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