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 * $DragonFly: src/sys/netinet/tcp_subr.c,v 1.63 2008/11/11 10:46:58 sephe Exp $ 69 */ 70 71 #include "opt_compat.h" 72 #include "opt_inet.h" 73 #include "opt_inet6.h" 74 #include "opt_ipsec.h" 75 #include "opt_tcpdebug.h" 76 77 #include <sys/param.h> 78 #include <sys/systm.h> 79 #include <sys/callout.h> 80 #include <sys/kernel.h> 81 #include <sys/sysctl.h> 82 #include <sys/malloc.h> 83 #include <sys/mpipe.h> 84 #include <sys/mbuf.h> 85 #ifdef INET6 86 #include <sys/domain.h> 87 #endif 88 #include <sys/proc.h> 89 #include <sys/priv.h> 90 #include <sys/socket.h> 91 #include <sys/socketvar.h> 92 #include <sys/protosw.h> 93 #include <sys/random.h> 94 #include <sys/in_cksum.h> 95 #include <sys/ktr.h> 96 97 #include <net/route.h> 98 #include <net/if.h> 99 #include <net/netisr.h> 100 101 #define _IP_VHL 102 #include <netinet/in.h> 103 #include <netinet/in_systm.h> 104 #include <netinet/ip.h> 105 #include <netinet/ip6.h> 106 #include <netinet/in_pcb.h> 107 #include <netinet6/in6_pcb.h> 108 #include <netinet/in_var.h> 109 #include <netinet/ip_var.h> 110 #include <netinet6/ip6_var.h> 111 #include <netinet/ip_icmp.h> 112 #ifdef INET6 113 #include <netinet/icmp6.h> 114 #endif 115 #include <netinet/tcp.h> 116 #include <netinet/tcp_fsm.h> 117 #include <netinet/tcp_seq.h> 118 #include <netinet/tcp_timer.h> 119 #include <netinet/tcp_timer2.h> 120 #include <netinet/tcp_var.h> 121 #include <netinet6/tcp6_var.h> 122 #include <netinet/tcpip.h> 123 #ifdef TCPDEBUG 124 #include <netinet/tcp_debug.h> 125 #endif 126 #include <netinet6/ip6protosw.h> 127 128 #ifdef IPSEC 129 #include <netinet6/ipsec.h> 130 #include <netproto/key/key.h> 131 #ifdef INET6 132 #include <netinet6/ipsec6.h> 133 #endif 134 #endif 135 136 #ifdef FAST_IPSEC 137 #include <netproto/ipsec/ipsec.h> 138 #ifdef INET6 139 #include <netproto/ipsec/ipsec6.h> 140 #endif 141 #define IPSEC 142 #endif 143 144 #include <sys/md5.h> 145 #include <machine/smp.h> 146 147 #include <sys/msgport2.h> 148 #include <sys/mplock2.h> 149 #include <net/netmsg2.h> 150 151 #if !defined(KTR_TCP) 152 #define KTR_TCP KTR_ALL 153 #endif 154 KTR_INFO_MASTER(tcp); 155 /* 156 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0); 157 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0); 158 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0); 159 */ 160 #define logtcp(name) KTR_LOG(tcp_ ## name) 161 162 struct inpcbinfo tcbinfo[MAXCPU]; 163 struct tcpcbackqhead tcpcbackq[MAXCPU]; 164 165 static struct lwkt_token tcp_port_token = 166 LWKT_TOKEN_MP_INITIALIZER(tcp_port_token); 167 168 int tcp_mssdflt = TCP_MSS; 169 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 170 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size"); 171 172 #ifdef INET6 173 int tcp_v6mssdflt = TCP6_MSS; 174 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW, 175 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6"); 176 #endif 177 178 /* 179 * Minimum MSS we accept and use. This prevents DoS attacks where 180 * we are forced to a ridiculous low MSS like 20 and send hundreds 181 * of packets instead of one. The effect scales with the available 182 * bandwidth and quickly saturates the CPU and network interface 183 * with packet generation and sending. Set to zero to disable MINMSS 184 * checking. This setting prevents us from sending too small packets. 185 */ 186 int tcp_minmss = TCP_MINMSS; 187 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW, 188 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size"); 189 190 #if 0 191 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ; 192 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW, 193 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time"); 194 #endif 195 196 int tcp_do_rfc1323 = 1; 197 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW, 198 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions"); 199 200 static int tcp_tcbhashsize = 0; 201 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD, 202 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable"); 203 204 static int do_tcpdrain = 1; 205 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0, 206 "Enable tcp_drain routine for extra help when low on mbufs"); 207 208 static int icmp_may_rst = 1; 209 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0, 210 "Certain ICMP unreachable messages may abort connections in SYN_SENT"); 211 212 static int tcp_isn_reseed_interval = 0; 213 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW, 214 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret"); 215 216 /* 217 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on 218 * by default, but with generous values which should allow maximal 219 * bandwidth. In particular, the slop defaults to 50 (5 packets). 220 * 221 * The reason for doing this is that the limiter is the only mechanism we 222 * have which seems to do a really good job preventing receiver RX rings 223 * on network interfaces from getting blown out. Even though GigE/10GigE 224 * is supposed to flow control it looks like either it doesn't actually 225 * do it or Open Source drivers do not properly enable it. 226 * 227 * People using the limiter to reduce bottlenecks on slower WAN connections 228 * should set the slop to 20 (2 packets). 229 */ 230 static int tcp_inflight_enable = 1; 231 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW, 232 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting"); 233 234 static int tcp_inflight_debug = 0; 235 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW, 236 &tcp_inflight_debug, 0, "Debug TCP inflight calculations"); 237 238 static int tcp_inflight_min = 6144; 239 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW, 240 &tcp_inflight_min, 0, "Lower bound for TCP inflight window"); 241 242 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT; 243 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW, 244 &tcp_inflight_max, 0, "Upper bound for TCP inflight window"); 245 246 static int tcp_inflight_stab = 50; 247 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW, 248 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 3 packets)"); 249 250 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives"); 251 static struct malloc_pipe tcptemp_mpipe; 252 253 static void tcp_willblock(void); 254 static void tcp_notify (struct inpcb *, int); 255 256 struct tcp_stats tcpstats_percpu[MAXCPU]; 257 #ifdef SMP 258 static int 259 sysctl_tcpstats(SYSCTL_HANDLER_ARGS) 260 { 261 int cpu, error = 0; 262 263 for (cpu = 0; cpu < ncpus; ++cpu) { 264 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu], 265 sizeof(struct tcp_stats)))) 266 break; 267 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu], 268 sizeof(struct tcp_stats)))) 269 break; 270 } 271 272 return (error); 273 } 274 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW), 275 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics"); 276 #else 277 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW, 278 &tcpstat, tcp_stats, "TCP statistics"); 279 #endif 280 281 /* 282 * Target size of TCP PCB hash tables. Must be a power of two. 283 * 284 * Note that this can be overridden by the kernel environment 285 * variable net.inet.tcp.tcbhashsize 286 */ 287 #ifndef TCBHASHSIZE 288 #define TCBHASHSIZE 512 289 #endif 290 291 /* 292 * This is the actual shape of what we allocate using the zone 293 * allocator. Doing it this way allows us to protect both structures 294 * using the same generation count, and also eliminates the overhead 295 * of allocating tcpcbs separately. By hiding the structure here, 296 * we avoid changing most of the rest of the code (although it needs 297 * to be changed, eventually, for greater efficiency). 