1 /*- 2 * Copyright (c) 2016-2020 Netflix, Inc. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. Redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution. 12 * 13 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 16 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 23 * SUCH DAMAGE. 24 * 25 */ 26 /* 27 * Author: Randall Stewart <rrs@netflix.com> 28 * This work is based on the ACM Queue paper 29 * BBR - Congestion Based Congestion Control 30 * and also numerous discussions with Neal, Yuchung and Van. 31 */ 32 33 #include <sys/cdefs.h> 34 __FBSDID("$FreeBSD$"); 35 36 #include "opt_inet.h" 37 #include "opt_inet6.h" 38 #include "opt_ipsec.h" 39 #include "opt_tcpdebug.h" 40 #include "opt_ratelimit.h" 41 #include "opt_kern_tls.h" 42 #include <sys/param.h> 43 #include <sys/arb.h> 44 #include <sys/module.h> 45 #include <sys/kernel.h> 46 #ifdef TCP_HHOOK 47 #include <sys/hhook.h> 48 #endif 49 #include <sys/malloc.h> 50 #include <sys/mbuf.h> 51 #include <sys/proc.h> 52 #include <sys/qmath.h> 53 #include <sys/socket.h> 54 #include <sys/socketvar.h> 55 #ifdef KERN_TLS 56 #include <sys/ktls.h> 57 #endif 58 #include <sys/sysctl.h> 59 #include <sys/systm.h> 60 #include <sys/tree.h> 61 #ifdef NETFLIX_STATS 62 #include <sys/stats.h> /* Must come after qmath.h and tree.h */ 63 #endif 64 #include <sys/refcount.h> 65 #include <sys/queue.h> 66 #include <sys/smp.h> 67 #include <sys/kthread.h> 68 #include <sys/lock.h> 69 #include <sys/mutex.h> 70 #include <sys/tim_filter.h> 71 #include <sys/time.h> 72 #include <vm/uma.h> 73 #include <sys/kern_prefetch.h> 74 75 #include <net/route.h> 76 #include <net/vnet.h> 77 #include <net/ethernet.h> 78 #include <net/bpf.h> 79 80 #define TCPSTATES /* for logging */ 81 82 #include <netinet/in.h> 83 #include <netinet/in_kdtrace.h> 84 #include <netinet/in_pcb.h> 85 #include <netinet/ip.h> 86 #include <netinet/ip_icmp.h> /* required for icmp_var.h */ 87 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */ 88 #include <netinet/ip_var.h> 89 #include <netinet/ip6.h> 90 #include <netinet6/in6_pcb.h> 91 #include <netinet6/ip6_var.h> 92 #include <netinet/tcp.h> 93 #include <netinet/tcp_fsm.h> 94 #include <netinet/tcp_seq.h> 95 #include <netinet/tcp_timer.h> 96 #include <netinet/tcp_var.h> 97 #include <netinet/tcpip.h> 98 #include <netinet/tcp_hpts.h> 99 #include <netinet/tcp_lro.h> 100 #include <netinet/cc/cc.h> 101 #include <netinet/tcp_log_buf.h> 102 #ifdef TCPDEBUG 103 #include <netinet/tcp_debug.h> 104 #endif /* TCPDEBUG */ 105 #ifdef TCP_OFFLOAD 106 #include <netinet/tcp_offload.h> 107 #endif 108 #ifdef INET6 109 #include <netinet6/tcp6_var.h> 110 #endif 111 #include <netinet/tcp_fastopen.h> 112 113 #include <netipsec/ipsec_support.h> 114 #include <net/if.h> 115 #include <net/if_var.h> 116 117 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 118 #include <netipsec/ipsec.h> 119 #include <netipsec/ipsec6.h> 120 #endif /* IPSEC */ 121 122 #include <netinet/udp.h> 123 #include <netinet/udp_var.h> 124 #include <machine/in_cksum.h> 125 126 #ifdef MAC 127 #include <security/mac/mac_framework.h> 128 #endif 129 #include "rack_bbr_common.h" 130 131 /* 132 * Common TCP Functions - These are shared by borth 133 * rack and BBR. 134 */ 135 #ifdef KERN_TLS 136 uint32_t 137 ctf_get_opt_tls_size(struct socket *so, uint32_t rwnd) 138 { 139 struct ktls_session *tls; 140 uint32_t len; 141 142 again: 143 tls = so->so_snd.sb_tls_info; 144 len = tls->params.max_frame_len; /* max tls payload */ 145 len += tls->params.tls_hlen; /* tls header len */ 146 len += tls->params.tls_tlen; /* tls trailer len */ 147 if ((len * 4) > rwnd) { 148 /* 149 * Stroke this will suck counter and what 150 * else should we do Drew? From the 151 * TCP perspective I am not sure 152 * what should be done... 153 */ 154 if (tls->params.max_frame_len > 4096) { 155 tls->params.max_frame_len -= 4096; 156 if (tls->params.max_frame_len < 4096) 157 tls->params.max_frame_len = 4096; 158 goto again; 159 } 160 } 161 return (len); 162 } 163 #endif 164 165 static int 166 ctf_get_enet_type(struct ifnet *ifp, struct mbuf *m) 167 { 168 struct ether_header *eh; 169 #ifdef INET6 170 struct ip6_hdr *ip6 = NULL; /* Keep compiler happy. */ 171 #endif 172 #ifdef INET 173 struct ip *ip = NULL; /* Keep compiler happy. */ 174 #endif 175 #if defined(INET) || defined(INET6) 176 struct tcphdr *th; 177 int32_t tlen; 178 uint16_t drop_hdrlen; 179 #endif 180 uint16_t etype; 181 #ifdef INET 182 uint8_t iptos; 183 #endif 184 185 /* Is it the easy way? */ 186 if (m->m_flags & M_LRO_EHDRSTRP) 187 return (m->m_pkthdr.lro_etype); 188 /* 189 * Ok this is the old style call, the ethernet header is here. 190 * This also means no checksum or BPF were done. This 191 * can happen if the race to setup the inp fails and 192 * LRO sees no INP at packet input, but by the time 193 * we queue the packets an INP gets there. Its rare 194 * but it can occur so we will handle it. Note that 195 * this means duplicated work but with the rarity of it 196 * its not worth worrying about. 197 */ 198 /* Let the BPF see the packet */ 199 if (bpf_peers_present(ifp->if_bpf)) 200 ETHER_BPF_MTAP(ifp, m); 201 /* Now the csum */ 202 eh = mtod(m, struct ether_header *); 203 etype = ntohs(eh->ether_type); 204 m_adj(m, sizeof(*eh)); 205 switch (etype) { 206 #ifdef INET6 207 case ETHERTYPE_IPV6: 208 { 209 if (m->m_len < (sizeof(*ip6) + sizeof(*th))) { 210 m = m_pullup(m, sizeof(*ip6) + sizeof(*th)); 211 if (m == NULL) { 212 KMOD_TCPSTAT_INC(tcps_rcvshort); 213 return (-1); 214 } 215 } 216 ip6 = (struct ip6_hdr *)(eh + 1); 217 th = (struct tcphdr *)(ip6 + 1); 218 drop_hdrlen = sizeof(*ip6); 219 tlen = ntohs(ip6->ip6_plen); 220 if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID_IPV6) { 221 if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) 222 th->th_sum = m->m_pkthdr.csum_data; 223 else 224 th->th_sum = in6_cksum_pseudo(ip6, tlen, 225 IPPROTO_TCP, 226 m->m_pkthdr.csum_data); 227 th->th_sum ^= 0xffff; 228 } else 229 th->th_sum = in6_cksum(m, IPPROTO_TCP, drop_hdrlen, tlen); 230 if (th->th_sum) { 231 KMOD_TCPSTAT_INC(tcps_rcvbadsum); 232 m_freem(m); 233 return (-1); 234 } 235 return (etype); 236 } 237 #endif 238 #ifdef INET 239 case ETHERTYPE_IP: 240 { 241 if (m->m_len < sizeof (struct tcpiphdr)) { 242 m = m_pullup(m, sizeof (struct tcpiphdr)); 243 if (m == NULL) { 244 KMOD_TCPSTAT_INC(tcps_rcvshort); 245 return (-1); 246 } 247 } 248 ip = (struct ip *)(eh + 1); 249 th = (struct tcphdr *)(ip + 1); 250 drop_hdrlen = sizeof(*ip); 251 iptos = ip->ip_tos; 252 tlen = ntohs(ip->ip_len) - sizeof(struct ip); 253 if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) { 254 if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) 255 th->th_sum = m->m_pkthdr.csum_data; 256 else 257 th->th_sum = in_pseudo(ip->ip_src.s_addr, 258 ip->ip_dst.s_addr, 259 htonl(m->m_pkthdr.csum_data + tlen + IPPROTO_TCP)); 260 th->th_sum ^= 0xffff; 261 } else { 262 int len; 263 struct ipovly *ipov = (struct ipovly *)ip; 264 /* 265 * Checksum extended TCP header and data. 266 */ 267 len = drop_hdrlen + tlen; 268 bzero(ipov->ih_x1, sizeof(ipov->ih_x1)); 269 ipov->ih_len = htons(tlen); 270 th->th_sum = in_cksum(m, len); 271 /* Reset length for SDT probes. */ 272 ip->ip_len = htons(len); 273 /* Reset TOS bits */ 274 ip->ip_tos = iptos; 275 /* Re-initialization for later version check */ 276 ip->ip_v = IPVERSION; 277 ip->ip_hl = sizeof(*ip) >> 2; 278 } 279 if (th->th_sum) { 280 KMOD_TCPSTAT_INC(tcps_rcvbadsum); 281 m_freem(m); 282 return (-1); 283 } 284 break; 285 } 286 #endif 287 }; 288 return (etype); 289 } 290 291 /* 292 * The function ctf_process_inbound_raw() is used by 293 * transport developers to do the steps needed to 294 * support