1 /* 2 * Copyright (c) 2003, 2004 Jeffrey M. Hsu. All rights reserved. 3 * Copyright (c) 2003, 2004 The DragonFly Project. All rights reserved. 4 * 5 * This code is derived from software contributed to The DragonFly Project 6 * by Jeffrey M. Hsu. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. Neither the name of The DragonFly Project nor the names of its 17 * contributors may be used to endorse or promote products derived 18 * from this software without specific, prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 22 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 23 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 24 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 25 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 26 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 27 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 28 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 29 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 30 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 */ 33 34 /* 35 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 36 * The Regents of the University of California. All rights reserved. 37 * 38 * Redistribution and use in source and binary forms, with or without 39 * modification, are permitted provided that the following conditions 40 * are met: 41 * 1. Redistributions of source code must retain the above copyright 42 * notice, this list of conditions and the following disclaimer. 43 * 2. Redistributions in binary form must reproduce the above copyright 44 * notice, this list of conditions and the following disclaimer in the 45 * documentation and/or other materials provided with the distribution. 46 * 3. Neither the name of the University nor the names of its contributors 47 * may be used to endorse or promote products derived from this software 48 * without specific prior written permission. 49 * 50 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 51 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 52 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 53 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 54 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 55 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 56 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 57 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 58 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 59 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 60 * SUCH DAMAGE. 61 * 62 * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95 63 * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $ 64 */ 65 66 #include "opt_compat.h" 67 #include "opt_inet.h" 68 #include "opt_inet6.h" 69 #include "opt_ipsec.h" 70 #include "opt_tcpdebug.h" 71 72 #include <sys/param.h> 73 #include <sys/systm.h> 74 #include <sys/callout.h> 75 #include <sys/kernel.h> 76 #include <sys/sysctl.h> 77 #include <sys/malloc.h> 78 #include <sys/mpipe.h> 79 #include <sys/mbuf.h> 80 #ifdef INET6 81 #include <sys/domain.h> 82 #endif 83 #include <sys/proc.h> 84 #include <sys/priv.h> 85 #include <sys/socket.h> 86 #include <sys/socketops.h> 87 #include <sys/socketvar.h> 88 #include <sys/protosw.h> 89 #include <sys/random.h> 90 #include <sys/in_cksum.h> 91 #include <sys/ktr.h> 92 93 #include <net/route.h> 94 #include <net/if.h> 95 #include <net/netisr2.h> 96 97 #define _IP_VHL 98 #include <netinet/in.h> 99 #include <netinet/in_systm.h> 100 #include <netinet/ip.h> 101 #include <netinet/ip6.h> 102 #include <netinet/in_pcb.h> 103 #include <netinet6/in6_pcb.h> 104 #include <netinet/in_var.h> 105 #include <netinet/ip_var.h> 106 #include <netinet6/ip6_var.h> 107 #include <netinet/ip_icmp.h> 108 #ifdef INET6 109 #include <netinet/icmp6.h> 110 #endif 111 #include <netinet/tcp.h> 112 #include <netinet/tcp_fsm.h> 113 #include <netinet/tcp_seq.h> 114 #include <netinet/tcp_timer.h> 115 #include <netinet/tcp_timer2.h> 116 #include <netinet/tcp_var.h> 117 #include <netinet6/tcp6_var.h> 118 #include <netinet/tcpip.h> 119 #ifdef TCPDEBUG 120 #include <netinet/tcp_debug.h> 121 #endif 122 #include <netinet6/ip6protosw.h> 123 124 #ifdef IPSEC 125 #include <netinet6/ipsec.h> 126 #include <netproto/key/key.h> 127 #ifdef INET6 128 #include <netinet6/ipsec6.h> 129 #endif 130 #endif 131 132 #ifdef FAST_IPSEC 133 #include <netproto/ipsec/ipsec.h> 134 #ifdef INET6 135 #include <netproto/ipsec/ipsec6.h> 136 #endif 137 #define IPSEC 138 #endif 139 140 #include <sys/md5.h> 141 #include <machine/smp.h> 142 143 #include <sys/msgport2.h> 144 #include <sys/mplock2.h> 145 #include <net/netmsg2.h> 146 147 #if !defined(KTR_TCP) 148 #define KTR_TCP KTR_ALL 149 #endif 150 /* 151 KTR_INFO_MASTER(tcp); 152 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0); 153 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0); 154 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0); 155 #define logtcp(name) KTR_LOG(tcp_ ## name) 156 */ 157 158 #define TCP_IW_MAXSEGS_DFLT 4 159 #define TCP_IW_CAPSEGS_DFLT 4 160 161 struct inpcbinfo tcbinfo[MAXCPU]; 162 struct tcpcbackqhead tcpcbackq[MAXCPU]; 163 164 int tcp_mssdflt = TCP_MSS; 165 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 166 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size"); 167 168 #ifdef INET6 169 int tcp_v6mssdflt = TCP6_MSS; 170 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW, 171 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6"); 172 #endif 173 174 /* 175 * Minimum MSS we accept and use. This prevents DoS attacks where 176 * we are forced to a ridiculous low MSS like 20 and send hundreds 177 * of packets instead of one. The effect scales with the available 178 * bandwidth and quickly saturates the CPU and network interface 179 * with packet generation and sending. Set to zero to disable MINMSS 180 * checking. This setting prevents us from sending too small packets. 181 */ 182 int tcp_minmss = TCP_MINMSS; 183 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW, 184 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size"); 185 186 #if 0 187 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ; 188 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW, 189 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time"); 190 #endif 191 192 int tcp_do_rfc1323 = 1; 193 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW, 194 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions"); 195 196 static int tcp_tcbhashsize = 0; 197 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD, 198 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable"); 199 200 static int do_tcpdrain = 1; 201 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0, 202 "Enable tcp_drain routine for extra help when low on mbufs"); 203 204 static int icmp_may_rst = 1; 205 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0, 206 "Certain ICMP unreachable messages may abort connections in SYN_SENT"); 207 208 static int tcp_isn_reseed_interval = 0; 209 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW, 210 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret"); 211 212 /* 213 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on 214 * by default, but with generous values which should allow maximal 215 * bandwidth. In particular, the slop defaults to 50 (5 packets). 216 * 217 * The reason for doing this is that the limiter is the only mechanism we 218 * have which seems to do a really good job preventing receiver RX rings 219 * on network interfaces from getting blown out. Even though GigE/10GigE 220 * is supposed to flow control it looks like either it doesn't actually 221 * do it or Open Source drivers do not properly enable it. 222 * 223 * People using the limiter to reduce bottlenecks on slower WAN connections 224 * should set the slop to 20 (2 packets). 225 */ 226 static int tcp_inflight_enable = 1; 227 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW, 228 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting"); 229 230 static int tcp_inflight_debug = 0; 231 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW, 232 &tcp_inflight_debug, 0, "Debug TCP inflight calculations"); 233 234 /* 235 * NOTE: tcp_inflight_start is essentially the starting receive window 236 * for a connection. If set too low then fetches over tcp 237 * connections will take noticably longer to ramp-up over 238 * high-latency connections. 6144 is too low for a default, 239 * use something more reasonable. 