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 off = th->th_off << 2; 419 if (off < sizeof (struct tcphdr) || off > tlen) { 420 printf("off:%d < hdrlen:%zu || > tlen:%u -- dump\n", 421 off, 422 sizeof(struct tcphdr), 423 tlen); 424 KMOD_TCPSTAT_INC(tcps_rcvbadoff); 425 m_freem(m); 426 goto skipped_pkt; 427 } 428 tlen -= off; 429 drop_hdrlen += off; 430 /* 431 * Now lets setup the timeval to be when we should 432 * have been called (if we can). 433 */ 434 m->m_pkthdr.lro_nsegs = 1; 435 /* Now what about next packet? */ 436 } else { 437 /* 438 * This mbuf is an array of acks that have 439 * been compressed. We assert the inp has 440 * the flag set to enable this! 441 */ 442 KASSERT((tp->t_flags2 & TF2_MBUF_ACKCMP), 443 ("tp:%p no TF2_MBUF_ACKCMP flags?", tp)); 444 tlen = 0; 445 drop_hdrlen = 0; 446 th = NULL; 447 iptos = 0; 448 } 449 tcp_get_usecs(&tv); 450 if (m_save || has_pkt) 451 nxt_pkt = 1; 452 else 453 nxt_pkt = 0; 454 if ((m->m_flags & M_ACKCMP) == 0) 455 KMOD_TCPSTAT_INC(tcps_rcvtotal); 456 else 457 KMOD_TCPSTAT_ADD(tcps_rcvtotal, (m->m_len / sizeof(struct tcp_ackent))); 458 retval = (*tp->t_fb->tfb_do_segment_nounlock)(tp, m, th, 459 drop_hdrlen, tlen, iptos, nxt_pkt, &tv); 460 if (retval) { 461 /* We lost the lock and tcb probably */ 462 m = m_save; 463 while(m) { 464 m_save = m->m_nextpkt; 465 m->m_nextpkt = NULL; 466 m_freem(m); 467 m = m_save; 468 } 469 if (no_vn == 0) { 470 CURVNET_RESTORE(); 471 } 472 INP_UNLOCK_ASSERT(inp); 473 return(retval); 474 } 475 skipped_pkt: 476 m = m_save; 477 } 478 if (no_vn == 0) { 479 CURVNET_RESTORE(); 480 } 481 return(retval); 482 } 483 484 int 485 ctf_do_queued_segments(struct tcpcb *tp, int have_pkt) 486 { 487 struct mbuf *m; 488 489 /* First lets see if we have old packets */ 490 if ((m = STAILQ_FIRST(&tp->t_inqueue)) != NULL) { 491 STAILQ_INIT(&tp->t_inqueue); 492 if (ctf_process_inbound_raw(tp, m, have_pkt)) { 493 /* We lost the tcpcb (maybe a RST came in)? */ 494 return(1); 495 } 496 } 497 return (0); 498 } 499 500 uint32_t 501 ctf_outstanding(struct tcpcb *tp) 502 { 503 uint32_t bytes_out; 504 505 bytes_out = tp->snd_max - tp->snd_una; 506 if (tp->t_state < TCPS_ESTABLISHED) 507 bytes_out++; 508 if (tp->t_flags & TF_SENTFIN) 509 bytes_out++; 510 return (bytes_out); 511 } 512 513 uint32_t 514 ctf_flight_size(struct tcpcb *tp, uint32_t rc_sacked) 515 { 516 if (rc_sacked <= ctf_outstanding(tp)) 517 return(ctf_outstanding(tp) - rc_sacked); 518 else { 519 return (0); 520 } 521 } 522 523 void 524 ctf_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, 525 int32_t rstreason, int32_t tlen) 526 { 527 if (tp != NULL) { 528 tcp_dropwithreset(m, th, tp, tlen, rstreason); 529 INP_WUNLOCK(tptoinpcb(tp)); 530 } else 531 tcp_dropwithreset(m, th, NULL, tlen, rstreason); 532 } 533 534 void 535 ctf_ack_war_checks(struct tcpcb *tp, uint32_t *ts, uint32_t *cnt) 536 { 537 if ((ts != NULL) && (cnt != NULL) && 538 (tcp_ack_war_time_window > 0) && 539 (tcp_ack_war_cnt > 0)) { 540 /* We are possibly doing ack war prevention */ 541 uint32_t cts; 542 543 /* 544 * We use a msec tick here which gives us 545 * roughly 49 days. We don't need the 546 * precision of a microsecond timestamp which 547 * would only give us hours. 