1 /*- 2 * Copyright (c) 2016-2020 Netflix, Inc. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. Redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution. 12 * 13 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 16 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 23 * SUCH DAMAGE. 24 * 25 */ 26 /* 27 * Author: Randall Stewart <rrs@netflix.com> 28 * This work is based on the ACM Queue paper 29 * BBR - Congestion Based Congestion Control 30 * and also numerous discussions with Neal, Yuchung and Van. 31 */ 32 33 #include <sys/cdefs.h> 34 __FBSDID("$FreeBSD$"); 35 36 #include "opt_inet.h" 37 #include "opt_inet6.h" 38 #include "opt_ipsec.h" 39 #include "opt_tcpdebug.h" 40 #include "opt_ratelimit.h" 41 #include "opt_kern_tls.h" 42 #include <sys/param.h> 43 #include <sys/arb.h> 44 #include <sys/module.h> 45 #include <sys/kernel.h> 46 #ifdef TCP_HHOOK 47 #include <sys/hhook.h> 48 #endif 49 #include <sys/malloc.h> 50 #include <sys/mbuf.h> 51 #include <sys/proc.h> 52 #include <sys/qmath.h> 53 #include <sys/socket.h> 54 #include <sys/socketvar.h> 55 #ifdef KERN_TLS 56 #include <sys/ktls.h> 57 #endif 58 #include <sys/sysctl.h> 59 #include <sys/systm.h> 60 #include <sys/tree.h> 61 #ifdef NETFLIX_STATS 62 #include <sys/stats.h> /* Must come after qmath.h and tree.h */ 63 #endif 64 #include <sys/refcount.h> 65 #include <sys/queue.h> 66 #include <sys/smp.h> 67 #include <sys/kthread.h> 68 #include <sys/lock.h> 69 #include <sys/mutex.h> 70 #include <sys/tim_filter.h> 71 #include <sys/time.h> 72 #include <vm/uma.h> 73 #include <sys/kern_prefetch.h> 74 75 #include <net/route.h> 76 #include <net/vnet.h> 77 #include <net/ethernet.h> 78 #include <net/bpf.h> 79 80 #define TCPSTATES /* for logging */ 81 82 #include <netinet/in.h> 83 #include <netinet/in_kdtrace.h> 84 #include <netinet/in_pcb.h> 85 #include <netinet/ip.h> 86 #include <netinet/ip_icmp.h> /* required for icmp_var.h */ 87 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */ 88 #include <netinet/ip_var.h> 89 #include <netinet/ip6.h> 90 #include <netinet6/in6_pcb.h> 91 #include <netinet6/ip6_var.h> 92 #include <netinet/tcp.h> 93 #include <netinet/tcp_fsm.h> 94 #include <netinet/tcp_seq.h> 95 #include <netinet/tcp_timer.h> 96 #include <netinet/tcp_var.h> 97 #include <netinet/tcpip.h> 98 #include <netinet/tcp_hpts.h> 99 #include <netinet/tcp_lro.h> 100 #include <netinet/cc/cc.h> 101 #include <netinet/tcp_log_buf.h> 102 #ifdef TCPDEBUG 103 #include <netinet/tcp_debug.h> 104 #endif /* TCPDEBUG */ 105 #ifdef TCP_OFFLOAD 106 #include <netinet/tcp_offload.h> 107 #endif 108 #ifdef INET6 109 #include <netinet6/tcp6_var.h> 110 #endif 111 #include <netinet/tcp_fastopen.h> 112 113 #include <netipsec/ipsec_support.h> 114 #include <net/if.h> 115 #include <net/if_var.h> 116 117 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 118 #include <netipsec/ipsec.h> 119 #include <netipsec/ipsec6.h> 120 #endif /* IPSEC */ 121 122 #include <netinet/udp.h> 123 #include <netinet/udp_var.h> 124 #include <machine/in_cksum.h> 125 126 #ifdef MAC 127 #include <security/mac/mac_framework.h> 128 #endif 129 #include "rack_bbr_common.h" 130 131 /* 132 * Common TCP Functions - These are shared by borth 133 * rack and BBR. 134 */ 135 #ifdef KERN_TLS 136 