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 #include "opt_inet.h" 35 #include "opt_inet6.h" 36 #include "opt_ipsec.h" 37 #include "opt_ratelimit.h" 38 #include <sys/param.h> 39 #include <sys/arb.h> 40 #include <sys/module.h> 41 #include <sys/kernel.h> 42 #ifdef TCP_HHOOK 43 #include <sys/hhook.h> 44 #endif 45 #include <sys/malloc.h> 46 #include <sys/mbuf.h> 47 #include <sys/proc.h> 48 #include <sys/qmath.h> 49 #include <sys/socket.h> 50 #include <sys/socketvar.h> 51 #include <sys/sysctl.h> 52 #include <sys/systm.h> 53 #include <sys/tree.h> 54 #ifdef NETFLIX_STATS 55 #include <sys/stats.h> /* Must come after qmath.h and tree.h */ 56 #endif 57 #include <sys/refcount.h> 58 #include <sys/queue.h> 59 #include <sys/smp.h> 60 #include <sys/kthread.h> 61 #include <sys/lock.h> 62 #include <sys/mutex.h> 63 #include <sys/tim_filter.h> 64 #include <sys/time.h> 65 #include <vm/uma.h> 66 #include <sys/kern_prefetch.h> 67 68 #include <net/route.h> 69 #include <net/vnet.h> 70 #include <net/ethernet.h> 71 #include <net/bpf.h> 72 73 #define TCPSTATES /* for logging */ 74 75 #include <netinet/in.h> 76 #include <netinet/in_kdtrace.h> 77 #include <netinet/in_pcb.h> 78 #include <netinet/ip.h> 79 #include <netinet/ip_icmp.h> /* required for icmp_var.h */ 80 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */ 81 #include <netinet/ip_var.h> 82 #include <netinet/ip6.h> 83 #include <netinet6/in6_pcb.h> 84 #include <netinet6/ip6_var.h> 85 #include <netinet/tcp.h> 86 #include <netinet/tcp_fsm.h> 87 #include <netinet/tcp_seq.h> 88 #include <netinet/tcp_timer.h> 89 #include <netinet/tcp_var.h> 90 #include <netinet/tcpip.h> 91 #include <netinet/tcp_ecn.h> 92 #include <netinet/tcp_hpts.h> 93 #include <netinet/tcp_lro.h> 94 #include <netinet/cc/cc.h> 95 #include <netinet/tcp_log_buf.h> 96 #ifdef TCP_OFFLOAD 97 #include <netinet/tcp_offload.h> 98 #endif 99 #ifdef INET6 100 #include <netinet6/tcp6_var.h> 101 #endif 102 #include <netinet/tcp_fastopen.h> 103 104 #include <netipsec/ipsec_support.h> 105 #include <net/if.h> 106 #include <net/if_var.h> 107 #include <net/if_private.h> 108 109 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 110 #include <netipsec/ipsec.h> 111 #include <netipsec/ipsec6.h> 112 #endif /* IPSEC */ 113 114 #include <netinet/udp.h> 115 #include <netinet/udp_var.h> 116 #include <machine/in_cksum.h> 117 118 #ifdef MAC 119 #include <security/mac/mac_framework.h> 120 #endif 121 #include "rack_bbr_common.h" 122 123 /* 124 * Common TCP Functions - These are shared by borth 125 * rack and BBR. 126 */ 127 static int 128 ctf_get_enet_type(struct ifnet *ifp, struct mbuf *m) 129 { 130 struct ether_header *eh; 131 #ifdef INET6 132 struct ip6_hdr *ip6 = NULL; /* Keep compiler happy. */ 133 #endif 134 #ifdef INET 135 struct ip *ip = NULL; /* Keep compiler happy. */ 136 #endif 137 #if defined(INET) || defined(INET6) 138 struct tcphdr *th; 139 int32_t tlen; 140 uint16_t drop_hdrlen; 141 #endif 142 uint16_t etype; 143 #ifdef INET 144 uint8_t iptos; 145 #endif 146 147 /* Is it the easy way? */ 148 if (m->m_flags & M_LRO_EHDRSTRP) 149 return (m->m_pkthdr.lro_etype); 150 /* 151 * Ok this is the old style call, the ethernet header is here. 152 * This also means no checksum or BPF were done. This 153 * can happen if the race to setup the inp fails and 154 * LRO sees no INP at packet input, but by the time 155 * we queue the packets an INP gets there. Its rare 156 * but it can occur so we will handle it. Note that 157 * this means duplicated work but with the rarity of it 158 * its not worth worrying about. 