1 #include <unistd.h> 2 #include <string.h> 3 4 #include "tcp_out.h" 5 #include "mtcp.h" 6 #include "ip_in.h" 7 #include "ip_out.h" 8 #include "tcp_in.h" 9 #include "tcp.h" 10 #include "tcp_stream.h" 11 #include "eventpoll.h" 12 #include "timer.h" 13 #include "debug.h" 14 #include "config.h" 15 16 #define TCP_CALCULATE_CHECKSUM TRUE 17 #define ACK_PIGGYBACK TRUE 18 /* Enable this for higher concurrency rate experiments */ 19 #define TRY_SEND_BEFORE_QUEUE /*FALSE*/ TRUE 20 21 #define TCP_MAX_WINDOW 65535 22 23 #define MAX(a, b) ((a)>(b)?(a):(b)) 24 #define MIN(a, b) ((a)<(b)?(a):(b)) 25 26 extern struct pkt_info * 27 ClonePacketCtx(struct pkt_info *to, unsigned char *frame, struct pkt_ctx *from); 28 29 /*----------------------------------------------------------------------------*/ 30 static inline uint16_t 31 CalculateOptionLength(uint8_t flags) 32 { 33 uint16_t optlen = 0; 34 35 if (flags & TCP_FLAG_SYN) { 36 optlen += TCP_OPT_MSS_LEN; 37 #if TCP_OPT_SACK_ENABLED 38 optlen += TCP_OPT_SACK_PERMIT_LEN; 39 #if !TCP_OPT_TIMESTAMP_ENABLED 40 optlen += 2; // insert NOP padding 41 #endif /* TCP_OPT_TIMESTAMP_ENABLED */ 42 #endif /* TCP_OPT_SACK_ENABLED */ 43 44 #if TCP_OPT_TIMESTAMP_ENABLED 45 optlen += TCP_OPT_TIMESTAMP_LEN; 46 #if !TCP_OPT_SACK_ENABLED 47 optlen += 2; // insert NOP padding 48 #endif /* TCP_OPT_SACK_ENABLED */ 49 #endif /* TCP_OPT_TIMESTAMP_ENABLED */ 50 51 optlen += TCP_OPT_WSCALE_LEN + 1; 52 53 } else { 54 55 #if TCP_OPT_TIMESTAMP_ENABLED 56 optlen += TCP_OPT_TIMESTAMP_LEN + 2; 57 #endif 58 59 #if TCP_OPT_SACK_ENABLED 60 if (flags & TCP_FLAG_SACK) { 61 optlen += TCP_OPT_SACK_LEN + 2; 62 } 63 #endif 64 } 65 66 assert(optlen % 4 == 0); 67 68 return optlen; 69 } 70 /*----------------------------------------------------------------------------*/ 71 static inline void 72 GenerateTCPTimestamp(tcp_stream *cur_stream, uint8_t *tcpopt, uint32_t cur_ts) 73 { 74 uint32_t *ts = (uint32_t *)(tcpopt + 2); 75 76 tcpopt[0] = TCP_OPT_TIMESTAMP; 77 tcpopt[1] = TCP_OPT_TIMESTAMP_LEN; 78 ts[0] = htonl(cur_ts); 79 ts[1] = htonl(cur_stream->rcvvar->ts_recent); 80 } 81 /*----------------------------------------------------------------------------*/ 82 static inline void 83 GenerateTCPOptions(tcp_stream *cur_stream, uint32_t cur_ts, 84 uint8_t flags, uint8_t *tcpopt, uint16_t optlen) 85 { 86 int i = 0; 87 88 if (flags & TCP_FLAG_SYN) { 89 uint16_t mss; 90 91 /* MSS option */ 92 mss = cur_stream->sndvar->mss; 93 tcpopt[i++] = TCP_OPT_MSS; 94 tcpopt[i++] = TCP_OPT_MSS_LEN; 95 tcpopt[i++] = mss >> 8; 96 tcpopt[i++] = mss % 256; 97 98 /* SACK permit */ 99 #if TCP_OPT_SACK_ENABLED 100 #if !TCP_OPT_TIMESTAMP_ENABLED 101 tcpopt[i++] = TCP_OPT_NOP; 102 tcpopt[i++] = TCP_OPT_NOP; 103 #endif /* TCP_OPT_TIMESTAMP_ENABLED */ 104 tcpopt[i++] = TCP_OPT_SACK_PERMIT; 105 tcpopt[i++] = TCP_OPT_SACK_PERMIT_LEN; 106 TRACE_SACK("Local SACK permited.