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. All advertising materials mentioning features or use of this software 47 * must display the following acknowledgement: 48 * This product includes software developed by the University of 49 * California, Berkeley and its contributors. 50 * 4. Neither the name of the University nor the names of its contributors 51 * may be used to endorse or promote products derived from this software 52 * without specific prior written permission. 53 * 54 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 55 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 56 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 57 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 58 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 59 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 60 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 61 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 62 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 63 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 64 * SUCH DAMAGE. 65 * 66 * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95 67 * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $ 68 */ 69 70 #include "opt_compat.h" 71 #include "opt_inet.h" 72 #include "opt_inet6.h" 73 #include "opt_ipsec.h" 74 #include "opt_tcpdebug.h" 75 76 #include <sys/param.h> 77 #include <sys/systm.h> 78 #include <sys/callout.h> 79 #include <sys/kernel.h> 80 #include <sys/sysctl.h> 81 #include <sys/malloc.h> 82 #include <sys/mpipe.h> 83 #include <sys/mbuf.h> 84 #ifdef INET6 85 #include <sys/domain.h> 86 #endif 87 #include <sys/proc.h> 88 #include <sys/priv.h> 89 #include <sys/socket.h> 90 #include <sys/socketvar.h> 91 #include <sys/protosw.h> 92 #include <sys/random.h> 93 #include <sys/in_cksum.h> 94 #include <sys/ktr.h> 95 96 #include <net/route.h> 97 #include <net/if.h> 98 #include <net/netisr.h> 99 100 #define _IP_VHL 101 #include <netinet/in.h> 102 #include <netinet/in_systm.h> 103 #include <netinet/ip.h> 104 #include <netinet/ip6.h> 105 #include <netinet/in_pcb.h> 106 #include <netinet6/in6_pcb.h> 107 #include <netinet/in_var.h> 108 #include <netinet/ip_var.h> 109 #include <netinet6/ip6_var.h> 110 #include <netinet/ip_icmp.h> 111 #ifdef INET6 112 #include <netinet/icmp6.h> 113 #endif 114 #include <netinet/tcp.h> 115 #include <netinet/tcp_fsm.h> 116 #include <netinet/tcp_seq.h> 117 #include <netinet/tcp_timer.h> 118 #include <netinet/tcp_timer2.h> 119 #include <netinet/tcp_var.h> 120 #include <netinet6/tcp6_var.h> 121 #include <netinet/tcpip.h> 122 #ifdef TCPDEBUG 123 #include <netinet/tcp_debug.h> 124 #endif 125 #include <netinet6/ip6protosw.h> 126 127 #ifdef IPSEC 128 #include <netinet6/ipsec.h> 129 #include <netproto/key/key.h> 130 #ifdef INET6 131 #include <netinet6/ipsec6.h> 132 #endif 133 #endif 134 135 #ifdef FAST_IPSEC 136 #include <netproto/ipsec/ipsec.h> 137 #ifdef INET6 138 #include <netproto/ipsec/ipsec6.h> 139 #endif 140 #define IPSEC 141 #endif 142 143 #include <sys/md5.h> 144 #include <machine/smp.h> 145 146 #include <sys/msgport2.h> 147 #include <sys/mplock2.h> 148 #include <net/netmsg2.h> 149 150 #if !defined(KTR_TCP) 151 #define KTR_TCP KTR_ALL 152 #endif 153 /* 154 KTR_INFO_MASTER(tcp); 155 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0); 156 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0); 157 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0); 158 #define logtcp(name) KTR_LOG(tcp_ ## name) 159 */ 160 161 #define TCP_IW_MAXSEGS_DFLT 4 162 #define TCP_IW_CAPSEGS_DFLT 3 163 164 struct inpcbinfo tcbinfo[MAXCPU]; 165 struct tcpcbackqhead tcpcbackq[MAXCPU]; 166 167 static struct lwkt_token tcp_port_token = 168 LWKT_TOKEN_INITIALIZER(tcp_port_token); 169 170 int tcp_mssdflt = TCP_MSS; 171 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 172 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size"); 173 174 #ifdef INET6 175 int tcp_v6mssdflt = TCP6_MSS; 176 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW, 177 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6"); 178 #endif 179 180 /* 181 * Minimum MSS we accept and use. This prevents DoS attacks where 182 * we are forced to a ridiculous low MSS like 20 and send hundreds 183 * of packets instead of one. The effect scales with the available 184 * bandwidth and quickly saturates the CPU and network interface 185 * with packet generation and sending. Set to zero to disable MINMSS 186 * checking. This setting prevents us from sending too small packets. 187 */ 188 int tcp_minmss = TCP_MINMSS; 189 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW, 190 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size"); 191 192 #if 0 193 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ; 194 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW, 195 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time"); 196 #endif 197 198 int tcp_do_rfc1323 = 1; 199 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW, 200 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions"); 201 202 static int tcp_tcbhashsize = 0; 203 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD, 204 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable"); 205 206 static int do_tcpdrain = 1; 207 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0, 208 "Enable tcp_drain routine for extra help when low on mbufs"); 209 210 static int icmp_may_rst = 1; 211 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0, 212 "Certain ICMP unreachable messages may abort connections in SYN_SENT"); 213 214 static int tcp_isn_reseed_interval = 0; 215 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW, 216 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret"); 217 218 /* 219 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on 220 * by default, but with generous values which should allow maximal 221 * bandwidth. In particular, the slop defaults to 50 (5 packets). 222 * 223 * The reason for doing this is that the limiter is the only mechanism we 224 * have which seems to do a really good job preventing receiver RX rings 225 * on network interfaces from getting blown out. Even though GigE/10GigE 226 * is supposed to flow control it looks like either it doesn't actually 227 * do it or Open Source drivers do not properly enable it. 228 * 229 * People using the limiter to reduce bottlenecks on slower WAN connections 230 * should set the slop to 20 (2 packets). 231 */ 232 static int tcp_inflight_enable = 1; 233 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW, 234 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting"); 235 236 static int tcp_inflight_debug = 0; 237 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW, 238 &tcp_inflight_debug, 0, "Debug TCP inflight calculations"); 239 240 static int tcp_inflight_min = 6144; 241 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW, 242 &tcp_inflight_min, 0, "Lower bound for TCP inflight window"); 243 244 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT; 245 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW, 246 &tcp_inflight_max, 0, "Upper bound for TCP inflight window"); 247 248 static int tcp_inflight_stab = 50; 249 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW, 250 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 3 packets)"); 251 252 static int tcp_do_rfc3390 = 1; 253 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_RW, 254 &tcp_do_rfc3390, 0, 255 "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)"); 256 257 static u_long tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 258 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwmaxsegs, CTLFLAG_RW, 259 &tcp_iw_maxsegs, 0, "TCP IW segments max"); 260 261 static u_long tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 262 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwcapsegs, CTLFLAG_RW, 263 &tcp_iw_capsegs, 0, "TCP IW segments"); 264 265 int tcp_low_rtobase = 1; 266 SYSCTL_INT(_net_inet_tcp, OID_AUTO, low_rtobase, CTLFLAG_RW, 267 &tcp_low_rtobase, 0, "Lowering the Initial RTO (RFC 6298)"); 268 269 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives"); 270 static struct malloc_pipe tcptemp_mpipe; 271 272 static void tcp_willblock(void); 273 static void tcp_notify (struct inpcb *, int); 274 275 struct tcp_stats tcpstats_percpu[MAXCPU]; 276 #ifdef SMP 277 static int 278 sysctl_tcpstats(SYSCTL_HANDLER_ARGS) 279 { 280 int cpu, error = 0; 281 282 for (cpu = 0; cpu < ncpus; ++cpu) { 283 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu], 284 sizeof(struct tcp_stats)))) 285 break; 286 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu], 287 sizeof(struct tcp_stats)))) 288 break; 289 } 290 291 return (error); 292 } 293 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW), 294 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics"); 295 #else 296 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW, 297 &tcpstat, tcp_stats, "TCP statistics"); 298 #endif 299 300 /* 301 * Target size of TCP PCB hash tables. Must be a power of two. 302 * 303 * Note that this can be overridden by the kernel environment 304 * variable net.inet.tcp.tcbhashsize 305 */ 306 #ifndef TCBHASHSIZE 307 #define TCBHASHSIZE 512 308 #endif 309 310 /* 311 * This is the actual shape of what we allocate using the zone 312 * allocator. Doing it this way allows us to protect both structures 313 * using the same generation count, and also eliminates the overhead 314 * of allocating tcpcbs separately. By hiding the structure here, 315 * we avoid changing most of the rest of the code (although it needs 316 * to be changed, eventually, for greater efficiency). 