159 */ 160 /* Let the BPF see the packet */ 161 if (bpf_peers_present(ifp->if_bpf)) 162 ETHER_BPF_MTAP(ifp, m); 163 /* Now the csum */ 164 eh = mtod(m, struct ether_header *); 165 etype = ntohs(eh->ether_type); 166 m_adj(m, sizeof(*eh)); 167 switch (etype) { 168 #ifdef INET6 169 case ETHERTYPE_IPV6: 170 { 171 if (m->m_len < (sizeof(*ip6) + sizeof(*th))) { 172 m = m_pullup(m, sizeof(*ip6) + sizeof(*th)); 173 if (m == NULL) { 174 KMOD_TCPSTAT_INC(tcps_rcvshort); 175 return (-1); 176 } 177 } 178 ip6 = (struct ip6_hdr *)(eh + 1); 179 th = (struct tcphdr *)(ip6 + 1); 180 drop_hdrlen = sizeof(*ip6); 181 tlen = ntohs(ip6->ip6_plen); 182 if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID_IPV6) { 183 if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) 184 th->th_sum = m->m_pkthdr.csum_data; 185 else 186 th->th_sum = in6_cksum_pseudo(ip6, tlen, 187 IPPROTO_TCP, 188 m->m_pkthdr.csum_data); 189 th->th_sum ^= 0xffff; 190 } else 191 th->th_sum = in6_cksum(m, IPPROTO_TCP, drop_hdrlen, tlen); 192 if (th->th_sum) { 193 KMOD_TCPSTAT_INC(tcps_rcvbadsum); 194 m_freem(m); 195 return (-1); 196 } 197 return (etype); 198 } 199 #endif 200 #ifdef INET 201 case ETHERTYPE_IP: 202 { 203 if (m->m_len < sizeof (struct tcpiphdr)) { 204 m = m_pullup(m, sizeof (struct tcpiphdr)); 205 if (m == NULL) { 206 KMOD_TCPSTAT_INC(tcps_rcvshort); 207 return (-1); 208 } 209 } 210 ip = (struct ip *)(eh + 1); 211 th = (struct tcphdr *)(ip + 1); 212 drop_hdrlen = sizeof(*ip); 213 iptos = ip->ip_tos; 214 tlen = ntohs(ip->ip_len) - sizeof(struct ip); 215 if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) { 216 if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) 217 th->th_sum = m->m_pkthdr.csum_data; 218 else 219 th->th_sum = in_pseudo(ip->ip_src.s_addr, 220 ip->ip_dst.s_addr, 221 htonl(m->m_pkthdr.csum_data + tlen + IPPROTO_TCP)); 222 th->th_sum ^= 0xffff; 223 } else { 224 int len; 225 struct ipovly *ipov = (struct ipovly *)ip; 226 /* 227 * Checksum extended TCP header and data. 228 */ 229 len = drop_hdrlen + tlen; 230 bzero(ipov->ih_x1, sizeof(ipov->ih_x1)); 231 ipov->ih_len = htons(tlen); 232 th->th_sum = in_cksum(m, len); 233 /* Reset length for SDT probes. */ 234 ip->ip_len = htons(len); 235 /* Reset TOS bits */ 236 ip->ip_tos = iptos; 237 /* Re-initialization for later version check */ 238 ip->ip_v = IPVERSION; 239 ip->ip_hl = sizeof(*ip) >> 2; 240 } 241 if (th->th_sum) { 242 KMOD_TCPSTAT_INC(tcps_rcvbadsum); 243 m_freem(m); 244 return (-1); 245 } 246 break; 247 } 248 #endif 249 }; 250 return (etype); 251 } 252 253 /* 254 * The function ctf_process_inbound_raw() is used by 255 * transport developers to do the steps needed to 256 * support MBUF Queuing i.e. the flags in 257 * inp->inp_flags2: 258 * 259 * - INP_SUPPORTS_MBUFQ 260 * - INP_MBUF_QUEUE_READY 261 * - INP_DONT_SACK_QUEUE 262 * - INP_MBUF_ACKCMP 263 * 264 * These flags help control how LRO will deliver 265 * packets to the transport. You first set in inp_flags2 266 * the INP_SUPPORTS_MBUFQ to tell the LRO code that you 267 * will gladly take a queue of packets instead of a compressed 268 * single packet. You also set in your t_fb pointer the 269 * tfb_do_queued_segments to point to ctf_process_inbound_raw. 270 * 271 * This then gets you lists of inbound ACK's/Data instead 272 * of a condensed compressed ACK/DATA packet. Why would you 273 * want that? This will get you access to all the arrival 274 * times of at least LRO and possibly at the Hardware (if 275 * the interface card supports that) of the actual ACK/DATA. 276 * In some transport designs this is important since knowing 277 * the actual time we got the packet is useful information. 