240 */ 241 static int tcp_inflight_start = 33792; 242 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_start, CTLFLAG_RW, 243 &tcp_inflight_start, 0, "Start value for TCP inflight window"); 244 245 static int tcp_inflight_min = 6144; 246 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW, 247 &tcp_inflight_min, 0, "Lower bound for TCP inflight window"); 248 249 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT; 250 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW, 251 &tcp_inflight_max, 0, "Upper bound for TCP inflight window"); 252 253 static int tcp_inflight_stab = 50; 254 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW, 255 &tcp_inflight_stab, 0, "Fudge bw 1/10% (50=5%)"); 256 257 static int tcp_inflight_adjrtt = 2; 258 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_adjrtt, CTLFLAG_RW, 259 &tcp_inflight_adjrtt, 0, "Slop for rtt 1/(hz*32)"); 260 261 static int tcp_do_rfc3390 = 1; 262 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_RW, 263 &tcp_do_rfc3390, 0, 264 "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)"); 265 266 static u_long tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 267 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwmaxsegs, CTLFLAG_RW, 268 &tcp_iw_maxsegs, 0, "TCP IW segments max"); 269 270 static u_long tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 271 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwcapsegs, CTLFLAG_RW, 272 &tcp_iw_capsegs, 0, "TCP IW segments"); 273 274 int tcp_low_rtobase = 1; 275 SYSCTL_INT(_net_inet_tcp, OID_AUTO, low_rtobase, CTLFLAG_RW, 276 &tcp_low_rtobase, 0, "Lowering the Initial RTO (RFC 6298)"); 277 278 static int tcp_do_ncr = 1; 279 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr, CTLFLAG_RW, 280 &tcp_do_ncr, 0, "Non-Congestion Robustness (RFC 4653)"); 281 282 int tcp_ncr_rxtthresh_max = 16; 283 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr_rxtthresh_max, CTLFLAG_RW, 284 &tcp_ncr_rxtthresh_max, 0, 285 "Non-Congestion Robustness (RFC 4653), DupThresh upper limit"); 286 287 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives"); 288 static struct malloc_pipe tcptemp_mpipe; 289 290 static void tcp_willblock(void); 291 static void tcp_notify (struct inpcb *, int); 292 293 struct tcp_stats tcpstats_percpu[MAXCPU] __cachealign; 294 295 static struct netmsg_base tcp_drain_netmsg[MAXCPU]; 296 static void tcp_drain_dispatch(netmsg_t nmsg); 297 298 static int 299 sysctl_tcpstats(SYSCTL_HANDLER_ARGS) 300 { 301 int cpu, error = 0; 302 303 for (cpu = 0; cpu < ncpus2; ++cpu) { 304 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu], 305 sizeof(struct tcp_stats)))) 306 break; 307 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu], 308 sizeof(struct tcp_stats)))) 309 break; 310 } 311 312 return (error); 313 } 314 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW), 315 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics"); 316 317 /* 318 * Target size of TCP PCB hash tables. Must be a power of two. 319 * 320 * Note that this can be overridden by the kernel environment 321 * variable net.inet.tcp.tcbhashsize 322 */ 323 #ifndef TCBHASHSIZE 324 #define TCBHASHSIZE 512 325 #endif 326 327 /* 328 * This is the actual shape of what we allocate using the zone 329 * allocator. Doing it this way allows us to protect both structures 330 * using the same generation count, and also eliminates the overhead 331 * of allocating tcpcbs separately. By hiding the structure here, 332 * we avoid changing most of the rest of the code (although it needs 333 * to be changed, eventually, for greater efficiency). 334 */ 335 #define ALIGNMENT 32 336 #define ALIGNM1 (ALIGNMENT - 1) 337 struct inp_tp { 338 union { 339 struct inpcb inp; 340 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1]; 341 } inp_tp_u; 342 struct tcpcb tcb; 343 struct tcp_callout inp_tp_rexmt; 344 struct tcp_callout inp_tp_persist; 345 struct tcp_callout inp_tp_keep; 346 struct tcp_callout inp_tp_2msl; 347 struct tcp_callout inp_tp_delack; 348 struct netmsg_tcp_timer inp_tp_timermsg; 349 struct netmsg_base inp_tp_sndmore; 350 }; 351 #undef ALIGNMENT 352 #undef ALIGNM1 353 354 /* 355 * Tcp initialization 356 */ 357 void 358 tcp_init(void) 359 { 360 struct inpcbportinfo *portinfo; 361 struct inpcbinfo *ticb; 362 int hashsize = TCBHASHSIZE; 363 int cpu; 364 365 /* 366 * note: tcptemp is used for keepalives, and it is ok for an 367 * allocation to fail so do not specify MPF_INT. 368 */ 369 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp), 370 25, -1, 0, NULL, NULL, NULL); 371 372 tcp_delacktime = TCPTV_DELACK; 373 tcp_keepinit = TCPTV_KEEP_INIT; 374 tcp_keepidle = TCPTV_KEEP_IDLE; 375 tcp_keepintvl = TCPTV_KEEPINTVL; 376 tcp_maxpersistidle = TCPTV_KEEP_IDLE; 377 tcp_msl = TCPTV_MSL; 378 tcp_rexmit_min = TCPTV_MIN; 379 tcp_rexmit_slop = TCPTV_CPU_VAR; 380 381 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize); 382 if (!powerof2(hashsize)) { 383 kprintf("WARNING: TCB hash size not a power of 2\n"); 384 hashsize = 512; /* safe default */ 385 } 386 tcp_tcbhashsize = hashsize; 387 388 portinfo = kmalloc_cachealign(sizeof(*portinfo) * ncpus2, M_PCB, 389 M_WAITOK); 390 391 for (cpu = 0; cpu < ncpus2; cpu++) { 392 ticb = &tcbinfo[cpu]; 393 in_pcbinfo_init(ticb, cpu, FALSE); 394 ticb->hashbase = hashinit(hashsize, M_PCB, 395 &ticb->hashmask); 396 in_pcbportinfo_init(&portinfo[cpu], hashsize, cpu); 397 ticb->portinfo = portinfo; 398 ticb->portinfo_mask = ncpus2_mask; 399 ticb->wildcardhashbase = hashinit(hashsize, M_PCB, 400 &ticb->wildcardhashmask); 401 ticb->localgrphashbase = hashinit(hashsize, M_PCB, 402 &ticb->localgrphashmask); 403 ticb->ipi_size = sizeof(struct inp_tp); 404 TAILQ_INIT(&tcpcbackq[cpu]); 405 } 406 407 tcp_reass_maxseg = nmbclusters / 16; 408 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg); 409 410 #ifdef INET6 411 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) 412 #else 413 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) 414 #endif 415 if (max_protohdr < TCP_MINPROTOHDR) 416 max_protohdr = TCP_MINPROTOHDR; 417 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) 418 panic("tcp_init"); 419 #undef TCP_MINPROTOHDR 420 421 /* 422 * Initialize TCP statistics counters for each CPU. 423 */ 424 for (cpu = 0; cpu < ncpus2; ++cpu) 425 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats)); 426 427 /* 428 * Initialize netmsgs for TCP drain 429 */ 430 for (cpu = 0; cpu < ncpus2; ++cpu) { 431 netmsg_init(&tcp_drain_netmsg[cpu], NULL, &netisr_adone_rport, 432 MSGF_PRIORITY, tcp_drain_dispatch); 433 } 434 435 syncache_init(); 436 netisr_register_rollup(tcp_willblock, NETISR_ROLLUP_PRIO_TCP); 437 } 438 439 static void 440 tcp_willblock(void) 441 { 442 struct tcpcb *tp; 443 int cpu = mycpu->gd_cpuid; 444 445 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) { 446 KKASSERT(tp->t_flags & TF_ONOUTPUTQ); 447 tp->t_flags &= ~TF_ONOUTPUTQ; 448 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq); 449 tcp_output(tp); 450 } 451 } 452 453 /* 454 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. 455 * tcp_template used to store this data in mbufs, but we now recopy it out 456 * of the tcpcb each time to conserve mbufs. 457 */ 458 void 459 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr, boolean_t tso) 460 { 461 struct inpcb *inp = tp->t_inpcb; 462 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr; 463 464 #ifdef INET6 465 if (INP_ISIPV6(inp)) { 466 struct ip6_hdr *ip6; 467 468 ip6 = (struct ip6_hdr *)ip_ptr; 469 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | 470 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK); 471 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | 472 (IPV6_VERSION & IPV6_VERSION_MASK); 473 ip6->ip6_nxt = IPPROTO_TCP; 474 ip6->ip6_plen = sizeof(struct tcphdr); 475 ip6->ip6_src = inp->in6p_laddr; 476 ip6->ip6_dst = inp->in6p_faddr; 477 tcp_hdr->th_sum = 0; 478 } else 479 #endif 480 { 481 struct ip *ip = (struct ip *) ip_ptr; 482 u_int plen; 483 484 ip->ip_vhl = IP_VHL_BORING; 485 ip->ip_tos = 0; 486 ip->ip_len = 0; 487 ip->ip_id = 0; 488 ip->ip_off = 0; 489 ip->ip_ttl = 0; 490 ip->ip_sum = 0; 491 ip->ip_p = IPPROTO_TCP; 492 ip->ip_src = inp->inp_laddr; 493 ip->ip_dst = inp->inp_faddr; 494 495 if (tso) 496 plen = htons(IPPROTO_TCP); 497 else 498 plen = htons(sizeof(struct tcphdr) + IPPROTO_TCP); 499 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr, 500 ip->ip_dst.s_addr, plen); 501 } 502 503 tcp_hdr->th_sport = inp->inp_lport; 504 tcp_hdr->th_dport = inp->inp_fport; 505 tcp_hdr->th_seq = 0; 506 tcp_hdr->th_ack = 0; 507 tcp_hdr->th_x2 = 0; 508 tcp_hdr->th_off = 5; 509 tcp_hdr->th_flags = 0; 510 tcp_hdr->th_win = 0; 511 tcp_hdr->th_urp = 0; 512 } 513 514 /* 515 * Create template to be used to send tcp packets on a connection. 516 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only 517 * use for this function is in keepalives, which use tcp_respond. 518 */ 519 struct tcptemp * 520 tcp_maketemplate(struct tcpcb *tp) 521 { 522 struct tcptemp *tmp; 523 524 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL) 525 return (NULL); 526 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t, FALSE); 527 return (tmp); 528 } 529 530 void 531 tcp_freetemplate(struct tcptemp *tmp) 532 { 533 mpipe_free(&tcptemp_mpipe, tmp); 534 } 535 536 /* 537 * Send a single message to the TCP at address specified by 538 * the given TCP/IP header. If m == NULL, then we make a copy 539 * of the tcpiphdr at ti and send directly to the addressed host. 540 * This is used to force keep alive messages out using the TCP 541 * template for a connection. If flags are given then we send 542 * a message back to the TCP which originated the * segment ti, 543 * and discard the mbuf containing it and any other attached mbufs. 544 * 545 * In any case the ack and sequence number of the transmitted 546 * segment are as specified by the parameters. 547 * 548 * NOTE: If m != NULL, then ti must point to *inside* the mbuf. 549 */ 550 void 551 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, 552 tcp_seq ack, tcp_seq seq, int flags) 553 { 554 int tlen; 555 long win = 0; 556 struct route *ro = NULL; 557 struct route sro; 558 struct ip *ip = ipgen; 559 struct tcphdr *nth; 560 int ipflags = 0; 561 struct route_in6 *ro6 = NULL; 562 struct route_in6 sro6; 563 struct ip6_hdr *ip6 = ipgen; 564 boolean_t use_tmpro = TRUE; 565 #ifdef INET6 566 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6); 567 #else 568 const boolean_t isipv6 = FALSE; 569 #endif 570 571 if (tp != NULL) { 572 if (!(flags & TH_RST)) { 573 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv); 574 if (win < 0) 575 win = 0; 576 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 577 win = (long)TCP_MAXWIN << tp->rcv_scale; 578 } 579 /* 580 * Don't use the route cache of a listen socket, 581 * it is not MPSAFE; use temporary route cache. 