sys/netinet/tcp_stacks/rack_bbr_common.c

   1 /*-
   2  * Copyright (c) 2016-9
   3  *      Netflix Inc.
   4  *      All rights reserved.
   5  *
   6  * Redistribution and use in source and binary forms, with or without
   7  * modification, are permitted provided that the following conditions
   8  * are met:
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice, this list of conditions and the following disclaimer.
  11  * 2. Redistributions in binary form must reproduce the above copyright
  12  *    notice, this list of conditions and the following disclaimer in the
  13  *    documentation and/or other materials provided with the distribution.
  14  *
  15  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  25  * SUCH DAMAGE.
  26  *
  27  */
  28 /*
  29  * Author: Randall Stewart <rrs@netflix.com>
  30  * This work is based on the ACM Queue paper
  31  * BBR - Congestion Based Congestion Control
  32  * and also numerous discussions with Neal, Yuchung and Van.
  33  */
  34
  35 #include <sys/cdefs.h>
  36 __FBSDID("$FreeBSD$");
  37
  38 #include "opt_inet.h"
  39 #include "opt_inet6.h"
  40 #include "opt_ipsec.h"
  41 #include "opt_tcpdebug.h"
  42 #include "opt_ratelimit.h"
  43 #include "opt_kern_tls.h"
  44 #include <sys/param.h>
  45 #include <sys/arb.h>
  46 #include <sys/module.h>
  47 #include <sys/kernel.h>
  48 #ifdef TCP_HHOOK
  49 #include <sys/hhook.h>
  50 #endif
  51 #include <sys/malloc.h>
  52 #include <sys/mbuf.h>
  53 #include <sys/proc.h>
  54 #include <sys/qmath.h>
  55 #include <sys/socket.h>
  56 #include <sys/socketvar.h>
  57 #ifdef KERN_TLS
  58 #include <sys/ktls.h>
  59 #endif
  60 #include <sys/sysctl.h>
  61 #include <sys/systm.h>
  62 #include <sys/tree.h>
  63 #ifdef NETFLIX_STATS
  64 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
  65 #endif
  66 #include <sys/refcount.h>
  67 #include <sys/queue.h>
  68 #include <sys/smp.h>
  69 #include <sys/kthread.h>
  70 #include <sys/lock.h>
  71 #include <sys/mutex.h>
  72 #include <sys/tim_filter.h>
  73 #include <sys/time.h>
  74 #include <vm/uma.h>
  75 #include <sys/kern_prefetch.h>
  76
  77 #include <net/route.h>
  78 #include <net/vnet.h>
  79 #include <net/ethernet.h>
  80 #include <net/bpf.h>
  81
  82 #define TCPSTATES               /* for logging */
  83
  84 #include <netinet/in.h>
  85 #include <netinet/in_kdtrace.h>
  86 #include <netinet/in_pcb.h>
  87 #include <netinet/ip.h>
  88 #include <netinet/ip_icmp.h>    /* required for icmp_var.h */
  89 #include <netinet/icmp_var.h>   /* for ICMP_BANDLIM */
  90 #include <netinet/ip_var.h>
  91 #include <netinet/ip6.h>
  92 #include <netinet6/in6_pcb.h>
  93 #include <netinet6/ip6_var.h>
  94 #include <netinet/tcp.h>
  95 #include <netinet/tcp_fsm.h>
  96 #include <netinet/tcp_seq.h>
  97 #include <netinet/tcp_timer.h>
  98 #include <netinet/tcp_var.h>
  99 #include <netinet/tcpip.h>
 100 #include <netinet/tcp_hpts.h>
 101 #include <netinet/cc/cc.h>
 102 #include <netinet/tcp_log_buf.h>
 103 #ifdef TCPDEBUG
 104 #include <netinet/tcp_debug.h>
 105 #endif                          /* TCPDEBUG */
 106 #ifdef TCP_OFFLOAD
 107 #include <netinet/tcp_offload.h>
 108 #endif
 109 #ifdef INET6
 110 #include <netinet6/tcp6_var.h>
 111 #endif
 112 #include <netinet/tcp_fastopen.h>
 113
 114 #include <netipsec/ipsec_support.h>
 115 #include <net/if.h>
 116 #include <net/if_var.h>
 117
 118 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 119 #include <netipsec/ipsec.h>
 120 #include <netipsec/ipsec6.h>
 121 #endif                          /* IPSEC */
 122
 123 #include <netinet/udp.h>
 124 #include <netinet/udp_var.h>
 125 #include <machine/in_cksum.h>
 126
 127 #ifdef MAC
 128 #include <security/mac/mac_framework.h>
 129 #endif
 130 #include "rack_bbr_common.h"
 131
 132 /*
 133  * Common TCP Functions - These are shared by borth
 134  * rack and BBR.