278 * 279 * A new special type of mbuf may also be supported by the transport 280 * if it has set the INP_MBUF_ACKCMP flag. If its set, LRO will 281 * possibly create a M_ACKCMP type mbuf. This is a mbuf with 282 * an array of "acks". One thing also to note is that when this 283 * occurs a subsequent LRO may find at the back of the untouched 284 * mbuf queue chain a M_ACKCMP and append on to it. This means 285 * that until the transport pulls in the mbuf chain queued 286 * for it more ack's may get on the mbufs that were already 287 * delivered. There currently is a limit of 6 acks condensed 288 * into 1 mbuf which means often when this is occuring, we 289 * don't get that effect but it does happen. 290 * 291 * Now there are some interesting Caveats that the transport 292 * designer needs to take into account when using this feature. 293 * 294 * 1) It is used with HPTS and pacing, when the pacing timer 295 * for output calls it will first call the input. 296 * 2) When you set INP_MBUF_QUEUE_READY this tells LRO 297 * queue normal packets, I am busy pacing out data and 298 * will process the queued packets before my tfb_tcp_output 299 * call from pacing. If a non-normal packet arrives, (e.g. sack) 300 * you will be awoken immediately. 301 * 3) Finally you can add the INP_DONT_SACK_QUEUE to not even 302 * be awoken if a SACK has arrived. You would do this when 303 * you were not only running a pacing for output timer 304 * but a Rack timer as well i.e. you know you are in recovery 305 * and are in the process (via the timers) of dealing with 306 * the loss. 307 * 308 * Now a critical thing you must be aware of here is that the 309 * use of the flags has a far greater scope then just your 310 * typical LRO. Why? Well thats because in the normal compressed 311 * LRO case at the end of a driver interupt all packets are going 312 * to get presented to the transport no matter if there is one 313 * or 100. With the MBUF_QUEUE model, this is not true. You will 314 * only be awoken to process the queue of packets when: 315 * a) The flags discussed above allow it. 316 * <or> 317 * b) You exceed a ack or data limit (by default the 318 * ack limit is infinity (64k acks) and the data 319 * limit is 64k of new TCP data) 320 * <or> 321 * c) The push bit has been set by the peer 322 */ 323 324 static int 325 ctf_process_inbound_raw(struct tcpcb *tp, struct mbuf *m, int has_pkt) 326 { 327 /* 328 * We are passed a raw change of mbuf packets 329 * that arrived in LRO. They are linked via 330 * the m_nextpkt link in the pkt-headers. 331 * 332 * We process each one by: 333 * a) saving off the next 334 * b) stripping off the ether-header 335 * c) formulating the arguments for tfb_do_segment_nounlock() 336 * d) calling each mbuf to tfb_do_segment_nounlock() 337 * after adjusting the time to match the arrival time. 338 * Note that the LRO code assures no IP options are present. 339 * 340 * The symantics for calling tfb_do_segment_nounlock() are the 341 * following: 342 * 1) It returns 0 if all went well and you (the caller) need 343 * to release the lock. 344 * 2) If nxt_pkt is set, then the function will surpress calls 345 * to tcp_output() since you are promising to call again 346 * with another packet. 347 * 3) If it returns 1, then you must free all the packets being 348 * shipped in, the tcb has been destroyed (or about to be destroyed). 