1 /*- 2 * Copyright (c) 2016-2020 Netflix, Inc. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. Redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution. 12 * 13 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 16 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 23 * SUCH DAMAGE. 24 * 25 */ 26 /* 27 * Author: Randall Stewart <rrs@netflix.com> 28 * This work is based on the ACM Queue paper 29 * BBR - Congestion Based Congestion Control 30 * and also numerous discussions with Neal, Yuchung and Van. 31 */ 32 33 #include <sys/cdefs.h> 34 #include "opt_inet.h" 35 #include "opt_inet6.h" 36 #include "opt_ipsec.h" 37 #include "opt_ratelimit.h" 38 #include <sys/param.h> 39 #include <sys/arb.h> 40 #include <sys/module.h> 41 #include <sys/kernel.h> 42 #ifdef TCP_HHOOK 43 #include <sys/hhook.h> 44 #endif 45 #include <sys/malloc.h> 46 #include <sys/mbuf.h> 47 #include <sys/proc.h> 48 #include <sys/qmath.h> 49 #include <sys/socket.h> 50 #include <sys/socketvar.h> 51 #include <sys/sysctl.h> 52 #include <sys/systm.h> 53 #include <sys/tree.h> 54 #ifdef NETFLIX_STATS 55 #include <sys/stats.h> /* Must come after qmath.h and tree.h */ 56 #endif 57 #include <sys/refcount.h> 58 #include <sys/queue.h> 59 #include <sys/smp.h> 60 #include <sys/kthread.h> 61 #include <sys/lock.h> 62 #include <sys/mutex.h> 63 #include <sys/tim_filter.h> 64 #include <sys/time.h> 65 #include <vm/uma.h> 66 #include <sys/kern_prefetch.h> 67 68 #include <net/route.h> 69 #include <net/vnet.h> 70 #include <net/ethernet.h> 71 #include <net/bpf.h> 72 73 #define TCPSTATES /* for logging */ 74 75 #include <netinet/in.h> 76 #include <netinet/in_kdtrace.h> 77 #include <netinet/in_pcb.h> 78 #include <netinet/ip.h> 79 #include <netinet/ip_icmp.h> /* required for icmp_var.h */ 80 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */ 81 #include <netinet/ip_var.h> 82 #include <netinet/ip6.h> 83 #include <netinet6/in6_pcb.h> 84 #include <netinet6/ip6_var.h> 85 #include <netinet/tcp.h> 86 #include <netinet/tcp_fsm.h> 87 #include <netinet/tcp_seq.h> 88 #include <netinet/tcp_timer.h> 89 #include <netinet/tcp_var.h> 90 #include <netinet/tcpip.h> 91 #include <netinet/tcp_ecn.h> 92 #include <netinet/tcp_hpts.h> 93 #include <netinet/tcp_lro.h> 94 #include <netinet/cc/cc.h> 95 #include <netinet/tcp_log_buf.h> 96 #ifdef TCP_OFFLOAD 97 #include <netinet/tcp_offload.h> 98 #endif 99 #ifdef INET6 100 #include <netinet6/tcp6_var.h> 101 #endif 102 #include <netinet/tcp_fastopen.h> 103 104 #include <netipsec/ipsec_support.h> 105 #include <net/if.h> 106 #include <net/if_var.h> 107 #include <net/if_private.h> 108 109 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 110 #include <netipsec/ipsec.h> 111 #include <netipsec/ipsec6.h> 112 #endif /* IPSEC */ 113 114 #include <netinet/udp.h> 115 #include <netinet/udp_var.h> 116 #include <machine/in_cksum.h> 117 118 #ifdef MAC 119 #include <security/mac/mac_framework.h> 120 #endif 121 #include "rack_bbr_common.h" 122 123 /* 124 * Common TCP Functions - These are shared by borth 125 * rack and BBR. 126 */ 127 static int 128 ctf_get_enet_type(struct ifnet *ifp, struct mbuf *m) 129 { 130 struct ether_header *eh; 131 #ifdef INET6 132 struct ip6_hdr *ip6 = NULL; /* Keep compiler happy. */ 133 #endif 134 #ifdef INET 135 struct ip *ip = NULL; /* Keep compiler happy. */ 136 #endif 137 #if defined(INET) || defined(INET6) 138 struct tcphdr *th; 139 int32_t tlen; 140 uint16_t drop_hdrlen; 141 #endif 142 uint16_t etype; 143 #ifdef INET 144 uint8_t iptos; 145 #endif 146 147 /* Is it the easy way? */ 148 if (m->m_flags & M_LRO_EHDRSTRP) 149 return (m->m_pkthdr.lro_etype); 150 /* 151 * Ok this is the old style call, the ethernet header is here. 152 * This also means no checksum or BPF were done. This 153 * can happen if the race to setup the inp fails and 154 * LRO sees no INP at packet input, but by the time 155 * we queue the packets an INP gets there. Its rare 156 * but it can occur so we will handle it. Note that 157 * this means duplicated work but with the rarity of it 158 * its not worth worrying about. 159 */ 160 /* Let the BPF see the packet */ 161 if (bpf_peers_present(ifp->if_bpf)) 162 ETHER_BPF_MTAP(ifp, m); 163 /* Now the csum */ 164 eh = mtod(m, struct ether_header *); 165 etype = ntohs(eh->ether_type); 166 m_adj(m, sizeof(*eh)); 167 switch (etype) { 168 #ifdef INET6 169 case ETHERTYPE_IPV6: 170 { 171 if (m->m_len < (sizeof(*ip6) + sizeof(*th))) { 172 m = m_pullup(m, sizeof(*ip6) + sizeof(*th)); 173 if (m == NULL) { 174 KMOD_TCPSTAT_INC(tcps_rcvshort); 175 return (-1); 176 } 177 } 178 ip6 = (struct ip6_hdr *)(eh + 1); 179 th = (struct tcphdr *)(ip6 + 1); 180 drop_hdrlen = sizeof(*ip6); 181 tlen = ntohs(ip6->ip6_plen); 182 if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID_IPV6) { 183 if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) 184 th->th_sum = m->m_pkthdr.csum_data; 185 else 186 th->th_sum = in6_cksum_pseudo(ip6, tlen, 187 IPPROTO_TCP, 188 m->m_pkthdr.csum_data); 189 th->th_sum ^= 0xffff; 190 } else 191 th->th_sum = in6_cksum(m, IPPROTO_TCP, drop_hdrlen, tlen); 192 if (th->th_sum) { 193 KMOD_TCPSTAT_INC(tcps_rcvbadsum); 194 m_freem(m); 195 return (-1); 196 } 197 return (etype); 198 } 199 #endif 200 #ifdef INET 201 case ETHERTYPE_IP: 202 { 203 if (m->m_len < sizeof (struct tcpiphdr)) { 204 m = m_pullup(m, sizeof (struct tcpiphdr)); 205 if (m == NULL) { 206 KMOD_TCPSTAT_INC(tcps_rcvshort); 207 return (-1); 208 } 209 } 210 ip = (struct ip *)(eh + 1); 211 th = (struct tcphdr *)(ip + 1); 212 drop_hdrlen = sizeof(*ip); 213 iptos = ip->ip_tos; 214 tlen = ntohs(ip->ip_len) - sizeof(struct ip); 215 if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) { 216 if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) 217 th->th_sum = m->m_pkthdr.csum_data; 218 else 219 th->th_sum = in_pseudo(ip->ip_src.s_addr, 220 ip->ip_dst.s_addr, 221 htonl(m->m_pkthdr.csum_data + tlen + IPPROTO_TCP)); 222 th->th_sum ^= 0xffff; 223 } else { 224 int len; 225 struct ipovly *ipov = (struct ipovly *)ip; 226 /* 227 * Checksum extended TCP header and data. 228 */ 229 len = drop_hdrlen + tlen; 230 bzero(ipov->ih_x1, sizeof(ipov->ih_x1)); 231 ipov->ih_len = htons(tlen); 232 th->th_sum = in_cksum(m, len); 233 /* Reset length for SDT probes. */ 234 ip->ip_len = htons(len); 235 /* Reset TOS bits */ 236 ip->ip_tos = iptos; 237 /* Re-initialization for later version check */ 238 ip->ip_v = IPVERSION; 239 ip->ip_hl = sizeof(*ip) >> 2; 240 } 241 if (th->th_sum) { 242 KMOD_TCPSTAT_INC(tcps_rcvbadsum); 243 m_freem(m); 244 return (-1); 245 } 246 break; 247 } 248 #endif 249 }; 250 return (etype); 251 } 252 253 /* 254 * The function ctf_process_inbound_raw() is used by 255 * transport developers to do the steps needed to 256 * support MBUF Queuing i.e. the flags in 257 * inp->inp_flags2: 258 * 259 * - INP_SUPPORTS_MBUFQ 260 * - INP_MBUF_QUEUE_READY 261 * - INP_DONT_SACK_QUEUE 262 * - INP_MBUF_ACKCMP 263 * 264 * These flags help control how LRO will deliver 265 * packets to the transport. You first set in inp_flags2 266 * the INP_SUPPORTS_MBUFQ to tell the LRO code that you 267 * will gladly take a queue of packets instead of a compressed 268 * single packet. You also set in your t_fb pointer the 269 * tfb_do_queued_segments to point to ctf_process_inbound_raw. 270 * 271 * This then gets you lists of inbound ACK's/Data instead 272 * of a condensed compressed ACK/DATA packet. Why would you 273 * want that? This will get you access to all the arrival 274 * times of at least LRO and possibly at the Hardware (if 275 * the interface card supports that) of the actual ACK/DATA. 276 * In some transport designs this is important since knowing 277 * the actual time we got the packet is useful information. 278 * 279 * A new special type of mbuf may also be supported by the transport 280 * if it has set the INP_MBUF_ACKCMP flag. If its set, LRO will 281 * possibly create a M_ACKCMP type mbuf. This is a mbuf with 282 * an array of "acks". One thing also to note is that when this 283 * occurs a subsequent LRO may find at the back of the untouched 284 * mbuf queue chain a M_ACKCMP and append on to it. This means 285 * that until the transport pulls in the mbuf chain queued 286 * for it more ack's may get on the mbufs that were already 287 * delivered. There currently is a limit of 6 acks condensed 288 * into 1 mbuf which means often when this is occuring, we 289 * don't get that effect but it does happen. 290 * 291 * Now there are some interesting Caveats that the transport 292 * designer needs to take into account when using this feature. 293 * 294 * 1) It is used with HPTS and pacing, when the pacing timer 295 * for output calls it will first call the input. 296 * 2) When you set INP_MBUF_QUEUE_READY this tells LRO 297 * queue normal packets, I am busy pacing out data and 298 * will process the queued packets before my tfb_tcp_output 299 * call from pacing. If a non-normal packet arrives, (e.g. sack) 300 * you will be awoken immediately. 301 * 3) Finally you can add the INP_DONT_SACK_QUEUE to not even 302 * be awoken if a SACK has arrived. You would do this when 303 * you were not only running a pacing for output timer 304 * but a Rack timer as well i.e. you know you are in recovery 305 * and are in the process (via the timers) of dealing with 306 * the loss. 307 * 308 * Now a critical thing you must be aware of here is that the 309 * use of the flags has a far greater scope then just your 310 * typical LRO. Why? Well thats because in the normal compressed 311 * LRO case at the end of a driver interupt all packets are going 312 * to get presented to the transport no matter if there is one 313 * or 100. With the MBUF_QUEUE model, this is not true. You will 314 * only be awoken to process the queue of packets when: 315 * a) The flags discussed above allow it. 316 * <or> 317 * b) You exceed a ack or data limit (by default the 318 * ack limit is infinity (64k acks) and the data 319 * limit is 64k of new TCP data) 320 * <or> 321 * c) The push bit has been set by the peer 322 */ 323 324 static int 325 ctf_process_inbound_raw(struct tcpcb *tp, struct mbuf *m, int has_pkt) 326 { 327 /* 328 * We are passed a raw change of mbuf packets 329 * that arrived in LRO. They are linked via 330 * the m_nextpkt link in the pkt-headers. 331 * 332 * We process each one by: 333 * a) saving off the next 334 * b) stripping off the ether-header 335 * c) formulating the arguments for tfb_do_segment_nounlock() 336 * d) calling each mbuf to tfb_do_segment_nounlock() 337 * after adjusting the time to match the arrival time. 338 * Note that the LRO code assures no IP options are present. 