548 */ 549 cts = tcp_ts_getticks(); 550 if (TSTMP_LT((*ts), cts)) { 551 /* Timestamp is in the past */ 552 *cnt = 0; 553 *ts = (cts + tcp_ack_war_time_window); 554 } 555 if (*cnt < tcp_ack_war_cnt) { 556 *cnt = (*cnt + 1); 557 tp->t_flags |= TF_ACKNOW; 558 } else 559 tp->t_flags &= ~TF_ACKNOW; 560 } else 561 tp->t_flags |= TF_ACKNOW; 562 } 563 564 /* 565 * ctf_drop_checks returns 1 for you should not proceed. It places 566 * in ret_val what should be returned 1/0 by the caller. The 1 indicates 567 * that the TCB is unlocked and probably dropped. The 0 indicates the 568 * TCB is still valid and locked. 569 */ 570 int 571 _ctf_drop_checks(struct tcpopt *to, struct mbuf *m, struct tcphdr *th, 572 struct tcpcb *tp, int32_t *tlenp, 573 int32_t *thf, int32_t *drop_hdrlen, int32_t *ret_val, 574 uint32_t *ts, uint32_t *cnt) 575 { 576 int32_t todrop; 577 int32_t thflags; 578 int32_t tlen; 579 580 thflags = *thf; 581 tlen = *tlenp; 582 todrop = tp->rcv_nxt - th->th_seq; 583 if (todrop > 0) { 584 if (thflags & TH_SYN) { 585 thflags &= ~TH_SYN; 586 th->th_seq++; 587 if (th->th_urp > 1) 588 th->th_urp--; 589 else 590 thflags &= ~TH_URG; 591 todrop--; 592 } 593 /* 594 * Following if statement from Stevens, vol. 2, p. 960. 595 */ 596 if (todrop > tlen 597 || (todrop == tlen && (thflags & TH_FIN) == 0)) { 598 /* 599 * Any valid FIN must be to the left of the window. 600 * At this point the FIN must be a duplicate or out 601 * of sequence; drop it. 602 */ 603 thflags &= ~TH_FIN; 604 /* 605 * Send an ACK to resynchronize and drop any data. 606 * But keep on processing for RST or ACK. 607 */ 608 ctf_ack_war_checks(tp, ts, cnt); 609 todrop = tlen; 610 KMOD_TCPSTAT_INC(tcps_rcvduppack); 611 KMOD_TCPSTAT_ADD(tcps_rcvdupbyte, todrop); 612 } else { 613 KMOD_TCPSTAT_INC(tcps_rcvpartduppack); 614 KMOD_TCPSTAT_ADD(tcps_rcvpartdupbyte, todrop); 615 } 616 /* 617 * DSACK - add SACK block for dropped range 618 */ 619 if ((todrop > 0) && (tp->t_flags & TF_SACK_PERMIT)) { 620 /* 621 * ACK now, as the next in-sequence segment 622 * will clear the DSACK block again 623 */ 624 ctf_ack_war_checks(tp, ts, cnt); 625 if (tp->t_flags & TF_ACKNOW) 626 tcp_update_sack_list(tp, th->th_seq, 627 th->th_seq + todrop); 628 } 629 *drop_hdrlen += todrop; /* drop from the top afterwards */ 630 th->th_seq += todrop; 631 tlen -= todrop; 632 if (th->th_urp > todrop) 633 th->th_urp -= todrop; 634 else { 635 thflags &= ~TH_URG; 636 th->th_urp = 0; 637 } 638 } 639 /* 640 * If segment ends after window, drop trailing data (and PUSH and 641 * FIN); if nothing left, just ACK. 642 */ 643 todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd); 644 if (todrop > 0) { 645 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin); 646 if (todrop >= tlen) { 647 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, tlen); 648 /* 649 * If window is closed can only take segments at 650 * window edge, and have to drop data and PUSH from 651 * incoming segments. Continue processing, but 652 * remember to ack. Otherwise, drop segment and 653 * ack. 654 */ 655 if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) { 656 ctf_ack_war_checks(tp, ts, cnt); 657 KMOD_TCPSTAT_INC(tcps_rcvwinprobe); 658 } else { 659 __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val, ts, cnt); 660 return (1); 661 } 662 } else 663 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 664 m_adj(m, -todrop); 665 tlen -= todrop; 666 thflags &= ~(TH_PUSH | TH_FIN); 667 } 668 *thf = thflags; 669 *tlenp = tlen; 670 return (0); 671 } 672 673 /* 674 * The value in ret_val informs the caller 675 * if we dropped the tcb (and lock) or not. 676 * 1 = we dropped it, 0 = the TCB is still locked 677 * and valid. 