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_tcpdebug.h" 69 70 #include <sys/param.h> 71 #include <sys/systm.h> 72 #include <sys/callout.h> 73 #include <sys/kernel.h> 74 #include <sys/sysctl.h> 75 #include <sys/malloc.h> 76 #include <sys/mpipe.h> 77 #include <sys/mbuf.h> 78 #ifdef INET6 79 #include <sys/domain.h> 80 #endif 81 #include <sys/proc.h> 82 #include <sys/priv.h> 83 #include <sys/socket.h> 84 #include <sys/socketops.h> 85 #include <sys/socketvar.h> 86 #include <sys/protosw.h> 87 #include <sys/random.h> 88 #include <sys/in_cksum.h> 89 #include <sys/ktr.h> 90 91 #include <net/route.h> 92 #include <net/if.h> 93 #include <net/netisr2.h> 94 95 #define _IP_VHL 96 #include <netinet/in.h> 97 #include <netinet/in_systm.h> 98 #include <netinet/ip.h> 99 #include <netinet/ip6.h> 100 #include <netinet/in_pcb.h> 101 #include <netinet6/in6_pcb.h> 102 #include <netinet/in_var.h> 103 #include <netinet/ip_var.h> 104 #include <netinet6/ip6_var.h> 105 #include <netinet/ip_icmp.h> 106 #ifdef INET6 107 #include <netinet/icmp6.h> 108 #endif 109 #include <netinet/tcp.h> 110 #include <netinet/tcp_fsm.h> 111 #include <netinet/tcp_seq.h> 112 #include <netinet/tcp_timer.h> 113 #include <netinet/tcp_timer2.h> 114 #include <netinet/tcp_var.h> 115 #include <netinet6/tcp6_var.h> 116 #include <netinet/tcpip.h> 117 #ifdef TCPDEBUG 118 #include <netinet/tcp_debug.h> 119 #endif 120 #include <netinet6/ip6protosw.h> 121 122 #include <sys/md5.h> 123 #include <machine/smp.h> 124 125 #include <sys/msgport2.h> 126 #include <sys/mplock2.h> 127 #include <net/netmsg2.h> 128 129 #if !defined(KTR_TCP) 130 #define KTR_TCP KTR_ALL 131 #endif 132 /* 133 KTR_INFO_MASTER(tcp); 134 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0); 135 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0); 136 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0); 137 #define logtcp(name) KTR_LOG(tcp_ ## name) 138 */ 139 140 #define TCP_IW_MAXSEGS_DFLT 4 141 #define TCP_IW_CAPSEGS_DFLT 4 142 143 struct tcp_reass_pcpu { 144 int draining; 145 struct netmsg_base drain_nmsg; 146 } __cachealign; 147 148 struct inpcbinfo tcbinfo[MAXCPU]; 149 struct tcpcbackq tcpcbackq[MAXCPU]; 150 struct tcp_reass_pcpu tcp_reassq[MAXCPU]; 151 152 int tcp_mssdflt = TCP_MSS; 153 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 154 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size"); 155 156 #ifdef INET6 157 int tcp_v6mssdflt = TCP6_MSS; 158 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW, 159 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6"); 160 #endif 161 162 /* 163 * Minimum MSS we accept and use. This prevents DoS attacks where 164 * we are forced to a ridiculous low MSS like 20 and send hundreds 165 * of packets instead of one. The effect scales with the available 166 * bandwidth and quickly saturates the CPU and network interface 167 * with packet generation and sending. Set to zero to disable MINMSS 168 * checking. This setting prevents us from sending too small packets. 169 */ 170 int tcp_minmss = TCP_MINMSS; 171 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW, 172 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size"); 173 174 #if 0 175 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ; 176 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW, 177 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time"); 178 #endif 179 180 int tcp_do_rfc1323 = 1; 181 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW, 182 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions"); 183 184 static int tcp_tcbhashsize = 0; 185 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD, 186 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable"); 187 188 static int do_tcpdrain = 1; 189 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0, 190 "Enable tcp_drain routine for extra help when low on mbufs"); 191 192 static int icmp_may_rst = 1; 193 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0, 194 "Certain ICMP unreachable messages may abort connections in SYN_SENT"); 195 196 /* 197 * Recommend 20 (6 times in two minutes) 198 * 199 * Lower values may cause the sequence space to cycle too quickly and lose 200 * its signed monotonically-increasing nature within the 2-minute TIMEDWAIT 201 * window. 202 */ 203 static int tcp_isn_reseed_interval = 20; 204 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW, 205 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret"); 206 207 /* 208 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on 209 * by default, but with generous values which should allow maximal 210 * bandwidth. In particular, the slop defaults to 50 (5 packets). 211 * 212 * The reason for doing this is that the limiter is the only mechanism we 213 * have which seems to do a really good job preventing receiver RX rings 214 * on network interfaces from getting blown out. Even though GigE/10GigE 215 * is supposed to flow control it looks like either it doesn't actually 216 * do it or Open Source drivers do not properly enable it. 217 * 218 * People using the limiter to reduce bottlenecks on slower WAN connections 219 * should set the slop to 20 (2 packets). 220 */ 221 static int tcp_inflight_enable = 1; 222 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW, 223 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting"); 224 225 static int tcp_inflight_debug = 0; 226 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW, 227 &tcp_inflight_debug, 0, "Debug TCP inflight calculations"); 228 229 /* 230 * NOTE: tcp_inflight_start is essentially the starting receive window 231 * for a connection. If set too low then fetches over tcp 232 * connections will take noticably longer to ramp-up over 233 * high-latency connections. 6144 is too low for a default, 234 * use something more reasonable. 235 */ 236 static int tcp_inflight_start = 33792; 237 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_start, CTLFLAG_RW, 238 &tcp_inflight_start, 0, "Start value for TCP inflight window"); 239 240 static int tcp_inflight_min = 6144; 241 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW, 242 &tcp_inflight_min, 0, "Lower bound for TCP inflight window"); 243 244 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT; 245 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW, 246 &tcp_inflight_max, 0, "Upper bound for TCP inflight window"); 247 248 static int tcp_inflight_stab = 50; 249 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW, 250 &tcp_inflight_stab, 0, "Fudge bw 1/10% (50=5%)"); 251 252 static int tcp_inflight_adjrtt = 2; 253 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_adjrtt, CTLFLAG_RW, 254 &tcp_inflight_adjrtt, 0, "Slop for rtt 1/(hz*32)"); 255 256 static int tcp_do_rfc3390 = 1; 257 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_RW, 258 &tcp_do_rfc3390, 0, 259 "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)"); 260 261 static u_long tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 262 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwmaxsegs, CTLFLAG_RW, 263 &tcp_iw_maxsegs, 0, "TCP IW segments max"); 264 265 static u_long tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 266 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwcapsegs, CTLFLAG_RW, 267 &tcp_iw_capsegs, 0, "TCP IW segments"); 268 269 int tcp_low_rtobase = 1; 270 SYSCTL_INT(_net_inet_tcp, OID_AUTO, low_rtobase, CTLFLAG_RW, 271 &tcp_low_rtobase, 0, "Lowering the Initial RTO (RFC 6298)"); 272 273 static int tcp_do_ncr = 1; 274 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr, CTLFLAG_RW, 275 &tcp_do_ncr, 0, "Non-Congestion Robustness (RFC 4653)"); 276 277 int tcp_ncr_linklocal = 0; 278 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr_linklocal, CTLFLAG_RW, 279 &tcp_ncr_linklocal, 0, 280 "Enable Non-Congestion Robustness (RFC 4653) on link local network"); 281 282 int tcp_ncr_rxtthresh_max = 16; 283 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr_rxtthresh_max, CTLFLAG_RW, 284 &tcp_ncr_rxtthresh_max, 0, 285 "Non-Congestion Robustness (RFC 4653), DupThresh upper limit"); 286 287 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives"); 288 static struct malloc_pipe tcptemp_mpipe; 289 290 static void tcp_willblock(void); 291 static void tcp_notify (struct inpcb *, int); 292 293 struct tcp_stats tcpstats_percpu[MAXCPU] __cachealign; 294 struct tcp_state_count tcpstate_count[MAXCPU] __cachealign; 295 296 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 < netisr_ncpus; ++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 CTASSERT(powerof2(TCBHASHSIZE)); 327 328 /* 329 * This is the actual shape of what we allocate using the zone 330 * allocator. Doing it this way allows us to protect both structures 331 * using the same generation count, and also eliminates the overhead 332 * of allocating tcpcbs separately. By hiding the structure here, 333 * we avoid changing most of the rest of the code (although it needs 334 * to be changed, eventually, for greater efficiency). 