339 * 340 * The symantics for calling tfb_do_segment_nounlock() are the 341 * following: 342 * 1) It returns 0 if all went well and you (the caller) need 343 * to release the lock. 344 * 2) If nxt_pkt is set, then the function will surpress calls 345 * to tcp_output() since you are promising to call again 346 * with another packet. 347 * 3) If it returns 1, then you must free all the packets being 348 * shipped in, the tcb has been destroyed (or about to be destroyed). 349 */ 350 struct mbuf *m_save; 351 struct tcphdr *th; 352 #ifdef INET6 353 struct ip6_hdr *ip6 = NULL; /* Keep compiler happy. */ 354 #endif 355 #ifdef INET 356 struct ip *ip = NULL; /* Keep compiler happy. */ 357 #endif 358 struct ifnet *ifp; 359 struct timeval tv; 360 struct inpcb *inp __diagused; 361 int32_t retval, nxt_pkt, tlen, off; 362 int etype = 0; 363 uint16_t drop_hdrlen; 364 uint8_t iptos; 365 366 inp = tptoinpcb(tp); 367 INP_WLOCK_ASSERT(inp); 368 NET_EPOCH_ASSERT(); 369 KASSERT(m != NULL, ("ctf_process_inbound_raw: m == NULL")); 370 ifp = m_rcvif(m); 371 KASSERT(ifp != NULL, ("ctf_process_inbound_raw: ifp == NULL")); 372 CURVNET_SET(ifp->if_vnet); 373 tcp_get_usecs(&tv); 374 while (m) { 375 m_save = m->m_nextpkt; 376 m->m_nextpkt = NULL; 377 if ((m->m_flags & M_ACKCMP) == 0) { 378 /* Now lets get the ether header */ 379 etype = ctf_get_enet_type(ifp, m); 380 if (etype == -1) { 381 /* Skip this packet it was freed by checksum */ 382 goto skipped_pkt; 383 } 384 KASSERT(((etype == ETHERTYPE_IPV6) || (etype == ETHERTYPE_IP)), 385 ("tp:%p m:%p etype:0x%x -- not IP or IPv6", tp, m, etype)); 386 /* Trim off the ethernet header */ 387 switch (etype) { 388 #ifdef INET6 389 case ETHERTYPE_IPV6: 390 ip6 = mtod(m, struct ip6_hdr *); 391 th = (struct tcphdr *)(ip6 + 1); 392 tlen = ntohs(ip6->ip6_plen); 393 drop_hdrlen = sizeof(*ip6); 394 iptos = (ntohl(ip6->ip6_flow) >> 20) & 0xff; 395 break; 396 #endif 397 #ifdef INET 398 case ETHERTYPE_IP: 399 ip = mtod(m, struct ip *); 400 th = (struct tcphdr *)(ip + 1); 401 drop_hdrlen = sizeof(*ip); 402 iptos = ip->ip_tos; 403 tlen = ntohs(ip->ip_len) - sizeof(struct ip); 404 break; 405 #endif 406 } /* end switch */ 407 off = th->th_off << 2; 408 if (off < sizeof (struct tcphdr) || off > tlen) { 409 printf("off:%d < hdrlen:%zu || > tlen:%u -- dump\n", 410 off, 411 sizeof(struct tcphdr), 412 tlen); 413 KMOD_TCPSTAT_INC(tcps_rcvbadoff); 414 m_freem(m); 415 goto skipped_pkt; 416 } 417 tlen -= off; 418 drop_hdrlen += off; 419 /* 420 * Now lets setup the timeval to be when we should 421 * have been called (if we can). 422 */ 423 m->m_pkthdr.lro_nsegs = 1; 424 /* Now what about next packet? */ 425 } else { 426 /* 427 * This mbuf is an array of acks that have 428 * been compressed. We assert the inp has 429 * the flag set to enable this! 430 */ 431 KASSERT((tp->t_flags2 & TF2_MBUF_ACKCMP), 432 ("tp:%p no TF2_MBUF_ACKCMP flags?", tp)); 433 tlen = 0; 434 drop_hdrlen = 0; 435 th = NULL; 436 iptos = 0; 437 } 438 tcp_get_usecs(&tv); 439 if (m_save || has_pkt) 440 nxt_pkt = 1; 441 else 442 nxt_pkt = 0; 443 if ((m->m_flags & M_ACKCMP) == 0) 444 KMOD_TCPSTAT_INC(tcps_rcvtotal); 445 else 446 KMOD_TCPSTAT_ADD(tcps_rcvtotal, (m->m_len / sizeof(struct tcp_ackent))); 447 retval = (*tp->t_fb->tfb_do_segment_nounlock)(tp, m, th, 448 drop_hdrlen, tlen, iptos, nxt_pkt, &tv); 449 if (retval) { 450 /* We lost the lock and tcb probably */ 451 m = m_save; 452 while(m) { 453 m_save = m->m_nextpkt; 454 m->m_nextpkt = NULL; 