349 */ 350 struct mbuf *m_save; 351 struct tcphdr *th; 352 #ifdef INET6 353 struct ip6_hdr *ip6 = NULL; /* Keep compiler happy. */ 354 #endif 355 #ifdef INET 356 struct ip *ip = NULL; /* Keep compiler happy. */ 357 #endif 358 struct ifnet *ifp; 359 struct timeval tv; 360 struct inpcb *inp __diagused; 361 int32_t retval, nxt_pkt, tlen, off; 362 int etype = 0; 363 uint16_t drop_hdrlen; 364 uint8_t iptos, no_vn=0; 365 366 inp = tptoinpcb(tp); 367 INP_WLOCK_ASSERT(inp); 368 NET_EPOCH_ASSERT(); 369 370 if (m) 371 ifp = m_rcvif(m); 372 else 373 ifp = NULL; 374 if (ifp == NULL) { 375 /* 376 * We probably should not work around 377 * but kassert, since lro alwasy sets rcvif. 378 */ 379 no_vn = 1; 380 goto skip_vnet; 381 } 382 CURVNET_SET(ifp->if_vnet); 383 skip_vnet: 384 tcp_get_usecs(&tv); 385 while (m) { 386 m_save = m->m_nextpkt; 387 m->m_nextpkt = NULL; 388 if ((m->m_flags & M_ACKCMP) == 0) { 389 /* Now lets get the ether header */ 390 etype = ctf_get_enet_type(ifp, m); 391 if (etype == -1) { 392 /* Skip this packet it was freed by checksum */ 393 goto skipped_pkt; 394 } 395 KASSERT(((etype == ETHERTYPE_IPV6) || (etype == ETHERTYPE_IP)), 396 ("tp:%p m:%p etype:0x%x -- not IP or IPv6", tp, m, etype)); 397 /* Trim off the ethernet header */ 398 switch (etype) { 399 #ifdef INET6 400 case ETHERTYPE_IPV6: 401 ip6 = mtod(m, struct ip6_hdr *); 402 th = (struct tcphdr *)(ip6 + 1); 403 tlen = ntohs(ip6->ip6_plen); 404 drop_hdrlen = sizeof(*ip6); 405 iptos = (ntohl(ip6->ip6_flow) >> 20) & 0xff; 406 break; 407 #endif 408 #ifdef INET 409 case ETHERTYPE_IP: 410 ip = mtod(m, struct ip *); 411 th = (struct tcphdr *)(ip + 1); 412 drop_hdrlen = sizeof(*ip); 413 iptos = ip->ip_tos; 414 tlen = ntohs(ip->ip_len) - sizeof(struct ip); 415 break; 416 #endif 417 } /* end switch */ 418 /* 419 * Convert TCP protocol specific fields to host format. 420 */ 421 tcp_fields_to_host(th); 422 off = th->th_off << 2; 423 if (off < sizeof (struct tcphdr) || off > tlen) { 424 printf("off:%d < hdrlen:%zu || > tlen:%u -- dump\n", 425 off, 426 sizeof(struct tcphdr), 427 tlen); 428 KMOD_TCPSTAT_INC(tcps_rcvbadoff); 429 m_freem(m); 430 goto skipped_pkt; 431 } 432 tlen -= off; 433 drop_hdrlen += off; 434 /* 435 * Now lets setup the timeval to be when we should 436 * have been called (if we can). 437 */ 438 m->m_pkthdr.lro_nsegs = 1; 439 /* Now what about next packet? */ 440 } else { 441 /* 442 * This mbuf is an array of acks that have 443 * been compressed. We assert the inp has 444 * the flag set to enable this! 445 */ 446 KASSERT((tp->t_flags2 & TF2_MBUF_ACKCMP), 447 ("tp:%p no TF2_MBUF_ACKCMP flags?", tp)); 448 tlen = 0; 449 drop_hdrlen = 0; 450 th = NULL; 451 iptos = 0; 452 } 453 tcp_get_usecs(&tv); 454 if (m_save || has_pkt) 455 nxt_pkt = 1; 456 else 457 nxt_pkt = 0; 458 if ((m->m_flags & M_ACKCMP) == 0) 459 KMOD_TCPSTAT_INC(tcps_rcvtotal); 460 else 461 KMOD_TCPSTAT_ADD(tcps_rcvtotal, (m->m_len / sizeof(struct tcp_ackent))); 462 retval = (*tp->t_fb->tfb_do_segment_nounlock)(tp, m, th, 463 drop_hdrlen, tlen, iptos, nxt_pkt, &tv); 464 if (retval) { 465 /* We lost the lock and tcb probably */ 466 m = m_save; 467 while(m) { 468 m_save = m->m_nextpkt; 469 m->m_nextpkt = NULL; 470 m_freem(m); 471 m = m_save; 472 } 473 if (no_vn == 0) { 474 CURVNET_RESTORE(); 475 } 476 INP_UNLOCK_ASSERT(inp); 477 return(retval); 478 } 479 skipped_pkt: 480 m = m_save; 481 } 482 if (no_vn == 0) { 483 CURVNET_RESTORE(); 484 } 485 return(retval); 486 } 487 488 int 489 ctf_do_queued_segments(struct tcpcb *tp, int have_pkt) 490 { 491 struct mbuf *m; 492 493 /* First lets see if we have old packets */ 494 if ((m = STAILQ_FIRST(&tp->t_inqueue)) != NULL) { 495 STAILQ_INIT(&tp->t_inqueue); 496 if (ctf_process_inbound_raw(tp, m, have_pkt)) { 497 /* We lost the tcpcb (maybe a RST came in)? */ 498 return(1); 499 } 500 } 501 return (0); 502 } 503 504 uint32_t 505 ctf_outstanding(struct tcpcb *tp) 506 { 507 uint32_t bytes_out; 508 509 bytes_out = tp->snd_max - tp->snd_una; 510 if (tp->t_state < TCPS_ESTABLISHED) 511 bytes_out++; 512 if (tp->t_flags & TF_SENTFIN) 513 bytes_out++; 514 return (bytes_out); 515 } 516 517 uint32_t 518 ctf_flight_size(struct tcpcb *tp, uint32_t rc_sacked) 519 { 520 if (rc_sacked <= ctf_outstanding(tp)) 521 return(ctf_outstanding(tp) - rc_sacked); 522 else { 523 return (0); 524 } 525 } 526 527 void 528 ctf_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, 529 int32_t rstreason, int32_t tlen) 530 { 531 if (tp != NULL) { 532 tcp_dropwithreset(m, th, tp, tlen, rstreason); 533 INP_WUNLOCK(tptoinpcb(tp)); 534 } else 535 tcp_dropwithreset(m, th, NULL, tlen, rstreason); 536 } 537 538 void 539 ctf_ack_war_checks(struct tcpcb *tp, uint32_t *ts, uint32_t *cnt) 540 { 541 if ((ts != NULL) && (cnt != NULL) && 542 (tcp_ack_war_time_window > 0) && 543 (tcp_ack_war_cnt > 0)) { 544 /* We are possibly doing ack war prevention */ 545 uint32_t cts; 546 547 /* 548 * We use a msec tick here which gives us 549 * roughly 49 days. We don't need the 550 * precision of a microsecond timestamp which 551 * would only give us hours. 552 */ 553 cts = tcp_ts_getticks(); 554 if (TSTMP_LT((*ts), cts)) { 555 /* Timestamp is in the past */ 556 *cnt = 0; 557 *ts = (cts + tcp_ack_war_time_window); 558 } 559 if (*cnt < tcp_ack_war_cnt) { 560 *cnt = (*cnt + 1); 561 tp->t_flags |= TF_ACKNOW; 562 } else 563 tp->t_flags &= ~TF_ACKNOW; 564 } else 565 tp->t_flags |= TF_ACKNOW; 566 } 567 568 /* 569 * ctf_drop_checks returns 1 for you should not proceed. It places 570 * in ret_val what should be returned 1/0 by the caller. The 1 indicates 571 * that the TCB is unlocked and probably dropped. The 0 indicates the 572 * TCB is still valid and locked. 573 */ 574 int 575 _ctf_drop_checks(struct tcpopt *to, struct mbuf *m, struct tcphdr *th, 576 struct tcpcb *tp, int32_t *tlenp, 577 int32_t *thf, int32_t *drop_hdrlen, int32_t *ret_val, 578 uint32_t *ts, uint32_t *cnt) 579 { 580 int32_t todrop; 581 int32_t thflags; 582 int32_t tlen; 583 584 thflags = *thf; 585 tlen = *tlenp; 586 todrop = tp->rcv_nxt - th->th_seq; 587 if (todrop > 0) { 588 if (thflags & TH_SYN) { 589 thflags &= ~TH_SYN; 590 th->th_seq++; 591 if (th->th_urp > 1) 592 th->th_urp--; 593 else 594 thflags &= ~TH_URG; 595 todrop--; 596 } 597 /* 598 * Following if statement from Stevens, vol. 2, p. 960. 599 */ 600 if (todrop > tlen 601 || (todrop == tlen && (thflags & TH_FIN) == 0)) { 602 /* 603 * Any valid FIN must be to the left of the window. 604 * At this point the FIN must be a duplicate or out 605 * of sequence; drop it. 606 */ 607 thflags &= ~TH_FIN; 608 /* 609 * Send an ACK to resynchronize and drop any data. 610 * But keep on processing for RST or ACK. 611 */ 612 ctf_ack_war_checks(tp, ts, cnt); 613 todrop = tlen; 614 KMOD_TCPSTAT_INC(tcps_rcvduppack); 615 KMOD_TCPSTAT_ADD(tcps_rcvdupbyte, todrop); 616 } else { 617 KMOD_TCPSTAT_INC(tcps_rcvpartduppack); 618 KMOD_TCPSTAT_ADD(tcps_rcvpartdupbyte, todrop); 619 } 620 /* 621 * DSACK - add SACK block for dropped range 622 */ 623 if ((todrop > 0) && (tp->t_flags & TF_SACK_PERMIT)) { 624 /* 625 * ACK now, as the next in-sequence segment 626 * will clear the DSACK block again 627 */ 628 ctf_ack_war_checks(tp, ts, cnt); 629 if (tp->t_flags & TF_ACKNOW) 630 tcp_update_sack_list(tp, th->th_seq, 631 th->th_seq + todrop); 632 } 633 *drop_hdrlen += todrop; /* drop from the top afterwards */ 634 th->th_seq += todrop; 635 tlen -= todrop; 636 if (th->th_urp > todrop) 637 th->th_urp -= todrop; 638 else { 639 thflags &= ~TH_URG; 640 th->th_urp = 0; 641 } 642 } 643 /* 644 * If segment ends after window, drop trailing data (and PUSH and 645 * FIN); if