335 */ 336 #define ALIGNMENT 32 337 #define ALIGNM1 (ALIGNMENT - 1) 338 struct inp_tp { 339 union { 340 struct inpcb inp; 341 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1]; 342 } inp_tp_u; 343 struct tcpcb tcb; 344 struct tcp_callout inp_tp_rexmt; 345 struct tcp_callout inp_tp_persist; 346 struct tcp_callout inp_tp_keep; 347 struct tcp_callout inp_tp_2msl; 348 struct tcp_callout inp_tp_delack; 349 struct netmsg_tcp_timer inp_tp_timermsg; 350 struct netmsg_base inp_tp_sndmore; 351 }; 352 #undef ALIGNMENT 353 #undef ALIGNM1 354 355 /* 356 * Tcp initialization 357 */ 358 void 359 tcp_init(void) 360 { 361 struct inpcbportinfo *portinfo; 362 struct inpcbinfo *ticb; 363 int hashsize = TCBHASHSIZE, portinfo_hsize; 364 int cpu; 365 366 /* 367 * note: tcptemp is used for keepalives, and it is ok for an 368 * allocation to fail so do not specify MPF_INT. 369 */ 370 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp), 371 25, -1, 0, NULL, NULL, NULL); 372 373 tcp_delacktime = TCPTV_DELACK; 374 tcp_keepinit = TCPTV_KEEP_INIT; 375 tcp_keepidle = TCPTV_KEEP_IDLE; 376 tcp_keepintvl = TCPTV_KEEPINTVL; 377 tcp_maxpersistidle = TCPTV_KEEP_IDLE; 378 tcp_msl = TCPTV_MSL; 379 tcp_rexmit_min = TCPTV_MIN; 380 if (tcp_rexmit_min < 1) /* if kern.hz is too low */ 381 tcp_rexmit_min = 1; 382 tcp_rexmit_slop = TCPTV_CPU_VAR; 383 384 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize); 385 if (!powerof2(hashsize)) { 386 kprintf("WARNING: TCB hash size not a power of 2\n"); 387 hashsize = TCBHASHSIZE; /* safe default */ 388 } 389 tcp_tcbhashsize = hashsize; 390 391 portinfo_hsize = 65536 / netisr_ncpus; 392 if (portinfo_hsize > hashsize) 393 portinfo_hsize = hashsize; 394 395 portinfo = kmalloc(sizeof(*portinfo) * netisr_ncpus, M_PCB, 396 M_WAITOK | M_CACHEALIGN); 397 398 for (cpu = 0; cpu < netisr_ncpus; cpu++) { 399 ticb = &tcbinfo[cpu]; 400 in_pcbinfo_init(ticb, cpu, FALSE); 401 ticb->hashbase = hashinit(hashsize, M_PCB, 402 &ticb->hashmask); 403 in_pcbportinfo_init(&portinfo[cpu], portinfo_hsize, cpu); 404 in_pcbportinfo_set(ticb, portinfo, netisr_ncpus); 405 ticb->wildcardhashbase = hashinit(hashsize, M_PCB, 406 &ticb->wildcardhashmask); 407 ticb->localgrphashbase = hashinit(hashsize, M_PCB, 408 &ticb->localgrphashmask); 409 ticb->ipi_size = sizeof(struct inp_tp); 410 TAILQ_INIT(&tcpcbackq[cpu].head); 411 } 412 413 tcp_reass_maxseg = nmbclusters / 16; 414 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg); 415 416 #ifdef INET6 417 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) 418 #else 419 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) 420 #endif 421 if (max_protohdr < TCP_MINPROTOHDR) 422 max_protohdr = TCP_MINPROTOHDR; 423 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) 424 panic("tcp_init"); 425 #undef TCP_MINPROTOHDR 426 427 /* 428 * Initialize TCP statistics counters for each CPU. 429 */ 430 for (cpu = 0; cpu < netisr_ncpus; ++cpu) 431 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats)); 432 433 /* 434 * Initialize netmsgs for TCP drain 435 */ 436 for (cpu = 0; cpu < netisr_ncpus; ++cpu) { 437 netmsg_init(&tcp_reassq[cpu].drain_nmsg, NULL, 438 &netisr_adone_rport, MSGF_PRIORITY, tcp_drain_dispatch); 439 } 440 441 syncache_init(); 442 netisr_register_rollup(tcp_willblock, NETISR_ROLLUP_PRIO_TCP); 443 } 444 445 static void 446 tcp_willblock(void) 447 { 448 struct tcpcb *tp; 449 int cpu = mycpuid; 450 451 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu].head)) != NULL) { 452 KKASSERT(tp->t_flags & TF_ONOUTPUTQ); 453 tp->t_flags &= ~TF_ONOUTPUTQ; 454 TAILQ_REMOVE(&tcpcbackq[cpu].head, tp, t_outputq); 455 tcp_output(tp); 456 } 457 } 458 459 /* 460 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. 461 * tcp_template used to store this data in mbufs, but we now recopy it out 462 * of the tcpcb each time to conserve mbufs. 463 */ 464 void 465 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr, boolean_t tso) 466 { 467 struct inpcb *inp = tp->t_inpcb; 468 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr; 469 470 #ifdef INET6 471 if (INP_ISIPV6(inp)) { 472 struct ip6_hdr *ip6; 473 474 ip6 = (struct ip6_hdr *)ip_ptr; 475 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | 476 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK); 477 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | 478 (IPV6_VERSION & IPV6_VERSION_MASK); 479 ip6->ip6_nxt = IPPROTO_TCP; 480 ip6->ip6_plen = sizeof(struct tcphdr); 481 ip6->ip6_src = inp->in6p_laddr; 482 ip6->ip6_dst = inp->in6p_faddr; 483 tcp_hdr->th_sum = 0; 484 } else 485 #endif 486 { 487 struct ip *ip = (struct ip *) ip_ptr; 488 u_int plen; 489 490 ip->ip_vhl = IP_VHL_BORING; 491 ip->ip_tos = 0; 492 ip->ip_len = 0; 493 ip->ip_id = 0; 494 ip->ip_off = 0; 495 ip->ip_ttl = 0; 496 ip->ip_sum = 0; 497 ip->ip_p = IPPROTO_TCP; 498 ip->ip_src = inp->inp_laddr; 499 ip->ip_dst = inp->inp_faddr; 500 501 if (tso) 502 plen = htons(IPPROTO_TCP); 503 else 504 plen = htons(sizeof(struct tcphdr) + IPPROTO_TCP); 505 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr, 506 ip->ip_dst.s_addr, plen); 507 } 508 509 tcp_hdr->th_sport = inp->inp_lport; 510 tcp_hdr->th_dport = inp->inp_fport; 511 tcp_hdr->th_seq = 0; 512 tcp_hdr->th_ack = 0; 513 tcp_hdr->th_x2 = 0; 514 tcp_hdr->th_off = 5; 515 tcp_hdr->th_flags = 0; 516 tcp_hdr->th_win = 0; 517 tcp_hdr->th_urp = 0; 518 } 519 520 /* 521 * Create template to be used to send tcp packets on a connection. 522 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only 523 * use for this function is in keepalives, which use tcp_respond. 524 */ 525 struct tcptemp * 526 tcp_maketemplate(struct tcpcb *tp) 527 { 528 struct tcptemp *tmp; 529 530 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL) 531 return (NULL); 532 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t, FALSE); 533 return (tmp); 534 } 535 536 void 537 tcp_freetemplate(struct tcptemp *tmp) 538 { 539 mpipe_free(&tcptemp_mpipe, tmp); 540 } 541 542 /* 543 * Send a single message to the TCP at address specified by 544 * the given TCP/IP header. If m == NULL, then we make a copy 545 * of the tcpiphdr at ti and send directly to the addressed host. 546 * This is used to force keep alive messages out using the TCP 547 * template for a connection. If flags are given then we send 548 * a message back to the TCP which originated the * segment ti, 549 * and discard the mbuf containing it and any other attached mbufs. 550 * 551 * In any case the ack and sequence number of the transmitted 552 * segment are as specified by the parameters. 553 * 554 * NOTE: If m != NULL, then ti must point to *inside* the mbuf. 555 */ 556 void 557 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, 558 tcp_seq ack, tcp_seq seq, int flags) 559 { 560 int tlen; 561 long win = 0; 562 struct route *ro = NULL; 563 struct route sro; 564 struct ip *ip = ipgen; 565 struct tcphdr *nth; 566 int ipflags = 0; 567 struct route_in6 *ro6 = NULL; 568 struct route_in6 sro6; 569 struct ip6_hdr *ip6 = ipgen; 570 struct inpcb *inp = NULL; 571 boolean_t use_tmpro = TRUE; 572 #ifdef INET6 573 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6); 574 #else 575 const boolean_t isipv6 = FALSE; 576 #endif 577 578 if (tp != NULL) { 579 inp = tp->t_inpcb; 580 if (!(flags & TH_RST)) { 581 win = ssb_space(&inp->inp_socket->so_rcv); 582 if (win < 0) 583 win = 0; 584 if (win > (long)TCP_MAXWIN << tp->rcv_scale) 585 win = (long)TCP_MAXWIN << tp->rcv_scale; 586 } 587 /* 588 * Don't use the route cache of a listen socket, 589 * it is not MPSAFE; use temporary route cache. 590 */ 591 if (tp->t_state != TCPS_LISTEN) { 592 if (isipv6) 593 ro6 = &inp->in6p_route; 594 else 595 ro = &inp->inp_route; 596 use_tmpro = FALSE; 597 } 598 } 599 if (use_tmpro) { 600 if (isipv6) { 601 ro6 = &sro6; 602 bzero(ro6, sizeof *ro6); 603 } else { 604 ro = &sro; 605 bzero(ro, sizeof *ro); 606 } 607 } 608 if (m == NULL) { 609 m = m_gethdr(M_NOWAIT, MT_HEADER); 610 if (m == NULL) 611 return; 612 tlen = 0; 613 m->m_data += max_linkhdr; 614 if (isipv6) { 615 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); 616 ip6 = mtod(m, struct ip6_hdr *); 617 nth = (struct tcphdr *)(ip6 + 1); 618 } else { 619 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip)); 620 ip = mtod(m, struct ip *); 621 nth = (struct tcphdr *)(ip + 1); 622 } 623 bcopy(th, nth, sizeof(struct tcphdr)); 624 flags = TH_ACK; 625 } else { 626 m_freem(m->m_next); 627 m->m_next = NULL; 628 m->m_data = (caddr_t)ipgen; 629 /* m_len is set later */ 630 tlen = 0; 631 #define xchg(a, b, type) { type t; t = a; a = b; b = t; } 632 if (isipv6) { 633 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); 634 nth = (struct tcphdr *)(ip6 + 1); 635 } else { 636 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long); 637 nth = (struct tcphdr *)(ip + 1); 638 } 639 if (th != nth) { 640 /* 641 * this is usually a case when an extension header 642 * exists between the IPv6 header and the 643 * TCP header. 644 */ 645 nth->th_sport = th->th_sport; 646 nth->th_dport = th->th_dport; 647 } 648 xchg(nth->th_dport, nth->th_sport, n_short); 649 #undef xchg 650 } 651 if (isipv6) { 652 ip6->ip6_flow = 0; 653 ip6->ip6_vfc = IPV6_VERSION; 654 ip6->ip6_nxt = IPPROTO_TCP; 655 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen)); 656 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 657 } else { 658 tlen += sizeof(struct tcpiphdr); 659 ip->ip_len = tlen; 660 ip->ip_ttl = ip_defttl; 661 } 662 m->m_len = tlen; 663 m->m_pkthdr.len = tlen; 664 m->m_pkthdr.rcvif = NULL; 665 nth->th_seq = htonl(seq); 666 nth->th_ack = htonl(ack); 667 nth->th_x2 = 0; 668 nth->th_off = sizeof(struct tcphdr) >> 2; 669 nth->th_flags = flags; 670 if (tp != NULL) 671 nth->th_win = htons((u_short) (win >> tp->rcv_scale)); 672 else 673 nth->th_win = htons((u_short)win); 674 nth->th_urp = 0; 675 if (isipv6) { 676 nth->th_sum = 0; 677 nth->th_sum = in6_cksum(m, IPPROTO_TCP, 678 sizeof(struct ip6_hdr), 679 tlen - sizeof(struct ip6_hdr)); 680 ip6->ip6_hlim = in6_selecthlim(inp, 681 (ro6 && ro6->ro_rt) ? ro6->ro_rt->rt_ifp : NULL); 682 } else { 683 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 684 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); 685 m->m_pkthdr.csum_flags = CSUM_TCP; 686 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 687 m->m_pkthdr.csum_thlen = sizeof(struct tcphdr); 688 } 689 #ifdef TCPDEBUG 690 if (tp == NULL || (inp->inp_socket->so_options & SO_DEBUG)) 691 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); 692 #endif 693 if (isipv6) { 694 ip6_output(m, NULL, ro6, ipflags, NULL, NULL, inp); 695 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) { 696 RTFREE(ro6->ro_rt); 697 ro6->ro_rt = NULL; 698 } 699 } else { 700 if (inp != NULL && (inp->inp_flags & INP_HASH)) 701 m_sethash(m, inp->inp_hashval); 702 ipflags |= IP_DEBUGROUTE; 703 ip_output(m, NULL, ro, ipflags, NULL, inp); 704 if ((ro == &sro) && (ro->ro_rt != NULL)) { 705 RTFREE(ro->ro_rt); 706 ro->ro_rt = NULL; 707 } 708 } 709 } 710 711 /* 712 * Create a new TCP control block, making an 713 * empty reassembly queue and hooking it to the argument 714 * protocol control block. The `inp' parameter must have 715 * come from the zone allocator set up in tcp_init(). 