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_ecn.h> 99 #include <netinet/tcp_hpts.h> 100 #include <netinet/tcp_lro.h> 101 #include <netinet/cc/cc.h> 102 #include <netinet/tcp_log_buf.h> 103 #ifdef TCPDEBUG 104 #include <netinet/tcp_debug.h> 105 #endif /* TCPDEBUG */ 106 #ifdef TCP_OFFLOAD 107 #include <netinet/tcp_offload.h> 108 #endif 109 #ifdef INET6 110 #include <netinet6/tcp6_var.h> 111 #endif 112 #include <netinet/tcp_fastopen.h> 113 114 #include <netipsec/ipsec_support.h> 115 #include <net/if.h> 116 #include <net/if_var.h> 117 118 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 119 #include <netipsec/ipsec.h> 120 #include <netipsec/ipsec6.h> 121 #endif /* IPSEC */ 122 123 #include <netinet/udp.h> 124 #include <netinet/udp_var.h> 125 #include <machine/in_cksum.h> 126 127 #ifdef MAC 128 #include <security/mac/mac_framework.h> 129 #endif 130 #include "rack_bbr_common.h" 131 132 /* 133 * Common TCP Functions - These are shared by borth 134 * rack and BBR. 135 */ 136 #ifdef KERN_TLS 137 uint32_t 138 ctf_get_opt_tls_size(struct socket *so, uint32_t rwnd) 139 { 140 struct ktls_session *tls; 141 uint32_t len; 142 143 again: 144 tls = so->so_snd.sb_tls_info; 145 len = tls->params.max_frame_len; /* max tls payload */ 146 len += tls->params.tls_hlen; /* tls header len */ 147 len += tls->params.tls_tlen; /* tls trailer len */ 148 if ((len * 4) > rwnd) { 149 /* 150 * Stroke this will suck counter and what 151 * else should we do Drew? From the 152 * TCP perspective I am not sure 153 * what should be done... 154 */ 155 if (tls->params.max_frame_len > 4096) { 156 tls->params.max_frame_len -= 4096; 157 if (tls->params.max_frame_len < 4096) 158 tls->params.max_frame_len = 4096; 159 goto again; 160 } 161 } 162 return (len); 163 } 164 #endif 165 166 static int 167 ctf_get_enet_type(struct ifnet *ifp, struct mbuf *m) 168 { 169 struct ether_header *eh; 170 #ifdef INET6 171 struct ip6_hdr *ip6 = NULL; /* Keep compiler happy. */ 172 #endif 173 #ifdef INET 174 struct ip *ip = NULL; /* Keep compiler happy. */ 175 #endif 176 #if defined(INET) || defined(INET6) 177 struct tcphdr *th; 178 int32_t tlen; 179 uint16_t drop_hdrlen; 180 #endif 181 uint16_t etype; 182 #ifdef INET 183 uint8_t iptos; 184 #endif 185 186 /* Is it the easy way? */ 187 if (m->m_flags & M_LRO_EHDRSTRP) 188 return (m->m_pkthdr.lro_etype); 189 /* 190 * Ok this is the old style call, the ethernet header is here. 191 * This also means no checksum or BPF were done. This 192 * can happen if the race to setup the inp fails and 193 * LRO sees no INP at packet input, but by the time 194 * we queue the packets an INP gets there. Its rare 195 * but it can occur so we will handle it. Note that 196 * this means duplicated work but with the rarity of it 197 * its not worth worrying about. 198 */ 199 /* Let the BPF see the packet */ 200 if (bpf_peers_present(ifp->if_bpf)) 201 ETHER_BPF_MTAP(ifp, m); 202 /* Now the csum */ 203 eh = mtod(m, struct ether_header *); 204 etype = ntohs(eh->ether_type); 205 m_adj(m, sizeof(*eh)); 206 switch (etype) { 207 #ifdef INET6 208 case ETHERTYPE_IPV6: 209 { 210 if (m->m_len < (sizeof(*ip6) + sizeof(*th))) { 211 m = m_pullup(m, sizeof(*ip6) + sizeof(*th)); 212 if (m == NULL) { 213 KMOD_TCPSTAT_INC(tcps_rcvshort); 214 return (-1); 215 } 216 } 217 ip6 = (struct ip6_hdr *)(eh + 1); 218 th = (struct tcphdr *)(ip6 + 1); 219 drop_hdrlen = sizeof(*ip6); 220 tlen = ntohs(ip6->ip6_plen); 221 if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID_IPV6) { 222 if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) 223 