\n"); 107 #endif /* TCP_OPT_SACK_ENABLED */ 108 109 /* Timestamp */ 110 #if TCP_OPT_TIMESTAMP_ENABLED 111 #if !TCP_OPT_SACK_ENABLED 112 tcpopt[i++] = TCP_OPT_NOP; 113 tcpopt[i++] = TCP_OPT_NOP; 114 #endif /* TCP_OPT_SACK_ENABLED */ 115 GenerateTCPTimestamp(cur_stream, tcpopt + i, cur_ts); 116 i += TCP_OPT_TIMESTAMP_LEN; 117 #endif /* TCP_OPT_TIMESTAMP_ENABLED */ 118 119 /* Window scale */ 120 tcpopt[i++] = TCP_OPT_NOP; 121 tcpopt[i++] = TCP_OPT_WSCALE; 122 tcpopt[i++] = TCP_OPT_WSCALE_LEN; 123 tcpopt[i++] = cur_stream->sndvar->wscale_mine; 124 125 } else { 126 127 #if TCP_OPT_TIMESTAMP_ENABLED 128 tcpopt[i++] = TCP_OPT_NOP; 129 tcpopt[i++] = TCP_OPT_NOP; 130 GenerateTCPTimestamp(cur_stream, tcpopt + i, cur_ts); 131 i += TCP_OPT_TIMESTAMP_LEN; 132 #endif 133 134 #if TCP_OPT_SACK_ENABLED 135 if (flags & TCP_OPT_SACK) { 136 // TODO: implement SACK support 137 } 138 #endif 139 } 140 141 assert (i == optlen); 142 } 143 /*----------------------------------------------------------------------------*/ 144 int 145 SendTCPPacketStandalone(struct mtcp_manager *mtcp, 146 uint32_t saddr, uint16_t sport, uint32_t daddr, uint16_t dport, 147 uint32_t seq, uint32_t ack_seq, uint16_t window, uint8_t flags, 148 uint8_t *payload, uint16_t payloadlen, 149 uint32_t cur_ts, uint32_t echo_ts) 150 { 151 struct tcphdr *tcph; 152 uint8_t *tcpopt; 153 uint32_t *ts; 154 uint16_t optlen; 155 struct pkt_ctx pctx; 156 int rc = -1; 157 158 memset(&pctx, 0, sizeof(pctx)); 159 optlen = CalculateOptionLength(flags); 160 if (payloadlen > TCP_DEFAULT_MSS + optlen) { 161 TRACE_ERROR("Payload size exceeds MSS.\n"); 162 assert(0); 163 return ERROR; 164 } 165 166 tcph = (struct tcphdr *)IPOutputStandalone(mtcp, 0, 167 saddr, daddr, TCP_HEADER_LEN + optlen + payloadlen, &pctx, cur_ts); 168 if (tcph == NULL) { 169 return ERROR; 170 } 171 memset(tcph, 0, TCP_HEADER_LEN + optlen); 172 173 tcph->source = sport; 174 tcph->dest = dport; 175 176 if (flags & TCP_FLAG_SYN) 177 tcph->syn = TRUE; 178 if (flags & TCP_FLAG_FIN) 179 tcph->fin = TRUE; 180 if (flags & TCP_FLAG_RST) 181 tcph->rst = TRUE; 182 if (flags & TCP_FLAG_PSH) 183 tcph->psh = TRUE; 184 185 tcph->seq = htonl(seq); 186 if (flags & TCP_FLAG_ACK) { 187 tcph->ack = TRUE; 188 tcph->ack_seq = htonl(ack_seq); 189 } 190 191 tcph->window = htons(MIN(window, TCP_MAX_WINDOW)); 192 193 tcpopt = (uint8_t *)tcph + TCP_HEADER_LEN; 194 ts = (uint32_t *)(tcpopt + 4); 195 196 tcpopt[0] = TCP_OPT_NOP; 197 tcpopt[1] = TCP_OPT_NOP; 198 tcpopt[2] = TCP_OPT_TIMESTAMP; 199 tcpopt[3] = TCP_OPT_TIMESTAMP_LEN; 200 ts[0] = htonl(cur_ts); 201 ts[1] = htonl(echo_ts); 202 203 tcph->doff = (TCP_HEADER_LEN + optlen) >> 2; 204 // copy payload if exist 205 if (payloadlen > 0) { 206 memcpy((uint8_t *)tcph + TCP_HEADER_LEN + optlen, payload, payloadlen); 207 } 208 209 #if TCP_CALCULATE_CHECKSUM 210 /* offload TCP checkum if possible */ 211 if (likely(mtcp->iom->dev_ioctl != NULL)) 212 rc = mtcp->iom->dev_ioctl(mtcp->ctx, 213 pctx.out_ifidx, 214 PKT_TX_TCP_CSUM, 215 pctx.p.iph); 216 /* otherwise calculate TCP checksum in S/W */ 217 if (rc == -1) 218 