317 */ 318 #define ALIGNMENT 32 319 #define ALIGNM1 (ALIGNMENT - 1) 320 struct inp_tp { 321 union { 322 struct inpcb inp; 323 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1]; 324 } inp_tp_u; 325 struct tcpcb tcb; 326 struct tcp_callout inp_tp_rexmt; 327 struct tcp_callout inp_tp_persist; 328 struct tcp_callout inp_tp_keep; 329 struct tcp_callout inp_tp_2msl; 330 struct tcp_callout inp_tp_delack; 331 struct netmsg_tcp_timer inp_tp_timermsg; 332 }; 333 #undef ALIGNMENT 334 #undef ALIGNM1 335 336 /* 337 * Tcp initialization 338 */ 339 void 340 tcp_init(void) 341 { 342 struct inpcbporthead *porthashbase; 343 struct inpcbinfo *ticb; 344 u_long porthashmask; 345 int hashsize = TCBHASHSIZE; 346 int cpu; 347 348 /* 349 * note: tcptemp is used for keepalives, and it is ok for an 350 * allocation to fail so do not specify MPF_INT. 351 */ 352 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp), 353 25, -1, 0, NULL, NULL, NULL); 354 355 tcp_delacktime = TCPTV_DELACK; 356 tcp_keepinit = TCPTV_KEEP_INIT; 357 tcp_keepidle = TCPTV_KEEP_IDLE; 358 tcp_keepintvl = TCPTV_KEEPINTVL; 359 tcp_maxpersistidle = TCPTV_KEEP_IDLE; 360 tcp_msl = TCPTV_MSL; 361 tcp_rexmit_min = TCPTV_MIN; 362 tcp_rexmit_slop = TCPTV_CPU_VAR; 363 364 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize); 365 if (!powerof2(hashsize)) { 366 kprintf("WARNING: TCB hash size not a power of 2\n"); 367 hashsize = 512; /* safe default */ 368 } 369 tcp_tcbhashsize = hashsize; 370 porthashbase = hashinit(hashsize, M_PCB, &porthashmask); 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 ticb->porthashbase = porthashbase; 379 ticb->porthashmask = porthashmask; 380 ticb->porttoken = &tcp_port_token; 381 #if 0 382 ticb->porthashbase = hashinit(hashsize, M_PCB, 383 &ticb->porthashmask); 384 #endif 385 ticb->wildcardhashbase = hashinit(hashsize, M_PCB, 386 &ticb->wildcardhashmask); 387 ticb->ipi_size = sizeof(struct inp_tp); 388 TAILQ_INIT(&tcpcbackq[cpu]); 389 } 390 391 tcp_reass_maxseg = nmbclusters / 16; 392 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg); 393 394 #ifdef INET6 395 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) 396 #else 397 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) 398 #endif 399 if (max_protohdr < TCP_MINPROTOHDR) 400 max_protohdr = TCP_MINPROTOHDR; 401 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) 402 panic("tcp_init"); 403 #undef TCP_MINPROTOHDR 404 405 /* 406 * Initialize TCP statistics counters for each CPU. 407 */ 408 #ifdef SMP 409 for (cpu = 0; cpu < ncpus; ++cpu) { 410 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats)); 411 } 412 #else 413 bzero(&tcpstat, sizeof(struct tcp_stats)); 414 #endif 415 416 syncache_init(); 417 netisr_register_rollup(tcp_willblock); 418 } 419 420 static void 421 tcp_willblock(void) 422 { 423 struct tcpcb *tp; 424 int cpu = mycpu->gd_cpuid; 425 426 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) { 427 KKASSERT(tp->t_flags & TF_ONOUTPUTQ); 428 tp->t_flags &= ~TF_ONOUTPUTQ; 429 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq); 430 tcp_output(tp); 431 } 432 } 433 434 /* 435 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. 436 * tcp_template used to store this data in mbufs, but we now recopy it out 437 * of the tcpcb each time to conserve mbufs. 438 */ 439 void 440 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr) 441 { 442 struct inpcb *inp = tp->t_inpcb; 443 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr; 444 445 #ifdef INET6 446 if (inp->inp_vflag & INP_IPV6) { 447 struct ip6_hdr *ip6; 448 449 ip6 = (struct ip6_hdr *)ip_ptr; 450 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | 451 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK); 452 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | 453 (IPV6_VERSION & IPV6_VERSION_MASK); 454 ip6->ip6_nxt = IPPROTO_TCP; 455 ip6->ip6_plen = sizeof(struct tcphdr); 456 ip6->ip6_src = inp->in6p_laddr; 457 ip6->ip6_dst = inp->in6p_faddr; 458 tcp_hdr->th_sum = 0; 459 } else 460 #endif 461 { 462 struct ip *ip = (struct ip *) ip_ptr; 463 464 ip->ip_vhl = IP_VHL_BORING; 465 ip->ip_tos = 0; 466 ip->ip_len = 0; 467 ip->ip_id = 0; 468 ip->ip_off = 0; 469 ip->ip_ttl = 0; 470 ip->ip_sum = 0; 471 ip->ip_p = IPPROTO_TCP; 472 ip->ip_src = inp->inp_laddr; 473 ip->ip_dst = inp->inp_faddr; 474 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr, 475 ip->ip_dst.s_addr, 476 htons(sizeof(struct tcphdr) + IPPROTO_TCP)); 477 } 478 479 tcp_hdr->th_sport = inp->inp_lport; 480 tcp_hdr->th_dport = inp->inp_fport; 481 tcp_hdr->th_seq = 0; 482 tcp_hdr->th_ack = 0; 483 tcp_hdr->th_x2 = 0; 484 tcp_hdr->th_off = 5; 485 tcp_hdr->th_flags = 0; 486 tcp_hdr->th_win = 0; 487 tcp_hdr->th_urp = 0; 488 } 489 490 /* 491 * Create template to be used to send tcp packets on a connection. 492 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only 493 * use for this function is in keepalives, which use tcp_respond. 494 */ 495 struct tcptemp * 496 tcp_maketemplate(struct tcpcb *tp) 497 { 498 struct tcptemp *tmp; 499 500 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL) 501 return (NULL); 502 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t); 503 return (tmp); 504 } 505 506 void 507 tcp_freetemplate(struct tcptemp *tmp) 508 { 509 mpipe_free(&tcptemp_mpipe, tmp); 510 } 511 512 /* 513 * Send a single message to the TCP at address specified by 514 * the given TCP/IP header. If m == NULL, then we make a copy 515 * of the tcpiphdr at ti and send directly to the addressed host. 516 * This is used to force keep alive messages out using the TCP 517 * template for a connection. If flags are given then we send 518 * a message back to the TCP which originated the * segment ti, 519 * and discard the mbuf containing it and any other attached mbufs. 520 * 521 * In any case the ack and sequence number of the transmitted 522 * segment are as specified by the parameters. 523 * 524 * NOTE: If m != NULL, then ti must point to *inside* the mbuf. 525 */ 526 void 527 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, 528 tcp_seq ack, tcp_seq seq, int flags) 529 { 530 int tlen; 531 int win = 0; 532 struct route *ro = NULL; 533 struct route sro; 534 struct ip *ip = ipgen; 535 struct tcphdr *nth; 536 int ipflags = 0; 537 struct route_in6 *ro6 = NULL; 538 struct route_in6 sro6; 539 struct ip6_hdr *ip6 = ipgen; 540 boolean_t use_tmpro = TRUE; 541 #ifdef INET6 542 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6); 543 #else 544 const boolean_t isipv6 = FALSE; 545 #endif 546 547 if (tp != NULL) { 548 if (!(flags & TH_RST)) { 549 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv); 550 if (win < 0) 551 win = 0; 552 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 553 win = (long)TCP_MAXWIN << tp->rcv_scale; 554 } 555 /* 556 * Don't use the route cache of a listen socket, 557 * it is not MPSAFE; use temporary route cache. 