716 */ 717 void 718 tcp_newtcpcb(struct inpcb *inp) 719 { 720 struct inp_tp *it; 721 struct tcpcb *tp; 722 #ifdef INET6 723 boolean_t isipv6 = INP_ISIPV6(inp); 724 #else 725 const boolean_t isipv6 = FALSE; 726 #endif 727 728 it = (struct inp_tp *)inp; 729 tp = &it->tcb; 730 bzero(tp, sizeof(struct tcpcb)); 731 TAILQ_INIT(&tp->t_segq); 732 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt; 733 tp->t_rxtthresh = tcprexmtthresh; 734 735 /* Set up our timeouts. */ 736 tp->tt_rexmt = &it->inp_tp_rexmt; 737 tp->tt_persist = &it->inp_tp_persist; 738 tp->tt_keep = &it->inp_tp_keep; 739 tp->tt_2msl = &it->inp_tp_2msl; 740 tp->tt_delack = &it->inp_tp_delack; 741 tcp_inittimers(tp); 742 743 /* 744 * Zero out timer message. We don't create it here, 745 * since the current CPU may not be the owner of this 746 * inpcb. 747 */ 748 tp->tt_msg = &it->inp_tp_timermsg; 749 bzero(tp->tt_msg, sizeof(*tp->tt_msg)); 750 751 tp->t_keepinit = tcp_keepinit; 752 tp->t_keepidle = tcp_keepidle; 753 tp->t_keepintvl = tcp_keepintvl; 754 tp->t_keepcnt = tcp_keepcnt; 755 tp->t_maxidle = tp->t_keepintvl * tp->t_keepcnt; 756 757 if (tcp_do_ncr) 758 tp->t_flags |= TF_NCR; 759 if (tcp_do_rfc1323) 760 tp->t_flags |= (TF_REQ_SCALE | TF_REQ_TSTMP); 761 762 tp->t_inpcb = inp; /* XXX */ 763 TCP_STATE_INIT(tp); 764 /* 765 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no 766 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives 767 * reasonable initial retransmit time. 768 */ 769 tp->t_srtt = TCPTV_SRTTBASE; 770 tp->t_rttvar = 771 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; 772 tp->t_rttmin = tcp_rexmit_min; 773 tp->t_rxtcur = TCPTV_RTOBASE; 774 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 775 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 776 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; 777 tp->snd_last = ticks; 778 tp->t_rcvtime = ticks; 779 /* 780 * IPv4 TTL initialization is necessary for an IPv6 socket as well, 781 * because the socket may be bound to an IPv6 wildcard address, 782 * which may match an IPv4-mapped IPv6 address. 783 */ 784 inp->inp_ip_ttl = ip_defttl; 785 inp->inp_ppcb = tp; 786 tcp_sack_tcpcb_init(tp); 787 788 tp->tt_sndmore = &it->inp_tp_sndmore; 789 tcp_output_init(tp); 790 } 791 792 /* 793 * Drop a TCP connection, reporting the specified error. 794 * If connection is synchronized, then send a RST to peer. 795 */ 796 struct tcpcb * 797 tcp_drop(struct tcpcb *tp, int error) 798 { 799 struct socket *so = tp->t_inpcb->inp_socket; 800 801 if (TCPS_HAVERCVDSYN(tp->t_state)) { 802 TCP_STATE_CHANGE(tp, TCPS_CLOSED); 803 tcp_output(tp); 804 tcpstat.tcps_drops++; 805 } else 806 tcpstat.tcps_conndrops++; 807 if (error == ETIMEDOUT && tp->t_softerror) 808 error = tp->t_softerror; 809 so->so_error = error; 810 return (tcp_close(tp)); 811 } 812 813 struct netmsg_listen_detach { 814 struct netmsg_base base; 815 struct tcpcb *nm_tp; 816 struct tcpcb *nm_tp_inh; 817 }; 818 819 static void 820 tcp_listen_detach_handler(netmsg_t msg) 821 { 822 struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg; 823 struct tcpcb *tp = nmsg->nm_tp; 824 int cpu = mycpuid, nextcpu; 825 826 if (tp->t_flags & TF_LISTEN) { 827 syncache_destroy(tp, nmsg->nm_tp_inh); 828 tcp_pcbport_merge_oncpu(tp); 829 } 830 831 in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]); 832 833 nextcpu = cpu + 1; 834 if (nextcpu < netisr_ncpus) 835 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nmsg->base.lmsg); 836 else 837 lwkt_replymsg(&nmsg->base.lmsg, 0); 838 } 839 840 /* 841 * Close a TCP control block: 842 * discard all space held by the tcp 843 * discard internet protocol block 844 * wake up any sleepers 845 */ 846 struct tcpcb * 847 tcp_close(struct tcpcb *tp) 848 { 849 struct tseg_qent *q; 850 struct inpcb *inp = tp->t_inpcb; 851 struct inpcb *inp_inh = NULL; 852 struct tcpcb *tp_inh = NULL; 853 struct socket *so = inp->inp_socket; 854 struct rtentry *rt; 855 boolean_t dosavessthresh; 856 #ifdef INET6 857 boolean_t isipv6 = INP_ISIPV6(inp); 858 #else 859 const boolean_t isipv6 = FALSE; 860 #endif 861 862 if (tp->t_flags & TF_LISTEN) { 863 /* 864 * Pending socket/syncache inheritance 865 * 866 * If this is a listen(2) socket, find another listen(2) 867 * socket in the same local group, which could inherit 868 * the syncache and sockets pending on the completion 869 * and incompletion queues. 870 * 871 * NOTE: 872 * Currently the inheritance could only happen on the 873 * listen(2) sockets w/ SO_REUSEPORT set. 874 */ 875 ASSERT_NETISR0; 876 inp_inh = in_pcblocalgroup_last(&tcbinfo[0], inp); 877 if (inp_inh != NULL) 878 tp_inh = intotcpcb(inp_inh); 879 } 880 881 /* 882 * INP_WILDCARD indicates that listen(2) has been called on 883 * this socket. This implies: 884 * - A wildcard inp's hash is replicated for each protocol thread. 885 * - Syncache for this inp grows independently in each protocol 886 * thread. 887 * - There is more than one cpu 888 * 889 * We have to chain a message to the rest of the protocol threads 890 * to cleanup the wildcard hash and the syncache. The cleanup 891 * in the current protocol thread is defered till the end of this 892 * function (syncache_destroy and in_pcbdetach). 893 * 894 * NOTE: 895 * After cleanup the inp's hash and syncache entries, this inp will 896 * no longer be available to the rest of the protocol threads, so we 897 * are safe to whack the inp in the following code. 898 */ 899 if ((inp->inp_flags & INP_WILDCARD) && netisr_ncpus > 1) { 900 struct netmsg_listen_detach nmsg; 901 902 KKASSERT(so->so_port == netisr_cpuport(0)); 903 ASSERT_NETISR0; 904 KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]); 905 906 netmsg_init(&nmsg.base, NULL, &curthread->td_msgport, 907 MSGF_PRIORITY, tcp_listen_detach_handler); 908 nmsg.nm_tp = tp; 909 nmsg.nm_tp_inh = tp_inh; 910 lwkt_domsg(netisr_cpuport(1), &nmsg.base.lmsg, 0); 911 } 912 913 TCP_STATE_TERM(tp); 914 915 /* 916 * Make sure that all of our timers are stopped before we 917 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL), 918 * timers are never used. If timer message is never created 919 * (tp->tt_msg->tt_tcb == NULL), timers are never used too. 920 */ 921 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) { 922 tcp_callout_terminate(tp, tp->tt_rexmt); 923 tcp_callout_terminate(tp, tp->tt_persist); 924 tcp_callout_terminate(tp, tp->tt_keep); 925 tcp_callout_terminate(tp, tp->tt_2msl); 926 tcp_callout_terminate(tp, tp->tt_delack); 927 } 928 929 if (tp->t_flags & TF_ONOUTPUTQ) { 930 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid); 931 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu].head, tp, t_outputq); 932 tp->t_flags &= ~TF_ONOUTPUTQ; 933 } 934 935 /* 936 * If we got enough samples through the srtt filter, 937 * save the rtt and rttvar in the routing entry. 938 * 'Enough' is arbitrarily defined as the 16 samples. 939 * 16 samples is enough for the srtt filter to converge 940 * to within 5% of the correct value; fewer samples and 941 * we could save a very bogus rtt. 942 * 943 * Don't update the default route's characteristics and don't 944 * update anything that the user "locked". 945 */ 946 if (tp->t_rttupdated >= 16) { 947 u_long i = 0; 948 949 if (isipv6) { 950 struct sockaddr_in6 *sin6; 951 952 if ((rt = inp->in6p_route.ro_rt) == NULL) 953 goto no_valid_rt; 954 sin6 = (struct sockaddr_in6 *)rt_key(rt); 955 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) 956 goto no_valid_rt; 957 } else 958 if ((rt = inp->inp_route.ro_rt) == NULL || 959 ((struct sockaddr_in *)rt_key(rt))-> 960 sin_addr.s_addr == INADDR_ANY) 961 goto no_valid_rt; 962 963 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) { 964 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE)); 965 if (rt->rt_rmx.rmx_rtt && i) 966 /* 967 * filter this update to half the old & half 968 * the new values, converting scale. 969 * See route.h and tcp_var.h for a 970 * description of the scaling constants. 971 */ 972 rt->rt_rmx.rmx_rtt = 973 (rt->rt_rmx.rmx_rtt + i) / 2; 974 else 975 rt->rt_rmx.rmx_rtt = i; 976 tcpstat.tcps_cachedrtt++; 977 } 978 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) { 979 i = tp->t_rttvar * 980 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE)); 981 if (rt->rt_rmx.rmx_rttvar && i) 982 rt->rt_rmx.rmx_rttvar = 983 (rt->rt_rmx.rmx_rttvar + i) / 2; 984 else 985 rt->rt_rmx.rmx_rttvar = i; 986 tcpstat.tcps_cachedrttvar++; 987 } 988 /* 989 * The old comment here said: 990 * update the pipelimit (ssthresh) if it has been updated 991 * already or if a pipesize was specified & the threshhold 992 * got below half the pipesize. I.e., wait for bad news 993 * before we start updating, then update on both good 994 * and bad news. 995 * 996 * But we want to save the ssthresh even if no pipesize is 997 * specified explicitly in the route, because such 998 * connections still have an implicit pipesize specified 999 * by the global tcp_sendspace. In the absence of a reliable 1000 * way to calculate the pipesize, it will have to do. 1001 */ 1002 i = tp->snd_ssthresh; 1003 if (rt->rt_rmx.rmx_sendpipe != 0) 1004 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2); 1005 else 1006 dosavessthresh = (i < so->so_snd.ssb_hiwat/2); 1007 if (dosavessthresh || 1008 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) && 1009 (rt->rt_rmx.rmx_ssthresh != 0))) { 1010 /* 1011 * convert the limit from user data bytes to 1012 * packets then to packet data bytes. 