 135  */
 136 #ifdef KERN_TLS
 137 uint32_t
 138 ctf_get_opt_tls_size(struct socket *so, uint32_t rwnd)
 139 {
 140         struct ktls_session *tls;
 141         uint32_t len;
 142
 143 again:
 144         tls = so->so_snd.sb_tls_info;
 145         len = tls->params.max_frame_len;         /* max tls payload */
 146         len += tls->params.tls_hlen;      /* tls header len  */
 147         len += tls->params.tls_tlen;      /* tls trailer len */
 148         if ((len * 4) > rwnd) {
 149                 /*
 150                  * Stroke this will suck counter and what
 151                  * else should we do Drew? From the
 152                  * TCP perspective I am not sure
 153                  * what should be done...
 154                  */
 155                 if (tls->params.max_frame_len > 4096) {
 156                         tls->params.max_frame_len -= 4096;
 157                         if (tls->params.max_frame_len < 4096)
 158                                 tls->params.max_frame_len = 4096;
 159                         goto again;
 160                 }
 161         }
 162         return (len);
 163 }
 164 #endif
 165
 166
 167 /*
 168  * The function ctf_process_inbound_raw() is used by
 169  * transport developers to do the steps needed to
 170  * support MBUF Queuing i.e. the flags in
 171  * inp->inp_flags2:
 172  *
 173  * - INP_SUPPORTS_MBUFQ
 174  * - INP_MBUF_QUEUE_READY
 175  * - INP_DONT_SACK_QUEUE
 176  *
 177  * These flags help control how LRO will deliver
 178  * packets to the transport. You first set in inp_flags2
 179  * the INP_SUPPORTS_MBUFQ to tell the LRO code that you
 180  * will gladly take a queue of packets instead of a compressed
 181  * single packet. You also set in your t_fb pointer the
 182  * tfb_do_queued_segments to point to ctf_process_inbound_raw.
 183  *
 184  * This then gets you lists of inbound ACK's/Data instead
 185  * of a condensed compressed ACK/DATA packet. Why would you
 186  * want that? This will get you access to all the arrival
 187  * times of at least LRO and possibly at the Hardware (if
 188  * the interface card supports that) of the actual ACK/DATA.
 189  * In some transport designs this is important since knowing
 190  * the actual time we got the packet is useful information.
 191  *
 192  * Now there are some interesting Caveats that the transport
 193  * designer needs to take into account when using this feature.
 194  *
 195  * 1) It is used with HPTS and pacing, when the pacing timer
 196  *    for output calls it will first call the input.
 197  * 2) When you set INP_MBUF_QUEUE_READY this tells LRO
 198  *    queue normal packets, I am busy pacing out data and
 199  *    will process the queued packets before my tfb_tcp_output
 200  *    call from pacing. If a non-normal packet arrives, (e.g. sack)
 201  *    you will be awoken immediately.
 202  * 3) Finally you can add the INP_DONT_SACK_QUEUE to not even
 203  *    be awoken if a SACK has arrived. You would do this when
 204  *    you were not only running a pacing for output timer
 205  *    but a Rack timer as well i.e. you know you are in recovery
 206  *    and are in the process (via the timers) of dealing with
 207  *    the loss.
 208  *
 209  * Now a critical thing you must be aware of here is that the
 210  * use of the flags has a far greater scope then just your
 211  * typical LRO. Why? Well thats because in the normal compressed
 212  * LRO case at the end of a driver interupt all packets are going
 213  * to get presented to the transport no matter if there is one
 214  * or 100. With the MBUF_QUEUE model, this is not true. You will
 215  * only be awoken to process the queue of packets when:
 216  *     a) The flags discussed above allow it.
 217  *          <or>
 218  *     b) You exceed a ack or data limit (by default the
 219  *        ack limit is infinity (64k acks) and the data
 220  *        limit is 64k of new TCP data)
 221  *         <or>
 222  *     c) The push bit has been set by the peer
 223  */
 224
 225 int
 226 ctf_process_inbound_raw(struct tcpcb *tp, struct socket *so, struct mbuf *m, int has_pkt)
 227 {
 228         /*
 229          * We are passed a raw change of mbuf packets
 230          * that arrived in LRO. They are linked via
 231          * the m_nextpkt link in the pkt-headers.