455 m_freem(m); 456 m = m_save; 457 } 458 CURVNET_RESTORE(); 459 INP_UNLOCK_ASSERT(inp); 460 return (retval); 461 } 462 skipped_pkt: 463 m = m_save; 464 } 465 CURVNET_RESTORE(); 466 return (0); 467 } 468 469 int 470 ctf_do_queued_segments(struct tcpcb *tp, int have_pkt) 471 { 472 struct mbuf *m; 473 474 /* First lets see if we have old packets */ 475 if ((m = STAILQ_FIRST(&tp->t_inqueue)) != NULL) { 476 STAILQ_INIT(&tp->t_inqueue); 477 if (ctf_process_inbound_raw(tp, m, have_pkt)) { 478 /* We lost the tcpcb (maybe a RST came in)? */ 479 return(1); 480 } 481 } 482 return (0); 483 } 484 485 uint32_t 486 ctf_outstanding(struct tcpcb *tp) 487 { 488 uint32_t bytes_out; 489 490 bytes_out = tp->snd_max - tp->snd_una; 491 if (tp->t_state < TCPS_ESTABLISHED) 492 bytes_out++; 493 if (tp->t_flags & TF_SENTFIN) 494 bytes_out++; 495 return (bytes_out); 496 } 497 498 uint32_t 499 ctf_flight_size(struct tcpcb *tp, uint32_t rc_sacked) 500 { 501 if (rc_sacked <= ctf_outstanding(tp)) 502 return(ctf_outstanding(tp) - rc_sacked); 503 else { 504 return (0); 505 } 506 } 507 508 void 509 ctf_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, 510 int32_t rstreason, int32_t tlen) 511 { 512 if (tp != NULL) { 513 tcp_dropwithreset(m, th, tp, tlen, rstreason); 514 INP_WUNLOCK(tptoinpcb(tp)); 515 } else 516 tcp_dropwithreset(m, th, NULL, tlen, rstreason); 517 } 518 519 void 520 ctf_ack_war_checks(struct tcpcb *tp) 521 { 522 sbintime_t now; 523 524 if ((V_tcp_ack_war_time_window > 0) && (V_tcp_ack_war_cnt > 0)) { 525 now = getsbinuptime(); 526 if (tp->t_challenge_ack_end < now) { 527 tp->t_challenge_ack_cnt = 0; 528 tp->t_challenge_ack_end = now + 529 V_tcp_ack_war_time_window * SBT_1MS; 530 } 531 if (tp->t_challenge_ack_cnt < V_tcp_ack_war_cnt) { 532 tp->t_challenge_ack_cnt++; 533 tp->t_flags |= TF_ACKNOW; 534 } else 535 tp->t_flags &= ~TF_ACKNOW; 536 } else 537 tp->t_flags |= TF_ACKNOW; 538 } 539 540 /* 541 * ctf_drop_checks returns 1 for you should not proceed. It places 542 * in ret_val what should be returned 1/0 by the caller. The 1 indicates 543 * that the TCB is unlocked and probably dropped. The 0 indicates the 544 * TCB is still valid and locked. 545 */ 546 int 547 ctf_drop_checks(struct tcpopt *to, struct mbuf *m, struct tcphdr *th, 548 struct tcpcb *tp, int32_t *tlenp, 549 int32_t *thf, int32_t *drop_hdrlen, int32_t *ret_val) 550 { 551 int32_t todrop; 552 int32_t thflags; 553 int32_t tlen; 554 555 thflags = *thf; 556 tlen = *tlenp; 557 todrop = tp->rcv_nxt - th->th_seq; 558 if (todrop > 0) { 559 if (thflags & TH_SYN) { 560 thflags &= ~TH_SYN; 561 th->th_seq++; 562 if (th->th_urp > 1) 563 th->th_urp--; 564 else 565 thflags &= ~TH_URG; 566 todrop--; 567 } 568 /* 569 * Following if statement from Stevens, vol. 2, p. 960. 570 */ 571 if (todrop > tlen 572 || (todrop == tlen && (thflags & TH_FIN) == 0)) { 573 /* 574 * Any valid FIN must be to the left of the window. 575 * At this point the FIN must be a duplicate or out 576 * of sequence; drop it. 577 */ 578 thflags &= ~TH_FIN; 579 /* 580 * Send an ACK to resynchronize and drop any data. 581 * But keep on processing for RST or ACK. 582 */ 583 ctf_ack_war_checks(tp); 584 todrop = tlen; 585 KMOD_TCPSTAT_INC(tcps_rcvduppack); 586 KMOD_TCPSTAT_ADD(tcps_rcvdupbyte, todrop); 587 } else { 588 KMOD_TCPSTAT_INC(tcps_rcvpartduppack); 589 KMOD_TCPSTAT_ADD(tcps_rcvpartdupbyte, todrop); 590 } 591 /* 592 * DSACK - add SACK block for dropped range 593 */ 594 if ((todrop > 0) && (tp->t_flags & TF_SACK_PERMIT)) { 595 /* 596 * ACK now, as