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 struct inpcb *inp = NULL; 565 boolean_t use_tmpro = TRUE; 566 #ifdef INET6 567 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6); 568 #else 569 const boolean_t isipv6 = FALSE; 570 #endif 571 572 if (tp != NULL) { 573 inp = tp->t_inpcb; 574 if (!(flags & TH_RST)) { 575 win = ssb_space(&inp->inp_socket->so_rcv); 576 if (win < 0) 577 win = 0; 578 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 579 win = (long)TCP_MAXWIN << tp->rcv_scale; 580 } 581 /* 582 * Don't use the route cache of a listen socket, 583 * it is not MPSAFE; use temporary route cache. 584 */ 585 if (tp->t_state != TCPS_LISTEN) { 586 if (isipv6) 587 ro6 = &inp->in6p_route; 588 else 589 ro = &inp->inp_route; 590 use_tmpro = FALSE; 591 } 592 } 593 if (use_tmpro) { 594 if (isipv6) { 595 ro6 = &sro6; 596 bzero(ro6, sizeof *ro6); 597 } else { 598 ro = &sro; 599 bzero(ro, sizeof *ro); 600 } 601 } 602 if (m == NULL) { 603 m = m_gethdr(M_NOWAIT, MT_HEADER); 604 if (m == NULL) 605 return; 606 tlen = 0; 607 m->m_data += max_linkhdr; 608 if (isipv6) { 609 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); 610 ip6 = mtod(m, struct ip6_hdr *); 611 nth = (struct tcphdr *)(ip6 + 1); 612 } else { 613 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip)); 614 ip = mtod(m, struct ip *); 615 nth = (struct tcphdr *)(ip + 1); 616 } 617 bcopy(th, nth, sizeof(struct tcphdr)); 618 flags = TH_ACK; 619 } else { 620 m_freem(m->m_next); 621 m->m_next = NULL; 622 m->m_data = (caddr_t)ipgen; 623 /* m_len is set later */ 624 tlen = 0; 625 #define xchg(a, b, type) { type t; t = a; a = b; b = t; } 626 if (isipv6) { 627 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); 628 nth = (struct tcphdr *)(ip6 + 1); 629 } else { 630 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long); 631 nth = (struct tcphdr *)(ip + 1); 632 } 633 if (th != nth) { 634 /* 635 * this is usually a case when an extension header 636 * exists between the IPv6 header and the 637 * TCP header. 638 */ 639 nth->th_sport = th->th_sport; 640 nth->th_dport = th->th_dport; 641 } 642 xchg(nth->th_dport, nth->th_sport, n_short); 643 #undef xchg 644 } 645 if (isipv6) { 646 ip6->ip6_flow = 0; 647 ip6->ip6_vfc = IPV6_VERSION; 648 ip6->ip6_nxt = IPPROTO_TCP; 649 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen)); 650 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 651 } else { 652 tlen += sizeof(struct tcpiphdr); 653 ip->ip_len = tlen; 654 ip->ip_ttl = ip_defttl; 655 } 656 m->m_len = tlen; 657 m->m_pkthdr.len = tlen; 658 m->m_pkthdr.rcvif = NULL; 659 nth->th_seq = htonl(seq); 660 nth->th_ack = htonl(ack); 661 nth->th_x2 = 0; 662 nth->th_off = sizeof(struct tcphdr) >> 2; 663 nth->th_flags = flags; 664 if (tp != NULL) 665 nth->th_win = htons((u_short) (win >> tp->rcv_scale)); 666 else 667 nth->th_win = htons((u_short)win); 668 nth->th_urp = 0; 669 if (isipv6) { 670 nth->th_sum = 0; 671 nth->th_sum = in6_cksum(m, IPPROTO_TCP, 672 sizeof(struct ip6_hdr), 673 tlen - sizeof(struct ip6_hdr)); 674 ip6->ip6_hlim = in6_selecthlim(inp, 675 (ro6 && ro6->ro_rt) ? ro6->ro_rt->rt_ifp : NULL); 676 } else { 677 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 678 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); 679 m->m_pkthdr.csum_flags = CSUM_TCP; 680 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 681 m->m_pkthdr.csum_thlen = sizeof(struct tcphdr); 682 } 683 #ifdef TCPDEBUG 684 if (tp == NULL || (inp->inp_socket->so_options & SO_DEBUG)) 685 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); 686 #endif 687 if (isipv6) { 688 ip6_output(m, NULL, ro6, ipflags, NULL, NULL, inp); 689 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) { 690 RTFREE(ro6->ro_rt); 691 ro6->ro_rt = NULL; 692 } 693 } else { 694 if (inp != NULL && (inp->inp_flags & INP_HASH)) 695 m_sethash(m, inp->inp_hashval); 696 ipflags |= IP_DEBUGROUTE; 697 ip_output(m, NULL, ro, ipflags, NULL, inp); 698 if ((ro == &sro) && (ro->ro_rt != NULL)) { 699 RTFREE(ro->ro_rt); 700 ro->ro_rt = NULL; 701 } 702 } 703 } 704 705 /* 706 * Create a new TCP control block, making an 707 * empty reassembly queue and hooking it to the argument 708 * protocol control block. The `inp' parameter must have 709 * come from the zone allocator set up in tcp_init(). 710 */ 711 void 712 tcp_newtcpcb(struct inpcb *inp) 713 { 714 struct inp_tp *it; 715 struct tcpcb *tp; 716 #ifdef INET6 717 boolean_t isipv6 = INP_ISIPV6(inp); 718 #else 719 const boolean_t isipv6 = FALSE; 720 #endif 721 722 it = (struct inp_tp *)inp; 723 tp = &it->tcb; 724 bzero(tp, sizeof(struct tcpcb)); 725 TAILQ_INIT(&tp->t_segq); 726 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 727 tp->t_rxtthresh = tcprexmtthresh; 728 729 /* Set up our timeouts. */ 730 tp->tt_rexmt = &it->inp_tp_rexmt; 731 tp->tt_persist = &it->inp_tp_persist; 732 tp->tt_keep = &it->inp_tp_keep; 733 tp->tt_2msl = &it->inp_tp_2msl; 734 tp->tt_delack = &it->inp_tp_delack; 735 tcp_inittimers(tp); 736 737 /* 738 * Zero out timer message. We don't create it here, 739 * since the current CPU may not be the owner of this 740 * inpcb. 741 */ 742 tp->tt_msg = &it->inp_tp_timermsg; 743 bzero(tp->tt_msg, sizeof(*tp->tt_msg)); 744 745 tp->t_keepinit = tcp_keepinit; 746 tp->t_keepidle = tcp_keepidle; 747 tp->t_keepintvl = tcp_keepintvl; 748 tp->t_keepcnt = tcp_keepcnt; 749 tp->t_maxidle = tp->t_keepintvl * tp->t_keepcnt; 750 751 if (tcp_do_ncr) 752 tp->t_flags |= TF_NCR; 753 if (tcp_do_rfc1323) 754 tp->t_flags |= (TF_REQ_SCALE | TF_REQ_TSTMP); 755 756 tp->t_inpcb = inp; /* XXX */ 757 TCP_STATE_INIT(tp); 758 /* 759 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 760 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 761 * reasonable initial retransmit time. 762 */ 763 tp->t_srtt = TCPTV_SRTTBASE; 764 tp->t_rttvar = 765 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 766 tp->t_rttmin = tcp_rexmit_min; 767 tp->t_rxtcur = TCPTV_RTOBASE; 768 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 769 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 770 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 771 tp->snd_last = ticks; 772 tp->t_rcvtime = ticks; 773 /* 774 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 775 * because the socket may be bound to an IPv6 wildcard address, 776 * which may match an IPv4-mapped IPv6 address. 777 */ 778 inp->inp_ip_ttl = ip_defttl; 779 inp->inp_ppcb = tp; 780 tcp_sack_tcpcb_init(tp); 781 782 tp->tt_sndmore = &it->inp_tp_sndmore; 783 tcp_output_init(tp); 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 TCP_STATE_CHANGE(tp, 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 TCP_STATE_TERM(tp); 908 909 /* 910 * Make sure that all of our timers are stopped before we 911 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL), 912 * timers are never used. If timer message is never created 913 * (tp->tt_msg->tt_tcb == NULL), timers are never used too. 914 */ 915 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) { 916 tcp_callout_stop(tp, tp->tt_rexmt); 917 tcp_callout_stop(tp, tp->tt_persist); 918 tcp_callout_stop(tp, tp->tt_keep); 919 tcp_callout_stop(tp, tp->tt_2msl); 920 tcp_callout_stop(tp, tp->tt_delack); 921 } 922 923 if (tp->t_flags & TF_ONOUTPUTQ) { 924 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 925 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq); 926 tp->t_flags &= ~TF_ONOUTPUTQ; 927 } 928 929 /* 930 * If we got enough samples through the srtt filter, 931 * save the rtt and rttvar in the routing entry. 