678 */ 679 void 680 __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) 681 { 682 /* 683 * Generate an ACK dropping incoming segment if it occupies sequence 684 * space, where the ACK reflects our state. 685 * 686 * We can now skip the test for the RST flag since all paths to this 687 * code happen after packets containing RST have been dropped. 688 * 689 * In the SYN-RECEIVED state, don't send an ACK unless the segment 690 * we received passes the SYN-RECEIVED ACK test. If it fails send a 691 * RST. This breaks the loop in the "LAND" DoS attack, and also 692 * prevents an ACK storm between two listening ports that have been 693 * sent forged SYN segments, each with the source address of the 694 * other. 695 */ 696 if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) && 697 (SEQ_GT(tp->snd_una, th->th_ack) || 698 SEQ_GT(th->th_ack, tp->snd_max))) { 699 *ret_val = 1; 700 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 701 return; 702 } else 703 *ret_val = 0; 704 ctf_ack_war_checks(tp, ts, cnt); 705 if (m) 706 m_freem(m); 707 } 708 709 void 710 ctf_do_drop(struct mbuf *m, struct tcpcb *tp) 711 { 712 713 /* 714 * Drop space held by incoming segment and return. 715 */ 716 if (tp != NULL) 717 INP_WUNLOCK(tptoinpcb(tp)); 718 if (m) 719 m_freem(m); 720 } 721 722 int 723 __ctf_process_rst(struct mbuf *m, struct tcphdr *th, struct socket *so, 724 struct tcpcb *tp, uint32_t *ts, uint32_t *cnt) 725 { 726 /* 727 * RFC5961 Section 3.2 728 * 729 * - RST drops connection only if SEG.SEQ == RCV.NXT. - If RST is in 730 * window, we send challenge ACK. 731 * 732 * Note: to take into account delayed ACKs, we should test against 733 * last_ack_sent instead of rcv_nxt. Note 2: we handle special case 734 * of closed window, not covered by the RFC. 735 */ 736 int dropped = 0; 737 738 if ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) && 739 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) || 740 (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq)) { 741 KASSERT(tp->t_state != TCPS_SYN_SENT, 742 ("%s: TH_RST for TCPS_SYN_SENT th %p tp %p", 743 __func__, th, tp)); 744 745 if (V_tcp_insecure_rst || 746 (tp->last_ack_sent == th->th_seq) || 747 (tp->rcv_nxt == th->th_seq)) { 748 KMOD_TCPSTAT_INC(tcps_drops); 749 /* Drop the connection. */ 750 switch (tp->t_state) { 751 case TCPS_SYN_RECEIVED: 752 so->so_error = ECONNREFUSED; 753 goto close; 754 case TCPS_ESTABLISHED: 755 case TCPS_FIN_WAIT_1: 756 case TCPS_FIN_WAIT_2: 757 case TCPS_CLOSE_WAIT: 758 case TCPS_CLOSING: 759 case TCPS_LAST_ACK: 760 so->so_error = ECONNRESET; 761 close: 762 tcp_state_change(tp, TCPS_CLOSED); 763 /* FALLTHROUGH */ 764 default: 765 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_RST); 766 tp = tcp_close(tp); 767 } 768 dropped = 1; 769 ctf_do_drop(m, tp); 770 } else { 771 int send_challenge; 772 773 KMOD_TCPSTAT_INC(tcps_badrst); 774 if ((ts != NULL) && (cnt != NULL) && 775 (tcp_ack_war_time_window > 0) && 776 (tcp_ack_war_cnt > 0)) { 777 /* We are possibly preventing an ack-rst war prevention */ 778 uint32_t cts; 779 780 /* 781 * We use a msec tick here which gives us 782 * roughly 49 days. We don't need the 783 * precision of a microsecond timestamp which 784 * would only give us hours. 