port, 1328 0, NULL); 1329 } else { 1330 inp = in6_pcblookup_hash(&tcbinfo[0], 1331 &addrs[1].sin6_addr, addrs[1].sin6_port, 1332 &addrs[0].sin6_addr, addrs[0].sin6_port, 1333 0, NULL); 1334 } 1335 if (inp == NULL || inp->inp_socket == NULL) { 1336 error = ENOENT; 1337 goto out; 1338 } 1339 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1340 out: 1341 crit_exit(); 1342 return (error); 1343 } 1344 1345 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1346 0, 0, 1347 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1348 #endif 1349 1350 struct netmsg_tcp_notify { 1351 struct netmsg_base base; 1352 void (*nm_notify)(struct inpcb *, int); 1353 struct in_addr nm_faddr; 1354 int nm_arg; 1355 }; 1356 1357 static void 1358 tcp_notifyall_oncpu(netmsg_t msg) 1359 { 1360 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg; 1361 int nextcpu; 1362 1363 in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nm->nm_faddr, 1364 nm->nm_arg, nm->nm_notify); 1365 1366 nextcpu = mycpuid + 1; 1367 if (nextcpu < ncpus2) 1368 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nm->base.lmsg); 1369 else 1370 lwkt_replymsg(&nm->base.lmsg, 0); 1371 } 1372 1373 void 1374 tcp_ctlinput(netmsg_t msg) 1375 { 1376 int cmd = msg->ctlinput.nm_cmd; 1377 struct sockaddr *sa = msg->ctlinput.nm_arg; 1378 struct ip *ip = msg->ctlinput.nm_extra; 1379 struct tcphdr *th; 1380 struct in_addr faddr; 1381 struct inpcb *inp; 1382 struct tcpcb *tp; 1383 void (*notify)(struct inpcb *, int) = tcp_notify; 1384 tcp_seq icmpseq; 1385 int arg, cpu; 1386 1387 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) { 1388 goto done; 1389 } 1390 1391 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1392 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1393 goto done; 1394 1395 arg = inetctlerrmap[cmd]; 1396 if (cmd == PRC_QUENCH) { 1397 notify = tcp_quench; 1398 } else if (icmp_may_rst && 1399 (cmd == PRC_UNREACH_ADMIN_PROHIB || 1400 cmd == PRC_UNREACH_PORT || 1401 cmd == PRC_TIMXCEED_INTRANS) && 1402 ip != NULL) { 1403 notify = tcp_drop_syn_sent; 1404 } else if (cmd == PRC_MSGSIZE) { 1405 struct icmp *icmp = (struct icmp *) 1406 ((caddr_t)ip - offsetof(struct icmp, icmp_ip)); 1407 1408 arg = ntohs(icmp->icmp_nextmtu); 1409 notify = tcp_mtudisc; 1410 } else if (PRC_IS_REDIRECT(cmd)) { 1411 ip = NULL; 1412 notify = in_rtchange; 1413 } else if (cmd == PRC_HOSTDEAD) { 1414 ip = NULL; 1415 } 1416 1417 if (ip != NULL) { 1418 crit_enter(); 1419 th = (struct tcphdr *)((caddr_t)ip + 1420 (IP_VHL_HL(ip->ip_vhl) << 2)); 1421 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport, 1422 ip->ip_src.s_addr, th->th_sport); 1423 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport, 1424 ip->ip_src, th->th_sport, 0, NULL); 1425 if ((inp != NULL) && (inp->inp_socket != NULL)) { 1426 icmpseq = htonl(th->th_seq); 1427 tp = intotcpcb(inp); 1428 if (SEQ_GEQ(icmpseq, tp->snd_una) && 1429 SEQ_LT(icmpseq, tp->snd_max)) 1430 (*notify)(inp, arg); 1431 } else { 1432 struct in_conninfo inc; 1433 1434 inc.inc_fport = th->th_dport; 1435 inc.inc_lport = th->th_sport; 1436 inc.inc_faddr = faddr; 1437 inc.inc_laddr = ip->ip_src; 1438 #ifdef INET6 1439 inc.inc_isipv6 = 0; 1440 #endif 1441 syncache_unreach(&inc, th); 1442 } 1443 crit_exit(); 1444 } else { 1445 struct netmsg_tcp_notify *nm; 1446 1447 KKASSERT(&curthread->td_msgport == netisr_cpuport(0)); 1448 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT); 1449 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1450 0, tcp_notifyall_oncpu); 1451 nm->nm_faddr = faddr; 1452 nm->nm_arg = arg; 1453 nm->nm_notify = notify; 1454 1455 lwkt_sendmsg(netisr_cpuport(0), &nm->base.lmsg); 1456 } 1457 done: 1458 lwkt_replymsg(&msg->lmsg, 0); 1459 } 1460 1461 #ifdef INET6 1462 1463 void 1464 tcp6_ctlinput(netmsg_t