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_inet.h" 67 #include "opt_inet6.h" 68 #include "opt_ipsec.h" 69 #include "opt_tcpdebug.h" 70 71 #include <sys/param.h> 72 #include <sys/systm.h> 73 #include <sys/callout.h> 74 #include <sys/kernel.h> 75 #include <sys/sysctl.h> 76 #include <sys/malloc.h> 77 #include <sys/mpipe.h> 78 #include <sys/mbuf.h> 79 #ifdef INET6 80 #include <sys/domain.h> 81 #endif 82 #include <sys/proc.h> 83 #include <sys/priv.h> 84 #include <sys/socket.h> 85 #include <sys/socketops.h> 86 #include <sys/socketvar.h> 87 #include <sys/protosw.h> 88 #include <sys/random.h> 89 #include <sys/in_cksum.h> 90 #include <sys/ktr.h> 91 92 #include <net/route.h> 93 #include <net/if.h> 94 #include <net/netisr2.h> 95 96 #define _IP_VHL 97 #include <netinet/in.h> 98 #include <netinet/in_systm.h> 99 #include <netinet/ip.h> 100 #include <netinet/ip6.h> 101 #include <netinet/in_pcb.h> 102 #include <netinet6/in6_pcb.h> 103 #include <netinet/in_var.h> 104 #include <netinet/ip_var.h> 105 #include <netinet6/ip6_var.h> 106 #include <netinet/ip_icmp.h> 107 #ifdef INET6 108 #include <netinet/icmp6.h> 109 #endif 110 #include <netinet/tcp.h> 111 #include <netinet/tcp_fsm.h> 112 #include <netinet/tcp_seq.h> 113 #include <netinet/tcp_timer.h> 114 #include <netinet/tcp_timer2.h> 115 #include <netinet/tcp_var.h> 116 #include <netinet6/tcp6_var.h> 117 #include <netinet/tcpip.h> 118 #ifdef TCPDEBUG 119 #include <netinet/tcp_debug.h> 120 #endif 121 #include <netinet6/ip6protosw.h> 122 123 #ifdef IPSEC 124 #include <netinet6/ipsec.h> 125 #include <netproto/key/key.h> 126 #ifdef INET6 127 #include <netinet6/ipsec6.h> 128 #endif 129 #endif 130 131 #ifdef FAST_IPSEC 132 #include <netproto/ipsec/ipsec.h> 133 #ifdef INET6 134 #include <netproto/ipsec/ipsec6.h> 135 #endif 136 #define IPSEC 137 #endif 138 139 #include <sys/md5.h> 140 #include <machine/smp.h> 141 142 #include <sys/msgport2.h> 143 #include <sys/mplock2.h> 144 #include <net/netmsg2.h> 145 146 #if !defined(KTR_TCP) 147 #define KTR_TCP KTR_ALL 148 #endif 149 /* 150 KTR_INFO_MASTER(tcp); 151 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0); 152 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0); 153 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0); 154 #define logtcp(name) KTR_LOG(tcp_ ## name) 155 */ 156 157 #define TCP_IW_MAXSEGS_DFLT 4 158 #define TCP_IW_CAPSEGS_DFLT 4 159 160 struct inpcbinfo tcbinfo[MAXCPU]; 161 struct tcpcbackqhead tcpcbackq[MAXCPU]; 162 163 int tcp_mssdflt = TCP_MSS; 164 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 165 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size"); 166 167 #ifdef INET6 168 int tcp_v6mssdflt = TCP6_MSS; 169 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW, 170 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6"); 171 #endif 172 173 /* 174 * Minimum MSS we accept and use. This prevents DoS attacks where 175 * we are forced to a ridiculous low MSS like 20 and send hundreds 176 * of packets instead of one. The effect scales with the available 177 * bandwidth and quickly saturates the CPU and network interface 178 * with packet generation and sending. Set to zero to disable MINMSS 179 * checking. This setting prevents us from sending too small packets. 180 */ 181 int tcp_minmss = TCP_MINMSS; 182 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW, 183 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size"); 184 185 #if 0 186 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ; 187 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW, 188 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time"); 189 #endif 190 191 int tcp_do_rfc1323 = 1; 192 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW, 193 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions"); 194 195 static int tcp_tcbhashsize = 0; 196 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD, 197 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable"); 198 199 static int do_tcpdrain = 1; 200 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0, 201 "Enable tcp_drain routine for extra help when low on mbufs"); 202 203 static int icmp_may_rst = 1; 204 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0, 205 "Certain ICMP unreachable messages may abort connections in SYN_SENT"); 206 207 static int tcp_isn_reseed_interval = 0; 208 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW, 209 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret"); 210 211 /* 212 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on 213 * by default, but with generous values which should allow maximal 214 * bandwidth. In particular, the slop defaults to 50 (5 packets). 215 * 216 * The reason for doing this is that the limiter is the only mechanism we 217 * have which seems to do a really good job preventing receiver RX rings 218 * on network interfaces from getting blown out. Even though GigE/10GigE 219 * is supposed to flow control it looks like either it doesn't actually 220 * do it or Open Source drivers do not properly enable it. 221 * 222 * People using the limiter to reduce bottlenecks on slower WAN connections 223 * should set the slop to 20 (2 packets). 224 */ 225 static int tcp_inflight_enable = 1; 226 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW, 227 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting"); 228 229 static int tcp_inflight_debug = 0; 230 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW, 231 &tcp_inflight_debug, 0, "Debug TCP inflight calculations"); 232 233 /* 234 * NOTE: tcp_inflight_start is essentially the starting receive window 235 * for a connection. If set too low then fetches over tcp 236 * connections will take noticably longer to ramp-up over 237 * high-latency connections. 6144 is too low for a default, 238 * use something more reasonable. 239 */ 240 static int tcp_inflight_start = 33792; 241 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_start, CTLFLAG_RW, 242 &tcp_inflight_start, 0, "Start value for TCP inflight window"); 243 244 static int tcp_inflight_min = 6144; 245 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW, 246 &tcp_inflight_min, 0, "Lower bound for TCP inflight window"); 247 248 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT; 249 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW, 250 &tcp_inflight_max, 0, "Upper bound for TCP inflight window"); 251 252 static int tcp_inflight_stab = 50; 253 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW, 254 &tcp_inflight_stab, 0, "Fudge bw 1/10% (50=5%)"); 255 256 static int tcp_inflight_adjrtt = 2; 257 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_adjrtt, CTLFLAG_RW, 258 &tcp_inflight_adjrtt, 0, "Slop for rtt 1/(hz*32)"); 259 260 static int tcp_do_rfc3390 = 1; 261 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_RW, 262 &tcp_do_rfc3390, 0, 263 "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)"); 264 265 static u_long tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 266 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwmaxsegs, CTLFLAG_RW, 267 &tcp_iw_maxsegs, 0, "TCP IW segments max"); 268 269 static u_long tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 270 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwcapsegs, CTLFLAG_RW, 271 &tcp_iw_capsegs, 0, "TCP IW segments"); 272 273 int tcp_low_rtobase = 1; 274 SYSCTL_INT(_net_inet_tcp, OID_AUTO, low_rtobase, CTLFLAG_RW, 275 &tcp_low_rtobase, 0, "Lowering the Initial RTO (RFC 6298)"); 276 277 static int tcp_do_ncr = 1; 278 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr, CTLFLAG_RW, 279 &tcp_do_ncr, 0, "Non-Congestion Robustness (RFC 4653)"); 280 281 int tcp_ncr_rxtthresh_max = 16; 282 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr_rxtthresh_max, CTLFLAG_RW, 283 &tcp_ncr_rxtthresh_max, 0, 284 "Non-Congestion Robustness (RFC 4653), DupThresh upper limit"); 285 286 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives"); 287 static struct malloc_pipe tcptemp_mpipe; 288 289 static void tcp_willblock(void); 290 static void tcp_notify (struct inpcb *, int); 291 292 struct tcp_stats tcpstats_percpu[MAXCPU] __cachealign; 293 struct tcp_state_count tcpstate_count[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 void 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 TCP_STATE_INIT(tp); 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 } 783 784 /* 785 * Drop a TCP connection, reporting the specified error. 786 * If connection is synchronized, then send a RST to peer. 787 */ 788 struct tcpcb * 789 tcp_drop(struct tcpcb *tp, int error) 790 { 791 struct socket *so = tp->t_inpcb->inp_socket; 792 793 if (TCPS_HAVERCVDSYN(tp->t_state)) { 794 TCP_STATE_CHANGE(tp, TCPS_CLOSED); 795 tcp_output(tp); 796 tcpstat.tcps_drops++; 797 } else 798 tcpstat.tcps_conndrops++; 799 if (error == ETIMEDOUT && tp->t_softerror) 800 error = tp->t_softerror; 801 so->so_error = error; 802 return (tcp_close(tp)); 803 } 804 805 struct netmsg_listen_detach { 806 struct netmsg_base base; 807 struct tcpcb *nm_tp; 808 struct tcpcb *nm_tp_inh; 809 }; 810 811 static void 812 tcp_listen_detach_handler(netmsg_t msg) 813 { 814 struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg; 815 struct tcpcb *tp = nmsg->nm_tp; 816 int cpu = mycpuid, nextcpu; 817 818 if (tp->t_flags & TF_LISTEN) { 819 syncache_destroy(tp, nmsg->nm_tp_inh); 820 tcp_pcbport_merge_oncpu(tp); 821 } 822 823 in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]); 824 825 nextcpu = cpu + 1; 826 if (nextcpu < ncpus2) 827 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nmsg->base.lmsg); 828 else 829 lwkt_replymsg(&nmsg->base.lmsg, 0); 830 } 831 832 /* 833 * Close a TCP control block: 834 * discard all space held by the tcp 835 * discard internet protocol block 836 * wake up any sleepers 837 */ 838 struct tcpcb * 839 tcp_close(struct tcpcb *tp) 840 { 841 struct tseg_qent *q; 842 struct inpcb *inp = tp->t_inpcb; 843 struct inpcb *inp_inh = NULL; 844 struct tcpcb *tp_inh = NULL; 845 struct socket *so = inp->inp_socket; 846 struct rtentry *rt; 847 boolean_t dosavessthresh; 848 #ifdef INET6 849 boolean_t isipv6 = INP_ISIPV6(inp); 850 #else 851 const boolean_t isipv6 = FALSE; 852 #endif 853 854 if (tp->t_flags & TF_LISTEN) { 855 /* 856 * Pending socket/syncache inheritance 857 * 858 * If this is a listen(2) socket, find another listen(2) 859 * socket in the same local group, which could inherit 860 * the syncache and sockets pending on the completion 861 * and incompletion queues. 862 * 863 * NOTE: 864 * Currently the inheritance could only happen on the 865 * listen(2) sockets w/ SO_REUSEPORT set. 866 */ 867 ASSERT_IN_NETISR(0); 868 inp_inh = in_pcblocalgroup_last(&tcbinfo[0], inp); 869 if (inp_inh != NULL) 870 tp_inh = intotcpcb(inp_inh); 871 } 872 873 /* 874 * INP_WILDCARD indicates that listen(2) has been called on 875 * this socket. This implies: 876 * - A wildcard inp's hash is replicated for each protocol thread. 877 * - Syncache for this inp grows independently in each protocol 878 * thread. 879 * - There is more than one cpu 880 * 881 * We have to chain a message to the rest of the protocol threads 882 * to cleanup the wildcard hash and the syncache. The cleanup 883 * in the current protocol thread is defered till the end of this 884 * function (syncache_destroy and in_pcbdetach). 885 * 886 * NOTE: 887 * After cleanup the inp's hash and syncache entries, this inp will 888 * no longer be available to the rest of the protocol threads, so we 889 * are safe to whack the inp in the following code. 890 */ 891 if ((inp->inp_flags & INP_WILDCARD) && ncpus2 > 1) { 892 struct netmsg_listen_detach nmsg; 893 894 KKASSERT(so->so_port == netisr_cpuport(0)); 895 ASSERT_IN_NETISR(0); 896 KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]); 897 898 netmsg_init(&nmsg.base, NULL, &curthread->td_msgport, 899 MSGF_PRIORITY, tcp_listen_detach_handler); 900 nmsg.nm_tp = tp; 901 nmsg.nm_tp_inh = tp_inh; 902 lwkt_domsg(netisr_cpuport(1), &nmsg.base.lmsg, 0); 903 } 904 905 TCP_STATE_TERM(tp); 906 907 /* 908 * Make sure that all of our timers are stopped before we 909 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL), 910 * timers are never used. If timer message is never created 911 * (tp->tt_msg->tt_tcb == NULL), timers are never used too. 912 */ 913 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) { 914 tcp_callout_stop(tp, tp->tt_rexmt); 915 tcp_callout_stop(tp, tp->tt_persist); 916 tcp_callout_stop(tp, tp->tt_keep); 917 tcp_callout_stop(tp, tp->tt_2msl); 918 tcp_callout_stop(tp, tp->tt_delack); 919 } 920 921 if (tp->t_flags & TF_ONOUTPUTQ) { 922 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 923 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 924 tp->t_flags &= ~TF_ONOUTPUTQ; 925 } 926 927 /* 928 * If we got enough samples through the srtt filter, 929 * save the rtt and rttvar in the routing entry. 930 * 'Enough' is arbitrarily defined as the 16 samples. 931 * 16 samples is enough for the srtt filter to converge 932 * to within 5% of the correct value; fewer samples and 933 * we could save a very bogus rtt. 934 * 935 * Don't update the default route's characteristics and don't 936 * update anything that the user "locked". 937 */ 938 if (tp->t_rttupdated >= 16) { 939 u_long i = 0; 940 941 if (isipv6) { 942 struct sockaddr_in6 *sin6; 943 944 if ((rt = inp->in6p_route.ro_rt) == NULL) 945 goto no_valid_rt; 946 sin6 = (struct sockaddr_in6 *)rt_key(rt); 947 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 948 goto no_valid_rt; 949 } else 950 if ((rt = inp->inp_route.ro_rt) == NULL || 951 ((struct sockaddr_in *)rt_key(rt))-> 952 sin_addr.s_addr == INADDR_ANY) 953 goto no_valid_rt; 954 955 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 956 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 957 if (rt->rt_rmx.rmx_rtt && i) 958 /* 959 * filter this update to half the old & half 960 * the new values, converting scale. 961 * See route.h and tcp_var.h for a 962 * description of the scaling constants. 963 */ 964 rt->rt_rmx.rmx_rtt = 965 (rt->rt_rmx.rmx_rtt + i) / 2; 966 else 967 rt->rt_rmx.rmx_rtt = i; 968 tcpstat.tcps_cachedrtt++; 969 } 970 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 971 i = tp->t_rttvar * 972 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 973 if (rt->rt_rmx.rmx_rttvar && i) 974 rt->rt_rmx.rmx_rttvar = 975 (rt->rt_rmx.rmx_rttvar + i) / 2; 976 else 977 rt->rt_rmx.rmx_rttvar = i; 978 tcpstat.tcps_cachedrttvar++; 979 } 980 /* 981 * The old comment here said: 982 * update the pipelimit (ssthresh) if it has been updated 983 * already or if a pipesize was specified & the threshhold 984 * got below half the pipesize. I.e., wait for bad news 985 * before we start updating, then update on both good 986 * and bad news. 987 * 988 * But we want to save the ssthresh even if no pipesize is 989 * specified explicitly in the route, because such 990 * connections still have an implicit pipesize specified 991 * by the global tcp_sendspace. In the absence of a reliable 992 * way to calculate the pipesize, it will have to do. 993 */ 994 i = tp->snd_ssthresh; 995 if (rt->rt_rmx.rmx_sendpipe != 0) 996 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 997 else 998 dosavessthresh = (i < so->so_snd.ssb_hiwat/2); 999 if (dosavessthresh || 1000 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 1001 (rt->rt_rmx.rmx_ssthresh != 0))) { 1002 /* 1003 * convert the limit from user data bytes to 1004 * packets then to packet data bytes. 1005 */ 1006 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 1007 if (i < 2) 1008 i = 2; 1009 i *= tp->t_maxseg + 1010 (isipv6 ? 