nothing left, just ACK. 646 */ 647 todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd); 648 if (todrop > 0) { 649 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin); 650 if (todrop >= tlen) { 651 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, tlen); 652 /* 653 * If window is closed can only take segments at 654 * window edge, and have to drop data and PUSH from 655 * incoming segments. Continue processing, but 656 * remember to ack. Otherwise, drop segment and 657 * ack. 658 */ 659 if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) { 660 ctf_ack_war_checks(tp, ts, cnt); 661 KMOD_TCPSTAT_INC(tcps_rcvwinprobe); 662 } else { 663 __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val, ts, cnt); 664 return (1); 665 } 666 } else 667 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 668 m_adj(m, -todrop); 669 tlen -= todrop; 670 thflags &= ~(TH_PUSH | TH_FIN); 671 } 672 *thf = thflags; 673 *tlenp = tlen; 674 return (0); 675 } 676 677 /* 678 * The value in ret_val informs the caller 679 * if we dropped the tcb (and lock) or not. 680 * 1 = we dropped it, 0 = the TCB is still locked 681 * and valid. 682 */ 683 void 684 __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) 685 { 686 /* 687 * Generate an ACK dropping incoming segment if it occupies sequence 688 * space, where the ACK reflects our state. 689 * 690 * We can now skip the test for the RST flag since all paths to this 691 * code happen after packets containing RST have been dropped. 692 * 693 * In the SYN-RECEIVED state, don't send an ACK unless the segment 694 * we received passes the SYN-RECEIVED ACK test. If it fails send a 695 * RST. This breaks the loop in the "LAND" DoS attack, and also 696 * prevents an ACK storm between two listening ports that have been 697 * sent forged SYN segments, each with the source address of the 698 * other. 699 */ 700 if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) && 701 (SEQ_GT(tp->snd_una, th->th_ack) || 702 SEQ_GT(th->th_ack, tp->snd_max))) { 703 *ret_val = 1; 704 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 705 return; 706 } else 707 *ret_val = 0; 708 ctf_ack_war_checks(tp, ts, cnt); 709 if (m) 710 m_freem(m); 711 } 712 713 void 714 ctf_do_drop(struct mbuf *m, struct tcpcb *tp) 715 { 716 717 /* 718 * Drop space held by incoming segment and return. 719 */ 720 if (tp != NULL) 721 INP_WUNLOCK(tptoinpcb(tp)); 722 if (m) 723 m_freem(m); 724 } 725 726 int 727 __ctf_process_rst(struct mbuf *m, struct tcphdr *th, struct socket *so, 728 struct tcpcb *tp, uint32_t *ts, uint32_t *cnt) 729 { 730 /* 731 * RFC5961 Section 3.2 732 * 733 * - RST drops connection only if SEG.SEQ == RCV.NXT. - If RST is in 734 * window, we send challenge ACK. 735 * 736 * Note: to take into account delayed ACKs, we should test against 737 * last_ack_sent instead of rcv_nxt. Note 2: we handle special case 738 * of closed window, not covered by the RFC. 739 */ 740 int dropped = 0; 741 742 if ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) && 743 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) || 744 (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq)) { 745 KASSERT(tp->t_state != TCPS_SYN_SENT, 746 ("%s: TH_RST for TCPS_SYN_SENT th %p tp %p", 747 __func__, th, tp)); 748 749 if (V_tcp_insecure_rst || 750 (tp->last_ack_sent == th->th_seq) || 751 (tp->rcv_nxt == th->th_seq)) { 752 KMOD_TCPSTAT_INC(tcps_drops); 753 /* Drop the connection. */ 754 switch (tp->t_state) { 755 case TCPS_SYN_RECEIVED: 756 so->so_error = ECONNREFUSED; 757 goto close; 758 case TCPS_ESTABLISHED: 759 case TCPS_FIN_WAIT_1: 760 case TCPS_FIN_WAIT_2: 761 case TCPS_CLOSE_WAIT: 