msg) 1465 { 1466 int cmd = msg->ctlinput.nm_cmd; 1467 struct sockaddr *sa = msg->ctlinput.nm_arg; 1468 void *d = msg->ctlinput.nm_extra; 1469 struct tcphdr th; 1470 void (*notify) (struct inpcb *, int) = tcp_notify; 1471 struct ip6_hdr *ip6; 1472 struct mbuf *m; 1473 struct ip6ctlparam *ip6cp = NULL; 1474 const struct sockaddr_in6 *sa6_src = NULL; 1475 int off; 1476 struct tcp_portonly { 1477 u_int16_t th_sport; 1478 u_int16_t th_dport; 1479 } *thp; 1480 int arg; 1481 1482 if (sa->sa_family != AF_INET6 || 1483 sa->sa_len != sizeof(struct sockaddr_in6)) { 1484 goto out; 1485 } 1486 1487 arg = 0; 1488 if (cmd == PRC_QUENCH) 1489 notify = tcp_quench; 1490 else if (cmd == PRC_MSGSIZE) { 1491 struct ip6ctlparam *ip6cp = d; 1492 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6; 1493 1494 arg = ntohl(icmp6->icmp6_mtu); 1495 notify = tcp_mtudisc; 1496 } else if (!PRC_IS_REDIRECT(cmd) && 1497 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) { 1498 goto out; 1499 } 1500 1501 /* if the parameter is from icmp6, decode it. */ 1502 if (d != NULL) { 1503 ip6cp = (struct ip6ctlparam *)d; 1504 m = ip6cp->ip6c_m; 1505 ip6 = ip6cp->ip6c_ip6; 1506 off = ip6cp->ip6c_off; 1507 sa6_src = ip6cp->ip6c_src; 1508 } else { 1509 m = NULL; 1510 ip6 = NULL; 1511 off = 0; /* fool gcc */ 1512 sa6_src = &sa6_any; 1513 } 1514 1515 if (ip6 != NULL) { 1516 struct in_conninfo inc; 1517 /* 1518 * XXX: We assume that when IPV6 is non NULL, 1519 * M and OFF are valid. 1520 */ 1521 1522 /* check if we can safely examine src and dst ports */ 1523 if (m->m_pkthdr.len < off + sizeof *thp) 1524 goto out; 1525 1526 bzero(&th, sizeof th); 1527 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1528 1529 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport, 1530 (struct sockaddr *)ip6cp->ip6c_src, 1531 th.th_sport, cmd, arg, notify); 1532 1533 inc.inc_fport = th.th_dport; 1534 inc.inc_lport = th.th_sport; 1535 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1536 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1537 inc.inc_isipv6 = 1; 1538 syncache_unreach(&inc, &th); 1539 } else { 1540 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0, 1541 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify); 1542 } 1543 out: 1544 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0); 1545 } 1546 1547 #endif 1548 1549 /* 1550 * Following is where TCP initial sequence number generation occurs. 1551 * 1552 * There are two places where we must use initial sequence numbers: 1553 * 1. In SYN-ACK packets. 1554 * 2. In SYN packets. 1555 * 1556 * All ISNs for SYN-ACK packets are generated by the syncache. See 1557 * tcp_syncache.c for details. 1558 * 1559 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1560 * depends on this property. In addition, these ISNs should be 1561 * unguessable so as to prevent connection hijacking. To satisfy 1562 * the requirements of this situation, the algorithm outlined in 1563 * RFC 1948 is used to generate sequence numbers. 1564 * 1565 * Implementation details: 1566 * 1567 * Time is based off the system timer, and is corrected so that it 1568 * increases by one megabyte per second. This allows for proper 1569 * recycling on high speed LANs while still leaving over an hour 1570 * before rollover. 1571 * 1572 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1573 * between seeding of isn_secret. This is normally set to zero, 1574 * as reseeding should not be necessary. 