298 */ 299 #define ALIGNMENT 32 300 #define ALIGNM1 (ALIGNMENT - 1) 301 struct inp_tp { 302 union { 303 struct inpcb inp; 304 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1]; 305 } inp_tp_u; 306 struct tcpcb tcb; 307 struct tcp_callout inp_tp_rexmt; 308 struct tcp_callout inp_tp_persist; 309 struct tcp_callout inp_tp_keep; 310 struct tcp_callout inp_tp_2msl; 311 struct tcp_callout inp_tp_delack; 312 struct netmsg_tcp_timer inp_tp_timermsg; 313 }; 314 #undef ALIGNMENT 315 #undef ALIGNM1 316 317 /* 318 * Tcp initialization 319 */ 320 void 321 tcp_init(void) 322 { 323 struct inpcbporthead *porthashbase; 324 struct inpcbinfo *ticb; 325 u_long porthashmask; 326 int hashsize = TCBHASHSIZE; 327 int cpu; 328 329 /* 330 * note: tcptemp is used for keepalives, and it is ok for an 331 * allocation to fail so do not specify MPF_INT. 332 */ 333 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp), 334 25, -1, 0, NULL); 335 336 tcp_delacktime = TCPTV_DELACK; 337 tcp_keepinit = TCPTV_KEEP_INIT; 338 tcp_keepidle = TCPTV_KEEP_IDLE; 339 tcp_keepintvl = TCPTV_KEEPINTVL; 340 tcp_maxpersistidle = TCPTV_KEEP_IDLE; 341 tcp_msl = TCPTV_MSL; 342 tcp_rexmit_min = TCPTV_MIN; 343 tcp_rexmit_slop = TCPTV_CPU_VAR; 344 345 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize); 346 if (!powerof2(hashsize)) { 347 kprintf("WARNING: TCB hash size not a power of 2\n"); 348 hashsize = 512; /* safe default */ 349 } 350 tcp_tcbhashsize = hashsize; 351 porthashbase = hashinit(hashsize, M_PCB, &porthashmask); 352 353 for (cpu = 0; cpu < ncpus2; cpu++) { 354 ticb = &tcbinfo[cpu]; 355 in_pcbinfo_init(ticb); 356 ticb->cpu = cpu; 357 ticb->hashbase = hashinit(hashsize, M_PCB, 358 &ticb->hashmask); 359 ticb->porthashbase = porthashbase; 360 ticb->porthashmask = porthashmask; 361 ticb->porttoken = &tcp_port_token; 362 #if 0 363 ticb->porthashbase = hashinit(hashsize, M_PCB, 364 &ticb->porthashmask); 365 #endif 366 ticb->wildcardhashbase = hashinit(hashsize, M_PCB, 367 &ticb->wildcardhashmask); 368 ticb->ipi_size = sizeof(struct inp_tp); 369 TAILQ_INIT(&tcpcbackq[cpu]); 370 } 371 372 tcp_reass_maxseg = nmbclusters / 16; 373 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg); 374 375 #ifdef INET6 376 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) 377 #else 378 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) 379 #endif 380 if (max_protohdr < TCP_MINPROTOHDR) 381 max_protohdr = TCP_MINPROTOHDR; 382 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) 383 panic("tcp_init"); 384 #undef TCP_MINPROTOHDR 385 386 /* 387 * Initialize TCP statistics counters for each CPU. 388 */ 389 #ifdef SMP 390 for (cpu = 0; cpu < ncpus; ++cpu) { 391 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats)); 392 } 393 #else 394 bzero(&tcpstat, sizeof(struct tcp_stats)); 395 #endif 396 397 syncache_init(); 398 netisr_register_rollup(tcp_willblock); 399 } 400 401 static void 402 tcp_willblock(void) 403 { 404 struct tcpcb *tp; 405 int cpu = mycpu->gd_cpuid; 406 407 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) { 408 KKASSERT(tp->t_flags & TF_ONOUTPUTQ); 409 tp->t_flags &= ~TF_ONOUTPUTQ; 410 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq); 411 tcp_output(tp); 412 } 413 } 414 415 /* 416 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. 417 * tcp_template used to store this data in mbufs, but we now recopy it out 418 * of the tcpcb each time to conserve mbufs. 419 */ 420 void 421 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr) 422 { 423 struct inpcb *inp = tp->t_inpcb; 424 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr; 425 426 #ifdef INET6 427 if (inp->inp_vflag & INP_IPV6) { 428 struct ip6_hdr *ip6; 429 430 ip6 = (struct ip6_hdr *)ip_ptr; 431 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | 432 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK); 433 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | 434 (IPV6_VERSION & IPV6_VERSION_MASK); 435 ip6->ip6_nxt = IPPROTO_TCP; 436 ip6->ip6_plen = sizeof(struct tcphdr); 437 ip6->ip6_src = inp->in6p_laddr; 438 ip6->ip6_dst = inp->in6p_faddr; 439 tcp_hdr->th_sum = 0; 440 } else 441 #endif 442 { 443 struct ip *ip = (struct ip *) ip_ptr; 444 445 ip->ip_vhl = IP_VHL_BORING; 446 ip->ip_tos = 0; 447 ip->ip_len = 0; 448 ip->ip_id = 0; 449 ip->ip_off = 0; 450 ip->ip_ttl = 0; 451 ip->ip_sum = 0; 452 ip->ip_p = IPPROTO_TCP; 453 ip->ip_src = inp->inp_laddr; 454 ip->ip_dst = inp->inp_faddr; 455 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr, 456 ip->ip_dst.s_addr, 457 htons(sizeof(struct tcphdr) + IPPROTO_TCP)); 458 } 459 460 tcp_hdr->th_sport = inp->inp_lport; 461 tcp_hdr->th_dport = inp->inp_fport; 462 tcp_hdr->th_seq = 0; 463 tcp_hdr->th_ack = 0; 464 tcp_hdr->th_x2 = 0; 465 tcp_hdr->th_off = 5; 466 tcp_hdr->th_flags = 0; 467 tcp_hdr->th_win = 0; 468 tcp_hdr->th_urp = 0; 469 } 470 471 /* 472 * Create template to be used to send tcp packets on a connection. 473 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only 474 * use for this function is in keepalives, which use tcp_respond. 475 */ 476 struct tcptemp * 477 tcp_maketemplate(struct tcpcb *tp) 478 { 479 struct tcptemp *tmp; 480 481 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL) 482 return (NULL); 483 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t); 484 return (tmp); 485 } 486 487 void 488 tcp_freetemplate(struct tcptemp *tmp) 489 { 490 mpipe_free(&tcptemp_mpipe, tmp); 491 } 492 493 /* 494 * Send a single message to the TCP at address specified by 495 * the given TCP/IP header. If m == NULL, then we make a copy 496 * of the tcpiphdr at ti and send directly to the addressed host. 497 * This is used to force keep alive messages out using the TCP 498 * template for a connection. If flags are given then we send 499 * a message back to the TCP which originated the * segment ti, 500 * and discard the mbuf containing it and any other attached mbufs. 501 * 502 * In any case the ack and sequence number of the transmitted 503 * segment are as specified by the parameters. 504 * 505 * NOTE: If m != NULL, then ti must point to *inside* the mbuf. 506 */ 507 void 508 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, 509 tcp_seq ack, tcp_seq seq, int flags) 510 { 511 int tlen; 512 int win = 0; 513 struct route *ro = NULL; 514 struct route sro; 515 struct ip *ip = ipgen; 516 struct tcphdr *nth; 517 int ipflags = 0; 518 struct route_in6 *ro6 = NULL; 519 struct route_in6 sro6; 520 struct ip6_hdr *ip6 = ipgen; 521 boolean_t use_tmpro = TRUE; 522 #ifdef INET6 523 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6); 524 #else 525 const boolean_t isipv6 = FALSE; 526 #endif 527 528 if (tp != NULL) { 529 if (!(flags & TH_RST)) { 530 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv); 531 if (win < 0) 532 win = 0; 533 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 534 win = (long)TCP_MAXWIN << tp->rcv_scale; 535 } 536 /* 537 * Don't use the route cache of a listen socket, 538 * it is not MPSAFE; use temporary route cache. 539 */ 540 if (tp->t_state != TCPS_LISTEN) { 541 if (isipv6) 542 ro6 = &tp->t_inpcb->in6p_route; 543 else 544 ro = &tp->t_inpcb->inp_route; 545 use_tmpro = FALSE; 546 } 547 } 548 if (use_tmpro) { 549 if (isipv6) { 550 ro6 = &sro6; 551 bzero(ro6, sizeof *ro6); 552 } else { 553 ro = &sro; 554 bzero(ro, sizeof *ro); 555 } 556 } 557 if (m == NULL) { 558 m = m_gethdr(MB_DONTWAIT, MT_HEADER); 559 if (m == NULL) 560 return; 561 tlen = 0; 562 m->m_data += max_linkhdr; 563 if (isipv6) { 564 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); 565 ip6 = mtod(m, struct ip6_hdr *); 566 nth = (struct tcphdr *)(ip6 + 1); 567 } else { 568 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip)); 569 ip = mtod(m, struct ip *); 570 nth = (struct tcphdr *)(ip + 1); 571 } 572 bcopy(th, nth, sizeof(struct tcphdr)); 573 flags = TH_ACK; 574 } else { 575 m_freem(m->m_next); 576 m->m_next = NULL; 577 m->m_data = (caddr_t)ipgen; 578 /* m_len is set later */ 579 tlen = 0; 580 #define xchg(a, b, type) { type t; t = a; a = b; b = t; } 581 if (isipv6) { 582 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); 583 nth = (struct tcphdr *)(ip6 + 1); 584 } else { 585 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long); 586 nth = (struct tcphdr *)(ip + 1); 587 } 588 if (th != nth) { 589 /* 590 * this is usually a case when an extension header 591 * exists between the IPv6 header and the 592 * TCP header. 