932 * 'Enough' is arbitrarily defined as the 16 samples. 933 * 16 samples is enough for the srtt filter to converge 934 * to within 5% of the correct value; fewer samples and 935 * we could save a very bogus rtt. 936 * 937 * Don't update the default route's characteristics and don't 938 * update anything that the user "locked". 939 */ 940 if (tp->t_rttupdated >= 16) { 941 u_long i = 0; 942 943 if (isipv6) { 944 struct sockaddr_in6 *sin6; 945 946 if ((rt = inp->in6p_route.ro_rt) == NULL) 947 goto no_valid_rt; 948 sin6 = (struct sockaddr_in6 *)rt_key(rt); 949 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 950 goto no_valid_rt; 951 } else 952 if ((rt = inp->inp_route.ro_rt) == NULL || 953 ((struct sockaddr_in *)rt_key(rt))-> 954 sin_addr.s_addr == INADDR_ANY) 955 goto no_valid_rt; 956 957 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 958 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 959 if (rt->rt_rmx.rmx_rtt && i) 960 /* 961 * filter this update to half the old & half 962 * the new values, converting scale. 963 * See route.h and tcp_var.h for a 964 * description of the scaling constants. 965 */ 966 rt->rt_rmx.rmx_rtt = 967 (rt->rt_rmx.rmx_rtt + i) / 2; 968 else 969 rt->rt_rmx.rmx_rtt = i; 970 tcpstat.tcps_cachedrtt++; 971 } 972 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 973 i = tp->t_rttvar * 974 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 975 if (rt->rt_rmx.rmx_rttvar && i) 976 rt->rt_rmx.rmx_rttvar = 977 (rt->rt_rmx.rmx_rttvar + i) / 2; 978 else 979 rt->rt_rmx.rmx_rttvar = i; 980 tcpstat.tcps_cachedrttvar++; 981 } 982 /* 983 * The old comment here said: 984 * update the pipelimit (ssthresh) if it has been updated 985 * already or if a pipesize was specified & the threshhold 986 * got below half the pipesize. I.e., wait for bad news 987 * before we start updating, then update on both good 988 * and bad news. 989 * 990 * But we want to save the ssthresh even if no pipesize is 991 * specified explicitly in the route, because such 992 * connections still have an implicit pipesize specified 993 * by the global tcp_sendspace. In the absence of a reliable 994 * way to calculate the pipesize, it will have to do. 995 */ 996 i = tp->snd_ssthresh; 997 if (rt->rt_rmx.rmx_sendpipe != 0) 998 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 999 else 1000 dosavessthresh = (i < so->so_snd.ssb_hiwat/2); 1001 if (dosavessthresh || 1002 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 1003 (rt->rt_rmx.rmx_ssthresh != 0))) { 1004 /* 1005 * convert the limit from user data bytes to 1006 * packets then to packet data bytes. 1007 */ 1008 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 1009 if (i < 2) 1010 i = 2; 1011 i *= tp->t_maxseg + 1012 (isipv6 ? 1013 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1014 sizeof(struct tcpiphdr)); 1015 if (rt->rt_rmx.rmx_ssthresh) 1016 rt->rt_rmx.rmx_ssthresh = 1017 (rt->rt_rmx.rmx_ssthresh + i) / 2; 1018 else 1019 rt->rt_rmx.rmx_ssthresh = i; 1020 tcpstat.tcps_cachedssthresh++; 1021 } 1022 } 1023 1024 no_valid_rt: 1025 /* free the reassembly queue, if any */ 1026 while((q = TAILQ_FIRST(&tp->t_segq)) != NULL) { 1027 TAILQ_REMOVE(&tp->t_segq, q, tqe_q); 1028 m_freem(q->tqe_m); 1029 kfree(q, M_TSEGQ); 1030 atomic_add_int(&tcp_reass_qsize, -1); 1031 } 1032 /* throw away SACK blocks in scoreboard*/ 1033 if (TCP_DO_SACK(tp)) 1034 tcp_sack_destroy(&tp->scb); 1035 1036 inp->inp_ppcb = NULL; 1037 soisdisconnected(so); 1038 /* note: pcb detached later on */ 1039 1040 tcp_destroy_timermsg(tp); 1041 tcp_output_cancel(tp); 1042 1043 if (tp->t_flags & TF_LISTEN) { 1044 syncache_destroy(tp, tp_inh); 1045 tcp_pcbport_merge_oncpu(tp); 1046 tcp_pcbport_destroy(tp); 1047 if (inp_inh != NULL && inp_inh->inp_socket != NULL) { 1048 /* 1049 * Pending sockets inheritance only needs 1050 * to be done once in the current thread, 1051 * i.e. netisr0. 1052 */ 1053 soinherit(so, inp_inh->inp_socket); 1054 } 1055 } 1056 KASSERT(tp->t_pcbport == NULL, ("tcpcb port cache is not destroyed")); 1057 1058 so_async_rcvd_drop(so); 1059 /* Drop the reference for the asynchronized pru_rcvd */ 1060 sofree(so); 1061 1062 /* 1063 * NOTE: 1064 * - Remove self from listen tcpcb per-cpu port cache _before_ 1065 * pcbdetach. 1066 * - pcbdetach removes any wildcard hash entry on the current CPU. 1067 */ 1068 tcp_pcbport_remove(inp); 1069 #ifdef INET6 1070 if (isipv6) 1071 in6_pcbdetach(inp); 1072 else 1073 #endif 1074 in_pcbdetach(inp); 1075 1076 tcpstat.tcps_closed++; 1077 return (NULL); 1078 } 1079 1080 static __inline void 1081 tcp_drain_oncpu(struct inpcbinfo *pcbinfo) 1082 { 1083 struct inpcbhead *head = &pcbinfo->pcblisthead; 1084 struct inpcb *inpb; 1085 1086 /* 1087 * Since we run in netisr, it is MP safe, even if 1088 * we block during the inpcb list iteration, i.e. 1089 * we don't need to use inpcb marker here. 1090 */ 1091 ASSERT_IN_NETISR(pcbinfo->cpu); 1092 1093 LIST_FOREACH(inpb, head, inp_list) { 1094 struct tcpcb *tcpb; 1095 struct tseg_qent *te; 1096 1097 if (inpb->inp_flags & INP_PLACEMARKER) 1098 continue; 1099 1100 tcpb = intotcpcb(inpb); 1101 KASSERT(tcpb != NULL, ("tcp_drain_oncpu: tcpb is NULL")); 1102 1103 if ((te = TAILQ_FIRST(&tcpb->t_segq)) != NULL) { 1104 TAILQ_REMOVE(&tcpb->t_segq, te, tqe_q); 1105 if (te->tqe_th->th_flags & TH_FIN) 1106 tcpb->t_flags &= ~TF_QUEDFIN; 1107 m_freem(te->tqe_m); 1108 kfree(te, M_TSEGQ); 1109 atomic_add_int(&tcp_reass_qsize, -1); 1110 /* retry */ 1111 } 1112 } 1113 } 1114 1115 static void 1116 tcp_drain_dispatch(netmsg_t nmsg) 1117 { 1118 crit_enter(); 1119 lwkt_replymsg(&nmsg->lmsg, 0); /* reply ASAP */ 1120 crit_exit(); 1121 1122 tcp_drain_oncpu(&tcbinfo[mycpuid]); 1123 } 1124 1125 static void 1126 tcp_drain_ipi(void *arg __unused) 1127 { 1128 int cpu = mycpuid; 1129 struct lwkt_msg *msg = &tcp_drain_netmsg[cpu].lmsg; 1130 1131 crit_enter(); 1132 if (msg->ms_flags & MSGF_DONE) 1133 lwkt_sendmsg_oncpu(netisr_cpuport(cpu), msg); 1134 crit_exit(); 1135 } 1136 1137 void 1138 tcp_drain(void) 1139 { 1140 cpumask_t mask; 1141 1142 if (!do_tcpdrain) 1143 return; 1144 1145 /* 1146 * Walk the tcpbs, if existing, and flush the reassembly queue, 1147 * if there is one... 1148 * XXX: The "Net/3" implementation doesn't imply that the TCP 1149 * reassembly queue should be flushed, but in a situation 1150 * where we're really low on mbufs, this is potentially 1151 * useful. 1152 * YYY: We may consider run tcp_drain_oncpu directly here, 1153 * however, that will require M_WAITOK memory allocation 1154 * for the inpcb marker. 1155 */ 1156 CPUMASK_ASSBMASK(mask, ncpus2); 1157 CPUMASK_ANDMASK(mask, smp_active_mask); 1158 if (CPUMASK_TESTNZERO(mask)) 1159 lwkt_send_ipiq_mask(mask, tcp_drain_ipi, NULL); 1160 } 1161 1162 /* 1163 * Notify a tcp user of an asynchronous error; 1164 * store error as soft error, but wake up user 1165 * (for now, won't do anything until can select for soft error). 