sys/netinet/tcp_ecn.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-3-Clause
   3  *
   4  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995
   5  *      The Regents of the University of California.  All rights reserved.
   6  * Copyright (c) 2007-2008,2010
   7  *      Swinburne University of Technology, Melbourne, Australia.
   8  * Copyright (c) 2009-2010 Lawrence Stewart <lstewart@freebsd.org>
   9  * Copyright (c) 2010 The FreeBSD Foundation
  10  * Copyright (c) 2010-2011 Juniper Networks, Inc.
  11  * Copyright (c) 2019 Richard Scheffenegger <srichard@netapp.com>
  12  * All rights reserved.
  13  *
  14  * Portions of this software were developed at the Centre for Advanced Internet
  15  * Architectures, Swinburne University of Technology, by Lawrence Stewart,
  16  * James Healy and David Hayes, made possible in part by a grant from the Cisco
  17  * University Research Program Fund at Community Foundation Silicon Valley.
  18  *
  19  * Portions of this software were developed at the Centre for Advanced
  20  * Internet Architectures, Swinburne University of Technology, Melbourne,
  21  * Australia by David Hayes under sponsorship from the FreeBSD Foundation.
  22  *
  23  * Portions of this software were developed by Robert N. M. Watson under
  24  * contract to Juniper Networks, Inc.
  25  *
  26  * Redistribution and use in source and binary forms, with or without
  27  * modification, are permitted provided that the following conditions
  28  * are met:
  29  * 1. Redistributions of source code must retain the above copyright
  30  *    notice, this list of conditions and the following disclaimer.
  31  * 2. Redistributions in binary form must reproduce the above copyright
  32  *    notice, this list of conditions and the following disclaimer in the
  33  *    documentation and/or other materials provided with the distribution.
  34  * 3. Neither the name of the University nor the names of its contributors
  35  *    may be used to endorse or promote products derived from this software
  36  *    without specific prior written permission.
  37  *
  38  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  39  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  40  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  41  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  42  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  43  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  44  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  45  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  46  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  47  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  48  * SUCH DAMAGE.
  49  *
  50  *      @(#)tcp_ecn.c 8.12 (Berkeley) 5/24/95
  51  */
  52
  53 /*
  54  * Utility functions to deal with Explicit Congestion Notification in TCP
  55  * implementing the essential parts of the Accurate ECN extension
  56  * https://tools.ietf.org/html/draft-ietf-tcpm-accurate-ecn-09
  57  */
  58
  59 #include <sys/cdefs.h>
  60 __FBSDID("$FreeBSD$");
  61
  62 #include "opt_inet.h"
  63 #include "opt_inet6.h"
  64 #include "opt_tcpdebug.h"
  65
  66 #include <sys/param.h>
  67 #include <sys/systm.h>
  68 #include <sys/kernel.h>
  69 #include <sys/sysctl.h>
  70 #include <sys/malloc.h>
  71 #include <sys/mbuf.h>
  72 #include <sys/socket.h>
  73 #include <sys/socketvar.h>
  74
  75 #include <machine/cpu.h>
  76
  77 #include <vm/uma.h>
  78
  79 #include <net/if.h>
  80 #include <net/if_var.h>
  81 #include <net/route.h>
  82 #include <net/vnet.h>
  83
  84 #include <netinet/in.h>
  85 #include <netinet/in_systm.h>
  86 #include <netinet/ip.h>
  87 #include <netinet/in_var.h>
  88 #include <netinet/in_pcb.h>
  89 #include <netinet/ip_var.h>
  90 #include <netinet/ip6.h>
  91 #include <netinet/icmp6.h>
  92 #include <netinet6/nd6.h>
  93 #include <netinet6/ip6_var.h>
  94 #include <netinet6/in6_pcb.h>
  95 #include <netinet/tcp.h>
  96 #include <netinet/tcp_fsm.h>
  97 #include <netinet/tcp_seq.h>
  98 #include <netinet/tcp_var.h>
  99 #include <netinet/tcp_syncache.h>
 100 #include <netinet/tcp_timer.h>
 101 #include <netinet6/tcp6_var.h>
 102 #include <netinet/tcpip.h>
 103 #include <netinet/tcp_ecn.h>
 104
 105
 106 /*
 107  * Process incoming SYN,ACK packet
 108  */
 109 void
 110 tcp_ecn_input_syn_sent(struct tcpcb *tp, uint16_t thflags, int iptos)
 111 {
 112
 113         if (V_tcp_do_ecn == 0)
 114                 return;
 115         if ((V_tcp_do_ecn == 1) ||
 116             (V_tcp_do_ecn == 2)) {
 117                 /* RFC3168 ECN handling */
 118                 if ((thflags & (TH_CWR | TH_ECE)) == (0 | TH_ECE)) {
 119                         tp->t_flags2 |= TF2_ECN_PERMIT;
 120                         TCPSTAT_INC(tcps_ecn_shs);
 121                 }
 122         } else
 123         /* decoding Accurate ECN according to table in section 3.1.1 */
 124         if ((V_tcp_do_ecn == 3) ||
 125             (V_tcp_do_ecn == 4)) {
 126                 /*
 127                  * on the SYN,ACK, process the AccECN
 128                  * flags indicating the state the SYN
 129                  * was delivered.