tcph->check = TCPCalcChecksum((uint16_t *)tcph, 219 TCP_HEADER_LEN + 220 optlen + payloadlen, 221 saddr, daddr); 222 #endif 223 224 if (tcph->syn || tcph->fin) { 225 payloadlen++; 226 } 227 228 struct mon_listener *walk; 229 /* callback for monitor raw socket */ 230 TAILQ_FOREACH(walk, &mtcp->monitors, link) 231 if (walk->socket->socktype == MOS_SOCK_MONITOR_RAW) 232 HandleCallback(mtcp, MOS_NULL, walk->socket, MOS_SIDE_BOTH, 233 &pctx, MOS_ON_PKT_IN); 234 return payloadlen; 235 } 236 /*----------------------------------------------------------------------------*/ 237 int 238 SendTCPPacket(struct mtcp_manager *mtcp, tcp_stream *cur_stream, 239 uint32_t cur_ts, uint8_t flags, uint8_t *payload, uint16_t payloadlen) 240 { 241 struct tcphdr *tcph; 242 uint16_t optlen; 243 uint8_t wscale = 0; 244 uint32_t window32 = 0; 245 struct pkt_ctx pctx; 246 int rc = -1; 247 248 memset(&pctx, 0, sizeof(pctx)); 249 optlen = CalculateOptionLength(flags); 250 if (payloadlen > cur_stream->sndvar->mss + optlen) { 251 TRACE_ERROR("Payload size exceeds MSS\n"); 252 return ERROR; 253 } 254 255 tcph = (struct tcphdr *)IPOutput(mtcp, cur_stream, 256 TCP_HEADER_LEN + optlen + payloadlen, &pctx, cur_ts); 257 if (tcph == NULL) { 258 return -2; 259 } 260 memset(tcph, 0, TCP_HEADER_LEN + optlen); 261 262 tcph->source = cur_stream->sport; 263 tcph->dest = cur_stream->dport; 264 265 if (flags & TCP_FLAG_SYN) { 266 tcph->syn = TRUE; 267 if (cur_stream->snd_nxt != cur_stream->sndvar->iss) { 268 TRACE_DBG("Stream %d: weird SYN sequence. " 269 "snd_nxt: %u, iss: %u\n", cur_stream->id, 270 cur_stream->snd_nxt, cur_stream->sndvar->iss); 271 } 272 TRACE_DBG("Stream %d: Sending SYN. seq: %u, ack_seq: %u\n", 273 cur_stream->id, cur_stream->snd_nxt, cur_stream->rcv_nxt); 274 } 275 if (flags & TCP_FLAG_RST) { 276 TRACE_FIN("Stream %d: Sending RST.\n", cur_stream->id); 277 tcph->rst = TRUE; 278 } 279 if (flags & TCP_FLAG_PSH) 280 tcph->psh = TRUE; 281 282 if (flags & TCP_FLAG_WACK) { 283 tcph->seq = htonl(cur_stream->snd_nxt - 1); 284 TRACE_CLWND("%u Sending ACK to get new window advertisement. " 285 "seq: %u, peer_wnd: %u, snd_nxt - snd_una: %u\n", 286 cur_stream->id, 287 cur_stream->snd_nxt - 1, cur_stream->sndvar->peer_wnd, 288 cur_stream->snd_nxt - cur_stream->sndvar->snd_una); 289 } else if (flags & TCP_FLAG_FIN) { 290 tcph->fin = TRUE; 291 292 if (cur_stream->sndvar->fss == 0) { 293 TRACE_ERROR("Stream %u: not fss set. closed: %u\n", 294 cur_stream->id, cur_stream->closed); 295 } 296 tcph->seq = htonl(cur_stream->sndvar->fss); 297 cur_stream->sndvar->is_fin_sent = TRUE; 298 TRACE_FIN("Stream %d: Sending FIN. seq: %u, ack_seq: %u\n", 299 cur_stream->id, cur_stream->snd_nxt, cur_stream->rcv_nxt); 300 } else { 301 tcph->seq = htonl(cur_stream->snd_nxt); 302 } 303 304 if (flags & TCP_FLAG_ACK) { 305 tcph->ack = TRUE; 306 tcph->ack_seq = htonl(cur_stream->rcv_nxt); 307 cur_stream->sndvar->ts_lastack_sent = cur_ts; 308 cur_stream->last_active_ts = cur_ts; 309 