1166 * 1167 * Do not wake up user since there currently is no mechanism for 1168 * reporting soft errors (yet - a kqueue filter may be added). 1169 */ 1170 static void 1171 tcp_notify(struct inpcb *inp, int error) 1172 { 1173 struct tcpcb *tp = intotcpcb(inp); 1174 1175 /* 1176 * Ignore some errors if we are hooked up. 1177 * If connection hasn't completed, has retransmitted several times, 1178 * and receives a second error, give up now. This is better 1179 * than waiting a long time to establish a connection that 1180 * can never complete. 1181 */ 1182 if (tp->t_state == TCPS_ESTABLISHED && 1183 (error == EHOSTUNREACH || error == ENETUNREACH || 1184 error == EHOSTDOWN)) { 1185 return; 1186 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1187 tp->t_softerror) 1188 tcp_drop(tp, error); 1189 else 1190 tp->t_softerror = error; 1191 #if 0 1192 wakeup(&so->so_timeo); 1193 sorwakeup(so); 1194 sowwakeup(so); 1195 #endif 1196 } 1197 1198 static int 1199 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1200 { 1201 int error, i, n; 1202 struct inpcb *marker; 1203 struct inpcb *inp; 1204 int origcpu, ccpu; 1205 1206 error = 0; 1207 n = 0; 1208 1209 /* 1210 * The process of preparing the TCB list is too time-consuming and 1211 * resource-intensive to repeat twice on every request. 1212 */ 1213 if (req->oldptr == NULL) { 1214 for (ccpu = 0; ccpu < ncpus2; ++ccpu) 1215 n += tcbinfo[ccpu].ipi_count; 1216 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1217 return (0); 1218 } 1219 1220 if (req->newptr != NULL) 1221 return (EPERM); 1222 1223 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1224 marker->inp_flags |= INP_PLACEMARKER; 1225 1226 /* 1227 * OK, now we're committed to doing something. Run the inpcb list 1228 * for each cpu in the system and construct the output. Use a 1229 * list placemarker to deal with list changes occuring during 1230 * copyout blockages (but otherwise depend on being on the correct 1231 * cpu to avoid races). 1232 */ 1233 origcpu = mycpu->gd_cpuid; 1234 for (ccpu = 0; ccpu < ncpus2 && error == 0; ++ccpu) { 1235 caddr_t inp_ppcb; 1236 struct xtcpcb xt; 1237 1238 lwkt_migratecpu(ccpu); 1239 1240 n = tcbinfo[ccpu].ipi_count; 1241 1242 LIST_INSERT_HEAD(&tcbinfo[ccpu].pcblisthead, marker, inp_list); 1243 i = 0; 1244 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1245 /* 1246 * process a snapshot of pcbs, ignoring placemarkers 1247 * and using our own to allow SYSCTL_OUT to block. 1248 */ 1249 LIST_REMOVE(marker, inp_list); 1250 LIST_INSERT_AFTER(inp, marker, inp_list); 1251 1252 if (inp->inp_flags & INP_PLACEMARKER) 1253 continue; 1254 if (prison_xinpcb(req->td, inp)) 1255 continue; 1256 1257 xt.xt_len = sizeof xt; 1258 bcopy(inp, &xt.xt_inp, sizeof *inp); 1259 inp_ppcb = inp->inp_ppcb; 1260 if (inp_ppcb != NULL) 1261 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1262 else 1263 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1264 if (inp->inp_socket) 1265 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1266 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1267 break; 1268 ++i; 1269 } 1270 LIST_REMOVE(marker, inp_list); 1271 if (error == 0 && i < n) { 1272 bzero(&xt, sizeof xt); 1273 xt.xt_len = sizeof xt; 1274 while (i < n) { 1275 error = SYSCTL_OUT(req, &xt, sizeof xt); 1276 if (error) 1277 break; 1278 ++i; 1279 } 1280 } 1281 } 1282 1283 /* 1284 * Make sure we are on the same cpu we were on originally, since 1285 * higher level callers expect this. Also don't pollute caches with 1286 * migrated userland data by (eventually) returning to userland 1287 * on a different cpu. 1288 */ 1289 lwkt_migratecpu(origcpu); 1290 kfree(marker, M_TEMP); 1291 return (error); 1292 } 1293 1294 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1295 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1296 1297 static int 1298 tcp_getcred(SYSCTL_HANDLER_ARGS) 1299 { 1300 struct sockaddr_in addrs[2]; 1301 struct ucred cred0, *cred = NULL; 1302 struct inpcb *inp; 1303 int cpu, origcpu, error; 1304 1305 error = priv_check(req->td, PRIV_ROOT); 1306 if (error != 0) 1307 return (error); 1308 error = SYSCTL_IN(req, addrs, sizeof addrs); 1309 if (error != 0) 1310 return (error); 1311 1312 origcpu = mycpuid; 1313 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1314 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1315 1316 lwkt_migratecpu(cpu); 1317 1318 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1319 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1320 if (inp == NULL || inp->inp_socket == NULL) { 1321 error = ENOENT; 1322 } else if (inp->inp_socket->so_cred != NULL) { 1323 cred0 = *(inp->inp_socket->so_cred); 1324 cred = &cred0; 1325 } 1326 1327 lwkt_migratecpu(origcpu); 1328 1329 if (error) 1330 return (error); 1331 1332 return SYSCTL_OUT(req, cred, sizeof(struct ucred)); 1333 } 1334 1335 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1336 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1337 1338 #ifdef INET6 1339 static int 1340 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1341 { 1342 struct sockaddr_in6 addrs[2]; 1343 struct inpcb *inp; 1344 int error; 1345 1346 error = priv_check(req->td, PRIV_ROOT); 1347 if (error != 0) 1348 return (error); 1349 error = SYSCTL_IN(req, addrs, sizeof addrs); 1350 if (error != 0) 1351 return (error); 1352 crit_enter(); 1353 inp = in6_pcblookup_hash(&tcbinfo[0], 1354 &addrs[1].sin6_addr, addrs[1].sin6_port, 1355 &addrs[0].sin6_addr, addrs[0].sin6_port, 0, NULL); 1356 if (inp == NULL || inp->inp_socket == NULL) { 1357 error = ENOENT; 1358 goto out; 1359 } 1360 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1361 out: 1362 crit_exit(); 1363 return (error); 1364 } 1365 1366 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1367 0, 0, 1368 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1369 #endif 1370 1371 struct netmsg_tcp_notify { 1372 struct netmsg_base base; 1373 inp_notify_t nm_notify; 1374 struct in_addr nm_faddr; 1375 int nm_arg; 1376 }; 1377 1378 static void 1379 tcp_notifyall_oncpu(netmsg_t msg) 1380 { 1381 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg; 1382 int nextcpu; 1383 1384 in_pcbnotifyall(&tcbinfo[mycpuid], nm->nm_faddr, 1385 nm->nm_arg, nm->nm_notify); 1386 1387 nextcpu = mycpuid + 1; 1388 if (nextcpu < ncpus2) 1389 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nm->base.lmsg); 1390 else 1391 lwkt_replymsg(&nm->base.lmsg, 0); 1392 } 1393 1394 inp_notify_t 1395 tcp_get_inpnotify(int cmd, const struct sockaddr *sa, 1396 int *arg, struct ip **ip0, int *cpuid) 1397 { 1398 struct ip *ip = *ip0; 1399 struct in_addr faddr; 1400 inp_notify_t notify = tcp_notify; 1401 1402 faddr = ((const struct sockaddr_in *)sa)->sin_addr; 1403 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1404 return NULL; 1405 1406 *arg = inetctlerrmap[cmd]; 1407 if (cmd == PRC_QUENCH) { 1408 notify = tcp_quench; 1409 } else if (icmp_may_rst && 1410 (cmd == PRC_UNREACH_ADMIN_PROHIB || 1411 cmd == PRC_UNREACH_PORT || 1412 cmd == PRC_TIMXCEED_INTRANS) && 1413 ip != NULL) { 1414 notify = tcp_drop_syn_sent; 1415 } else if (cmd == PRC_MSGSIZE) { 1416 const struct icmp *icmp = (const struct icmp *) 1417 ((caddr_t)ip - offsetof(struct