1013 */ 1014 i = (i + tp->t_maxseg / 2) / tp->t_maxseg; 1015 if (i < 2) 1016 i = 2; 1017 i *= tp->t_maxseg + 1018 (isipv6 ? 1019 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1020 sizeof(struct tcpiphdr)); 1021 if (rt->rt_rmx.rmx_ssthresh) 1022 rt->rt_rmx.rmx_ssthresh = 1023 (rt->rt_rmx.rmx_ssthresh + i) / 2; 1024 else 1025 rt->rt_rmx.rmx_ssthresh = i; 1026 tcpstat.tcps_cachedssthresh++; 1027 } 1028 } 1029 1030 no_valid_rt: 1031 /* free the reassembly queue, if any */ 1032 while((q = TAILQ_FIRST(&tp->t_segq)) != NULL) { 1033 TAILQ_REMOVE(&tp->t_segq, q, tqe_q); 1034 m_freem(q->tqe_m); 1035 kfree(q, M_TSEGQ); 1036 atomic_add_int(&tcp_reass_qsize, -1); 1037 } 1038 /* throw away SACK blocks in scoreboard*/ 1039 if (TCP_DO_SACK(tp)) 1040 tcp_sack_destroy(&tp->scb); 1041 1042 inp->inp_ppcb = NULL; 1043 soisdisconnected(so); 1044 /* note: pcb detached later on */ 1045 1046 tcp_destroy_timermsg(tp); 1047 tcp_output_cancel(tp); 1048 1049 if (tp->t_flags & TF_LISTEN) { 1050 syncache_destroy(tp, tp_inh); 1051 tcp_pcbport_merge_oncpu(tp); 1052 tcp_pcbport_destroy(tp); 1053 if (inp_inh != NULL && inp_inh->inp_socket != NULL) { 1054 /* 1055 * Pending sockets inheritance only needs 1056 * to be done once in the current thread, 1057 * i.e. netisr0. 1058 */ 1059 soinherit(so, inp_inh->inp_socket); 1060 } 1061 } 1062 KASSERT(tp->t_pcbport == NULL, ("tcpcb port cache is not destroyed")); 1063 1064 so_async_rcvd_drop(so); 1065 /* Drop the reference for the asynchronized pru_rcvd */ 1066 sofree(so); 1067 1068 /* 1069 * NOTE: 1070 * - Remove self from listen tcpcb per-cpu port cache _before_ 1071 * pcbdetach. 1072 * - pcbdetach removes any wildcard hash entry on the current CPU. 1073 */ 1074 tcp_pcbport_remove(inp); 1075 #ifdef INET6 1076 if (isipv6) 1077 in6_pcbdetach(inp); 1078 else 1079 #endif 1080 in_pcbdetach(inp); 1081 1082 tcpstat.tcps_closed++; 1083 return (NULL); 1084 } 1085 1086 /* 1087 * Walk the tcpbs, if existing, and flush the reassembly queue, 1088 * if there is one... 1089 */ 1090 static void 1091 tcp_drain_oncpu(struct inpcbinfo *pcbinfo) 1092 { 1093 struct inpcbhead *head = &pcbinfo->pcblisthead; 1094 struct inpcb *inpb; 1095 1096 /* 1097 * Since we run in netisr, it is MP safe, even if 1098 * we block during the inpcb list iteration, i.e. 1099 * we don't need to use inpcb marker here. 1100 */ 1101 ASSERT_NETISR_NCPUS(pcbinfo->cpu); 1102 1103 LIST_FOREACH(inpb, head, inp_list) { 1104 struct tcpcb *tcpb; 1105 struct tseg_qent *te; 1106 1107 if (inpb->inp_flags & INP_PLACEMARKER) 1108 continue; 1109 1110 tcpb = intotcpcb(inpb); 1111 KASSERT(tcpb != NULL, ("tcp_drain_oncpu: tcpb is NULL")); 1112 1113 if ((te = TAILQ_FIRST(&tcpb->t_segq)) != NULL) { 1114 TAILQ_REMOVE(&tcpb->t_segq, te, tqe_q); 1115 if (te->tqe_th->th_flags & TH_FIN) 1116 tcpb->t_flags &= ~TF_QUEDFIN; 1117 m_freem(te->tqe_m); 1118 kfree(te, M_TSEGQ); 1119 atomic_add_int(&tcp_reass_qsize, -1); 1120 /* retry */ 1121 } 1122 } 1123 } 1124 1125 static void 1126 tcp_drain_dispatch(netmsg_t nmsg) 1127 { 1128 crit_enter(); 1129 lwkt_replymsg(&nmsg->lmsg, 0); /* reply ASAP */ 1130 crit_exit(); 1131 1132 tcp_drain_oncpu(&tcbinfo[mycpuid]); 1133 tcp_reassq[mycpuid].draining = 0; 1134 } 1135 1136 static void 1137 tcp_drain_ipi(void *arg __unused) 1138 { 1139 int cpu = mycpuid; 1140 struct lwkt_msg *msg = &tcp_reassq[cpu].drain_nmsg.lmsg; 1141 1142 crit_enter(); 1143 if (msg->ms_flags & MSGF_DONE) 1144 lwkt_sendmsg_oncpu(netisr_cpuport(cpu), msg); 1145 crit_exit(); 1146 } 1147 1148 void 1149 tcp_drain(void) 1150 { 1151 cpumask_t mask; 1152 int cpu; 1153 1154 if (!do_tcpdrain) 1155 return; 1156 1157 if (tcp_reass_qsize == 0) 1158 return; 1159 1160 CPUMASK_ASSBMASK(mask, netisr_ncpus); 1161 CPUMASK_ANDMASK(mask, smp_active_mask); 1162 1163 cpu = mycpuid; 1164 if (IN_NETISR_NCPUS(cpu)) { 1165 tcp_drain_oncpu(&tcbinfo[cpu]); 1166 CPUMASK_NANDBIT(mask, cpu); 1167 } 1168 1169 if (tcp_reass_qsize < netisr_ncpus) { 1170 /* Does not worth the trouble. */ 1171 return; 1172 } 1173 1174 for (cpu = 0; cpu < netisr_ncpus; ++cpu) { 1175 if (!CPUMASK_TESTBIT(mask, cpu)) 1176 continue; 1177 1178 if (tcp_reassq[cpu].draining) { 1179 /* Draining; skip this cpu. */ 1180 CPUMASK_NANDBIT(mask, cpu); 1181 continue; 1182 } 1183 tcp_reassq[cpu].draining = 1; 1184 } 1185 1186 if (CPUMASK_TESTNZERO(mask)) 1187 lwkt_send_ipiq_mask(mask, tcp_drain_ipi, NULL); 1188 } 1189 1190 /* 1191 * Notify a tcp user of an asynchronous error; 1192 * store error as soft error, but wake up user 1193 * (for now, won't do anything until can select for soft error). 1194 * 1195 * Do not wake up user since there currently is no mechanism for 1196 * reporting soft errors (yet - a kqueue filter may be added). 1197 */ 1198 static void 1199 tcp_notify(struct inpcb *inp, int error) 1200 { 1201 struct tcpcb *tp = intotcpcb(inp); 1202 1203 /* 1204 * Ignore some errors if we are hooked up. 1205 * If connection hasn't completed, has retransmitted several times, 1206 * and receives a second error, give up now. This is better 1207 * than waiting a long time to establish a connection that 1208 * can never complete. 1209 */ 1210 if (tp->t_state == TCPS_ESTABLISHED && 1211 (error == EHOSTUNREACH || error == ENETUNREACH || 1212 error == EHOSTDOWN)) { 1213 return; 1214 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && 1215 tp->t_softerror) 1216 tcp_drop(tp, error); 1217 else 1218 tp->t_softerror = error; 1219 #if 0 1220 wakeup(&so->so_timeo); 1221 sorwakeup(so); 1222 sowwakeup(so); 1223 #endif 1224 } 1225 1226 static int 1227 tcp_pcblist(SYSCTL_HANDLER_ARGS) 1228 { 1229 int error, i, n; 1230 struct inpcb *marker; 1231 struct inpcb *inp; 1232 int origcpu, ccpu; 1233 1234 error = 0; 1235 n = 0; 1236 1237 /* 1238 * The process of preparing the TCB list is too time-consuming and 1239 * resource-intensive to repeat twice on every request. 1240 */ 1241 if (req->oldptr == NULL) { 1242 for (ccpu = 0; ccpu < netisr_ncpus; ++ccpu) 1243 n += tcbinfo[ccpu].ipi_count; 1244 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb); 1245 return (0); 1246 } 1247 1248 if (req->newptr != NULL) 1249 return (EPERM); 1250 1251 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO); 1252 marker->inp_flags |= INP_PLACEMARKER; 1253 1254 /* 1255 * OK, now we're committed to doing something. Run the inpcb list 1256 * for each cpu in the system and construct the output. Use a 1257 * list placemarker to deal with list changes occuring during 1258 * copyout blockages (but otherwise depend on being on the correct 1259 * cpu to avoid races). 1260 */ 1261 origcpu = mycpu->gd_cpuid; 1262 for (ccpu = 0; ccpu < netisr_ncpus && error == 0; ++ccpu) { 1263 caddr_t inp_ppcb; 1264 struct xtcpcb xt; 1265 1266 lwkt_migratecpu(ccpu); 1267 1268 n = tcbinfo[ccpu].ipi_count; 1269 1270 LIST_INSERT_HEAD(&tcbinfo[ccpu].pcblisthead, marker, inp_list); 1271 i = 0; 1272 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) { 1273 /* 1274 * process a snapshot of pcbs, ignoring placemarkers 1275 * and using our own to allow SYSCTL_OUT to block. 1276 */ 1277 LIST_REMOVE(marker, inp_list); 1278 LIST_INSERT_AFTER(inp, marker, inp_list); 1279 1280 if (inp->inp_flags & INP_PLACEMARKER) 1281 continue; 1282 if (prison_xinpcb(req->td, inp)) 1283 continue; 1284 1285 xt.xt_len = sizeof xt; 1286 bcopy(inp, &xt.xt_inp, sizeof *inp); 1287 inp_ppcb = inp->inp_ppcb; 1288 if (inp_ppcb != NULL) 1289 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp); 1290 else 1291 bzero(&xt.xt_tp, sizeof xt.xt_tp); 1292 if (inp->inp_socket) 1293 sotoxsocket(inp->inp_socket, &xt.xt_socket); 1294 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0) 1295 break; 1296 ++i; 1297 } 1298 LIST_REMOVE(marker, inp_list); 1299 if (error == 0 && i < n) { 1300 bzero(&xt, sizeof xt); 1301 xt.xt_len = sizeof xt; 1302 while (i < n) { 1303 error = SYSCTL_OUT(req, &xt, sizeof xt); 1304 if (error) 1305 break; 1306 ++i; 1307 } 1308 } 1309 } 1310 1311 /* 1312 * Make sure we are on the same cpu we were on originally, since 1313 * higher level callers expect this. Also don't pollute caches with 1314 * migrated userland data by (eventually) returning to userland 1315 * on a different cpu. 1316 */ 1317 lwkt_migratecpu(origcpu); 1318 kfree(marker, M_TEMP); 1319 return (error); 1320 } 1321 1322 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0, 1323 tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); 1324 1325 static int 1326 tcp_getcred(SYSCTL_HANDLER_ARGS) 1327 { 1328 struct sockaddr_in addrs[2]; 1329 struct ucred cred0, *cred = NULL; 1330 struct inpcb *inp; 1331 int cpu, origcpu, error; 1332 1333 error = priv_check(req->td, PRIV_ROOT); 1334 if (error != 0) 1335 return (error); 1336 error = SYSCTL_IN(req, addrs, sizeof addrs); 1337 if (error != 0) 1338 return (error); 1339 1340 origcpu = mycpuid; 1341 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port, 1342 addrs[0].sin_addr.s_addr, addrs[0].sin_port); 1343 1344 lwkt_migratecpu(cpu); 1345 1346 