MBUF Queuing i.e. the flags in 295 * inp->inp_flags2: 296 * 297 * - INP_SUPPORTS_MBUFQ 298 * - INP_MBUF_QUEUE_READY 299 * - INP_DONT_SACK_QUEUE 300 * - INP_MBUF_ACKCMP 301 * 302 * These flags help control how LRO will deliver 303 * packets to the transport. You first set in inp_flags2 304 * the INP_SUPPORTS_MBUFQ to tell the LRO code that you 305 * will gladly take a queue of packets instead of a compressed 306 * single packet. You also set in your t_fb pointer the 307 * tfb_do_queued_segments to point to ctf_process_inbound_raw. 308 * 309 * This then gets you lists of inbound ACK's/Data instead 310 * of a condensed compressed ACK/DATA packet. Why would you 311 * want that? This will get you access to all the arrival 312 * times of at least LRO and possibly at the Hardware (if 313 * the interface card supports that) of the actual ACK/DATA. 314 * In some transport designs this is important since knowing 315 * the actual time we got the packet is useful information. 316 * 317 * A new special type of mbuf may also be supported by the transport 318 * if it has set the INP_MBUF_ACKCMP flag. If its set, LRO will 319 * possibly create a M_ACKCMP type mbuf. This is a mbuf with 320 * an array of "acks". One thing also to note is that when this 321 * occurs a subsequent LRO may find at the back of the untouched 322 * mbuf queue chain a M_ACKCMP and append on to it. This means 323 * that until the transport pulls in the mbuf chain queued 324 * for it more ack's may get on the mbufs that were already 325 * delivered. There currently is a limit of 6 acks condensed 326 * into 1 mbuf which means often when this is occuring, we 327 * don't get that effect but it does happen. 328 * 329 * Now there are some interesting Caveats that the transport 330 * designer needs to take into account when using this feature. 331 * 332 * 1) It is used with HPTS and pacing, when the pacing timer 333 * for output calls it will first call the input. 334 * 2) When you set INP_MBUF_QUEUE_READY this tells LRO 335 * queue normal packets, I am busy pacing out data and 336 * will process the queued packets before my tfb_tcp_output 337 * call from pacing. If a non-normal packet arrives, (e.g. sack) 338 * you will be awoken immediately. 339 * 3) Finally you can add the INP_DONT_SACK_QUEUE to not even 340 * be awoken if a SACK has arrived. You would do this when 341 * you were not only running a pacing for output timer 342 * but a Rack timer as well i.e. you know you are in recovery 343 * and are in the process (via the timers) of dealing with 344 * the loss. 345 * 346 * Now a critical thing you must be aware of here is that the 347 * use of the flags has a far greater scope then just your 348 * typical LRO. Why? Well thats because in the normal compressed 349 * LRO case at the end of a driver interupt all packets are going 350 * to get presented to the transport no matter if there is one 351 * or 100. With the MBUF_QUEUE model, this is not true. You will 352 * only be awoken to process the queue of packets when: 353 * a) The flags discussed above allow it. 354 * <or> 355 * b) You exceed a ack or data limit (by default the 356 * ack limit is infinity (64k acks) and the data 357 * limit is 64k of new TCP data) 358 * <or> 359 * c) The push bit has been set by the peer 360 */ 361 362 int 363 ctf_process_inbound_raw(struct tcpcb *tp, struct socket *so, struct mbuf *m, int has_pkt) 364 { 365 /* 366 * We are passed a raw change of mbuf packets 367 * that arrived in LRO. They are linked via 368 * the m_nextpkt link in the pkt-headers. 369 * 370 * We process each one by: 371 * a) saving off the next 372 * b) stripping off the ether-header 373 * c) formulating the arguments for 374 * the tfb_tcp_hpts_do_segment 375 * d) calling each mbuf to tfb_tcp_hpts_do_segment 376 * after adjusting the time to match the arrival time. 377 * Note that the LRO code assures no IP options are present. 