the next in-sequence segment 597 * will clear the DSACK block again 598 */ 599 ctf_ack_war_checks(tp); 600 if (tp->t_flags & TF_ACKNOW) 601 tcp_update_sack_list(tp, th->th_seq, 602 th->th_seq + todrop); 603 } 604 *drop_hdrlen += todrop; /* drop from the top afterwards */ 605 th->th_seq += todrop; 606 tlen -= todrop; 607 if (th->th_urp > todrop) 608 th->th_urp -= todrop; 609 else { 610 thflags &= ~TH_URG; 611 th->th_urp = 0; 612 } 613 } 614 /* 615 * If segment ends after window, drop trailing data (and PUSH and 616 * FIN); if nothing left, just ACK. 617 */ 618 todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd); 619 if (todrop > 0) { 620 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin); 621 if (todrop >= tlen) { 622 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, tlen); 623 /* 624 * If window is closed can only take segments at 625 * window edge, and have to drop data and PUSH from 626 * incoming segments. Continue processing, but 627 * remember to ack. Otherwise, drop segment and 628 * ack. 629 */ 630 if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) { 631 ctf_ack_war_checks(tp); 632 KMOD_TCPSTAT_INC(tcps_rcvwinprobe); 633 } else { 634 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 635 return (1); 636 } 637 } else 638 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 639 m_adj(m, -todrop); 640 tlen -= todrop; 641 thflags &= ~(TH_PUSH | TH_FIN); 642 } 643 *thf = thflags; 644 *tlenp = tlen; 645 return (0); 646 } 647 648 /* 649 * The value in ret_val informs the caller 650 * if we dropped the tcb (and lock) or not. 651 * 1 = we dropped it, 0 = the TCB is still locked 652 * and valid. 653 */ 654 void 655 ctf_do_dropafterack(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, int32_t thflags, int32_t tlen, int32_t *ret_val) 656 { 657 /* 658 * Generate an ACK dropping incoming segment if it occupies sequence 659 * space, where the ACK reflects our state. 660 * 661 * We can now skip the test for the RST flag since all paths to this 662 * code happen after packets containing RST have been dropped. 663 * 664 * In the SYN-RECEIVED state, don't send an ACK unless the segment 665 * we received passes the SYN-RECEIVED ACK test. If it fails send a 666 * RST. This breaks the loop in the "LAND" DoS attack, and also 667 * prevents an ACK storm between two listening ports that have been 668 * sent forged SYN segments, each with the source address of the 669 * other. 670 */ 671 if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) && 672 (SEQ_GT(tp->snd_una, th->th_ack) || 673 SEQ_GT(th->th_ack, tp->snd_max))) { 674 *ret_val = 1; 675 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 676 return; 677 } else 678 *ret_val = 0; 679 ctf_ack_war_checks(tp); 680 if (m) 681 m_freem(m); 682 } 683 684 void 685 ctf_do_drop(struct mbuf *m, struct tcpcb *tp) 686 { 687 688 /* 689 * Drop space held by incoming segment and return. 690 */ 691 if (tp != NULL) 692 INP_WUNLOCK(tptoinpcb(tp)); 693 if (m) 694 m_freem(m); 695 } 696 697 int 698 ctf_process_rst(struct mbuf *m, struct tcphdr *th, struct socket *so, 699 struct tcpcb *tp) 700 { 701 /* 702 * RFC5961 Section 3.2 703 * 704 * - RST drops connection only if SEG.SEQ == RCV.NXT. - If RST is in 705 * window, we send challenge ACK. 706 * 707 * Note: to take into account delayed ACKs, we should test against 708 * last_ack_sent instead of rcv_nxt. Note 2: we handle special case 709 * of closed window, not covered by the RFC. 710 */ 711 int dropped = 0; 712 713 if ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) && 