582 */ 583 if (tp->t_state != TCPS_LISTEN) { 584 if (isipv6) 585 ro6 = &tp->t_inpcb->in6p_route; 586 else 587 ro = &tp->t_inpcb->inp_route; 588 use_tmpro = FALSE; 589 } 590 } 591 if (use_tmpro) { 592 if (isipv6) { 593 ro6 = &sro6; 594 bzero(ro6, sizeof *ro6); 595 } else { 596 ro = &sro; 597 bzero(ro, sizeof *ro); 598 } 599 } 600 if (m == NULL) { 601 m = m_gethdr(M_NOWAIT, MT_HEADER); 602 if (m == NULL) 603 return; 604 tlen = 0; 605 m->m_data += max_linkhdr; 606 if (isipv6) { 607 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); 608 ip6 = mtod(m, struct ip6_hdr *); 609 nth = (struct tcphdr *)(ip6 + 1); 610 } else { 611 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip)); 612 ip = mtod(m, struct ip *); 613 nth = (struct tcphdr *)(ip + 1); 614 } 615 bcopy(th, nth, sizeof(struct tcphdr)); 616 flags = TH_ACK; 617 } else { 618 m_freem(m->m_next); 619 m->m_next = NULL; 620 m->m_data = (caddr_t)ipgen; 621 /* m_len is set later */ 622 tlen = 0; 623 #define xchg(a, b, type) { type t; t = a; a = b; b = t; } 624 if (isipv6) { 625 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); 626 nth = (struct tcphdr *)(ip6 + 1); 627 } else { 628 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long); 629 nth = (struct tcphdr *)(ip + 1); 630 } 631 if (th != nth) { 632 /* 633 * this is usually a case when an extension header 634 * exists between the IPv6 header and the 635 * TCP header. 636 */ 637 nth->th_sport = th->th_sport; 638 nth->th_dport = th->th_dport; 639 } 640 xchg(nth->th_dport, nth->th_sport, n_short); 641 #undef xchg 642 } 643 if (isipv6) { 644 ip6->ip6_flow = 0; 645 ip6->ip6_vfc = IPV6_VERSION; 646 ip6->ip6_nxt = IPPROTO_TCP; 647 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen)); 648 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 649 } else { 650 tlen += sizeof(struct tcpiphdr); 651 ip->ip_len = tlen; 652 ip->ip_ttl = ip_defttl; 653 } 654 m->m_len = tlen; 655 m->m_pkthdr.len = tlen; 656 m->m_pkthdr.rcvif = NULL; 657 nth->th_seq = htonl(seq); 658 nth->th_ack = htonl(ack); 659 nth->th_x2 = 0; 660 nth->th_off = sizeof(struct tcphdr) >> 2; 661 nth->th_flags = flags; 662 if (tp != NULL) 663 nth->th_win = htons((u_short) (win >> tp->rcv_scale)); 664 else 665 nth->th_win = htons((u_short)win); 666 nth->th_urp = 0; 667 if (isipv6) { 668 nth->th_sum = 0; 669 nth->th_sum = in6_cksum(m, IPPROTO_TCP, 670 sizeof(struct ip6_hdr), 671 tlen - sizeof(struct ip6_hdr)); 672 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL, 673 (ro6 && ro6->ro_rt) ? 674 ro6->ro_rt->rt_ifp : NULL); 675 } else { 676 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 677 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); 678 m->m_pkthdr.csum_flags = CSUM_TCP; 679 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 680 m->m_pkthdr.csum_thlen = sizeof(struct tcphdr); 681 } 682 #ifdef TCPDEBUG 683 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) 684 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); 685 #endif 686 if (isipv6) { 687 ip6_output(m, NULL, ro6, ipflags, NULL, NULL, 688 tp ? tp->t_inpcb : NULL); 689 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) { 690 RTFREE(ro6->ro_rt); 691 ro6->ro_rt = NULL; 692 } 693 } else { 694 ipflags |= IP_DEBUGROUTE; 695 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL); 696 if ((ro == &sro) && (ro->ro_rt != NULL)) { 697 RTFREE(ro->ro_rt); 698 ro->ro_rt = NULL; 699 } 700 } 701 } 702 703 /* 704 * Create a new TCP control block, making an 705 * empty reassembly queue and hooking it to the argument 706 * protocol control block. The `inp' parameter must have 707 * come from the zone allocator set up in tcp_init(). 708 */ 709 struct tcpcb * 710 tcp_newtcpcb(struct inpcb *inp) 711 { 712 struct inp_tp *it; 713 struct tcpcb *tp; 714 #ifdef INET6 715 boolean_t isipv6 = INP_ISIPV6(inp); 716 #else 717 const boolean_t isipv6 = FALSE; 718 #endif 719 720 it = (struct inp_tp *)inp; 721 tp = &it->tcb; 722 bzero(tp, sizeof(struct tcpcb)); 723 TAILQ_INIT(&tp->t_segq); 724 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 725 tp->t_rxtthresh = tcprexmtthresh; 726 727 /* Set up our timeouts. */ 728 tp->tt_rexmt = &it->inp_tp_rexmt; 729 tp->tt_persist = &it->inp_tp_persist; 730 tp->tt_keep = &it->inp_tp_keep; 731 tp->tt_2msl = &it->inp_tp_2msl; 732 tp->tt_delack = &it->inp_tp_delack; 733 tcp_inittimers(tp); 734 735 /* 736 * Zero out timer message. We don't create it here, 737 * since the current CPU may not be the owner of this 738 * inpcb. 739 */ 740 tp->tt_msg = &it->inp_tp_timermsg; 741 bzero(tp->tt_msg, sizeof(*tp->tt_msg)); 742 743 tp->t_keepinit = tcp_keepinit; 744 tp->t_keepidle = tcp_keepidle; 745 tp->t_keepintvl = tcp_keepintvl; 746 tp->t_keepcnt = tcp_keepcnt; 747 tp->t_maxidle = tp->t_keepintvl * tp->t_keepcnt; 748 749 if (tcp_do_ncr) 750 tp->t_flags |= TF_NCR; 751 if (tcp_do_rfc1323) 752 tp->t_flags |= (TF_REQ_SCALE | TF_REQ_TSTMP); 753 754 tp->t_inpcb = inp; /* XXX */ 755 tp->t_state = TCPS_CLOSED; 756 /* 757 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 758 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 759 * reasonable initial retransmit time. 760 */ 761 tp->t_srtt = TCPTV_SRTTBASE; 762 tp->t_rttvar = 763 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 764 tp->t_rttmin = tcp_rexmit_min; 765 tp->t_rxtcur = TCPTV_RTOBASE; 766 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 767 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 768 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 769 tp->snd_last = ticks; 770 tp->t_rcvtime = ticks; 771 /* 772 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 773 * because the socket may be bound to an IPv6 wildcard address, 774 * which may match an IPv4-mapped IPv6 address. 775 */ 776 inp->inp_ip_ttl = ip_defttl; 777 inp->inp_ppcb = tp; 778 tcp_sack_tcpcb_init(tp); 779 780 tp->tt_sndmore = &it->inp_tp_sndmore; 781 tcp_output_init(tp); 782 783 return (tp); /* XXX */ 784 } 785 786 /* 787 * Drop a TCP connection, reporting the specified error. 788 * If connection is synchronized, then send a RST to peer. 789 */ 790 struct tcpcb * 791 tcp_drop(struct tcpcb *tp, int error) 792 { 793 struct socket *so = tp->t_inpcb->inp_socket; 794 795 if (TCPS_HAVERCVDSYN(tp->t_state)) { 796 tp->t_state = TCPS_CLOSED; 797 tcp_output(tp); 798 tcpstat.tcps_drops++; 799 } else 800 tcpstat.tcps_conndrops++; 801 if (error == ETIMEDOUT && tp->t_softerror) 802 error = tp->t_softerror; 803 so->so_error = error; 804 return (tcp_close(tp)); 805 } 806 807 struct netmsg_listen_detach { 808 struct netmsg_base base; 809 struct tcpcb *nm_tp; 810 struct tcpcb *nm_tp_inh; 811 }; 812 813 static void 814 tcp_listen_detach_handler(netmsg_t msg) 815 { 816 struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg; 817 struct tcpcb *tp = nmsg->nm_tp; 818 int cpu = mycpuid, nextcpu; 819 820 if (tp->t_flags & TF_LISTEN) { 821 syncache_destroy(tp, nmsg->nm_tp_inh); 822 tcp_pcbport_merge_oncpu(tp); 823 } 824 825 in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]); 826 827 nextcpu = cpu + 1; 828 if (nextcpu < ncpus2) 829 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nmsg->base.lmsg); 830 else 831 lwkt_replymsg(&nmsg->base.lmsg, 0); 832 } 833 834 /* 835 * Close a TCP control block: 836 * discard all space held by the tcp 837 * discard internet protocol block 838 * wake up any sleepers 839 */ 840 struct tcpcb * 841 tcp_close(struct tcpcb *tp) 842 { 843 struct tseg_qent *q; 844 struct inpcb *inp = tp->t_inpcb; 845 struct inpcb *inp_inh = NULL; 846 struct tcpcb *tp_inh = NULL; 847 struct socket *so = inp->inp_socket; 848 struct rtentry *rt; 849 boolean_t dosavessthresh; 850 #ifdef INET6 851 boolean_t isipv6 = INP_ISIPV6(inp); 852 #else 853 const boolean_t isipv6 = FALSE; 854 #endif 855 856 if (tp->t_flags & TF_LISTEN) { 857 /* 858 * Pending socket/syncache inheritance 859 * 860 * If this is a listen(2) socket, find another listen(2) 861 * socket in the same local group, which could inherit 862 * the syncache and sockets pending on the completion 863 * and incompletion queues. 864 * 865 * NOTE: 866 * Currently the inheritance could only happen on the 867 * listen(2) sockets w/ SO_REUSEPORT set. 868 */ 869 ASSERT_IN_NETISR(0); 870 inp_inh = in_pcblocalgroup_last(&tcbinfo[0], inp); 871 if (inp_inh != NULL) 872 tp_inh = intotcpcb(inp_inh); 873 } 874 875 /* 876 * INP_WILDCARD indicates that listen(2) has been called on 877 * this socket. This implies: 878 * - A wildcard inp's hash is replicated for each protocol thread. 879 * - Syncache for this inp grows independently in each protocol 880 * thread. 881 * - There is more than one cpu 882 * 883 * We have to chain a message to the rest of the protocol threads 884 * to cleanup the wildcard hash and the syncache. The cleanup 885 * in the current protocol thread is defered till the end of this 886 * function (syncache_destroy and in_pcbdetach). 887 * 888 * NOTE: 889 * After cleanup the inp's hash and syncache entries, this inp will 890 * no longer be available to the rest of the protocol threads, so we 891 * are safe to whack the inp in the following code. 892 */ 893 if ((inp->inp_flags & INP_WILDCARD) && ncpus2 > 1) { 894 struct netmsg_listen_detach nmsg; 895 896 KKASSERT(so->so_port == netisr_cpuport(0)); 897 ASSERT_IN_NETISR(0); 898 KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]); 899 900 netmsg_init(&nmsg.base, NULL, &curthread->td_msgport, 901 MSGF_PRIORITY, tcp_listen_detach_handler); 902 nmsg.nm_tp = tp; 903 nmsg.nm_tp_inh = tp_inh; 904 lwkt_domsg(netisr_cpuport(1), &nmsg.base.lmsg, 0); 905 } 906 907 KKASSERT(tp->t_state != TCPS_TERMINATING); 908 tp->t_state = TCPS_TERMINATING; 909 910 /* 911 * Make sure that all of our timers are stopped before we 912 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL), 913 * timers are never used. If timer message is never created 914 * (tp->tt_msg->tt_tcb == NULL), timers are never used too. 915 */ 916 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) { 917 tcp_callout_stop(tp, tp->tt_rexmt); 918 tcp_callout_stop(tp, tp->tt_persist); 919 tcp_callout_stop(tp, tp->tt_keep); 920 tcp_callout_stop(tp, tp->tt_2msl); 921 tcp_callout_stop(tp, tp->tt_delack); 922 } 923 924 if (tp->t_flags & TF_ONOUTPUTQ) { 925 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 926 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 927 tp->t_flags &= ~TF_ONOUTPUTQ; 928 } 929 930 /* 931 * If we got enough samples through the srtt filter, 932 * save the rtt and rttvar in the routing entry. 