 232          *
 233          * We process each one by:
 234          * a) saving off the next
 235          * b) stripping off the ether-header
 236          * c) formulating the arguments for
 237          *    the tfb_tcp_hpts_do_segment
 238          * d) calling each mbuf to tfb_tcp_hpts_do_segment
 239          *    after adjusting the time to match the arrival time.
 240          * Note that the LRO code assures no IP options are present.
 241          *
 242          * The symantics for calling tfb_tcp_hpts_do_segment are the
 243          * following:
 244          * 1) It returns 0 if all went well and you (the caller) need
 245          *    to release the lock.
 246          * 2) If nxt_pkt is set, then the function will surpress calls
 247          *    to tfb_tcp_output() since you are promising to call again
 248          *    with another packet.
 249          * 3) If it returns 1, then you must free all the packets being
 250          *    shipped in, the tcb has been destroyed (or about to be destroyed).
 251          */
 252         struct mbuf *m_save;
 253         struct ether_header *eh;
 254         struct tcphdr *th;
 255 #ifdef INET6
 256         struct ip6_hdr *ip6 = NULL;     /* Keep compiler happy. */
 257 #endif
 258 #ifdef INET
 259         struct ip *ip = NULL;           /* Keep compiler happy. */
 260 #endif
 261         struct ifnet *ifp;
 262         struct timeval tv;
 263         int32_t retval, nxt_pkt, tlen, off;
 264         uint16_t etype;
 265         uint16_t drop_hdrlen;
 266         uint8_t iptos, no_vn=0, bpf_req=0;
 267
 268         NET_EPOCH_ASSERT();
 269
 270         if (m && m->m_pkthdr.rcvif)
 271                 ifp = m->m_pkthdr.rcvif;
 272         else
 273                 ifp = NULL;
 274         if (ifp) {
 275                 bpf_req = bpf_peers_present(ifp->if_bpf);
 276         } else  {
 277                 /*
 278                  * We probably should not work around
 279                  * but kassert, since lro alwasy sets rcvif.
 280                  */
 281                 no_vn = 1;
 282                 goto skip_vnet;
 283         }
 284         CURVNET_SET(ifp->if_vnet);
 285 skip_vnet:
 286         while (m) {
 287                 m_save = m->m_nextpkt;
 288                 m->m_nextpkt = NULL;
 289                 /* Now lets get the ether header */
 290                 eh = mtod(m, struct ether_header *);
 291                 etype = ntohs(eh->ether_type);
 292                 /* Let the BPF see the packet */
 293                 if (bpf_req && ifp)
 294                         ETHER_BPF_MTAP(ifp, m);
 295                 m_adj(m,  sizeof(*eh));
 296                 /* Trim off the ethernet header */
 297                 switch (etype) {
 298 #ifdef INET6
 299                 case ETHERTYPE_IPV6:
 300                 {
 301                         if (m->m_len < (sizeof(*ip6) + sizeof(*th))) {
 302                                 m = m_pullup(m, sizeof(*ip6) + sizeof(*th));
 303                                 if (m == NULL) {
 304                                         TCPSTAT_INC(tcps_rcvshort);
 305                                         m_freem(m);
 306                                         goto skipped_pkt;
 307                                 }
 308                         }
 309                         ip6 = (struct ip6_hdr *)(eh + 1);
 310                         th = (struct tcphdr *)(ip6 + 1);
 311                         tlen = ntohs(ip6->ip6_plen);
 312                         drop_hdrlen = sizeof(*ip6);
 313                         if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID_IPV6) {
 314                                 if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR)
 315                                         th->th_sum = m->m_pkthdr.csum_data;
 316                                 else
 317                                         th->th_sum = in6_cksum_pseudo(ip6, tlen,
 318                                                                       IPPROTO_TCP, m->m_pkthdr.csum_data);
 319                                 th->th_sum ^= 0xffff;
 320                         } else
 321                                 th->th_sum = in6_cksum(m, IPPROTO_TCP, drop_hdrlen, tlen);
 322                         if (th->th_sum) {
 323                                 TCPSTAT_INC(tcps_rcvbadsum);
 324                                 m_freem(m);
 325                                 goto skipped_pkt;
 326                         }
 327                         /*
 328                          * Be proactive about unspecified IPv6 address in source.
 329                          * As we use all-zero to indicate unbounded/unconnected pcb,
 330                          * unspecified IPv6 address can be used to confuse us.