378 * 379 * The symantics for calling tfb_tcp_hpts_do_segment are the 380 * following: 381 * 1) It returns 0 if all went well and you (the caller) need 382 * to release the lock. 383 * 2) If nxt_pkt is set, then the function will surpress calls 384 * to tcp_output() since you are promising to call again 385 * with another packet. 386 * 3) If it returns 1, then you must free all the packets being 387 * shipped in, the tcb has been destroyed (or about to be destroyed). 388 */ 389 struct mbuf *m_save; 390 struct tcphdr *th; 391 #ifdef INET6 392 struct ip6_hdr *ip6 = NULL; /* Keep compiler happy. */ 393 #endif 394 #ifdef INET 395 struct ip *ip = NULL; /* Keep compiler happy. */ 396 #endif 397 struct ifnet *ifp; 398 struct timeval tv; 399 struct inpcb *inp __diagused; 400 int32_t retval, nxt_pkt, tlen, off; 401 int etype = 0; 402 uint16_t drop_hdrlen; 403 uint8_t iptos, no_vn=0; 404 405 NET_EPOCH_ASSERT(); 406 if (m) 407 ifp = m_rcvif(m); 408 else 409 ifp = NULL; 410 if (ifp == NULL) { 411 /* 412 * We probably should not work around 413 * but kassert, since lro alwasy sets rcvif. 414 */ 415 no_vn = 1; 416 goto skip_vnet; 417 } 418 CURVNET_SET(ifp->if_vnet); 419 skip_vnet: 420 tcp_get_usecs(&tv); 421 while (m) { 422 m_save = m->m_nextpkt; 423 m->m_nextpkt = NULL; 424 if ((m->m_flags & M_ACKCMP) == 0) { 425 /* Now lets get the ether header */ 426 etype = ctf_get_enet_type(ifp, m); 427 if (etype == -1) { 428 /* Skip this packet it was freed by checksum */ 429 goto skipped_pkt; 430 } 431 KASSERT(((etype == ETHERTYPE_IPV6) || (etype == ETHERTYPE_IP)), 432 ("tp:%p m:%p etype:0x%x -- not IP or IPv6", tp, m, etype)); 433 /* Trim off the ethernet header */ 434 switch (etype) { 435 #ifdef INET6 436 case ETHERTYPE_IPV6: 437 ip6 = mtod(m, struct ip6_hdr *); 438 th = (struct tcphdr *)(ip6 + 1); 439 tlen = ntohs(ip6->ip6_plen); 440 drop_hdrlen = sizeof(*ip6); 441 iptos = (ntohl(ip6->ip6_flow) >> 20) & 0xff; 442 break; 443 #endif 444 #ifdef INET 445 case ETHERTYPE_IP: 446 ip = mtod(m, struct ip *); 447 th = (struct tcphdr *)(ip + 1); 448 drop_hdrlen = sizeof(*ip); 449 iptos = ip->ip_tos; 450 tlen = ntohs(ip->ip_len) - sizeof(struct ip); 451 break; 452 #endif 453 } /* end switch */ 454 /* 455 * Convert TCP protocol specific fields to host format. 456 */ 457 tcp_fields_to_host(th); 458 off = th->th_off << 2; 459 if (off < sizeof (struct tcphdr) || off > tlen) { 460 printf("off:%d < hdrlen:%zu || > tlen:%u -- dump\n", 461 off, 462 sizeof(struct tcphdr), 463 tlen); 464 KMOD_TCPSTAT_INC(tcps_rcvbadoff); 465 m_freem(m); 466 goto skipped_pkt; 467 } 468 tlen -= off; 469 drop_hdrlen += off; 470 /* 471 * Now lets setup the timeval to be when we should 472 * have been called (if we can). 473 */ 474 m->m_pkthdr.lro_nsegs = 1; 475 /* Now what about next packet? */ 476 } else { 477 /* 478 * This mbuf is an array of acks that have 479 * been compressed. We assert the inp has 480 * the flag set to enable this! 481 */ 482 KASSERT((tp->t_inpcb->inp_flags2 & INP_MBUF_ACKCMP), 483 ("tp:%p inp:%p no INP_MBUF_ACKCMP flags?", tp, tp->t_inpcb)); 484 tlen = 0; 485 drop_hdrlen = 0; 486 th = NULL; 487 iptos = 0; 488 } 489 tcp_get_usecs(&tv); 490 if (m_save || has_pkt) 491 nxt_pkt = 1; 492 else 493 nxt_pkt = 0; 494 if ((m->m_flags & M_ACKCMP) == 0) 495 KMOD_TCPSTAT_INC(tcps_rcvtotal); 496 else 497 KMOD_TCPSTAT_ADD(tcps_rcvtotal, (m->m_len / sizeof(struct tcp_ackent))); 498 inp = tp->t_inpcb; 499 INP_WLOCK_ASSERT(inp); 500 retval = (*tp->t_fb->tfb_do_segment_nounlock)(m, th, so, tp, drop_hdrlen, tlen, 501 iptos, nxt_pkt, &tv); 502 if (retval) { 503 /* We lost the lock and tcb probably */ 504 m = m_save; 505 while(m) { 506 m_save = m->m_nextpkt; 507 m->m_nextpkt = NULL; 508 m_freem(m); 509 m = m_save; 510 } 511 if (no_vn == 0) { 512 CURVNET_RESTORE(); 