933 * 'Enough' is arbitrarily defined as the 16 samples. 934 * 16 samples is enough for the srtt filter to converge 935 * to within 5% of the correct value; fewer samples and 936 * we could save a very bogus rtt. 937 * 938 * Don't update the default route's characteristics and don't 939 * update anything that the user "locked". 940 */ 941 if (tp->t_rttupdated >= 16) { 942 u_long i = 0; 943 944 if (isipv6) { 945 struct sockaddr_in6 *sin6; 946 947 if ((rt = inp->in6p_route.ro_rt) == NULL) 948 goto no_valid_rt; 949 sin6 = (struct sockaddr_in6 *)rt_key(rt); 950 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 951 goto no_valid_rt; 952 } else 953 if ((rt = inp->inp_route.ro_rt) == NULL || 954 ((struct sockaddr_in *)rt_key(rt))-> 955 sin_addr.s_addr == INADDR_ANY) 956 goto no_valid_rt; 957 958 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 959 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 960 if (rt->rt_rmx.rmx_rtt && i) 961 /* 962 * filter this update to half the old & half 963 * the new values, converting scale. 964 * See route.h and tcp_var.h for a 965 * description of the scaling constants. 966 */ 967 rt->rt_rmx.rmx_rtt = 968 (rt->rt_rmx.rmx_rtt + i) / 2; 969 else 970 rt->rt_rmx.rmx_rtt = i; 971 tcpstat.tcps_cachedrtt++; 972 } 973 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 974 i = tp->t_rttvar * 975 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 976 if (rt->rt_rmx.rmx_rttvar && i) 977 rt->rt_rmx.rmx_rttvar = 978 (rt->rt_rmx.rmx_rttvar + i) / 2; 979 else 980 rt->rt_rmx.rmx_rttvar = i; 981 tcpstat.tcps_cachedrttvar++; 982 } 983 /* 984 * The old comment here said: 985 * update the pipelimit (ssthresh) if it has been updated 986 * already or if a pipesize was specified & the threshhold 987 * got below half the pipesize. I.e., wait for bad news 988 * before we start updating, then update on both good 989 * and bad news. 990 * 991 * But we want to save the ssthresh even if no pipesize is 992 * specified explicitly in the route, because such 993 * connections still have an implicit pipesize specified 994 * by the global tcp_sendspace. In the absence of a reliable 995 * way to calculate the pipesize, it will have to do. 996 */ 997 i = tp->snd_ssthresh; 998 if (rt->rt_rmx.rmx_sendpipe != 0) 999 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 1000 else 1001 dosavessthresh = (i < so->so_snd.ssb_hiwat/2); 1002 if (dosavessthresh || 1003 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 1004 (rt->rt_rmx.rmx_ssthresh != 0))) { 1005 /* 1006 * convert the limit from user data bytes to 1007 * packets then to packet data bytes. 1008 */ 1009 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 1010 if (i < 2) 1011 i = 2; 1012 i *= tp->t_maxseg + 1013 (isipv6 ? 1014 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1015 sizeof(struct tcpiphdr)); 1016 if (rt->rt_rmx.rmx_ssthresh) 1017 rt->rt_rmx.rmx_ssthresh = 1018 (rt->rt_rmx.rmx_ssthresh + i) / 2; 1019 else 1020 rt->rt_rmx.rmx_ssthresh = i; 1021 tcpstat.tcps_cachedssthresh++; 1022 } 1023 } 1024 1025 no_valid_rt: 1026 /* free the reassembly queue, if any */ 1027 while((q = TAILQ_FIRST(&tp->t_segq)) != NULL) { 1028 TAILQ_REMOVE(&tp->t_segq, q, tqe_q); 1029 m_freem(q->tqe_m); 1030 kfree(q, M_TSEGQ); 1031 atomic_add_int(&tcp_reass_qsize, -1); 1032 } 1033 /* throw away SACK blocks in scoreboard*/ 1034 if (TCP_DO_SACK(tp)) 1035 tcp_sack_destroy(&tp->scb); 1036 1037 inp->inp_ppcb = NULL; 1038 soisdisconnected(so); 1039 /* note: pcb detached later on */ 1040 1041 tcp_destroy_timermsg(tp); 1042 tcp_output_cancel(tp); 1043 1044 if (tp->t_flags & TF_LISTEN) { 1045 syncache_destroy(tp, tp_inh); 1046 tcp_pcbport_merge_oncpu(tp); 1047 tcp_pcbport_destroy(tp); 1048 if (inp_inh != NULL && inp_inh->inp_socket != NULL) { 1049 /* 1050 * Pending sockets inheritance only needs 1051 * to be done once in the current thread, 1052 * i.e. netisr0. 1053 */ 1054 soinherit(so, inp_inh->inp_socket); 1055 } 1056 } 1057 KASSERT(tp->t_pcbport == NULL, ("tcpcb port cache is not destroyed")); 1058 1059 so_async_rcvd_drop(so); 1060 /* Drop the reference for the asynchronized pru_rcvd */ 1061 sofree(so); 1062 1063 /* 1064 * NOTE: 1065 * - Remove self from listen tcpcb per-cpu port cache _before_ 1066 * pcbdetach. 1067 * - pcbdetach removes any wildcard hash entry on the current CPU. 1068 */ 1069 tcp_pcbport_remove(inp); 1070 #ifdef INET6 1071 if (isipv6) 1072 in6_pcbdetach(inp); 1073 else 1074 #endif 1075 in_pcbdetach(inp); 1076 1077 tcpstat.tcps_closed++; 1078 return (NULL); 1079 } 1080 1081 static __inline void 1082 tcp_drain_oncpu(struct inpcbinfo *pcbinfo) 1083 { 1084 struct inpcbhead *head = &pcbinfo->pcblisthead; 1085 struct inpcb *inpb; 1086 1087 /* 1088 * Since we run in netisr, it is MP safe, even if 1089 * we block during the inpcb list iteration, i.e. 1090 * we don't need to use inpcb marker here. 1091 */ 1092 ASSERT_IN_NETISR(pcbinfo->cpu); 1093 1094 LIST_FOREACH(inpb, head, inp_list) { 1095 struct tcpcb *tcpb; 1096 struct tseg_qent *te; 1097 1098 if (inpb->inp_flags & INP_PLACEMARKER) 1099 continue; 1100 1101 tcpb = intotcpcb(inpb); 1102 KASSERT(tcpb != NULL, ("tcp_drain_oncpu: tcpb is NULL")); 1103 1104 if ((te = TAILQ_FIRST(&tcpb->t_segq)) != NULL) { 1105 TAILQ_REMOVE(&tcpb->t_segq, te, tqe_q); 1106 if (te->tqe_th->th_flags & TH_FIN) 1107 tcpb->t_flags &= ~TF_QUEDFIN; 1108 m_freem(te->tqe_m); 1109 kfree(te, M_TSEGQ); 1110 atomic_add_int(&tcp_reass_qsize, -1); 1111 /* retry */ 1112 } 1113 } 1114 } 1115 1116 static void 1117 tcp_drain_dispatch(netmsg_t nmsg) 1118 { 1119 crit_enter(); 1120 lwkt_replymsg(&nmsg->lmsg, 0); /* reply ASAP */ 1121 crit_exit(); 1122 1123 tcp_drain_oncpu(&tcbinfo[mycpuid]); 1124 } 1125 1126 static void 1127 tcp_drain_ipi(void *arg __unused) 1128 { 1129 int cpu = mycpuid; 1130 struct lwkt_msg *msg = &tcp_drain_netmsg[cpu].lmsg; 1131 1132 crit_enter(); 1133 if (msg->ms_flags & MSGF_DONE) 1134 lwkt_sendmsg_oncpu(netisr_cpuport(cpu), msg); 1135 crit_exit(); 1136 } 1137 1138 void 1139 tcp_drain(void) 1140 { 1141 cpumask_t mask; 1142 1143 if (!do_tcpdrain) 1144 return; 1145 1146 /* 1147 * Walk the tcpbs, if existing, and flush the reassembly queue, 1148 * if there is one... 1149 * XXX: The "Net/3" implementation doesn't imply that the TCP 1150 * reassembly queue should be flushed, but in a situation 1151 * where we're really low on mbufs, this is potentially 1152 * useful. 1153 * YYY: We may consider run tcp_drain_oncpu directly here, 1154 * however, that will require M_WAITOK memory allocation 1155 * for the inpcb marker. 1156 */ 1157 CPUMASK_ASSBMASK(mask, ncpus2); 1158 CPUMASK_ANDMASK(mask, smp_active_mask); 1159 if (CPUMASK_TESTNZERO(mask)) 1160 lwkt_send_ipiq_mask(mask, tcp_drain_ipi, NULL); 1161 } 1162 1163 /* 1164 * Notify a tcp user of an asynchronous error; 1165 * store error as soft error, but wake up user 1166 * (for now, won't do anything until can select for soft error). 1167 * 1168 * Do not wake up user since there currently is no mechanism for 1169 * reporting soft errors (yet - a kqueue filter may be added). 1170 */ 1171 static void 1172 tcp_notify(struct inpcb *inp, int error) 1173 { 1174 struct tcpcb *tp = intotcpcb(inp); 1175 1176 /* 1177 * Ignore some errors if we are hooked up. 1178 * If connection hasn't completed, has retransmitted several times, 1179 * and receives a second error, give up now. This is better 1180 * than waiting a long time to establish a connection that 1181 * can never complete. 1182 */ 1183 if (tp->t_state == TCPS_ESTABLISHED && 1184 (error == EHOSTUNREACH || error == ENETUNREACH || 1185 error == EHOSTDOWN)) { 1186 return; 1187 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1188 tp->t_softerror) 1189 tcp_drop(tp, error); 1190 else 1191 tp->t_softerror = error; 1192 #if 0 1193 wakeup(&so->so_timeo); 1194 sorwakeup(so); 1195 sowwakeup(so); 1196 #endif 1197 } 1198 1199 static int 1200 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1201 { 1202 int error, i, n; 1203 struct inpcb *marker; 1204 struct inpcb *inp; 1205 int origcpu, ccpu; 1206 1207 error = 0; 1208 n = 0; 1209 1210 /* 1211 * The process of preparing the TCB list is too time-consuming and 1212 * resource-intensive to repeat twice on every request. 1213 */ 1214 if (req->oldptr == NULL) { 1215 for (ccpu = 0; ccpu < ncpus2; ++ccpu) 1216 n += tcbinfo[ccpu].ipi_count; 1217 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1218 return (0); 1219 } 1220 1221 if (req->newptr != NULL) 1222 return (EPERM); 1223 1224 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1225 marker->inp_flags |= INP_PLACEMARKER; 1226 1227 /* 1228 * OK, now we're committed to doing something. Run the inpcb list 1229 * for each cpu in the system and construct the output. Use a 1230 * list placemarker to deal with list changes occuring during 1231 * copyout blockages (but otherwise depend on being on the correct 1232 * cpu to avoid races). 1233 */ 1234 origcpu = mycpu->gd_cpuid; 1235 for (ccpu = 0; ccpu < ncpus2 && error == 0; ++ccpu) { 1236 caddr_t inp_ppcb; 1237 struct xtcpcb xt; 1238 1239 lwkt_migratecpu(ccpu); 1240 1241 n = tcbinfo[ccpu].ipi_count; 1242 1243 LIST_INSERT_HEAD(&tcbinfo[ccpu].pcblisthead, marker, inp_list); 1244 i = 0; 1245 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1246 /* 1247 * process a snapshot of pcbs, ignoring placemarkers 1248 * and using our own to allow SYSCTL_OUT to block. 