th->th_sum = m->m_pkthdr.csum_data; 224 else 225 th->th_sum = in6_cksum_pseudo(ip6, tlen, 226 IPPROTO_TCP, 227 m->m_pkthdr.csum_data); 228 th->th_sum ^= 0xffff; 229 } else 230 th->th_sum = in6_cksum(m, IPPROTO_TCP, drop_hdrlen, tlen); 231 if (th->th_sum) { 232 KMOD_TCPSTAT_INC(tcps_rcvbadsum); 233 m_freem(m); 234 return (-1); 235 } 236 return (etype); 237 } 238 #endif 239 #ifdef INET 240 case ETHERTYPE_IP: 241 { 242 if (m->m_len < sizeof (struct tcpiphdr)) { 243 m = m_pullup(m, sizeof (struct tcpiphdr)); 244 if (m == NULL) { 245 KMOD_TCPSTAT_INC(tcps_rcvshort); 246 return (-1); 247 } 248 } 249 ip = (struct ip *)(eh + 1); 250 th = (struct tcphdr *)(ip + 1); 251 drop_hdrlen = sizeof(*ip); 252 iptos = ip->ip_tos; 253 tlen = ntohs(ip->ip_len) - sizeof(struct ip); 254 if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) { 255 if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) 256 th->th_sum = m->m_pkthdr.csum_data; 257 else 258 th->th_sum = in_pseudo(ip->ip_src.s_addr, 259 ip->ip_dst.s_addr, 260 htonl(m->m_pkthdr.csum_data + tlen + IPPROTO_TCP)); 261 th->th_sum ^= 0xffff; 262 } else { 263 int len; 264 struct ipovly *ipov = (struct ipovly *)ip; 265 /* 266 * Checksum extended TCP header and data. 267 */ 268 len = drop_hdrlen + tlen; 269 bzero(ipov->ih_x1, sizeof(ipov->ih_x1)); 270 ipov->ih_len = htons(tlen); 271 th->th_sum = in_cksum(m, len); 272 /* Reset length for SDT probes. */ 273 ip->ip_len = htons(len); 274 /* Reset TOS bits */ 275 ip->ip_tos = iptos; 276 /* Re-initialization for later version check */ 277 ip->ip_v = IPVERSION; 278 ip->ip_hl = sizeof(*ip) >> 2; 279 } 280 if (th->th_sum) { 281 KMOD_TCPSTAT_INC(tcps_rcvbadsum); 282 m_freem(m); 283 return (-1); 284 } 285 break; 286 } 287 #endif 288 }; 289 return (etype); 290 } 291 292 /* 293 * The function ctf_process_inbound_raw() is used by 294 * transport developers to do the steps needed to 295 * support MBUF Queuing i.e. the flags in 296 * inp->inp_flags2: 297 * 298 * - INP_SUPPORTS_MBUFQ 299 * - INP_MBUF_QUEUE_READY 300 * - INP_DONT_SACK_QUEUE 301 * - INP_MBUF_ACKCMP 302 * 303 * These flags help control how LRO will deliver 304 * packets to the transport. You first set in inp_flags2 305 * the INP_SUPPORTS_MBUFQ to tell the LRO code that you 306 * will gladly take a queue of packets instead of a compressed 307 * single packet. You also set in your t_fb pointer the 308 * tfb_do_queued_segments to point to ctf_process_inbound_raw. 309 * 310 * This then gets you lists of inbound ACK's/Data instead 311 * of a condensed compressed ACK/DATA packet. Why would you 312 * want that? This will get you access to all the arrival 313 * times of at least LRO and possibly at the Hardware (if 314 * the interface card supports that) of the actual ACK/DATA. 315 * In some transport designs this is important since knowing 316 * the actual time we got the packet is useful information. 317 * 318 * A new special type of mbuf may also be supported by the transport 319 * if it has set the INP_MBUF_ACKCMP flag. If its set, LRO will 320 * possibly create a M_ACKCMP type mbuf. This is a mbuf with 321 * an array of "acks". One thing also to note is that when this 322 * occurs a subsequent LRO may find at the back of the untouched 323 * mbuf queue chain a M_ACKCMP and append on to it. This means 324 * that until the transport pulls in the mbuf chain queued 325 * for it more ack's may get on the mbufs that were already 326 * delivered. There currently is a limit of 6 acks condensed 327 * into 1 mbuf which means often when this is occuring, we 328 * don't get that effect but it does happen. 