1011 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1012 sizeof(struct tcpiphdr)); 1013 if (rt->rt_rmx.rmx_ssthresh) 1014 rt->rt_rmx.rmx_ssthresh = 1015 (rt->rt_rmx.rmx_ssthresh + i) / 2; 1016 else 1017 rt->rt_rmx.rmx_ssthresh = i; 1018 tcpstat.tcps_cachedssthresh++; 1019 } 1020 } 1021 1022 no_valid_rt: 1023 /* free the reassembly queue, if any */ 1024 while((q = TAILQ_FIRST(&tp->t_segq)) != NULL) { 1025 TAILQ_REMOVE(&tp->t_segq, q, tqe_q); 1026 m_freem(q->tqe_m); 1027 kfree(q, M_TSEGQ); 1028 atomic_add_int(&tcp_reass_qsize, -1); 1029 } 1030 /* throw away SACK blocks in scoreboard*/ 1031 if (TCP_DO_SACK(tp)) 1032 tcp_sack_destroy(&tp->scb); 1033 1034 inp->inp_ppcb = NULL; 1035 soisdisconnected(so); 1036 /* note: pcb detached later on */ 1037 1038 tcp_destroy_timermsg(tp); 1039 tcp_output_cancel(tp); 1040 1041 if (tp->t_flags & TF_LISTEN) { 1042 syncache_destroy(tp, tp_inh); 1043 tcp_pcbport_merge_oncpu(tp); 1044 tcp_pcbport_destroy(tp); 1045 if (inp_inh != NULL && inp_inh->inp_socket != NULL) { 1046 /* 1047 * Pending sockets inheritance only needs 1048 * to be done once in the current thread, 1049 * i.e. netisr0. 1050 */ 1051 soinherit(so, inp_inh->inp_socket); 1052 } 1053 } 1054 KASSERT(tp->t_pcbport == NULL, ("tcpcb port cache is not destroyed")); 1055 1056 so_async_rcvd_drop(so); 1057 /* Drop the reference for the asynchronized pru_rcvd */ 1058 sofree(so); 1059 1060 /* 1061 * NOTE: 1062 * - Remove self from listen tcpcb per-cpu port cache _before_ 1063 * pcbdetach. 1064 * - pcbdetach removes any wildcard hash entry on the current CPU. 1065 */ 1066 tcp_pcbport_remove(inp); 1067 #ifdef INET6 1068 if (isipv6) 1069 in6_pcbdetach(inp); 1070 else 1071 #endif 1072 in_pcbdetach(inp); 1073 1074 tcpstat.tcps_closed++; 1075 return (NULL); 1076 } 1077 1078 static __inline void 1079 tcp_drain_oncpu(struct inpcbinfo *pcbinfo) 1080 { 1081 struct inpcbhead *head = &pcbinfo->pcblisthead; 1082 struct inpcb *inpb; 1083 1084 /* 1085 * Since we run in netisr, it is MP safe, even if 1086 * we block during the inpcb list iteration, i.e. 1087 * we don't need to use inpcb marker here. 1088 */ 1089 ASSERT_IN_NETISR(pcbinfo->cpu); 1090 1091 LIST_FOREACH(inpb, head, inp_list) { 1092 struct tcpcb *tcpb; 1093 struct tseg_qent *te; 1094 1095 if (inpb->inp_flags & INP_PLACEMARKER) 1096 continue; 1097 1098 tcpb = intotcpcb(inpb); 1099 KASSERT(tcpb != NULL, ("tcp_drain_oncpu: tcpb is NULL")); 1100 1101 if ((te = TAILQ_FIRST(&tcpb->t_segq)) != NULL) { 1102 TAILQ_REMOVE(&tcpb->t_segq, te, tqe_q); 1103 if (te->tqe_th->th_flags & TH_FIN) 1104 tcpb->t_flags &= ~TF_QUEDFIN; 1105 m_freem(te->tqe_m); 1106 kfree(te, M_TSEGQ); 1107 atomic_add_int(&tcp_reass_qsize, -1); 1108 /* retry */ 1109 } 1110 } 1111 } 1112 1113 static void 1114 tcp_drain_dispatch(netmsg_t nmsg) 1115 { 1116 crit_enter(); 1117 lwkt_replymsg(&nmsg->lmsg, 0); /* reply ASAP */ 1118 crit_exit(); 1119 1120 tcp_drain_oncpu(&tcbinfo[mycpuid]); 1121 } 1122 1123 static void 1124 tcp_drain_ipi(void *arg __unused) 1125 { 1126 int cpu = mycpuid; 1127 struct lwkt_msg *msg = &tcp_drain_netmsg[cpu].lmsg; 1128 1129 crit_enter(); 1130 if (msg->ms_flags & MSGF_DONE) 1131 lwkt_sendmsg_oncpu(netisr_cpuport(cpu), msg); 1132 crit_exit(); 1133 } 1134 1135 void 1136 tcp_drain(void) 1137 { 1138 cpumask_t mask; 1139 1140 if (!do_tcpdrain) 1141 return; 1142 1143 /* 1144 * Walk the tcpbs, if existing, and flush the reassembly queue, 1145 * if there is one... 1146 * XXX: The "Net/3" implementation doesn't imply that the TCP 1147 * reassembly queue should be flushed, but in a situation 1148 * where we're really low on mbufs, this is potentially 1149 * useful. 1150 * YYY: We may consider run tcp_drain_oncpu directly here, 1151 * however, that will require M_WAITOK memory allocation 1152 * for the inpcb marker. 1153 */ 1154 CPUMASK_ASSBMASK(mask, ncpus2); 1155 CPUMASK_ANDMASK(mask, smp_active_mask); 1156 if (CPUMASK_TESTNZERO(mask)) 1157 lwkt_send_ipiq_mask(mask, tcp_drain_ipi, NULL); 1158 } 1159 1160 /* 1161 * Notify a tcp user of an asynchronous error; 1162 * store error as soft error, but wake up user 1163 * (for now, won't do anything until can select for soft error). 1164 * 1165 * Do not wake up user since there currently is no mechanism for 1166 * reporting soft errors (yet - a kqueue filter may be added). 1167 */ 1168 static void 1169 tcp_notify(struct inpcb *inp, int error) 1170 { 1171 struct tcpcb *tp = intotcpcb(inp); 1172 1173 /* 1174 * Ignore some errors if we are hooked up. 1175 * If connection hasn't completed, has retransmitted several times, 1176 * and receives a second error, give up now. This is better 1177 * than waiting a long time to establish a connection that 1178 * can never complete. 1179 */ 1180 if (tp->t_state == TCPS_ESTABLISHED && 1181 (error == EHOSTUNREACH || error == ENETUNREACH || 1182 error == EHOSTDOWN)) { 1183 return; 1184 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1185 tp->t_softerror) 1186 tcp_drop(tp, error); 1187 else 1188 tp->t_softerror = error; 1189 #if 0 1190 wakeup(&so->so_timeo); 1191 sorwakeup(so); 1192 sowwakeup(so); 1193 #endif 1194 } 1195 1196 static int 1197 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1198 { 1199 int error, i, n; 1200 struct inpcb *marker; 1201 struct inpcb *inp; 1202 int origcpu, ccpu; 1203 1204 error = 0; 1205 n = 0; 1206 1207 /* 1208 * The process of preparing the TCB list is too time-consuming and 1209 * resource-intensive to repeat twice on every request. 1210 */ 1211 if (req->oldptr == NULL) { 1212 for (ccpu = 0; ccpu < ncpus2; ++ccpu) 1213 n += tcbinfo[ccpu].ipi_count; 1214 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1215 return (0); 1216 } 1217 1218 if (req->newptr != NULL) 1219 return (EPERM); 1220 1221 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1222 marker->inp_flags |= INP_PLACEMARKER; 1223 1224 /* 1225 * OK, now we're committed to doing something. Run the inpcb list 1226 * for each cpu in the system and construct the output. Use a 1227 * list placemarker to deal with list changes occuring during 1228 * copyout blockages (but otherwise depend on being on the correct 1229 * cpu to avoid races). 1230 */ 1231 origcpu = mycpu->gd_cpuid; 1232 for (ccpu = 0; ccpu < ncpus2 && error == 0; ++ccpu) { 1233 caddr_t inp_ppcb; 1234 struct xtcpcb xt; 1235 1236 lwkt_migratecpu(ccpu); 1237 1238 n = tcbinfo[ccpu].ipi_count; 1239 1240 LIST_INSERT_HEAD(&tcbinfo[ccpu].pcblisthead, marker, inp_list); 1241 i = 0; 1242 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1243 /* 1244 * process a snapshot of pcbs, ignoring placemarkers 1245 * and using our own to allow SYSCTL_OUT to block. 1246 */ 1247 LIST_REMOVE(marker, inp_list); 1248 LIST_INSERT_AFTER(inp, marker, inp_list); 1249 1250 if (inp->inp_flags & INP_PLACEMARKER) 1251 continue; 1252 if (prison_xinpcb(req->td, inp)) 1253 continue; 1254 1255 xt.xt_len = sizeof xt; 1256 bcopy(inp, &xt.xt_inp, sizeof *inp); 1257 inp_ppcb = inp->inp_ppcb; 1258 if (inp_ppcb != NULL) 1259 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1260 else 1261 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1262 if (inp->inp_socket) 1263 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1264 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1265 break; 1266 ++i; 1267 } 1268 LIST_REMOVE(marker, inp_list); 1269 if (error == 0 && i < n) { 1270 bzero(&xt, sizeof xt); 1271 xt.xt_len = sizeof xt; 1272 while (i < n) { 1273 error = SYSCTL_OUT(req, &xt, sizeof xt); 1274 if (error) 1275 break; 1276 ++i; 1277 } 1278 } 1279 } 1280 1281 /* 1282 * Make sure we are on the same cpu we were on originally, since 1283 * higher level callers expect this. Also don't pollute caches with 1284 * migrated userland data by (eventually) returning to userland 1285 * on a different cpu. 1286 */ 1287 lwkt_migratecpu(origcpu); 1288 kfree(marker, M_TEMP); 1289 return (error); 1290 } 1291 1292 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1293 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1294 1295 static int 