1249 */ 1250 LIST_REMOVE(marker, inp_list); 1251 LIST_INSERT_AFTER(inp, marker, inp_list); 1252 1253 if (inp->inp_flags & INP_PLACEMARKER) 1254 continue; 1255 if (prison_xinpcb(req->td, inp)) 1256 continue; 1257 1258 xt.xt_len = sizeof xt; 1259 bcopy(inp, &xt.xt_inp, sizeof *inp); 1260 inp_ppcb = inp->inp_ppcb; 1261 if (inp_ppcb != NULL) 1262 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1263 else 1264 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1265 if (inp->inp_socket) 1266 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1267 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1268 break; 1269 ++i; 1270 } 1271 LIST_REMOVE(marker, inp_list); 1272 if (error == 0 && i < n) { 1273 bzero(&xt, sizeof xt); 1274 xt.xt_len = sizeof xt; 1275 while (i < n) { 1276 error = SYSCTL_OUT(req, &xt, sizeof xt); 1277 if (error) 1278 break; 1279 ++i; 1280 } 1281 } 1282 } 1283 1284 /* 1285 * Make sure we are on the same cpu we were on originally, since 1286 * higher level callers expect this. Also don't pollute caches with 1287 * migrated userland data by (eventually) returning to userland 1288 * on a different cpu. 1289 */ 1290 lwkt_migratecpu(origcpu); 1291 kfree(marker, M_TEMP); 1292 return (error); 1293 } 1294 1295 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1296 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1297 1298 static int 1299 tcp_getcred(SYSCTL_HANDLER_ARGS) 1300 { 1301 struct sockaddr_in addrs[2]; 1302 struct ucred cred0, *cred = NULL; 1303 struct inpcb *inp; 1304 int cpu, origcpu, error; 1305 1306 error = priv_check(req->td, PRIV_ROOT); 1307 if (error != 0) 1308 return (error); 1309 error = SYSCTL_IN(req, addrs, sizeof addrs); 1310 if (error != 0) 1311 return (error); 1312 1313 origcpu = mycpuid; 1314 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1315 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1316 1317 lwkt_migratecpu(cpu); 1318 1319 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1320 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1321 if (inp == NULL || inp->inp_socket == NULL) { 1322 error = ENOENT; 1323 } else if (inp->inp_socket->so_cred != NULL) { 1324 cred0 = *(inp->inp_socket->so_cred); 1325 cred = &cred0; 1326 } 1327 1328 lwkt_migratecpu(origcpu); 1329 1330 if (error) 1331 return (error); 1332 1333 return SYSCTL_OUT(req, cred, sizeof(struct ucred)); 1334 } 1335 1336 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1337 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1338 1339 #ifdef INET6 1340 static int 1341 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1342 { 1343 struct sockaddr_in6 addrs[2]; 1344 struct inpcb *inp; 1345 int error; 1346 1347 error = priv_check(req->td, PRIV_ROOT); 1348 if (error != 0) 1349 return (error); 1350 error = SYSCTL_IN(req, addrs, sizeof addrs); 1351 if (error != 0) 1352 return (error); 1353 crit_enter(); 1354 inp = in6_pcblookup_hash(&tcbinfo[0], 1355 &addrs[1].sin6_addr, addrs[1].sin6_port, 1356 &addrs[0].sin6_addr, addrs[0].sin6_port, 0, NULL); 1357 if (inp == NULL || inp->inp_socket == NULL) { 1358 error = ENOENT; 1359 goto out; 1360 } 1361 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1362 out: 1363 crit_exit(); 1364 return (error); 1365 } 1366 1367 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1368 0, 0, 1369 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1370 #endif 1371 1372 struct netmsg_tcp_notify { 1373 struct netmsg_base base; 1374 inp_notify_t nm_notify; 1375 struct in_addr nm_faddr; 1376 int nm_arg; 1377 }; 1378 1379 static void 1380 tcp_notifyall_oncpu(netmsg_t msg) 1381 { 1382 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg; 1383 int nextcpu; 1384 1385 in_pcbnotifyall(&tcbinfo[mycpuid], nm->nm_faddr, 1386 nm->nm_arg, nm->nm_notify); 1387 1388 nextcpu = mycpuid + 1; 1389 if (nextcpu < ncpus2) 1390 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nm->base.lmsg); 1391 else 1392 lwkt_replymsg(&nm->base.lmsg, 0); 1393 } 1394 1395 inp_notify_t 1396 tcp_get_inpnotify(int cmd, const struct sockaddr *sa, 1397 int *arg, struct ip **ip0, int *cpuid) 1398 { 1399 struct ip *ip = *ip0; 1400 struct in_addr faddr; 1401 inp_notify_t notify = tcp_notify; 1402 1403 faddr = ((const struct sockaddr_in *)sa)->sin_addr; 1404 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1405 return NULL; 1406 1407 *arg = inetctlerrmap[cmd]; 1408 if (cmd == PRC_QUENCH) { 1409 notify = tcp_quench; 1410 } else if (icmp_may_rst && 1411 (cmd == PRC_UNREACH_ADMIN_PROHIB || 1412 cmd == PRC_UNREACH_PORT || 1413 cmd == PRC_TIMXCEED_INTRANS) && 1414 ip != NULL) { 1415 notify = tcp_drop_syn_sent; 1416 } else if (cmd == PRC_MSGSIZE) { 1417 const struct icmp *icmp = (const struct icmp *) 1418 ((caddr_t)ip - offsetof(struct icmp, icmp_ip)); 1419 1420 *arg = ntohs(icmp->icmp_nextmtu); 1421 notify = tcp_mtudisc; 1422 } else if (PRC_IS_REDIRECT(cmd)) { 1423 ip = NULL; 1424 notify = in_rtchange; 1425 } else if (cmd == PRC_HOSTDEAD) { 1426 ip = NULL; 1427 } else if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) { 1428 return NULL; 1429 } 1430 1431 if (cpuid != NULL) { 1432 if (ip == NULL) { 1433 /* Go through all CPUs */ 1434 *cpuid = ncpus; 1435 } else { 1436 const struct tcphdr *th; 1437 1438 th = (const struct tcphdr *) 1439 ((caddr_t)ip + (IP_VHL_HL(ip->ip_vhl) << 2)); 1440 *cpuid = tcp_addrcpu(faddr.s_addr, th->th_dport, 1441 ip->ip_src.s_addr, th->th_sport); 1442 } 1443 } 1444 1445 *ip0 = ip; 1446 return notify; 1447 } 1448 1449 void 1450 tcp_ctlinput(netmsg_t msg) 1451 { 1452 int cmd = msg->ctlinput.nm_cmd; 1453 struct sockaddr *sa = msg->ctlinput.nm_arg; 1454 struct ip *ip = msg->ctlinput.nm_extra; 1455 struct in_addr faddr; 1456 inp_notify_t notify; 1457 int arg, cpuid; 1458 1459 notify = tcp_get_inpnotify(cmd, sa, &arg, &ip, &cpuid); 1460 if (notify == NULL) 1461 goto done; 1462 1463 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1464 if (ip != NULL) { 1465 const struct tcphdr *th; 1466 struct inpcb *inp; 1467 1468 if (cpuid != mycpuid) 1469 goto done; 1470 1471 th = (const struct tcphdr *) 1472 ((caddr_t)ip + (IP_VHL_HL(ip->ip_vhl) << 2)); 1473 inp = in_pcblookup_hash(&tcbinfo[mycpuid], faddr, th->th_dport, 1474 ip->ip_src, th->th_sport, 0, NULL); 1475 if (inp != NULL && inp->inp_socket != NULL) { 1476 tcp_seq icmpseq = htonl(th->th_seq); 1477 struct tcpcb *tp = intotcpcb(inp); 1478 1479 if (SEQ_GEQ(icmpseq, tp->snd_una) && 1480 SEQ_LT(icmpseq, tp->snd_max)) 1481 notify(inp, arg); 1482 } else { 1483 struct in_conninfo inc; 1484 1485 inc.inc_fport = th->th_dport; 1486 inc.inc_lport = th->th_sport; 1487 inc.inc_faddr = faddr; 1488 inc.inc_laddr = ip->ip_src; 1489 #ifdef INET6 1490 inc.inc_isipv6 = 0; 1491 #endif 1492 syncache_unreach(&inc, th); 1493 } 1494 } else if (msg->ctlinput.nm_direct) { 1495 if (cpuid != ncpus && cpuid != mycpuid) 1496 goto done; 1497 if (mycpuid >= ncpus2) 1498 goto done; 1499 1500 in_pcbnotifyall(&tcbinfo[mycpuid], faddr, arg, notify); 1501 } else { 1502 struct netmsg_tcp_notify *nm; 1503 1504 ASSERT_IN_NETISR(0); 1505 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT); 1506 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1507 0, tcp_notifyall_oncpu); 1508 nm->nm_faddr = faddr; 1509 nm->nm_arg = arg; 1510 nm->nm_notify = notify; 1511 1512 lwkt_sendmsg(netisr_cpuport(0), &nm->base.lmsg); 1513 } 1514 done: 1515 lwkt_replymsg(&msg->lmsg, 0); 1516 } 1517 1518 #ifdef INET6 1519 1520 void 1521 tcp6_ctlinput(netmsg_t msg) 1522 { 1523 int cmd = msg->ctlinput.nm_cmd; 1524 struct sockaddr *sa = msg->ctlinput.nm_arg; 1525 void *d = msg->ctlinput.nm_extra; 1526 struct tcphdr th; 1527 inp_notify_t notify = tcp_notify; 1528 struct ip6_hdr *ip6; 1529 struct mbuf *m; 1530 struct ip6ctlparam *ip6cp = NULL; 1531 const struct sockaddr_in6 *sa6_src = NULL; 1532 int off; 1533 struct tcp_portonly { 1534 u_int16_t th_sport; 1535 u_int16_t th_dport; 1536 } *thp; 1537 int arg; 1538 1539 if (sa->sa_family != AF_INET6 || 1540 sa->sa_len != sizeof(struct sockaddr_in6)) { 1541 goto out; 1542 } 1543 1544 arg = 0; 1545 if (cmd == PRC_QUENCH) 1546 notify = tcp_quench; 1547 else if (cmd == PRC_MSGSIZE) { 1548 struct ip6ctlparam *ip6cp = d; 1549 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6; 1550 1551 arg = ntohl(icmp6->icmp6_mtu); 1552 notify = tcp_mtudisc; 1553 } else if (!PRC_IS_REDIRECT(cmd) && 1554 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) { 1555 goto out; 1556 } 1557 1558 /* if the parameter is from icmp6, decode it. */ 1559 if (d != NULL) { 1560 ip6cp = (struct ip6ctlparam *)d; 1561 m = ip6cp->ip6c_m; 1562 ip6 = ip6cp->ip6c_ip6; 1563 off = ip6cp->ip6c_off; 1564 sa6_src = ip6cp->ip6c_src; 1565 } else { 1566 m = NULL; 1567 ip6 = NULL; 1568 off = 0; /* fool gcc */ 1569 sa6_src = &sa6_any; 1570 } 1571 1572 if (ip6 != NULL) { 1573 struct in_conninfo inc; 1574 /* 1575 * XXX: We assume that when IPV6 is non NULL, 1576 * M and OFF are valid. 1577 */ 1578 1579 /* check if we can safely examine src and dst ports */ 1580 if (m->m_pkthdr.len < off + sizeof *thp) 1581 goto out; 1582 1583 bzero(&th, sizeof th); 1584 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1585 1586 in6_pcbnotify(&tcbinfo[0], sa, th.th_dport, 1587 (struct sockaddr *)ip6cp->ip6c_src, 1588 th.th_sport, cmd, arg, notify); 1589 1590 inc.inc_fport = th.th_dport; 1591 inc.inc_lport = th.th_sport; 1592 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1593 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1594 inc.inc_isipv6 = 1; 1595 syncache_unreach(&inc, &th); 1596 } else { 1597 in6_pcbnotify(&tcbinfo[0], sa, 0, 1598 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify); 1599 } 1600 out: 1601 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0); 1602 } 1603 1604 #endif 1605 1606 /* 1607 * Following is where TCP initial sequence number generation occurs. 1608 * 1609 * There are two places where we must use initial sequence numbers: 1610 * 1. In SYN-ACK packets. 1611 * 2. In SYN packets. 1612 * 1613 * All ISNs for SYN-ACK packets are generated by the syncache. See 1614 * tcp_syncache.c for details. 1615 * 1616 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1617 * depends on this property. In addition, these ISNs should be 1618 * unguessable so as to prevent connection hijacking. To satisfy 1619 * the requirements of this situation, the algorithm outlined in 1620 * RFC 1948 is used to generate sequence numbers. 