714 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) || 715 (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq)) { 716 KASSERT(tp->t_state != TCPS_SYN_SENT, 717 ("%s: TH_RST for TCPS_SYN_SENT th %p tp %p", 718 __func__, th, tp)); 719 720 if (V_tcp_insecure_rst || 721 (tp->last_ack_sent == th->th_seq) || 722 (tp->rcv_nxt == th->th_seq)) { 723 KMOD_TCPSTAT_INC(tcps_drops); 724 /* Drop the connection. */ 725 switch (tp->t_state) { 726 case TCPS_SYN_RECEIVED: 727 so->so_error = ECONNREFUSED; 728 goto close; 729 case TCPS_ESTABLISHED: 730 case TCPS_FIN_WAIT_1: 731 case TCPS_FIN_WAIT_2: 732 case TCPS_CLOSE_WAIT: 733 case TCPS_CLOSING: 734 case TCPS_LAST_ACK: 735 so->so_error = ECONNRESET; 736 close: 737 tcp_state_change(tp, TCPS_CLOSED); 738 /* FALLTHROUGH */ 739 default: 740 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_RST); 741 tp = tcp_close(tp); 742 } 743 dropped = 1; 744 ctf_do_drop(m, tp); 745 } else { 746 KMOD_TCPSTAT_INC(tcps_badrst); 747 tcp_send_challenge_ack(tp, th, m); 748 } 749 } else { 750 m_freem(m); 751 } 752 return (dropped); 753 } 754 755 /* 756 * The value in ret_val informs the caller 757 * if we dropped the tcb (and lock) or not. 758 * 1 = we dropped it, 0 = the TCB is still locked 759 * and valid. 760 */ 761 void 762 ctf_challenge_ack(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, uint8_t iptos, int32_t * ret_val) 763 { 764 765 NET_EPOCH_ASSERT(); 766 767 KMOD_TCPSTAT_INC(tcps_badsyn); 768 if (V_tcp_insecure_syn && 769 SEQ_GEQ(th->th_seq, tp->last_ack_sent) && 770 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) { 771 tp = tcp_drop(tp, ECONNRESET); 772 *ret_val = 1; 773 ctf_do_drop(m, tp); 774 } else { 775 tcp_ecn_input_syn_sent(tp, tcp_get_flags(th), iptos); 776 /* Send challenge ACK. */ 777 tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt, 778 tp->snd_nxt, TH_ACK); 779 tp->last_ack_sent = tp->rcv_nxt; 780 m = NULL; 781 *ret_val = 0; 782 ctf_do_drop(m, NULL); 783 } 784 } 785 786 /* 787 * ctf_ts_check returns 1 for you should not proceed, the state 788 * machine should return. It places in ret_val what should 789 * be returned 1/0 by the caller (hpts_do_segment). The 1 indicates 790 * that the TCB is unlocked and probably dropped. The 0 indicates the 791 * TCB is still valid and locked. 792 */ 793 int 794 ctf_ts_check(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, 795 int32_t tlen, int32_t thflags, int32_t * ret_val) 796 { 797 798 if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { 799 /* 800 * Invalidate ts_recent. If this segment updates ts_recent, 801 * the age will be reset later and ts_recent will get a 802 * valid value. If it does not, setting ts_recent to zero 803 * will at least satisfy the requirement that zero be placed 804 * in the timestamp echo reply when ts_recent isn't valid. 805 * The age isn't reset until we get a valid ts_recent 806 * because we don't want out-of-order segments to be dropped 807 * when ts_recent is old. 808 */ 809 tp->ts_recent = 0; 810 } else { 811 KMOD_TCPSTAT_INC(tcps_rcvduppack); 812 KMOD_TCPSTAT_ADD(tcps_rcvdupbyte, tlen); 813 KMOD_TCPSTAT_INC(tcps_pawsdrop); 814 *ret_val = 0; 815 if (tlen) { 816 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 817 } else { 818 ctf_do_drop(m, NULL); 819 } 820 return (1); 821 } 822 return (0); 823 } 824 825 int 826 ctf_ts_check_ac(struct tcpcb *tp, int32_t thflags) 827 { 828 829 if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { 830 /* 831 * Invalidate ts_recent. If this segment updates ts_recent, 832 * the age will be reset later and ts_recent will get a 833 * valid value. If it does not, setting ts_recent to zero 834 * will at least satisfy the requirement that zero be placed 835 * in the timestamp echo reply when ts_recent isn't valid. 