329 * 330 * Now there are some interesting Caveats that the transport 331 * designer needs to take into account when using this feature. 332 * 333 * 1) It is used with HPTS and pacing, when the pacing timer 334 * for output calls it will first call the input. 335 * 2) When you set INP_MBUF_QUEUE_READY this tells LRO 336 * queue normal packets, I am busy pacing out data and 337 * will process the queued packets before my tfb_tcp_output 338 * call from pacing. If a non-normal packet arrives, (e.g. sack) 339 * you will be awoken immediately. 340 * 3) Finally you can add the INP_DONT_SACK_QUEUE to not even 341 * be awoken if a SACK has arrived. You would do this when 342 * you were not only running a pacing for output timer 343 * but a Rack timer as well i.e. you know you are in recovery 344 * and are in the process (via the timers) of dealing with 345 * the loss. 346 * 347 * Now a critical thing you must be aware of here is that the 348 * use of the flags has a far greater scope then just your 349 * typical LRO. Why? Well thats because in the normal compressed 350 * LRO case at the end of a driver interupt all packets are going 351 * to get presented to the transport no matter if there is one 352 * or 100. With the MBUF_QUEUE model, this is not true. You will 353 * only be awoken to process the queue of packets when: 354 * a) The flags discussed above allow it. 355 * <or> 356 * b) You exceed a ack or data limit (by default the 357 * ack limit is infinity (64k acks) and the data 358 * limit is 64k of new TCP data) 359 * <or> 360 * c) The push bit has been set by the peer 361 */ 362 363 int 364 ctf_process_inbound_raw(struct tcpcb *tp, struct socket *so, struct mbuf *m, int has_pkt) 365 { 366 /* 367 * We are passed a raw change of mbuf packets 368 * that arrived in LRO. They are linked via 369 * the m_nextpkt link in the pkt-headers. 370 * 371 * We process each one by: 372 * a) saving off the next 373 * b) stripping off the ether-header 374 * c) formulating the arguments for 375 * the tfb_tcp_hpts_do_segment 376 * d) calling each mbuf to tfb_tcp_hpts_do_segment 377 * after adjusting the time to match the arrival time. 378 * Note that the LRO code assures no IP options are present. 379 * 380 * The symantics for calling tfb_tcp_hpts_do_segment are the 381 * following: 382 * 1) It returns 0 if all went well and you (the caller) need 383 * to release the lock. 384 * 2) If nxt_pkt is set, then the function will surpress calls 385 * to tcp_output() since you are promising to call again 386 * with another packet. 387 * 3) If it returns 1, then you must free all the packets being 388 * shipped in, the tcb has been destroyed (or about to be destroyed). 389 */ 390 struct mbuf *m_save; 391 struct tcphdr *th; 392 #ifdef INET6 393 struct ip6_hdr *ip6 = NULL; /* Keep compiler happy. */ 394 #endif 395 #ifdef INET 396 struct ip *ip = NULL; /* Keep compiler happy. */ 397 #endif 398 struct ifnet *ifp; 399 struct timeval tv; 400 struct inpcb *inp __diagused; 401 int32_t retval, nxt_pkt, tlen, off; 402 int etype = 0; 403 uint16_t drop_hdrlen; 404 uint8_t iptos, no_vn=0; 405 406 NET_EPOCH_ASSERT(); 407 if (m) 408 ifp = m_rcvif(m); 409 else 410 ifp = NULL; 411 if (ifp == NULL) { 412 /* 413 * We probably should not work around 414 * but kassert, since lro alwasy sets rcvif. 