 331                          *
 332                          * Note that packets with unspecified IPv6 destination is
 333                          * already dropped in ip6_input.
 334                          */
 335                         if (IN6_IS_ADDR_UNSPECIFIED(&ip6->ip6_src)) {
 336                                 /* XXX stat */
 337                                 m_freem(m);
 338                                 goto skipped_pkt;
 339                         }
 340                         iptos = (ntohl(ip6->ip6_flow) >> 20) & 0xff;
 341                         break;
 342                 }
 343 #endif
 344 #ifdef INET
 345                 case ETHERTYPE_IP:
 346                 {
 347                         if (m->m_len < sizeof (struct tcpiphdr)) {
 348                                 if ((m = m_pullup(m, sizeof (struct tcpiphdr)))
 349                                     == NULL) {
 350                                         TCPSTAT_INC(tcps_rcvshort);
 351                                         m_freem(m);
 352                                         goto skipped_pkt;
 353                                 }
 354                         }
 355                         ip = (struct ip *)(eh + 1);
 356                         th = (struct tcphdr *)(ip + 1);
 357                         drop_hdrlen = sizeof(*ip);
 358                         iptos = ip->ip_tos;
 359                         tlen = ntohs(ip->ip_len) - sizeof(struct ip);
 360                         if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) {
 361                                 if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR)
 362                                         th->th_sum = m->m_pkthdr.csum_data;
 363                                 else
 364                                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
 365                                                                ip->ip_dst.s_addr,
 366                                                                htonl(m->m_pkthdr.csum_data + tlen +
 367                                                                      IPPROTO_TCP));
 368                                 th->th_sum ^= 0xffff;
 369                         } else {
 370                                 int len;
 371                                 struct ipovly *ipov = (struct ipovly *)ip;
 372                                 /*
 373                                  * Checksum extended TCP header and data.
 374                                  */
 375                                 len = drop_hdrlen + tlen;
 376                                 bzero(ipov->ih_x1, sizeof(ipov->ih_x1));
 377                                 ipov->ih_len = htons(tlen);
 378                                 th->th_sum = in_cksum(m, len);
 379                                 /* Reset length for SDT probes. */
 380                                 ip->ip_len = htons(len);
 381                                 /* Reset TOS bits */
 382                                 ip->ip_tos = iptos;
 383                                 /* Re-initialization for later version check */
 384                                 ip->ip_v = IPVERSION;
 385                                 ip->ip_hl = sizeof(*ip) >> 2;
 386                         }
 387                         if (th->th_sum) {
 388                                 TCPSTAT_INC(tcps_rcvbadsum);
 389                                 m_freem(m);
 390                                 goto skipped_pkt;
 391                         }
 392                         break;
 393                 }
 394 #endif
 395                 }
 396                 /*
 397                  * Convert TCP protocol specific fields to host format.
 398                  */
 399                 tcp_fields_to_host(th);
 400
 401                 off = th->th_off << 2;
 402                 if (off < sizeof (struct tcphdr) || off > tlen) {
 403                         TCPSTAT_INC(tcps_rcvbadoff);
 404                                 m_freem(m);
 405                                 goto skipped_pkt;
 406                 }
 407                 tlen -= off;
 408                 drop_hdrlen += off;
 409                 /*
 410                  * Now lets setup the timeval to be when we should
 411                  * have been called (if we can).