uint32_t 137 ctf_get_opt_tls_size(struct socket *so, uint32_t rwnd) 138 { 139 struct ktls_session *tls; 140 uint32_t len; 141 142 again: 143 tls = so->so_snd.sb_tls_info; 144 len = tls->params.max_frame_len; /* max tls payload */ 145 len += tls->params.tls_hlen; /* tls header len */ 146 len += tls->params.tls_tlen; /* tls trailer len */ 147 if ((len * 4) > rwnd) { 148 /* 149 * Stroke this will suck counter and what 150 * else should we do Drew? From the 151 * TCP perspective I am not sure 152 * what should be done... 153 */ 154 if (tls->params.max_frame_len > 4096) { 155 tls->params.max_frame_len -= 4096; 156 if (tls->params.max_frame_len < 4096) 157 tls->params.max_frame_len = 4096; 158 goto again; 159 } 160 } 161 return (len); 162 } 163 #endif 164 165 static int 166 ctf_get_enet_type(struct ifnet *ifp, struct mbuf *m) 167 { 168 struct ether_header *eh; 169 struct tcphdr *th; 170 #ifdef INET6 171 struct ip6_hdr *ip6 = NULL; /* Keep compiler happy. */ 172 #endif 173 #ifdef INET 174 struct ip *ip = NULL; /* Keep compiler happy. */ 175 #endif 176 int32_t tlen; 177 uint16_t drop_hdrlen; 178 uint16_t etype; 179 uint8_t iptos; 180 181 /* Is it the easy way? */ 182 if (m->m_flags & M_LRO_EHDRSTRP) 183 return (m->m_pkthdr.lro_etype); 184 /* 185 * Ok this is the old style call, the ethernet header is here. 186 * This also means no checksum or BPF were done. This 187 * can happen if the race to setup the inp fails and 188 * LRO sees no INP at packet input, but by the time 189 * we queue the packets an INP gets there. Its rare 190 * but it can occur so we will handle it. Note that 191 * this means duplicated work but with the rarity of it 192 * its not worth worrying about. 193 */ 194 /* Let the BPF see the packet */ 195 if (bpf_peers_present(ifp->if_bpf)) 196 ETHER_BPF_MTAP(ifp, m); 197 /* Now the csum */ 198 eh = mtod(m, struct ether_header *); 199 etype = ntohs(eh->ether_type); 200 m_adj(m, sizeof(*eh)); 201 switch (etype) { 202 #ifdef INET6 203 case ETHERTYPE_IPV6: 204 { 205 if (m->m_len < (sizeof(*ip6) + sizeof(*th))) { 206 m = m_pullup(m, sizeof(*ip6) + sizeof(*th)); 207 if (m == NULL) { 208 KMOD_TCPSTAT_INC(tcps_rcvshort); 209 m_freem(m); 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 m_freem(m); 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 tfb_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; 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 NET_EPOCH_ASSERT(); 404 if (m) 405 ifp = m_rcvif(m); 406 else 407 ifp = NULL; 408 if (ifp == NULL) { 409 /* 410 * We probably should not work around 411 * but kassert, since lro alwasy sets rcvif. 412 */ 413 no_vn = 1; 414 goto skip_vnet; 415 } 416 CURVNET_SET(ifp->if_vnet); 417 skip_vnet: 418 tcp_get_usecs(&tv); 419 while (m) { 420 m_save = m->m_nextpkt; 421 m->m_nextpkt = NULL; 422 if ((m->m_flags & M_ACKCMP) == 0) { 423 /* Now lets get the ether header */ 424 etype = ctf_get_enet_type(ifp, m); 425 if (etype == -1) { 426 /* Skip this packet it was freed by checksum */ 427 goto skipped_pkt; 428 } 429 KASSERT(((etype == ETHERTYPE_IPV6) || (etype == ETHERTYPE_IP)), 430 ("tp:%p m:%p etype:0x%x -- not IP or IPv6", tp, m, etype)); 431 /* Trim off the ethernet header */ 432 switch (etype) { 433 #ifdef INET6 434 case ETHERTYPE_IPV6: 