558 */ 559 if (tp->t_state != TCPS_LISTEN) { 560 if (isipv6) 561 ro6 = &tp->t_inpcb->in6p_route; 562 else 563 ro = &tp->t_inpcb->inp_route; 564 use_tmpro = FALSE; 565 } 566 } 567 if (use_tmpro) { 568 if (isipv6) { 569 ro6 = &sro6; 570 bzero(ro6, sizeof *ro6); 571 } else { 572 ro = &sro; 573 bzero(ro, sizeof *ro); 574 } 575 } 576 if (m == NULL) { 577 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 578 if (m == NULL) 579 return; 580 tlen = 0; 581 m->m_data += max_linkhdr; 582 if (isipv6) { 583 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); 584 ip6 = mtod(m, struct ip6_hdr *); 585 nth = (struct tcphdr *)(ip6 + 1); 586 } else { 587 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip)); 588 ip = mtod(m, struct ip *); 589 nth = (struct tcphdr *)(ip + 1); 590 } 591 bcopy(th, nth, sizeof(struct tcphdr)); 592 flags = TH_ACK; 593 } else { 594 m_freem(m->m_next); 595 m->m_next = NULL; 596 m->m_data = (caddr_t)ipgen; 597 /* m_len is set later */ 598 tlen = 0; 599 #define xchg(a, b, type) { type t; t = a; a = b; b = t; } 600 if (isipv6) { 601 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); 602 nth = (struct tcphdr *)(ip6 + 1); 603 } else { 604 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long); 605 nth = (struct tcphdr *)(ip + 1); 606 } 607 if (th != nth) { 608 /* 609 * this is usually a case when an extension header 610 * exists between the IPv6 header and the 611 * TCP header. 612 */ 613 nth->th_sport = th->th_sport; 614 nth->th_dport = th->th_dport; 615 } 616 xchg(nth->th_dport, nth->th_sport, n_short); 617 #undef xchg 618 } 619 if (isipv6) { 620 ip6->ip6_flow = 0; 621 ip6->ip6_vfc = IPV6_VERSION; 622 ip6->ip6_nxt = IPPROTO_TCP; 623 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen)); 624 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 625 } else { 626 tlen += sizeof(struct tcpiphdr); 627 ip->ip_len = tlen; 628 ip->ip_ttl = ip_defttl; 629 } 630 m->m_len = tlen; 631 m->m_pkthdr.len = tlen; 632 m->m_pkthdr.rcvif = NULL; 633 nth->th_seq = htonl(seq); 634 nth->th_ack = htonl(ack); 635 nth->th_x2 = 0; 636 nth->th_off = sizeof(struct tcphdr) >> 2; 637 nth->th_flags = flags; 638 if (tp != NULL) 639 nth->th_win = htons((u_short) (win >> tp->rcv_scale)); 640 else 641 nth->th_win = htons((u_short)win); 642 nth->th_urp = 0; 643 if (isipv6) { 644 nth->th_sum = 0; 645 nth->th_sum = in6_cksum(m, IPPROTO_TCP, 646 sizeof(struct ip6_hdr), 647 tlen - sizeof(struct ip6_hdr)); 648 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL, 649 (ro6 && ro6->ro_rt) ? 650 ro6->ro_rt->rt_ifp : NULL); 651 } else { 652 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 653 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); 654 m->m_pkthdr.csum_flags = CSUM_TCP; 655 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 656 } 657 #ifdef TCPDEBUG 658 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) 659 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); 660 #endif 661 if (isipv6) { 662 ip6_output(m, NULL, ro6, ipflags, NULL, NULL, 663 tp ? tp->t_inpcb : NULL); 664 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) { 665 RTFREE(ro6->ro_rt); 666 ro6->ro_rt = NULL; 667 } 668 } else { 669 ipflags |= IP_DEBUGROUTE; 670 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL); 671 if ((ro == &sro) && (ro->ro_rt != NULL)) { 672 RTFREE(ro->ro_rt); 673 ro->ro_rt = NULL; 674 } 675 } 676 } 677 678 /* 679 * Create a new TCP control block, making an 680 * empty reassembly queue and hooking it to the argument 681 * protocol control block. The `inp' parameter must have 682 * come from the zone allocator set up in tcp_init(). 683 */ 684 struct tcpcb * 685 tcp_newtcpcb(struct inpcb *inp) 686 { 687 struct inp_tp *it; 688 struct tcpcb *tp; 689 #ifdef INET6 690 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 691 #else 692 const boolean_t isipv6 = FALSE; 693 #endif 694 695 it = (struct inp_tp *)inp; 696 tp = &it->tcb; 697 bzero(tp, sizeof(struct tcpcb)); 698 LIST_INIT(&tp->t_segq); 699 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 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_rfc1323) 724 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP); 725 tp->t_inpcb = inp; /* XXX */ 726 tp->t_state = TCPS_CLOSED; 727 /* 728 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 729 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 730 * reasonable initial retransmit time. 731 */ 732 tp->t_srtt = TCPTV_SRTTBASE; 733 tp->t_rttvar = 734 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 735 tp->t_rttmin = tcp_rexmit_min; 736 tp->t_rxtcur = TCPTV_RTOBASE; 737 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 738 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 739 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 740 tp->snd_last = ticks; 741 tp->t_rcvtime = ticks; 742 /* 743 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 744 * because the socket may be bound to an IPv6 wildcard address, 745 * which may match an IPv4-mapped IPv6 address. 746 */ 747 inp->inp_ip_ttl = ip_defttl; 748 inp->inp_ppcb = tp; 749 tcp_sack_tcpcb_init(tp); 750 return (tp); /* XXX */ 751 } 752 753 /* 754 * Drop a TCP connection, reporting the specified error. 755 * If connection is synchronized, then send a RST to peer. 756 */ 757 struct tcpcb * 758 tcp_drop(struct tcpcb *tp, int error) 759 { 760 struct socket *so = tp->t_inpcb->inp_socket; 761 762 if (TCPS_HAVERCVDSYN(tp->t_state)) { 763 tp->t_state = TCPS_CLOSED; 764 tcp_output(tp); 765 tcpstat.tcps_drops++; 766 } else 767 tcpstat.tcps_conndrops++; 768 if (error == ETIMEDOUT && tp->t_softerror) 769 error = tp->t_softerror; 770 so->so_error = error; 771 return (tcp_close(tp)); 772 } 773 774 #ifdef SMP 775 776 struct netmsg_listen_detach { 777 struct netmsg_base base; 778 struct tcpcb *nm_tp; 779 }; 780 781 static void 782 tcp_listen_detach_handler(netmsg_t msg) 783 { 784 struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg; 785 struct tcpcb *tp = nmsg->nm_tp; 786 int cpu = mycpuid, nextcpu; 787 788 if (tp->t_flags & TF_LISTEN) 789 syncache_destroy(tp); 790 791 in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]); 792 793 nextcpu = cpu + 1; 794 if (nextcpu < ncpus2) 795 lwkt_forwardmsg(cpu_portfn(nextcpu), &nmsg->base.lmsg); 796 else 797 lwkt_replymsg(&nmsg->base.lmsg, 0); 798 } 799 800 #endif 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 socket *so = inp->inp_socket; 814 struct rtentry *rt; 815 boolean_t dosavessthresh; 816 #ifdef INET6 817 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 818 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0); 819 #else 820 const boolean_t isipv6 = FALSE; 821 #endif 822 823 #ifdef SMP 824 /* 825 * INP_WILDCARD_MP indicates that listen(2) has been called on 826 * this socket. This implies: 827 * - A wildcard inp's hash is replicated for each protocol thread. 828 * - Syncache for this inp grows independently in each protocol 829 * thread. 830 * - There is more than one cpu 831 * 832 * We have to chain a message to the rest of the protocol threads 833 * to cleanup the wildcard hash and the syncache. The cleanup 834 * in the current protocol thread is defered till the end of this 835 * function. 836 * 837 * NOTE: 838 * After cleanup the inp's hash and syncache entries, this inp will 839 * no longer be available to the rest of the protocol threads, so we 840 * are safe to whack the inp in the following code. 841 */ 842 if (inp->inp_flags & INP_WILDCARD_MP) { 843 struct netmsg_listen_detach nmsg; 844 845 KKASSERT(so->so_port == cpu_portfn(0)); 846 KKASSERT(&curthread->td_msgport == cpu_portfn(0)); 847 KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]); 848 849 netmsg_init(&nmsg.base, NULL, &curthread->td_msgport, 850 MSGF_PRIORITY, tcp_listen_detach_handler); 851 nmsg.nm_tp = tp; 852 lwkt_domsg(cpu_portfn(1), &nmsg.base.lmsg, 0); 853 854 inp->inp_flags &= ~INP_WILDCARD_MP; 855 } 856 #endif 857 858 KKASSERT(tp->t_state != TCPS_TERMINATING); 859 tp->t_state = TCPS_TERMINATING; 860 861 /* 862 * Make sure that all of our timers are stopped before we 863 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL), 864 * timers are never used. If timer message is never created 865 * (tp->tt_msg->tt_tcb == NULL), timers are never used too. 866 */ 867 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) { 868 tcp_callout_stop(tp, tp->tt_rexmt); 869 tcp_callout_stop(tp, tp->tt_persist); 870 tcp_callout_stop(tp, tp->tt_keep); 871 tcp_callout_stop(tp, tp->tt_2msl); 872 tcp_callout_stop(tp, tp->tt_delack); 873 } 874 875 if (tp->t_flags & TF_ONOUTPUTQ) { 876 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 877 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 878 tp->t_flags &= ~TF_ONOUTPUTQ; 879 } 880 881 /* 882 * If we got enough samples through the srtt filter, 883 * save the rtt and rttvar in the routing entry. 