593 */ 594 nth->th_sport = th->th_sport; 595 nth->th_dport = th->th_dport; 596 } 597 xchg(nth->th_dport, nth->th_sport, n_short); 598 #undef xchg 599 } 600 if (isipv6) { 601 ip6->ip6_flow = 0; 602 ip6->ip6_vfc = IPV6_VERSION; 603 ip6->ip6_nxt = IPPROTO_TCP; 604 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen)); 605 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 606 } else { 607 tlen += sizeof(struct tcpiphdr); 608 ip->ip_len = tlen; 609 ip->ip_ttl = ip_defttl; 610 } 611 m->m_len = tlen; 612 m->m_pkthdr.len = tlen; 613 m->m_pkthdr.rcvif = NULL; 614 nth->th_seq = htonl(seq); 615 nth->th_ack = htonl(ack); 616 nth->th_x2 = 0; 617 nth->th_off = sizeof(struct tcphdr) >> 2; 618 nth->th_flags = flags; 619 if (tp != NULL) 620 nth->th_win = htons((u_short) (win >> tp->rcv_scale)); 621 else 622 nth->th_win = htons((u_short)win); 623 nth->th_urp = 0; 624 if (isipv6) { 625 nth->th_sum = 0; 626 nth->th_sum = in6_cksum(m, IPPROTO_TCP, 627 sizeof(struct ip6_hdr), 628 tlen - sizeof(struct ip6_hdr)); 629 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL, 630 (ro6 && ro6->ro_rt) ? 631 ro6->ro_rt->rt_ifp : NULL); 632 } else { 633 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 634 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); 635 m->m_pkthdr.csum_flags = CSUM_TCP; 636 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 637 } 638 #ifdef TCPDEBUG 639 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) 640 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); 641 #endif 642 if (isipv6) { 643 ip6_output(m, NULL, ro6, ipflags, NULL, NULL, 644 tp ? tp->t_inpcb : NULL); 645 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) { 646 RTFREE(ro6->ro_rt); 647 ro6->ro_rt = NULL; 648 } 649 } else { 650 ipflags |= IP_DEBUGROUTE; 651 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL); 652 if ((ro == &sro) && (ro->ro_rt != NULL)) { 653 RTFREE(ro->ro_rt); 654 ro->ro_rt = NULL; 655 } 656 } 657 } 658 659 /* 660 * Create a new TCP control block, making an 661 * empty reassembly queue and hooking it to the argument 662 * protocol control block. The `inp' parameter must have 663 * come from the zone allocator set up in tcp_init(). 664 */ 665 struct tcpcb * 666 tcp_newtcpcb(struct inpcb *inp) 667 { 668 struct inp_tp *it; 669 struct tcpcb *tp; 670 #ifdef INET6 671 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 672 #else 673 const boolean_t isipv6 = FALSE; 674 #endif 675 676 it = (struct inp_tp *)inp; 677 tp = &it->tcb; 678 bzero(tp, sizeof(struct tcpcb)); 679 LIST_INIT(&tp->t_segq); 680 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 681 682 /* Set up our timeouts. */ 683 tp->tt_rexmt = &it->inp_tp_rexmt; 684 tp->tt_persist = &it->inp_tp_persist; 685 tp->tt_keep = &it->inp_tp_keep; 686 tp->tt_2msl = &it->inp_tp_2msl; 687 tp->tt_delack = &it->inp_tp_delack; 688 tcp_inittimers(tp); 689 690 /* 691 * Zero out timer message. We don't create it here, 692 * since the current CPU may not be the owner of this 693 * inpcb. 694 */ 695 tp->tt_msg = &it->inp_tp_timermsg; 696 bzero(tp->tt_msg, sizeof(*tp->tt_msg)); 697 698 if (tcp_do_rfc1323) 699 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP); 700 tp->t_inpcb = inp; /* XXX */ 701 tp->t_state = TCPS_CLOSED; 702 /* 703 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 704 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 705 * reasonable initial retransmit time. 706 */ 707 tp->t_srtt = TCPTV_SRTTBASE; 708 tp->t_rttvar = 709 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 710 tp->t_rttmin = tcp_rexmit_min; 711 tp->t_rxtcur = TCPTV_RTOBASE; 712 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 713 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 714 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 715 tp->t_rcvtime = ticks; 716 /* 717 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 718 * because the socket may be bound to an IPv6 wildcard address, 719 * which may match an IPv4-mapped IPv6 address. 720 */ 721 inp->inp_ip_ttl = ip_defttl; 722 inp->inp_ppcb = tp; 723 tcp_sack_tcpcb_init(tp); 724 return (tp); /* XXX */ 725 } 726 727 /* 728 * Drop a TCP connection, reporting the specified error. 729 * If connection is synchronized, then send a RST to peer. 730 */ 731 struct tcpcb * 732 tcp_drop(struct tcpcb *tp, int error) 733 { 734 struct socket *so = tp->t_inpcb->inp_socket; 735 736 if (TCPS_HAVERCVDSYN(tp->t_state)) { 737 tp->t_state = TCPS_CLOSED; 738 tcp_output(tp); 739 tcpstat.tcps_drops++; 740 } else 741 tcpstat.tcps_conndrops++; 742 if (error == ETIMEDOUT && tp->t_softerror) 743 error = tp->t_softerror; 744 so->so_error = error; 745 return (tcp_close(tp)); 746 } 747 748 #ifdef SMP 749 750 struct netmsg_remwildcard { 751 struct netmsg_base base; 752 struct inpcb *nm_inp; 753 struct inpcbinfo *nm_pcbinfo; 754 #if defined(INET6) 755 int nm_isinet6; 756 #else 757 int nm_unused01; 758 #endif 759 }; 760 761 /* 762 * Wildcard inpcb's on SMP boxes must be removed from all cpus before the 763 * inp can be detached. We do this by cycling through the cpus, ending up 764 * on the cpu controlling the inp last and then doing the disconnect. 765 */ 766 static void 767 in_pcbremwildcardhash_handler(netmsg_t msg) 768 { 769 struct netmsg_remwildcard *nmsg = (struct netmsg_remwildcard *)msg; 770 int cpu; 771 772 cpu = nmsg->nm_pcbinfo->cpu; 773 774 if (cpu == nmsg->nm_inp->inp_pcbinfo->cpu) { 775 /* note: detach removes any wildcard hash entry */ 776 #ifdef INET6 777 if (nmsg->nm_isinet6) 778 in6_pcbdetach(nmsg->nm_inp); 779 else 780 #endif 781 in_pcbdetach(nmsg->nm_inp); 782 lwkt_replymsg(&nmsg->base.lmsg, 0); 783 } else { 784 in_pcbremwildcardhash_oncpu(nmsg->nm_inp, nmsg->nm_pcbinfo); 785 cpu = (cpu + 1) % ncpus2; 786 nmsg->nm_pcbinfo = &tcbinfo[cpu]; 787 lwkt_forwardmsg(cpu_portfn(cpu), &nmsg->base.lmsg); 788 } 789 } 790 791 #endif 792 793 /* 794 * Close a TCP control block: 795 * discard all space held by the tcp 796 * discard internet protocol block 797 * wake up any sleepers 798 */ 799 struct tcpcb * 800 tcp_close(struct tcpcb *tp) 801 { 802 struct tseg_qent *q; 803 struct inpcb *inp = tp->t_inpcb; 804 struct socket *so = inp->inp_socket; 805 struct rtentry *rt; 806 boolean_t dosavessthresh; 807 #ifdef SMP 808 int cpu; 809 #endif 810 #ifdef INET6 811 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0); 812 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0); 813 #else 814 const boolean_t isipv6 = FALSE; 815 #endif 816 817 /* 818 * The tp is not instantly destroyed in the wildcard case. Setting 819 * the state to TCPS_TERMINATING will prevent the TCP stack from 820 * messing with it, though it should be noted that this change may 821 * not take effect on other cpus until we have chained the wildcard 822 * hash removal. 823 * 824 * XXX we currently depend on the BGL to synchronize the tp->t_state 825 * update and prevent other tcp protocol threads from accepting new 826 * connections on the listen socket we might be trying to close down. 827 */ 828 KKASSERT(tp->t_state != TCPS_TERMINATING); 829 tp->t_state = TCPS_TERMINATING; 830 831 /* 832 * Make sure that all of our timers are stopped before we 833 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL), 834 * timers are never used. If timer message is never created 835 * (tp->tt_msg->tt_tcb == NULL), timers are never used too. 836 */ 837 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) { 838 tcp_callout_stop(tp, tp->tt_rexmt); 839 tcp_callout_stop(tp, tp->tt_persist); 840 tcp_callout_stop(tp, tp->tt_keep); 841 tcp_callout_stop(tp, tp->tt_2msl); 842 tcp_callout_stop(tp, tp->tt_delack); 843 } 844 845 if (tp->t_flags & TF_ONOUTPUTQ) { 846 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 847 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 848 tp->t_flags &= ~TF_ONOUTPUTQ; 849 } 850 851 /* 852 * If we got enough samples through the srtt filter, 853 * save the rtt and rttvar in the routing entry. 854 * 'Enough' is arbitrarily defined as the 16 samples. 855 * 16 samples is enough for the srtt filter to converge 856 * to within 5% of the correct value; fewer samples and 857 * we could save a very bogus rtt. 858 * 859 * Don't update the default route's characteristics and don't 860 * update anything that the user "locked". 