435 ip6 = mtod(m, struct ip6_hdr *); 436 th = (struct tcphdr *)(ip6 + 1); 437 tlen = ntohs(ip6->ip6_plen); 438 drop_hdrlen = sizeof(*ip6); 439 iptos = (ntohl(ip6->ip6_flow) >> 20) & 0xff; 440 break; 441 #endif 442 #ifdef INET 443 case ETHERTYPE_IP: 444 ip = mtod(m, struct ip *); 445 th = (struct tcphdr *)(ip + 1); 446 drop_hdrlen = sizeof(*ip); 447 iptos = ip->ip_tos; 448 tlen = ntohs(ip->ip_len) - sizeof(struct ip); 449 break; 450 #endif 451 } /* end switch */ 452 /* 453 * Convert TCP protocol specific fields to host format. 454 */ 455 tcp_fields_to_host(th); 456 off = th->th_off << 2; 457 if (off < sizeof (struct tcphdr) || off > tlen) { 458 printf("off:%d < hdrlen:%zu || > tlen:%u -- dump\n", 459 off, 460 sizeof(struct tcphdr), 461 tlen); 462 KMOD_TCPSTAT_INC(tcps_rcvbadoff); 463 m_freem(m); 464 goto skipped_pkt; 465 } 466 tlen -= off; 467 drop_hdrlen += off; 468 /* 469 * Now lets setup the timeval to be when we should 470 * have been called (if we can). 471 */ 472 m->m_pkthdr.lro_nsegs = 1; 473 /* Now what about next packet? */ 474 } else { 475 /* 476 * This mbuf is an array of acks that have 477 * been compressed. We assert the inp has 478 * the flag set to enable this! 479 */ 480 KASSERT((tp->t_inpcb->inp_flags2 & INP_MBUF_ACKCMP), 481 ("tp:%p inp:%p no INP_MBUF_ACKCMP flags?", tp, tp->t_inpcb)); 482 tlen = 0; 483 drop_hdrlen = 0; 484 th = NULL; 485 iptos = 0; 486 } 487 tcp_get_usecs(&tv); 488 if (m_save || has_pkt) 489 nxt_pkt = 1; 490 else 491 nxt_pkt = 0; 492 if ((m->m_flags & M_ACKCMP) == 0) 493 KMOD_TCPSTAT_INC(tcps_rcvtotal); 494 else 495 KMOD_TCPSTAT_ADD(tcps_rcvtotal, (m->m_len / sizeof(struct tcp_ackent))); 496 inp = tp->t_inpcb; 497 INP_WLOCK_ASSERT(inp); 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 INP_UNLOCK_ASSERT(inp); 512 return(retval); 513 } 514 skipped_pkt: 515 m = m_save; 516 } 517 if (no_vn == 0) 518 CURVNET_RESTORE(); 519 return(retval); 520 } 521 522 int 523 ctf_do_queued_segments(struct socket *so, struct tcpcb *tp, int have_pkt) 524 { 525 struct mbuf *m; 526 527 /* First lets see if we have old packets */ 528 if (tp->t_in_pkt) { 529 m = tp->t_in_pkt; 530 tp->t_in_pkt = NULL; 531 tp->t_tail_pkt = NULL; 532 if (ctf_process_inbound_raw(tp, so, m, have_pkt)) { 533 /* We lost the tcpcb (maybe a RST came in)? */ 534 return(1); 535 } 536 } 537 return (0); 538 } 539 540 uint32_t 541 ctf_outstanding(struct tcpcb *tp) 542 { 543 uint32_t bytes_out; 544 545 bytes_out = tp->snd_max - tp->snd_una; 546 if (tp->t_state < TCPS_ESTABLISHED) 547 bytes_out++; 548 if (tp->t_flags & TF_SENTFIN) 549 bytes_out++; 550 return (bytes_out); 551 } 552 553 uint32_t 554 ctf_flight_size(struct tcpcb *tp, uint32_t rc_sacked) 555 { 556 if (rc_sacked <= ctf_outstanding(tp)) 557 return(ctf_outstanding(tp) - rc_sacked); 558 else { 559 return (0); 560 } 561 } 562 563 void 564 ctf_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, 565 int32_t rstreason, int32_t tlen) 566 { 567 if (tp != NULL) { 568 tcp_dropwithreset(m, th, tp, tlen, rstreason); 569 INP_WUNLOCK(tp->t_inpcb); 570 } else 571 tcp_dropwithreset(m, th, NULL, tlen, rstreason); 572 } 573 574 void 575 ctf_ack_war_checks(struct tcpcb *tp, uint32_t *ts, uint32_t *cnt) 576 { 577 if ((ts != NULL) && (cnt != NULL) && 578 (tcp_ack_war_time_window > 0) && 579 (tcp_ack_war_cnt > 0)) { 580 /* We are possibly doing ack war prevention */ 581 uint32_t cts; 582 583 /* 584 * We use a msec tick here which gives us 585 * roughly 49 days. We don't need the 586 * precision of a microsecond timestamp which 587 * would only give us hours. 