UpdateTimeoutList(mtcp, cur_stream); 310 } 311 312 if (flags & TCP_FLAG_SYN) { 313 wscale = 0; 314 } else { 315 wscale = cur_stream->sndvar->wscale_mine; 316 } 317 318 window32 = cur_stream->rcvvar->rcv_wnd >> wscale; 319 tcph->window = htons((uint16_t)MIN(window32, TCP_MAX_WINDOW)); 320 /* if the advertised window is 0, we need to advertise again later */ 321 if (window32 == 0) { 322 cur_stream->need_wnd_adv = TRUE; 323 } 324 325 GenerateTCPOptions(cur_stream, cur_ts, flags, 326 (uint8_t *)tcph + TCP_HEADER_LEN, optlen); 327 328 tcph->doff = (TCP_HEADER_LEN + optlen) >> 2; 329 // copy payload if exist 330 if (payloadlen > 0) { 331 memcpy((uint8_t *)tcph + TCP_HEADER_LEN + optlen, payload, payloadlen); 332 } 333 334 #if TCP_CALCULATE_CHECKSUM 335 if (likely(mtcp->iom->dev_ioctl != NULL)) 336 rc = mtcp->iom->dev_ioctl(mtcp->ctx, 337 pctx.out_ifidx, 338 PKT_TX_TCP_CSUM, 339 pctx.p.iph); 340 if (rc == -1) 341 tcph->check = TCPCalcChecksum((uint16_t *)tcph, 342 TCP_HEADER_LEN + 343 optlen + payloadlen, 344 cur_stream->saddr, 345 cur_stream->daddr); 346 #endif 347 cur_stream->snd_nxt += payloadlen; 348 349 if (tcph->syn || tcph->fin) { 350 cur_stream->snd_nxt++; 351 payloadlen++; 352 } 353 354 if (payloadlen > 0) { 355 if (cur_stream->state > TCP_ST_ESTABLISHED) { 356 TRACE_FIN("Payload after ESTABLISHED: length: %d, snd_nxt: %u\n", 357 payloadlen, cur_stream->snd_nxt); 358 } 359 360 /* update retransmission timer if have payload */ 361 cur_stream->sndvar->ts_rto = cur_ts + cur_stream->sndvar->rto; 362 TRACE_RTO("Updating retransmission timer. " 363 "cur_ts: %u, rto: %u, ts_rto: %u\n", 364 cur_ts, cur_stream->sndvar->rto, cur_stream->sndvar->ts_rto); 365 AddtoRTOList(mtcp, cur_stream); 366 } 367 368 struct mon_listener *walk; 369 /* callback for monitor raw socket */ 370 TAILQ_FOREACH(walk, &mtcp->monitors, link) 371 if (walk->socket->socktype == MOS_SOCK_MONITOR_RAW) 372 HandleCallback(mtcp, MOS_NULL, walk->socket, MOS_SIDE_BOTH, 373 &pctx, MOS_ON_PKT_IN); 374 375 if (mtcp->num_msp /* this means that stream monitor is on */) { 376 FillPacketContextTCPInfo(&pctx, tcph); 377 378 /* New abstraction for monitor stream */ 379 struct tcp_stream *recvside_stream = cur_stream->pair_stream; 380 struct tcp_stream *sendside_stream = cur_stream; 381 382 #ifdef NEWPPEEK 383 if (recvside_stream->rcvvar && recvside_stream->rcvvar->rcvbuf) 384 pctx.p.offset = (uint64_t)seq2loff(recvside_stream->rcvvar->rcvbuf, 385 pctx.p.seq, recvside_stream->rcvvar->irs + 1); 386 #endif 387 388 UpdateMonitor(mtcp, sendside_stream, recvside_stream, &pctx, false); 389 } 390 391 #ifdef PKTDUMP 392 DumpPacket(mtcp, 393 (char *)tcph - sizeof(struct iphdr) - sizeof(struct ethhdr), 394 payloadlen + sizeof(struct iphdr) + sizeof(struct ethhdr), 395 "OUT", -1); 396 #endif 397 398 399 return payloadlen; 400 } 401 /*----------------------------------------------------------------------------*/ 402 static int 403 FlushTCPSendingBuffer(mtcp_manager_t mtcp, tcp_stream *cur_stream, uint32_t cur_ts) 404 { 405 struct tcp_send_vars *sndvar = cur_stream->sndvar; 406 const uint32_t maxlen = sndvar->mss - CalculateOptionLength(TCP_FLAG_ACK); 407 uint8_t *data; 408 uint32_t buffered_len; 409 uint32_t seq; 410 uint16_t len; 411 int16_t sndlen; 412 uint32_t window; 413 int packets = 0; 414 415 if (!sndvar->sndbuf) { 416 TRACE_ERROR("Stream %d: No send buffer available.