 412                  */
 413                 m->m_pkthdr.lro_nsegs = 1;
 414                 if (m->m_flags & M_TSTMP_LRO) {
 415                         tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
 416                         tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
 417                 } else {
 418                         /* Should not be should we kassert instead? */
 419                         tcp_get_usecs(&tv);
 420                 }
 421                 /* Now what about next packet? */
 422                 if (m_save || has_pkt)
 423                         nxt_pkt = 1;
 424                 else
 425                         nxt_pkt = 0;
 426                 retval = (*tp->t_fb->tfb_do_segment_nounlock)(m, th, so, tp, drop_hdrlen, tlen,
 427                                                               iptos, nxt_pkt, &tv);
 428                 if (retval) {
 429                         /* We lost the lock and tcb probably */
 430                         m = m_save;
 431                         while(m) {
 432                                 m_save = m->m_nextpkt;
 433                                 m->m_nextpkt = NULL;
 434                                 m_freem(m);
 435                                 m = m_save;
 436                         }
 437                         if (no_vn == 0)
 438                                 CURVNET_RESTORE();
 439                         return(retval);
 440                 }
 441 skipped_pkt:
 442                 m = m_save;
 443         }
 444         if (no_vn == 0)
 445                 CURVNET_RESTORE();
 446         return(retval);
 447 }
 448
 449 int
 450 ctf_do_queued_segments(struct socket *so, struct tcpcb *tp, int have_pkt)
 451 {
 452         struct mbuf *m;
 453
 454         /* First lets see if we have old packets */
 455         if (tp->t_in_pkt) {
 456                 m = tp->t_in_pkt;
 457                 tp->t_in_pkt = NULL;
 458                 tp->t_tail_pkt = NULL;
 459                 if (ctf_process_inbound_raw(tp, so, m, have_pkt)) {
 460                         /* We lost the tcpcb (maybe a RST came in)? */
 461                         return(1);
 462                 }
 463         }
 464         return (0);
 465 }
 466
 467 uint32_t
 468 ctf_outstanding(struct tcpcb *tp)
 469 {
 470         return(tp->snd_max - tp->snd_una);
 471 }
 472
 473 uint32_t
 474 ctf_flight_size(struct tcpcb *tp, uint32_t rc_sacked)
 475 {
 476         if (rc_sacked <= ctf_outstanding(tp))
 477                 return(ctf_outstanding(tp) - rc_sacked);
 478         else {
 479                 /* TSNH */
 480 #ifdef INVARIANTS
 481                 panic("tp:%p rc_sacked:%d > out:%d",
 482                       tp, rc_sacked, ctf_outstanding(tp));
 483 #endif
 484                 return (0);
 485         }
 486 }
 487
 488 void
 489 ctf_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th,
 490     int32_t rstreason, int32_t tlen)
 491 {
 492         if (tp != NULL) {
 493                 tcp_dropwithreset(m, th, tp, tlen, rstreason);
 494                 INP_WUNLOCK(tp->t_inpcb);
 495         } else
 496                 tcp_dropwithreset(m, th, NULL, tlen, rstreason);
 497 }
 498
 499 /*
 500  * ctf_drop_checks returns 1 for you should not proceed. It places
 501  * in ret_val what should be returned 1/0 by the caller. The 1 indicates
 502  * that the TCB is unlocked and probably dropped. The 0 indicates the
 503  * TCB is still valid and locked.
 504  */
 505 int
 506 ctf_drop_checks(struct tcpopt *to, struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t * tlenp,  int32_t * thf, int32_t * drop_hdrlen, int32_t * ret_val)
 507 {
 508         int32_t todrop;
 509         int32_t thflags;
 510         int32_t tlen;
 511
 512         thflags = *thf;
 513         tlen = *tlenp;
 514         todrop = tp->rcv_nxt - th->th_seq;
 515         if (todrop > 0) {
 516                 if (thflags & TH_SYN) {
 517                         thflags &= ~TH_SYN;
 518                         th->th_seq++;
 519                         if (th->th_urp > 1)
 520                                 th->th_urp--;
 521                         else
 522                                 thflags &= ~TH_URG;
 523                         todrop--;
 524                 }
 525                 /*
 526                  * Following if statement from Stevens, vol. 2, p. 960.
 527                  */
 528                 if (todrop > tlen
 529                     || (todrop == tlen && (thflags & TH_FIN) == 0)) {
 530                         /*
 531                          * Any valid FIN must be to the left of the window.
 532                          * At this point the FIN must be a duplicate or out
 533                          * of sequence; drop it.
 534                          */
 535                         thflags &= ~TH_FIN;
 536                         /*
 537                          * Send an ACK to resynchronize and drop any data.
 538                          * But keep on processing for RST or ACK.
 539                          */
 540                         tp->t_flags |= TF_ACKNOW;
 541                         todrop = tlen;
 542                         TCPSTAT_INC(tcps_rcvduppack);
 543                         TCPSTAT_ADD(tcps_rcvdupbyte, todrop);
 544                 } else {
 545                         TCPSTAT_INC(tcps_rcvpartduppack);
 546                         TCPSTAT_ADD(tcps_rcvpartdupbyte, todrop);
 547                 }
 548                 /*
 549                  * DSACK - add SACK block for dropped range
 550                  */
 551                 if ((todrop > 0) && (tp->t_flags & TF_SACK_PERMIT)) {
 552                         tcp_update_sack_list(tp, th->th_seq,
 553                             th->th_seq + todrop);
 554                         /*
 555                          * ACK now, as the next in-sequence segment
 556                          * will clear the DSACK block again
 557                          */
 558                         tp->t_flags |= TF_ACKNOW;
 559                 }
 560                 *drop_hdrlen += todrop; /* drop from the top afterwards */
 561                 th->th_seq += todrop;
 562                 tlen -= todrop;
 563                 if (th->th_urp > todrop)
 564                         th->th_urp -= todrop;
 565                 else {
 566                         thflags &= ~TH_URG;
 567                         th->th_urp = 0;
 568                 }
 569         }
 570         /*
 571          * If segment ends after window, drop trailing data (and PUSH and
 572          * FIN); if nothing left, just ACK.