1575 * 1576 */ 1577 1578 #define ISN_BYTES_PER_SECOND 1048576 1579 1580 u_char isn_secret[32]; 1581 int isn_last_reseed; 1582 MD5_CTX isn_ctx; 1583 1584 tcp_seq 1585 tcp_new_isn(struct tcpcb *tp) 1586 { 1587 u_int32_t md5_buffer[4]; 1588 tcp_seq new_isn; 1589 1590 /* Seed if this is the first use, reseed if requested. */ 1591 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1592 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1593 < (u_int)ticks))) { 1594 read_random_unlimited(&isn_secret, sizeof isn_secret); 1595 isn_last_reseed = ticks; 1596 } 1597 1598 /* Compute the md5 hash and return the ISN. */ 1599 MD5Init(&isn_ctx); 1600 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1601 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1602 #ifdef INET6 1603 if (tp->t_inpcb->inp_vflag & INP_IPV6) { 1604 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1605 sizeof(struct in6_addr)); 1606 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1607 sizeof(struct in6_addr)); 1608 } else 1609 #endif 1610 { 1611 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1612 sizeof(struct in_addr)); 1613 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1614 sizeof(struct in_addr)); 1615 } 1616 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1617 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1618 new_isn = (tcp_seq) md5_buffer[0]; 1619 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1620 return (new_isn); 1621 } 1622 1623 /* 1624 * When a source quench is received, close congestion window 1625 * to one segment. We will gradually open it again as we proceed. 1626 */ 1627 void 1628 tcp_quench(struct inpcb *inp, int error) 1629 { 1630 struct tcpcb *tp = intotcpcb(inp); 1631 1632 if (tp != NULL) { 1633 tp->snd_cwnd = tp->t_maxseg; 1634 tp->snd_wacked = 0; 1635 } 1636 } 1637 1638 /* 1639 * When a specific ICMP unreachable message is received and the 1640 * connection state is SYN-SENT, drop the connection. This behavior 1641 * is controlled by the icmp_may_rst sysctl. 1642 */ 1643 void 1644 tcp_drop_syn_sent(struct inpcb *inp, int error) 1645 { 1646 struct tcpcb *tp = intotcpcb(inp); 1647 1648 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT)) 1649 tcp_drop(tp, error); 1650 } 1651 1652 /* 1653 * When a `need fragmentation' ICMP is received, update our idea of the MSS 1654 * based on the new value in the route. Also nudge TCP to send something, 1655 * since we know the packet we just sent was dropped. 1656 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1657 */ 1658 void 1659 tcp_mtudisc(struct inpcb *inp, int mtu) 1660 { 1661 struct tcpcb *tp = intotcpcb(inp); 1662 struct rtentry *rt; 1663 struct socket *so = inp->inp_socket; 1664 int maxopd, mss; 1665 #ifdef INET6 1666 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0); 1667 #else 1668 const boolean_t isipv6 = FALSE; 1669 #endif 1670 1671 if (tp == NULL) 1672 return; 1673 1674 /* 1675 * If no MTU is provided in the ICMP message, use the 1676 * next lower likely value, as specified in RFC 1191. 1677 */ 1678 if (mtu == 0) { 1679 int oldmtu; 1680 1681 oldmtu = tp->t_maxopd + 1682 (isipv6 ? 1683 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1684 sizeof(struct tcpiphdr)); 1685 mtu = ip_next_mtu(oldmtu, 0); 1686 } 1687 1688 if (isipv6) 1689 rt = tcp_rtlookup6(&inp->inp_inc); 1690 else 1691 rt = tcp_rtlookup(&inp->inp_inc); 1692 if (rt != NULL) { 1693 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu) 1694 mtu = rt->rt_rmx.rmx_mtu; 1695 1696 maxopd = mtu - 1697 (isipv6 ? 1698 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1699 sizeof(struct tcpiphdr)); 1700 1701 /* 1702 * XXX - The following conditional probably violates the TCP 1703 * spec. The problem is that, since we don't know the 1704 * other end's MSS, we are supposed to use a conservative 1705 * default. But, if we do that, then MTU discovery will 1706 * never actually take place, because the conservative 1707 * default is much less than the MTUs typically seen 1708 * on the Internet today. For the moment, we'll sweep 1709 * this under the carpet. 