588 */ 589 cts = tcp_ts_getticks(); 590 if (TSTMP_LT((*ts), cts)) { 591 /* Timestamp is in the past */ 592 *cnt = 0; 593 *ts = (cts + tcp_ack_war_time_window); 594 } 595 if (*cnt < tcp_ack_war_cnt) { 596 *cnt = (*cnt + 1); 597 tp->t_flags |= TF_ACKNOW; 598 } else 599 tp->t_flags &= ~TF_ACKNOW; 600 } else 601 tp->t_flags |= TF_ACKNOW; 602 } 603 604 /* 605 * ctf_drop_checks returns 1 for you should not proceed. It places 606 * in ret_val what should be returned 1/0 by the caller. The 1 indicates 607 * that the TCB is unlocked and probably dropped. The 0 indicates the 608 * TCB is still valid and locked. 609 */ 610 int 611 _ctf_drop_checks(struct tcpopt *to, struct mbuf *m, struct tcphdr *th, 612 struct tcpcb *tp, int32_t *tlenp, 613 int32_t *thf, int32_t *drop_hdrlen, int32_t *ret_val, 614 uint32_t *ts, uint32_t *cnt) 615 { 616 int32_t todrop; 617 int32_t thflags; 618 int32_t tlen; 619 620 thflags = *thf; 621 tlen = *tlenp; 622 todrop = tp->rcv_nxt - th->th_seq; 623 if (todrop > 0) { 624 if (thflags & TH_SYN) { 625 thflags &= ~TH_SYN; 626 th->th_seq++; 627 if (th->th_urp > 1) 628 th->th_urp--; 629 else 630 thflags &= ~TH_URG; 631 todrop--; 632 } 633 /* 634 * Following if statement from Stevens, vol. 2, p. 960. 635 */ 636 if (todrop > tlen 637 || (todrop == tlen && (thflags & TH_FIN) == 0)) { 638 /* 639 * Any valid FIN must be to the left of the window. 640 * At this point the FIN must be a duplicate or out 641 * of sequence; drop it. 642 */ 643 thflags &= ~TH_FIN; 644 /* 645 * Send an ACK to resynchronize and drop any data. 646 * But keep on processing for RST or ACK. 647 */ 648 ctf_ack_war_checks(tp, ts, cnt); 649 todrop = tlen; 650 KMOD_TCPSTAT_INC(tcps_rcvduppack); 651 KMOD_TCPSTAT_ADD(tcps_rcvdupbyte, todrop); 652 } else { 653 KMOD_TCPSTAT_INC(tcps_rcvpartduppack); 654 KMOD_TCPSTAT_ADD(tcps_rcvpartdupbyte, todrop); 655 } 656 /* 657 * DSACK - add SACK block for dropped range 658 */ 659 if ((todrop > 0) && (tp->t_flags & TF_SACK_PERMIT)) { 660 /* 661 * ACK now, as the next in-sequence segment 662 * will clear the DSACK block again 663 */ 664 ctf_ack_war_checks(tp, ts, cnt); 665 if (tp->t_flags & TF_ACKNOW) 666 tcp_update_sack_list(tp, th->th_seq, 667 th->th_seq + todrop); 668 } 669 *drop_hdrlen += todrop; /* drop from the top afterwards */ 670 th->th_seq += todrop; 671 tlen -= todrop; 672 if (th->th_urp > todrop) 673 th->th_urp -= todrop; 674 else { 675 thflags &= ~TH_URG; 676 th->th_urp = 0; 677 } 678 } 679 /* 680 * If segment ends after window, drop trailing data (and PUSH and 681 * FIN); if nothing left, just ACK. 682 */ 683 todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd); 684 if (todrop > 0) { 685 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin); 686 if (todrop >= tlen) { 687 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, tlen); 688 /* 689 * If window is closed can only take segments at 690 * window edge, and have to drop data and PUSH from 691 * incoming segments. Continue processing, but 692 * remember to ack. Otherwise, drop segment and 693 * ack. 694 */ 695 if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) { 696 ctf_ack_war_checks(tp, ts, cnt); 697 KMOD_TCPSTAT_INC(tcps_rcvwinprobe); 698 } else { 699 __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val, ts, cnt); 700 return (1); 701 } 702 } else 703 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 704 m_adj(m, -todrop); 705 tlen -= todrop; 706 thflags &= ~(TH_PUSH | TH_FIN); 707 } 708 *thf = thflags; 709 *tlenp = tlen; 710 return (0); 711 } 712 713 /* 714 * The value in ret_val informs the caller 715 * if we dropped the tcb (and lock) or not. 716 * 1 = we dropped it, 0 = the TCB is still locked 717 * and valid. 718 */ 719 void 720 __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) 721 { 722 /* 723 * Generate an ACK dropping incoming segment if it occupies sequence 724 * space, where the ACK reflects our state. 725 * 726 * We can now skip the test for the RST flag since all paths to this 727 * code happen after packets containing RST have been dropped. 728 * 729 * In the SYN-RECEIVED state, don't send an ACK unless the segment 730 * we received passes the SYN-RECEIVED ACK test. If it fails send a 731 * RST. This breaks the loop in the "LAND" DoS attack, and also 732 * prevents an ACK storm between two listening ports that have been 733 * sent forged SYN segments, each with the source address of the 734 * other. 735 */ 736 if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) && 737 (SEQ_GT(tp->snd_una, th->th_ack) || 738 SEQ_GT(th->th_ack, tp->snd_max))) { 739 *ret_val = 1; 740 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 741 return; 742 } else 743 *ret_val = 0; 744 ctf_ack_war_checks(tp, ts, cnt); 745 if (m) 746 m_freem(m); 747 } 748 749 void 750 ctf_do_drop(struct mbuf *m, struct tcpcb *tp) 751 { 752 753 /* 754 * Drop space held by incoming segment and return. 755 */ 756 if (tp != NULL) 757 INP_WUNLOCK(tp->t_inpcb); 758 if (m) 759 m_freem(m); 760 } 761 762 int 763 ctf_process_rst(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp) 764 { 765 /* 766 * RFC5961 Section 3.2 767 * 768 * - RST drops connection only if SEG.SEQ == RCV.NXT. - If RST is in 769 * window, we send challenge ACK. 770 * 771 * Note: to take into account delayed ACKs, we should test against 772 * last_ack_sent instead of rcv_nxt. Note 2: we handle special case 773 * of closed window, not covered by the RFC. 774 */ 775 int dropped = 0; 776 777 if ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) && 778 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) || 779 (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq)) { 780 KASSERT(tp->t_state != TCPS_SYN_SENT, 781 ("%s: TH_RST for TCPS_SYN_SENT th %p tp %p", 782 __func__, th, tp)); 783 784 if (V_tcp_insecure_rst || 785 (tp->last_ack_sent == th->th_seq) || 786 (tp->rcv_nxt == th->th_seq)) { 787 KMOD_TCPSTAT_INC(tcps_drops); 788 /* Drop the connection. */ 789 switch (tp->t_state) { 790 case TCPS_SYN_RECEIVED: 791 so->so_error = ECONNREFUSED; 792 goto close; 793 case TCPS_ESTABLISHED: 794 case TCPS_FIN_WAIT_1: 795 case TCPS_FIN_WAIT_2: 796 case TCPS_CLOSE_WAIT: 797 case TCPS_CLOSING: 798 case TCPS_LAST_ACK: 799 