 573          */
 574         todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd);
 575         if (todrop > 0) {
 576                 TCPSTAT_INC(tcps_rcvpackafterwin);
 577                 if (todrop >= tlen) {
 578                         TCPSTAT_ADD(tcps_rcvbyteafterwin, tlen);
 579                         /*
 580                          * If window is closed can only take segments at
 581                          * window edge, and have to drop data and PUSH from
 582                          * incoming segments.  Continue processing, but
 583                          * remember to ack.  Otherwise, drop segment and
 584                          * ack.
 585                          */
 586                         if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) {
 587                                 tp->t_flags |= TF_ACKNOW;
 588                                 TCPSTAT_INC(tcps_rcvwinprobe);
 589                         } else {
 590                                 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
 591                                 return (1);
 592                         }
 593                 } else
 594                         TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
 595                 m_adj(m, -todrop);
 596                 tlen -= todrop;
 597                 thflags &= ~(TH_PUSH | TH_FIN);
 598         }
 599         *thf = thflags;
 600         *tlenp = tlen;
 601         return (0);
 602 }
 603
 604 /*
 605  * The value in ret_val informs the caller
 606  * if we dropped the tcb (and lock) or not.
 607  * 1 = we dropped it, 0 = the TCB is still locked
 608  * and valid.
 609  */
 610 void
 611 ctf_do_dropafterack(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, int32_t thflags, int32_t tlen, int32_t * ret_val)
 612 {
 613         /*
 614          * Generate an ACK dropping incoming segment if it occupies sequence
 615          * space, where the ACK reflects our state.
 616          *
 617          * We can now skip the test for the RST flag since all paths to this
 618          * code happen after packets containing RST have been dropped.
 619          *
 620          * In the SYN-RECEIVED state, don't send an ACK unless the segment
 621          * we received passes the SYN-RECEIVED ACK test. If it fails send a
 622          * RST.  This breaks the loop in the "LAND" DoS attack, and also
 623          * prevents an ACK storm between two listening ports that have been
 624          * sent forged SYN segments, each with the source address of the
 625          * other.
 626          */
 627         if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) &&
 628             (SEQ_GT(tp->snd_una, th->th_ack) ||
 629             SEQ_GT(th->th_ack, tp->snd_max))) {
 630                 *ret_val = 1;
 631                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
 632                 return;
 633         } else
 634                 *ret_val = 0;
 635         tp->t_flags |= TF_ACKNOW;
 636         if (m)
 637                 m_freem(m);
 638 }
 639
 640 void
 641 ctf_do_drop(struct mbuf *m, struct tcpcb *tp)
 642 {
 643
 644         /*
 645          * Drop space held by incoming segment and return.
 646          */
 647         if (tp != NULL)
 648                 INP_WUNLOCK(tp->t_inpcb);
 649         if (m)
 650                 m_freem(m);
 651 }
 652
 653 int
 654 ctf_process_rst(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp)
 655 {
 656         /*
 657          * RFC5961 Section 3.2
 658          *
 659          * - RST drops connection only if SEG.SEQ == RCV.NXT. - If RST is in
 660          * window, we send challenge ACK.
 661          *
 662          * Note: to take into account delayed ACKs, we should test against
 663          * last_ack_sent instead of rcv_nxt. Note 2: we handle special case
 664          * of closed window, not covered by the RFC.