1710 * 1711 * The conservative default might not actually be a problem 1712 * if the only case this occurs is when sending an initial 1713 * SYN with options and data to a host we've never talked 1714 * to before. Then, they will reply with an MSS value which 1715 * will get recorded and the new parameters should get 1716 * recomputed. For Further Study. 1717 */ 1718 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd) 1719 maxopd = rt->rt_rmx.rmx_mssopt; 1720 } else 1721 maxopd = mtu - 1722 (isipv6 ? 1723 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1724 sizeof(struct tcpiphdr)); 1725 1726 if (tp->t_maxopd <= maxopd) 1727 return; 1728 tp->t_maxopd = maxopd; 1729 1730 mss = maxopd; 1731 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) == 1732 (TF_REQ_TSTMP | TF_RCVD_TSTMP)) 1733 mss -= TCPOLEN_TSTAMP_APPA; 1734 1735 /* round down to multiple of MCLBYTES */ 1736 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */ 1737 if (mss > MCLBYTES) 1738 mss &= ~(MCLBYTES - 1); 1739 #else 1740 if (mss > MCLBYTES) 1741 mss = (mss / MCLBYTES) * MCLBYTES; 1742 #endif 1743 1744 if (so->so_snd.ssb_hiwat < mss) 1745 mss = so->so_snd.ssb_hiwat; 1746 1747 tp->t_maxseg = mss; 1748 tp->t_rtttime = 0; 1749 tp->snd_nxt = tp->snd_una; 1750 tcp_output(tp); 1751 tcpstat.tcps_mturesent++; 1752 } 1753 1754 /* 1755 * Look-up the routing entry to the peer of this inpcb. If no route 1756 * is found and it cannot be allocated the return NULL. This routine 1757 * is called by TCP routines that access the rmx structure and by tcp_mss 1758 * to get the interface MTU. 1759 */ 1760 struct rtentry * 1761 tcp_rtlookup(struct in_conninfo *inc) 1762 { 1763 struct route *ro = &inc->inc_route; 1764 1765 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1766 /* No route yet, so try to acquire one */ 1767 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1768 /* 1769 * unused portions of the structure MUST be zero'd 1770 * out because rtalloc() treats it as opaque data 1771 */ 1772 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1773 ro->ro_dst.sa_family = AF_INET; 1774 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1775 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1776 inc->inc_faddr; 1777 rtalloc(ro); 1778 } 1779 } 1780 return (ro->ro_rt); 1781 } 1782 1783 #ifdef INET6 1784 struct rtentry * 1785 tcp_rtlookup6(struct in_conninfo *inc) 1786 { 1787 struct route_in6 *ro6 = &inc->inc6_route; 1788 1789 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1790 /* No route yet, so try to acquire one */ 1791 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1792 /* 1793 * unused portions of the structure MUST be zero'd 1794 * out because rtalloc() treats it as opaque data 1795 */ 1796 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1797 ro6->ro_dst.sin6_family = AF_INET6; 1798 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1799 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1800 rtalloc((struct route *)ro6); 1801 } 1802 } 1803 return (ro6->ro_rt); 1804 } 1805 #endif 1806 1807 #ifdef IPSEC 1808 /* compute ESP/AH header size for TCP, including outer IP header. */ 1809 size_t 1810 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1811 { 1812 struct inpcb *inp; 1813 struct mbuf *m; 1814 size_t hdrsiz; 1815 struct ip *ip; 1816 struct tcphdr *th; 1817 1818 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1819 return (0); 1820 MGETHDR(m, MB_DONTWAIT, MT_DATA); 1821 if (!m) 1822 return (0); 1823 1824 #ifdef INET6 1825 if (inp->inp_vflag & INP_IPV6) { 1826 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1827 1828 th = (struct tcphdr *)(ip6 + 1); 1829 m->m_pkthdr.len = m->m_len = 1830 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1831 tcp_fillheaders(tp, ip6, th, FALSE); 1832 