861 */ 862 if (tp->t_rttupdated >= 16) { 863 u_long i = 0; 864 865 if (isipv6) { 866 struct sockaddr_in6 *sin6; 867 868 if ((rt = inp->in6p_route.ro_rt) == NULL) 869 goto no_valid_rt; 870 sin6 = (struct sockaddr_in6 *)rt_key(rt); 871 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 872 goto no_valid_rt; 873 } else 874 if ((rt = inp->inp_route.ro_rt) == NULL || 875 ((struct sockaddr_in *)rt_key(rt))-> 876 sin_addr.s_addr == INADDR_ANY) 877 goto no_valid_rt; 878 879 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 880 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 881 if (rt->rt_rmx.rmx_rtt && i) 882 /* 883 * filter this update to half the old & half 884 * the new values, converting scale. 885 * See route.h and tcp_var.h for a 886 * description of the scaling constants. 887 */ 888 rt->rt_rmx.rmx_rtt = 889 (rt->rt_rmx.rmx_rtt + i) / 2; 890 else 891 rt->rt_rmx.rmx_rtt = i; 892 tcpstat.tcps_cachedrtt++; 893 } 894 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 895 i = tp->t_rttvar * 896 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 897 if (rt->rt_rmx.rmx_rttvar && i) 898 rt->rt_rmx.rmx_rttvar = 899 (rt->rt_rmx.rmx_rttvar + i) / 2; 900 else 901 rt->rt_rmx.rmx_rttvar = i; 902 tcpstat.tcps_cachedrttvar++; 903 } 904 /* 905 * The old comment here said: 906 * update the pipelimit (ssthresh) if it has been updated 907 * already or if a pipesize was specified & the threshhold 908 * got below half the pipesize. I.e., wait for bad news 909 * before we start updating, then update on both good 910 * and bad news. 911 * 912 * But we want to save the ssthresh even if no pipesize is 913 * specified explicitly in the route, because such 914 * connections still have an implicit pipesize specified 915 * by the global tcp_sendspace. In the absence of a reliable 916 * way to calculate the pipesize, it will have to do. 917 */ 918 i = tp->snd_ssthresh; 919 if (rt->rt_rmx.rmx_sendpipe != 0) 920 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 921 else 922 dosavessthresh = (i < so->so_snd.ssb_hiwat/2); 923 if (dosavessthresh || 924 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 925 (rt->rt_rmx.rmx_ssthresh != 0))) { 926 /* 927 * convert the limit from user data bytes to 928 * packets then to packet data bytes. 929 */ 930 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 931 if (i < 2) 932 i = 2; 933 i *= tp->t_maxseg + 934 (isipv6 ? 935 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 936 sizeof(struct tcpiphdr)); 937 if (rt->rt_rmx.rmx_ssthresh) 938 rt->rt_rmx.rmx_ssthresh = 939 (rt->rt_rmx.rmx_ssthresh + i) / 2; 940 else 941 rt->rt_rmx.rmx_ssthresh = i; 942 tcpstat.tcps_cachedssthresh++; 943 } 944 } 945 946 no_valid_rt: 947 /* free the reassembly queue, if any */ 948 while((q = LIST_FIRST(&tp->t_segq)) != NULL) { 949 LIST_REMOVE(q, tqe_q); 950 m_freem(q->tqe_m); 951 FREE(q, M_TSEGQ); 952 atomic_add_int(&tcp_reass_qsize, -1); 953 } 954 /* throw away SACK blocks in scoreboard*/ 955 if (TCP_DO_SACK(tp)) 956 tcp_sack_cleanup(&tp->scb); 957 958 inp->inp_ppcb = NULL; 959 soisdisconnected(so); 960 /* note: pcb detached later on */ 961 962 tcp_destroy_timermsg(tp); 963 if (tp->t_flags & TF_SYNCACHE) 964 syncache_destroy(tp); 965 966 /* 967 * Discard the inp. In the SMP case a wildcard inp's hash (created 968 * by a listen socket or an INADDR_ANY udp socket) is replicated 969 * for each protocol thread and must be removed in the context of 970 * that thread. This is accomplished by chaining the message 971 * through the cpus. 972 * 973 * If the inp is not wildcarded we simply detach, which will remove 974 * the any hashes still present for this inp. 975 */ 976 #ifdef SMP 977 if (inp->inp_flags & INP_WILDCARD_MP) { 978 struct netmsg_remwildcard *nmsg; 979 980 cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2; 981 nmsg = kmalloc(sizeof(struct netmsg_remwildcard), 982 M_LWKTMSG, M_INTWAIT); 983 netmsg_init(&nmsg->base, NULL, &netisr_afree_rport, 984 0, in_pcbremwildcardhash_handler); 985 #ifdef INET6 986 nmsg->nm_isinet6 = isafinet6; 987 #endif 988 nmsg->nm_inp = inp; 989 nmsg->nm_pcbinfo = &tcbinfo[cpu]; 990 lwkt_sendmsg(cpu_portfn(cpu), &nmsg->base.lmsg); 991 } else 992 #endif 993 { 994 /* note: detach removes any wildcard hash entry */ 995 #ifdef INET6 996 if (isafinet6) 997 in6_pcbdetach(inp); 998 else 999 #endif 1000 in_pcbdetach(inp); 1001 } 1002 tcpstat.tcps_closed++; 1003 return (NULL); 1004 } 1005 1006 static __inline void 1007 tcp_drain_oncpu(struct inpcbhead *head) 1008 { 1009 struct inpcb *marker; 1010 struct inpcb *inpb; 1011 struct tcpcb *tcpb; 1012 struct tseg_qent *te; 1013 1014 /* 1015 * Allows us to block while running the list 1016 */ 1017 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1018 marker->inp_flags |= INP_PLACEMARKER; 1019 LIST_INSERT_HEAD(head, marker, inp_list); 1020 1021 while ((inpb = LIST_NEXT(marker, inp_list)) != NULL) { 1022 if ((inpb->inp_flags & INP_PLACEMARKER) == 0 && 1023 (tcpb = intotcpcb(inpb)) != NULL && 1024 (te = LIST_FIRST(&tcpb->t_segq)) != NULL) { 1025 LIST_REMOVE(te, tqe_q); 1026 m_freem(te->tqe_m); 1027 FREE(te, M_TSEGQ); 1028 atomic_add_int(&tcp_reass_qsize, -1); 1029 /* retry */ 1030 } else { 1031 LIST_REMOVE(marker, inp_list); 1032 LIST_INSERT_AFTER(inpb, marker, inp_list); 1033 } 1034 } 1035 LIST_REMOVE(marker, inp_list); 1036 kfree(marker, M_TEMP); 1037 } 1038 1039 #ifdef SMP 1040 struct netmsg_tcp_drain { 1041 struct netmsg_base base; 1042 struct inpcbhead *nm_head; 1043 }; 1044 1045 static void 1046 tcp_drain_handler(netmsg_t msg) 1047 { 1048 struct netmsg_tcp_drain *nm = (void *)msg; 1049 1050 tcp_drain_oncpu(nm->nm_head); 1051 lwkt_replymsg(&nm->base.lmsg, 0); 1052 } 1053 #endif 1054 1055 void 1056 tcp_drain(void) 1057 { 1058 #ifdef SMP 1059 int cpu; 1060 #endif 1061 1062 if (!do_tcpdrain) 1063 return; 1064 1065 /* 1066 * Walk the tcpbs, if existing, and flush the reassembly queue, 1067 * if there is one... 1068 * XXX: The "Net/3" implementation doesn't imply that the TCP 1069 * reassembly queue should be flushed, but in a situation 1070 * where we're really low on mbufs, this is potentially 1071 * useful. 1072 */ 1073 #ifdef SMP 1074 for (cpu = 0; cpu < ncpus2; cpu++) { 1075 struct netmsg_tcp_drain *nm; 1076 1077 if (cpu == mycpu->gd_cpuid) { 1078 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead); 1079 } else { 1080 nm = kmalloc(sizeof(struct netmsg_tcp_drain), 1081 M_LWKTMSG, M_NOWAIT); 1082 if (nm == NULL) 1083 continue; 1084 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1085 0, tcp_drain_handler); 1086 nm->nm_head = &tcbinfo[cpu].pcblisthead; 1087 lwkt_sendmsg(cpu_portfn(cpu), &nm->base.lmsg); 1088 } 1089 } 1090 #else 1091 tcp_drain_oncpu(&tcbinfo[0].pcblisthead); 1092 #endif 1093 } 1094 1095 /* 1096 * Notify a tcp user of an asynchronous error; 1097 * store error as soft error, but wake up user 1098 * (for now, won't do anything until can select for soft error). 1099 * 1100 * Do not wake up user since there currently is no mechanism for 1101 * reporting soft errors (yet - a kqueue filter may be added). 1102 */ 1103 static void 1104 tcp_notify(struct inpcb *inp, int error) 1105 { 1106 struct tcpcb *tp = intotcpcb(inp); 1107 1108 /* 1109 * Ignore some errors if we are hooked up. 1110 * If connection hasn't completed, has retransmitted several times, 1111 * and receives a second error, give up now. This is better 1112 * than waiting a long time to establish a connection that 1113 * can never complete. 1114 */ 1115 if (tp->t_state == TCPS_ESTABLISHED && 1116 (error == EHOSTUNREACH || error == ENETUNREACH || 1117 error == EHOSTDOWN)) { 1118 return; 1119 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1120 tp->t_softerror) 1121 tcp_drop(tp, error); 1122 else 1123 tp->t_softerror = error; 1124 #if 0 1125 wakeup(&so->so_timeo); 1126 sorwakeup(so); 1127 sowwakeup(so); 1128 #endif 1129 } 1130 1131 static int 1132 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1133 { 1134 int error, i, n; 1135 struct inpcb *marker; 1136 struct inpcb *inp; 1137 globaldata_t gd; 1138 int origcpu, ccpu; 1139 1140 error = 0; 1141 n = 0; 1142 1143 /* 1144 * The process of preparing the TCB list is too time-consuming and 1145 * resource-intensive to repeat twice on every request. 