 665          */
 666         int dropped = 0;
 667
 668         if ((SEQ_GEQ(th->th_seq, (tp->last_ack_sent - 1)) &&
 669             SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) ||
 670             (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq)) {
 671
 672                 KASSERT(tp->t_state != TCPS_SYN_SENT,
 673                     ("%s: TH_RST for TCPS_SYN_SENT th %p tp %p",
 674                     __func__, th, tp));
 675
 676                 if (V_tcp_insecure_rst ||
 677                     (tp->last_ack_sent == th->th_seq) ||
 678                     (tp->rcv_nxt == th->th_seq) ||
 679                     ((tp->last_ack_sent - 1) == th->th_seq)) {
 680                         TCPSTAT_INC(tcps_drops);
 681                         /* Drop the connection. */
 682                         switch (tp->t_state) {
 683                         case TCPS_SYN_RECEIVED:
 684                                 so->so_error = ECONNREFUSED;
 685                                 goto close;
 686                         case TCPS_ESTABLISHED:
 687                         case TCPS_FIN_WAIT_1:
 688                         case TCPS_FIN_WAIT_2:
 689                         case TCPS_CLOSE_WAIT:
 690                         case TCPS_CLOSING:
 691                         case TCPS_LAST_ACK:
 692                                 so->so_error = ECONNRESET;
 693                 close:
 694                                 tcp_state_change(tp, TCPS_CLOSED);
 695                                 /* FALLTHROUGH */
 696                         default:
 697                                 tp = tcp_close(tp);
 698                         }
 699                         dropped = 1;
 700                         ctf_do_drop(m, tp);
 701                 } else {
 702                         TCPSTAT_INC(tcps_badrst);
 703                         /* Send challenge ACK. */
 704                         tcp_respond(tp, mtod(m, void *), th, m,
 705                             tp->rcv_nxt, tp->snd_nxt, TH_ACK);
 706                         tp->last_ack_sent = tp->rcv_nxt;
 707                 }
 708         } else {
 709                 m_freem(m);
 710         }
 711         return (dropped);
 712 }
 713
 714 /*
 715  * The value in ret_val informs the caller
 716  * if we dropped the tcb (and lock) or not.
 717  * 1 = we dropped it, 0 = the TCB is still locked
 718  * and valid.
 719  */
 720 void
 721 ctf_challenge_ack(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t * ret_val)
 722 {
 723
 724         NET_EPOCH_ASSERT();
 725
 726         TCPSTAT_INC(tcps_badsyn);
 727         if (V_tcp_insecure_syn &&
 728             SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
 729             SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) {
 730                 tp = tcp_drop(tp, ECONNRESET);
 731                 *ret_val = 1;
 732                 ctf_do_drop(m, tp);
 733         } else {
 734                 /* Send challenge ACK. */
 735                 tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt,
 736                     tp->snd_nxt, TH_ACK);
 737                 tp->last_ack_sent = tp->rcv_nxt;
 738                 m = NULL;
 739                 *ret_val = 0;
 740                 ctf_do_drop(m, NULL);
 741         }
 742 }
 743
 744 /*
 745  * bbr_ts_check returns 1 for you should not proceed, the state
 746  * machine should return. It places in ret_val what should
 747  * be returned 1/0 by the caller (hpts_do_segment). The 1 indicates
 748  * that the TCB is unlocked and probably dropped. The 0 indicates the
 749  * TCB is still valid and locked.
 750  */
 751 int
 752 ctf_ts_check(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp,
 753     int32_t tlen, int32_t thflags, int32_t * ret_val)
 754 {
 755
 756         if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) {
 757                 /*
 758                  * Invalidate ts_recent.  If this segment updates ts_recent,
 759                  * the age will be reset later and ts_recent will get a
 760                  * valid value.  If it does not, setting ts_recent to zero
 761                  * will at least satisfy the requirement that zero be placed
 762                  * in the timestamp echo reply when ts_recent isn't valid.
 763                  * The age isn't reset until we get a valid ts_recent
 764                  * because we don't want out-of-order segments to be dropped
 765                  * when ts_recent is old.
 766                  */
 767                 tp->ts_recent = 0;
 768         } else {
 769                 TCPSTAT_INC(tcps_rcvduppack);
 770                 TCPSTAT_ADD(tcps_rcvdupbyte, tlen);
 771                 TCPSTAT_INC(tcps_pawsdrop);
 772                 *ret_val = 0;
 773                 if (tlen) {
 774                         ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
 775                 } else {
 776                         ctf_do_drop(m, NULL);
 777                 }
 778                 return (1);
 779         }
 780         return (0);
 781 }
 782
 783 void
 784 ctf_calc_rwin(struct socket *so, struct tcpcb *tp)
 785 {
 786         int32_t win;
 787
 788         /*
 789          * Calculate amount of space in receive window, and then do TCP
 790          * input processing. Receive window is amount of space in rcv queue,
 791          * but not less than advertised window.