513 } 514 INP_UNLOCK_ASSERT(inp); 515 return(retval); 516 } 517 skipped_pkt: 518 m = m_save; 519 } 520 if (no_vn == 0) { 521 CURVNET_RESTORE(); 522 } 523 return(retval); 524 } 525 526 int 527 ctf_do_queued_segments(struct socket *so, struct tcpcb *tp, int have_pkt) 528 { 529 struct mbuf *m; 530 531 /* First lets see if we have old packets */ 532 if (tp->t_in_pkt) { 533 m = tp->t_in_pkt; 534 tp->t_in_pkt = NULL; 535 tp->t_tail_pkt = NULL; 536 if (ctf_process_inbound_raw(tp, so, m, have_pkt)) { 537 /* We lost the tcpcb (maybe a RST came in)? */ 538 return(1); 539 } 540 } 541 return (0); 542 } 543 544 uint32_t 545 ctf_outstanding(struct tcpcb *tp) 546 { 547 uint32_t bytes_out; 548 549 bytes_out = tp->snd_max - tp->snd_una; 550 if (tp->t_state < TCPS_ESTABLISHED) 551 bytes_out++; 552 if (tp->t_flags & TF_SENTFIN) 553 bytes_out++; 554 return (bytes_out); 555 } 556 557 uint32_t 558 ctf_flight_size(struct tcpcb *tp, uint32_t rc_sacked) 559 { 560 if (rc_sacked <= ctf_outstanding(tp)) 561 return(ctf_outstanding(tp) - rc_sacked); 562 else { 563 return (0); 564 } 565 } 566 567 void 568 ctf_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, 569 int32_t rstreason, int32_t tlen) 570 { 571 if (tp != NULL) { 572 tcp_dropwithreset(m, th, tp, tlen, rstreason); 573 INP_WUNLOCK(tp->t_inpcb); 574 } else 575 tcp_dropwithreset(m, th, NULL, tlen, rstreason); 576 } 577 578 void 579 ctf_ack_war_checks(struct tcpcb *tp, uint32_t *ts, uint32_t *cnt) 580 { 581 if ((ts != NULL) && (cnt != NULL) && 582 (tcp_ack_war_time_window > 0) && 583 (tcp_ack_war_cnt > 0)) { 584 /* We are possibly doing ack war prevention */ 585 uint32_t cts; 586 587 /* 588 * We use a msec tick here which gives us 589 * roughly 49 days. We don't need the 590 * precision of a microsecond timestamp which 591 * would only give us hours. 592 */ 593 cts = tcp_ts_getticks(); 594 if (TSTMP_LT((*ts), cts)) { 595 /* Timestamp is in the past */ 596 *cnt = 0; 597 *ts = (cts + tcp_ack_war_time_window); 598 } 599 if (*cnt < tcp_ack_war_cnt) { 600 *cnt = (*cnt + 1); 601 tp->t_flags |= TF_ACKNOW; 602 } else 603 tp->t_flags &= ~TF_ACKNOW; 604 } else 605 tp->t_flags |= TF_ACKNOW; 606 } 607 608 /* 609 * ctf_drop_checks returns 1 for you should not proceed. It places 610 * in ret_val what should be returned 1/0 by the caller. The 1 indicates 611 * that the TCB is unlocked and probably dropped. The 0 indicates the 612 * TCB is still valid and locked. 613 */ 614 int 615 _ctf_drop_checks(struct tcpopt *to, struct mbuf *m, struct tcphdr *th, 616 struct tcpcb *tp, int32_t *tlenp, 617 int32_t *thf, int32_t *drop_hdrlen, int32_t *ret_val, 618 uint32_t *ts, uint32_t *cnt) 619 { 620 int32_t todrop; 621 int32_t thflags; 622 int32_t tlen; 623 624 thflags = *thf; 625 tlen = *tlenp; 626 todrop = tp->rcv_nxt - th->th_seq; 627 if (todrop > 0) { 628 if (thflags & TH_SYN) { 629 thflags &= ~TH_SYN; 630 th->th_seq++; 631 if (th->th_urp > 1) 632 th->th_urp--; 633 else 634 thflags &= ~TH_URG; 635 todrop--; 636 } 637 /* 638 * Following if statement from Stevens, vol. 2, p. 960. 639 */ 640 if (todrop > tlen 641 || (todrop == tlen && (thflags & TH_FIN) == 0)) { 642 /* 643 * Any valid FIN must be to the left of the window. 644 * At this point the FIN must be a duplicate or out 645 * of sequence; drop it. 646 */ 647 thflags &= ~TH_FIN; 648 /* 649 * Send an ACK to resynchronize and drop any data. 650 * But keep on processing for RST or ACK. 651 */ 652 ctf_ack_war_checks(tp, ts, cnt); 653 todrop = tlen; 654 KMOD_TCPSTAT_INC(tcps_rcvduppack); 655 KMOD_TCPSTAT_ADD(tcps_rcvdupbyte, todrop); 656 } else { 657 KMOD_TCPSTAT_INC(tcps_rcvpartduppack); 658 KMOD_TCPSTAT_ADD(tcps_rcvpartdupbyte, todrop); 659 } 660 /* 661 * DSACK - add SACK block for dropped range 662 */ 663 if ((todrop > 0) && (tp->t_flags & TF_SACK_PERMIT)) { 664 /* 665 * ACK now, as the next in-sequence segment 