884 * 'Enough' is arbitrarily defined as the 16 samples. 885 * 16 samples is enough for the srtt filter to converge 886 * to within 5% of the correct value; fewer samples and 887 * we could save a very bogus rtt. 888 * 889 * Don't update the default route's characteristics and don't 890 * update anything that the user "locked". 891 */ 892 if (tp->t_rttupdated >= 16) { 893 u_long i = 0; 894 895 if (isipv6) { 896 struct sockaddr_in6 *sin6; 897 898 if ((rt = inp->in6p_route.ro_rt) == NULL) 899 goto no_valid_rt; 900 sin6 = (struct sockaddr_in6 *)rt_key(rt); 901 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 902 goto no_valid_rt; 903 } else 904 if ((rt = inp->inp_route.ro_rt) == NULL || 905 ((struct sockaddr_in *)rt_key(rt))-> 906 sin_addr.s_addr == INADDR_ANY) 907 goto no_valid_rt; 908 909 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 910 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 911 if (rt->rt_rmx.rmx_rtt && i) 912 /* 913 * filter this update to half the old & half 914 * the new values, converting scale. 915 * See route.h and tcp_var.h for a 916 * description of the scaling constants. 917 */ 918 rt->rt_rmx.rmx_rtt = 919 (rt->rt_rmx.rmx_rtt + i) / 2; 920 else 921 rt->rt_rmx.rmx_rtt = i; 922 tcpstat.tcps_cachedrtt++; 923 } 924 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 925 i = tp->t_rttvar * 926 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 927 if (rt->rt_rmx.rmx_rttvar && i) 928 rt->rt_rmx.rmx_rttvar = 929 (rt->rt_rmx.rmx_rttvar + i) / 2; 930 else 931 rt->rt_rmx.rmx_rttvar = i; 932 tcpstat.tcps_cachedrttvar++; 933 } 934 /* 935 * The old comment here said: 936 * update the pipelimit (ssthresh) if it has been updated 937 * already or if a pipesize was specified & the threshhold 938 * got below half the pipesize. I.e., wait for bad news 939 * before we start updating, then update on both good 940 * and bad news. 941 * 942 * But we want to save the ssthresh even if no pipesize is 943 * specified explicitly in the route, because such 944 * connections still have an implicit pipesize specified 945 * by the global tcp_sendspace. In the absence of a reliable 946 * way to calculate the pipesize, it will have to do. 947 */ 948 i = tp->snd_ssthresh; 949 if (rt->rt_rmx.rmx_sendpipe != 0) 950 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 951 else 952 dosavessthresh = (i < so->so_snd.ssb_hiwat/2); 953 if (dosavessthresh || 954 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 955 (rt->rt_rmx.rmx_ssthresh != 0))) { 956 /* 957 * convert the limit from user data bytes to 958 * packets then to packet data bytes. 959 */ 960 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 961 if (i < 2) 962 i = 2; 963 i *= tp->t_maxseg + 964 (isipv6 ? 965 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 966 sizeof(struct tcpiphdr)); 967 if (rt->rt_rmx.rmx_ssthresh) 968 rt->rt_rmx.rmx_ssthresh = 969 (rt->rt_rmx.rmx_ssthresh + i) / 2; 970 else 971 rt->rt_rmx.rmx_ssthresh = i; 972 tcpstat.tcps_cachedssthresh++; 973 } 974 } 975 976 no_valid_rt: 977 /* free the reassembly queue, if any */ 978 while((q = LIST_FIRST(&tp->t_segq)) != NULL) { 979 LIST_REMOVE(q, tqe_q); 980 m_freem(q->tqe_m); 981 kfree(q, M_TSEGQ); 982 atomic_add_int(&tcp_reass_qsize, -1); 983 } 984 /* throw away SACK blocks in scoreboard*/ 985 if (TCP_DO_SACK(tp)) 986 tcp_sack_destroy(&tp->scb); 987 988 inp->inp_ppcb = NULL; 989 soisdisconnected(so); 990 /* note: pcb detached later on */ 991 992 tcp_destroy_timermsg(tp); 993 994 if (tp->t_flags & TF_LISTEN) 995 syncache_destroy(tp); 996 997 /* 998 * NOTE: 999 * pcbdetach removes any wildcard hash entry on the current CPU. 1000 */ 1001 #ifdef INET6 1002 if (isafinet6) 1003 in6_pcbdetach(inp); 1004 else 1005 #endif 1006 in_pcbdetach(inp); 1007 1008 tcpstat.tcps_closed++; 1009 return (NULL); 1010 } 1011 1012 static __inline void 1013 tcp_drain_oncpu(struct inpcbhead *head) 1014 { 1015 struct inpcb *marker; 1016 struct inpcb *inpb; 1017 struct tcpcb *tcpb; 1018 struct tseg_qent *te; 1019 1020 /* 1021 * Allows us to block while running the list 1022 */ 1023 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1024 marker->inp_flags |= INP_PLACEMARKER; 1025 LIST_INSERT_HEAD(head, marker, inp_list); 1026 1027 while ((inpb = LIST_NEXT(marker, inp_list)) != NULL) { 1028 if ((inpb->inp_flags & INP_PLACEMARKER) == 0 && 1029 (tcpb = intotcpcb(inpb)) != NULL && 1030 (te = LIST_FIRST(&tcpb->t_segq)) != NULL) { 1031 LIST_REMOVE(te, tqe_q); 1032 m_freem(te->tqe_m); 1033 kfree(te, M_TSEGQ); 1034 atomic_add_int(&tcp_reass_qsize, -1); 1035 /* retry */ 1036 } else { 1037 LIST_REMOVE(marker, inp_list); 1038 LIST_INSERT_AFTER(inpb, marker, inp_list); 1039 } 1040 } 1041 LIST_REMOVE(marker, inp_list); 1042 kfree(marker, M_TEMP); 1043 } 1044 1045 #ifdef SMP 1046 struct netmsg_tcp_drain { 1047 struct netmsg_base base; 1048 struct inpcbhead *nm_head; 1049 }; 1050 1051 static void 1052 tcp_drain_handler(netmsg_t msg) 1053 { 1054 struct netmsg_tcp_drain *nm = (void *)msg; 1055 1056 tcp_drain_oncpu(nm->nm_head); 1057 lwkt_replymsg(&nm->base.lmsg, 0); 1058 } 1059 #endif 1060 1061 void 1062 tcp_drain(void) 1063 { 1064 #ifdef SMP 1065 int cpu; 1066 #endif 1067 1068 if (!do_tcpdrain) 1069 return; 1070 1071 /* 1072 * Walk the tcpbs, if existing, and flush the reassembly queue, 1073 * if there is one... 1074 * XXX: The "Net/3" implementation doesn't imply that the TCP 1075 * reassembly queue should be flushed, but in a situation 1076 * where we're really low on mbufs, this is potentially 1077 * useful. 1078 */ 1079 #ifdef SMP 1080 for (cpu = 0; cpu < ncpus2; cpu++) { 1081 struct netmsg_tcp_drain *nm; 1082 1083 if (cpu == mycpu->gd_cpuid) { 1084 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead); 1085 } else { 1086 nm = kmalloc(sizeof(struct netmsg_tcp_drain), 1087 M_LWKTMSG, M_NOWAIT); 1088 if (nm == NULL) 1089 continue; 1090 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1091 0, tcp_drain_handler); 1092 nm->nm_head = &tcbinfo[cpu].pcblisthead; 1093 lwkt_sendmsg(cpu_portfn(cpu), &nm->base.lmsg); 1094 } 1095 } 1096 #else 1097 tcp_drain_oncpu(&tcbinfo[0].pcblisthead); 1098 #endif 1099 } 1100 1101 /* 1102 * Notify a tcp user of an asynchronous error; 1103 * store error as soft error, but wake up user 1104 * (for now, won't do anything until can select for soft error). 1105 * 1106 * Do not wake up user since there currently is no mechanism for 1107 * reporting soft errors (yet - a kqueue filter may be added). 1108 */ 1109 static void 1110 tcp_notify(struct inpcb *inp, int error) 1111 { 1112 struct tcpcb *tp = intotcpcb(inp); 1113 1114 /* 1115 * Ignore some errors if we are hooked up. 1116 * If connection hasn't completed, has retransmitted several times, 1117 * and receives a second error, give up now. This is better 1118 * than waiting a long time to establish a connection that 1119 * can never complete. 1120 */ 1121 if (tp->t_state == TCPS_ESTABLISHED && 1122 (error == EHOSTUNREACH || error == ENETUNREACH || 1123 error == EHOSTDOWN)) { 1124 return; 1125 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1126 tp->t_softerror) 1127 tcp_drop(tp, error); 1128 else 1129 tp->t_softerror = error; 1130 #if 0 1131 wakeup(&so->so_timeo); 1132 sorwakeup(so); 1133 sowwakeup(so); 1134 #endif 1135 } 1136 1137 static int 1138 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1139 { 1140 int error, i, n; 1141 struct inpcb *marker; 1142 struct inpcb *inp; 1143 globaldata_t gd; 1144 int origcpu, ccpu; 1145 1146 error = 0; 1147 n = 0; 1148 1149 /* 1150 * The process of preparing the TCB list is too time-consuming and 1151 * resource-intensive to repeat twice on every request. 1152 */ 1153 if (req->oldptr == NULL) { 1154 for (ccpu = 0; ccpu < ncpus; ++ccpu) { 1155 gd = globaldata_find(ccpu); 1156 n += tcbinfo[gd->gd_cpuid].ipi_count; 1157 } 1158 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1159 return (0); 1160 } 1161 1162 if (req->newptr != NULL) 1163 return (EPERM); 1164 1165 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1166 marker->inp_flags |= INP_PLACEMARKER; 1167 1168 /* 1169 * OK, now we're committed to doing something. Run the inpcb list 1170 * for each cpu in the system and construct the output. Use a 1171 * list placemarker to deal with list changes occuring during 1172 * copyout blockages (but otherwise depend on being on the correct 1173 * cpu to avoid races). 