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