836 * The age isn't reset until we get a valid ts_recent 837 * because we don't want out-of-order segments to be dropped 838 * when ts_recent is old. 839 */ 840 tp->ts_recent = 0; 841 } else { 842 KMOD_TCPSTAT_INC(tcps_rcvduppack); 843 KMOD_TCPSTAT_INC(tcps_pawsdrop); 844 return (1); 845 } 846 return (0); 847 } 848 849 850 851 void 852 ctf_calc_rwin(struct socket *so, struct tcpcb *tp) 853 { 854 int32_t win; 855 856 /* 857 * Calculate amount of space in receive window, and then do TCP 858 * input processing. Receive window is amount of space in rcv queue, 859 * but not less than advertised window. 860 */ 861 win = sbspace(&so->so_rcv); 862 if (win < 0) 863 win = 0; 864 tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt)); 865 } 866 867 void 868 ctf_do_dropwithreset_conn(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, 869 int32_t rstreason, int32_t tlen) 870 { 871 872 tcp_dropwithreset(m, th, tp, tlen, rstreason); 873 tp = tcp_drop(tp, ETIMEDOUT); 874 if (tp) 875 INP_WUNLOCK(tptoinpcb(tp)); 876 } 877 878 uint32_t 879 ctf_fixed_maxseg(struct tcpcb *tp) 880 { 881 return (tcp_fixed_maxseg(tp)); 882 } 883 884 void 885 ctf_log_sack_filter(struct tcpcb *tp, int num_sack_blks, struct sackblk *sack_blocks) 886 { 887 if (tcp_bblogging_on(tp)) { 888 union tcp_log_stackspecific log; 889 struct timeval tv; 890 891 memset(&log, 0, sizeof(log)); 892 log.u_bbr.timeStamp = tcp_get_usecs(&tv); 893 log.u_bbr.flex8 = num_sack_blks; 894 if (num_sack_blks > 0) { 895 log.u_bbr.flex1 = sack_blocks[0].start; 896 log.u_bbr.flex2 = sack_blocks[0].end; 897 } 898 if (num_sack_blks > 1) { 899 log.u_bbr.flex3 = sack_blocks[1].start; 900 log.u_bbr.flex4 = sack_blocks[1].end; 901 } 902 if (num_sack_blks > 2) { 903 log.u_bbr.flex5 = sack_blocks[2].start; 904 log.u_bbr.flex6 = sack_blocks[2].end; 905 } 906 if (num_sack_blks > 3) { 907 log.u_bbr.applimited = sack_blocks[3].start; 908 log.u_bbr.pkts_out = sack_blocks[3].end; 909 } 910 TCP_LOG_EVENTP(tp, NULL, 911 &tptosocket(tp)->so_rcv, 912 &tptosocket(tp)->so_snd, 913 TCP_SACK_FILTER_RES, 0, 914 0, &log, false, &tv); 915 } 916 } 917 918 uint32_t 919 ctf_decay_count(uint32_t count, uint32_t decay) 920 { 921 /* 922 * Given a count, decay it by a set percentage. The 923 * percentage is in thousands i.e. 100% = 1000, 924 * 19.3% = 193. 925 */ 926 uint64_t perc_count, decay_per; 927 uint32_t decayed_count; 928 if (decay > 1000) { 929 /* We don't raise it */ 930 return (count); 931 } 932 perc_count = count; 933 decay_per = decay; 934 perc_count *= decay_per; 935 perc_count /= 1000; 936 /* 937 * So now perc_count holds the 938 * count decay value. 939 */ 940 decayed_count = count - (uint32_t)perc_count; 941 return(decayed_count); 942 } 943 944 int32_t 945 ctf_progress_timeout_check(struct tcpcb *tp, bool log) 946 { 947 if (tp->t_maxunacktime && tp->t_acktime && TSTMP_GT(ticks, tp->t_acktime)) { 948 if ((ticks - tp->t_acktime) >= tp->t_maxunacktime) { 949 /* 950 * There is an assumption that the caller 951 * will drop the connection so we will 952 * increment the counters here. 953 */ 954 if (log) 955 tcp_log_end_status(tp, TCP_EI_STATUS_PROGRESS); 956 #ifdef NETFLIX_STATS 957 KMOD_TCPSTAT_INC(tcps_progdrops); 958 #endif 959 return (1); 960 } 961 } 962 return (0); 963 } 964