1174 */ 1175 origcpu = mycpu->gd_cpuid; 1176 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) { 1177 globaldata_t rgd; 1178 caddr_t inp_ppcb; 1179 struct xtcpcb xt; 1180 int cpu_id; 1181 1182 cpu_id = (origcpu + ccpu) % ncpus; 1183 if ((smp_active_mask & CPUMASK(cpu_id)) == 0) 1184 continue; 1185 rgd = globaldata_find(cpu_id); 1186 lwkt_setcpu_self(rgd); 1187 1188 n = tcbinfo[cpu_id].ipi_count; 1189 1190 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list); 1191 i = 0; 1192 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1193 /* 1194 * process a snapshot of pcbs, ignoring placemarkers 1195 * and using our own to allow SYSCTL_OUT to block. 1196 */ 1197 LIST_REMOVE(marker, inp_list); 1198 LIST_INSERT_AFTER(inp, marker, inp_list); 1199 1200 if (inp->inp_flags & INP_PLACEMARKER) 1201 continue; 1202 if (prison_xinpcb(req->td, inp)) 1203 continue; 1204 1205 xt.xt_len = sizeof xt; 1206 bcopy(inp, &xt.xt_inp, sizeof *inp); 1207 inp_ppcb = inp->inp_ppcb; 1208 if (inp_ppcb != NULL) 1209 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1210 else 1211 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1212 if (inp->inp_socket) 1213 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1214 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1215 break; 1216 ++i; 1217 } 1218 LIST_REMOVE(marker, inp_list); 1219 if (error == 0 && i < n) { 1220 bzero(&xt, sizeof xt); 1221 xt.xt_len = sizeof xt; 1222 while (i < n) { 1223 error = SYSCTL_OUT(req, &xt, sizeof xt); 1224 if (error) 1225 break; 1226 ++i; 1227 } 1228 } 1229 } 1230 1231 /* 1232 * Make sure we are on the same cpu we were on originally, since 1233 * higher level callers expect this. Also don't pollute caches with 1234 * migrated userland data by (eventually) returning to userland 1235 * on a different cpu. 1236 */ 1237 lwkt_setcpu_self(globaldata_find(origcpu)); 1238 kfree(marker, M_TEMP); 1239 return (error); 1240 } 1241 1242 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1243 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1244 1245 static int 1246 tcp_getcred(SYSCTL_HANDLER_ARGS) 1247 { 1248 struct sockaddr_in addrs[2]; 1249 struct inpcb *inp; 1250 int cpu; 1251 int error; 1252 1253 error = priv_check(req->td, PRIV_ROOT); 1254 if (error != 0) 1255 return (error); 1256 error = SYSCTL_IN(req, addrs, sizeof addrs); 1257 if (error != 0) 1258 return (error); 1259 crit_enter(); 1260 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1261 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1262 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1263 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1264 if (inp == NULL || inp->inp_socket == NULL) { 1265 error = ENOENT; 1266 goto out; 1267 } 1268 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1269 out: 1270 crit_exit(); 1271 return (error); 1272 } 1273 1274 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1275 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1276 1277 #ifdef INET6 1278 static int 1279 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1280 { 1281 struct sockaddr_in6 addrs[2]; 1282 struct inpcb *inp; 1283 int error; 1284 boolean_t mapped = FALSE; 1285 1286 error = priv_check(req->td, PRIV_ROOT); 1287 if (error != 0) 1288 return (error); 1289 error = SYSCTL_IN(req, addrs, sizeof addrs); 1290 if (error != 0) 1291 return (error); 1292 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { 1293 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) 1294 mapped = TRUE; 1295 else 1296 return (EINVAL); 1297 } 1298 crit_enter(); 1299 if (mapped) { 1300 inp = in_pcblookup_hash(&tcbinfo[0], 1301 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], 1302 addrs[1].sin6_port, 1303 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], 1304 addrs[0].sin6_port, 1305 0, NULL); 1306 } else { 1307 inp = in6_pcblookup_hash(&tcbinfo[0], 1308 &addrs[1].sin6_addr, addrs[1].sin6_port, 1309 &addrs[0].sin6_addr, addrs[0].sin6_port, 1310 0, NULL); 1311 } 1312 if (inp == NULL || inp->inp_socket == NULL) { 1313 error = ENOENT; 1314 goto out; 1315 } 1316 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1317 out: 1318 crit_exit(); 1319 return (error); 1320 } 1321 1322 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1323 0, 0, 1324 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1325 #endif 1326 1327 struct netmsg_tcp_notify { 1328 struct netmsg_base base; 1329 void (*nm_notify)(struct inpcb *, int); 1330 struct in_addr nm_faddr; 1331 int nm_arg; 1332 }; 1333 1334 static void 1335 tcp_notifyall_oncpu(netmsg_t msg) 1336 { 1337 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg; 1338 int nextcpu; 1339 1340 in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nm->nm_faddr, 1341 nm->nm_arg, nm->nm_notify); 1342 1343 nextcpu = mycpuid + 1; 1344 if (nextcpu < ncpus2) 1345 lwkt_forwardmsg(cpu_portfn(nextcpu), &nm->base.lmsg); 1346 else 1347 lwkt_replymsg(&nm->base.lmsg, 0); 1348 } 1349 1350 void 1351 tcp_ctlinput(netmsg_t msg) 1352 { 1353 int cmd = msg->ctlinput.nm_cmd; 1354 struct sockaddr *sa = msg->ctlinput.nm_arg; 1355 struct ip *ip = msg->ctlinput.nm_extra; 1356 struct tcphdr *th; 1357 struct in_addr faddr; 1358 struct inpcb *inp; 1359 struct tcpcb *tp; 1360 void (*notify)(struct inpcb *, int) = tcp_notify; 1361 tcp_seq icmpseq; 1362 int arg, cpu; 1363 1364 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) { 1365 goto done; 1366 } 1367 1368 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1369 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1370 goto done; 1371 1372 arg = inetctlerrmap[cmd]; 1373 if (cmd == PRC_QUENCH) { 1374 notify = tcp_quench; 1375 } else if (icmp_may_rst && 1376 (cmd == PRC_UNREACH_ADMIN_PROHIB || 1377 cmd == PRC_UNREACH_PORT || 1378 cmd == PRC_TIMXCEED_INTRANS) && 1379 ip != NULL) { 1380 notify = tcp_drop_syn_sent; 1381 } else if (cmd == PRC_MSGSIZE) { 1382 struct icmp *icmp = (struct icmp *) 1383 ((caddr_t)ip - offsetof(struct icmp, icmp_ip)); 1384 1385 arg = ntohs(icmp->icmp_nextmtu); 1386 notify = tcp_mtudisc; 1387 } else if (PRC_IS_REDIRECT(cmd)) { 1388 ip = NULL; 1389 notify = in_rtchange; 1390 } else if (cmd == PRC_HOSTDEAD) { 1391 ip = NULL; 1392 } 1393 1394 if (ip != NULL) { 1395 crit_enter(); 1396 th = (struct tcphdr *)((caddr_t)ip + 1397 (IP_VHL_HL(ip->ip_vhl) << 2)); 1398 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport, 1399 ip->ip_src.s_addr, th->th_sport); 1400 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport, 1401 ip->ip_src, th->th_sport, 0, NULL); 1402 if ((inp != NULL) && (inp->inp_socket != NULL)) { 1403 icmpseq = htonl(th->th_seq); 1404 tp = intotcpcb(inp); 1405 if (SEQ_GEQ(icmpseq, tp->snd_una) && 1406 SEQ_LT(icmpseq, tp->snd_max)) 1407 (*notify)(inp, arg); 1408 } else { 1409 struct in_conninfo inc; 1410 1411 inc.inc_fport = th->th_dport; 1412 inc.inc_lport = th->th_sport; 1413 inc.inc_faddr = faddr; 1414 inc.inc_laddr = ip->ip_src; 1415 #ifdef INET6 1416 inc.inc_isipv6 = 0; 1417 #endif 1418 syncache_unreach(&inc, th); 1419 } 1420 crit_exit(); 1421 } else { 1422 struct netmsg_tcp_notify *nm; 1423 1424 KKASSERT(&curthread->td_msgport == cpu_portfn(0)); 1425 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT); 1426 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1427 0, tcp_notifyall_oncpu); 1428 nm->nm_faddr = faddr; 1429 nm->nm_arg = arg; 1430 nm->nm_notify = notify; 1431 1432 lwkt_sendmsg(cpu_portfn(0), &nm->base.lmsg); 1433 } 1434 done: 1435 lwkt_replymsg(&msg->lmsg, 0); 1436 } 1437 1438 #ifdef INET6 1439 1440 void 1441 tcp6_ctlinput(netmsg_t msg) 1442 { 1443 int cmd = msg->ctlinput.nm_cmd; 1444 struct sockaddr *sa = msg->ctlinput.nm_arg; 1445 void *d = msg->ctlinput.nm_extra; 1446 struct tcphdr th; 1447 void (*notify) (struct inpcb *, int) = tcp_notify; 1448 struct ip6_hdr *ip6; 1449 struct mbuf *m; 1450 struct ip6ctlparam *ip6cp = NULL; 1451 const struct sockaddr_in6 *sa6_src = NULL; 1452 int off; 1453 struct tcp_portonly { 1454 u_int16_t th_sport; 1455 u_int16_t th_dport; 1456 } *thp; 1457 int arg; 1458 1459 if (sa->sa_family != AF_INET6 || 1460 sa->sa_len != sizeof(struct sockaddr_in6)) { 1461 goto out; 1462 } 1463 1464 arg = 0; 1465 if (cmd == PRC_QUENCH) 1466 notify = tcp_quench; 1467 else if (cmd == PRC_MSGSIZE) { 1468 struct ip6ctlparam *ip6cp = d; 1469 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6; 1470 1471 arg = ntohl(icmp6->icmp6_mtu); 1472 notify = tcp_mtudisc; 1473 } else if (!PRC_IS_REDIRECT(cmd) && 1474 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) { 1475 goto out; 1476 } 1477 1478 /* if the parameter is from icmp6, decode it. */ 1479 if (d != NULL) { 1480 ip6cp = (struct ip6ctlparam *)d; 1481 m = ip6cp->ip6c_m; 1482 ip6 = ip6cp->ip6c_ip6; 1483 off = ip6cp->ip6c_off; 1484 sa6_src = ip6cp->ip6c_src; 1485 } else { 1486 m = NULL; 1487 ip6 = NULL; 1488 off = 0; /* fool gcc */ 1489 sa6_src = &sa6_any; 1490 } 1491 1492 if (ip6 != NULL) { 1493 struct in_conninfo inc; 1494 /* 1495 * XXX: We assume that when IPV6 is non NULL, 1496 * M and OFF are valid. 