icmp, icmp_ip)); 1418 1419 *arg = ntohs(icmp->icmp_nextmtu); 1420 notify = tcp_mtudisc; 1421 } else if (PRC_IS_REDIRECT(cmd)) { 1422 ip = NULL; 1423 notify = in_rtchange; 1424 } else if (cmd == PRC_HOSTDEAD) { 1425 ip = NULL; 1426 } else if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) { 1427 return NULL; 1428 } 1429 1430 if (cpuid != NULL) { 1431 if (ip == NULL) { 1432 /* Go through all CPUs */ 1433 *cpuid = ncpus; 1434 } else { 1435 const struct tcphdr *th; 1436 1437 th = (const struct tcphdr *) 1438 ((caddr_t)ip + (IP_VHL_HL(ip->ip_vhl) << 2)); 1439 *cpuid = tcp_addrcpu(faddr.s_addr, th->th_dport, 1440 ip->ip_src.s_addr, th->th_sport); 1441 } 1442 } 1443 1444 *ip0 = ip; 1445 return notify; 1446 } 1447 1448 void 1449 tcp_ctlinput(netmsg_t msg) 1450 { 1451 int cmd = msg->ctlinput.nm_cmd; 1452 struct sockaddr *sa = msg->ctlinput.nm_arg; 1453 struct ip *ip = msg->ctlinput.nm_extra; 1454 struct in_addr faddr; 1455 inp_notify_t notify; 1456 int arg, cpuid; 1457 1458 notify = tcp_get_inpnotify(cmd, sa, &arg, &ip, &cpuid); 1459 if (notify == NULL) 1460 goto done; 1461 1462 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1463 if (ip != NULL) { 1464 const struct tcphdr *th; 1465 struct inpcb *inp; 1466 1467 if (cpuid != mycpuid) 1468 goto done; 1469 1470 th = (const struct tcphdr *) 1471 ((caddr_t)ip + (IP_VHL_HL(ip->ip_vhl) << 2)); 1472 inp = in_pcblookup_hash(&tcbinfo[mycpuid], faddr, th->th_dport, 1473 ip->ip_src, th->th_sport, 0, NULL); 1474 if (inp != NULL && inp->inp_socket != NULL) { 1475 tcp_seq icmpseq = htonl(th->th_seq); 1476 struct tcpcb *tp = intotcpcb(inp); 1477 1478 if (SEQ_GEQ(icmpseq, tp->snd_una) && 1479 SEQ_LT(icmpseq, tp->snd_max)) 1480 notify(inp, arg); 1481 } else { 1482 struct in_conninfo inc; 1483 1484 inc.inc_fport = th->th_dport; 1485 inc.inc_lport = th->th_sport; 1486 inc.inc_faddr = faddr; 1487 inc.inc_laddr = ip->ip_src; 1488 #ifdef INET6 1489 inc.inc_isipv6 = 0; 1490 #endif 1491 syncache_unreach(&inc, th); 1492 } 1493 } else if (msg->ctlinput.nm_direct) { 1494 if (cpuid != ncpus && cpuid != mycpuid) 1495 goto done; 1496 if (mycpuid >= ncpus2) 1497 goto done; 1498 1499 in_pcbnotifyall(&tcbinfo[mycpuid], faddr, arg, notify); 1500 } else { 1501 struct netmsg_tcp_notify *nm; 1502 1503 ASSERT_IN_NETISR(0); 1504 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT); 1505 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1506 0, tcp_notifyall_oncpu); 1507 nm->nm_faddr = faddr; 1508 nm->nm_arg = arg; 1509 nm->nm_notify = notify; 1510 1511 lwkt_sendmsg(netisr_cpuport(0), &nm->base.lmsg); 1512 } 1513 done: 1514 lwkt_replymsg(&msg->lmsg, 0); 1515 } 1516 1517 #ifdef INET6 1518 1519 void 1520 tcp6_ctlinput(netmsg_t msg) 1521 { 1522 int cmd = msg->ctlinput.nm_cmd; 1523 struct sockaddr *sa = msg->ctlinput.nm_arg; 1524 void *d = msg->ctlinput.nm_extra; 1525 struct tcphdr th; 1526 inp_notify_t notify = tcp_notify; 1527 struct ip6_hdr *ip6; 1528 struct mbuf *m; 1529 struct ip6ctlparam *ip6cp = NULL; 1530 const struct sockaddr_in6 *sa6_src = NULL; 1531 int off; 1532 struct tcp_portonly { 1533 u_int16_t th_sport; 1534 u_int16_t th_dport; 1535 } *thp; 1536 int arg; 1537 1538 if (sa->sa_family != AF_INET6 || 1539 sa->sa_len != sizeof(struct sockaddr_in6)) { 1540 goto out; 1541 } 1542 1543 arg = 0; 1544 if (cmd == PRC_QUENCH) 1545 notify = tcp_quench; 1546 else if (cmd == PRC_MSGSIZE) { 1547 struct ip6ctlparam *ip6cp = d; 1548 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6; 1549 1550 arg = ntohl(icmp6->icmp6_mtu); 1551 notify = tcp_mtudisc; 1552 } else if (!PRC_IS_REDIRECT(cmd) && 1553 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) { 1554 goto out; 1555 } 1556 1557 /* if the parameter is from icmp6, decode it. */ 1558 if (d != NULL) { 1559 ip6cp = (struct ip6ctlparam *)d; 1560 m = ip6cp->ip6c_m; 1561 ip6 = ip6cp->ip6c_ip6; 1562 off = ip6cp->ip6c_off; 1563 sa6_src = ip6cp->ip6c_src; 1564 } else { 1565 m = NULL; 1566 ip6 = NULL; 1567 off = 0; /* fool gcc */ 1568 sa6_src = &sa6_any; 1569 } 1570 1571 if (ip6 != NULL) { 1572 struct in_conninfo inc; 1573 /* 1574 * XXX: We assume that when IPV6 is non NULL, 1575 * M and OFF are valid. 1576 */ 1577 1578 /* check if we can safely examine src and dst ports */ 1579 if (m->m_pkthdr.len < off + sizeof *thp) 1580 goto out; 1581 1582 bzero(&th, sizeof th); 1583 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1584 1585 in6_pcbnotify(&tcbinfo[0], sa, th.th_dport, 1586 (struct sockaddr *)ip6cp->ip6c_src, 1587 th.th_sport, cmd, arg, notify); 1588 1589 inc.inc_fport = th.th_dport; 1590 inc.inc_lport = th.th_sport; 1591 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1592 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1593 inc.inc_isipv6 = 1; 1594 syncache_unreach(&inc, &th); 1595 } else { 1596 in6_pcbnotify(&tcbinfo[0], sa, 0, 1597 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify); 1598 } 1599 out: 1600 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0); 1601 } 1602 1603 #endif 1604 1605 /* 1606 * Following is where TCP initial sequence number generation occurs. 1607 * 1608 * There are two places where we must use initial sequence numbers: 1609 * 1. In SYN-ACK packets. 1610 * 2. In SYN packets. 1611 * 1612 * All ISNs for SYN-ACK packets are generated by the syncache. See 1613 * tcp_syncache.c for details. 1614 * 1615 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1616 * depends on this property. In addition, these ISNs should be 1617 * unguessable so as to prevent connection hijacking. To satisfy 1618 * the requirements of this situation, the algorithm outlined in 1619 * RFC 1948 is used to generate sequence numbers. 1620 * 1621 * Implementation details: 1622 * 1623 * Time is based off the system timer, and is corrected so that it 1624 * increases by one megabyte per second. This allows for proper 1625 * recycling on high speed LANs while still leaving over an hour 1626 * before rollover. 1627 * 1628 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1629 * between seeding of isn_secret. This is normally set to zero, 1630 * as reseeding should not be necessary. 1631 * 1632 */ 1633 1634 #define ISN_BYTES_PER_SECOND 1048576 1635 1636 u_char isn_secret[32]; 1637 int isn_last_reseed; 1638 MD5_CTX isn_ctx; 1639 1640 tcp_seq 1641 tcp_new_isn(struct tcpcb *tp) 1642 { 1643 u_int32_t md5_buffer[4]; 1644 tcp_seq new_isn; 1645 1646 /* Seed if this is the first use, reseed if requested. */ 1647 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) && 1648 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz) 1649 < (u_int)ticks))) { 1650 read_random_unlimited(&isn_secret, sizeof isn_secret); 1651 isn_last_reseed = ticks; 1652 } 1653 1654 /* Compute the md5 hash and return the ISN. */ 1655 MD5Init(&isn_ctx); 1656 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short)); 1657 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short)); 1658 #ifdef INET6 1659 if (INP_ISIPV6(tp->t_inpcb)) { 1660 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr, 1661 sizeof(struct in6_addr)); 1662 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr, 1663 sizeof(struct in6_addr)); 1664 } else 1665 #endif 1666 { 1667 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr, 1668 sizeof(struct in_addr)); 1669 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr, 1670 sizeof(struct in_addr)); 1671 } 1672 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret)); 1673 MD5Final((u_char *) &md5_buffer, &isn_ctx); 1674 new_isn = (tcp_seq) md5_buffer[0]; 1675 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz); 1676 return (new_isn); 1677 } 1678 1679 /* 1680 * When a source quench is received, close congestion window 1681 * to one segment. We will gradually open it again as we proceed. 