so->so_error = ECONNRESET; 800 close: 801 tcp_state_change(tp, TCPS_CLOSED); 802 /* FALLTHROUGH */ 803 default: 804 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_RST); 805 tp = tcp_close(tp); 806 } 807 dropped = 1; 808 ctf_do_drop(m, tp); 809 } else { 810 KMOD_TCPSTAT_INC(tcps_badrst); 811 /* Send challenge ACK. */ 812 tcp_respond(tp, mtod(m, void *), th, m, 813 tp->rcv_nxt, tp->snd_nxt, TH_ACK); 814 tp->last_ack_sent = tp->rcv_nxt; 815 } 816 } else { 817 m_freem(m); 818 } 819 return (dropped); 820 } 821 822 /* 823 * The value in ret_val informs the caller 824 * if we dropped the tcb (and lock) or not. 825 * 1 = we dropped it, 0 = the TCB is still locked 826 * and valid. 827 */ 828 void 829 ctf_challenge_ack(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t * ret_val) 830 { 831 832 NET_EPOCH_ASSERT(); 833 834 KMOD_TCPSTAT_INC(tcps_badsyn); 835 if (V_tcp_insecure_syn && 836 SEQ_GEQ(th->th_seq, tp->last_ack_sent) && 837 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) { 838 tp = tcp_drop(tp, ECONNRESET); 839 *ret_val = 1; 840 ctf_do_drop(m, tp); 841 } else { 842 /* Send challenge ACK. */ 843 tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt, 844 tp->snd_nxt, TH_ACK); 845 tp->last_ack_sent = tp->rcv_nxt; 846 m = NULL; 847 *ret_val = 0; 848 ctf_do_drop(m, NULL); 849 } 850 } 851 852 /* 853 * ctf_ts_check returns 1 for you should not proceed, the state 854 * machine should return. It places in ret_val what should 855 * be returned 1/0 by the caller (hpts_do_segment). The 1 indicates 856 * that the TCB is unlocked and probably dropped. The 0 indicates the 857 * TCB is still valid and locked. 858 */ 859 int 860 ctf_ts_check(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, 861 int32_t tlen, int32_t thflags, int32_t * ret_val) 862 { 863 864 if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { 865 /* 866 * Invalidate ts_recent. If this segment updates ts_recent, 867 * the age will be reset later and ts_recent will get a 868 * valid value. If it does not, setting ts_recent to zero 869 * will at least satisfy the requirement that zero be placed 870 * in the timestamp echo reply when ts_recent isn't valid. 871 * The age isn't reset until we get a valid ts_recent 872 * because we don't want out-of-order segments to be dropped 873 * when ts_recent is old. 874 */ 875 tp->ts_recent = 0; 876 } else { 877 KMOD_TCPSTAT_INC(tcps_rcvduppack); 878 KMOD_TCPSTAT_ADD(tcps_rcvdupbyte, tlen); 879 KMOD_TCPSTAT_INC(tcps_pawsdrop); 880 *ret_val = 0; 881 if (tlen) { 882 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 883 } else { 884 ctf_do_drop(m, NULL); 885 } 886 return (1); 887 } 888 return (0); 889 } 890 891 int 892 ctf_ts_check_ac(struct tcpcb *tp, int32_t thflags) 893 { 894 895 if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { 896 /* 897 * Invalidate ts_recent. If this segment updates ts_recent, 898 * the age will be reset later and ts_recent will get a 899 * valid value. If it does not, setting ts_recent to zero 900 * will at least satisfy the requirement that zero be placed 901 * in the timestamp echo reply when ts_recent isn't valid. 