 130                  * Reactions to Path ECN mangling can
 131                  * come here.
 132                  */
 133                 switch (thflags & (TH_AE | TH_CWR | TH_ECE)) {
 134                 /* RFC3168 SYN */
 135                 case (0|0|TH_ECE):
 136                         tp->t_flags2 |= TF2_ECN_PERMIT;
 137                         TCPSTAT_INC(tcps_ecn_shs);
 138                         break;
 139                 /* non-ECT SYN */
 140                 case (0|TH_CWR|0):
 141                         tp->t_flags2 |= TF2_ACE_PERMIT;
 142                         tp->t_scep = 5;
 143                         TCPSTAT_INC(tcps_ecn_shs);
 144                         TCPSTAT_INC(tcps_ace_nect);
 145                         break;
 146                 /* ECT0 SYN */
 147                 case (TH_AE|0|0):
 148                         tp->t_flags2 |= TF2_ACE_PERMIT;
 149                         tp->t_scep = 5;
 150                         TCPSTAT_INC(tcps_ecn_shs);
 151                         TCPSTAT_INC(tcps_ace_ect0);
 152                         break;
 153                 /* ECT1 SYN */
 154                 case (0|TH_CWR|TH_ECE):
 155                         tp->t_flags2 |= TF2_ACE_PERMIT;
 156                         tp->t_scep = 5;
 157                         TCPSTAT_INC(tcps_ecn_shs);
 158                         TCPSTAT_INC(tcps_ace_ect1);
 159                         break;
 160                 /* CE SYN */
 161                 case (TH_AE|TH_CWR|0):
 162                         tp->t_flags2 |= TF2_ACE_PERMIT;
 163                         tp->t_scep = 6;
 164                         /*
 165                          * reduce the IW to 2 MSS (to
 166                          * account for delayed acks) if
 167                          * the SYN,ACK was CE marked
 168                          */
 169                         tp->snd_cwnd = 2 * tcp_maxseg(tp);
 170                         TCPSTAT_INC(tcps_ecn_shs);
 171                         TCPSTAT_INC(tcps_ace_nect);
 172                         break;
 173                 default:
 174                         break;
 175                 }
 176                 /*
 177                  * Set the AccECN Codepoints on
 178                  * the outgoing <ACK> to the ECN
 179                  * state of the <SYN,ACK>
 180                  * according to table 3 in the
 181                  * AccECN draft
 182                  */
 183                 switch (iptos & IPTOS_ECN_MASK) {
 184                 case (IPTOS_ECN_NOTECT):
 185                         tp->t_rcep = 0b010;
 186                         break;
 187                 case (IPTOS_ECN_ECT0):
 188                         tp->t_rcep = 0b100;
 189                         break;
 190                 case (IPTOS_ECN_ECT1):
 191                         tp->t_rcep = 0b011;
 192                         break;
 193                 case (IPTOS_ECN_CE):
 194                         tp->t_rcep = 0b110;
 195                         break;
 196                 }
 197         }
 198 }
 199
 200 /*
 201  * Handle parallel SYN for ECN
 202  */
 203 void
 204 tcp_ecn_input_parallel_syn(struct tcpcb *tp, uint16_t thflags, int iptos)
 205 {
 206         if (thflags & TH_ACK)
 207                 return;
 208         if (V_tcp_do_ecn == 0)
 209                 return;
 210         if ((V_tcp_do_ecn == 1) ||
 211             (V_tcp_do_ecn == 2)) {
 212                 /* RFC3168 ECN handling */
 213                 if ((thflags & (TH_CWR | TH_ECE)) == (TH_CWR | TH_ECE)) {
 214                         tp->t_flags2 |= TF2_ECN_PERMIT;
 215                         tp->t_flags2 |= TF2_ECN_SND_ECE;
 216                         TCPSTAT_INC(tcps_ecn_shs);
 217                 }
 218         } else
 219         if ((V_tcp_do_ecn == 3) ||
 220             (V_tcp_do_ecn == 4)) {
 221                 /* AccECN handling */
 222                 switch (thflags & (TH_AE | TH_CWR | TH_ECE)) {
 223                 default:
 224                 case (0|0|0):
 225                         break;
 226                 case (0|TH_CWR|TH_ECE):
 227                         tp->t_flags2 |= TF2_ECN_PERMIT;
 228                         tp->t_flags2 |= TF2_ECN_SND_ECE;
 229                         TCPSTAT_INC(tcps_ecn_shs);
 230                         break;
 231                 case (TH_AE|TH_CWR|TH_ECE):
 232                         tp->t_flags2 |= TF2_ACE_PERMIT;
 233                         TCPSTAT_INC(tcps_ecn_shs);
 234                         /*
 235                          * Set the AccECN Codepoints on
 236                          * the outgoing <ACK> to the ECN
 237                          * state of the <SYN,ACK>