1682 */ 1683 void 1684 tcp_quench(struct inpcb *inp, int error) 1685 { 1686 struct tcpcb *tp = intotcpcb(inp); 1687 1688 KASSERT(tp != NULL, ("tcp_quench: tp is NULL")); 1689 tp->snd_cwnd = tp->t_maxseg; 1690 tp->snd_wacked = 0; 1691 } 1692 1693 /* 1694 * When a specific ICMP unreachable message is received and the 1695 * connection state is SYN-SENT, drop the connection. This behavior 1696 * is controlled by the icmp_may_rst sysctl. 1697 */ 1698 void 1699 tcp_drop_syn_sent(struct inpcb *inp, int error) 1700 { 1701 struct tcpcb *tp = intotcpcb(inp); 1702 1703 KASSERT(tp != NULL, ("tcp_drop_syn_sent: tp is NULL")); 1704 if (tp->t_state == TCPS_SYN_SENT) 1705 tcp_drop(tp, error); 1706 } 1707 1708 /* 1709 * When a `need fragmentation' ICMP is received, update our idea of the MSS 1710 * based on the new value in the route. Also nudge TCP to send something, 1711 * since we know the packet we just sent was dropped. 1712 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1713 */ 1714 void 1715 tcp_mtudisc(struct inpcb *inp, int mtu) 1716 { 1717 struct tcpcb *tp = intotcpcb(inp); 1718 struct rtentry *rt; 1719 struct socket *so = inp->inp_socket; 1720 int maxopd, mss; 1721 #ifdef INET6 1722 boolean_t isipv6 = INP_ISIPV6(inp); 1723 #else 1724 const boolean_t isipv6 = FALSE; 1725 #endif 1726 1727 KASSERT(tp != NULL, ("tcp_mtudisc: tp is NULL")); 1728 1729 /* 1730 * If no MTU is provided in the ICMP message, use the 1731 * next lower likely value, as specified in RFC 1191. 1732 */ 1733 if (mtu == 0) { 1734 int oldmtu; 1735 1736 oldmtu = tp->t_maxopd + 1737 (isipv6 ? 1738 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1739 sizeof(struct tcpiphdr)); 1740 mtu = ip_next_mtu(oldmtu, 0); 1741 } 1742 1743 if (isipv6) 1744 rt = tcp_rtlookup6(&inp->inp_inc); 1745 else 1746 rt = tcp_rtlookup(&inp->inp_inc); 1747 if (rt != NULL) { 1748 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu) 1749 mtu = rt->rt_rmx.rmx_mtu; 1750 1751 maxopd = mtu - 1752 (isipv6 ? 1753 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1754 sizeof(struct tcpiphdr)); 1755 1756 /* 1757 * XXX - The following conditional probably violates the TCP 1758 * spec. The problem is that, since we don't know the 1759 * other end's MSS, we are supposed to use a conservative 1760 * default. But, if we do that, then MTU discovery will 1761 * never actually take place, because the conservative 1762 * default is much less than the MTUs typically seen 1763 * on the Internet today. For the moment, we'll sweep 1764 * this under the carpet. 1765 * 1766 * The conservative default might not actually be a problem 1767 * if the only case this occurs is when sending an initial 1768 * SYN with options and data to a host we've never talked 1769 * to before. Then, they will reply with an MSS value which 1770 * will get recorded and the new parameters should get 1771 * recomputed. For Further Study. 1772 */ 1773 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd) 1774 maxopd = rt->rt_rmx.rmx_mssopt; 1775 } else 1776 maxopd = mtu - 1777 (isipv6 ? 1778 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1779 sizeof(struct tcpiphdr)); 1780 1781 if (tp->t_maxopd <= maxopd) 1782 return; 1783 tp->t_maxopd = maxopd; 1784 1785 mss = maxopd; 1786 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) == 1787 (TF_REQ_TSTMP | TF_RCVD_TSTMP)) 1788 mss -= TCPOLEN_TSTAMP_APPA; 1789 1790 /* round down to multiple of MCLBYTES */ 1791 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */ 1792 if (mss > MCLBYTES) 1793 mss &= ~(MCLBYTES - 1); 1794 #else 1795 if (mss > MCLBYTES) 1796 mss = (mss / MCLBYTES) * MCLBYTES; 1797 #endif 1798 1799 if (so->so_snd.ssb_hiwat < mss) 1800 mss = so->so_snd.ssb_hiwat; 1801 1802 tp->t_maxseg = mss; 1803 tp->t_rtttime = 0; 1804 tp->snd_nxt = tp->snd_una; 1805 tcp_output(tp); 1806 tcpstat.tcps_mturesent++; 1807 } 1808 1809 /* 1810 * Look-up the routing entry to the peer of this inpcb. If no route 1811 * is found and it cannot be allocated the return NULL. This routine 1812 * is called by TCP routines that access the rmx structure and by tcp_mss 1813 * to get the interface MTU. 1814 */ 1815 struct rtentry * 1816 tcp_rtlookup(struct in_conninfo *inc) 1817 { 1818 struct route *ro = &inc->inc_route; 1819 1820 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1821 /* No route yet, so try to acquire one */ 1822 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1823 /* 1824 * unused portions of the structure MUST be zero'd 1825 * out because rtalloc() treats it as opaque data 1826 */ 1827 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1828 ro->ro_dst.sa_family = AF_INET; 1829 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1830 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1831 inc->inc_faddr; 1832 rtalloc(ro); 1833 } 1834 } 1835 return (ro->ro_rt); 1836 } 1837 1838 #ifdef INET6 1839 struct rtentry * 1840 tcp_rtlookup6(struct in_conninfo *inc) 1841 { 1842 struct route_in6 *ro6 = &inc->inc6_route; 1843 1844 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1845 /* No route yet, so try to acquire one */ 1846 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1847 /* 1848 * unused portions of the structure MUST be zero'd 1849 * out because rtalloc() treats it as opaque data 1850 */ 1851 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1852 ro6->ro_dst.sin6_family = AF_INET6; 1853 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1854 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1855 rtalloc((struct route *)ro6); 1856 } 1857 } 1858 return (ro6->ro_rt); 1859 } 1860 #endif 1861 1862 #ifdef IPSEC 1863 /* compute ESP/AH header size for TCP, including outer IP header. */ 1864 size_t 1865 ipsec_hdrsiz_tcp(struct tcpcb *tp) 1866 { 1867 struct inpcb *inp; 1868 struct mbuf *m; 1869 size_t hdrsiz; 1870 struct ip *ip; 1871 struct tcphdr *th; 1872 1873 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL)) 1874 return (0); 1875 MGETHDR(m, M_NOWAIT, MT_DATA); 1876 if (!m) 1877 return (0); 1878 1879 #ifdef INET6 1880 if (INP_ISIPV6(inp)) { 1881 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); 1882 1883 th = (struct tcphdr *)(ip6 + 1); 1884 m->m_pkthdr.len = m->m_len = 1885 sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1886 tcp_fillheaders(tp, ip6, th, FALSE); 1887 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1888 } else 1889 #endif 1890 { 1891 ip = mtod(m, struct ip *); 1892 th = (struct tcphdr *)(ip + 1); 1893 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr); 1894 tcp_fillheaders(tp, ip, th, FALSE); 1895 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp); 1896 } 1897 1898 m_free(m); 1899 return (hdrsiz); 1900 } 1901 #endif 1902 1903 /* 1904 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1905 * 1906 * This code attempts to calculate the bandwidth-delay product as a 1907 * means of determining the optimal window size to maximize bandwidth, 1908 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1909 * routers. This code also does a fairly good job keeping RTTs in check 1910 * across slow links like modems. We implement an algorithm which is very 1911 * similar (but not meant to be) TCP/Vegas. The code operates on the 1912 * transmitter side of a TCP connection and so only effects the transmit 1913 * side of the connection. 