415 */ 416 no_vn = 1; 417 goto skip_vnet; 418 } 419 CURVNET_SET(ifp->if_vnet); 420 skip_vnet: 421 tcp_get_usecs(&tv); 422 while (m) { 423 m_save = m->m_nextpkt; 424 m->m_nextpkt = NULL; 425 if ((m->m_flags & M_ACKCMP) == 0) { 426 /* Now lets get the ether header */ 427 etype = ctf_get_enet_type(ifp, m); 428 if (etype == -1) { 429 /* Skip this packet it was freed by checksum */ 430 goto skipped_pkt; 431 } 432 KASSERT(((etype == ETHERTYPE_IPV6) || (etype == ETHERTYPE_IP)), 433 ("tp:%p m:%p etype:0x%x -- not IP or IPv6", tp, m, etype)); 434 /* Trim off the ethernet header */ 435 switch (etype) { 436 #ifdef INET6 437 case ETHERTYPE_IPV6: 438 ip6 = mtod(m, struct ip6_hdr *); 439 th = (struct tcphdr *)(ip6 + 1); 440 tlen = ntohs(ip6->ip6_plen); 441 drop_hdrlen = sizeof(*ip6); 442 iptos = (ntohl(ip6->ip6_flow) >> 20) & 0xff; 443 break; 444 #endif 445 #ifdef INET 446 case ETHERTYPE_IP: 447 ip = mtod(m, struct ip *); 448 th = (struct tcphdr *)(ip + 1); 449 drop_hdrlen = sizeof(*ip); 450 iptos = ip->ip_tos; 451 tlen = ntohs(ip->ip_len) - sizeof(struct ip); 452 break; 453 #endif 454 } /* end switch */ 455 /* 456 * Convert TCP protocol specific fields to host format. 457 */ 458 tcp_fields_to_host(th); 459 off = th->th_off << 2; 460 if (off < sizeof (struct tcphdr) || off > tlen) { 461 printf("off:%d < hdrlen:%zu || > tlen:%u -- dump\n", 462 off, 463 sizeof(struct tcphdr), 464 tlen); 465 KMOD_TCPSTAT_INC(tcps_rcvbadoff); 466 m_freem(m); 467 goto skipped_pkt; 468 } 469 tlen -= off; 470 drop_hdrlen += off; 471 /* 472 * Now lets setup the timeval to be when we should 473 * have been called (if we can). 474 */ 475 m->m_pkthdr.lro_nsegs = 1; 476 /* Now what about next packet? */ 477 } else { 478 /* 479 * This mbuf is an array of acks that have 480 * been compressed. We assert the inp has 481 * the flag set to enable this! 482 */ 483 KASSERT((tp->t_inpcb->inp_flags2 & INP_MBUF_ACKCMP), 484 ("tp:%p inp:%p no INP_MBUF_ACKCMP flags?", tp, tp->t_inpcb)); 485 tlen = 0; 486 drop_hdrlen = 0; 487 th = NULL; 488 iptos = 0; 489 } 490 tcp_get_usecs(&tv); 491 if (m_save || has_pkt) 492 nxt_pkt = 1; 493 else 494 nxt_pkt = 0; 495 if ((m->m_flags & M_ACKCMP) == 0) 496 KMOD_TCPSTAT_INC(tcps_rcvtotal); 497 else 498 KMOD_TCPSTAT_ADD(tcps_rcvtotal, (m->m_len / sizeof(struct tcp_ackent))); 499 inp = tp->t_inpcb; 500 INP_WLOCK_ASSERT(inp); 501 retval = (*tp->t_fb->tfb_do_segment_nounlock)(m, th, so, tp, drop_hdrlen, tlen, 502 iptos, nxt_pkt, &tv); 503 if (retval) { 504 /* We lost the lock and tcb probably */ 505 m = m_save; 506 while(m) { 507 m_save = m->m_nextpkt; 508 m->m_nextpkt = NULL; 509 m_freem(m); 510 m = m_save; 511 } 512 if (no_vn == 0) { 513 CURVNET_RESTORE(); 514 } 515 INP_UNLOCK_ASSERT(inp); 516 return(retval); 517 } 518 skipped_pkt: 519 m = m_save; 520 } 521 if (no_vn == 0) { 522 CURVNET_RESTORE(); 523 } 524 return(retval); 525 } 526 527 int 528 ctf_do_queued_segments(struct socket *so, struct tcpcb *tp, int have_pkt) 529 { 530 struct mbuf *m; 531 532 /* First lets see if we have old packets */ 533 if (tp->t_in_pkt) { 534 m = tp->t_in_pkt; 535 tp->t_in_pkt = NULL; 536 tp->t_tail_pkt = NULL; 537 if (ctf_process_inbound_raw(tp, so, m, have_pkt)) { 538 /* We lost the tcpcb (maybe a RST came in)? */ 539 return(1); 540 } 541 } 542 return (0); 543 } 544 545 uint32_t 546 ctf_outstanding(struct tcpcb *tp) 547 { 548 uint32_t bytes_out; 549 550 bytes_out = tp->snd_max - tp->snd_una; 551 if (tp->t_state < TCPS_ESTABLISHED) 552 bytes_out++; 553 if (tp->t_flags & TF_SENTFIN) 554 bytes_out++; 555 return (bytes_out); 556 } 557 558 uint32_t 559 ctf_flight_size(struct tcpcb *tp, uint32_t rc_sacked) 560 { 561 if (rc_sacked <= ctf_outstanding(tp)) 562 return(ctf_outstanding(tp) - rc_sacked); 563 else { 564 return (0); 565 } 566 } 567 568 void 569 