1296 tcp_getcred(SYSCTL_HANDLER_ARGS) 1297 { 1298 struct sockaddr_in addrs[2]; 1299 struct ucred cred0, *cred = NULL; 1300 struct inpcb *inp; 1301 int cpu, origcpu, error; 1302 1303 error = priv_check(req->td, PRIV_ROOT); 1304 if (error != 0) 1305 return (error); 1306 error = SYSCTL_IN(req, addrs, sizeof addrs); 1307 if (error != 0) 1308 return (error); 1309 1310 origcpu = mycpuid; 1311 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1312 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1313 1314 lwkt_migratecpu(cpu); 1315 1316 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1317 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1318 if (inp == NULL || inp->inp_socket == NULL) { 1319 error = ENOENT; 1320 } else if (inp->inp_socket->so_cred != NULL) { 1321 cred0 = *(inp->inp_socket->so_cred); 1322 cred = &cred0; 1323 } 1324 1325 lwkt_migratecpu(origcpu); 1326 1327 if (error) 1328 return (error); 1329 1330 return SYSCTL_OUT(req, cred, sizeof(struct ucred)); 1331 } 1332 1333 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1334 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1335 1336 #ifdef INET6 1337 static int 1338 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1339 { 1340 struct sockaddr_in6 addrs[2]; 1341 struct inpcb *inp; 1342 int error; 1343 1344 error = priv_check(req->td, PRIV_ROOT); 1345 if (error != 0) 1346 return (error); 1347 error = SYSCTL_IN(req, addrs, sizeof addrs); 1348 if (error != 0) 1349 return (error); 1350 crit_enter(); 1351 inp = in6_pcblookup_hash(&tcbinfo[0], 1352 &addrs[1].sin6_addr, addrs[1].sin6_port, 1353 &addrs[0].sin6_addr, addrs[0].sin6_port, 0, NULL); 1354 if (inp == NULL || inp->inp_socket == NULL) { 1355 error = ENOENT; 1356 goto out; 1357 } 1358 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1359 out: 1360 crit_exit(); 1361 return (error); 1362 } 1363 1364 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1365 0, 0, 1366 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1367 #endif 1368 1369 struct netmsg_tcp_notify { 1370 struct netmsg_base base; 1371 inp_notify_t nm_notify; 1372 struct in_addr nm_faddr; 1373 int nm_arg; 1374 }; 1375 1376 static void 1377 tcp_notifyall_oncpu(netmsg_t msg) 1378 { 1379 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg; 1380 int nextcpu; 1381 1382 in_pcbnotifyall(&tcbinfo[mycpuid], nm->nm_faddr, 1383 nm->nm_arg, nm->nm_notify); 1384 1385 nextcpu = mycpuid + 1; 1386 if (nextcpu < ncpus2) 1387 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nm->base.lmsg); 1388 else 1389 lwkt_replymsg(&nm->base.lmsg, 0); 1390 } 1391 1392 inp_notify_t 1393 tcp_get_inpnotify(int cmd, const struct sockaddr *sa, 1394 int *arg, struct ip **ip0, int *cpuid) 1395 { 1396 struct ip *ip = *ip0; 1397 struct in_addr faddr; 1398 inp_notify_t notify = tcp_notify; 1399 1400 faddr = ((const struct sockaddr_in *)sa)->sin_addr; 1401 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1402 return NULL; 1403 1404 *arg = inetctlerrmap[cmd]; 1405 if (cmd == PRC_QUENCH) { 1406 notify = tcp_quench; 1407 } else if (icmp_may_rst && 1408 (cmd == PRC_UNREACH_ADMIN_PROHIB || 1409 cmd == PRC_UNREACH_PORT || 1410 cmd == PRC_TIMXCEED_INTRANS) && 1411 ip != NULL) { 1412 notify = tcp_drop_syn_sent; 1413 } else if (cmd == PRC_MSGSIZE) { 1414 const struct icmp *icmp = (const struct icmp *) 1415 ((caddr_t)ip - offsetof(struct icmp, icmp_ip)); 1416 1417 *arg = ntohs(icmp->icmp_nextmtu); 1418 notify = tcp_mtudisc; 1419 } else if (PRC_IS_REDIRECT(cmd)) { 1420 ip = NULL; 1421 notify = in_rtchange; 1422 } else if (cmd == PRC_HOSTDEAD) { 1423 ip = NULL; 1424 } else if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) { 1425 return NULL; 1426 } 1427 1428 if (cpuid != NULL) { 1429 if (ip == NULL) { 1430 /* Go through all CPUs */ 1431 *cpuid = ncpus; 1432 } else { 1433 const struct tcphdr *th; 1434 1435 th = (const struct tcphdr *) 1436 ((caddr_t)ip + (IP_VHL_HL(ip->ip_vhl) << 2)); 1437 *cpuid = tcp_addrcpu(faddr.s_addr, th->th_dport, 1438 ip->ip_src.s_addr, th->th_sport); 1439 } 1440 } 1441 1442 *ip0 = ip; 1443 return notify; 1444 } 1445 1446 void 1447 tcp_ctlinput(netmsg_t msg) 1448 { 1449 int cmd = msg->ctlinput.nm_cmd; 1450 struct sockaddr *sa = msg->ctlinput.nm_arg; 1451 struct ip *ip = msg->ctlinput.nm_extra; 1452 struct in_addr faddr; 1453 inp_notify_t notify; 1454 int arg, cpuid; 1455 1456 notify = tcp_get_inpnotify(cmd, sa, &arg, &ip, &cpuid); 1457 if (notify == NULL) 1458 goto done; 1459 1460 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1461 if (ip != NULL) { 1462 const struct tcphdr *th; 1463 struct inpcb *inp; 1464 1465 if (cpuid != mycpuid) 1466 goto done; 1467 1468 th = (const struct tcphdr *) 1469 ((caddr_t)ip + (IP_VHL_HL(ip->ip_vhl) << 2)); 1470 inp = in_pcblookup_hash(&tcbinfo[mycpuid], faddr, th->th_dport, 1471 ip->ip_src, th->th_sport, 0, NULL); 1472 if (inp != NULL && inp->inp_socket != NULL) { 1473 tcp_seq icmpseq = htonl(th->th_seq); 1474 struct tcpcb *tp = intotcpcb(inp); 1475 1476 if (SEQ_GEQ(icmpseq, tp->snd_una) && 1477 SEQ_LT(icmpseq, tp->snd_max)) 1478 notify(inp, arg); 1479 } else { 1480 struct in_conninfo inc; 1481 1482 inc.inc_fport = th->th_dport; 1483 inc.inc_lport = th->th_sport; 1484 inc.inc_faddr = faddr; 1485 inc.inc_laddr = ip->ip_src; 1486 #ifdef INET6 1487 inc.inc_isipv6 = 0; 1488 #endif 1489 syncache_unreach(&inc, th); 1490 } 1491 } else if (msg->ctlinput.nm_direct) { 1492 if (cpuid != ncpus && cpuid != mycpuid) 1493 goto done; 1494 if (mycpuid >= ncpus2) 1495 goto done; 1496 1497 in_pcbnotifyall(&tcbinfo[mycpuid], faddr, arg, notify); 1498 } else { 1499 struct netmsg_tcp_notify *nm; 1500 1501 ASSERT_IN_NETISR(0); 1502 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT); 1503 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1504 0, tcp_notifyall_oncpu); 1505 nm->nm_faddr = faddr; 1506 nm->nm_arg = arg; 1507 nm->nm_notify = notify; 1508 1509 lwkt_sendmsg(netisr_cpuport(0), &nm->base.lmsg); 1510 } 1511 done: 1512 lwkt_replymsg(&msg->lmsg, 0); 1513 } 1514 1515 #ifdef INET6 1516 1517 void 1518 tcp6_ctlinput(netmsg_t msg) 1519 { 1520 int cmd = msg->ctlinput.nm_cmd; 1521 struct sockaddr *sa = msg->ctlinput.nm_arg; 1522 void *d = msg->ctlinput.nm_extra; 1523 struct tcphdr th; 1524 inp_notify_t notify = tcp_notify; 1525 struct ip6_hdr *ip6; 1526 struct mbuf *m; 1527 struct ip6ctlparam *ip6cp = NULL; 1528 const struct sockaddr_in6 *sa6_src = NULL; 1529 int off; 1530 struct tcp_portonly { 1531 u_int16_t th_sport; 1532 u_int16_t th_dport; 1533 } *thp; 1534 int arg; 1535 1536 if (sa->sa_family != AF_INET6 || 1537 sa->sa_len != sizeof(struct sockaddr_in6)) { 1538 goto out; 1539 } 1540 1541 arg = 0; 1542 if (cmd == PRC_QUENCH) 1543 notify = tcp_quench; 1544 else if (cmd == PRC_MSGSIZE) { 1545 struct ip6ctlparam *ip6cp = d; 1546 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6; 1547 1548 arg = ntohl(icmp6->icmp6_mtu); 1549 notify = tcp_mtudisc; 1550 } else if (!PRC_IS_REDIRECT(cmd) && 1551 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) { 1552 goto out; 1553 } 1554 1555 /* if the parameter is from icmp6, decode it. */ 1556 if (d != NULL) { 1557 ip6cp = (struct ip6ctlparam *)d; 1558 m = ip6cp->ip6c_m; 1559 ip6 = ip6cp->ip6c_ip6; 1560 off = ip6cp->ip6c_off; 1561 sa6_src = ip6cp->ip6c_src; 1562 } else { 1563 m = NULL; 1564 ip6 = NULL; 1565 off = 0; /* fool gcc */ 1566 sa6_src = &sa6_any; 1567 } 1568 1569 if (ip6 != NULL) { 1570 struct in_conninfo inc; 1571 /* 1572 * XXX: We assume that when IPV6 is non NULL, 1573 * M and OFF are valid. 