1914 * 1915 * BACKGROUND: TCP makes no provision for the management of buffer space 1916 * at the end points or at the intermediate routers and switches. A TCP 1917 * stream, whether using NewReno or not, will eventually buffer as 1918 * many packets as it is able and the only reason this typically works is 1919 * due to the fairly small default buffers made available for a connection 1920 * (typicaly 16K or 32K). As machines use larger windows and/or window 1921 * scaling it is now fairly easy for even a single TCP connection to blow-out 1922 * all available buffer space not only on the local interface, but on 1923 * intermediate routers and switches as well. NewReno makes a misguided 1924 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1925 * then backing off, then steadily increasing the window again until another 1926 * failure occurs, ad-infinitum. This results in terrible oscillation that 1927 * is only made worse as network loads increase and the idea of intentionally 1928 * blowing out network buffers is, frankly, a terrible way to manage network 1929 * resources. 1930 * 1931 * It is far better to limit the transmit window prior to the failure 1932 * condition being achieved. There are two general ways to do this: First 1933 * you can 'scan' through different transmit window sizes and locate the 1934 * point where the RTT stops increasing, indicating that you have filled the 1935 * pipe, then scan backwards until you note that RTT stops decreasing, then 1936 * repeat ad-infinitum. This method works in principle but has severe 1937 * implementation issues due to RTT variances, timer granularity, and 1938 * instability in the algorithm which can lead to many false positives and 1939 * create oscillations as well as interact badly with other TCP streams 1940 * implementing the same algorithm. 1941 * 1942 * The second method is to limit the window to the bandwidth delay product 1943 * of the link. This is the method we implement. RTT variances and our 1944 * own manipulation of the congestion window, bwnd, can potentially 1945 * destabilize the algorithm. For this reason we have to stabilize the 1946 * elements used to calculate the window. We do this by using the minimum 1947 * observed RTT, the long term average of the observed bandwidth, and 1948 * by adding two segments worth of slop. It isn't perfect but it is able 1949 * to react to changing conditions and gives us a very stable basis on 1950 * which to extend the algorithm. 1951 */ 1952 void 1953 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1954 { 1955 u_long bw; 1956 u_long ibw; 1957 u_long bwnd; 1958 int save_ticks; 1959 int delta_ticks; 1960 1961 /* 1962 * If inflight_enable is disabled in the middle of a tcp connection, 1963 * make sure snd_bwnd is effectively disabled. 1964 */ 1965 if (!tcp_inflight_enable) { 1966 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1967 tp->snd_bandwidth = 0; 1968 return; 1969 } 1970 1971 /* 1972 * Validate the delta time. If a connection is new or has been idle 1973 * a long time we have to reset the bandwidth calculator. 1974 */ 1975 save_ticks = ticks; 1976 cpu_ccfence(); 1977 delta_ticks = save_ticks - tp->t_bw_rtttime; 1978 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1979 tp->t_bw_rtttime = save_ticks; 1980 tp->t_bw_rtseq = ack_seq; 1981 if (tp->snd_bandwidth == 0) 1982 tp->snd_bandwidth = tcp_inflight_start; 1983 return; 1984 } 1985 1986 /* 1987 * A delta of at least 1 tick is required. Waiting 2 ticks will 1988 * result in better (bw) accuracy. More than that and the ramp-up 1989 * will be too slow. 1990 */ 1991 if (delta_ticks == 0 || delta_ticks == 1) 1992 return; 1993 1994 /* 1995 * Sanity check, plus ignore pure window update acks. 1996 */ 1997 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 1998 return; 1999 2000 /* 2001 * Figure out the bandwidth. Due to the tick granularity this 2002 * is a very rough number and it MUST be averaged over a fairly 2003 * long period of time. XXX we need to take into account a link 2004 * that is not using all available bandwidth, but for now our 2005 * slop will ramp us up if this case occurs and the bandwidth later 2006 * increases. 2007 */ 2008 ibw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 2009 tp->t_bw_rtttime = save_ticks; 2010 tp->t_bw_rtseq = ack_seq; 2011 bw = ((int64_t)tp->snd_bandwidth * 15 + ibw) >> 4; 2012 2013 tp->snd_bandwidth = bw; 2014 2015 /* 2016 * Calculate the semi-static bandwidth delay product, plus two maximal 2017 * segments. The additional slop puts us squarely in the sweet 2018 * spot and also handles the bandwidth run-up case. Without the 2019 * slop we could be locking ourselves into a lower bandwidth. 2020 * 2021 * At very high speeds the bw calculation can become overly sensitive 2022 * and error prone when delta_ticks is low (e.g. usually 1). To deal 2023 * with the problem the stab must be scaled to the bw. A stab of 50 2024 * (the default) increases the bw for the purposes of the bwnd 2025 * calculation by 5%. 2026 * 2027 * Situations Handled: 2028 * (1) Prevents over-queueing of packets on LANs, especially on 2029 * high speed LANs, allowing larger TCP buffers to be 2030 * specified, and also does a good job preventing 2031 * over-queueing of packets over choke points like modems 2032 * (at least for the transmit side). 2033 * 2034 * (2) Is able to handle changing network loads (bandwidth 2035 * drops so bwnd drops, bandwidth increases so bwnd 2036 * increases). 2037 * 2038 * (3) Theoretically should stabilize in the face of multiple 2039 * connections implementing the same algorithm (this may need 2040 * a little work). 2041 * 2042 * (4) Stability value (defaults to 20 = 2 maximal packets) can 2043 * be adjusted with a sysctl but typically only needs to be on 2044 * very slow connections. A value no smaller then 5 should 2045 * be used, but only reduce this default if you have no other 2046 * choice. 