1621 * 1622 * Implementation details: 1623 * 1624 * Time is based off the system timer, and is corrected so that it 1625 * increases by one megabyte per second. This allows for proper 1626 * recycling on high speed LANs while still leaving over an hour 1627 * before rollover. 1628 * 1629 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1630 * between seeding of isn_secret. This is normally set to zero, 1631 * as reseeding should not be necessary. 1632 * 1633 */ 1634 1635 #define ISN_BYTES_PER_SECOND 1048576 1636 1637 u_char isn_secret[32]; 1638 int isn_last_reseed; 1639 MD5_CTX isn_ctx; 1640 1641 tcp_seq 1642 tcp_new_isn(struct tcpcb *tp) 1643 { 1644 u_int32_t md5_buffer[4]; 1645 tcp_seq new_isn; 1646 1647 /* Seed if this is the first use, reseed if requested. */ 1648 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1649 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1650 < (u_int)ticks))) { 1651 read_random_unlimited(&isn_secret, sizeof isn_secret); 1652 isn_last_reseed = ticks; 1653 } 1654 1655 /* Compute the md5 hash and return the ISN. */ 1656 MD5Init(&isn_ctx); 1657 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1658 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1659 #ifdef INET6 1660 if (INP_ISIPV6(tp->t_inpcb)) { 1661 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1662 sizeof(struct in6_addr)); 1663 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1664 sizeof(struct in6_addr)); 1665 } else 1666 #endif 1667 { 1668 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1669 sizeof(struct in_addr)); 1670 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1671 sizeof(struct in_addr)); 1672 } 1673 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1674 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1675 new_isn = (tcp_seq) md5_buffer[0]; 1676 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1677 return (new_isn); 1678 } 1679 1680 /* 1681 * When a source quench is received, close congestion window 1682 * to one segment. We will gradually open it again as we proceed. 1683 */ 1684 void 1685 tcp_quench(struct inpcb *inp, int error) 1686 { 1687 struct tcpcb *tp = intotcpcb(inp); 1688 1689 KASSERT(tp != NULL, ("tcp_quench: tp is NULL")); 1690 tp->snd_cwnd = tp->t_maxseg; 1691 tp->snd_wacked = 0; 1692 } 1693 1694 /* 1695 * When a specific ICMP unreachable message is received and the 1696 * connection state is SYN-SENT, drop the connection. This behavior 1697 * is controlled by the icmp_may_rst sysctl. 1698 */ 1699 void 1700 tcp_drop_syn_sent(struct inpcb *inp, int error) 1701 { 1702 struct tcpcb *tp = intotcpcb(inp); 1703 1704 KASSERT(tp != NULL, ("tcp_drop_syn_sent: tp is NULL")); 1705 if (tp->t_state == TCPS_SYN_SENT) 1706 tcp_drop(tp, error); 1707 } 1708 1709 /* 1710 * When a `need fragmentation' ICMP is received, update our idea of the MSS 1711 * based on the new value in the route. Also nudge TCP to send something, 1712 * since we know the packet we just sent was dropped. 1713 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1714 */ 1715 void 1716 tcp_mtudisc(struct inpcb *inp, int mtu) 1717 { 1718 struct tcpcb *tp = intotcpcb(inp); 1719 struct rtentry *rt; 1720 struct socket *so = inp->inp_socket; 1721 int maxopd, mss; 1722 #ifdef INET6 1723 boolean_t isipv6 = INP_ISIPV6(inp); 1724 #else 1725 const boolean_t isipv6 = FALSE; 1726 #endif 1727 1728 KASSERT(tp != NULL, ("tcp_mtudisc: tp is NULL")); 1729 1730 /* 1731 * If no MTU is provided in the ICMP message, use the 1732 * next lower likely value, as specified in RFC 1191. 1733 */ 1734 if (mtu == 0) { 1735 int oldmtu; 1736 1737 oldmtu = tp->t_maxopd + 1738 (isipv6 ? 1739 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1740 sizeof(struct tcpiphdr)); 1741 mtu = ip_next_mtu(oldmtu, 0); 1742 } 1743 1744 if (isipv6) 1745 rt = tcp_rtlookup6(&inp->inp_inc); 1746 else 1747 rt = tcp_rtlookup(&inp->inp_inc); 1748 if (rt != NULL) { 1749 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu) 1750 mtu = rt->rt_rmx.rmx_mtu; 1751 1752 maxopd = mtu - 1753 (isipv6 ? 1754 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1755 sizeof(struct tcpiphdr)); 1756 1757 /* 1758 * XXX - The following conditional probably violates the TCP 1759 * spec. The problem is that, since we don't know the 1760 * other end's MSS, we are supposed to use a conservative 1761 * default. But, if we do that, then MTU discovery will 1762 * never actually take place, because the conservative 1763 * default is much less than the MTUs typically seen 1764 * on the Internet today. For the moment, we'll sweep 1765 * this under the carpet. 1766 * 1767 * The conservative default might not actually be a problem 1768 * if the only case this occurs is when sending an initial 1769 * SYN with options and data to a host we've never talked 1770 * to before. Then, they will reply with an MSS value which 1771 * will get recorded and the new parameters should get 1772 * recomputed. For Further Study. 1773 */ 1774 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd) 1775 maxopd = rt->rt_rmx.rmx_mssopt; 1776 } else 1777 maxopd = mtu - 1778 (isipv6 ? 1779 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1780 sizeof(struct tcpiphdr)); 1781 1782 if (tp->t_maxopd <= maxopd) 1783 return; 1784 tp->t_maxopd = maxopd; 1785 1786 mss = maxopd; 1787 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) == 1788 (TF_REQ_TSTMP | TF_RCVD_TSTMP)) 1789 mss -= TCPOLEN_TSTAMP_APPA; 1790 1791 /* round down to multiple of MCLBYTES */ 1792 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */ 1793 if (mss > MCLBYTES) 1794 mss &= ~(MCLBYTES - 1); 1795 #else 1796 if (mss > MCLBYTES) 1797 mss = (mss / MCLBYTES) * MCLBYTES; 1798 #endif 1799 1800 if (so->so_snd.ssb_hiwat < mss) 1801 mss = so->so_snd.ssb_hiwat; 1802 1803 tp->t_maxseg = mss; 1804 tp->t_rtttime = 0; 1805 tp->snd_nxt = tp->snd_una; 1806 tcp_output(tp); 1807 tcpstat.tcps_mturesent++; 1808 } 1809 1810 /* 1811 * Look-up the routing entry to the peer of this inpcb. If no route 1812 * is found and it cannot be allocated the return NULL. This routine 1813 * is called by TCP routines that access the rmx structure and by tcp_mss 1814 * to get the interface MTU. 1815 */ 1816 struct rtentry * 1817 tcp_rtlookup(struct in_conninfo *inc) 1818 { 1819 struct route *ro = &inc->inc_route; 1820 1821 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1822 /* No route yet, so try to acquire one */ 1823 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1824 /* 1825 * unused portions of the structure MUST be zero'd 1826 * out because rtalloc() treats it as opaque data 1827 */ 1828 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1829 ro->ro_dst.sa_family = AF_INET; 1830 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1831 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1832 inc->inc_faddr; 1833 rtalloc(ro); 1834 } 1835 } 1836 return (ro->ro_rt); 1837 } 1838 1839 #ifdef INET6 1840 struct rtentry * 1841 tcp_rtlookup6(struct in_conninfo *inc) 1842 { 1843 struct route_in6 *ro6 = &inc->inc6_route; 1844 1845 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1846 /* No route yet, so try to acquire one */ 1847 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1848 /* 1849 * unused portions of the structure MUST be zero'd 1850 * out because rtalloc() treats it as opaque data 1851 */ 1852 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1853 ro6->ro_dst.sin6_family = AF_INET6; 1854 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1855 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1856 rtalloc((struct route *)ro6); 1857 } 1858 } 1859 return (ro6->ro_rt); 1860 } 1861 #endif 1862 1863 #ifdef IPSEC 1864 /* compute ESP/AH header size for TCP, including outer IP header. */ 1865 size_t 1866 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1867 { 1868 struct inpcb *inp; 1869 struct mbuf *m; 1870 size_t hdrsiz; 1871 struct ip *ip; 1872 struct tcphdr *th; 1873 1874 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1875 return (0); 1876 MGETHDR(m, M_NOWAIT, MT_DATA); 1877 if (!m) 1878 return (0); 1879 1880 #ifdef INET6 1881 if (INP_ISIPV6(inp)) { 1882 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1883 1884 th = (struct tcphdr *)(ip6 + 1); 1885 m->m_pkthdr.len = m->m_len = 1886 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1887 tcp_fillheaders(tp, ip6, th, FALSE); 1888 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1889 } else 1890 #endif 1891 { 1892 ip = mtod(m, struct ip *); 1893 th = (struct tcphdr *)(ip + 1); 1894 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1895 tcp_fillheaders(tp, ip, th, FALSE); 1896 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1897 } 1898 1899 m_free(m); 1900 return (hdrsiz); 1901 } 1902 #endif 1903 1904 /* 1905 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1906 * 1907 * This code attempts to calculate the bandwidth-delay product as a 1908 * means of determining the optimal window size to maximize bandwidth, 1909 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1910 * routers. This code also does a fairly good job keeping RTTs in check 1911 * across slow links like modems. We implement an algorithm which is very 1912 * similar (but not meant to be) TCP/Vegas. The code operates on the 1913 * transmitter side of a TCP connection and so only effects the transmit 1914 * side of the connection. 1915 * 1916 * BACKGROUND: TCP makes no provision for the management of buffer space 1917 * at the end points or at the intermediate routers and switches. A TCP 1918 * stream, whether using NewReno or not, will eventually buffer as 1919 * many packets as it is able and the only reason this typically works is 1920 * due to the fairly small default buffers made available for a connection 1921 * (typicaly 16K or 32K). As machines use larger windows and/or window 1922 * scaling it is now fairly easy for even a single TCP connection to blow-out 1923 * all available buffer space not only on the local interface, but on 1924 * intermediate routers and switches as well. NewReno makes a misguided 1925 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1926 * then backing off, then steadily increasing the window again until another 1927 * failure occurs, ad-infinitum. This results in terrible oscillation that 1928 * is only made worse as network loads increase and the idea of intentionally 1929 * blowing out network buffers is, frankly, a terrible way to manage network 1930 * resources. 