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1833 } else 1834 #endif 1835 { 1836 ip = mtod(m, struct ip *); 1837 th = (struct tcphdr *)(ip + 1); 1838 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1839 tcp_fillheaders(tp, ip, th, FALSE); 1840 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1841 } 1842 1843 m_free(m); 1844 return (hdrsiz); 1845 } 1846 #endif 1847 1848 /* 1849 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1850 * 1851 * This code attempts to calculate the bandwidth-delay product as a 1852 * means of determining the optimal window size to maximize bandwidth, 1853 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1854 * routers. This code also does a fairly good job keeping RTTs in check 1855 * across slow links like modems. We implement an algorithm which is very 1856 * similar (but not meant to be) TCP/Vegas. The code operates on the 1857 * transmitter side of a TCP connection and so only effects the transmit 1858 * side of the connection. 1859 * 1860 * BACKGROUND: TCP makes no provision for the management of buffer space 1861 * at the end points or at the intermediate routers and switches. A TCP 1862 * stream, whether using NewReno or not, will eventually buffer as 1863 * many packets as it is able and the only reason this typically works is 1864 * due to the fairly small default buffers made available for a connection 1865 * (typicaly 16K or 32K). As machines use larger windows and/or window 1866 * scaling it is now fairly easy for even a single TCP connection to blow-out 1867 * all available buffer space not only on the local interface, but on 1868 * intermediate routers and switches as well. NewReno makes a misguided 1869 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1870 * then backing off, then steadily increasing the window again until another 1871 * failure occurs, ad-infinitum. This results in terrible oscillation that 1872 * is only made worse as network loads increase and the idea of intentionally 1873 * blowing out network buffers is, frankly, a terrible way to manage network 1874 * resources. 1875 * 1876 * It is far better to limit the transmit window prior to the failure 1877 * condition being achieved. There are two general ways to do this: First 1878 * you can 'scan' through different transmit window sizes and locate the 1879 * point where the RTT stops increasing, indicating that you have filled the 1880 * pipe, then scan backwards until you note that RTT stops decreasing, then 1881 * repeat ad-infinitum. This method works in principle but has severe 1882 * implementation issues due to RTT variances, timer granularity, and 1883 * instability in the algorithm which can lead to many false positives and 1884 * create oscillations as well as interact badly with other TCP streams 1885 * implementing the same algorithm. 1886 * 1887 * The second method is to limit the window to the bandwidth delay product 1888 * of the link. This is the method we implement. RTT variances and our 1889 * own manipulation of the congestion window, bwnd, can potentially 1890 * destabilize the algorithm. For this reason we have to stabilize the 1891 * elements used to calculate the window. We do this by using the minimum 1892 * observed RTT, the long term average of the observed bandwidth, and 1893 * by adding two segments worth of slop. It isn't perfect but it is able 1894 * to react to changing conditions and gives us a very stable basis on 1895 * which to extend the algorithm. 1896 */ 1897 void 1898 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1899 { 1900 u_long bw; 1901 u_long bwnd; 1902 int save_ticks; 1903 int delta_ticks; 1904 1905 /* 1906 * If inflight_enable is disabled in the middle of a tcp connection, 1907 * make sure snd_bwnd is effectively disabled. 