1574 */ 1575 1576 /* check if we can safely examine src and dst ports */ 1577 if (m->m_pkthdr.len < off + sizeof *thp) 1578 goto out; 1579 1580 bzero(&th, sizeof th); 1581 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1582 1583 in6_pcbnotify(&tcbinfo[0], sa, th.th_dport, 1584 (struct sockaddr *)ip6cp->ip6c_src, 1585 th.th_sport, cmd, arg, notify); 1586 1587 inc.inc_fport = th.th_dport; 1588 inc.inc_lport = th.th_sport; 1589 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1590 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1591 inc.inc_isipv6 = 1; 1592 syncache_unreach(&inc, &th); 1593 } else { 1594 in6_pcbnotify(&tcbinfo[0], sa, 0, 1595 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify); 1596 } 1597 out: 1598 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0); 1599 } 1600 1601 #endif 1602 1603 /* 1604 * Following is where TCP initial sequence number generation occurs. 1605 * 1606 * There are two places where we must use initial sequence numbers: 1607 * 1. In SYN-ACK packets. 1608 * 2. In SYN packets. 1609 * 1610 * All ISNs for SYN-ACK packets are generated by the syncache. See 1611 * tcp_syncache.c for details. 1612 * 1613 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1614 * depends on this property. In addition, these ISNs should be 1615 * unguessable so as to prevent connection hijacking. To satisfy 1616 * the requirements of this situation, the algorithm outlined in 1617 * RFC 1948 is used to generate sequence numbers. 1618 * 1619 * Implementation details: 1620 * 1621 * Time is based off the system timer, and is corrected so that it 1622 * increases by one megabyte per second. This allows for proper 1623 * recycling on high speed LANs while still leaving over an hour 1624 * before rollover. 1625 * 1626 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1627 * between seeding of isn_secret. This is normally set to zero, 1628 * as reseeding should not be necessary. 1629 * 1630 */ 1631 1632 #define ISN_BYTES_PER_SECOND 1048576 1633 1634 u_char isn_secret[32]; 1635 int isn_last_reseed; 1636 MD5_CTX isn_ctx; 1637 1638 tcp_seq 1639 tcp_new_isn(struct tcpcb *tp) 1640 { 1641 u_int32_t md5_buffer[4]; 1642 tcp_seq new_isn; 1643 1644 /* Seed if this is the first use, reseed if requested. */ 1645 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1646 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1647 < (u_int)ticks))) { 1648 read_random_unlimited(&isn_secret, sizeof isn_secret); 1649 isn_last_reseed = ticks; 1650 } 1651 1652 /* Compute the md5 hash and return the ISN. */ 1653 MD5Init(&isn_ctx); 1654 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1655 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1656 #ifdef INET6 1657 if (INP_ISIPV6(tp->t_inpcb)) { 1658 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1659 sizeof(struct in6_addr)); 1660 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1661 sizeof(struct in6_addr)); 1662 } else 1663 #endif 1664 { 1665 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1666 sizeof(struct in_addr)); 1667 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1668 sizeof(struct in_addr)); 1669 } 1670 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1671 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1672 new_isn = (tcp_seq) md5_buffer[0]; 1673 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1674 return (new_isn); 1675 } 1676 1677 /* 1678 * When a source quench is received, close congestion window 1679 * to one segment. We will gradually open it again as we proceed. 1680 */ 1681 void 1682 tcp_quench(struct inpcb *inp, int error) 1683 { 1684 struct tcpcb *tp = intotcpcb(inp); 1685 1686 KASSERT(tp != NULL, ("tcp_quench: tp is NULL")); 1687 tp->snd_cwnd = tp->t_maxseg; 1688 tp->snd_wacked = 0; 1689 } 1690 1691 /* 1692 * When a specific ICMP unreachable message is received and the 1693 * connection state is SYN-SENT, drop the connection. This behavior 1694 * is controlled by the icmp_may_rst sysctl. 1695 */ 1696 void 1697 tcp_drop_syn_sent(struct inpcb *inp, int error) 1698 { 1699 struct tcpcb *tp = intotcpcb(inp); 1700 1701 KASSERT(tp != NULL, ("tcp_drop_syn_sent: tp is NULL")); 1702 if (tp->t_state == TCPS_SYN_SENT) 1703 tcp_drop(tp, error); 1704 } 1705 1706 /* 1707 * When a `need fragmentation' ICMP is received, update our idea of the MSS 1708 * based on the new value in the route. Also nudge TCP to send something, 1709 * since we know the packet we just sent was dropped. 1710 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1711 */ 1712 void 1713 tcp_mtudisc(struct inpcb *inp, int mtu) 1714 { 1715 struct tcpcb *tp = intotcpcb(inp); 1716 struct rtentry *rt; 1717 struct socket *so = inp->inp_socket; 1718 int maxopd, mss; 1719 #ifdef INET6 1720 boolean_t isipv6 = INP_ISIPV6(inp); 1721 #else 1722 const boolean_t isipv6 = FALSE; 1723 #endif 1724 1725 KASSERT(tp != NULL, ("tcp_mtudisc: tp is NULL")); 1726 1727 /* 1728 * If no MTU is provided in the ICMP message, use the 1729 * next lower likely value, as specified in RFC 1191. 1730 */ 1731 if (mtu == 0) { 1732 int oldmtu; 1733 1734 oldmtu = tp->t_maxopd + 1735 (isipv6 ? 1736 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1737 sizeof(struct tcpiphdr)); 1738 mtu = ip_next_mtu(oldmtu, 0); 1739 } 1740 1741 if (isipv6) 1742 rt = tcp_rtlookup6(&inp->inp_inc); 1743 else 1744 rt = tcp_rtlookup(&inp->inp_inc); 1745 if (rt != NULL) { 1746 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu) 1747 mtu = rt->rt_rmx.rmx_mtu; 1748 1749 maxopd = mtu - 1750 (isipv6 ? 1751 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1752 sizeof(struct tcpiphdr)); 1753 1754 /* 1755 * XXX - The following conditional probably violates the TCP 1756 * spec. The problem is that, since we don't know the 1757 * other end's MSS, we are supposed to use a conservative 1758 * default. But, if we do that, then MTU discovery will 1759 * never actually take place, because the conservative 1760 * default is much less than the MTUs typically seen 1761 * on the Internet today. For the moment, we'll sweep 1762 * this under the carpet. 1763 * 1764 * The conservative default might not actually be a problem 1765 * if the only case this occurs is when sending an initial 1766 * SYN with options and data to a host we've never talked 1767 * to before. Then, they will reply with an MSS value which 1768 * will get recorded and the new parameters should get 1769 * recomputed. For Further Study. 1770 */ 1771 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd) 1772 maxopd = rt->rt_rmx.rmx_mssopt; 1773 } else 1774 maxopd = mtu - 1775 (isipv6 ? 1776 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1777 sizeof(struct tcpiphdr)); 1778 1779 if (tp->t_maxopd <= maxopd) 1780 return; 1781 tp->t_maxopd = maxopd; 1782 1783 mss = maxopd; 1784 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) == 1785 (TF_REQ_TSTMP | TF_RCVD_TSTMP)) 1786 mss -= TCPOLEN_TSTAMP_APPA; 1787 1788 /* round down to multiple of MCLBYTES */ 1789 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */ 1790 if (mss > MCLBYTES) 1791 mss &= ~(MCLBYTES - 1); 1792 #else 1793 if (mss > MCLBYTES) 1794 mss = (mss / MCLBYTES) * MCLBYTES; 1795 #endif 1796 1797 if (so->so_snd.ssb_hiwat < mss) 1798 mss = so->so_snd.ssb_hiwat; 1799 1800 tp->t_maxseg = mss; 1801 tp->t_rtttime = 0; 1802 tp->snd_nxt = tp->snd_una; 1803 tcp_output(tp); 1804 tcpstat.tcps_mturesent++; 1805 } 1806 1807 /* 1808 * Look-up the routing entry to the peer of this inpcb. If no route 1809 * is found and it cannot be allocated the return NULL. This routine 1810 * is called by TCP routines that access the rmx structure and by tcp_mss 1811 * to get the interface MTU. 1812 */ 1813 struct rtentry * 1814 tcp_rtlookup(struct in_conninfo *inc) 1815 { 1816 struct route *ro = &inc->inc_route; 1817 1818 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1819 /* No route yet, so try to acquire one */ 1820 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1821 /* 1822 * unused portions of the structure MUST be zero'd 1823 * out because rtalloc() treats it as opaque data 1824 */ 1825 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1826 ro->ro_dst.sa_family = AF_INET; 1827 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1828 