902 * The age isn't reset until we get a valid ts_recent 903 * because we don't want out-of-order segments to be dropped 904 * when ts_recent is old. 905 */ 906 tp->ts_recent = 0; 907 } else { 908 KMOD_TCPSTAT_INC(tcps_rcvduppack); 909 KMOD_TCPSTAT_INC(tcps_pawsdrop); 910 return (1); 911 } 912 return (0); 913 } 914 915 916 917 void 918 ctf_calc_rwin(struct socket *so, struct tcpcb *tp) 919 { 920 int32_t win; 921 922 /* 923 * Calculate amount of space in receive window, and then do TCP 924 * input processing. Receive window is amount of space in rcv queue, 925 * but not less than advertised window. 926 */ 927 win = sbspace(&so->so_rcv); 928 if (win < 0) 929 win = 0; 930 tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt)); 931 } 932 933 void 934 ctf_do_dropwithreset_conn(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, 935 int32_t rstreason, int32_t tlen) 936 { 937 938 if (tp->t_inpcb) { 939 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); 940 } 941 tcp_dropwithreset(m, th, tp, tlen, rstreason); 942 INP_WUNLOCK(tp->t_inpcb); 943 } 944 945 uint32_t 946 ctf_fixed_maxseg(struct tcpcb *tp) 947 { 948 return (tcp_fixed_maxseg(tp)); 949 } 950 951 void 952 ctf_log_sack_filter(struct tcpcb *tp, int num_sack_blks, struct sackblk *sack_blocks) 953 { 954 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 955 union tcp_log_stackspecific log; 956 struct timeval tv; 957 958 memset(&log, 0, sizeof(log)); 959 log.u_bbr.timeStamp = tcp_get_usecs(&tv); 960 log.u_bbr.flex8 = num_sack_blks; 961 if (num_sack_blks > 0) { 962 log.u_bbr.flex1 = sack_blocks[0].start; 963 log.u_bbr.flex2 = sack_blocks[0].end; 964 } 965 if (num_sack_blks > 1) { 966 log.u_bbr.flex3 = sack_blocks[1].start; 967 log.u_bbr.flex4 = sack_blocks[1].end; 968 } 969 if (num_sack_blks > 2) { 970 log.u_bbr.flex5 = sack_blocks[2].start; 971 log.u_bbr.flex6 = sack_blocks[2].end; 972 } 973 if (num_sack_blks > 3) { 974 log.u_bbr.applimited = sack_blocks[3].start; 975 log.u_bbr.pkts_out = sack_blocks[3].end; 976 } 977 TCP_LOG_EVENTP(tp, NULL, 978 &tp->t_inpcb->inp_socket->so_rcv, 979 &tp->t_inpcb->inp_socket->so_snd, 980 TCP_SACK_FILTER_RES, 0, 981 0, &log, false, &tv); 982 } 983 } 984 985 uint32_t 986 ctf_decay_count(uint32_t count, uint32_t decay) 987 { 988 /* 989 * Given a count, decay it by a set percentage. The 990 * percentage is in thousands i.e. 100% = 1000, 991 * 19.3% = 193. 992 */ 993 uint64_t perc_count, decay_per; 994 uint32_t decayed_count; 995 if (decay > 1000) { 996 /* We don't raise it */ 997 return (count); 998 } 999 perc_count = count; 1000 decay_per = decay; 1001 perc_count *= decay_per; 1002 perc_count /= 1000; 1003 /* 1004 * So now perc_count holds the 1005 * count decay value. 1006 */ 1007 decayed_count = count - (uint32_t)perc_count; 1008 return(decayed_count); 1009 } 1010 1011 int32_t 1012 ctf_progress_timeout_check(struct tcpcb *tp, bool log) 1013 { 1014 if (tp->t_maxunacktime && tp->t_acktime && TSTMP_GT(ticks, tp->t_acktime)) { 1015 if ((ticks - tp->t_acktime) >= tp->t_maxunacktime) { 1016 /* 1017 * There is an assumption that the caller 1018 * will drop the connection so we will 1019 * increment the counters here. 1020 */ 1021 if (log) 1022 tcp_log_end_status(tp, TCP_EI_STATUS_PROGRESS); 1023 #ifdef NETFLIX_STATS 1024 KMOD_TCPSTAT_INC(tcps_progdrops); 1025 #endif 1026 return (1); 1027 } 1028 } 1029 return (0); 1030 } 1031