666 * will clear the DSACK block again 667 */ 668 ctf_ack_war_checks(tp, ts, cnt); 669 if (tp->t_flags & TF_ACKNOW) 670 tcp_update_sack_list(tp, th->th_seq, 671 th->th_seq + todrop); 672 } 673 *drop_hdrlen += todrop; /* drop from the top afterwards */ 674 th->th_seq += todrop; 675 tlen -= todrop; 676 if (th->th_urp > todrop) 677 th->th_urp -= todrop; 678 else { 679 thflags &= ~TH_URG; 680 th->th_urp = 0; 681 } 682 } 683 /* 684 * If segment ends after window, drop trailing data (and PUSH and 685 * FIN); if nothing left, just ACK. 686 */ 687 todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd); 688 if (todrop > 0) { 689 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin); 690 if (todrop >= tlen) { 691 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, tlen); 692 /* 693 * If window is closed can only take segments at 694 * window edge, and have to drop data and PUSH from 695 * incoming segments. Continue processing, but 696 * remember to ack. Otherwise, drop segment and 697 * ack. 698 */ 699 if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) { 700 ctf_ack_war_checks(tp, ts, cnt); 701 KMOD_TCPSTAT_INC(tcps_rcvwinprobe); 702 } else { 703 __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val, ts, cnt); 704 return (1); 705 } 706 } else 707 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 708 m_adj(m, -todrop); 709 tlen -= todrop; 710 thflags &= ~(TH_PUSH | TH_FIN); 711 } 712 *thf = thflags; 713 *tlenp = tlen; 714 return (0); 715 } 716 717 /* 718 * The value in ret_val informs the caller 719 * if we dropped the tcb (and lock) or not. 720 * 1 = we dropped it, 0 = the TCB is still locked 721 * and valid. 722 */ 723 void 724 __ctf_do_dropafterack(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, int32_t thflags, int32_t tlen, int32_t *ret_val, uint32_t *ts, uint32_t *cnt) 725 { 726 /* 727 * Generate an ACK dropping incoming segment if it occupies sequence 728 * space, where the ACK reflects our state. 729 * 730 * We can now skip the test for the RST flag since all paths to this 731 * code happen after packets containing RST have been dropped. 732 * 733 * In the SYN-RECEIVED state, don't send an ACK unless the segment 734 * we received passes the SYN-RECEIVED ACK test. If it fails send a 735 * RST. This breaks the loop in the "LAND" DoS attack, and also 736 * prevents an ACK storm between two listening ports that have been 737 * sent forged SYN segments, each with the source address of the 738 * other. 739 */ 740 if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) && 741 (SEQ_GT(tp->snd_una, th->th_ack) || 742 SEQ_GT(th->th_ack, tp->snd_max))) { 743 *ret_val = 1; 744 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 745 return; 746 } else 747 *ret_val = 0; 748 ctf_ack_war_checks(tp, ts, cnt); 749 if (m) 750 m_freem(m); 751 } 752 753 void 754 ctf_do_drop(struct mbuf *m, struct tcpcb *tp) 755 { 756 757 /* 758 * Drop space held by incoming segment and return. 759 */ 760 if (tp != NULL) 761 INP_WUNLOCK(tp->t_inpcb); 762 if (m) 763 m_freem(m); 764 } 765 766 int 767 __ctf_process_rst(struct mbuf *m, struct tcphdr *th, struct socket *so, 768 struct tcpcb *tp, uint32_t *ts, uint32_t *cnt) 769 { 770 /* 771 * RFC5961 Section 3.2 772 * 773 * - RST drops connection only if SEG.SEQ == RCV.NXT. - If RST is in 774 * window, we send challenge ACK. 775 * 776 * Note: to take into account delayed ACKs, we should test against 777 * last_ack_sent instead of rcv_nxt. Note 2: we handle special case 778 * of closed window, not covered by the RFC. 779 */ 780 int dropped = 0; 781 782 if ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) && 783 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) || 784 (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq)) { 785 KASSERT(tp->t_state != TCPS_SYN_SENT, 786 ("%s: TH_RST for TCPS_SYN_SENT th %p tp %p", 