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr, 1347 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL); 1348 if (inp == NULL || inp->inp_socket == NULL) { 1349 error = ENOENT; 1350 } else if (inp->inp_socket->so_cred != NULL) { 1351 cred0 = *(inp->inp_socket->so_cred); 1352 cred = &cred0; 1353 } 1354 1355 lwkt_migratecpu(origcpu); 1356 1357 if (error) 1358 return (error); 1359 1360 return SYSCTL_OUT(req, cred, sizeof(struct ucred)); 1361 } 1362 1363 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1364 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection"); 1365 1366 #ifdef INET6 1367 static int 1368 tcp6_getcred(SYSCTL_HANDLER_ARGS) 1369 { 1370 struct sockaddr_in6 addrs[2]; 1371 struct inpcb *inp; 1372 int error; 1373 1374 error = priv_check(req->td, PRIV_ROOT); 1375 if (error != 0) 1376 return (error); 1377 error = SYSCTL_IN(req, addrs, sizeof addrs); 1378 if (error != 0) 1379 return (error); 1380 crit_enter(); 1381 inp = in6_pcblookup_hash(&tcbinfo[0], 1382 &addrs[1].sin6_addr, addrs[1].sin6_port, 1383 &addrs[0].sin6_addr, addrs[0].sin6_port, 0, NULL); 1384 if (inp == NULL || inp->inp_socket == NULL) { 1385 error = ENOENT; 1386 goto out; 1387 } 1388 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred)); 1389 out: 1390 crit_exit(); 1391 return (error); 1392 } 1393 1394 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW), 1395 0, 0, 1396 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection"); 1397 #endif 1398 1399 struct netmsg_tcp_notify { 1400 struct netmsg_base base; 1401 inp_notify_t nm_notify; 1402 struct in_addr nm_faddr; 1403 int nm_arg; 1404 }; 1405 1406 static void 1407 tcp_notifyall_oncpu(netmsg_t msg) 1408 { 1409 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg; 1410 int nextcpu; 1411 1412 ASSERT_NETISR_NCPUS(mycpuid); 1413 1414 in_pcbnotifyall(&tcbinfo[mycpuid], nm->nm_faddr, 1415 nm->nm_arg, nm->nm_notify); 1416 1417 nextcpu = mycpuid + 1; 1418 if (nextcpu < netisr_ncpus) 1419 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nm->base.lmsg); 1420 else 1421 lwkt_replymsg(&nm->base.lmsg, 0); 1422 } 1423 1424 inp_notify_t 1425 tcp_get_inpnotify(int cmd, const struct sockaddr *sa, 1426 int *arg, struct ip **ip0, int *cpuid) 1427 { 1428 struct ip *ip = *ip0; 1429 struct in_addr faddr; 1430 inp_notify_t notify = tcp_notify; 1431 1432 faddr = ((const struct sockaddr_in *)sa)->sin_addr; 1433 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) 1434 return NULL; 1435 1436 *arg = inetctlerrmap[cmd]; 1437 if (cmd == PRC_QUENCH) { 1438 notify = tcp_quench; 1439 } else if (icmp_may_rst && 1440 (cmd == PRC_UNREACH_ADMIN_PROHIB || 1441 cmd == PRC_UNREACH_PORT || 1442 cmd == PRC_TIMXCEED_INTRANS) && 1443 ip != NULL) { 1444 notify = tcp_drop_syn_sent; 1445 } else if (cmd == PRC_MSGSIZE) { 1446 const struct icmp *icmp = (const struct icmp *) 1447 ((caddr_t)ip - offsetof(struct icmp, icmp_ip)); 1448 1449 *arg = ntohs(icmp->icmp_nextmtu); 1450 notify = tcp_mtudisc; 1451 } else if (PRC_IS_REDIRECT(cmd)) { 1452 ip = NULL; 1453 notify = in_rtchange; 1454 } else if (cmd == PRC_HOSTDEAD) { 1455 ip = NULL; 1456 } else if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) { 1457 return NULL; 1458 } 1459 1460 if (cpuid != NULL) { 1461 if (ip == NULL) { 1462 /* Go through all effective netisr CPUs. */ 1463 *cpuid = netisr_ncpus; 1464 } else { 1465 const struct tcphdr *th; 1466 1467 th = (const struct tcphdr *) 1468 ((caddr_t)ip + (IP_VHL_HL(ip->ip_vhl) << 2)); 1469 *cpuid = tcp_addrcpu(faddr.s_addr, th->th_dport, 1470 ip->ip_src.s_addr, th->th_sport); 1471 } 1472 } 1473 1474 *ip0 = ip; 1475 return notify; 1476 } 1477 1478 void 1479 tcp_ctlinput(netmsg_t msg) 1480 { 1481 int cmd = msg->ctlinput.nm_cmd; 1482 struct sockaddr *sa = msg->ctlinput.nm_arg; 1483 struct ip *ip = msg->ctlinput.nm_extra; 1484 struct in_addr faddr; 1485 inp_notify_t notify; 1486 int arg, cpuid; 1487 1488 ASSERT_NETISR_NCPUS(mycpuid); 1489 1490 notify = tcp_get_inpnotify(cmd, sa, &arg, &ip, &cpuid); 1491 if (notify == NULL) 1492 goto done; 1493 1494 faddr = ((struct sockaddr_in *)sa)->sin_addr; 1495 if (ip != NULL) { 1496 const struct tcphdr *th; 1497 struct inpcb *inp; 1498 1499 if (cpuid != mycpuid) 1500 goto done; 1501 1502 th = (const struct tcphdr *) 1503 ((caddr_t)ip + (IP_VHL_HL(ip->ip_vhl) << 2)); 1504 inp = in_pcblookup_hash(&tcbinfo[mycpuid], faddr, th->th_dport, 1505 ip->ip_src, th->th_sport, 0, NULL); 1506 if (inp != NULL && inp->inp_socket != NULL) { 1507 tcp_seq icmpseq = htonl(th->th_seq); 1508 struct tcpcb *tp = intotcpcb(inp); 1509 1510 if (SEQ_GEQ(icmpseq, tp->snd_una) && 1511 SEQ_LT(icmpseq, tp->snd_max)) 1512 notify(inp, arg); 1513 } else { 1514 struct in_conninfo inc; 1515 1516 inc.inc_fport = th->th_dport; 1517 inc.inc_lport = th->th_sport; 1518 inc.inc_faddr = faddr; 1519 inc.inc_laddr = ip->ip_src; 1520 #ifdef INET6 1521 inc.inc_isipv6 = 0; 1522 #endif 1523 syncache_unreach(&inc, th); 1524 } 1525 } else if (msg->ctlinput.nm_direct) { 1526 if (cpuid != netisr_ncpus && cpuid != mycpuid) 1527 goto done; 1528 1529 in_pcbnotifyall(&tcbinfo[mycpuid], faddr, arg, notify); 1530 } else { 1531 struct netmsg_tcp_notify *nm; 1532 1533 ASSERT_NETISR0; 1534 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT); 1535 netmsg_init(&nm->base, NULL, &netisr_afree_rport, 1536 0, tcp_notifyall_oncpu); 1537 nm->nm_faddr = faddr; 1538 nm->nm_arg = arg; 1539 nm->nm_notify = notify; 1540 1541 lwkt_sendmsg(netisr_cpuport(0), &nm->base.lmsg); 1542 } 1543 done: 1544 lwkt_replymsg(&msg->lmsg, 0); 1545 } 1546 1547 #ifdef INET6 1548 1549 void 1550 tcp6_ctlinput(netmsg_t msg) 1551 { 1552 int cmd = msg->ctlinput.nm_cmd; 1553 struct sockaddr *sa = msg->ctlinput.nm_arg; 1554 void *d = msg->ctlinput.nm_extra; 1555 struct tcphdr th; 1556 inp_notify_t notify = tcp_notify; 1557 struct ip6_hdr *ip6; 1558 struct mbuf *m; 1559 struct ip6ctlparam *ip6cp = NULL; 1560 const struct sockaddr_in6 *sa6_src = NULL; 1561 int off; 1562 struct tcp_portonly { 1563 u_int16_t th_sport; 1564 u_int16_t th_dport; 1565 } *thp; 1566 int arg; 1567 1568 if (sa->sa_family != AF_INET6 || 1569 sa->sa_len != sizeof(struct sockaddr_in6)) { 1570 goto out; 1571 } 1572 1573 arg = 0; 1574 if (cmd == PRC_QUENCH) 1575 notify = tcp_quench; 1576 else if (cmd == PRC_MSGSIZE) { 1577 struct ip6ctlparam *ip6cp = d; 1578 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6; 1579 1580 arg = ntohl(icmp6->icmp6_mtu); 1581 notify = tcp_mtudisc; 1582 } else if (!PRC_IS_REDIRECT(cmd) && 1583 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) { 1584 goto out; 1585 } 1586 1587 /* if the parameter is from icmp6, decode it. */ 1588 if (d != NULL) { 1589 ip6cp = (struct ip6ctlparam *)d; 1590 m = ip6cp->ip6c_m; 1591 ip6 = ip6cp->ip6c_ip6; 1592 off = ip6cp->ip6c_off; 1593 sa6_src = ip6cp->ip6c_src; 1594 } else { 1595 m = NULL; 1596 ip6 = NULL; 1597 off = 0; /* fool gcc */ 1598 sa6_src = &sa6_any; 1599 } 1600 1601 if (ip6 != NULL) { 1602 struct in_conninfo inc; 1603 /* 1604 * XXX: We assume that when IPV6 is non NULL, 1605 * M and OFF are valid. 1606 */ 1607 1608 /* check if we can safely examine src and dst ports */ 1609 if (m->m_pkthdr.len < off + sizeof *thp) 1610 goto out; 1611 1612 bzero(&th, sizeof th); 1613 m_copydata(m, off, sizeof *thp, (caddr_t)&th); 1614 1615 in6_pcbnotify(&tcbinfo[0], sa, th.th_dport, 1616 (struct sockaddr *)ip6cp->ip6c_src, 1617 th.th_sport, cmd, arg, notify); 1618 1619 inc.inc_fport = th.th_dport; 1620 inc.inc_lport = th.th_sport; 1621 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr; 1622 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr; 1623 inc.inc_isipv6 = 1; 1624 syncache_unreach(&inc, &th); 1625 } else { 1626 in6_pcbnotify(&tcbinfo[0], sa, 0, 1627 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify); 1628 } 1629 out: 1630 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0); 1631 } 1632 1633 #endif 1634 1635 /* 1636 * Following is where TCP initial sequence number generation occurs. 1637 * 1638 * There are two places where we must use initial sequence numbers: 1639 * 1. In SYN-ACK packets. 1640 * 2. In SYN packets. 1641 * 1642 * All ISNs for SYN-ACK packets are generated by the syncache. See 1643 * tcp_syncache.c for details. 1644 * 1645 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling 1646 * depends on this property. In addition, these ISNs should be 1647 * unguessable so as to prevent connection hijacking. To satisfy 1648 * the requirements of this situation, the algorithm outlined in 1649 * RFC 1948 is used to generate sequence numbers. 1650 * 1651 * Implementation details: 1652 * 1653 * net.inet.tcp.isn_reseed_interval controls the number of seconds 1654 * between the seeding of isn_secret. On every reseed we jump the 1655 * ISN by a lot. 1656 */ 1657 struct tcp_isn { 1658 u_char secret[16]; 1659 MD5_CTX ctx; 1660 int last_reseed; 1661 int last_offset; 1662 } __cachealign; 1663 1664 struct tcp_isn tcp_isn_ary[MAXCPU]; 1665 1666 tcp_seq 1667 tcp_new_isn(struct tcpcb *tp) 1668 { 1669 struct tcp_isn *isn; 1670 tcp_seq new_isn; 1671 tcp_seq digest[16 / sizeof(tcp_seq)]; 1672 int n; 1673 1674 isn = &tcp_isn_ary[mycpuid]; 1675 1676 /* 1677 * Reseed every 20 seconds. 6 reseeds per 2-minute interval in 1678 * order to retain our monotonic offset. 1679 * 1680 * The initial seed randomizes last_offset with all 32 bits. 1681 * 1682 * Note that the md5 digest is masked with 0x0FFFFFFF, so we must 1683 * add 1/16 of our full range (1/8 of our signed range) to ensure 1684 * monotonic operation. 