ctf_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, 570 int32_t rstreason, int32_t tlen) 571 { 572 if (tp != NULL) { 573 tcp_dropwithreset(m, th, tp, tlen, rstreason); 574 INP_WUNLOCK(tp->t_inpcb); 575 } else 576 tcp_dropwithreset(m, th, NULL, tlen, rstreason); 577 } 578 579 void 580 ctf_ack_war_checks(struct tcpcb *tp, uint32_t *ts, uint32_t *cnt) 581 { 582 if ((ts != NULL) && (cnt != NULL) && 583 (tcp_ack_war_time_window > 0) && 584 (tcp_ack_war_cnt > 0)) { 585 /* We are possibly doing ack war prevention */ 586 uint32_t cts; 587 588 /* 589 * We use a msec tick here which gives us 590 * roughly 49 days. We don't need the 591 * precision of a microsecond timestamp which 592 * would only give us hours. 593 */ 594 cts = tcp_ts_getticks(); 595 if (TSTMP_LT((*ts), cts)) { 596 /* Timestamp is in the past */ 597 *cnt = 0; 598 *ts = (cts + tcp_ack_war_time_window); 599 } 600 if (*cnt < tcp_ack_war_cnt) { 601 *cnt = (*cnt + 1); 602 tp->t_flags |= TF_ACKNOW; 603 } else 604 tp->t_flags &= ~TF_ACKNOW; 605 } else 606 tp->t_flags |= TF_ACKNOW; 607 } 608 609 /* 610 * ctf_drop_checks returns 1 for you should not proceed. It places 611 * in ret_val what should be returned 1/0 by the caller. The 1 indicates 612 * that the TCB is unlocked and probably dropped. The 0 indicates the 613 * TCB is still valid and locked. 614 */ 615 int 616 _ctf_drop_checks(struct tcpopt *to, struct mbuf *m, struct tcphdr *th, 617 struct tcpcb *tp, int32_t *tlenp, 618 int32_t *thf, int32_t *drop_hdrlen, int32_t *ret_val, 619 uint32_t *ts, uint32_t *cnt) 620 { 621 int32_t todrop; 622 int32_t thflags; 623 int32_t tlen; 624 625 thflags = *thf; 626 tlen = *tlenp; 627 todrop = tp->rcv_nxt - th->th_seq; 628 if (todrop > 0) { 629 if (thflags & TH_SYN) { 630 thflags &= ~TH_SYN; 631 th->th_seq++; 632 if (th->th_urp > 1) 633 th->th_urp--; 634 else 635 thflags &= ~TH_URG; 636 todrop--; 637 } 638 /* 639 * Following if statement from Stevens, vol. 2, p. 960. 640 */ 641 if (todrop > tlen 642 || (todrop == tlen && (thflags & TH_FIN) == 0)) { 643 /* 644 * Any valid FIN must be to the left of the window. 645 * At this point the FIN must be a duplicate or out 646 * of sequence; drop it. 647 */ 648 thflags &= ~TH_FIN; 649 /* 650 * Send an ACK to resynchronize and drop any data. 651 * But keep on processing for RST or ACK. 652 */ 653 ctf_ack_war_checks(tp, ts, cnt); 654 todrop = tlen; 655 KMOD_TCPSTAT_INC(tcps_rcvduppack); 656 KMOD_TCPSTAT_ADD(tcps_rcvdupbyte, todrop); 657 } else { 658 KMOD_TCPSTAT_INC(tcps_rcvpartduppack); 659 KMOD_TCPSTAT_ADD(tcps_rcvpartdupbyte, todrop); 660 } 661 /* 662 * DSACK - add SACK block for dropped range 663 */ 664 if ((todrop > 0) && (tp->t_flags & TF_SACK_PERMIT)) { 665 /* 666 * ACK now, as the next in-sequence segment 667 * will clear the DSACK block again 668 */ 669 ctf_ack_war_checks(tp, ts, cnt); 670 if (tp->t_flags & TF_ACKNOW) 671 tcp_update_sack_list(tp, th->th_seq, 672 th->th_seq + todrop); 673 } 674 *drop_hdrlen += todrop; /* drop from the top afterwards */ 675 th->th_seq += todrop; 676 tlen -= todrop; 677 if (th->th_urp > todrop) 678 th->th_urp -= todrop; 679 else { 680 thflags &= ~TH_URG; 681 th->th_urp = 0; 682 } 683 } 684 /* 685 * If segment ends after window, drop trailing data (and PUSH and 686 * FIN); if nothing left, just ACK. 687 */ 688 todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd); 689 if (todrop > 0) { 690 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin); 691 if (todrop >= tlen) { 692 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, tlen); 693 /* 694 * If window is closed can