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1829 inc->inc_faddr; 1830 rtalloc(ro); 1831 } 1832 } 1833 return (ro->ro_rt); 1834 } 1835 1836 #ifdef INET6 1837 struct rtentry * 1838 tcp_rtlookup6(struct in_conninfo *inc) 1839 { 1840 struct route_in6 *ro6 = &inc->inc6_route; 1841 1842 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1843 /* No route yet, so try to acquire one */ 1844 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1845 /* 1846 * unused portions of the structure MUST be zero'd 1847 * out because rtalloc() treats it as opaque data 1848 */ 1849 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1850 ro6->ro_dst.sin6_family = AF_INET6; 1851 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1852 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1853 rtalloc((struct route *)ro6); 1854 } 1855 } 1856 return (ro6->ro_rt); 1857 } 1858 #endif 1859 1860 #ifdef IPSEC 1861 /* compute ESP/AH header size for TCP, including outer IP header. */ 1862 size_t 1863 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1864 { 1865 struct inpcb *inp; 1866 struct mbuf *m; 1867 size_t hdrsiz; 1868 struct ip *ip; 1869 struct tcphdr *th; 1870 1871 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1872 return (0); 1873 MGETHDR(m, M_NOWAIT, MT_DATA); 1874 if (!m) 1875 return (0); 1876 1877 #ifdef INET6 1878 if (INP_ISIPV6(inp)) { 1879 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1880 1881 th = (struct tcphdr *)(ip6 + 1); 1882 m->m_pkthdr.len = m->m_len = 1883 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1884 tcp_fillheaders(tp, ip6, th, FALSE); 1885 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1886 } else 1887 #endif 1888 { 1889 ip = mtod(m, struct ip *); 1890 th = (struct tcphdr *)(ip + 1); 1891 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1892 tcp_fillheaders(tp, ip, th, FALSE); 1893 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1894 } 1895 1896 m_free(m); 1897 return (hdrsiz); 1898 } 1899 #endif 1900 1901 /* 1902 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1903 * 1904 * This code attempts to calculate the bandwidth-delay product as a 1905 * means of determining the optimal window size to maximize bandwidth, 1906 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1907 * routers. This code also does a fairly good job keeping RTTs in check 1908 * across slow links like modems. We implement an algorithm which is very 1909 * similar (but not meant to be) TCP/Vegas. The code operates on the 1910 * transmitter side of a TCP connection and so only effects the transmit 1911 * side of the connection. 1912 * 1913 * BACKGROUND: TCP makes no provision for the management of buffer space 1914 * at the end points or at the intermediate routers and switches. A TCP 1915 * stream, whether using NewReno or not, will eventually buffer as 1916 * many packets as it is able and the only reason this typically works is 1917 * due to the fairly small default buffers made available for a connection 1918 * (typicaly 16K or 32K). As machines use larger windows and/or window 1919 * scaling it is now fairly easy for even a single TCP connection to blow-out 1920 * all available buffer space not only on the local interface, but on 1921 * intermediate routers and switches as well. NewReno makes a misguided 1922 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1923 * then backing off, then steadily increasing the window again until another 1924 * failure occurs, ad-infinitum. This results in terrible oscillation that 1925 * is only made worse as network loads increase and the idea of intentionally 1926 * blowing out network buffers is, frankly, a terrible way to manage network 1927 * resources. 1928 * 1929 * It is far better to limit the transmit window prior to the failure 1930 * condition being achieved. There are two general ways to do this: First 1931 * you can 'scan' through different transmit window sizes and locate the 1932 * point where the RTT stops increasing, indicating that you have filled the 1933 * pipe, then scan backwards until you note that RTT stops decreasing, then 1934 * repeat ad-infinitum. This method works in principle but has severe 1935 * implementation issues due to RTT variances, timer granularity, and 1936 * instability in the algorithm which can lead to many false positives and 1937 * create oscillations as well as interact badly with other TCP streams 1938 * implementing the same algorithm. 1939 * 1940 * The second method is to limit the window to the bandwidth delay product 1941 * of the link. This is the method we implement. RTT variances and our 1942 * own manipulation of the congestion window, bwnd, can potentially 1943 * destabilize the algorithm. For this reason we have to stabilize the 1944 * elements used to calculate the window. We do this by using the minimum 1945 * observed RTT, the long term average of the observed bandwidth, and 1946 * by adding two segments worth of slop. It isn't perfect but it is able 1947 * to react to changing conditions and gives us a very stable basis on 1948 * which to extend the algorithm. 1949 */ 1950 void 1951 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1952 { 1953 u_long bw; 1954 u_long ibw; 1955 u_long bwnd; 1956 int save_ticks; 1957 int delta_ticks; 1958 1959 /* 1960 * If inflight_enable is disabled in the middle of a tcp connection, 1961 * make sure snd_bwnd is effectively disabled. 1962 */ 1963 if (!tcp_inflight_enable) { 1964 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1965 tp->snd_bandwidth = 0; 1966 return; 1967 } 1968 1969 /* 1970 * Validate the delta time. If a connection is new or has been idle 1971 * a long time we have to reset the bandwidth calculator. 1972 */ 1973 save_ticks = ticks; 1974 cpu_ccfence(); 1975 delta_ticks = save_ticks - tp->t_bw_rtttime; 1976 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1977 tp->t_bw_rtttime = save_ticks; 1978 tp->t_bw_rtseq = ack_seq; 1979 if (tp->snd_bandwidth == 0) 1980 tp->snd_bandwidth = tcp_inflight_start; 1981 return; 1982 } 1983 1984 /* 1985 * A delta of at least 1 tick is required. Waiting 2 ticks will 1986 * result in better (bw) accuracy. More than that and the ramp-up 1987 * will be too slow. 1988 */ 1989 if (delta_ticks == 0 || delta_ticks == 1) 1990 return; 1991 1992 /* 1993 * Sanity check, plus ignore pure window update acks. 1994 */ 1995 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1996 return; 1997 1998 /* 1999 * Figure out the bandwidth. Due to the tick granularity this 2000 * is a very rough number and it MUST be averaged over a fairly 2001 * long period of time. XXX we need to take into account a link 2002 * that is not using all available bandwidth, but for now our 2003 * slop will ramp us up if this case occurs and the bandwidth later 2004 * increases. 2005 */ 2006 ibw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 2007 tp->t_bw_rtttime = save_ticks; 2008 tp->t_bw_rtseq = ack_seq; 2009 bw = ((int64_t)tp->snd_bandwidth * 15 + ibw) >> 4; 2010 2011 tp->snd_bandwidth = bw; 2012 2013 /* 2014 * Calculate the semi-static bandwidth delay product, plus two maximal 2015 * segments. The additional slop puts us squarely in the sweet 2016 * spot and also handles the bandwidth run-up case. Without the 2017 * slop we could be locking ourselves into a lower bandwidth. 2018 * 2019 * At very high speeds the bw calculation can become overly sensitive 2020 * and error prone when delta_ticks is low (e.g. usually 1). To deal 2021 * with the problem the stab must be scaled to the bw. A stab of 50 2022 * (the default) increases the bw for the purposes of the bwnd 2023 * calculation by 5%. 2024 * 2025 * Situations Handled: 2026 * (1) Prevents over-queueing of packets on LANs, especially on 2027 * high speed LANs, allowing larger TCP buffers to be 2028 * specified, and also does a good job preventing 2029 * over-queueing of packets over choke points like modems 2030 * (at least for the transmit side). 