1146 */ 1147 if (req->oldptr == NULL) { 1148 for (ccpu = 0; ccpu < ncpus; ++ccpu) { 1149 gd = globaldata_find(ccpu); 1150 n += tcbinfo[gd->gd_cpuid].ipi_count; 1151 } 1152 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1153 return (0); 1154 } 1155 1156 if (req->newptr != NULL) 1157 return (EPERM); 1158 1159 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1160 marker->inp_flags |= INP_PLACEMARKER; 1161 1162 /* 1163 * OK, now we're committed to doing something. Run the inpcb list 1164 * for each cpu in the system and construct the output. Use a 1165 * list placemarker to deal with list changes occuring during 1166 * copyout blockages (but otherwise depend on being on the correct 1167 * cpu to avoid races). 1168 */ 1169 origcpu = mycpu->gd_cpuid; 1170 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) { 1171 globaldata_t rgd; 1172 caddr_t inp_ppcb; 1173 struct xtcpcb xt; 1174 int cpu_id; 1175 1176 cpu_id = (origcpu + ccpu) % ncpus; 1177 if ((smp_active_mask & CPUMASK(cpu_id)) == 0) 1178 continue; 1179 rgd = globaldata_find(cpu_id); 1180 lwkt_setcpu_self(rgd); 1181 1182 n = tcbinfo[cpu_id].ipi_count; 1183 1184 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list); 1185 i = 0; 1186 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1187 /* 1188 * process a snapshot of pcbs, ignoring placemarkers 1189 * and using our own to allow SYSCTL_OUT to block. 1190 */ 1191 LIST_REMOVE(marker, inp_list); 1192 LIST_INSERT_AFTER(inp, marker, inp_list); 1193 1194 if (inp->inp_flags & INP_PLACEMARKER) 1195 continue; 1196 if (prison_xinpcb(req->td, inp)) 1197 continue; 1198 1199 xt.xt_len = sizeof xt; 1200 bcopy(inp, &xt.xt_inp, sizeof *inp); 1201 inp_ppcb = inp->inp_ppcb; 1202 if (inp_ppcb != NULL) 1203 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1204 else 1205 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1206 if (inp->inp_socket) 1207 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1208 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1209 break; 1210 ++i; 1211 } 1212 LIST_REMOVE(marker, inp_list); 1213 if (error == 0 && i < n) { 1214 bzero(&xt, sizeof xt); 1215 xt.xt_len = sizeof xt; 1216 while (i < n) { 1217 error = SYSCTL_OUT(req, &xt, sizeof xt); 1218 if (error) 1219 break; 1220 ++i; 1221 } 1222 } 1223 } 1224 1225 /* 1226 * Make sure we are on the same cpu we were on originally, since 1227 * higher level callers expect this. Also don't pollute caches with 1228 * migrated userland data by (eventually) returning to userland 1229 * on a different cpu. 1230 */ 1231 lwkt_setcpu_self(globaldata_find(origcpu)); 1232 kfree(marker, M_TEMP); 1233 return (error); 1234 } 1235 1236 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1237 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1238 1239 static int 1240 tcp_getcred(SYSCTL_HANDLER_ARGS) 1241 { 1242 struct sockaddr_in addrs[2]; 1243 struct inpcb *inp; 1244 int cpu; 1245 int error; 1246 1247 error = priv_check(req->td, PRIV_ROOT); 1248 if (error != 0) 1249 return (error); 1250 error = SYSCTL_IN(req, addrs, sizeof addrs); 1251 if (error != 0) 1252 return (error); 1253 crit_enter(); 1254 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1255 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1256 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1257 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1258 if (inp == NULL || inp->inp_socket == NULL) { 1259 error = ENOENT; 1260 goto out; 1261 } 1262 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1263 out: 1264 crit_exit(); 1265 return (error); 1266 } 1267 1268 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1269 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1270 1271 #ifdef INET6 1272 static int 1273 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1274 { 1275 struct sockaddr_in6 addrs[2]; 1276 struct inpcb *inp; 1277 int error; 1278 boolean_t mapped = FALSE; 1279 1280 error = priv_check(req->td, PRIV_ROOT); 1281 if (error != 0) 1282 return (error); 1283 error = SYSCTL_IN(req, addrs, sizeof addrs); 1284 if (error != 0) 1285 return (error); 1286 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { 1287 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) 1288 mapped = TRUE; 1289 else 1290 return (EINVAL); 1291 } 1292 crit_enter(); 1293 if (mapped) { 1294 inp = in_pcblookup_hash(&tcbinfo[0], 1295 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], 1296 addrs[1].sin6_port, 1297 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], 1298 addrs[0].sin6_port, 1299 0, NULL); 1300 } else { 1301 inp = in6_pcblookup_hash(&tcbinfo[0], 1302 &addrs[1].sin6_addr, addrs[1].sin6_port, 1303 &addrs[0].sin6_addr, addrs[0].sin6_port, 1304 0, NULL); 1305 } 1306 if (inp == NULL || inp->inp_socket == NULL) { 1307 error = ENOENT; 1308 goto out; 1309 } 1310 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1311 out: 1312 crit_exit(); 1313 return (error); 1314 } 1315 1316 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1317 0, 0, 1318 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1319 #endif 1320 1321 struct netmsg_tcp_notify { 1322 struct netmsg_base base; 1323 void (*nm_notify)(struct inpcb *, int); 1324 struct in_addr nm_faddr; 1325 int nm_arg; 1326 }; 1327 1328 static void 1329 tcp_notifyall_oncpu(netmsg_t msg) 1330 { 1331 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg; 1332 int nextcpu; 1333 1334 in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nm->nm_faddr, 1335 nm->nm_arg, nm->nm_notify); 1336 1337 nextcpu = mycpuid + 1; 1338 if (nextcpu < ncpus2) 1339 lwkt_forwardmsg(cpu_portfn(nextcpu), &nm->base.lmsg); 1340 else 1341 lwkt_replymsg(&nm->base.lmsg, 0); 1342 } 1343 1344 void 1345 tcp_ctlinput(netmsg_t msg) 1346 { 1347 int cmd = msg->ctlinput.nm_cmd; 1348 struct sockaddr *sa = msg->ctlinput.nm_arg; 1349 struct ip *ip = msg->ctlinput.nm_extra; 1350 struct tcphdr *th; 1351 struct in_addr faddr; 1352 struct inpcb *inp; 1353 struct tcpcb *tp; 1354 void (*notify)(struct inpcb *, int) = tcp_notify; 1355 tcp_seq icmpseq; 1356 int arg, cpu; 1357 1358 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) { 1359 goto done; 1360 } 1361 1362 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1363 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1364 goto done; 1365 1366 arg = inetctlerrmap[cmd]; 1367 if (cmd == PRC_QUENCH) { 1368 notify = tcp_quench; 1369 } else if (icmp_may_rst && 1370 (cmd == PRC_UNREACH_ADMIN_PROHIB || 1371 cmd == PRC_UNREACH_PORT || 1372 cmd == PRC_TIMXCEED_INTRANS) && 1373 ip != NULL) { 1374 notify = tcp_drop_syn_sent; 1375 } else if (cmd == PRC_MSGSIZE) { 1376 struct icmp *icmp = (struct icmp *) 1377 ((caddr_t)ip - offsetof(struct icmp, icmp_ip)); 1378 1379 arg = ntohs(icmp->icmp_nextmtu); 1380 notify = tcp_mtudisc; 1381 } else if (PRC_IS_REDIRECT(cmd)) { 1382 ip = NULL; 1383 notify = in_rtchange; 1384 } else if (cmd == PRC_HOSTDEAD) { 1385 ip = NULL; 1386 } 1387 1388 if (ip != NULL) { 1389 crit_enter(); 1390 th = (struct tcphdr *)((caddr_t)ip + 1391 (IP_VHL_HL(ip->ip_vhl) << 2)); 1392 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport, 1393 ip->ip_src.s_addr, th->th_sport); 1394 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport, 1395 ip->ip_src, th->th_sport, 0, NULL); 1396 if ((inp != NULL) && (inp->inp_socket != NULL)) { 1397 icmpseq = htonl(th->th_seq); 1398 tp = intotcpcb(inp); 1399 if (SEQ_GEQ(icmpseq, tp->snd_una) && 1400 SEQ_LT(icmpseq, tp->snd_max)) 1401 (*notify)(inp, arg); 1402 } else { 1403 struct in_conninfo inc; 1404 1405 inc.inc_fport = th->th_dport; 1406 inc.inc_lport = th->th_sport; 1407 inc.inc_faddr = faddr; 1408 inc.inc_laddr = ip->ip_src; 1409 #ifdef INET6 1410 inc.inc_isipv6 = 0; 1411 #endif 1412 syncache_unreach(&inc, th); 1413 } 1414 crit_exit(); 1415 } else { 1416 struct netmsg_tcp_notify *nm; 1417 1418 KKASSERT(&curthread->td_msgport == cpu_portfn(0)); 1419 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT); 1420 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1421 0, tcp_notifyall_oncpu); 1422 nm->nm_faddr = faddr; 1423 nm->nm_arg = arg; 1424 nm->nm_notify = notify; 1425 1426 lwkt_sendmsg(cpu_portfn(0), &nm->base.lmsg); 1427 } 1428 done: 1429 lwkt_replymsg(&msg->lmsg, 0); 1430 } 1431 1432 #ifdef INET6 1433 1434 void 1435 tcp6_ctlinput(netmsg_t msg) 1436 { 1437 int cmd = msg->ctlinput.nm_cmd; 1438 struct sockaddr *sa = msg->ctlinput.nm_arg; 1439 void *d = msg->ctlinput.nm_extra; 1440 struct tcphdr th; 1441 void (*notify) (struct inpcb *, int) = tcp_notify; 1442 struct ip6_hdr *ip6; 1443 struct mbuf *m; 1444 struct ip6ctlparam *ip6cp = NULL; 1445 const struct sockaddr_in6 *sa6_src = NULL; 1446 int off; 1447 struct tcp_portonly { 1448 u_int16_t th_sport; 1449 u_int16_t th_dport; 1450 } *thp; 1451 int arg; 1452 1453 if (sa->sa_family != AF_INET6 || 1454 sa->sa_len != sizeof(struct sockaddr_in6)) { 1455 goto out; 1456 } 1457 1458 arg = 0; 1459 if (cmd == PRC_QUENCH) 1460 notify = tcp_quench; 1461 else if (cmd == PRC_MSGSIZE) { 1462 struct ip6ctlparam *ip6cp = d; 1463 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6; 1464 1465 arg = ntohl(icmp6->icmp6_mtu); 1466 notify = tcp_mtudisc; 1467 } else if (!PRC_IS_REDIRECT(cmd) && 1468 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) { 1469 goto out; 1470 } 1471 1472 /* if the parameter is from icmp6, decode it. */ 1473 if (d != NULL) { 1474 ip6cp = (struct ip6ctlparam *)d; 1475 m = ip6cp->ip6c_m; 1476 ip6 = ip6cp->ip6c_ip6; 1477 off = ip6cp->ip6c_off; 1478 sa6_src = ip6cp->ip6c_src; 1479 } else { 1480 m = NULL; 1481 ip6 = NULL; 1482 off = 0; /* fool gcc */ 1483 sa6_src = &sa6_any; 1484 } 1485 1486 if (ip6 != NULL) { 1487 struct in_conninfo inc; 1488 /* 1489 * XXX: We assume that when IPV6 is non NULL, 1490 * M and OFF are valid. 1491 */ 1492 1493 /* check if we can safely examine src and dst ports */ 1494 if (m->m_pkthdr.len < off + sizeof *thp) 1495 goto out; 1496 1497 bzero(&th, sizeof th); 1498 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1499 1500 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport, 1501 (struct sockaddr *)ip6cp->ip6c_src, 1502 th.th_sport, cmd, arg, notify); 1503 1504 inc.inc_fport = th.th_dport; 1505 inc.inc_lport = th.th_sport; 1506 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1507 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1508 inc.inc_isipv6 = 1; 1509 syncache_unreach(&inc, &th); 1510 } else { 1511 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0, 1512 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify); 1513 } 1514 out: 1515 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0); 1516 } 1517 1518 #endif 1519 1520 /* 1521 * Following is where TCP initial sequence number generation occurs. 1522 * 1523 * There are two places where we must use initial sequence numbers: 1524 * 1. In SYN-ACK packets. 1525 * 2. In SYN packets. 1526 * 1527 * All ISNs for SYN-ACK packets are generated by the syncache. See 1528 * tcp_syncache.c for details. 1529 * 1530 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1531 * depends on this property. In addition, these ISNs should be 1532 * unguessable so as to prevent connection hijacking. To satisfy 1533 * the requirements of this situation, the algorithm outlined in 1534 * RFC 1948 is used to generate sequence numbers. 1535 * 1536 * Implementation details: 1537 * 1538 * Time is based off the system timer, and is corrected so that it 1539 * increases by one megabyte per second. This allows for proper 1540 * recycling on high speed LANs while still leaving over an hour 1541 * before rollover. 1542 * 1543 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1544 * between seeding of isn_secret. This is normally set to zero, 1545 * as reseeding should not be necessary. 1546 * 1547 */ 1548 1549 #define ISN_BYTES_PER_SECOND 1048576 1550 1551 u_char isn_secret[32]; 1552 int isn_last_reseed; 1553 MD5_CTX isn_ctx; 1554 1555 tcp_seq 1556 tcp_new_isn(struct tcpcb *tp) 1557 { 1558 u_int32_t md5_buffer[4]; 1559 tcp_seq new_isn; 1560 1561 /* Seed if this is the first use, reseed if requested. */ 1562 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1563 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1564 < (u_int)ticks))) { 1565 read_random_unlimited(&isn_secret, sizeof isn_secret); 1566 isn_last_reseed = ticks; 1567 } 1568 1569 /* Compute the md5 hash and return the ISN. */ 1570 MD5Init(&isn_ctx); 1571 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1572 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1573 #ifdef INET6 1574 if (tp->t_inpcb->inp_vflag & INP_IPV6) { 1575 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1576 sizeof(struct in6_addr)); 1577 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1578 sizeof(struct in6_addr)); 1579 } else 1580 #endif 1581 { 1582 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1583 sizeof(struct in_addr)); 1584 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1585 sizeof(struct in_addr)); 1586 } 1587 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1588 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1589 new_isn = (tcp_seq) md5_buffer[0]; 1590 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1591 return (new_isn); 1592 } 1593 1594 /* 1595 * When a source quench is received, close congestion window 1596 * to one segment. We will gradually open it again as we proceed. 1597 */ 1598 void 1599 tcp_quench(struct inpcb *inp, int error) 1600 { 1601 struct tcpcb *tp = intotcpcb(inp); 1602 1603 if (tp != NULL) { 1604 tp->snd_cwnd = tp->t_maxseg; 1605 tp->snd_wacked = 0; 1606 } 1607 } 1608 1609 /* 1610 * When a specific ICMP unreachable message is received and the 1611 * connection state is SYN-SENT, drop the connection. This behavior 1612 * is controlled by the icmp_may_rst sysctl. 1613 */ 1614 void 1615 tcp_drop_syn_sent(struct inpcb *inp, int error) 1616 { 1617 struct tcpcb *tp = intotcpcb(inp); 1618 1619 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT)) 1620 tcp_drop(tp, error); 1621 } 1622 1623 /* 1624 * When a `need fragmentation' ICMP is received, update our idea of the MSS 1625 * based on the new value in the route. Also nudge TCP to send something, 1626 * since we know the packet we just sent was dropped. 1627 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1628 */ 1629 void 1630 tcp_mtudisc(struct inpcb *inp, int mtu) 1631 { 1632 struct tcpcb *tp = intotcpcb(inp); 1633 struct rtentry *rt; 1634 struct socket *so = inp->inp_socket; 1635 int maxopd, mss; 1636 #ifdef INET6 1637 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0); 1638 #else 1639 const boolean_t isipv6 = FALSE; 1640 #endif 1641 1642 if (tp == NULL) 1643 return; 1644 1645 /* 1646 * If no MTU is provided in the ICMP message, use the 1647 * next lower likely value, as specified in RFC 1191. 1648 */ 1649 if (mtu == 0) { 1650 int oldmtu; 1651 1652 oldmtu = tp->t_maxopd + 1653 (isipv6 ? 1654 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1655 sizeof(struct tcpiphdr)); 1656 mtu = ip_next_mtu(oldmtu, 0); 1657 } 1658 1659 if (isipv6) 1660 rt = tcp_rtlookup6(&inp->inp_inc); 1661 else 1662 rt = tcp_rtlookup(&inp->inp_inc); 1663 if (rt != NULL) { 1664 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu) 1665 mtu = rt->rt_rmx.rmx_mtu; 1666 1667 maxopd = mtu - 1668 (isipv6 ? 1669 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1670 sizeof(struct tcpiphdr)); 1671 1672 /* 1673 * XXX - The following conditional probably violates the TCP 1674 * spec. The problem is that, since we don't know the 1675 * other end's MSS, we are supposed to use a conservative 1676 * default. But, if we do that, then MTU discovery will 1677 * never actually take place, because the conservative 1678 * default is much less than the MTUs typically seen 1679 * on the Internet today. For the moment, we'll sweep 1680 * this under the carpet. 