only take segments at 695 * window edge, and have to drop data and PUSH from 696 * incoming segments. Continue processing, but 697 * remember to ack. Otherwise, drop segment and 698 * ack. 699 */ 700 if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) { 701 ctf_ack_war_checks(tp, ts, cnt); 702 KMOD_TCPSTAT_INC(tcps_rcvwinprobe); 703 } else { 704 __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val, ts, cnt); 705 return (1); 706 } 707 } else 708 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 709 m_adj(m, -todrop); 710 tlen -= todrop; 711 thflags &= ~(TH_PUSH | TH_FIN); 712 } 713 *thf = thflags; 714 *tlenp = tlen; 715 return (0); 716 } 717 718 /* 719 * The value in ret_val informs the caller 720 * if we dropped the tcb (and lock) or not. 721 * 1 = we dropped it, 0 = the TCB is still locked 722 * and valid. 723 */ 724 void 725 __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) 726 { 727 /* 728 * Generate an ACK dropping incoming segment if it occupies sequence 729 * space, where the ACK reflects our state. 730 * 731 * We can now skip the test for the RST flag since all paths to this 732 * code happen after packets containing RST have been dropped. 733 * 734 * In the SYN-RECEIVED state, don't send an ACK unless the segment 735 * we received passes the SYN-RECEIVED ACK test. If it fails send a 736 * RST. This breaks the loop in the "LAND" DoS attack, and also 737 * prevents an ACK storm between two listening ports that have been 738 * sent forged SYN segments, each with the source address of the 739 * other. 740 */ 741 if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) && 742 (SEQ_GT(tp->snd_una, th->th_ack) || 743 SEQ_GT(th->th_ack, tp->snd_max))) { 744 *ret_val = 1; 745 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 746 return; 747 } else 748 *ret_val = 0; 749 ctf_ack_war_checks(tp, ts, cnt); 750 if (m) 751 m_freem(m); 752 } 753 754 void 755 ctf_do_drop(struct mbuf *m, struct tcpcb *tp) 756 { 757 758 /* 759 * Drop space held by incoming segment and return. 760 */ 761 if (tp != NULL) 762 INP_WUNLOCK(tp->t_inpcb); 763 if (m) 764 m_freem(m); 765 } 766 767 int 768 __ctf_process_rst(struct mbuf *m, struct tcphdr *th, struct socket *so, 769 struct tcpcb *tp, uint32_t *ts, uint32_t *cnt) 770 { 771 /* 772 * RFC5961 Section 3.2 773 * 774 * - RST drops connection only if SEG.SEQ == RCV.NXT. - If RST is in 775 * window, we send challenge ACK. 776 * 777 * Note: to take into account delayed ACKs, we should test against 778 * last_ack_sent instead of rcv_nxt. Note 2: we handle special case 779 * of closed window, not covered by the RFC. 780 */ 781 int dropped = 0; 782 783 if ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) && 784 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) || 785 (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq)) { 786 KASSERT(tp->t_state != TCPS_SYN_SENT, 787 ("%s: TH_RST for TCPS_SYN_SENT th %p tp %p", 788 __func__, th, tp)); 789 790 if (V_tcp_insecure_rst || 791 (tp->last_ack_sent == th->th_seq) || 792 (tp->rcv_nxt == th->th_seq)) { 793 KMOD_TCPSTAT_INC(tcps_drops); 794 /* Drop the connection. */ 795 switch (tp->t_state) { 796 case TCPS_SYN_RECEIVED: 797 so->so_error = ECONNREFUSED; 798 goto close; 799 case TCPS_ESTABLISHED: 800 case TCPS_FIN_WAIT_1: 801 case TCPS_FIN_WAIT_2: 802 case TCPS_CLOSE_WAIT: 803 case TCPS_CLOSING: 804 case TCPS_LAST_ACK: 805 so->so_error = ECONNRESET; 806 close: 807 tcp_state_change(tp, TCPS_CLOSED); 808 /* FALLTHROUGH */ 809 default: 810 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_RST); 811 tp = tcp_close(tp); 812 } 813 dropped = 1; 814 ctf_do_drop(m, tp); 815 } else { 816 int send_challenge; 817 818 KMOD_TCPSTAT_INC(tcps_badrst); 819 if ((ts != NULL) && (cnt != NULL) && 820 (tcp_ack_war_time_window > 0) && 821 (tcp_ack_war_cnt > 0)) { 822 /* We are possibly preventing an ack-rst war prevention */ 823 uint32_t cts; 824 825 /* 826 * We use a msec tick here which gives us 827 * roughly 49 days. We don't need the 828 * precision of a microsecond timestamp which 829 * would only give us hours. 