1685 */ 1686 if (isn->last_reseed == 0 || 1687 (u_int)(ticks - isn->last_reseed) > tcp_isn_reseed_interval * hz) { 1688 if (isn->last_reseed == 0) { 1689 read_random(&isn->last_offset, 1690 sizeof(isn->last_offset), 1); 1691 } 1692 read_random(&isn->secret, sizeof(isn->secret), 1); 1693 isn->last_reseed = ticks; 1694 isn->last_offset += 0x10000000; 1695 } 1696 1697 /* 1698 * Compute the md5 hash, giving us a deterministic result for the 1699 * port/address pair for any given secret. 1700 */ 1701 MD5Init(&isn->ctx); 1702 MD5Update(&isn->ctx, isn->secret, sizeof(isn->secret)); 1703 MD5Update(&isn->ctx, (u_char *)&tp->t_inpcb->inp_fport, 2); 1704 MD5Update(&isn->ctx, (u_char *)&tp->t_inpcb->inp_lport, 2); 1705 #ifdef INET6 1706 if (INP_ISIPV6(tp->t_inpcb)) { 1707 MD5Update(&isn->ctx, (u_char *)&tp->t_inpcb->in6p_faddr, 1708 sizeof(struct in6_addr)); 1709 MD5Update(&isn->ctx, (u_char *)&tp->t_inpcb->in6p_laddr, 1710 sizeof(struct in6_addr)); 1711 } else 1712 #endif 1713 { 1714 MD5Update(&isn->ctx, (u_char *)&tp->t_inpcb->inp_faddr, 1715 sizeof(struct in_addr)); 1716 MD5Update(&isn->ctx, (u_char *)&tp->t_inpcb->inp_laddr, 1717 sizeof(struct in_addr)); 1718 } 1719 MD5Final((char *)digest, &isn->ctx); 1720 1721 /* 1722 * Add a random component 0-1048575 plus advance by 1048576. 1723 * 1724 * The sequence space is simply too small, in modern times we also 1725 * must depend on the receive-side being a bit smarter when recycling 1726 * ports in TIME_WAIT. 1727 */ 1728 read_random(&n, sizeof(n), 1); 1729 isn->last_offset += (n & 0x000FFFFF) + 0x00100000; 1730 new_isn = (digest[0] & 0x0FFFFFFF) + isn->last_offset; 1731 1732 return (new_isn); 1733 } 1734 1735 /* 1736 * When a source quench is received, close congestion window 1737 * to one segment. We will gradually open it again as we proceed. 1738 */ 1739 void 1740 tcp_quench(struct inpcb *inp, int error) 1741 { 1742 struct tcpcb *tp = intotcpcb(inp); 1743 1744 KASSERT(tp != NULL, ("tcp_quench: tp is NULL")); 1745 tp->snd_cwnd = tp->t_maxseg; 1746 tp->snd_wacked = 0; 1747 } 1748 1749 /* 1750 * When a specific ICMP unreachable message is received and the 1751 * connection state is SYN-SENT, drop the connection. This behavior 1752 * is controlled by the icmp_may_rst sysctl. 1753 */ 1754 void 1755 tcp_drop_syn_sent(struct inpcb *inp, int error) 1756 { 1757 struct tcpcb *tp = intotcpcb(inp); 1758 1759 KASSERT(tp != NULL, ("tcp_drop_syn_sent: tp is NULL")); 1760 if (tp->t_state == TCPS_SYN_SENT) 1761 tcp_drop(tp, error); 1762 } 1763 1764 /* 1765 * When a `need fragmentation' ICMP is received, update our idea of the MSS 1766 * based on the new value in the route. Also nudge TCP to send something, 1767 * since we know the packet we just sent was dropped. 1768 * This duplicates some code in the tcp_mss() function in tcp_input.c. 1769 */ 1770 void 1771 tcp_mtudisc(struct inpcb *inp, int mtu) 1772 { 1773 struct tcpcb *tp = intotcpcb(inp); 1774 struct rtentry *rt; 1775 struct socket *so = inp->inp_socket; 1776 int maxopd, mss; 1777 #ifdef INET6 1778 boolean_t isipv6 = INP_ISIPV6(inp); 1779 #else 1780 const boolean_t isipv6 = FALSE; 1781 #endif 1782 1783 KASSERT(tp != NULL, ("tcp_mtudisc: tp is NULL")); 1784 1785 /* 1786 * If no MTU is provided in the ICMP message, use the 1787 * next lower likely value, as specified in RFC 1191. 1788 */ 1789 if (mtu == 0) { 1790 int oldmtu; 1791 1792 oldmtu = tp->t_maxopd + 1793 (isipv6 ? 1794 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1795 sizeof(struct tcpiphdr)); 1796 mtu = ip_next_mtu(oldmtu, 0); 1797 } 1798 1799 if (isipv6) 1800 rt = tcp_rtlookup6(&inp->inp_inc); 1801 else 1802 rt = tcp_rtlookup(&inp->inp_inc); 1803 if (rt != NULL) { 1804 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu) 1805 mtu = rt->rt_rmx.rmx_mtu; 1806 1807 maxopd = mtu - 1808 (isipv6 ? 1809 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1810 sizeof(struct tcpiphdr)); 1811 1812 /* 1813 * XXX - The following conditional probably violates the TCP 1814 * spec. The problem is that, since we don't know the 1815 * other end's MSS, we are supposed to use a conservative 1816 * default. But, if we do that, then MTU discovery will 1817 * never actually take place, because the conservative 1818 * default is much less than the MTUs typically seen 1819 * on the Internet today. For the moment, we'll sweep 1820 * this under the carpet. 1821 * 1822 * The conservative default might not actually be a problem 1823 * if the only case this occurs is when sending an initial 1824 * SYN with options and data to a host we've never talked 1825 * to before. Then, they will reply with an MSS value which 1826 * will get recorded and the new parameters should get 1827 * recomputed. For Further Study. 1828 */ 1829 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd) 1830 maxopd = rt->rt_rmx.rmx_mssopt; 1831 } else 1832 maxopd = mtu - 1833 (isipv6 ? 1834 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) : 1835 sizeof(struct tcpiphdr)); 1836 1837 if (tp->t_maxopd <= maxopd) 1838 return; 1839 tp->t_maxopd = maxopd; 1840 1841 mss = maxopd; 1842 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) == 1843 (TF_REQ_TSTMP | TF_RCVD_TSTMP)) 1844 mss -= TCPOLEN_TSTAMP_APPA; 1845 1846 /* round down to multiple of MCLBYTES */ 1847 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */ 1848 if (mss > MCLBYTES) 1849 mss &= ~(MCLBYTES - 1); 1850 #else 1851 if (mss > MCLBYTES) 1852 mss = rounddown(mss, MCLBYTES); 1853 #endif 1854 1855 if (so->so_snd.ssb_hiwat < mss) 1856 mss = so->so_snd.ssb_hiwat; 1857 1858 tp->t_maxseg = mss; 1859 tp->t_rtttime = 0; 1860 tp->snd_nxt = tp->snd_una; 1861 tcp_output(tp); 1862 tcpstat.tcps_mturesent++; 1863 } 1864 1865 /* 1866 * Look-up the routing entry to the peer of this inpcb. If no route 1867 * is found and it cannot be allocated the return NULL. This routine 1868 * is called by TCP routines that access the rmx structure and by tcp_mss 1869 * to get the interface MTU. 1870 */ 1871 struct rtentry * 1872 tcp_rtlookup(struct in_conninfo *inc) 1873 { 1874 struct route *ro = &inc->inc_route; 1875 1876 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) { 1877 /* No route yet, so try to acquire one */ 1878 if (inc->inc_faddr.s_addr != INADDR_ANY) { 1879 /* 1880 * unused portions of the structure MUST be zero'd 1881 * out because rtalloc() treats it as opaque data 1882 */ 1883 bzero(&ro->ro_dst, sizeof(struct sockaddr_in)); 1884 ro->ro_dst.sa_family = AF_INET; 1885 ro->ro_dst.sa_len = sizeof(struct sockaddr_in); 1886 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr = 1887 inc->inc_faddr; 1888 rtalloc(ro); 1889 } 1890 } 1891 return (ro->ro_rt); 1892 } 1893 1894 #ifdef INET6 1895 struct rtentry * 1896 tcp_rtlookup6(struct in_conninfo *inc) 1897 { 1898 struct route_in6 *ro6 = &inc->inc6_route; 1899 1900 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) { 1901 /* No route yet, so try to acquire one */ 1902 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { 1903 /* 1904 * unused portions of the structure MUST be zero'd 1905 * out because rtalloc() treats it as opaque data 1906 */ 1907 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6)); 1908 ro6->ro_dst.sin6_family = AF_INET6; 1909 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6); 1910 ro6->ro_dst.sin6_addr = inc->inc6_faddr; 1911 rtalloc((struct route *)ro6); 1912 } 1913 } 1914 return (ro6->ro_rt); 1915 } 1916 #endif 1917 1918 /* 1919 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING 1920 * 1921 * This code attempts to calculate the bandwidth-delay product as a 1922 * means of determining the optimal window size to maximize bandwidth, 1923 * minimize RTT, and avoid the over-allocation of buffers on interfaces and 1924 * routers. This code also does a fairly good job keeping RTTs in check 1925 * across slow links like modems. We implement an algorithm which is very 1926 * similar (but not meant to be) TCP/Vegas. The code operates on the 1927 * transmitter side of a TCP connection and so only effects the transmit 1928 * side of the connection. 1929 * 1930 * BACKGROUND: TCP makes no provision for the management of buffer space 1931 * at the end points or at the intermediate routers and switches. A TCP 1932 * stream, whether using NewReno or not, will eventually buffer as 1933 * many packets as it is able and the only reason this typically works is 1934 * due to the fairly small default buffers made available for a connection 1935 * (typicaly 16K or 32K). As machines use larger windows and/or window 1936 * scaling it is now fairly easy for even a single TCP connection to blow-out 1937 * all available buffer space not only on the local interface, but on 1938 * intermediate routers and switches as well. NewReno makes a misguided 1939 * attempt to 'solve' this problem by waiting for an actual failure to occur, 1940 * then backing off, then steadily increasing the window again until another 1941 * failure occurs, ad-infinitum. This results in terrible oscillation that 1942 * is only made worse as network loads increase and the idea of intentionally 1943 * blowing out network buffers is, frankly, a terrible way to manage network 1944 * resources. 1945 * 1946 * It is far better to limit the transmit window prior to the failure 1947 * condition being achieved. There are two general ways to do this: First 1948 * you can 'scan' through different transmit window sizes and locate the 1949 * point where the RTT stops increasing, indicating that you have filled the 1950 * pipe, then scan backwards until you note that RTT stops decreasing, then 1951 * repeat ad-infinitum. This method works in principle but has severe 1952 * implementation issues due to RTT variances, timer granularity, and 1953 * instability in the algorithm which can lead to many false positives and 1954 * create oscillations as well as interact badly with other TCP streams 1955 * implementing the same algorithm. 