1497 */ 1498 1499 /* check if we can safely examine src and dst ports */ 1500 if (m->m_pkthdr.len < off + sizeof *thp) 1501 goto out; 1502 1503 bzero(&th, sizeof th); 1504 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1505 1506 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport, 1507 (struct sockaddr *)ip6cp->ip6c_src, 1508 th.th_sport, cmd, arg, notify); 1509 1510 inc.inc_fport = th.th_dport; 1511 inc.inc_lport = th.th_sport; 1512 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1513 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1514 inc.inc_isipv6 = 1; 1515 syncache_unreach(&inc, &th); 1516 } else { 1517 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0, 1518 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify); 1519 } 1520 out: 1521 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0); 1522 } 1523 1524 #endif 1525 1526 /* 1527 * Following is where TCP initial sequence number generation occurs. 1528 * 1529 * There are two places where we must use initial sequence numbers: 1530 * 1. In SYN-ACK packets. 1531 * 2. In SYN packets. 1532 * 1533 * All ISNs for SYN-ACK packets are generated by the syncache. See 1534 * tcp_syncache.c for details. 1535 * 1536 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1537 * depends on this property. In addition, these ISNs should be 1538 * unguessable so as to prevent connection hijacking. To satisfy 1539 * the requirements of this situation, the algorithm outlined in 1540 * RFC 1948 is used to generate sequence numbers. 1541 * 1542 * Implementation details: 1543 * 1544 * Time is based off the system timer, and is corrected so that it 1545 * increases by one megabyte per second. This allows for proper 1546 * recycling on high speed LANs while still leaving over an hour 1547 * before rollover. 1548 * 1549 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1550 * between seeding of isn_secret. This is normally set to zero, 1551 * as reseeding should not be necessary. 1552 * 1553 */ 1554 1555 #define ISN_BYTES_PER_SECOND 1048576 1556 1557 u_char isn_secret[32]; 1558 int isn_last_reseed; 1559 MD5_CTX isn_ctx; 1560 1561 tcp_seq 1562 tcp_new_isn(struct tcpcb *tp) 1563 { 1564 u_int32_t md5_buffer[4]; 1565 tcp_seq new_isn; 1566 1567 /* Seed if this is the first use, reseed if requested. */ 1568 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1569 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1570 < (u_int)ticks))) { 1571 read_random_unlimited(&isn_secret, sizeof isn_secret); 1572 isn_last_reseed = ticks; 1573 } 1574 1575 /* Compute the md5 hash and return the ISN. */ 1576 MD5Init(&isn_ctx); 1577 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1578 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1579 #ifdef INET6 1580 if (tp->t_inpcb->inp_vflag & INP_IPV6) { 1581 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1582 sizeof(struct in6_addr)); 1583 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1584 sizeof(struct in6_addr)); 1585 } else 1586 #endif 1587 { 1588 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1589 sizeof(struct in_addr)); 1590 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1591 sizeof(struct in_addr)); 1592 } 1593 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1594 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1595 new_isn = (tcp_seq) md5_buffer[0]; 1596 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1597 return (new_isn); 1598 } 1599 1600 /* 1601 * When a source quench is received, close congestion window 1602 * to one segment. We will gradually open it again as we proceed. 1603 */ 1604 void 1605 tcp_quench(struct inpcb *inp, int error) 1606 { 1607 struct tcpcb *tp = intotcpcb(inp); 1608 1609 if (tp != NULL) { 1610 tp->snd_cwnd = tp->t_maxseg; 1611 tp->snd_wacked = 0; 1612 } 1613 } 1614 1615 /* 1616 * When a specific ICMP unreachable message is received and the 1617 * connection state is SYN-SENT, drop the connection. This behavior 1618 * is controlled by the icmp_may_rst sysctl. 1619 */ 1620 void 1621 tcp_drop_syn_sent(struct inpcb *inp, int error) 1622 { 1623 struct tcpcb *tp = intotcpcb(inp); 1624 1625 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT)) 1626 tcp_drop(tp, error); 1627 } 1628 1629 /* 1630 * When a `need fragmentation' ICMP is received, update our idea of the MSS 1631 * based on the new value in the route. Also nudge TCP to send something, 1632 * since we know the packet we just sent was dropped. 1633 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1634 */ 1635 void 1636 tcp_mtudisc(struct inpcb *inp, int mtu) 1637 { 1638 struct tcpcb *tp = intotcpcb(inp); 1639 struct rtentry *rt; 1640 struct socket *so = inp->inp_socket; 1641 int maxopd, mss; 1642 #ifdef INET6 1643 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0); 1644 #else 1645 const boolean_t isipv6 = FALSE; 1646 #endif 1647 1648 if (tp == NULL) 1649 return; 1650 1651 /* 1652 * If no MTU is provided in the ICMP message, use the 1653 * next lower likely value, as specified in RFC 1191. 1654 */ 1655 if (mtu == 0) { 1656 int oldmtu; 1657 1658 oldmtu = tp->t_maxopd + 1659 (isipv6 ? 1660 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1661 sizeof(struct tcpiphdr)); 1662 mtu = ip_next_mtu(oldmtu, 0); 1663 } 1664 1665 if (isipv6) 1666 rt = tcp_rtlookup6(&inp->inp_inc); 1667 else 1668 rt = tcp_rtlookup(&inp->inp_inc); 1669 if (rt != NULL) { 1670 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu) 1671 mtu = rt->rt_rmx.rmx_mtu; 1672 1673 maxopd = mtu - 1674 (isipv6 ? 1675 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1676 sizeof(struct tcpiphdr)); 1677 1678 /* 1679 * XXX - The following conditional probably violates the TCP 1680 * spec. The problem is that, since we don't know the 1681 * other end's MSS, we are supposed to use a conservative 1682 * default. But, if we do that, then MTU discovery will 1683 * never actually take place, because the conservative 1684 * default is much less than the MTUs typically seen 1685 * on the Internet today. For the moment, we'll sweep 1686 * this under the carpet. 1687 * 1688 * The conservative default might not actually be a problem 1689 * if the only case this occurs is when sending an initial 1690 * SYN with options and data to a host we've never talked 1691 * to before. Then, they will reply with an MSS value which 1692 * will get recorded and the new parameters should get 1693 * recomputed. For Further Study. 1694 */ 1695 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd) 1696 maxopd = rt->rt_rmx.rmx_mssopt; 1697 } else 1698 maxopd = mtu - 1699 (isipv6 ? 1700 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1701 sizeof(struct tcpiphdr)); 1702 1703 if (tp->t_maxopd <= maxopd) 1704 return; 1705 tp->t_maxopd = maxopd; 1706 1707 mss = maxopd; 1708 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) == 1709 (TF_REQ_TSTMP | TF_RCVD_TSTMP)) 1710 mss -= TCPOLEN_TSTAMP_APPA; 1711 1712 /* round down to multiple of MCLBYTES */ 1713 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */ 1714 if (mss > MCLBYTES) 1715 mss &= ~(MCLBYTES - 1); 1716 #else 1717 if (mss > MCLBYTES) 1718 mss = (mss / MCLBYTES) * MCLBYTES; 1719 #endif 1720 1721 if (so->so_snd.ssb_hiwat < mss) 1722 mss = so->so_snd.ssb_hiwat; 1723 1724 tp->t_maxseg = mss; 1725 tp->t_rtttime = 0; 1726 tp->snd_nxt = tp->snd_una; 1727 tcp_output(tp); 1728 tcpstat.tcps_mturesent++; 1729 } 1730 1731 /* 1732 * Look-up the routing entry to the peer of this inpcb. If no route 1733 * is found and it cannot be allocated the return NULL. This routine 1734 * is called by TCP routines that access the rmx structure and by tcp_mss 1735 * to get the interface MTU. 1736 */ 1737 struct rtentry * 1738 tcp_rtlookup(struct in_conninfo *inc) 1739 { 1740 struct route *ro = &inc->inc_route; 1741 1742 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1743 /* No route yet, so try to acquire one */ 1744 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1745 /* 1746 * unused portions of the structure MUST be zero'd 1747 * out because