 238                          * according to table 3 in the
 239                          * AccECN draft
 240                          */
 241                         switch (iptos & IPTOS_ECN_MASK) {
 242                         case (IPTOS_ECN_NOTECT):
 243                                 tp->t_rcep = 0b010;
 244                                 break;
 245                         case (IPTOS_ECN_ECT0):
 246                                 tp->t_rcep = 0b100;
 247                                 break;
 248                         case (IPTOS_ECN_ECT1):
 249                                 tp->t_rcep = 0b011;
 250                                 break;
 251                         case (IPTOS_ECN_CE):
 252                                 tp->t_rcep = 0b110;
 253                                 break;
 254                         }
 255                         break;
 256                 }
 257         }
 258 }
 259
 260 /*
 261  * TCP ECN processing.
 262  */
 263 int
 264 tcp_ecn_input_segment(struct tcpcb *tp, uint16_t thflags, int iptos)
 265 {
 266         int delta_ace = 0;
 267
 268         if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT)) {
 269                 switch (iptos & IPTOS_ECN_MASK) {
 270                 case IPTOS_ECN_CE:
 271                         TCPSTAT_INC(tcps_ecn_ce);
 272                         break;
 273                 case IPTOS_ECN_ECT0:
 274                         TCPSTAT_INC(tcps_ecn_ect0);
 275                         break;
 276                 case IPTOS_ECN_ECT1:
 277                         TCPSTAT_INC(tcps_ecn_ect1);
 278                         break;
 279                 }
 280
 281                 if (tp->t_flags2 & TF2_ACE_PERMIT) {
 282                         if ((iptos & IPTOS_ECN_MASK) == IPTOS_ECN_CE)
 283                                 tp->t_rcep += 1;
 284                         if (tp->t_flags2 & TF2_ECN_PERMIT) {
 285                                 delta_ace = (tcp_ecn_get_ace(thflags) + 8 -
 286                                             (tp->t_scep & 0x07)) & 0x07;
 287                                 tp->t_scep += delta_ace;
 288                         } else {
 289                                 /*
 290                                  * process the final ACK of the 3WHS
 291                                  * see table 3 in draft-ietf-tcpm-accurate-ecn
 292                                  */
 293                                 switch (tcp_ecn_get_ace(thflags)) {
 294                                 case 0b010:
 295                                         /* nonECT SYN or SYN,ACK */
 296                                         /* Fallthrough */
 297                                 case 0b011:
 298                                         /* ECT1 SYN or SYN,ACK */
 299                                         /* Fallthrough */
 300                                 case 0b100:
 301                                         /* ECT0 SYN or SYN,ACK */
 302                                         tp->t_scep = 5;
 303                                         break;
 304                                 case 0b110:
 305                                         /* CE SYN or SYN,ACK */
 306                                         tp->t_scep = 6;
 307                                         tp->snd_cwnd = 2 * tcp_maxseg(tp);
 308                                         break;
 309                                 default:
 310                                         /* mangled AccECN handshake */
 311                                         tp->t_scep = 5;
 312                                         break;
 313                                 }
 314                                 tp->t_flags2 |= TF2_ECN_PERMIT;
 315                         }
 316                 } else {
 317                         /* RFC3168 ECN handling */
 318                         if (thflags & TH_ECE)
 319                                 delta_ace = 1;
 320                         if (thflags & TH_CWR) {
 321                                 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
 322                                 tp->t_flags |= TF_ACKNOW;
 323                         }
 324                         if ((iptos & IPTOS_ECN_MASK) == IPTOS_ECN_CE)
 325                                 tp->t_flags2 |= TF2_ECN_SND_ECE;
 326                 }
 327
 328                 /* Process a packet differently from RFC3168. */
 329                 cc_ecnpkt_handler_flags(tp, thflags, iptos);
 330         }
 331
 332         return delta_ace;
 333 }
 334
 335 /*
 336  * Send ECN setup <SYN> packet header flags
 337  */
 338 uint16_t