1956 * 1957 * The second method is to limit the window to the bandwidth delay product 1958 * of the link. This is the method we implement. RTT variances and our 1959 * own manipulation of the congestion window, bwnd, can potentially 1960 * destabilize the algorithm. For this reason we have to stabilize the 1961 * elements used to calculate the window. We do this by using the minimum 1962 * observed RTT, the long term average of the observed bandwidth, and 1963 * by adding two segments worth of slop. It isn't perfect but it is able 1964 * to react to changing conditions and gives us a very stable basis on 1965 * which to extend the algorithm. 1966 */ 1967 void 1968 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq) 1969 { 1970 u_long bw; 1971 u_long ibw; 1972 u_long bwnd; 1973 int save_ticks; 1974 int delta_ticks; 1975 1976 /* 1977 * If inflight_enable is disabled in the middle of a tcp connection, 1978 * make sure snd_bwnd is effectively disabled. 1979 */ 1980 if (!tcp_inflight_enable) { 1981 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; 1982 tp->snd_bandwidth = 0; 1983 return; 1984 } 1985 1986 /* 1987 * Validate the delta time. If a connection is new or has been idle 1988 * a long time we have to reset the bandwidth calculator. 1989 */ 1990 save_ticks = ticks; 1991 cpu_ccfence(); 1992 delta_ticks = save_ticks - tp->t_bw_rtttime; 1993 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) { 1994 tp->t_bw_rtttime = save_ticks; 1995 tp->t_bw_rtseq = ack_seq; 1996 if (tp->snd_bandwidth == 0) 1997 tp->snd_bandwidth = tcp_inflight_start; 1998 return; 1999 } 2000 2001 /* 2002 * A delta of at least 1 tick is required. Waiting 2 ticks will 2003 * result in better (bw) accuracy. More than that and the ramp-up 2004 * will be too slow. 2005 */ 2006 if (delta_ticks == 0 || delta_ticks == 1) 2007 return; 2008 2009 /* 2010 * Sanity check, plus ignore pure window update acks. 2011 */ 2012 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0) 2013 return; 2014 2015 /* 2016 * Figure out the bandwidth. Due to the tick granularity this 2017 * is a very rough number and it MUST be averaged over a fairly 2018 * long period of time. XXX we need to take into account a link 2019 * that is not using all available bandwidth, but for now our 2020 * slop will ramp us up if this case occurs and the bandwidth later 2021 * increases. 2022 */ 2023 ibw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks; 2024 tp->t_bw_rtttime = save_ticks; 2025 tp->t_bw_rtseq = ack_seq; 2026 bw = ((int64_t)tp->snd_bandwidth * 15 + ibw) >> 4; 2027 2028 tp->snd_bandwidth = bw; 2029 2030 /* 2031 * Calculate the semi-static bandwidth delay product, plus two maximal 2032 * segments. The additional slop puts us squarely in the sweet 2033 * spot and also handles the bandwidth run-up case. Without the 2034 * slop we could be locking ourselves into a lower bandwidth. 2035 * 2036 * At very high speeds the bw calculation can become overly sensitive 2037 * and error prone when delta_ticks is low (e.g. usually 1). To deal 2038 * with the problem the stab must be scaled to the bw. A stab of 50 2039 * (the default) increases the bw for the purposes of the bwnd 2040 * calculation by 5%. 2041 * 2042 * Situations Handled: 2043 * (1) Prevents over-queueing of packets on LANs, especially on 2044 * high speed LANs, allowing larger TCP buffers to be 2045 * specified, and also does a good job preventing 2046 * over-queueing of packets over choke points like modems 2047 * (at least for the transmit side). 2048 * 2049 * (2) Is able to handle changing network loads (bandwidth 2050 * drops so bwnd drops, bandwidth increases so bwnd 2051 * increases). 2052 * 2053 * (3) Theoretically should stabilize in the face of multiple 2054 * connections implementing the same algorithm (this may need 2055 * a little work). 2056 * 2057 * (4) Stability value (defaults to 20 = 2 maximal packets) can 2058 * be adjusted with a sysctl but typically only needs to be on 2059 * very slow connections. A value no smaller then 5 should 2060 * be used, but only reduce this default if you have no other 2061 * choice. 2062 */ 2063 2064 #define USERTT ((tp->t_srtt + tp->t_rttvar) + tcp_inflight_adjrtt) 2065 bw += bw * tcp_inflight_stab / 1000; 2066 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + 2067 (int)tp->t_maxseg * 2; 2068 #undef USERTT 2069 2070 if (tcp_inflight_debug > 0) { 2071 static int ltime; 2072 if ((u_int)(save_ticks - ltime) >= hz / tcp_inflight_debug) { 2073 ltime = save_ticks; 2074 kprintf("%p ibw %ld bw %ld rttvar %d srtt %d " 2075 "bwnd %ld delta %d snd_win %ld\n", 2076 tp, ibw, bw, tp->t_rttvar, tp->t_srtt, 2077 bwnd, delta_ticks, tp->snd_wnd); 2078 } 2079 } 2080 if ((long)bwnd < tcp_inflight_min) 2081 bwnd = tcp_inflight_min; 2082 if (bwnd > tcp_inflight_max) 2083 bwnd = tcp_inflight_max; 2084 if ((long)bwnd < tp->t_maxseg * 2) 2085 bwnd = tp->t_maxseg * 2; 2086 tp->snd_bwnd = bwnd; 2087 } 2088 2089 static void 2090 tcp_rmx_iwsegs(struct tcpcb *tp, u_long *maxsegs, u_long *capsegs) 2091 { 2092 struct rtentry *rt; 2093 struct inpcb *inp = tp->t_inpcb; 2094 #ifdef INET6 2095 boolean_t isipv6 = INP_ISIPV6(inp); 2096 #else 2097 const boolean_t isipv6 = FALSE; 2098 #endif 2099 2100 /* XXX */ 2101 if (tcp_iw_maxsegs < TCP_IW_MAXSEGS_DFLT) 2102 tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT; 2103 if (tcp_iw_capsegs < TCP_IW_CAPSEGS_DFLT) 2104 tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT; 2105 2106 if (isipv6) 2107 rt = tcp_rtlookup6(&inp->inp_inc); 2108 else 2109 rt = tcp_rtlookup(&inp->inp_inc); 2110 if (rt == NULL || 2111 rt->rt_rmx.rmx_iwmaxsegs < TCP_IW_MAXSEGS_DFLT || 2112 rt->rt_rmx.rmx_iwcapsegs < TCP_IW_CAPSEGS_DFLT) { 2113 *maxsegs = tcp_iw_maxsegs; 2114 *capsegs = tcp_iw_capsegs; 2115 return; 2116 } 2117 *maxsegs = rt->rt_rmx.rmx_iwmaxsegs; 2118 *capsegs = rt->rt_rmx.rmx_iwcapsegs; 2119 } 2120 2121 u_long 2122 tcp_initial_window(struct tcpcb *tp) 2123 { 2124 if (tcp_do_rfc3390) { 2125 /* 2126 * RFC3390: 2127 * "If the SYN or SYN/ACK is lost, the initial window 2128 * used by a sender after a correctly transmitted SYN 2129 * MUST be one segment consisting of MSS bytes." 2130 * 2131 * However, we do something a little bit more aggressive 2132 * then RFC3390 here: 2133 * - Only if time spent in the SYN or SYN|ACK retransmition 2134 * >= 3 seconds, the IW is reduced. We do this mainly 2135 * because when RFC3390 is published, the initial RTO is 2136 * still 3 seconds (the threshold we test here), while 2137 * after RFC6298, the initial RTO is 1 second. This 2138 * behaviour probably still falls within the spirit of 2139 * RFC3390. 2140 * - When IW is reduced, 2*MSS is used instead of 1*MSS. 2141 * Mainly to avoid sender and receiver deadlock until 2142 * delayed ACK timer expires. And even RFC2581 does not 2143 * try to reduce IW upon SYN or SYN|ACK retransmition 2144 * timeout. 2145 * 2146 * See also: 2147 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03 2148 */ 2149 if (tp->t_rxtsyn >= TCPTV_RTOBASE3) { 2150 return (2 * tp->t_maxseg); 2151 } else { 2152 u_long maxsegs, capsegs; 2153 2154 tcp_rmx_iwsegs(tp, &maxsegs, &capsegs); 2155 return min(maxsegs * tp->t_maxseg, 2156 max(2 * tp->t_maxseg, capsegs * 1460)); 2157 } 2158 } else { 2159 /* 2160 * Even RFC2581 (back to 1999) allows 2*SMSS IW. 2161 * 2162 * Mainly to avoid sender and receiver deadlock 2163 * until delayed ACK timer expires. 2164 */ 2165 return (2 * tp->t_maxseg); 2166 } 2167 } 2168 2169 #ifdef TCP_SIGNATURE 2170 /* 2171 * Compute TCP-MD5 hash of a TCP segment. (RFC2385) 2172 * 2173 * We do this over ip, tcphdr, segment data, and the key in the SADB. 2174 * When called from tcp_input(), we can be sure that th_sum has been 2175 * zeroed out and verified already. 2176 * 2177 * Return 0 if successful, otherwise return -1. 2178 * 2179 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a 2180 * search with the destination IP address, and a 'magic SPI' to be 2181 * determined by the application. This is hardcoded elsewhere to 1179 2182 * right now. Another branch of this code exists which uses the SPD to 2183 * specify per-application flows but it is unstable. 2184 */ 2185 int 2186 tcpsignature_compute( 2187 struct mbuf *m, /* mbuf chain */ 2188 int len, /* length of TCP data */ 2189 int optlen, /* length of TCP options */ 2190 u_char *buf, /* storage for MD5 digest */ 2191 u_int direction) /* direction of flow */ 2192 { 2193 struct ippseudo ippseudo; 2194 MD5_CTX ctx; 2195 int doff; 2196 struct ip *ip; 2197 struct ipovly *ipovly; 2198 struct secasvar *sav; 2199 struct tcphdr *th; 2200 #ifdef INET6 2201 struct ip6_hdr *ip6; 2202 struct in6_addr in6; 2203 uint32_t plen; 2204 uint16_t nhdr; 2205 #endif /* INET6 */ 2206 u_short savecsum; 2207 2208 KASSERT(m != NULL, ("passed NULL mbuf. Game over.")); 2209 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature")); 2210 /* 2211 * Extract the destination from the IP header in the mbuf. 2212 */ 2213 ip = mtod(m, struct ip *); 2214 #ifdef INET6 2215 ip6 = NULL; /* Make the compiler happy. */ 2216 #endif /* INET6 */ 2217 /* 2218 * Look up an SADB entry which matches the address found in 2219 * the segment. 2220 */ 2221 switch (IP_VHL_V(ip->ip_vhl)) { 2222 case IPVERSION: 2223 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst, 2224 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2225 break; 2226 #ifdef INET6 2227 case (IPV6_VERSION >> 4): 2228 ip6 = mtod(m, struct ip6_hdr *); 2229 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst, 2230 IPPROTO_TCP, htonl(TCP_SIG_SPI)); 2231 break; 2232 #endif /* INET6 */ 2233 default: 2234 return (EINVAL); 2235 /* NOTREACHED */ 2236 break; 2237 } 2238 if (sav == NULL) { 2239 kprintf("%s: SADB lookup failed\n", __func__); 2240 return (EINVAL); 2241 } 2242 MD5Init(&ctx); 2243 2244 /* 2245 * Step 1: Update MD5 hash with IP pseudo-header. 