1931 * 1932 * It is far better to limit the transmit window prior to the failure 1933 * condition being achieved. There are two general ways to do this: First 1934 * you can 'scan' through different transmit window sizes and locate the 1935 * point where the RTT stops increasing, indicating that you have filled the 1936 * pipe, then scan backwards until you note that RTT stops decreasing, then 1937 * repeat ad-infinitum. This method works in principle but has severe 1938 * implementation issues due to RTT variances, timer granularity, and 1939 * instability in the algorithm which can lead to many false positives and 1940 * create oscillations as well as interact badly with other TCP streams 1941 * implementing the same algorithm. 1942 * 1943 * The second method is to limit the window to the bandwidth delay product 1944 * of the link. This is the method we implement. RTT variances and our 1945 * own manipulation of the congestion window, bwnd, can potentially 1946 * destabilize the algorithm. For this reason we have to stabilize the 1947 * elements used to calculate the window. We do this by using the minimum 1948 * observed RTT, the long term average of the observed bandwidth, and 1949 * by adding two segments worth of slop. It isn't perfect but it is able 1950 * to react to changing conditions and gives us a very stable basis on 1951 * which to extend the algorithm. 1952 */ 1953 void 1954 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1955 { 1956 u_long bw; 1957 u_long ibw; 1958 u_long bwnd; 1959 int save_ticks; 1960 int delta_ticks; 1961 1962 /* 1963 * If inflight_enable is disabled in the middle of a tcp connection, 1964 * make sure snd_bwnd is effectively disabled. 1965 */ 1966 if (!tcp_inflight_enable) { 1967 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1968 tp->snd_bandwidth = 0; 1969 return; 1970 } 1971 1972 /* 1973 * Validate the delta time. If a connection is new or has been idle 1974 * a long time we have to reset the bandwidth calculator. 1975 */ 1976 save_ticks = ticks; 1977 cpu_ccfence(); 1978 delta_ticks = save_ticks - tp->t_bw_rtttime; 1979 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1980 tp->t_bw_rtttime = save_ticks; 1981 tp->t_bw_rtseq = ack_seq; 1982 if (tp->snd_bandwidth == 0) 1983 tp->snd_bandwidth = tcp_inflight_start; 1984 return; 1985 } 1986 1987 /* 1988 * A delta of at least 1 tick is required. Waiting 2 ticks will 1989 * result in better (bw) accuracy. More than that and the ramp-up 1990 * will be too slow. 1991 */ 1992 if (delta_ticks == 0 || delta_ticks == 1) 1993 return; 1994 1995 /* 1996 * Sanity check, plus ignore pure window update acks. 1997 */ 1998 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1999 return; 2000 2001 /* 2002 * Figure out the bandwidth. Due to the tick granularity this 2003 * is a very rough number and it MUST be averaged over a fairly 2004 * long period of time. XXX we need to take into account a link 2005 * that is not using all available bandwidth, but for now our 2006 * slop will ramp us up if this case occurs and the bandwidth later 2007 * increases. 2008 */ 2009 ibw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 2010 tp->t_bw_rtttime = save_ticks; 2011 tp->t_bw_rtseq = ack_seq; 2012 bw = ((int64_t)tp->snd_bandwidth * 15 + ibw) >> 4; 2013 2014 tp->snd_bandwidth = bw; 2015 2016 /* 2017 * Calculate the semi-static bandwidth delay product, plus two maximal 2018 * segments. The additional slop puts us squarely in the sweet 2019 * spot and also handles the bandwidth run-up case. Without the 2020 * slop we could be locking ourselves into a lower bandwidth. 2021 * 2022 * At very high speeds the bw calculation can become overly sensitive 2023 * and error prone when delta_ticks is low (e.g. usually 1). To deal 2024 * with the problem the stab must be scaled to the bw. A stab of 50 2025 * (the default) increases the bw for the purposes of the bwnd 2026 * calculation by 5%. 2027 * 2028 * Situations Handled: 2029 * (1) Prevents over-queueing of packets on LANs, especially on 2030 * high speed LANs, allowing larger TCP buffers to be 2031 * specified, and also does a good job preventing 2032 * over-queueing of packets over choke points like modems 2033 * (at least for the transmit side). 2034 * 2035 * (2) Is able to handle changing network loads (bandwidth 2036 * drops so bwnd drops, bandwidth increases so bwnd 2037 * increases). 2038 * 2039 * (3) Theoretically should stabilize in the face of multiple 2040 * connections implementing the same algorithm (this may need 2041 * a little work). 2042 * 2043 * (4) Stability value (defaults to 20 = 2 maximal packets) can 2044 * be adjusted with a sysctl but typically only needs to be on 2045 * very slow connections. A value no smaller then 5 should 2046 * be used, but only reduce this default if you have no other 2047 * choice. 2048 */ 2049 2050 #define USERTT ((tp->t_srtt + tp->t_rttvar) + tcp_inflight_adjrtt) 2051 bw += bw * tcp_inflight_stab / 1000; 2052 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 2053 (int)tp->t_maxseg * 2; 2054 #undef USERTT 2055 2056 if (tcp_inflight_debug > 0) { 2057 static int ltime; 2058 if ((u_int)(save_ticks - ltime) >= hz / tcp_inflight_debug) { 2059 ltime = save_ticks; 2060 kprintf("%p ibw %ld bw %ld rttvar %d srtt %d " 2061 "bwnd %ld delta %d snd_win %ld\n", 2062 tp, ibw, bw, tp->t_rttvar, tp->t_srtt, 2063 bwnd, delta_ticks, tp->snd_wnd); 2064 } 2065 } 2066 if ((long)bwnd < tcp_inflight_min) 2067 bwnd = tcp_inflight_min; 2068 if (bwnd > tcp_inflight_max) 2069 bwnd = tcp_inflight_max; 2070 if ((long)bwnd < tp->t_maxseg * 2) 2071 bwnd = tp->t_maxseg * 2; 2072 tp->snd_bwnd = bwnd; 2073 } 2074 2075 static void 2076 tcp_rmx_iwsegs(struct tcpcb *tp, u_long *maxsegs, u_long *capsegs) 2077 { 2078 struct rtentry *rt; 2079 struct inpcb *inp = tp->t_inpcb; 2080 #ifdef INET6 2081 boolean_t isipv6 = INP_ISIPV6(inp); 2082 #else 2083 const boolean_t isipv6 = FALSE; 2084 #endif 2085 2086 /* XXX */ 2087 if (tcp_iw_maxsegs < TCP_IW_MAXSEGS_DFLT) 2088 tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 2089 if (tcp_iw_capsegs < TCP_IW_CAPSEGS_DFLT) 2090 tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 2091 2092 if (isipv6) 2093 rt = tcp_rtlookup6(&inp->inp_inc); 2094 else 2095 rt = tcp_rtlookup(&inp->inp_inc); 2096 if (rt == NULL || 2097 rt->rt_rmx.rmx_iwmaxsegs < TCP_IW_MAXSEGS_DFLT || 2098 rt->rt_rmx.rmx_iwcapsegs < TCP_IW_CAPSEGS_DFLT) { 2099 *maxsegs = tcp_iw_maxsegs; 2100 *capsegs = tcp_iw_capsegs; 2101 return; 2102 } 2103 *maxsegs = rt->rt_rmx.rmx_iwmaxsegs; 2104 *capsegs = rt->rt_rmx.rmx_iwcapsegs; 2105 } 2106 2107 u_long 2108 tcp_initial_window(struct tcpcb *tp) 2109 { 2110 if (tcp_do_rfc3390) { 2111 /* 2112 * RFC3390: 2113 * "If the SYN or SYN/ACK is lost, the initial window 2114 * used by a sender after a correctly transmitted SYN 2115 * MUST be one segment consisting of MSS bytes." 2116 * 2117 * However, we do something a little bit more aggressive 2118 * then RFC3390 here: 2119 * - Only if time spent in the SYN or SYN|ACK retransmition 2120 * >= 3 seconds, the IW is reduced. We do this mainly 2121 * because when RFC3390 is published, the initial RTO is 2122 * still 3 seconds (the threshold we test here), while 2123 * after RFC6298, the initial RTO is 1 second. This 2124 * behaviour probably still falls within the spirit of 2125 * RFC3390. 2126 * - When IW is reduced, 2*MSS is used instead of 1*MSS. 2127 * Mainly to avoid sender and receiver deadlock until 2128 * delayed ACK timer expires. And even RFC2581 does not 2129 * try to reduce IW upon SYN or SYN|ACK retransmition 2130 * timeout. 2131 * 2132 * See also: 2133 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03 2134 */ 2135 if (tp->t_rxtsyn >= TCPTV_RTOBASE3) { 2136 return (2 * tp->t_maxseg); 2137 } else { 2138 u_long maxsegs, capsegs; 2139 2140 tcp_rmx_iwsegs(tp, &maxsegs, &capsegs); 2141 return min(maxsegs * tp->t_maxseg, 2142 max(2 * tp->t_maxseg, capsegs * 1460)); 2143 } 2144 } else { 2145 /* 2146 * Even RFC2581 (back to 1999) allows 2*SMSS IW. 2147 * 2148 * Mainly to avoid sender and receiver deadlock 2149 * until delayed ACK timer expires. 2150 */ 2151 return (2 * tp->t_maxseg); 2152 } 2153 } 2154 2155 #ifdef TCP_SIGNATURE 2156 /* 2157 * Compute TCP-MD5 hash of a TCP segment. (RFC2385) 2158 * 2159 * We do this over ip, tcphdr, segment data, and the key in the SADB. 2160 * When called from tcp_input(), we can be sure that th_sum has been 2161 * zeroed out and verified already. 2162 * 2163 * Return 0 if successful, otherwise return -1. 2164 * 2165 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a 2166 * search with the destination IP address, and a 'magic SPI' to be 2167 * determined by the application. This is hardcoded elsewhere to 1179 2168 * right now. Another branch of this code exists which uses the SPD to 2169 * specify per-application flows but it is unstable. 