1908 */ 1909 if (!tcp_inflight_enable) { 1910 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1911 tp->snd_bandwidth = 0; 1912 return; 1913 } 1914 1915 /* 1916 * Validate the delta time. If a connection is new or has been idle 1917 * a long time we have to reset the bandwidth calculator. 1918 */ 1919 save_ticks = ticks; 1920 delta_ticks = save_ticks - tp->t_bw_rtttime; 1921 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1922 tp->t_bw_rtttime = ticks; 1923 tp->t_bw_rtseq = ack_seq; 1924 if (tp->snd_bandwidth == 0) 1925 tp->snd_bandwidth = tcp_inflight_min; 1926 return; 1927 } 1928 if (delta_ticks == 0) 1929 return; 1930 1931 /* 1932 * Sanity check, plus ignore pure window update acks. 1933 */ 1934 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1935 return; 1936 1937 /* 1938 * Figure out the bandwidth. Due to the tick granularity this 1939 * is a very rough number and it MUST be averaged over a fairly 1940 * long period of time. XXX we need to take into account a link 1941 * that is not using all available bandwidth, but for now our 1942 * slop will ramp us up if this case occurs and the bandwidth later 1943 * increases. 1944 */ 1945 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 1946 tp->t_bw_rtttime = save_ticks; 1947 tp->t_bw_rtseq = ack_seq; 1948 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4; 1949 1950 tp->snd_bandwidth = bw; 1951 1952 /* 1953 * Calculate the semi-static bandwidth delay product, plus two maximal 1954 * segments. The additional slop puts us squarely in the sweet 1955 * spot and also handles the bandwidth run-up case. Without the 1956 * slop we could be locking ourselves into a lower bandwidth. 1957 * 1958 * Situations Handled: 1959 * (1) Prevents over-queueing of packets on LANs, especially on 1960 * high speed LANs, allowing larger TCP buffers to be 1961 * specified, and also does a good job preventing 1962 * over-queueing of packets over choke points like modems 1963 * (at least for the transmit side). 1964 * 1965 * (2) Is able to handle changing network loads (bandwidth 1966 * drops so bwnd drops, bandwidth increases so bwnd 1967 * increases). 1968 * 1969 * (3) Theoretically should stabilize in the face of multiple 1970 * connections implementing the same algorithm (this may need 1971 * a little work). 1972 * 1973 * (4) Stability value (defaults to 20 = 2 maximal packets) can 1974 * be adjusted with a sysctl but typically only needs to be on 1975 * very slow connections. A value no smaller then 5 should 1976 * be used, but only reduce this default if you have no other 1977 * choice. 1978 */ 1979 1980 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2) 1981 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 1982 tcp_inflight_stab * (int)tp->t_maxseg / 10; 1983 #undef USERTT 1984 1985 if (tcp_inflight_debug > 0) { 1986 static int ltime; 1987 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) { 1988 ltime = ticks; 1989 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n", 1990 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd); 1991 } 1992 } 1993 if ((long)bwnd < tcp_inflight_min) 1994 bwnd = tcp_inflight_min; 1995 if (bwnd > tcp_inflight_max) 1996 bwnd = tcp_inflight_max; 1997 if ((long)bwnd < tp->t_maxseg * 2) 1998 bwnd = tp->t_maxseg * 2; 1999 tp->snd_bwnd = bwnd; 2000 } 2001 2002 static void 2003 tcp_rmx_iwsegs(struct tcpcb *tp, u_long *maxsegs, u_long *capsegs) 2004 { 2005 struct rtentry *rt; 2006 struct inpcb *inp = tp->t_inpcb; 2007 #ifdef INET6 2008 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE); 2009 #else 2010 const boolean_t isipv6 = FALSE; 2011 #endif 2012 2013 /* XXX */ 2014 if (tcp_iw_maxsegs < TCP_IW_MAXSEGS_DFLT) 2015 tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 2016 if (tcp_iw_capsegs < TCP_IW_CAPSEGS_DFLT) 2017 tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 2018 2019 if (isipv6) 2020 rt = tcp_rtlookup6(&inp->inp_inc); 2021 else 2022 rt = tcp_rtlookup(&inp->inp_inc); 2023 if (rt == NULL || 2024 rt->rt_rmx.rmx_iwmaxsegs < TCP_IW_MAXSEGS_DFLT || 2025 rt->rt_rmx.rmx_iwcapsegs < TCP_IW_CAPSEGS_DFLT) { 2026 *maxsegs = tcp_iw_maxsegs; 2027 *capsegs = tcp_iw_capsegs; 2028 return; 2029 } 2030 *maxsegs = rt->rt_rmx.rmx_iwmaxsegs; 2031 *capsegs = rt->rt_rmx.rmx_iwcapsegs; 2032 } 2033 2034 u_long 2035 tcp_initial_window(struct tcpcb *tp) 2036 { 2037 if (tcp_do_rfc3390) { 2038 /* 2039 * RFC3390: 2040 * "If the SYN or SYN/ACK is lost, the initial window 2041 * used by a sender after a correctly transmitted SYN 2042 * MUST be one segment consisting of MSS bytes." 2043 * 2044 * However, we do something a little bit more aggressive 2045 * then RFC3390 here: 2046 * - Only if time spent in the SYN or SYN|ACK retransmition 2047 * >= 3 seconds, the IW is reduced. We do this mainly 2048 * because when RFC3390 is published, the initial RTO is 2049 * still 3 seconds (the threshold we test here), while 2050 * after RFC6298, the initial RTO is 1 second. This 2051 * behaviour probably still falls within the spirit of 2052 * RFC3390. 2053 * - When IW is reduced, 2*MSS is used instead of 1*MSS. 2054 * Mainly to avoid sender and receiver deadlock until 2055 * delayed ACK timer expires. And even RFC2581 does not 2056 * try to reduce IW upon SYN or SYN|ACK retransmition 2057 * timeout. 2058 * 2059 * See also: 2060 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03 2061 */ 2062 if (tp->t_rxtsyn >= TCPTV_RTOBASE3) { 2063 return (2 * tp->t_maxseg); 2064 } else { 2065 u_long maxsegs, capsegs; 2066 2067 tcp_rmx_iwsegs(tp, &maxsegs, &capsegs); 2068 return min(maxsegs * tp->t_maxseg, 2069 max(2 * tp->t_maxseg, capsegs * 1460)); 2070 } 2071 } else { 2072 /* 2073 * Even RFC2581 (back to 1999) allows 2*SMSS IW. 2074 * 2075 * Mainly to avoid sender and receiver deadlock 2076 * until delayed ACK timer expires. 2077 */ 2078 return (2 * tp->t_maxseg); 2079 } 2080 } 2081 2082 #ifdef TCP_SIGNATURE 2083 /* 2084 * Compute TCP-MD5 hash of a TCP segment. (RFC2385) 2085 * 2086 * We do this over ip, tcphdr, segment data, and the key in the SADB. 2087 * When called from tcp_input(), we can be sure that th_sum has been 2088 * zeroed out and verified already. 2089 * 2090 * Return 0 if successful, otherwise return -1. 2091 * 2092 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a 2093 * search with the destination IP address, and a 'magic SPI' to be 2094 * determined by the application. This is hardcoded elsewhere to 1179 2095 * right now. Another branch of this code exists which uses the SPD to 2096 * specify per-application flows but it is unstable. 2097 */ 2098 int 2099 tcpsignature_compute( 2100 struct mbuf *m, /* mbuf chain */ 2101 int len, /* length of TCP data */ 2102 int optlen, /* length of TCP options */ 2103 u_char *buf, /* storage for MD5 digest */ 2104 u_int direction) /* direction of flow */ 2105 { 2106 struct ippseudo ippseudo; 2107 MD5_CTX ctx; 2108 int doff; 2109 struct ip *ip; 2110 struct ipovly *ipovly; 2111 struct secasvar *sav; 2112 struct tcphdr *th; 2113 #ifdef INET6 2114 struct ip6_hdr *ip6; 2115 struct in6_addr in6; 2116 uint32_t plen; 2117 uint16_t nhdr; 2118 #endif /* INET6 */ 2119 u_short savecsum; 2120 2121 KASSERT(m != NULL, ("passed NULL mbuf. Game over.")); 2122 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature")); 2123 /* 2124 * Extract the destination from the IP header in the mbuf. 2125 */ 2126 ip = mtod(m, struct ip *); 2127 #ifdef INET6 2128 ip6 = NULL; /* Make the compiler happy. */ 2129 #endif /* INET6 */ 2130 /* 2131 * Look up an SADB entry which matches the address found in 2132 * the segment. 