1681 * 1682 * The conservative default might not actually be a problem 1683 * if the only case this occurs is when sending an initial 1684 * SYN with options and data to a host we've never talked 1685 * to before. Then, they will reply with an MSS value which 1686 * will get recorded and the new parameters should get 1687 * recomputed. For Further Study. 1688 */ 1689 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd) 1690 maxopd = rt->rt_rmx.rmx_mssopt; 1691 } else 1692 maxopd = mtu - 1693 (isipv6 ? 1694 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1695 sizeof(struct tcpiphdr)); 1696 1697 if (tp->t_maxopd <= maxopd) 1698 return; 1699 tp->t_maxopd = maxopd; 1700 1701 mss = maxopd; 1702 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) == 1703 (TF_REQ_TSTMP | TF_RCVD_TSTMP)) 1704 mss -= TCPOLEN_TSTAMP_APPA; 1705 1706 /* round down to multiple of MCLBYTES */ 1707 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */ 1708 if (mss > MCLBYTES) 1709 mss &= ~(MCLBYTES - 1); 1710 #else 1711 if (mss > MCLBYTES) 1712 mss = (mss / MCLBYTES) * MCLBYTES; 1713 #endif 1714 1715 if (so->so_snd.ssb_hiwat < mss) 1716 mss = so->so_snd.ssb_hiwat; 1717 1718 tp->t_maxseg = mss; 1719 tp->t_rtttime = 0; 1720 tp->snd_nxt = tp->snd_una; 1721 tcp_output(tp); 1722 tcpstat.tcps_mturesent++; 1723 } 1724 1725 /* 1726 * Look-up the routing entry to the peer of this inpcb. If no route 1727 * is found and it cannot be allocated the return NULL. This routine 1728 * is called by TCP routines that access the rmx structure and by tcp_mss 1729 * to get the interface MTU. 1730 */ 1731 struct rtentry * 1732 tcp_rtlookup(struct in_conninfo *inc) 1733 { 1734 struct route *ro = &inc->inc_route; 1735 1736 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1737 /* No route yet, so try to acquire one */ 1738 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1739 /* 1740 * unused portions of the structure MUST be zero'd 1741 * out because rtalloc() treats it as opaque data 1742 */ 1743 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1744 ro->ro_dst.sa_family = AF_INET; 1745 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1746 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1747 inc->inc_faddr; 1748 rtalloc(ro); 1749 } 1750 } 1751 return (ro->ro_rt); 1752 } 1753 1754 #ifdef INET6 1755 struct rtentry * 1756 tcp_rtlookup6(struct in_conninfo *inc) 1757 { 1758 struct route_in6 *ro6 = &inc->inc6_route; 1759 1760 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1761 /* No route yet, so try to acquire one */ 1762 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1763 /* 1764 * unused portions of the structure MUST be zero'd 1765 * out because rtalloc() treats it as opaque data 1766 */ 1767 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1768 ro6->ro_dst.sin6_family = AF_INET6; 1769 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1770 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1771 rtalloc((struct route *)ro6); 1772 } 1773 } 1774 return (ro6->ro_rt); 1775 } 1776 #endif 1777 1778 #ifdef IPSEC 1779 /* compute ESP/AH header size for TCP, including outer IP header. */ 1780 size_t 1781 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1782 { 1783 struct inpcb *inp; 1784 struct mbuf *m; 1785 size_t hdrsiz; 1786 struct ip *ip; 1787 struct tcphdr *th; 1788 1789 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1790 return (0); 1791 MGETHDR(m, MB_DONTWAIT, MT_DATA); 1792 if (!m) 1793 return (0); 1794 1795 #ifdef INET6 1796 if (inp->inp_vflag & INP_IPV6) { 1797 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1798 1799 th = (struct tcphdr *)(ip6 + 1); 1800 m->m_pkthdr.len = m->m_len = 1801 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1802 tcp_fillheaders(tp, ip6, th); 1803 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1804 } else 1805 #endif 1806 { 1807 ip = mtod(m, struct ip *); 1808 th = (struct tcphdr *)(ip + 1); 1809 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1810 tcp_fillheaders(tp, ip, th); 1811 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1812 } 1813 1814 m_free(m); 1815 return (hdrsiz); 1816 } 1817 #endif 1818 1819 /* 1820 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1821 * 1822 * This code attempts to calculate the bandwidth-delay product as a 1823 * means of determining the optimal window size to maximize bandwidth, 1824 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1825 * routers. This code also does a fairly good job keeping RTTs in check 1826 * across slow links like modems. We implement an algorithm which is very 1827 * similar (but not meant to be) TCP/Vegas. The code operates on the 1828 * transmitter side of a TCP connection and so only effects the transmit 1829 * side of the connection. 1830 * 1831 * BACKGROUND: TCP makes no provision for the management of buffer space 1832 * at the end points or at the intermediate routers and switches. A TCP 1833 * stream, whether using NewReno or not, will eventually buffer as 1834 * many packets as it is able and the only reason this typically works is 1835 * due to the fairly small default buffers made available for a connection 1836 * (typicaly 16K or 32K). As machines use larger windows and/or window 1837 * scaling it is now fairly easy for even a single TCP connection to blow-out 1838 * all available buffer space not only on the local interface, but on 1839 * intermediate routers and switches as well. NewReno makes a misguided 1840 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1841 * then backing off, then steadily increasing the window again until another 1842 * failure occurs, ad-infinitum. This results in terrible oscillation that 1843 * is only made worse as network loads increase and the idea of intentionally 1844 * blowing out network buffers is, frankly, a terrible way to manage network 1845 * resources. 1846 * 1847 * It is far better to limit the transmit window prior to the failure 1848 * condition being achieved. There are two general ways to do this: First 1849 * you can 'scan' through different transmit window sizes and locate the 1850 * point where the RTT stops increasing, indicating that you have filled the 1851 * pipe, then scan backwards until you note that RTT stops decreasing, then 1852 * repeat ad-infinitum. This method works in principle but has severe 1853 * implementation issues due to RTT variances, timer granularity, and 1854 * instability in the algorithm which can lead to many false positives and 1855 * create oscillations as well as interact badly with other TCP streams 1856 * implementing the same algorithm. 1857 * 1858 * The second method is to limit the window to the bandwidth delay product 1859 * of the link. This is the method we implement. RTT variances and our 1860 * own manipulation of the congestion window, bwnd, can potentially 1861 * destabilize the algorithm. For this reason we have to stabilize the 1862 * elements used to calculate the window. We do this by using the minimum 1863 * observed RTT, the long term average of the observed bandwidth, and 1864 * by adding two segments worth of slop. It isn't perfect but it is able 1865 * to react to changing conditions and gives us a very stable basis on 1866 * which to extend the algorithm. 1867 */ 1868 void 1869 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1870 { 1871 u_long bw; 1872 u_long bwnd; 1873 int save_ticks; 1874 int delta_ticks; 1875 1876 /* 1877 * If inflight_enable is disabled in the middle of a tcp connection, 1878 * make sure snd_bwnd is effectively disabled. 1879 */ 1880 if (!tcp_inflight_enable) { 1881 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1882 tp->snd_bandwidth = 0; 1883 return; 1884 } 1885 1886 /* 1887 * Validate the delta time. If a connection is new or has been idle 1888 * a long time we have to reset the bandwidth calculator. 