 339 tcp_ecn_output_syn_sent(struct tcpcb *tp)
 340 {
 341         uint16_t thflags = 0;
 342
 343         if (V_tcp_do_ecn == 0)
 344                 return thflags;
 345         if (V_tcp_do_ecn == 1) {
 346                 /* Send a RFC3168 ECN setup <SYN> packet */
 347                 if (tp->t_rxtshift >= 1) {
 348                         if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
 349                                 thflags = TH_ECE|TH_CWR;
 350                 } else
 351                         thflags = TH_ECE|TH_CWR;
 352         } else
 353         if (V_tcp_do_ecn == 3) {
 354                 /* Send an Accurate ECN setup <SYN> packet */
 355                 if (tp->t_rxtshift >= 1) {
 356                         if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
 357                                 thflags = TH_ECE|TH_CWR|TH_AE;
 358                 } else
 359                         thflags = TH_ECE|TH_CWR|TH_AE;
 360         }
 361
 362         return thflags;
 363 }
 364
 365 /*
 366  * output processing of ECN feature
 367  * returning IP ECN header codepoint
 368  */
 369 int
 370 tcp_ecn_output_established(struct tcpcb *tp, uint16_t *thflags, int len, bool rxmit)
 371 {
 372         int ipecn = IPTOS_ECN_NOTECT;
 373         bool newdata;
 374
 375         /*
 376          * If the peer has ECN, mark data packets with
 377          * ECN capable transmission (ECT).
 378          * Ignore pure control packets, retransmissions
 379          * and window probes.
 380          */
 381         newdata = (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
 382                     !rxmit &&
 383                     !((tp->t_flags & TF_FORCEDATA) && len == 1));
 384         /* RFC3168 ECN marking, only new data segments */
 385         if (newdata) {
 386                 ipecn = IPTOS_ECN_ECT0;
 387                 TCPSTAT_INC(tcps_ecn_ect0);
 388         }
 389         /*
 390          * Reply with proper ECN notifications.
 391          */
 392         if (tp->t_flags2 & TF2_ACE_PERMIT) {
 393                 *thflags &= ~(TH_AE|TH_CWR|TH_ECE);
 394                 if (tp->t_rcep & 0x01)
 395                         *thflags |= TH_ECE;
 396                 if (tp->t_rcep & 0x02)
 397                         *thflags |= TH_CWR;
 398                 if (tp->t_rcep & 0x04)
 399                         *thflags |= TH_AE;
 400                 if (!(tp->t_flags2 & TF2_ECN_PERMIT)) {
 401                         /*
 402                          * here we process the final
 403                          * ACK of the 3WHS
 404                          */
 405                         if (tp->t_rcep == 0b110) {
 406                                 tp->t_rcep = 6;
 407                         } else {
 408                                 tp->t_rcep = 5;
 409                         }
 410                         tp->t_flags2 |= TF2_ECN_PERMIT;
 411                 }
 412         } else {
 413                 if (newdata &&
 414                     (tp->t_flags2 & TF2_ECN_SND_CWR)) {
 415                         *thflags |= TH_CWR;
 416                         tp->t_flags2 &= ~TF2_ECN_SND_CWR;
 417                 }
 418                 if (tp->t_flags2 & TF2_ECN_SND_ECE)
 419                         *thflags |= TH_ECE;
 420         }
 421
 422         return ipecn;
 423 }
 424
 425 /*
 426  * Set up the ECN related tcpcb fields from
 427  * a syncache entry
 428  */
 429 void
 430 tcp_ecn_syncache_socket(struct tcpcb *tp, struct syncache *sc)
 431 {
 432         if (sc->sc_flags & SCF_ECN_MASK) {
 433                 switch (sc->sc_flags & SCF_ECN_MASK) {
 434                 case SCF_ECN:
 435                         tp->t_flags2 |= TF2_ECN_PERMIT;
 436                         break;
 437                 case SCF_ACE_N:
 438                         /* Fallthrough */
 439                 case SCF_ACE_0:
 440                         /* Fallthrough */
 441                 case SCF_ACE_1:
 442                         tp->t_flags2 |= TF2_ACE_PERMIT;
 443                         tp->t_scep = 5;
 444                         tp->t_rcep = 5;
 445                         break;
 446                 case SCF_ACE_CE:
 447                         tp->t_flags2 |= TF2_ACE_PERMIT;
 448                         tp->t_scep = 6;
 449                         tp->t_rcep = 6;
 450                         break;
 451                 /* undefined SCF codepoint */
 452                 default:
 453                         break;
 454                 }
 455         }
 456 }
 457
 458 /*
 459  * Process a <SYN> packets ECN information, and provide the
 460  * syncache with the relevant information.