762 case TCPS_CLOSING: 763 case TCPS_LAST_ACK: 764 so->so_error = ECONNRESET; 765 close: 766 tcp_state_change(tp, TCPS_CLOSED); 767 /* FALLTHROUGH */ 768 default: 769 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_RST); 770 tp = tcp_close(tp); 771 } 772 dropped = 1; 773 ctf_do_drop(m, tp); 774 } else { 775 int send_challenge; 776 777 KMOD_TCPSTAT_INC(tcps_badrst); 778 if ((ts != NULL) && (cnt != NULL) && 779 (tcp_ack_war_time_window > 0) && 780 (tcp_ack_war_cnt > 0)) { 781 /* We are possibly preventing an ack-rst war prevention */ 782 uint32_t cts; 783 784 /* 785 * We use a msec tick here which gives us 786 * roughly 49 days. We don't need the 787 * precision of a microsecond timestamp which 788 * would only give us hours. 789 */ 790 cts = tcp_ts_getticks(); 791 if (TSTMP_LT((*ts), cts)) { 792 /* Timestamp is in the past */ 793 *cnt = 0; 794 *ts = (cts + tcp_ack_war_time_window); 795 } 796 if (*cnt < tcp_ack_war_cnt) { 797 *cnt = (*cnt + 1); 798 send_challenge = 1; 799 } else 800 send_challenge = 0; 801 } else 802 send_challenge = 1; 803 if (send_challenge) { 804 /* Send challenge ACK. */ 805 tcp_respond(tp, mtod(m, void *), th, m, 806 tp->rcv_nxt, tp->snd_nxt, TH_ACK); 807 tp->last_ack_sent = tp->rcv_nxt; 808 } 809 } 810 } else { 811 m_freem(m); 812 } 813 return (dropped); 814 } 815 816 /* 817 * The value in ret_val informs the caller 818 * if we dropped the tcb (and lock) or not. 819 * 1 = we dropped it, 0 = the TCB is still locked 820 * and valid. 821 */ 822 void 823 ctf_challenge_ack(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, uint8_t iptos, int32_t * ret_val) 824 { 825 826 NET_EPOCH_ASSERT(); 827 828 KMOD_TCPSTAT_INC(tcps_badsyn); 829 if (V_tcp_insecure_syn && 830 SEQ_GEQ(th->th_seq, tp->last_ack_sent) && 831 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) { 832 tp = tcp_drop(tp, ECONNRESET); 833 *ret_val = 1; 834 ctf_do_drop(m, tp); 835 } else { 836 tcp_ecn_input_syn_sent(tp, tcp_get_flags(th), iptos); 837 /* Send challenge ACK. */ 838 tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt, 839 tp->snd_nxt, TH_ACK); 840 tp->last_ack_sent = tp->rcv_nxt; 841 m = NULL; 842 *ret_val = 0; 843 ctf_do_drop(m, NULL); 844 } 845 } 846 847 /* 848 * ctf_ts_check returns 1 for you should not proceed, the state 849 * machine should return. It places in ret_val what should 850 * be returned 1/0 by the caller (hpts_do_segment). The 1 indicates 851 * that the TCB is unlocked and probably dropped. The 0 indicates the 852 * TCB is still valid and locked. 853 */ 854 int 855 ctf_ts_check(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, 856 int32_t tlen, int32_t thflags, int32_t * ret_val) 857 { 858 859 if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { 860 /* 861 * Invalidate ts_recent. If this segment updates ts_recent, 862 * the age will be reset later and ts_recent will get a 863 * valid value. If it does not, setting ts_recent to zero 864 * will at least satisfy the requirement that zero be placed 865 * in the timestamp echo reply when ts_recent isn't valid. 866 * The age isn't reset until we get a valid ts_recent 867 * because we don't want out-of-order segments to be dropped 868 * when ts_recent is old. 869 */ 870 tp->ts_recent = 0; 871 } else { 872 KMOD_TCPSTAT_INC(tcps_rcvduppack); 873 KMOD_TCPSTAT_ADD(tcps_rcvdupbyte, tlen); 874 KMOD_TCPSTAT_INC(tcps_pawsdrop); 875 *ret_val = 0; 876 if (tlen) { 877 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 878 } else { 879 ctf_do_drop(m, NULL); 880 } 881 return (1); 882 } 883 return (0); 884 } 885 886 int 887 ctf_ts_check_ac(struct tcpcb *tp, int32_t thflags) 888 { 889 890 if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { 891 /* 892 * Invalidate ts_recent. If this segment updates ts_recent, 893 * the age will be reset later and ts_recent will get a 894 * valid value. If it does not, setting ts_recent to zero 895 * will at least satisfy the requirement that zero be placed 896 * in the timestamp echo reply when ts_recent isn't valid. 