2031 * 2032 * (2) Is able to handle changing network loads (bandwidth 2033 * drops so bwnd drops, bandwidth increases so bwnd 2034 * increases). 2035 * 2036 * (3) Theoretically should stabilize in the face of multiple 2037 * connections implementing the same algorithm (this may need 2038 * a little work). 2039 * 2040 * (4) Stability value (defaults to 20 = 2 maximal packets) can 2041 * be adjusted with a sysctl but typically only needs to be on 2042 * very slow connections. A value no smaller then 5 should 2043 * be used, but only reduce this default if you have no other 2044 * choice. 2045 */ 2046 2047 #define USERTT ((tp->t_srtt + tp->t_rttvar) + tcp_inflight_adjrtt) 2048 bw += bw * tcp_inflight_stab / 1000; 2049 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 2050 (int)tp->t_maxseg * 2; 2051 #undef USERTT 2052 2053 if (tcp_inflight_debug > 0) { 2054 static int ltime; 2055 if ((u_int)(save_ticks - ltime) >= hz / tcp_inflight_debug) { 2056 ltime = save_ticks; 2057 kprintf("%p ibw %ld bw %ld rttvar %d srtt %d " 2058 "bwnd %ld delta %d snd_win %ld\n", 2059 tp, ibw, bw, tp->t_rttvar, tp->t_srtt, 2060 bwnd, delta_ticks, tp->snd_wnd); 2061 } 2062 } 2063 if ((long)bwnd < tcp_inflight_min) 2064 bwnd = tcp_inflight_min; 2065 if (bwnd > tcp_inflight_max) 2066 bwnd = tcp_inflight_max; 2067 if ((long)bwnd < tp->t_maxseg * 2) 2068 bwnd = tp->t_maxseg * 2; 2069 tp->snd_bwnd = bwnd; 2070 } 2071 2072 static void 2073 tcp_rmx_iwsegs(struct tcpcb *tp, u_long *maxsegs, u_long *capsegs) 2074 { 2075 struct rtentry *rt; 2076 struct inpcb *inp = tp->t_inpcb; 2077 #ifdef INET6 2078 boolean_t isipv6 = INP_ISIPV6(inp); 2079 #else 2080 const boolean_t isipv6 = FALSE; 2081 #endif 2082 2083 /* XXX */ 2084 if (tcp_iw_maxsegs < TCP_IW_MAXSEGS_DFLT) 2085 tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 2086 if (tcp_iw_capsegs < TCP_IW_CAPSEGS_DFLT) 2087 tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 2088 2089 if (isipv6) 2090 rt = tcp_rtlookup6(&inp->inp_inc); 2091 else 2092 rt = tcp_rtlookup(&inp->inp_inc); 2093 if (rt == NULL || 2094 rt->rt_rmx.rmx_iwmaxsegs < TCP_IW_MAXSEGS_DFLT || 2095 rt->rt_rmx.rmx_iwcapsegs < TCP_IW_CAPSEGS_DFLT) { 2096 *maxsegs = tcp_iw_maxsegs; 2097 *capsegs = tcp_iw_capsegs; 2098 return; 2099 } 2100 *maxsegs = rt->rt_rmx.rmx_iwmaxsegs; 2101 *capsegs = rt->rt_rmx.rmx_iwcapsegs; 2102 } 2103 2104 u_long 2105 tcp_initial_window(struct tcpcb *tp) 2106 { 2107 if (tcp_do_rfc3390) { 2108 /* 2109 * RFC3390: 2110 * "If the SYN or SYN/ACK is lost, the initial window 2111 * used by a sender after a correctly transmitted SYN 2112 * MUST be one segment consisting of MSS bytes." 2113 * 2114 * However, we do something a little bit more aggressive 2115 * then RFC3390 here: 2116 * - Only if time spent in the SYN or SYN|ACK retransmition 2117 * >= 3 seconds, the IW is reduced. We do this mainly 2118 * because when RFC3390 is published, the initial RTO is 2119 * still 3 seconds (the threshold we test here), while 2120 * after RFC6298, the initial RTO is 1 second. This 2121 * behaviour probably still falls within the spirit of 2122 * RFC3390. 2123 * - When IW is reduced, 2*MSS is used instead of 1*MSS. 2124 * Mainly to avoid sender and receiver deadlock until 2125 * delayed ACK timer expires. And even RFC2581 does not 2126 * try to reduce IW upon SYN or SYN|ACK retransmition 2127 * timeout. 2128 * 2129 * See also: 2130 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03 2131 */ 2132 if (tp->t_rxtsyn >= TCPTV_RTOBASE3) { 2133 return (2 * tp->t_maxseg); 2134 } else { 2135 u_long maxsegs, capsegs; 2136 2137 tcp_rmx_iwsegs(tp, &maxsegs, &capsegs); 2138 return min(maxsegs * tp->t_maxseg, 2139 max(2 * tp->t_maxseg, capsegs * 1460)); 2140 } 2141 } else { 2142 /* 2143 * Even RFC2581 (back to 1999) allows 2*SMSS IW. 2144 * 2145 * Mainly to avoid sender and receiver deadlock 2146 * until delayed ACK timer expires. 2147 */ 2148 return (2 * tp->t_maxseg); 2149 } 2150 } 2151 2152 #ifdef TCP_SIGNATURE 2153 /* 2154 * Compute TCP-MD5 hash of a TCP segment. (RFC2385) 2155 * 2156 * We do this over ip, tcphdr, segment data, and the key in the SADB. 2157 * When called from tcp_input(), we can be sure that th_sum has been 2158 * zeroed out and verified already. 2159 * 2160 * Return 0 if successful, otherwise return -1. 2161 * 2162 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a 2163 * search with the destination IP address, and a 'magic SPI' to be 2164 * determined by the application. This is hardcoded elsewhere to 1179 2165 * right now. Another branch of this code exists which uses the SPD to 2166 * specify per-application flows but it is unstable. 2167 */ 2168 int 2169 tcpsignature_compute( 2170 struct mbuf *m, /* mbuf chain */ 2171 int len, /* length of TCP data */ 2172 int optlen, /* length of TCP options */ 2173 u_char *buf, /* storage for MD5 digest */ 2174 u_int direction) /* direction of flow */ 2175 { 2176 struct ippseudo ippseudo; 2177 MD5_CTX ctx; 2178 int doff; 2179 struct ip *ip; 2180 struct ipovly *ipovly; 2181 struct secasvar *sav; 2182 struct tcphdr *th; 2183 #ifdef INET6 2184 struct ip6_hdr *ip6; 2185 struct in6_addr in6; 2186 uint32_t plen; 2187 uint16_t nhdr; 2188 #endif /* INET6 */ 2189 u_short savecsum; 2190 2191 KASSERT(m != NULL, ("passed NULL mbuf. Game over.")); 2192 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature")); 2193 /* 2194 * Extract the destination from the IP header in the mbuf. 2195 */ 2196 ip = mtod(m, struct ip *); 2197 #ifdef INET6 2198 ip6 = NULL; /* Make the compiler happy. */ 2199 #endif /* INET6 */ 2200 /* 2201 * Look up an SADB entry which matches the address found in 2202 * the segment. 2203 */ 2204 switch (IP_VHL_V(ip->ip_vhl)) { 2205 case IPVERSION: 2206 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst, 2207 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2208 break; 2209 #ifdef INET6 2210 case (IPV6_VERSION >> 4): 2211 ip6 = mtod(m, struct ip6_hdr *); 2212 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst, 2213 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2214 break; 2215 #endif /* INET6 */ 2216 default: 2217 return (EINVAL); 2218 /* NOTREACHED */ 2219 break; 2220 } 2221 if (sav == NULL) { 2222 kprintf("%s: SADB lookup failed\n", __func__); 2223 return (EINVAL); 2224 } 2225 MD5Init(&ctx); 2226 2227 /* 2228 * Step 1: Update MD5 hash with IP pseudo-header. 2229 * 2230 * XXX The ippseudo header MUST be digested in network byte order, 2231 * or else we'll fail the regression test. Assume all fields we've 2232 * been doing arithmetic on have been in host byte order. 2233 * XXX One cannot depend on ipovly->ih_len here. When called from 2234 * tcp_output(), the underlying ip_len member has not yet been set. 2235 */ 2236 switch (IP_VHL_V(ip->ip_vhl)) { 2237 case IPVERSION: 2238 ipovly = (struct ipovly *)ip; 2239 ippseudo.ippseudo_src = ipovly->ih_src; 2240 ippseudo.ippseudo_dst = ipovly->ih_dst; 2241 ippseudo.ippseudo_pad = 0; 2242 ippseudo.ippseudo_p = IPPROTO_TCP; 2243 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen); 2244 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo)); 2245 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip)); 2246 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen; 2247 break; 2248 #ifdef INET6 2249 /* 2250 * RFC 2385, 2.0 Proposal 2251 * For IPv6, the pseudo-header is as described in RFC 2460, namely the 2252 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero- 2253 * extended next header value (to form 32 bits), and 32-bit segment 2254 * length. 2255 * Note: Upper-Layer Packet Length comes before Next Header. 