785 */ 786 cts = tcp_ts_getticks(); 787 if (TSTMP_LT((*ts), cts)) { 788 /* Timestamp is in the past */ 789 *cnt = 0; 790 *ts = (cts + tcp_ack_war_time_window); 791 } 792 if (*cnt < tcp_ack_war_cnt) { 793 *cnt = (*cnt + 1); 794 send_challenge = 1; 795 } else 796 send_challenge = 0; 797 } else 798 send_challenge = 1; 799 if (send_challenge) { 800 /* Send challenge ACK. */ 801 tcp_respond(tp, mtod(m, void *), th, m, 802 tp->rcv_nxt, tp->snd_nxt, TH_ACK); 803 tp->last_ack_sent = tp->rcv_nxt; 804 } 805 } 806 } else { 807 m_freem(m); 808 } 809 return (dropped); 810 } 811 812 /* 813 * The value in ret_val informs the caller 814 * if we dropped the tcb (and lock) or not. 815 * 1 = we dropped it, 0 = the TCB is still locked 816 * and valid. 817 */ 818 void 819 ctf_challenge_ack(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, uint8_t iptos, int32_t * ret_val) 820 { 821 822 NET_EPOCH_ASSERT(); 823 824 KMOD_TCPSTAT_INC(tcps_badsyn); 825 if (V_tcp_insecure_syn && 826 SEQ_GEQ(th->th_seq, tp->last_ack_sent) && 827 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) { 828 tp = tcp_drop(tp, ECONNRESET); 829 *ret_val = 1; 830 ctf_do_drop(m, tp); 831 } else { 832 tcp_ecn_input_syn_sent(tp, tcp_get_flags(th), iptos); 833 /* Send challenge ACK. */ 834 tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt, 835 tp->snd_nxt, TH_ACK); 836 tp->last_ack_sent = tp->rcv_nxt; 837 m = NULL; 838 *ret_val = 0; 839 ctf_do_drop(m, NULL); 840 } 841 } 842 843 /* 844 * ctf_ts_check returns 1 for you should not proceed, the state 845 * machine should return. It places in ret_val what should 846 * be returned 1/0 by the caller (hpts_do_segment). The 1 indicates 847 * that the TCB is unlocked and probably dropped. The 0 indicates the 848 * TCB is still valid and locked. 849 */ 850 int 851 ctf_ts_check(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, 852 int32_t tlen, int32_t thflags, int32_t * ret_val) 853 { 854 855 if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { 856 /* 857 * Invalidate ts_recent. If this segment updates ts_recent, 858 * the age will be reset later and ts_recent will get a 859 * valid value. If it does not, setting ts_recent to zero 860 * will at least satisfy the requirement that zero be placed 861 * in the timestamp echo reply when ts_recent isn't valid. 862 * The age isn't reset until we get a valid ts_recent 863 * because we don't want out-of-order segments to be dropped 864 * when ts_recent is old. 865 */ 866 tp->ts_recent = 0; 867 } else { 868 KMOD_TCPSTAT_INC(tcps_rcvduppack); 869 KMOD_TCPSTAT_ADD(tcps_rcvdupbyte, tlen); 870 KMOD_TCPSTAT_INC(tcps_pawsdrop); 871 *ret_val = 0; 872 if (tlen) { 873 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 874 } else { 875 ctf_do_drop(m, NULL); 876 } 877 return (1); 878 } 879 return (0); 880 } 881 882 int 883 ctf_ts_check_ac(struct tcpcb *tp, int32_t thflags) 884 { 885 886 if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { 887 /* 888 * Invalidate ts_recent. If this segment updates ts_recent, 889 * the age will be reset later and ts_recent will get a 890 * valid value. If it does not, setting ts_recent to zero 891 * will at least satisfy the requirement that zero be placed 892 * in the timestamp echo reply when ts_recent isn't valid. 