 792          */
 793         win = sbspace(&so->so_rcv);
 794         if (win < 0)
 795                 win = 0;
 796         tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt));
 797 }
 798
 799 void
 800 ctf_do_dropwithreset_conn(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th,
 801     int32_t rstreason, int32_t tlen)
 802 {
 803
 804         if (tp->t_inpcb) {
 805                 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
 806         }
 807         tcp_dropwithreset(m, th, tp, tlen, rstreason);
 808         INP_WUNLOCK(tp->t_inpcb);
 809 }
 810
 811 uint32_t
 812 ctf_fixed_maxseg(struct tcpcb *tp)
 813 {
 814         int optlen;
 815
 816         if (tp->t_flags & TF_NOOPT)
 817                 return (tp->t_maxseg);
 818
 819         /*
 820          * Here we have a simplified code from tcp_addoptions(),
 821          * without a proper loop, and having most of paddings hardcoded.
 822          * We only consider fixed options that we would send every
 823          * time I.e. SACK is not considered.
 824          *
 825          */
 826 #define PAD(len)        ((((len) / 4) + !!((len) % 4)) * 4)
 827         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
 828                 if (tp->t_flags & TF_RCVD_TSTMP)
 829                         optlen = TCPOLEN_TSTAMP_APPA;
 830                 else
 831                         optlen = 0;
 832 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
 833                 if (tp->t_flags & TF_SIGNATURE)
 834                         optlen += PAD(TCPOLEN_SIGNATURE);
 835 #endif
 836         } else {
 837                 if (tp->t_flags & TF_REQ_TSTMP)
 838                         optlen = TCPOLEN_TSTAMP_APPA;
 839                 else
 840                         optlen = PAD(TCPOLEN_MAXSEG);
 841                 if (tp->t_flags & TF_REQ_SCALE)
 842                         optlen += PAD(TCPOLEN_WINDOW);
 843 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
 844                 if (tp->t_flags & TF_SIGNATURE)
 845                         optlen += PAD(TCPOLEN_SIGNATURE);
 846 #endif
 847                 if (tp->t_flags & TF_SACK_PERMIT)
 848                         optlen += PAD(TCPOLEN_SACK_PERMITTED);
 849         }
 850 #undef PAD
 851         optlen = min(optlen, TCP_MAXOLEN);
 852         return (tp->t_maxseg - optlen);
 853 }
 854
 855 void
 856 ctf_log_sack_filter(struct tcpcb *tp, int num_sack_blks, struct sackblk *sack_blocks)
 857 {
 858         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
 859                 union tcp_log_stackspecific log;
 860                 struct timeval tv;
 861
 862                 memset(&log, 0, sizeof(log));
 863                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
 864                 log.u_bbr.flex8 = num_sack_blks;
 865                 if (num_sack_blks > 0) {
 866                         log.u_bbr.flex1 = sack_blocks[0].start;
 867                         log.u_bbr.flex2 = sack_blocks[0].end;
 868                 }
 869                 if (num_sack_blks > 1) {
 870                         log.u_bbr.flex3 = sack_blocks[1].start;
 871                         log.u_bbr.flex4 = sack_blocks[1].end;
 872                 }
 873                 if (num_sack_blks > 2) {
 874                         log.u_bbr.flex5 = sack_blocks[2].start;
 875                         log.u_bbr.flex6 = sack_blocks[2].end;
 876                 }
 877                 if (num_sack_blks > 3) {
 878                         log.u_bbr.applimited = sack_blocks[3].start;
 879                         log.u_bbr.pkts_out = sack_blocks[3].end;
 880                 }
 881                 TCP_LOG_EVENTP(tp, NULL,
 882                     &tp->t_inpcb->inp_socket->so_rcv,
 883                     &tp->t_inpcb->inp_socket->so_snd,
 884                     TCP_SACK_FILTER_RES, 0,
 885                     0, &log, false, &tv);
 886         }
 887 }
 888
 889 uint32_t
 890 ctf_decay_count(uint32_t count, uint32_t decay)
 891 {
 892         /*
 893          * Given a count, decay it by a set percentage. The
 894          * percentage is in thousands i.e. 100% = 1000,
 895          * 19.3% = 193.
 896          */
 897         uint64_t perc_count, decay_per;
 898         uint32_t decayed_count;
 899         if (decay > 1000) {
 900                 /* We don't raise it */
 901                 return (count);
 902         }
 903         perc_count = count;
 904         decay_per = decay;
 905         perc_count *= decay_per;
 906         perc_count /= 1000;
 907         /*
 908          * So now perc_count holds the
 909          * count decay value.
 910          */
 911         decayed_count = count - (uint32_t)perc_count;
 912         return(decayed_count);
 913 }