2047 */ 2048 2049 #define USERTT ((tp->t_srtt + tp->t_rttvar) + tcp_inflight_adjrtt) 2050 bw += bw * tcp_inflight_stab / 1000; 2051 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 2052 (int)tp->t_maxseg * 2; 2053 #undef USERTT 2054 2055 if (tcp_inflight_debug > 0) { 2056 static int ltime; 2057 if ((u_int)(save_ticks - ltime) >= hz / tcp_inflight_debug) { 2058 ltime = save_ticks; 2059 kprintf("%p ibw %ld bw %ld rttvar %d srtt %d " 2060 "bwnd %ld delta %d snd_win %ld\n", 2061 tp, ibw, bw, tp->t_rttvar, tp->t_srtt, 2062 bwnd, delta_ticks, tp->snd_wnd); 2063 } 2064 } 2065 if ((long)bwnd < tcp_inflight_min) 2066 bwnd = tcp_inflight_min; 2067 if (bwnd > tcp_inflight_max) 2068 bwnd = tcp_inflight_max; 2069 if ((long)bwnd < tp->t_maxseg * 2) 2070 bwnd = tp->t_maxseg * 2; 2071 tp->snd_bwnd = bwnd; 2072 } 2073 2074 static void 2075 tcp_rmx_iwsegs(struct tcpcb *tp, u_long *maxsegs, u_long *capsegs) 2076 { 2077 struct rtentry *rt; 2078 struct inpcb *inp = tp->t_inpcb; 2079 #ifdef INET6 2080 boolean_t isipv6 = INP_ISIPV6(inp); 2081 #else 2082 const boolean_t isipv6 = FALSE; 2083 #endif 2084 2085 /* XXX */ 2086 if (tcp_iw_maxsegs < TCP_IW_MAXSEGS_DFLT) 2087 tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 2088 if (tcp_iw_capsegs < TCP_IW_CAPSEGS_DFLT) 2089 tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 2090 2091 if (isipv6) 2092 rt = tcp_rtlookup6(&inp->inp_inc); 2093 else 2094 rt = tcp_rtlookup(&inp->inp_inc); 2095 if (rt == NULL || 2096 rt->rt_rmx.rmx_iwmaxsegs < TCP_IW_MAXSEGS_DFLT || 2097 rt->rt_rmx.rmx_iwcapsegs < TCP_IW_CAPSEGS_DFLT) { 2098 *maxsegs = tcp_iw_maxsegs; 2099 *capsegs = tcp_iw_capsegs; 2100 return; 2101 } 2102 *maxsegs = rt->rt_rmx.rmx_iwmaxsegs; 2103 *capsegs = rt->rt_rmx.rmx_iwcapsegs; 2104 } 2105 2106 u_long 2107 tcp_initial_window(struct tcpcb *tp) 2108 { 2109 if (tcp_do_rfc3390) { 2110 /* 2111 * RFC3390: 2112 * "If the SYN or SYN/ACK is lost, the initial window 2113 * used by a sender after a correctly transmitted SYN 2114 * MUST be one segment consisting of MSS bytes." 2115 * 2116 * However, we do something a little bit more aggressive 2117 * then RFC3390 here: 2118 * - Only if time spent in the SYN or SYN|ACK retransmition 2119 * >= 3 seconds, the IW is reduced. We do this mainly 2120 * because when RFC3390 is published, the initial RTO is 2121 * still 3 seconds (the threshold we test here), while 2122 * after RFC6298, the initial RTO is 1 second. This 2123 * behaviour probably still falls within the spirit of 2124 * RFC3390. 2125 * - When IW is reduced, 2*MSS is used instead of 1*MSS. 2126 * Mainly to avoid sender and receiver deadlock until 2127 * delayed ACK timer expires. And even RFC2581 does not 2128 * try to reduce IW upon SYN or SYN|ACK retransmition 2129 * timeout. 2130 * 2131 * See also: 2132 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03 2133 */ 2134 if (tp->t_rxtsyn >= TCPTV_RTOBASE3) { 2135 return (2 * tp->t_maxseg); 2136 } else { 2137 u_long maxsegs, capsegs; 2138 2139 tcp_rmx_iwsegs(tp, &maxsegs, &capsegs); 2140 return min(maxsegs * tp->t_maxseg, 2141 max(2 * tp->t_maxseg, capsegs * 1460)); 2142 } 2143 } else { 2144 /* 2145 * Even RFC2581 (back to 1999) allows 2*SMSS IW. 2146 * 2147 * Mainly to avoid sender and receiver deadlock 2148 * until delayed ACK timer expires. 2149 */ 2150 return (2 * tp->t_maxseg); 2151 } 2152 } 2153 2154 #ifdef TCP_SIGNATURE 2155 /* 2156 * Compute TCP-MD5 hash of a TCP segment. (RFC2385) 2157 * 2158 * We do this over ip, tcphdr, segment data, and the key in the SADB. 2159 * When called from tcp_input(), we can be sure that th_sum has been 2160 * zeroed out and verified already. 2161 * 2162 * Return 0 if successful, otherwise return -1. 2163 * 2164 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a 2165 * search with the destination IP address, and a 'magic SPI' to be 2166 * determined by the application. This is hardcoded elsewhere to 1179 2167 * right now. Another branch of this code exists which uses the SPD to 2168 * specify per-application flows but it is unstable. 2169 */ 2170 int 2171 tcpsignature_compute( 2172 struct mbuf *m, /* mbuf chain */ 2173 int len, /* length of TCP data */ 2174 int optlen, /* length of TCP options */ 2175 u_char *buf, /* storage for MD5 digest */ 2176 u_int direction) /* direction of flow */ 2177 { 2178 struct ippseudo ippseudo; 2179 MD5_CTX ctx; 2180 int doff; 2181 struct ip *ip; 2182 struct ipovly *ipovly; 2183 struct secasvar *sav; 2184 struct tcphdr *th; 2185 #ifdef INET6 2186 struct ip6_hdr *ip6; 2187 struct in6_addr in6; 2188 uint32_t plen; 2189 uint16_t nhdr; 2190 #endif /* INET6 */ 2191 u_short savecsum; 2192 2193 KASSERT(m != NULL, ("passed NULL mbuf. Game over.")); 2194 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature")); 2195 /* 2196 * Extract the destination from the IP header in the mbuf. 2197 */ 2198 ip = mtod(m, struct ip *); 2199 #ifdef INET6 2200 ip6 = NULL; /* Make the compiler happy. */ 2201 #endif /* INET6 */ 2202 /* 2203 * Look up an SADB entry which matches the address found in 2204 * the segment. 2205 */ 2206 switch (IP_VHL_V(ip->ip_vhl)) { 2207 case IPVERSION: 2208 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst, 2209 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2210 break; 2211 #ifdef INET6 2212 case (IPV6_VERSION >> 4): 2213 ip6 = mtod(m, struct ip6_hdr *); 2214 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst, 2215 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2216 break; 2217 #endif /* INET6 */ 2218 default: 2219 return (EINVAL); 2220 /* NOTREACHED */ 2221 break; 2222 } 2223 if (sav == NULL) { 2224 kprintf("%s: SADB lookup failed\n", __func__); 2225 return (EINVAL); 2226 } 2227 MD5Init(&ctx); 2228 2229 /* 2230 * Step 1: Update MD5 hash with IP pseudo-header. 2231 * 2232 * XXX The ippseudo header MUST be digested in network byte order, 2233 * or else we'll fail the regression test. Assume all fields we've 2234 * been doing arithmetic on have been in host byte order. 2235 * XXX One cannot depend on ipovly->ih_len here. When called from 2236 * tcp_output(), the underlying ip_len member has not yet been set. 2237 */ 2238 switch (IP_VHL_V(ip->ip_vhl)) { 2239 case IPVERSION: 2240 ipovly = (struct ipovly *)ip; 2241 ippseudo.ippseudo_src = ipovly->ih_src; 2242 ippseudo.ippseudo_dst = ipovly->ih_dst; 2243 ippseudo.ippseudo_pad = 0; 2244 ippseudo.ippseudo_p = IPPROTO_TCP; 2245 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen); 2246 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo)); 2247 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip)); 2248 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen; 2249 break; 2250 #ifdef INET6 2251 /* 2252 * RFC 2385, 2.0 Proposal 2253 * For IPv6, the pseudo-header is as described in RFC 2460, namely the 2254 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero- 2255 * extended next header value (to form 32 bits), and 32-bit segment 2256 * length. 2257 * Note: Upper-Layer Packet Length comes before Next Header. 2258 */ 2259 case (IPV6_VERSION >> 4): 2260 in6 = ip6->ip6_src; 2261 in6_clearscope(&in6); 2262 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2263 in6 = ip6->ip6_dst; 2264 in6_clearscope(&in6); 2265 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2266 plen = htonl(len + sizeof(struct tcphdr) + optlen); 2267 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t)); 2268 nhdr = 0; 2269 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2270 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2271 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2272 nhdr = IPPROTO_TCP; 2273 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2274 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr)); 2275 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen; 2276 break; 2277 #endif /* INET6 */ 2278 default: 2279 return (EINVAL); 2280 /* NOTREACHED */ 2281 break; 2282 } 2283 /* 2284 * Step 2: Update MD5 hash with TCP header, excluding options. 