830 */ 831 cts = tcp_ts_getticks(); 832 if (TSTMP_LT((*ts), cts)) { 833 /* Timestamp is in the past */ 834 *cnt = 0; 835 *ts = (cts + tcp_ack_war_time_window); 836 } 837 if (*cnt < tcp_ack_war_cnt) { 838 *cnt = (*cnt + 1); 839 send_challenge = 1; 840 } else 841 send_challenge = 0; 842 } else 843 send_challenge = 1; 844 if (send_challenge) { 845 /* Send challenge ACK. */ 846 tcp_respond(tp, mtod(m, void *), th, m, 847 tp->rcv_nxt, tp->snd_nxt, TH_ACK); 848 tp->last_ack_sent = tp->rcv_nxt; 849 } 850 } 851 } else { 852 m_freem(m); 853 } 854 return (dropped); 855 } 856 857 /* 858 * The value in ret_val informs the caller 859 * if we dropped the tcb (and lock) or not. 860 * 1 = we dropped it, 0 = the TCB is still locked 861 * and valid. 862 */ 863 void 864 ctf_challenge_ack(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, uint8_t iptos, int32_t * ret_val) 865 { 866 867 NET_EPOCH_ASSERT(); 868 869 KMOD_TCPSTAT_INC(tcps_badsyn); 870 if (V_tcp_insecure_syn && 871 SEQ_GEQ(th->th_seq, tp->last_ack_sent) && 872 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) { 873 tp = tcp_drop(tp, ECONNRESET); 874 *ret_val = 1; 875 ctf_do_drop(m, tp); 876 } else { 877 tcp_ecn_input_syn_sent(tp, tcp_get_flags(th), iptos); 878 /* Send challenge ACK. */ 879 tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt, 880 tp->snd_nxt, TH_ACK); 881 tp->last_ack_sent = tp->rcv_nxt; 882 m = NULL; 883 *ret_val = 0; 884 ctf_do_drop(m, NULL); 885 } 886 } 887 888 /* 889 * ctf_ts_check returns 1 for you should not proceed, the state 890 * machine should return. It places in ret_val what should 891 * be returned 1/0 by the caller (hpts_do_segment). The 1 indicates 892 * that the TCB is unlocked and probably dropped. The 0 indicates the 893 * TCB is still valid and locked. 894 */ 895 int 896 ctf_ts_check(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, 897 int32_t tlen, int32_t thflags, int32_t * ret_val) 898 { 899 900 if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { 901 /* 902 * Invalidate ts_recent. If this segment updates ts_recent, 903 * the age will be reset later and ts_recent will get a 904 * valid value. If it does not, setting ts_recent to zero 905 * will at least satisfy the requirement that zero be placed 906 * in the timestamp echo reply when ts_recent isn't valid. 907 * The age isn't reset until we get a valid ts_recent 908 * because we don't want out-of-order segments to be dropped 909 * when ts_recent is old. 910 */ 911 tp->ts_recent = 0; 912 } else { 913 KMOD_TCPSTAT_INC(tcps_rcvduppack); 914 KMOD_TCPSTAT_ADD(tcps_rcvdupbyte, tlen); 915 KMOD_TCPSTAT_INC(tcps_pawsdrop); 916 *ret_val = 0; 917 if (tlen) { 918 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 919 } else { 920 ctf_do_drop(m, NULL); 921 } 922 return (1); 923 } 924 return (0); 925 } 926 927 int 928 ctf_ts_check_ac(struct tcpcb *tp, int32_t thflags) 929 { 930 931 if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { 932 /* 933 * Invalidate ts_recent. If this segment updates ts_recent, 934 * the age will be reset later and ts_recent will get a 935 * valid value. If it does not, setting ts_recent to zero 936 * will at least satisfy the requirement that zero be placed 937 * in the timestamp echo reply when ts_recent isn't valid. 