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