897 * The age isn't reset until we get a valid ts_recent 898 * because we don't want out-of-order segments to be dropped 899 * when ts_recent is old. 900 */ 901 tp->ts_recent = 0; 902 } else { 903 KMOD_TCPSTAT_INC(tcps_rcvduppack); 904 KMOD_TCPSTAT_INC(tcps_pawsdrop); 905 return (1); 906 } 907 return (0); 908 } 909 910 911 912 void 913 ctf_calc_rwin(struct socket *so, struct tcpcb *tp) 914 { 915 int32_t win; 916 917 /* 918 * Calculate amount of space in receive window, and then do TCP 919 * input processing. Receive window is amount of space in rcv queue, 920 * but not less than advertised window. 921 */ 922 win = sbspace(&so->so_rcv); 923 if (win < 0) 924 win = 0; 925 tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt)); 926 } 927 928 void 929 ctf_do_dropwithreset_conn(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, 930 int32_t rstreason, int32_t tlen) 931 { 932 933 tcp_dropwithreset(m, th, tp, tlen, rstreason); 934 tp = tcp_drop(tp, ETIMEDOUT); 935 if (tp) 936 INP_WUNLOCK(tptoinpcb(tp)); 937 } 938 939 uint32_t 940 ctf_fixed_maxseg(struct tcpcb *tp) 941 { 942 return (tcp_fixed_maxseg(tp)); 943 } 944 945 void 946 ctf_log_sack_filter(struct tcpcb *tp, int num_sack_blks, struct sackblk *sack_blocks) 947 { 948 if (tcp_bblogging_on(tp)) { 949 union tcp_log_stackspecific log; 950 struct timeval tv; 951 952 memset(&log, 0, sizeof(log)); 953 log.u_bbr.timeStamp = tcp_get_usecs(&tv); 954 log.u_bbr.flex8 = num_sack_blks; 955 if (num_sack_blks > 0) { 956 log.u_bbr.flex1 = sack_blocks[0].start; 957 log.u_bbr.flex2 = sack_blocks[0].end; 958 } 959 if (num_sack_blks > 1) { 960 log.u_bbr.flex3 = sack_blocks[1].start; 961 log.u_bbr.flex4 = sack_blocks[1].end; 962 } 963 if (num_sack_blks > 2) { 964 log.u_bbr.flex5 = sack_blocks[2].start; 965 log.u_bbr.flex6 = sack_blocks[2].end; 966 } 967 if (num_sack_blks > 3) { 968 log.u_bbr.applimited = sack_blocks[3].start; 969 log.u_bbr.pkts_out = sack_blocks[3].end; 970 } 971 TCP_LOG_EVENTP(tp, NULL, 972 &tptosocket(tp)->so_rcv, 973 &tptosocket(tp)->so_snd, 974 TCP_SACK_FILTER_RES, 0, 975 0, &log, false, &tv); 976 } 977 } 978 979 uint32_t 980 ctf_decay_count(uint32_t count, uint32_t decay) 981 { 982 /* 983 * Given a count, decay it by a set percentage. The 984 * percentage is in thousands i.e. 100% = 1000, 985 * 19.3% = 193. 986 */ 987 uint64_t perc_count, decay_per; 988 uint32_t decayed_count; 989 if (decay > 1000) { 990 /* We don't raise it */ 991 return (count); 992 } 993 perc_count = count; 994 decay_per = decay; 995 perc_count *= decay_per; 996 perc_count /= 1000; 997 /* 998 * So now perc_count holds the 999 * count decay value. 1000 */ 1001 decayed_count = count - (uint32_t)perc_count; 1002 return(decayed_count); 1003 } 1004 1005 int32_t 1006 ctf_progress_timeout_check(struct tcpcb *tp, bool log) 1007 { 1008 if (tp->t_maxunacktime && tp->t_acktime && TSTMP_GT(ticks, tp->t_acktime)) { 1009 if ((ticks - tp->t_acktime) >= tp->t_maxunacktime) { 1010 /* 1011 * There is an assumption that the caller 1012 * will drop the connection so we will 1013 * increment the counters here. 1014 */ 1015 if (log) 1016 tcp_log_end_status(tp, TCP_EI_STATUS_PROGRESS); 1017 #ifdef NETFLIX_STATS 1018 KMOD_TCPSTAT_INC(tcps_progdrops); 1019 #endif 1020 return (1); 1021 } 1022 } 1023 return (0); 1024 } 1025