2170 */ 2171 int 2172 tcpsignature_compute( 2173 struct mbuf *m, /* mbuf chain */ 2174 int len, /* length of TCP data */ 2175 int optlen, /* length of TCP options */ 2176 u_char *buf, /* storage for MD5 digest */ 2177 u_int direction) /* direction of flow */ 2178 { 2179 struct ippseudo ippseudo; 2180 MD5_CTX ctx; 2181 int doff; 2182 struct ip *ip; 2183 struct ipovly *ipovly; 2184 struct secasvar *sav; 2185 struct tcphdr *th; 2186 #ifdef INET6 2187 struct ip6_hdr *ip6; 2188 struct in6_addr in6; 2189 uint32_t plen; 2190 uint16_t nhdr; 2191 #endif /* INET6 */ 2192 u_short savecsum; 2193 2194 KASSERT(m != NULL, ("passed NULL mbuf. Game over.")); 2195 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature")); 2196 /* 2197 * Extract the destination from the IP header in the mbuf. 2198 */ 2199 ip = mtod(m, struct ip *); 2200 #ifdef INET6 2201 ip6 = NULL; /* Make the compiler happy. */ 2202 #endif /* INET6 */ 2203 /* 2204 * Look up an SADB entry which matches the address found in 2205 * the segment. 2206 */ 2207 switch (IP_VHL_V(ip->ip_vhl)) { 2208 case IPVERSION: 2209 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst, 2210 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2211 break; 2212 #ifdef INET6 2213 case (IPV6_VERSION >> 4): 2214 ip6 = mtod(m, struct ip6_hdr *); 2215 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst, 2216 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2217 break; 2218 #endif /* INET6 */ 2219 default: 2220 return (EINVAL); 2221 /* NOTREACHED */ 2222 break; 2223 } 2224 if (sav == NULL) { 2225 kprintf("%s: SADB lookup failed\n", __func__); 2226 return (EINVAL); 2227 } 2228 MD5Init(&ctx); 2229 2230 /* 2231 * Step 1: Update MD5 hash with IP pseudo-header. 2232 * 2233 * XXX The ippseudo header MUST be digested in network byte order, 2234 * or else we'll fail the regression test. Assume all fields we've 2235 * been doing arithmetic on have been in host byte order. 2236 * XXX One cannot depend on ipovly->ih_len here. When called from 2237 * tcp_output(), the underlying ip_len member has not yet been set. 2238 */ 2239 switch (IP_VHL_V(ip->ip_vhl)) { 2240 case IPVERSION: 2241 ipovly = (struct ipovly *)ip; 2242 ippseudo.ippseudo_src = ipovly->ih_src; 2243 ippseudo.ippseudo_dst = ipovly->ih_dst; 2244 ippseudo.ippseudo_pad = 0; 2245 ippseudo.ippseudo_p = IPPROTO_TCP; 2246 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen); 2247 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo)); 2248 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip)); 2249 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen; 2250 break; 2251 #ifdef INET6 2252 /* 2253 * RFC 2385, 2.0 Proposal 2254 * For IPv6, the pseudo-header is as described in RFC 2460, namely the 2255 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero- 2256 * extended next header value (to form 32 bits), and 32-bit segment 2257 * length. 2258 * Note: Upper-Layer Packet Length comes before Next Header. 2259 */ 2260 case (IPV6_VERSION >> 4): 2261 in6 = ip6->ip6_src; 2262 in6_clearscope(&in6); 2263 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2264 in6 = ip6->ip6_dst; 2265 in6_clearscope(&in6); 2266 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2267 plen = htonl(len + sizeof(struct tcphdr) + optlen); 2268 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t)); 2269 nhdr = 0; 2270 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2271 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2272 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2273 nhdr = IPPROTO_TCP; 2274 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2275 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr)); 2276 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen; 2277 break; 2278 #endif /* INET6 */ 2279 default: 2280 return (EINVAL); 2281 /* NOTREACHED */ 2282 break; 2283 } 2284 /* 2285 * Step 2: Update MD5 hash with TCP header, excluding options. 2286 * The TCP checksum must be set to zero. 2287 */ 2288 savecsum = th->th_sum; 2289 th->th_sum = 0; 2290 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr)); 2291 th->th_sum = savecsum; 2292 /* 2293 * Step 3: Update MD5 hash with TCP segment data. 2294 * Use m_apply() to avoid an early m_pullup(). 2295 */ 2296 if (len > 0) 2297 m_apply(m, doff, len, tcpsignature_apply, &ctx); 2298 /* 2299 * Step 4: Update MD5 hash with shared secret. 2300 */ 2301 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth)); 2302 MD5Final(buf, &ctx); 2303 key_sa_recordxfer(sav, m); 2304 key_freesav(sav); 2305 return (0); 2306 } 2307 2308 int 2309 tcpsignature_apply(void *fstate, void *data, unsigned int len) 2310 { 2311 2312 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len); 2313 return (0); 2314 } 2315 #endif /* TCP_SIGNATURE */ 2316 2317 static void 2318 tcp_drop_sysctl_dispatch(netmsg_t nmsg) 2319 { 2320 struct lwkt_msg *lmsg = &nmsg->lmsg; 2321 /* addrs[0] is a foreign socket, addrs[1] is a local one. */ 2322 struct sockaddr_storage *addrs = lmsg->u.ms_resultp; 2323 int error; 2324 struct sockaddr_in *fin, *lin; 2325 #ifdef INET6 2326 struct sockaddr_in6 *fin6, *lin6; 2327 struct in6_addr f6, l6; 2328 #endif 2329 struct inpcb *inp; 2330 2331 switch (addrs[0].ss_family) { 2332 #ifdef INET6 2333 case AF_INET6: 2334 fin6 = (struct sockaddr_in6 *)&addrs[0]; 2335 lin6 = (struct sockaddr_in6 *)&addrs[1]; 2336 error = in6_embedscope(&f6, fin6, NULL, NULL); 2337 if (error) 2338 goto done; 2339 error = in6_embedscope(&l6, lin6, NULL, NULL); 2340 if (error) 2341 goto done; 2342 inp = in6_pcblookup_hash(&tcbinfo[mycpuid], &f6, 2343 fin6->sin6_port, &l6, lin6->sin6_port, FALSE, NULL); 2344 break; 2345 #endif 2346 #ifdef INET 2347 case AF_INET: 2348 fin = (struct sockaddr_in *)&addrs[0]; 2349 lin = (struct sockaddr_in *)&addrs[1]; 2350 inp = in_pcblookup_hash(&tcbinfo[mycpuid], fin->sin_addr, 2351 fin->sin_port, lin->sin_addr, lin->sin_port, FALSE, NULL); 2352 break; 2353 #endif 2354 default: 2355 /* 2356 * Must not reach here, since the address family was 2357 * checked in sysctl handler. 2358 */ 2359 panic("unknown address family %d", addrs[0].ss_family); 2360 } 2361 if (inp != NULL) { 2362 struct tcpcb *tp = intotcpcb(inp); 2363 2364 KASSERT((inp->inp_flags & INP_WILDCARD) == 0, 2365 ("in wildcard hash")); 2366 KASSERT(tp != NULL, ("tcp_drop_sysctl_dispatch: tp is NULL")); 2367 KASSERT((tp->t_flags & TF_LISTEN) == 0, ("listen socket")); 2368 tcp_drop(tp, ECONNABORTED); 2369 error = 0; 2370 } else { 2371 error = ESRCH; 2372 } 2373 #ifdef INET6 2374 done: 2375 #endif 2376 lwkt_replymsg(lmsg, error); 2377 } 2378 2379 static int 2380 sysctl_tcp_drop(SYSCTL_HANDLER_ARGS) 2381 { 2382 /* addrs[0] is a foreign socket, addrs[1] is a local one. */ 2383 struct sockaddr_storage addrs[2]; 2384 struct sockaddr_in *fin, *lin; 2385 #ifdef INET6 2386 struct sockaddr_in6 *fin6, *lin6; 2387 #endif 2388 struct netmsg_base nmsg; 2389 struct lwkt_msg *lmsg = &nmsg.lmsg; 2390 struct lwkt_port *port = NULL; 2391 int error; 2392 2393 fin = lin = NULL; 2394 #ifdef INET6 2395 fin6 = lin6 = NULL; 2396 #endif 2397 error = 0; 2398 2399 if (req->oldptr != NULL || req->oldlen != 0) 2400 return (EINVAL); 2401 if (req->newptr == NULL) 2402 return (EPERM); 2403 if (req->newlen < sizeof(addrs)) 2404 return (ENOMEM); 2405 error = SYSCTL_IN(req, &addrs, sizeof(addrs)); 2406 if (error) 2407 return (error); 2408 2409 switch (addrs[0].ss_family) { 2410 #ifdef INET6 2411 case AF_INET6: 2412 fin6 = (struct sockaddr_in6 *)&addrs[0]; 2413 lin6 = (struct sockaddr_in6 *)&addrs[1]; 2414 if (fin6->sin6_len != sizeof(struct sockaddr_in6) || 2415 lin6->sin6_len != sizeof(struct sockaddr_in6)) 2416 return (EINVAL); 2417 if (IN6_IS_ADDR_V4MAPPED(&fin6->sin6_addr) || 2418 IN6_IS_ADDR_V4MAPPED(&lin6->sin6_addr)) 2419 return (EADDRNOTAVAIL); 2420 #if 0 2421 error = sa6_embedscope(fin6, V_ip6_use_defzone); 2422 if (error) 2423 return (error); 2424 error = sa6_embedscope(lin6, V_ip6_use_defzone); 2425 if (error) 2426 return (error); 2427 #endif 2428 port = tcp6_addrport(); 2429 break; 2430 #endif 2431 #ifdef INET 2432 case AF_INET: 2433 fin = (struct sockaddr_in *)&addrs[0]; 2434 lin = (struct sockaddr_in *)&addrs[1]; 2435 if (fin->sin_len != sizeof(struct sockaddr_in) || 2436 lin->sin_len != sizeof(struct sockaddr_in)) 2437 return (EINVAL); 2438 port = tcp_addrport(fin->sin_addr.s_addr, fin->sin_port, 2439 lin->sin_addr.s_addr, lin->sin_port); 2440 break; 2441 #endif 2442 default: 2443 return (EINVAL); 2444 } 2445 2446 netmsg_init(&nmsg, NULL, &curthread->td_msgport, 0, 2447 tcp_drop_sysctl_dispatch); 2448 lmsg->u.ms_resultp = addrs; 2449 return lwkt_domsg(port, lmsg, 0); 2450 } 2451 2452 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, drop, 2453 CTLTYPE_STRUCT | CTLFLAG_WR | CTLFLAG_SKIP, NULL, 2454 0, sysctl_tcp_drop, "", "Drop TCP connection"); 2455 2456 void 2457 tcp_pcbport_create(struct tcpcb *tp) 2458 { 2459 int cpu; 2460 2461 KASSERT((tp->t_flags & TF_LISTEN) && tp->t_state == TCPS_LISTEN, 2462 ("not a listen tcpcb")); 2463 2464 KASSERT(tp->t_pcbport == NULL, ("tcpcb port cache was created")); 2465 tp->t_pcbport = kmalloc_cachealign(sizeof(struct tcp_pcbport) * ncpus2, 2466 M_PCB, M_WAITOK); 2467 2468 for (cpu = 0; cpu < ncpus2; ++cpu) { 2469 struct inpcbport *phd; 2470 2471 phd = &tp->t_pcbport[cpu].t_phd; 2472 LIST_INIT(&phd->phd_pcblist); 2473 /* Though, not used ... */ 2474 phd->phd_port = tp->t_inpcb->inp_lport; 2475 } 2476 } 2477 2478 void 2479 tcp_pcbport_merge_oncpu(struct tcpcb *tp) 2480 { 2481 struct inpcbport *phd; 2482 struct inpcb *inp; 2483 int cpu = mycpuid; 2484 2485 KASSERT(cpu < ncpus2, ("invalid cpu%d", cpu)); 2486 phd = &tp->t_pcbport[cpu].t_phd; 2487 2488 while ((inp = LIST_FIRST(&phd->phd_pcblist)) != NULL) { 2489 KASSERT(inp->inp_phd == phd && inp->inp_porthash == NULL, 2490 ("not on tcpcb port cache")); 2491 LIST_REMOVE(inp, inp_portlist); 2492 in_pcbinsporthash_lport(inp); 2493 KASSERT(inp->inp_phd == tp->t_inpcb->inp_phd && 2494 inp->inp_porthash == tp->t_inpcb->inp_porthash, 2495 ("tcpcb port cache merge failed")); 2496 } 2497 } 2498 2499 void 2500 tcp_pcbport_destroy(struct tcpcb *tp) 2501 { 2502 #ifdef INVARIANTS 2503 int cpu; 2504 2505 for (cpu = 0; cpu < ncpus2; ++cpu) { 2506 KASSERT(LIST_EMPTY(&tp->t_pcbport[cpu].t_phd.phd_pcblist), 2507 ("tcpcb port cache is not empty")); 2508 } 2509 #endif 2510 kfree(tp->t_pcbport, M_PCB); 2511 tp->t_pcbport = NULL; 2512 } 2513