1889 */ 1890 save_ticks = ticks; 1891 delta_ticks = save_ticks - tp->t_bw_rtttime; 1892 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1893 tp->t_bw_rtttime = ticks; 1894 tp->t_bw_rtseq = ack_seq; 1895 if (tp->snd_bandwidth == 0) 1896 tp->snd_bandwidth = tcp_inflight_min; 1897 return; 1898 } 1899 if (delta_ticks == 0) 1900 return; 1901 1902 /* 1903 * Sanity check, plus ignore pure window update acks. 1904 */ 1905 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1906 return; 1907 1908 /* 1909 * Figure out the bandwidth. Due to the tick granularity this 1910 * is a very rough number and it MUST be averaged over a fairly 1911 * long period of time. XXX we need to take into account a link 1912 * that is not using all available bandwidth, but for now our 1913 * slop will ramp us up if this case occurs and the bandwidth later 1914 * increases. 1915 */ 1916 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 1917 tp->t_bw_rtttime = save_ticks; 1918 tp->t_bw_rtseq = ack_seq; 1919 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4; 1920 1921 tp->snd_bandwidth = bw; 1922 1923 /* 1924 * Calculate the semi-static bandwidth delay product, plus two maximal 1925 * segments. The additional slop puts us squarely in the sweet 1926 * spot and also handles the bandwidth run-up case. Without the 1927 * slop we could be locking ourselves into a lower bandwidth. 1928 * 1929 * Situations Handled: 1930 * (1) Prevents over-queueing of packets on LANs, especially on 1931 * high speed LANs, allowing larger TCP buffers to be 1932 * specified, and also does a good job preventing 1933 * over-queueing of packets over choke points like modems 1934 * (at least for the transmit side). 1935 * 1936 * (2) Is able to handle changing network loads (bandwidth 1937 * drops so bwnd drops, bandwidth increases so bwnd 1938 * increases). 1939 * 1940 * (3) Theoretically should stabilize in the face of multiple 1941 * connections implementing the same algorithm (this may need 1942 * a little work). 1943 * 1944 * (4) Stability value (defaults to 20 = 2 maximal packets) can 1945 * be adjusted with a sysctl but typically only needs to be on 1946 * very slow connections. A value no smaller then 5 should 1947 * be used, but only reduce this default if you have no other 1948 * choice. 1949 */ 1950 1951 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2) 1952 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 1953 tcp_inflight_stab * (int)tp->t_maxseg / 10; 1954 #undef USERTT 1955 1956 if (tcp_inflight_debug > 0) { 1957 static int ltime; 1958 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) { 1959 ltime = ticks; 1960 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n", 1961 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd); 1962 } 1963 } 1964 if ((long)bwnd < tcp_inflight_min) 1965 bwnd = tcp_inflight_min; 1966 if (bwnd > tcp_inflight_max) 1967 bwnd = tcp_inflight_max; 1968 if ((long)bwnd < tp->t_maxseg * 2) 1969 bwnd = tp->t_maxseg * 2; 1970 tp->snd_bwnd = bwnd; 1971 } 1972 1973 #ifdef TCP_SIGNATURE 1974 /* 1975 * Compute TCP-MD5 hash of a TCP segment. (RFC2385) 1976 * 1977 * We do this over ip, tcphdr, segment data, and the key in the SADB. 1978 * When called from tcp_input(), we can be sure that th_sum has been 1979 * zeroed out and verified already. 1980 * 1981 * Return 0 if successful, otherwise return -1. 1982 * 1983 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a 1984 * search with the destination IP address, and a 'magic SPI' to be 1985 * determined by the application. This is hardcoded elsewhere to 1179 1986 * right now. Another branch of this code exists which uses the SPD to 1987 * specify per-application flows but it is unstable. 1988 */ 1989 int 1990 tcpsignature_compute( 1991 struct mbuf *m, /* mbuf chain */ 1992 int len, /* length of TCP data */ 1993 int optlen, /* length of TCP options */ 1994 u_char *buf, /* storage for MD5 digest */ 1995 u_int direction) /* direction of flow */ 1996 { 1997 struct ippseudo ippseudo; 1998 MD5_CTX ctx; 1999 int doff; 2000 struct ip *ip; 2001 struct ipovly *ipovly; 2002 struct secasvar *sav; 2003 struct tcphdr *th; 2004 #ifdef INET6 2005 struct ip6_hdr *ip6; 2006 struct in6_addr in6; 2007 uint32_t plen; 2008 uint16_t nhdr; 2009 #endif /* INET6 */ 2010 u_short savecsum; 2011 2012 KASSERT(m != NULL, ("passed NULL mbuf. Game over.")); 2013 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature")); 2014 /* 2015 * Extract the destination from the IP header in the mbuf. 2016 */ 2017 ip = mtod(m, struct ip *); 2018 #ifdef INET6 2019 ip6 = NULL; /* Make the compiler happy. */ 2020 #endif /* INET6 */ 2021 /* 2022 * Look up an SADB entry which matches the address found in 2023 * the segment. 2024 */ 2025 switch (IP_VHL_V(ip->ip_vhl)) { 2026 case IPVERSION: 2027 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst, 2028 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2029 break; 2030 #ifdef INET6 2031 case (IPV6_VERSION >> 4): 2032 ip6 = mtod(m, struct ip6_hdr *); 2033 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst, 2034 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2035 break; 2036 #endif /* INET6 */ 2037 default: 2038 return (EINVAL); 2039 /* NOTREACHED */ 2040 break; 2041 } 2042 if (sav == NULL) { 2043 kprintf("%s: SADB lookup failed\n", __func__); 2044 return (EINVAL); 2045 } 2046 MD5Init(&ctx); 2047 2048 /* 2049 * Step 1: Update MD5 hash with IP pseudo-header. 2050 * 2051 * XXX The ippseudo header MUST be digested in network byte order, 2052 * or else we'll fail the regression test. Assume all fields we've 2053 * been doing arithmetic on have been in host byte order. 2054 * XXX One cannot depend on ipovly->ih_len here. When called from 2055 * tcp_output(), the underlying ip_len member has not yet been set. 2056 */ 2057 switch (IP_VHL_V(ip->ip_vhl)) { 2058 case IPVERSION: 2059 ipovly = (struct ipovly *)ip; 2060 ippseudo.ippseudo_src = ipovly->ih_src; 2061 ippseudo.ippseudo_dst = ipovly->ih_dst; 2062 ippseudo.ippseudo_pad = 0; 2063 ippseudo.ippseudo_p = IPPROTO_TCP; 2064 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen); 2065 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo)); 2066 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip)); 2067 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen; 2068 break; 2069 #ifdef INET6 2070 /* 2071 * RFC 2385, 2.0 Proposal 2072 * For IPv6, the pseudo-header is as described in RFC 2460, namely the 2073 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero- 2074 * extended next header value (to form 32 bits), and 32-bit segment 2075 * length. 2076 * Note: Upper-Layer Packet Length comes before Next Header. 2077 */ 2078 case (IPV6_VERSION >> 4): 2079 in6 = ip6->ip6_src; 2080 in6_clearscope(&in6); 2081 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2082 in6 = ip6->ip6_dst; 2083 in6_clearscope(&in6); 2084 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2085 plen = htonl(len + sizeof(struct tcphdr) + optlen); 2086 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t)); 2087 nhdr = 0; 2088 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2089 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2090 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2091 nhdr = IPPROTO_TCP; 2092 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2093 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr)); 2094 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen; 2095 break; 2096 #endif /* INET6 */ 2097 default: 2098 return (EINVAL); 2099 /* NOTREACHED */ 2100 break; 2101 } 2102 /* 2103 * Step 2: Update MD5 hash with TCP header, excluding options. 2104 * The TCP checksum must be set to zero. 2105 */ 2106 savecsum = th->th_sum; 2107 th->th_sum = 0; 2108 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr)); 2109 th->th_sum = savecsum; 2110 /* 2111 * Step 3: Update MD5 hash with TCP segment data. 2112 * Use m_apply() to avoid an early m_pullup(). 2113 */ 2114 if (len > 0) 2115 m_apply(m, doff, len, tcpsignature_apply, &ctx); 2116 /* 2117 * Step 4: Update MD5 hash with shared secret. 2118 */ 2119 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth)); 2120 MD5Final(buf, &ctx); 2121 key_sa_recordxfer(sav, m); 2122 key_freesav(sav); 2123 return (0); 2124 } 2125 2126 int 2127 tcpsignature_apply(void *fstate, void *data, unsigned int len) 2128 { 2129 2130 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len); 2131 return (0); 2132 } 2133 #endif /* TCP_SIGNATURE */ 2134