893 * The age isn't reset until we get a valid ts_recent 894 * because we don't want out-of-order segments to be dropped 895 * when ts_recent is old. 896 */ 897 tp->ts_recent = 0; 898 } else { 899 KMOD_TCPSTAT_INC(tcps_rcvduppack); 900 KMOD_TCPSTAT_INC(tcps_pawsdrop); 901 return (1); 902 } 903 return (0); 904 } 905 906 907 908 void 909 ctf_calc_rwin(struct socket *so, struct tcpcb *tp) 910 { 911 int32_t win; 912 913 /* 914 * Calculate amount of space in receive window, and then do TCP 915 * input processing. Receive window is amount of space in rcv queue, 916 * but not less than advertised window. 917 */ 918 win = sbspace(&so->so_rcv); 919 if (win < 0) 920 win = 0; 921 tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt)); 922 } 923 924 void 925 ctf_do_dropwithreset_conn(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, 926 int32_t rstreason, int32_t tlen) 927 { 928 929 tcp_dropwithreset(m, th, tp, tlen, rstreason); 930 tp = tcp_drop(tp, ETIMEDOUT); 931 if (tp) 932 INP_WUNLOCK(tptoinpcb(tp)); 933 } 934 935 uint32_t 936 ctf_fixed_maxseg(struct tcpcb *tp) 937 { 938 return (tcp_fixed_maxseg(tp)); 939 } 940 941 void 942 ctf_log_sack_filter(struct tcpcb *tp, int num_sack_blks, struct sackblk *sack_blocks) 943 { 944 if (tcp_bblogging_on(tp)) { 945 union tcp_log_stackspecific log; 946 struct timeval tv; 947 948 memset(&log, 0, sizeof(log)); 949 log.u_bbr.timeStamp = tcp_get_usecs(&tv); 950 log.u_bbr.flex8 = num_sack_blks; 951 if (num_sack_blks > 0) { 952 log.u_bbr.flex1 = sack_blocks[0].start; 953 log.u_bbr.flex2 = sack_blocks[0].end; 954 } 955 if (num_sack_blks > 1) { 956 log.u_bbr.flex3 = sack_blocks[1].start; 957 log.u_bbr.flex4 = sack_blocks[1].end; 958 } 959 if (num_sack_blks > 2) { 960 log.u_bbr.flex5 = sack_blocks[2].start; 961 log.u_bbr.flex6 = sack_blocks[2].end; 962 } 963 if (num_sack_blks > 3) { 964 log.u_bbr.applimited = sack_blocks[3].start; 965 log.u_bbr.pkts_out = sack_blocks[3].end; 966 } 967 TCP_LOG_EVENTP(tp, NULL, 968 &tptosocket(tp)->so_rcv, 969 &tptosocket(tp)->so_snd, 970 TCP_SACK_FILTER_RES, 0, 971 0, &log, false, &tv); 972 } 973 } 974 975 uint32_t 976 ctf_decay_count(uint32_t count, uint32_t decay) 977 { 978 /* 979 * Given a count, decay it by a set percentage. The 980 * percentage is in thousands i.e. 100% = 1000, 981 * 19.3% = 193. 982 */ 983 uint64_t perc_count, decay_per; 984 uint32_t decayed_count; 985 if (decay > 1000) { 986 /* We don't raise it */ 987 return (count); 988 } 989 perc_count = count; 990 decay_per = decay; 991 perc_count *= decay_per; 992 perc_count /= 1000; 993 /* 994 * So now perc_count holds the 995 * count decay value. 996 */ 997 decayed_count = count - (uint32_t)perc_count; 998 return(decayed_count); 999 } 1000 1001 int32_t 1002 ctf_progress_timeout_check(struct tcpcb *tp, bool log) 1003 { 1004 if (tp->t_maxunacktime && tp->t_acktime && TSTMP_GT(ticks, tp->t_acktime)) { 1005 if ((ticks - tp->t_acktime) >= tp->t_maxunacktime) { 1006 /* 1007 * There is an assumption that the caller 1008 * will drop the connection so we will 1009 * increment the counters here. 1010 */ 1011 if (log) 1012 tcp_log_end_status(tp, TCP_EI_STATUS_PROGRESS); 1013 #ifdef NETFLIX_STATS 1014 KMOD_TCPSTAT_INC(tcps_progdrops); 1015 #endif 1016 return (1); 1017 } 1018 } 1019 return (0); 1020 } 1021