787 __func__, th, tp)); 788 789 if (V_tcp_insecure_rst || 790 (tp->last_ack_sent == th->th_seq) || 791 (tp->rcv_nxt == th->th_seq)) { 792 KMOD_TCPSTAT_INC(tcps_drops); 793 /* Drop the connection. */ 794 switch (tp->t_state) { 795 case TCPS_SYN_RECEIVED: 796 so->so_error = ECONNREFUSED; 797 goto close; 798 case TCPS_ESTABLISHED: 799 case TCPS_FIN_WAIT_1: 800 case TCPS_FIN_WAIT_2: 801 case TCPS_CLOSE_WAIT: 802 case TCPS_CLOSING: 803 case TCPS_LAST_ACK: 804 so->so_error = ECONNRESET; 805 close: 806 tcp_state_change(tp, TCPS_CLOSED); 807 /* FALLTHROUGH */ 808 default: 809 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_RST); 810 tp = tcp_close(tp); 811 } 812 dropped = 1; 813 ctf_do_drop(m, tp); 814 } else { 815 int send_challenge; 816 817 KMOD_TCPSTAT_INC(tcps_badrst); 818 if ((ts != NULL) && (cnt != NULL) && 819 (tcp_ack_war_time_window > 0) && 820 (tcp_ack_war_cnt > 0)) { 821 /* We are possibly preventing an ack-rst war prevention */ 822 uint32_t cts; 823 824 /* 825 * We use a msec tick here which gives us 826 * roughly 49 days. We don't need the 827 * precision of a microsecond timestamp which 828 * would only give us hours. 829 */ 830 cts = tcp_ts_getticks(); 831 if (TSTMP_LT((*ts), cts)) { 832 /* Timestamp is in the past */ 833 *cnt = 0; 834 *ts = (cts + tcp_ack_war_time_window); 835 } 836 if (*cnt < tcp_ack_war_cnt) { 837 *cnt = (*cnt + 1); 838 send_challenge = 1; 839 } else 840 send_challenge = 0; 841 } else 842 send_challenge = 1; 843 if (send_challenge) { 844 /* Send challenge ACK. */ 845 tcp_respond(tp, mtod(m, void *), th, m, 846 tp->rcv_nxt, tp->snd_nxt, TH_ACK); 847 tp->last_ack_sent = tp->rcv_nxt; 848 } 849 } 850 } else { 851 m_freem(m); 852 } 853 return (dropped); 854 } 855 856 /* 857 * The value in ret_val informs the caller 858 * if we dropped the tcb (and lock) or not. 859 * 1 = we dropped it, 0 = the TCB is still locked 860 * and valid. 861 */ 862 void 863 ctf_challenge_ack(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t * ret_val) 864 { 865 866 NET_EPOCH_ASSERT(); 867 868 KMOD_TCPSTAT_INC(tcps_badsyn); 869 if (V_tcp_insecure_syn && 870 SEQ_GEQ(th->th_seq, tp->last_ack_sent) && 871 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) { 872 tp = tcp_drop(tp, ECONNRESET); 873 *ret_val = 1; 874 ctf_do_drop(m, tp); 875 } else { 876 /* Send challenge ACK. */ 877 tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt, 878 tp->snd_nxt, TH_ACK); 879 tp->last_ack_sent = tp->rcv_nxt; 880 m = NULL; 881 *ret_val = 0; 882 ctf_do_drop(m, NULL); 883 } 884 } 885 886 /* 887 * ctf_ts_check returns 1 for you should not proceed, the state 888 * machine should return. It places in ret_val what should 889 * be returned 1/0 by the caller (hpts_do_segment). The 1 indicates 890 * that the TCB is unlocked and probably dropped. The 0 indicates the 891 * TCB is still valid and locked. 892 */ 893 int 894 ctf_ts_check(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, 895 int32_t tlen, int32_t thflags, int32_t * ret_val) 896 { 897 898 if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { 899 /* 900 * Invalidate ts_recent. If this segment updates ts_recent, 901 * the age will be reset later and ts_recent will get a 902 * valid value. If it does not, setting ts_recent to zero 903 * will at least satisfy the requirement that zero be placed 904 * in the timestamp echo reply when ts_recent isn't valid. 