2246 * 2247 * XXX The ippseudo header MUST be digested in network byte order, 2248 * or else we'll fail the regression test. Assume all fields we've 2249 * been doing arithmetic on have been in host byte order. 2250 * XXX One cannot depend on ipovly->ih_len here. When called from 2251 * tcp_output(), the underlying ip_len member has not yet been set. 2252 */ 2253 switch (IP_VHL_V(ip->ip_vhl)) { 2254 case IPVERSION: 2255 ipovly = (struct ipovly *)ip; 2256 ippseudo.ippseudo_src = ipovly->ih_src; 2257 ippseudo.ippseudo_dst = ipovly->ih_dst; 2258 ippseudo.ippseudo_pad = 0; 2259 ippseudo.ippseudo_p = IPPROTO_TCP; 2260 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen); 2261 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo)); 2262 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip)); 2263 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen; 2264 break; 2265 #ifdef INET6 2266 /* 2267 * RFC 2385, 2.0 Proposal 2268 * For IPv6, the pseudo-header is as described in RFC 2460, namely the 2269 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero- 2270 * extended next header value (to form 32 bits), and 32-bit segment 2271 * length. 2272 * Note: Upper-Layer Packet Length comes before Next Header. 2273 */ 2274 case (IPV6_VERSION >> 4): 2275 in6 = ip6->ip6_src; 2276 in6_clearscope(&in6); 2277 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2278 in6 = ip6->ip6_dst; 2279 in6_clearscope(&in6); 2280 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr)); 2281 plen = htonl(len + sizeof(struct tcphdr) + optlen); 2282 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t)); 2283 nhdr = 0; 2284 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2285 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2286 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2287 nhdr = IPPROTO_TCP; 2288 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t)); 2289 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr)); 2290 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen; 2291 break; 2292 #endif /* INET6 */ 2293 default: 2294 return (EINVAL); 2295 /* NOTREACHED */ 2296 break; 2297 } 2298 /* 2299 * Step 2: Update MD5 hash with TCP header, excluding options. 2300 * The TCP checksum must be set to zero. 2301 */ 2302 savecsum = th->th_sum; 2303 th->th_sum = 0; 2304 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr)); 2305 th->th_sum = savecsum; 2306 /* 2307 * Step 3: Update MD5 hash with TCP segment data. 2308 * Use m_apply() to avoid an early m_pullup(). 2309 */ 2310 if (len > 0) 2311 m_apply(m, doff, len, tcpsignature_apply, &ctx); 2312 /* 2313 * Step 4: Update MD5 hash with shared secret. 2314 */ 2315 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth)); 2316 MD5Final(buf, &ctx); 2317 key_sa_recordxfer(sav, m); 2318 key_freesav(sav); 2319 return (0); 2320 } 2321 2322 int 2323 tcpsignature_apply(void *fstate, void *data, unsigned int len) 2324 { 2325 2326 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len); 2327 return (0); 2328 } 2329 #endif /* TCP_SIGNATURE */ 2330 2331 static void 2332 tcp_drop_sysctl_dispatch(netmsg_t nmsg) 2333 { 2334 struct lwkt_msg *lmsg = &nmsg->lmsg; 2335 /* addrs[0] is a foreign socket, addrs[1] is a local one. */ 2336 struct sockaddr_storage *addrs = lmsg->u.ms_resultp; 2337 int error; 2338 struct sockaddr_in *fin, *lin; 2339 #ifdef INET6 2340 struct sockaddr_in6 *fin6, *lin6; 2341 struct in6_addr f6, l6; 2342 #endif 2343 struct inpcb *inp; 2344 2345 switch (addrs[0].ss_family) { 2346 #ifdef INET6 2347 case AF_INET6: 2348 fin6 = (struct sockaddr_in6 *)&addrs[0]; 2349 lin6 = (struct sockaddr_in6 *)&addrs[1]; 2350 error = in6_embedscope(&f6, fin6, NULL, NULL); 2351 if (error) 2352 goto done; 2353 error = in6_embedscope(&l6, lin6, NULL, NULL); 2354 if (error) 2355 goto done; 2356 inp = in6_pcblookup_hash(&tcbinfo[mycpuid], &f6, 2357 fin6->sin6_port, &l6, lin6->sin6_port, FALSE, NULL); 2358 break; 2359 #endif 2360 #ifdef INET 2361 case AF_INET: 2362 fin = (struct sockaddr_in *)&addrs[0]; 2363 lin = (struct sockaddr_in *)&addrs[1]; 2364 inp = in_pcblookup_hash(&tcbinfo[mycpuid], fin->sin_addr, 2365 fin->sin_port, lin->sin_addr, lin->sin_port, FALSE, NULL); 2366 break; 2367 #endif 2368 default: 2369 /* 2370 * Must not reach here, since the address family was 2371 * checked in sysctl handler. 2372 */ 2373 panic("unknown address family %d", addrs[0].ss_family); 2374 } 2375 if (inp != NULL) { 2376 struct tcpcb *tp = intotcpcb(inp); 2377 2378 KASSERT((inp->inp_flags & INP_WILDCARD) == 0, 2379 ("in wildcard hash")); 2380 KASSERT(tp != NULL, ("tcp_drop_sysctl_dispatch: tp is NULL")); 2381 KASSERT((tp->t_flags & TF_LISTEN) == 0, ("listen socket")); 2382 tcp_drop(tp, ECONNABORTED); 2383 error = 0; 2384 } else { 2385 error = ESRCH; 2386 } 2387 #ifdef INET6 2388 done: 2389 #endif 2390 lwkt_replymsg(lmsg, error); 2391 } 2392 2393 static int 2394 sysctl_tcp_drop(SYSCTL_HANDLER_ARGS) 2395 { 2396 /* addrs[0] is a foreign socket, addrs[1] is a local one. */ 2397 struct sockaddr_storage addrs[2]; 2398 struct sockaddr_in *fin, *lin; 2399 #ifdef INET6 2400 struct sockaddr_in6 *fin6, *lin6; 2401 #endif 2402 struct netmsg_base nmsg; 2403 struct lwkt_msg *lmsg = &nmsg.lmsg; 2404 struct lwkt_port *port = NULL; 2405 int error; 2406 2407 fin = lin = NULL; 2408 #ifdef INET6 2409 fin6 = lin6 = NULL; 2410 #endif 2411 error = 0; 2412 2413 if (req->oldptr != NULL || req->oldlen != 0) 2414 return (EINVAL); 2415 if (req->newptr == NULL) 2416 return (EPERM); 2417 if (req->newlen < sizeof(addrs)) 2418 return (ENOMEM); 2419 error = SYSCTL_IN(req, &addrs, sizeof(addrs)); 2420 if (error) 2421 return (error); 2422 2423 switch (addrs[0].ss_family) { 2424 #ifdef INET6 2425 case AF_INET6: 2426 fin6 = (struct sockaddr_in6 *)&addrs[0]; 2427 lin6 = (struct sockaddr_in6 *)&addrs[1]; 2428 if (fin6->sin6_len != sizeof(struct sockaddr_in6) || 2429 lin6->sin6_len != sizeof(struct sockaddr_in6)) 2430 return (EINVAL); 2431 if (IN6_IS_ADDR_V4MAPPED(&fin6->sin6_addr) || 2432 IN6_IS_ADDR_V4MAPPED(&lin6->sin6_addr)) 2433 return (EADDRNOTAVAIL); 2434 #if 0 2435 error = sa6_embedscope(fin6, V_ip6_use_defzone); 2436 if (error) 2437 return (error); 2438 error = sa6_embedscope(lin6, V_ip6_use_defzone); 2439 if (error) 2440 return (error); 2441 #endif 2442 port = tcp6_addrport(); 2443 break; 2444 #endif 2445 #ifdef INET 2446 case AF_INET: 2447 fin = (struct sockaddr_in *)&addrs[0]; 2448 lin = (struct sockaddr_in *)&addrs[1]; 2449 if (fin->sin_len != sizeof(struct sockaddr_in) || 2450 lin->sin_len != sizeof(struct sockaddr_in)) 2451 return (EINVAL); 2452 port = tcp_addrport(fin->sin_addr.s_addr, fin->sin_port, 2453 lin->sin_addr.s_addr, lin->sin_port); 2454 break; 2455 #endif 2456 default: 2457 return (EINVAL); 2458 } 2459 2460 netmsg_init(&nmsg, NULL, &curthread->td_msgport, 0, 2461 tcp_drop_sysctl_dispatch); 2462 lmsg->u.ms_resultp = addrs; 2463 return lwkt_domsg(port, lmsg, 0); 2464 } 2465 2466 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, drop, 2467 CTLTYPE_STRUCT | CTLFLAG_WR | CTLFLAG_SKIP, NULL, 2468 0, sysctl_tcp_drop, "", "Drop TCP connection"); 2469 2470 static int 2471 sysctl_tcps_count(SYSCTL_HANDLER_ARGS) 2472 { 2473 u_long state_count[TCP_NSTATES]; 2474 int cpu; 2475 2476 memset(state_count, 0, sizeof(state_count)); 2477 for (cpu = 0; cpu < netisr_ncpus; ++cpu) { 2478 int i; 2479 2480 for (i = 0; i < TCP_NSTATES; ++i) 2481 state_count[i] += tcpstate_count[cpu].tcps_count[i]; 2482 } 2483 2484 return sysctl_handle_opaque(oidp, state_count, sizeof(state_count), req); 2485 } 2486 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, state_count, 2487 CTLTYPE_OPAQUE | CTLFLAG_RD, NULL, 0, 2488 sysctl_tcps_count, "LU", "TCP connection counts by state"); 2489 2490 void 2491 tcp_pcbport_create(struct tcpcb *tp) 2492 { 2493 int cpu; 2494 2495 KASSERT((tp->t_flags & TF_LISTEN) && tp->t_state == TCPS_LISTEN, 2496 ("not a listen tcpcb")); 2497 2498 KASSERT(tp->t_pcbport == NULL, ("tcpcb port cache was created")); 2499 tp->t_pcbport = 2500 kmalloc(sizeof(struct tcp_pcbport) * netisr_ncpus, 2501 M_PCB, 2502 M_WAITOK | M_CACHEALIGN); 2503 2504 for (cpu = 0; cpu < netisr_ncpus; ++cpu) { 2505 struct inpcbport *phd; 2506 2507 phd = &tp->t_pcbport[cpu].t_phd; 2508 LIST_INIT(&phd->phd_pcblist); 2509 /* Though, not used ... */ 2510 phd->phd_port = tp->t_inpcb->inp_lport; 2511 } 2512 } 2513 2514 void 2515 tcp_pcbport_merge_oncpu(struct tcpcb *tp) 2516 { 2517 struct inpcbport *phd; 2518 struct inpcb *inp; 2519 int cpu = mycpuid; 2520 2521 KASSERT(cpu < netisr_ncpus, ("invalid cpu%d", cpu)); 2522 phd = &tp->t_pcbport[cpu].t_phd; 2523 2524 while ((inp = LIST_FIRST(&phd->phd_pcblist)) != NULL) { 2525 KASSERT(inp->inp_phd == phd && inp->inp_porthash == NULL, 2526 ("not on tcpcb port cache")); 2527 LIST_REMOVE(inp, inp_portlist); 2528 in_pcbinsporthash_lport(inp); 2529 KASSERT(inp->inp_phd == tp->t_inpcb->inp_phd && 2530 inp->inp_porthash == tp->t_inpcb->inp_porthash, 2531 ("tcpcb port cache merge failed")); 2532 } 2533 } 2534 2535 void 2536 tcp_pcbport_destroy(struct tcpcb *tp) 2537 { 2538 #ifdef INVARIANTS 2539 int cpu; 2540 2541 for (cpu = 0; cpu < netisr_ncpus; ++cpu) { 2542 KASSERT(LIST_EMPTY(&tp->t_pcbport[cpu].t_phd.phd_pcblist), 2543 ("tcpcb port cache is not empty")); 2544 } 2545 #endif 2546 kfree(tp->t_pcbport, M_PCB); 2547 tp->t_pcbport = NULL; 2548 } 2549