938 * The age isn't reset until we get a valid ts_recent 939 * because we don't want out-of-order segments to be dropped 940 * when ts_recent is old. 941 */ 942 tp->ts_recent = 0; 943 } else { 944 KMOD_TCPSTAT_INC(tcps_rcvduppack); 945 KMOD_TCPSTAT_INC(tcps_pawsdrop); 946 return (1); 947 } 948 return (0); 949 } 950 951 952 953 void 954 ctf_calc_rwin(struct socket *so, struct tcpcb *tp) 955 { 956 int32_t win; 957 958 /* 959 * Calculate amount of space in receive window, and then do TCP 960 * input processing. Receive window is amount of space in rcv queue, 961 * but not less than advertised window. 962 */ 963 win = sbspace(&so->so_rcv); 964 if (win < 0) 965 win = 0; 966 tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt)); 967 } 968 969 void 970 ctf_do_dropwithreset_conn(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, 971 int32_t rstreason, int32_t tlen) 972 { 973 974 tcp_dropwithreset(m, th, tp, tlen, rstreason); 975 tp = tcp_drop(tp, ETIMEDOUT); 976 if (tp) 977 INP_WUNLOCK(tp->t_inpcb); 978 } 979 980 uint32_t 981 ctf_fixed_maxseg(struct tcpcb *tp) 982 { 983 return (tcp_fixed_maxseg(tp)); 984 } 985 986 void 987 ctf_log_sack_filter(struct tcpcb *tp, int num_sack_blks, struct sackblk *sack_blocks) 988 { 989 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 990 union tcp_log_stackspecific log; 991 struct timeval tv; 992 993 memset(&log, 0, sizeof(log)); 994 log.u_bbr.timeStamp = tcp_get_usecs(&tv); 995 log.u_bbr.flex8 = num_sack_blks; 996 if (num_sack_blks > 0) { 997 log.u_bbr.flex1 = sack_blocks[0].start; 998 log.u_bbr.flex2 = sack_blocks[0].end; 999 } 1000 if (num_sack_blks > 1) { 1001 log.u_bbr.flex3 = sack_blocks[1].start; 1002 log.u_bbr.flex4 = sack_blocks[1].end; 1003 } 1004 if (num_sack_blks > 2) { 1005 log.u_bbr.flex5 = sack_blocks[2].start; 1006 log.u_bbr.flex6 = sack_blocks[2].end; 1007 } 1008 if (num_sack_blks > 3) { 1009 log.u_bbr.applimited = sack_blocks[3].start; 1010 log.u_bbr.pkts_out = sack_blocks[3].end; 1011 } 1012 TCP_LOG_EVENTP(tp, NULL, 1013 &tp->t_inpcb->inp_socket->so_rcv, 1014 &tp->t_inpcb->inp_socket->so_snd, 1015 TCP_SACK_FILTER_RES, 0, 1016 0, &log, false, &tv); 1017 } 1018 } 1019 1020 uint32_t 1021 ctf_decay_count(uint32_t count, uint32_t decay) 1022 { 1023 /* 1024 * Given a count, decay it by a set percentage. The 1025 * percentage is in thousands i.e. 100% = 1000, 1026 * 19.3% = 193. 1027 */ 1028 uint64_t perc_count, decay_per; 1029 uint32_t decayed_count; 1030 if (decay > 1000) { 1031 /* We don't raise it */ 1032 return (count); 1033 } 1034 perc_count = count; 1035 decay_per = decay; 1036 perc_count *= decay_per; 1037 perc_count /= 1000; 1038 /* 1039 * So now perc_count holds the 1040 * count decay value. 1041 */ 1042 decayed_count = count - (uint32_t)perc_count; 1043 return(decayed_count); 1044 } 1045 1046 int32_t 1047 ctf_progress_timeout_check(struct tcpcb *tp, bool log) 1048 { 1049 if (tp->t_maxunacktime && tp->t_acktime && TSTMP_GT(ticks, tp->t_acktime)) { 1050 if ((ticks - tp->t_acktime) >= tp->t_maxunacktime) { 1051 /* 1052 * There is an assumption that the caller 1053 * will drop the connection so we will 1054 * increment the counters here. 1055 */ 1056 if (log) 1057 tcp_log_end_status(tp, TCP_EI_STATUS_PROGRESS); 1058 #ifdef NETFLIX_STATS 1059 KMOD_TCPSTAT_INC(tcps_progdrops); 1060 #endif 1061 return (1); 1062 } 1063 } 1064 return (0); 1065 } 1066