2285 * The TCP checksum must be set to zero. 2286 */ 2287 savecsum = th->th_sum; 2288 th->th_sum = 0; 2289 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr)); 2290 th->th_sum = savecsum; 2291 /* 2292 * Step 3: Update MD5 hash with TCP segment data. 2293 * Use m_apply() to avoid an early m_pullup(). 2294 */ 2295 if (len > 0) 2296 m_apply(m, doff, len, tcpsignature_apply, &ctx); 2297 /* 2298 * Step 4: Update MD5 hash with shared secret. 2299 */ 2300 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth)); 2301 MD5Final(buf, &ctx); 2302 key_sa_recordxfer(sav, m); 2303 key_freesav(sav); 2304 return (0); 2305 } 2306 2307 int 2308 tcpsignature_apply(void *fstate, void *data, unsigned int len) 2309 { 2310 2311 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len); 2312 return (0); 2313 } 2314 #endif /* TCP_SIGNATURE */ 2315 2316 static void 2317 tcp_drop_sysctl_dispatch(netmsg_t nmsg) 2318 { 2319 struct lwkt_msg *lmsg = &nmsg->lmsg; 2320 /* addrs[0] is a foreign socket, addrs[1] is a local one. */ 2321 struct sockaddr_storage *addrs = lmsg->u.ms_resultp; 2322 int error; 2323 struct sockaddr_in *fin, *lin; 2324 #ifdef INET6 2325 struct sockaddr_in6 *fin6, *lin6; 2326 struct in6_addr f6, l6; 2327 #endif 2328 struct inpcb *inp; 2329 2330 switch (addrs[0].ss_family) { 2331 #ifdef INET6 2332 case AF_INET6: 2333 fin6 = (struct sockaddr_in6 *)&addrs[0]; 2334 lin6 = (struct sockaddr_in6 *)&addrs[1]; 2335 error = in6_embedscope(&f6, fin6, NULL, NULL); 2336 if (error) 2337 goto done; 2338 error = in6_embedscope(&l6, lin6, NULL, NULL); 2339 if (error) 2340 goto done; 2341 inp = in6_pcblookup_hash(&tcbinfo[mycpuid], &f6, 2342 fin6->sin6_port, &l6, lin6->sin6_port, FALSE, NULL); 2343 break; 2344 #endif 2345 #ifdef INET 2346 case AF_INET: 2347 fin = (struct sockaddr_in *)&addrs[0]; 2348 lin = (struct sockaddr_in *)&addrs[1]; 2349 inp = in_pcblookup_hash(&tcbinfo[mycpuid], fin->sin_addr, 2350 fin->sin_port, lin->sin_addr, lin->sin_port, FALSE, NULL); 2351 break; 2352 #endif 2353 default: 2354 /* 2355 * Must not reach here, since the address family was 2356 * checked in sysctl handler. 2357 */ 2358 panic("unknown address family %d", addrs[0].ss_family); 2359 } 2360 if (inp != NULL) { 2361 struct tcpcb *tp = intotcpcb(inp); 2362 2363 KASSERT((inp->inp_flags & INP_WILDCARD) == 0, 2364 ("in wildcard hash")); 2365 KASSERT(tp != NULL, ("tcp_drop_sysctl_dispatch: tp is NULL")); 2366 KASSERT((tp->t_flags & TF_LISTEN) == 0, ("listen socket")); 2367 tcp_drop(tp, ECONNABORTED); 2368 error = 0; 2369 } else { 2370 error = ESRCH; 2371 } 2372 #ifdef INET6 2373 done: 2374 #endif 2375 lwkt_replymsg(lmsg, error); 2376 } 2377 2378 static int 2379 sysctl_tcp_drop(SYSCTL_HANDLER_ARGS) 2380 { 2381 /* addrs[0] is a foreign socket, addrs[1] is a local one. */ 2382 struct sockaddr_storage addrs[2]; 2383 struct sockaddr_in *fin, *lin; 2384 #ifdef INET6 2385 struct sockaddr_in6 *fin6, *lin6; 2386 #endif 2387 struct netmsg_base nmsg; 2388 struct lwkt_msg *lmsg = &nmsg.lmsg; 2389 struct lwkt_port *port = NULL; 2390 int error; 2391 2392 fin = lin = NULL; 2393 #ifdef INET6 2394 fin6 = lin6 = NULL; 2395 #endif 2396 error = 0; 2397 2398 if (req->oldptr != NULL || req->oldlen != 0) 2399 return (EINVAL); 2400 if (req->newptr == NULL) 2401 return (EPERM); 2402 if (req->newlen < sizeof(addrs)) 2403 return (ENOMEM); 2404 error = SYSCTL_IN(req, &addrs, sizeof(addrs)); 2405 if (error) 2406 return (error); 2407 2408 switch (addrs[0].ss_family) { 2409 #ifdef INET6 2410 case AF_INET6: 2411 fin6 = (struct sockaddr_in6 *)&addrs[0]; 2412 lin6 = (struct sockaddr_in6 *)&addrs[1]; 2413 if (fin6->sin6_len != sizeof(struct sockaddr_in6) || 2414 lin6->sin6_len != sizeof(struct sockaddr_in6)) 2415 return (EINVAL); 2416 if (IN6_IS_ADDR_V4MAPPED(&fin6->sin6_addr) || 2417 IN6_IS_ADDR_V4MAPPED(&lin6->sin6_addr)) 2418 return (EADDRNOTAVAIL); 2419 #if 0 2420 error = sa6_embedscope(fin6, V_ip6_use_defzone); 2421 if (error) 2422 return (error); 2423 error = sa6_embedscope(lin6, V_ip6_use_defzone); 2424 if (error) 2425 return (error); 2426 #endif 2427 port = tcp6_addrport(); 2428 break; 2429 #endif 2430 #ifdef INET 2431 case AF_INET: 2432 fin = (struct sockaddr_in *)&addrs[0]; 2433 lin = (struct sockaddr_in *)&addrs[1]; 2434 if (fin->sin_len != sizeof(struct sockaddr_in) || 2435 lin->sin_len != sizeof(struct sockaddr_in)) 2436 return (EINVAL); 2437 port = tcp_addrport(fin->sin_addr.s_addr, fin->sin_port, 2438 lin->sin_addr.s_addr, lin->sin_port); 2439 break; 2440 #endif 2441 default: 2442 return (EINVAL); 2443 } 2444 2445 netmsg_init(&nmsg, NULL, &curthread->td_msgport, 0, 2446 tcp_drop_sysctl_dispatch); 2447 lmsg->u.ms_resultp = addrs; 2448 return lwkt_domsg(port, lmsg, 0); 2449 } 2450 2451 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, drop, 2452 CTLTYPE_STRUCT | CTLFLAG_WR | CTLFLAG_SKIP, NULL, 2453 0, sysctl_tcp_drop, "", "Drop TCP connection"); 2454 2455 static int 2456 sysctl_tcps_count(SYSCTL_HANDLER_ARGS) 2457 { 2458 u_long state_count[TCP_NSTATES]; 2459 int cpu; 2460 2461 memset(state_count, 0, sizeof(state_count)); 2462 for (cpu = 0; cpu < ncpus2; ++cpu) { 2463 int i; 2464 2465 for (i = 0; i < TCP_NSTATES; ++i) 2466 state_count[i] += tcpstate_count[cpu].tcps_count[i]; 2467 } 2468 2469 return sysctl_handle_opaque(oidp, state_count, sizeof(state_count), req); 2470 } 2471 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, state_count, 2472 CTLTYPE_OPAQUE | CTLFLAG_RD, NULL, 0, 2473 sysctl_tcps_count, "LU", "TCP connection counts by state"); 2474 2475 void 2476 tcp_pcbport_create(struct tcpcb *tp) 2477 { 2478 int cpu; 2479 2480 KASSERT((tp->t_flags & TF_LISTEN) && tp->t_state == TCPS_LISTEN, 2481 ("not a listen tcpcb")); 2482 2483 KASSERT(tp->t_pcbport == NULL, ("tcpcb port cache was created")); 2484 tp->t_pcbport = kmalloc_cachealign(sizeof(struct tcp_pcbport) * ncpus2, 2485 M_PCB, M_WAITOK); 2486 2487 for (cpu = 0; cpu < ncpus2; ++cpu) { 2488 struct inpcbport *phd; 2489 2490 phd = &tp->t_pcbport[cpu].t_phd; 2491 LIST_INIT(&phd->phd_pcblist); 2492 /* Though, not used ... */ 2493 phd->phd_port = tp->t_inpcb->inp_lport; 2494 } 2495 } 2496 2497 void 2498 tcp_pcbport_merge_oncpu(struct tcpcb *tp) 2499 { 2500 struct inpcbport *phd; 2501 struct inpcb *inp; 2502 int cpu = mycpuid; 2503 2504 KASSERT(cpu < ncpus2, ("invalid cpu%d", cpu)); 2505 phd = &tp->t_pcbport[cpu].t_phd; 2506 2507 while ((inp = LIST_FIRST(&phd->phd_pcblist)) != NULL) { 2508 KASSERT(inp->inp_phd == phd && inp->inp_porthash == NULL, 2509 ("not on tcpcb port cache")); 2510 LIST_REMOVE(inp, inp_portlist); 2511 in_pcbinsporthash_lport(inp); 2512 KASSERT(inp->inp_phd == tp->t_inpcb->inp_phd && 2513 inp->inp_porthash == tp->t_inpcb->inp_porthash, 2514 ("tcpcb port cache merge failed")); 2515 } 2516 } 2517 2518 void 2519 tcp_pcbport_destroy(struct tcpcb *tp) 2520 { 2521 #ifdef INVARIANTS 2522 int cpu; 2523 2524 for (cpu = 0; cpu < ncpus2; ++cpu) { 2525 KASSERT(LIST_EMPTY(&tp->t_pcbport[cpu].t_phd.phd_pcblist), 2526 ("tcpcb port cache is not empty")); 2527 } 2528 #endif 2529 kfree(tp->t_pcbport, M_PCB); 2530 tp->t_pcbport = NULL; 2531 } 2532