2256 */ 2257 case (IPV6_VERSION >> 4): 2258 in6 = ip6->ip6_src; 2259 in6_clearscope(&in6); 2260 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2261 in6 = ip6->ip6_dst; 2262 in6_clearscope(&in6); 2263 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2264 plen = htonl(len + sizeof(struct tcphdr) + optlen); 2265 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t)); 2266 nhdr = 0; 2267 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2268 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2269 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2270 nhdr = IPPROTO_TCP; 2271 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2272 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr)); 2273 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen; 2274 break; 2275 #endif /* INET6 */ 2276 default: 2277 return (EINVAL); 2278 /* NOTREACHED */ 2279 break; 2280 } 2281 /* 2282 * Step 2: Update MD5 hash with TCP header, excluding options. 2283 * The TCP checksum must be set to zero. 2284 */ 2285 savecsum = th->th_sum; 2286 th->th_sum = 0; 2287 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr)); 2288 th->th_sum = savecsum; 2289 /* 2290 * Step 3: Update MD5 hash with TCP segment data. 2291 * Use m_apply() to avoid an early m_pullup(). 2292 */ 2293 if (len > 0) 2294 m_apply(m, doff, len, tcpsignature_apply, &ctx); 2295 /* 2296 * Step 4: Update MD5 hash with shared secret. 2297 */ 2298 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth)); 2299 MD5Final(buf, &ctx); 2300 key_sa_recordxfer(sav, m); 2301 key_freesav(sav); 2302 return (0); 2303 } 2304 2305 int 2306 tcpsignature_apply(void *fstate, void *data, unsigned int len) 2307 { 2308 2309 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len); 2310 return (0); 2311 } 2312 #endif /* TCP_SIGNATURE */ 2313 2314 static void 2315 tcp_drop_sysctl_dispatch(netmsg_t nmsg) 2316 { 2317 struct lwkt_msg *lmsg = &nmsg->lmsg; 2318 /* addrs[0] is a foreign socket, addrs[1] is a local one. */ 2319 struct sockaddr_storage *addrs = lmsg->u.ms_resultp; 2320 int error; 2321 struct sockaddr_in *fin, *lin; 2322 #ifdef INET6 2323 struct sockaddr_in6 *fin6, *lin6; 2324 struct in6_addr f6, l6; 2325 #endif 2326 struct inpcb *inp; 2327 2328 switch (addrs[0].ss_family) { 2329 #ifdef INET6 2330 case AF_INET6: 2331 fin6 = (struct sockaddr_in6 *)&addrs[0]; 2332 lin6 = (struct sockaddr_in6 *)&addrs[1]; 2333 error = in6_embedscope(&f6, fin6, NULL, NULL); 2334 if (error) 2335 goto done; 2336 error = in6_embedscope(&l6, lin6, NULL, NULL); 2337 if (error) 2338 goto done; 2339 inp = in6_pcblookup_hash(&tcbinfo[mycpuid], &f6, 2340 fin6->sin6_port, &l6, lin6->sin6_port, FALSE, NULL); 2341 break; 2342 #endif 2343 #ifdef INET 2344 case AF_INET: 2345 fin = (struct sockaddr_in *)&addrs[0]; 2346 lin = (struct sockaddr_in *)&addrs[1]; 2347 inp = in_pcblookup_hash(&tcbinfo[mycpuid], fin->sin_addr, 2348 fin->sin_port, lin->sin_addr, lin->sin_port, FALSE, NULL); 2349 break; 2350 #endif 2351 default: 2352 /* 2353 * Must not reach here, since the address family was 2354 * checked in sysctl handler. 2355 */ 2356 panic("unknown address family %d", addrs[0].ss_family); 2357 } 2358 if (inp != NULL) { 2359 struct tcpcb *tp = intotcpcb(inp); 2360 2361 KASSERT((inp->inp_flags & INP_WILDCARD) == 0, 2362 ("in wildcard hash")); 2363 KASSERT(tp != NULL, ("tcp_drop_sysctl_dispatch: tp is NULL")); 2364 KASSERT((tp->t_flags & TF_LISTEN) == 0, ("listen socket")); 2365 tcp_drop(tp, ECONNABORTED); 2366 error = 0; 2367 } else { 2368 error = ESRCH; 2369 } 2370 #ifdef INET6 2371 done: 2372 #endif 2373 lwkt_replymsg(lmsg, error); 2374 } 2375 2376 static int 2377 sysctl_tcp_drop(SYSCTL_HANDLER_ARGS) 2378 { 2379 /* addrs[0] is a foreign socket, addrs[1] is a local one. */ 2380 struct sockaddr_storage addrs[2]; 2381 struct sockaddr_in *fin, *lin; 2382 #ifdef INET6 2383 struct sockaddr_in6 *fin6, *lin6; 2384 #endif 2385 struct netmsg_base nmsg; 2386 struct lwkt_msg *lmsg = &nmsg.lmsg; 2387 struct lwkt_port *port = NULL; 2388 int error; 2389 2390 fin = lin = NULL; 2391 #ifdef INET6 2392 fin6 = lin6 = NULL; 2393 #endif 2394 error = 0; 2395 2396 if (req->oldptr != NULL || req->oldlen != 0) 2397 return (EINVAL); 2398 if (req->newptr == NULL) 2399 return (EPERM); 2400 if (req->newlen < sizeof(addrs)) 2401 return (ENOMEM); 2402 error = SYSCTL_IN(req, &addrs, sizeof(addrs)); 2403 if (error) 2404 return (error); 2405 2406 switch (addrs[0].ss_family) { 2407 #ifdef INET6 2408 case AF_INET6: 2409 fin6 = (struct sockaddr_in6 *)&addrs[0]; 2410 lin6 = (struct sockaddr_in6 *)&addrs[1]; 2411 if (fin6->sin6_len != sizeof(struct sockaddr_in6) || 2412 lin6->sin6_len != sizeof(struct sockaddr_in6)) 2413 return (EINVAL); 2414 if (IN6_IS_ADDR_V4MAPPED(&fin6->sin6_addr) || 2415 IN6_IS_ADDR_V4MAPPED(&lin6->sin6_addr)) 2416 return (EADDRNOTAVAIL); 2417 #if 0 2418 error = sa6_embedscope(fin6, V_ip6_use_defzone); 2419 if (error) 2420 return (error); 2421 error = sa6_embedscope(lin6, V_ip6_use_defzone); 2422 if (error) 2423 return (error); 2424 #endif 2425 port = tcp6_addrport(); 2426 break; 2427 #endif 2428 #ifdef INET 2429 case AF_INET: 2430 fin = (struct sockaddr_in *)&addrs[0]; 2431 lin = (struct sockaddr_in *)&addrs[1]; 2432 if (fin->sin_len != sizeof(struct sockaddr_in) || 2433 lin->sin_len != sizeof(struct sockaddr_in)) 2434 return (EINVAL); 2435 port = tcp_addrport(fin->sin_addr.s_addr, fin->sin_port, 2436 lin->sin_addr.s_addr, lin->sin_port); 2437 break; 2438 #endif 2439 default: 2440 return (EINVAL); 2441 } 2442 2443 netmsg_init(&nmsg, NULL, &curthread->td_msgport, 0, 2444 tcp_drop_sysctl_dispatch); 2445 lmsg->u.ms_resultp = addrs; 2446 return lwkt_domsg(port, lmsg, 0); 2447 } 2448 2449 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, drop, 2450 CTLTYPE_STRUCT | CTLFLAG_WR | CTLFLAG_SKIP, NULL, 2451 0, sysctl_tcp_drop, "", "Drop TCP connection"); 2452 2453 static int 2454 sysctl_tcps_count(SYSCTL_HANDLER_ARGS) 2455 { 2456 u_long state_count[TCP_NSTATES]; 2457 int cpu; 2458 2459 memset(state_count, 0, sizeof(state_count)); 2460 for (cpu = 0; cpu < ncpus2; ++cpu) { 2461 int i; 2462 2463 for (i = 0; i < TCP_NSTATES; ++i) 2464 state_count[i] += tcpstate_count[cpu].tcps_count[i]; 2465 } 2466 2467 return sysctl_handle_opaque(oidp, state_count, sizeof(state_count), req); 2468 } 2469 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, state_count, 2470 CTLTYPE_OPAQUE | CTLFLAG_RD, NULL, 0, 2471 sysctl_tcps_count, "LU", "TCP connection counts by state"); 2472 2473 void 2474 tcp_pcbport_create(struct tcpcb *tp) 2475 { 2476 int cpu; 2477 2478 KASSERT((tp->t_flags & TF_LISTEN) && tp->t_state == TCPS_LISTEN, 2479 ("not a listen tcpcb")); 2480 2481 KASSERT(tp->t_pcbport == NULL, ("tcpcb port cache was created")); 2482 tp->t_pcbport = kmalloc_cachealign(sizeof(struct tcp_pcbport) * ncpus2, 2483 M_PCB, M_WAITOK); 2484 2485 for (cpu = 0; cpu < ncpus2; ++cpu) { 2486 struct inpcbport *phd; 2487 2488 phd = &tp->t_pcbport[cpu].t_phd; 2489 LIST_INIT(&phd->phd_pcblist); 2490 /* Though, not used ... */ 2491 phd->phd_port = tp->t_inpcb->inp_lport; 2492 } 2493 } 2494 2495 void 2496 tcp_pcbport_merge_oncpu(struct tcpcb *tp) 2497 { 2498 struct inpcbport *phd; 2499 struct inpcb *inp; 2500 int cpu = mycpuid; 2501 2502 KASSERT(cpu < ncpus2, ("invalid cpu%d", cpu)); 2503 phd = &tp->t_pcbport[cpu].t_phd; 2504 2505 while ((inp = LIST_FIRST(&phd->phd_pcblist)) != NULL) { 2506 KASSERT(inp->inp_phd == phd && inp->inp_porthash == NULL, 2507 ("not on tcpcb port cache")); 2508 LIST_REMOVE(inp, inp_portlist); 2509 in_pcbinsporthash_lport(inp); 2510 KASSERT(inp->inp_phd == tp->t_inpcb->inp_phd && 2511 inp->inp_porthash == tp->t_inpcb->inp_porthash, 2512 ("tcpcb port cache merge failed")); 2513 } 2514 } 2515 2516 void 2517 tcp_pcbport_destroy(struct tcpcb *tp) 2518 { 2519 #ifdef INVARIANTS 2520 int cpu; 2521 2522 for (cpu = 0; cpu < ncpus2; ++cpu) { 2523 KASSERT(LIST_EMPTY(&tp->t_pcbport[cpu].t_phd.phd_pcblist), 2524 ("tcpcb port cache is not empty")); 2525 } 2526 #endif 2527 kfree(tp->t_pcbport, M_PCB); 2528 tp->t_pcbport = NULL; 2529 } 2530