rtalloc() treats it as opaque data 1748 */ 1749 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1750 ro->ro_dst.sa_family = AF_INET; 1751 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1752 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1753 inc->inc_faddr; 1754 rtalloc(ro); 1755 } 1756 } 1757 return (ro->ro_rt); 1758 } 1759 1760 #ifdef INET6 1761 struct rtentry * 1762 tcp_rtlookup6(struct in_conninfo *inc) 1763 { 1764 struct route_in6 *ro6 = &inc->inc6_route; 1765 1766 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1767 /* No route yet, so try to acquire one */ 1768 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1769 /* 1770 * unused portions of the structure MUST be zero'd 1771 * out because rtalloc() treats it as opaque data 1772 */ 1773 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1774 ro6->ro_dst.sin6_family = AF_INET6; 1775 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1776 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1777 rtalloc((struct route *)ro6); 1778 } 1779 } 1780 return (ro6->ro_rt); 1781 } 1782 #endif 1783 1784 #ifdef IPSEC 1785 /* compute ESP/AH header size for TCP, including outer IP header. */ 1786 size_t 1787 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1788 { 1789 struct inpcb *inp; 1790 struct mbuf *m; 1791 size_t hdrsiz; 1792 struct ip *ip; 1793 struct tcphdr *th; 1794 1795 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1796 return (0); 1797 MGETHDR(m, MB_DONTWAIT, MT_DATA); 1798 if (!m) 1799 return (0); 1800 1801 #ifdef INET6 1802 if (inp->inp_vflag & INP_IPV6) { 1803 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1804 1805 th = (struct tcphdr *)(ip6 + 1); 1806 m->m_pkthdr.len = m->m_len = 1807 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1808 tcp_fillheaders(tp, ip6, th); 1809 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1810 } else 1811 #endif 1812 { 1813 ip = mtod(m, struct ip *); 1814 th = (struct tcphdr *)(ip + 1); 1815 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1816 tcp_fillheaders(tp, ip, th); 1817 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1818 } 1819 1820 m_free(m); 1821 return (hdrsiz); 1822 } 1823 #endif 1824 1825 /* 1826 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1827 * 1828 * This code attempts to calculate the bandwidth-delay product as a 1829 * means of determining the optimal window size to maximize bandwidth, 1830 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1831 * routers. This code also does a fairly good job keeping RTTs in check 1832 * across slow links like modems. We implement an algorithm which is very 1833 * similar (but not meant to be) TCP/Vegas. The code operates on the 1834 * transmitter side of a TCP connection and so only effects the transmit 1835 * side of the connection. 1836 * 1837 * BACKGROUND: TCP makes no provision for the management of buffer space 1838 * at the end points or at the intermediate routers and switches. A TCP 1839 * stream, whether using NewReno or not, will eventually buffer as 1840 * many packets as it is able and the only reason this typically works is 1841 * due to the fairly small default buffers made available for a connection 1842 * (typicaly 16K or 32K). As machines use larger windows and/or window 1843 * scaling it is now fairly easy for even a single TCP connection to blow-out 1844 * all available buffer space not only on the local interface, but on 1845 * intermediate routers and switches as well. NewReno makes a misguided 1846 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1847 * then backing off, then steadily increasing the window again until another 1848 * failure occurs, ad-infinitum. This results in terrible oscillation that 1849 * is only made worse as network loads increase and the idea of intentionally 1850 * blowing out network buffers is, frankly, a terrible way to manage network 1851 * resources. 1852 * 1853 * It is far better to limit the transmit window prior to the failure 1854 * condition being achieved. There are two general ways to do this: First 1855 * you can 'scan' through different transmit window sizes and locate the 1856 * point where the RTT stops increasing, indicating that you have filled the 1857 * pipe, then scan backwards until you note that RTT stops decreasing, then 1858 * repeat ad-infinitum. This method works in principle but has severe 1859 * implementation issues due to RTT variances, timer granularity, and 1860 * instability in the algorithm which can lead to many false positives and 1861 * create oscillations as well as interact badly with other TCP streams 1862 * implementing the same algorithm. 1863 * 1864 * The second method is to limit the window to the bandwidth delay product 1865 * of the link. This is the method we implement. RTT variances and our 1866 * own manipulation of the congestion window, bwnd, can potentially 1867 * destabilize the algorithm. For this reason we have to stabilize the 1868 * elements used to calculate the window. We do this by using the minimum 1869 * observed RTT, the long term average of the observed bandwidth, and 1870 * by adding two segments worth of slop. It isn't perfect but it is able 1871 * to react to changing conditions and gives us a very stable basis on 1872 * which to extend the algorithm. 1873 */ 1874 void 1875 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1876 { 1877 u_long bw; 1878 u_long bwnd; 1879 int save_ticks; 1880 int delta_ticks; 1881 1882 /* 1883 * If inflight_enable is disabled in the middle of a tcp connection, 1884 * make sure snd_bwnd is effectively disabled. 1885 */ 1886 if (!tcp_inflight_enable) { 1887 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1888 tp->snd_bandwidth = 0; 1889 return; 1890 } 1891 1892 /* 1893 * Validate the delta time. If a connection is new or has been idle 1894 * a long time we have to reset the bandwidth calculator. 1895 */ 1896 save_ticks = ticks; 1897 delta_ticks = save_ticks - tp->t_bw_rtttime; 1898 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1899 tp->t_bw_rtttime = ticks; 1900 tp->t_bw_rtseq = ack_seq; 1901 if (tp->snd_bandwidth == 0) 1902 tp->snd_bandwidth = tcp_inflight_min; 1903 return; 1904 } 1905 if (delta_ticks == 0) 1906 return; 1907 1908 /* 1909 * Sanity check, plus ignore pure window update acks. 1910 */ 1911 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1912 return; 1913 1914 /* 1915 * Figure out the bandwidth. Due to the tick granularity this 1916 * is a very rough number and it MUST be averaged over a fairly 1917 * long period of time. XXX we need to take into account a link 1918 * that is not using all available bandwidth, but for now our 1919 * slop will ramp us up if this case occurs and the bandwidth later 1920 * increases. 1921 */ 1922 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 1923 tp->t_bw_rtttime = save_ticks; 1924 tp->t_bw_rtseq = ack_seq; 1925 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4; 1926 1927 tp->snd_bandwidth = bw; 1928 1929 /* 1930 * Calculate the semi-static bandwidth delay product, plus two maximal 1931 * segments. The additional slop puts us squarely in the sweet 1932 * spot and also handles the bandwidth run-up case. Without the 1933 * slop we could be locking ourselves into a lower bandwidth. 1934 * 1935 * Situations Handled: 1936 * (1) Prevents over-queueing of packets on LANs, especially on 1937 * high speed LANs, allowing larger TCP buffers to be 1938 * specified, and also does a good job preventing 1939 * over-queueing of packets over choke points like modems 1940 * (at least for the transmit side). 1941 * 1942 * (2) Is able to handle changing network loads (bandwidth 1943 * drops so bwnd drops, bandwidth increases so bwnd 1944 * increases). 1945 * 1946 * (3) Theoretically should stabilize in the face of multiple 1947 * connections implementing the same algorithm (this may need 1948 * a little work). 1949 * 1950 * (4) Stability value (defaults to 20 = 2 maximal packets) can 1951 * be adjusted with a sysctl but typically only needs to be on 1952 * very slow connections. A value no smaller then 5 should 1953 * be used, but only reduce this default if you have no other 1954 * choice. 