 461  */
 462 int
 463 tcp_ecn_syncache_add(uint16_t thflags, int iptos)
 464 {
 465         int scflags = 0;
 466
 467         switch (thflags & (TH_AE|TH_CWR|TH_ECE)) {
 468         /* no ECN */
 469         case (0|0|0):
 470                 break;
 471         /* legacy ECN */
 472         case (0|TH_CWR|TH_ECE):
 473                 scflags = SCF_ECN;
 474                 break;
 475         /* Accurate ECN */
 476         case (TH_AE|TH_CWR|TH_ECE):
 477                 if ((V_tcp_do_ecn == 3) ||
 478                     (V_tcp_do_ecn == 4)) {
 479                         switch (iptos & IPTOS_ECN_MASK) {
 480                         case IPTOS_ECN_CE:
 481                                 scflags = SCF_ACE_CE;
 482                                 break;
 483                         case IPTOS_ECN_ECT0:
 484                                 scflags = SCF_ACE_0;
 485                                 break;
 486                         case IPTOS_ECN_ECT1:
 487                                 scflags = SCF_ACE_1;
 488                                 break;
 489                         case IPTOS_ECN_NOTECT:
 490                                 scflags = SCF_ACE_N;
 491                                 break;
 492                         }
 493                 } else
 494                         scflags = SCF_ECN;
 495                 break;
 496         /* Default Case (section 3.1.2) */
 497         default:
 498                 if ((V_tcp_do_ecn == 3) ||
 499                     (V_tcp_do_ecn == 4)) {
 500                         switch (iptos & IPTOS_ECN_MASK) {
 501                         case IPTOS_ECN_CE:
 502                                 scflags = SCF_ACE_CE;
 503                                 break;
 504                         case IPTOS_ECN_ECT0:
 505                                 scflags = SCF_ACE_0;
 506                                 break;
 507                         case IPTOS_ECN_ECT1:
 508                                 scflags = SCF_ACE_1;
 509                                 break;
 510                         case IPTOS_ECN_NOTECT:
 511                                 scflags = SCF_ACE_N;
 512                                 break;
 513                         }
 514                 }
 515                 break;
 516         }
 517         return scflags;
 518 }
 519
 520 /*
 521  * Set up the ECN information for the <SYN,ACK> from
 522  * syncache information.
 523  */
 524 uint16_t
 525 tcp_ecn_syncache_respond(uint16_t thflags, struct syncache *sc)
 526 {
 527         if ((thflags & TH_SYN) &&
 528             (sc->sc_flags & SCF_ECN_MASK)) {
 529                 switch (sc->sc_flags & SCF_ECN_MASK) {
 530                 case SCF_ECN:
 531                         thflags |= (0 | 0 | TH_ECE);
 532                         TCPSTAT_INC(tcps_ecn_shs);
 533                         break;
 534                 case SCF_ACE_N:
 535                         thflags |= (0 | TH_CWR | 0);
 536                         TCPSTAT_INC(tcps_ecn_shs);
 537                         TCPSTAT_INC(tcps_ace_nect);
 538                         break;
 539                 case SCF_ACE_0:
 540                         thflags |= (TH_AE | 0 | 0);
 541                         TCPSTAT_INC(tcps_ecn_shs);
 542                         TCPSTAT_INC(tcps_ace_ect0);
 543                         break;
 544                 case SCF_ACE_1:
 545                         thflags |= (0 | TH_ECE | TH_CWR);
 546                         TCPSTAT_INC(tcps_ecn_shs);
 547                         TCPSTAT_INC(tcps_ace_ect1);
 548                         break;
 549                 case SCF_ACE_CE:
 550                         thflags |= (TH_AE | TH_CWR | 0);
 551                         TCPSTAT_INC(tcps_ecn_shs);
 552                         TCPSTAT_INC(tcps_ace_ce);
 553                         break;
 554                 /* undefined SCF codepoint */
 555                 default:
 556                         break;
 557                 }
 558         }
 559         return thflags;
 560 }
 561
 562 int
 563 tcp_ecn_get_ace(uint16_t thflags)
 564 {
 565         int ace = 0;
 566
 567         if (thflags & TH_ECE)
 568                 ace += 1;
 569         if (thflags & TH_CWR)
 570                 ace += 2;
 571         if (thflags & TH_AE)
 572                 ace += 4;
 573         return ace;
 574 }