2133 */ 2134 switch (IP_VHL_V(ip->ip_vhl)) { 2135 case IPVERSION: 2136 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst, 2137 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2138 break; 2139 #ifdef INET6 2140 case (IPV6_VERSION >> 4): 2141 ip6 = mtod(m, struct ip6_hdr *); 2142 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst, 2143 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2144 break; 2145 #endif /* INET6 */ 2146 default: 2147 return (EINVAL); 2148 /* NOTREACHED */ 2149 break; 2150 } 2151 if (sav == NULL) { 2152 kprintf("%s: SADB lookup failed\n", __func__); 2153 return (EINVAL); 2154 } 2155 MD5Init(&ctx); 2156 2157 /* 2158 * Step 1: Update MD5 hash with IP pseudo-header. 2159 * 2160 * XXX The ippseudo header MUST be digested in network byte order, 2161 * or else we'll fail the regression test. Assume all fields we've 2162 * been doing arithmetic on have been in host byte order. 2163 * XXX One cannot depend on ipovly->ih_len here. When called from 2164 * tcp_output(), the underlying ip_len member has not yet been set. 2165 */ 2166 switch (IP_VHL_V(ip->ip_vhl)) { 2167 case IPVERSION: 2168 ipovly = (struct ipovly *)ip; 2169 ippseudo.ippseudo_src = ipovly->ih_src; 2170 ippseudo.ippseudo_dst = ipovly->ih_dst; 2171 ippseudo.ippseudo_pad = 0; 2172 ippseudo.ippseudo_p = IPPROTO_TCP; 2173 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen); 2174 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo)); 2175 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip)); 2176 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen; 2177 break; 2178 #ifdef INET6 2179 /* 2180 * RFC 2385, 2.0 Proposal 2181 * For IPv6, the pseudo-header is as described in RFC 2460, namely the 2182 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero- 2183 * extended next header value (to form 32 bits), and 32-bit segment 2184 * length. 2185 * Note: Upper-Layer Packet Length comes before Next Header. 2186 */ 2187 case (IPV6_VERSION >> 4): 2188 in6 = ip6->ip6_src; 2189 in6_clearscope(&in6); 2190 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2191 in6 = ip6->ip6_dst; 2192 in6_clearscope(&in6); 2193 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2194 plen = htonl(len + sizeof(struct tcphdr) + optlen); 2195 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t)); 2196 nhdr = 0; 2197 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2198 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2199 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2200 nhdr = IPPROTO_TCP; 2201 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2202 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr)); 2203 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen; 2204 break; 2205 #endif /* INET6 */ 2206 default: 2207 return (EINVAL); 2208 /* NOTREACHED */ 2209 break; 2210 } 2211 /* 2212 * Step 2: Update MD5 hash with TCP header, excluding options. 2213 * The TCP checksum must be set to zero. 2214 */ 2215 savecsum = th->th_sum; 2216 th->th_sum = 0; 2217 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr)); 2218 th->th_sum = savecsum; 2219 /* 2220 * Step 3: Update MD5 hash with TCP segment data. 2221 * Use m_apply() to avoid an early m_pullup(). 2222 */ 2223 if (len > 0) 2224 m_apply(m, doff, len, tcpsignature_apply, &ctx); 2225 /* 2226 * Step 4: Update MD5 hash with shared secret. 2227 */ 2228 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth)); 2229 MD5Final(buf, &ctx); 2230 key_sa_recordxfer(sav, m); 2231 key_freesav(sav); 2232 return (0); 2233 } 2234 2235 int 2236 tcpsignature_apply(void *fstate, void *data, unsigned int len) 2237 { 2238 2239 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len); 2240 return (0); 2241 } 2242 #endif /* TCP_SIGNATURE */ 2243