1955 */ 1956 1957 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2) 1958 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 1959 tcp_inflight_stab * (int)tp->t_maxseg / 10; 1960 #undef USERTT 1961 1962 if (tcp_inflight_debug > 0) { 1963 static int ltime; 1964 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) { 1965 ltime = ticks; 1966 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n", 1967 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd); 1968 } 1969 } 1970 if ((long)bwnd < tcp_inflight_min) 1971 bwnd = tcp_inflight_min; 1972 if (bwnd > tcp_inflight_max) 1973 bwnd = tcp_inflight_max; 1974 if ((long)bwnd < tp->t_maxseg * 2) 1975 bwnd = tp->t_maxseg * 2; 1976 tp->snd_bwnd = bwnd; 1977 } 1978 1979 static void 1980 tcp_rmx_iwsegs(struct tcpcb *tp, u_long *maxsegs, u_long *capsegs) 1981 { 1982 struct rtentry *rt; 1983 struct inpcb *inp = tp->t_inpcb; 1984 #ifdef INET6 1985 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE); 1986 #else 1987 const boolean_t isipv6 = FALSE; 1988 #endif 1989 1990 /* XXX */ 1991 if (tcp_iw_maxsegs < TCP_IW_MAXSEGS_DFLT) 1992 tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 1993 if (tcp_iw_capsegs < TCP_IW_CAPSEGS_DFLT) 1994 tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 1995 1996 if (isipv6) 1997 rt = tcp_rtlookup6(&inp->inp_inc); 1998 else 1999 rt = tcp_rtlookup(&inp->inp_inc); 2000 if (rt == NULL || 2001 rt->rt_rmx.rmx_iwmaxsegs < TCP_IW_MAXSEGS_DFLT || 2002 rt->rt_rmx.rmx_iwcapsegs < TCP_IW_CAPSEGS_DFLT) { 2003 *maxsegs = tcp_iw_maxsegs; 2004 *capsegs = tcp_iw_capsegs; 2005 return; 2006 } 2007 *maxsegs = rt->rt_rmx.rmx_iwmaxsegs; 2008 *capsegs = rt->rt_rmx.rmx_iwcapsegs; 2009 } 2010 2011 u_long 2012 tcp_initial_window(struct tcpcb *tp) 2013 { 2014 if (tcp_do_rfc3390) { 2015 /* 2016 * RFC3390: 2017 * "If the SYN or SYN/ACK is lost, the initial window 2018 * used by a sender after a correctly transmitted SYN 2019 * MUST be one segment consisting of MSS bytes." 2020 * 2021 * However, we do something a little bit more aggressive 2022 * then RFC3390 here: 2023 * - Only if time spent in the SYN or SYN|ACK retransmition 2024 * >= 3 seconds, the IW is reduced. We do this mainly 2025 * because when RFC3390 is published, the initial RTO is 2026 * still 3 seconds (the threshold we test here), while 2027 * after RFC6298, the initial RTO is 1 second. This 2028 * behaviour probably still falls within the spirit of 2029 * RFC3390. 2030 * - When IW is reduced, 2*MSS is used instead of 1*MSS. 2031 * Mainly to avoid sender and receiver deadlock until 2032 * delayed ACK timer expires. And even RFC2581 does not 2033 * try to reduce IW upon SYN or SYN|ACK retransmition 2034 * timeout. 2035 * 2036 * See also: 2037 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03 2038 */ 2039 if (tp->t_rxtsyn >= TCPTV_RTOBASE3) { 2040 return (2 * tp->t_maxseg); 2041 } else { 2042 u_long maxsegs, capsegs; 2043 2044 tcp_rmx_iwsegs(tp, &maxsegs, &capsegs); 2045 return min(maxsegs * tp->t_maxseg, 2046 max(2 * tp->t_maxseg, capsegs * 1460)); 2047 } 2048 } else { 2049 /* 2050 * Even RFC2581 (back to 1999) allows 2*SMSS IW. 2051 * 2052 * Mainly to avoid sender and receiver deadlock 2053 * until delayed ACK timer expires. 2054 */ 2055 return (2 * tp->t_maxseg); 2056 } 2057 } 2058 2059 #ifdef TCP_SIGNATURE 2060 /* 2061 * Compute TCP-MD5 hash of a TCP segment. (RFC2385) 2062 * 2063 * We do this over ip, tcphdr, segment data, and the key in the SADB. 2064 * When called from tcp_input(), we can be sure that th_sum has been 2065 * zeroed out and verified already. 2066 * 2067 * Return 0 if successful, otherwise return -1. 2068 * 2069 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a 2070 * search with the destination IP address, and a 'magic SPI' to be 2071 * determined by the application. This is hardcoded elsewhere to 1179 2072 * right now. Another branch of this code exists which uses the SPD to 2073 * specify per-application flows but it is unstable. 2074 */ 2075 int 2076 tcpsignature_compute( 2077 struct mbuf *m, /* mbuf chain */ 2078 int len, /* length of TCP data */ 2079 int optlen, /* length of TCP options */ 2080 u_char *buf, /* storage for MD5 digest */ 2081 u_int direction) /* direction of flow */ 2082 { 2083 struct ippseudo ippseudo; 2084 MD5_CTX ctx; 2085 int doff; 2086 struct ip *ip; 2087 struct ipovly *ipovly; 2088 struct secasvar *sav; 2089 struct tcphdr *th; 2090 #ifdef INET6 2091 struct ip6_hdr *ip6; 2092 struct in6_addr in6; 2093 uint32_t plen; 2094 uint16_t nhdr; 2095 #endif /* INET6 */ 2096 u_short savecsum; 2097 2098 KASSERT(m != NULL, ("passed NULL mbuf. Game over.")); 2099 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature")); 2100 /* 2101 * Extract the destination from the IP header in the mbuf. 2102 */ 2103 ip = mtod(m, struct ip *); 2104 #ifdef INET6 2105 ip6 = NULL; /* Make the compiler happy. */ 2106 #endif /* INET6 */ 2107 /* 2108 * Look up an SADB entry which matches the address found in 2109 * the segment. 2110 */ 2111 switch (IP_VHL_V(ip->ip_vhl)) { 2112 case IPVERSION: 2113 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst, 2114 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2115 break; 2116 #ifdef INET6 2117 case (IPV6_VERSION >> 4): 2118 ip6 = mtod(m, struct ip6_hdr *); 2119 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst, 2120 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2121 break; 2122 #endif /* INET6 */ 2123 default: 2124 return (EINVAL); 2125 /* NOTREACHED */ 2126 break; 2127 } 2128 if (sav == NULL) { 2129 kprintf("%s: SADB lookup failed\n", __func__); 2130 return (EINVAL); 2131 } 2132 MD5Init(&ctx); 2133 2134 /* 2135 * Step 1: Update MD5 hash with IP pseudo-header. 2136 * 2137 * XXX The ippseudo header MUST be digested in network byte order, 2138 * or else we'll fail the regression test. Assume all fields we've 2139 * been doing arithmetic on have been in host byte order. 2140 * XXX One cannot depend on ipovly->ih_len here. When called from 2141 * tcp_output(), the underlying ip_len member has not yet been set. 2142 */ 2143 switch (IP_VHL_V(ip->ip_vhl)) { 2144 case IPVERSION: 2145 ipovly = (struct ipovly *)ip; 2146 ippseudo.ippseudo_src = ipovly->ih_src; 2147 ippseudo.ippseudo_dst = ipovly->ih_dst; 2148 ippseudo.ippseudo_pad = 0; 2149 ippseudo.ippseudo_p = IPPROTO_TCP; 2150 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen); 2151 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo)); 2152 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip)); 2153 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen; 2154 break; 2155 #ifdef INET6 2156 /* 2157 * RFC 2385, 2.0 Proposal 2158 * For IPv6, the pseudo-header is as described in RFC 2460, namely the 2159 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero- 2160 * extended next header value (to form 32 bits), and 32-bit segment 2161 * length. 2162 * Note: Upper-Layer Packet Length comes before Next Header. 2163 */ 2164 case (IPV6_VERSION >> 4): 2165 in6 = ip6->ip6_src; 2166 in6_clearscope(&in6); 2167 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2168 in6 = ip6->ip6_dst; 2169 in6_clearscope(&in6); 2170 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2171 plen = htonl(len + sizeof(struct tcphdr) + optlen); 2172 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t)); 2173 nhdr = 0; 2174 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2175 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2176 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2177 nhdr = IPPROTO_TCP; 2178 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2179 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr)); 2180 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen; 2181 break; 2182 #endif /* INET6 */ 2183 default: 2184 return (EINVAL); 2185 /* NOTREACHED */ 2186 break; 2187 } 2188 /* 2189 * Step 2: Update MD5 hash with TCP header, excluding options. 2190 * The TCP checksum must be set to zero. 2191 */ 2192 savecsum = th->th_sum; 2193 th->th_sum = 0; 2194 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr)); 2195 th->th_sum = savecsum; 2196 /* 2197 * Step 3: Update MD5 hash with TCP segment data. 2198 * Use m_apply() to avoid an early m_pullup(). 2199 */ 2200 if (len > 0) 2201 m_apply(m, doff, len, tcpsignature_apply, &ctx); 2202 /* 2203 * Step 4: Update MD5 hash with shared secret. 2204 */ 2205 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth)); 2206 MD5Final(buf, &ctx); 2207 key_sa_recordxfer(sav, m); 2208 key_freesav(sav); 2209 return (0); 2210 } 2211 2212 int 2213 tcpsignature_apply(void *fstate, void *data, unsigned int len) 2214 { 2215 2216 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len); 2217 return (0); 2218 } 2219 #endif /* TCP_SIGNATURE */ 2220