905 * The age isn't reset until we get a valid ts_recent 906 * because we don't want out-of-order segments to be dropped 907 * when ts_recent is old. 908 */ 909 tp->ts_recent = 0; 910 } else { 911 KMOD_TCPSTAT_INC(tcps_rcvduppack); 912 KMOD_TCPSTAT_ADD(tcps_rcvdupbyte, tlen); 913 KMOD_TCPSTAT_INC(tcps_pawsdrop); 914 *ret_val = 0; 915 if (tlen) { 916 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 917 } else { 918 ctf_do_drop(m, NULL); 919 } 920 return (1); 921 } 922 return (0); 923 } 924 925 int 926 ctf_ts_check_ac(struct tcpcb *tp, int32_t thflags) 927 { 928 929 if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { 930 /* 931 * Invalidate ts_recent. If this segment updates ts_recent, 932 * the age will be reset later and ts_recent will get a 933 * valid value. If it does not, setting ts_recent to zero 934 * will at least satisfy the requirement that zero be placed 935 * in the timestamp echo reply when ts_recent isn't valid. 936 * The age isn't reset until we get a valid ts_recent 937 * because we don't want out-of-order segments to be dropped 938 * when ts_recent is old. 939 */ 940 tp->ts_recent = 0; 941 } else { 942 KMOD_TCPSTAT_INC(tcps_rcvduppack); 943 KMOD_TCPSTAT_INC(tcps_pawsdrop); 944 return (1); 945 } 946 return (0); 947 } 948 949 950 951 void 952 ctf_calc_rwin(struct socket *so, struct tcpcb *tp) 953 { 954 int32_t win; 955 956 /* 957 * Calculate amount of space in receive window, and then do TCP 958 * input processing. Receive window is amount of space in rcv queue, 959 * but not less than advertised window. 960 */ 961 win = sbspace(&so->so_rcv); 962 if (win < 0) 963 win = 0; 964 tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt)); 965 } 966 967 void 968 ctf_do_dropwithreset_conn(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, 969 int32_t rstreason, int32_t tlen) 970 { 971 972 tcp_dropwithreset(m, th, tp, tlen, rstreason); 973 tp = tcp_drop(tp, ETIMEDOUT); 974 if (tp) 975 INP_WUNLOCK(tp->t_inpcb); 976 } 977 978 uint32_t 979 ctf_fixed_maxseg(struct tcpcb *tp) 980 { 981 return (tcp_fixed_maxseg(tp)); 982 } 983 984 void 985 ctf_log_sack_filter(struct tcpcb *tp, int num_sack_blks, struct sackblk *sack_blocks) 986 { 987 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 988 union tcp_log_stackspecific log; 989 struct timeval tv; 990 991 memset(&log, 0, sizeof(log)); 992 log.u_bbr.timeStamp = tcp_get_usecs(&tv); 993 log.u_bbr.flex8 = num_sack_blks; 994 if (num_sack_blks > 0) { 995 log.u_bbr.flex1 = sack_blocks[0].start; 996 log.u_bbr.flex2 = sack_blocks[0].end; 997 } 998 if (num_sack_blks > 1) { 999 log.u_bbr.flex3 = sack_blocks[1].start; 1000 log.u_bbr.flex4 = sack_blocks[1].end; 1001 } 1002 if (num_sack_blks > 2) { 1003 log.u_bbr.flex5 = sack_blocks[2].start; 1004 log.u_bbr.flex6 = sack_blocks[2].end; 1005 } 1006 if (num_sack_blks > 3) { 1007 log.u_bbr.applimited = sack_blocks[3].start; 1008 log.u_bbr.pkts_out = sack_blocks[3].end; 1009 } 1010 TCP_LOG_EVENTP(tp, NULL, 1011 &tp->t_inpcb->inp_socket->so_rcv, 1012 &tp->t_inpcb->inp_socket->so_snd, 1013 TCP_SACK_FILTER_RES, 0, 1014 0, &log, false, &tv); 1015 } 1016 } 1017 1018 uint32_t 1019 ctf_decay_count(uint32_t count, uint32_t decay) 1020 { 1021 /* 1022 * Given a count, decay it by a set percentage. The 1023 * percentage is in thousands i.e. 100% = 1000, 1024 * 19.3% = 193. 1025 */ 1026 uint64_t perc_count, decay_per; 1027 uint32_t decayed_count; 1028 if (decay > 1000) { 1029 /* We don't raise it */ 1030 return (count); 1031 } 1032 perc_count = count; 1033 decay_per = decay; 1034 perc_count *= decay_per; 1035 perc_count /= 1000; 1036 /* 1037 * So now perc_count holds the 1038 * count decay value. 1039 */ 1040 decayed_count = count - (uint32_t)perc_count; 1041 return(decayed_count); 1042 } 1043 1044 int32_t 1045 ctf_progress_timeout_check(struct tcpcb *tp, bool log) 1046 { 1047 if (tp->t_maxunacktime && tp->t_acktime && TSTMP_GT(ticks, tp->t_acktime)) { 1048 if ((ticks - tp->t_acktime) >= tp->t_maxunacktime) { 1049 /* 1050 * There is an assumption that the caller 1051 * will drop the connection so we will 1052 * increment the counters here. 1053 */ 1054 if (log) 1055 tcp_log_end_status(tp, TCP_EI_STATUS_PROGRESS); 1056 #ifdef NETFLIX_STATS 1057 KMOD_TCPSTAT_INC(tcps_progdrops); 1058 #endif 1059 return (1); 1060 } 1061 } 1062 return (0); 1063 } 1064