\n", cur_stream->id); 417 assert(0); 418 return 0; 419 } 420 421 if (sndvar->sndbuf->len == 0) { 422 return 0; 423 } 424 425 window = MIN(sndvar->cwnd, sndvar->peer_wnd); 426 427 while (1) { 428 seq = cur_stream->snd_nxt; 429 430 if (TCP_SEQ_LT(seq, sndvar->sndbuf->head_seq)) { 431 TRACE_ERROR("Stream %d: Invalid sequence to send. " 432 "state: %s, seq: %u, head_seq: %u.\n", 433 cur_stream->id, TCPStateToString(cur_stream), 434 seq, sndvar->sndbuf->head_seq); 435 assert(0); 436 break; 437 } 438 buffered_len = sndvar->sndbuf->head_seq + sndvar->sndbuf->len - seq; 439 if (cur_stream->state > TCP_ST_ESTABLISHED) { 440 TRACE_FIN("head_seq: %u, len: %u, seq: %u, " 441 "buffered_len: %u\n", sndvar->sndbuf->head_seq, 442 sndvar->sndbuf->len, seq, buffered_len); 443 } 444 if (buffered_len == 0) 445 break; 446 447 data = sndvar->sndbuf->head + 448 (seq - sndvar->sndbuf->head_seq); 449 450 if (buffered_len > maxlen) { 451 len = maxlen; 452 } else { 453 len = buffered_len; 454 } 455 456 if (len <= 0) 457 break; 458 459 if (cur_stream->state > TCP_ST_ESTABLISHED) { 460 TRACE_FIN("Flushing after ESTABLISHED: seq: %u, len: %u, " 461 "buffered_len: %u\n", seq, len, buffered_len); 462 } 463 464 if (seq - sndvar->snd_una + len > window) { 465 /* Ask for new window advertisement to peer */ 466 if (seq - sndvar->snd_una + len > sndvar->peer_wnd) { 467 TRACE_DBG("Full peer window. " 468 "peer_wnd: %u, (snd_nxt-snd_una): %u\n", 469 sndvar->peer_wnd, seq - sndvar->snd_una); 470 if (TS_TO_MSEC(cur_ts - sndvar->ts_lastack_sent) > 500) { 471 EnqueueACK(mtcp, cur_stream, cur_ts, ACK_OPT_WACK); 472 } 473 } 474 return -3; 475 } 476 477 sndlen = SendTCPPacket(mtcp, cur_stream, cur_ts, 478 TCP_FLAG_ACK, data, len); 479 if (sndlen < 0) { 480 return sndlen; 481 } 482 packets++; 483 } 484 485 return packets; 486 } 487 /*----------------------------------------------------------------------------*/ 488 static inline int 489 SendControlPacket(mtcp_manager_t mtcp, tcp_stream *cur_stream, uint32_t cur_ts) 490 { 491 struct tcp_send_vars *sndvar = cur_stream->sndvar; 492 int ret = 0; 493 int flag = 0; 494 495 switch (cur_stream->state) { 496 case TCP_ST_SYN_SENT: /* Send SYN here */ 497 flag = TCP_FLAG_SYN; 498 break; 499 case TCP_ST_SYN_RCVD: /* Send SYN/ACK here */ 500 cur_stream->snd_nxt = sndvar->iss; 501 flag = TCP_FLAG_SYN | TCP_FLAG_ACK; 502 break; 503 case TCP_ST_ESTABLISHED: /* Send ACK here */ 504 case TCP_ST_CLOSE_WAIT: /* Send ACK for the FIN here */ 505 case TCP_ST_FIN_WAIT_2: /* Send ACK here */ 506 case TCP_ST_TIME_WAIT: /* Send ACK here */ 507 flag = TCP_FLAG_ACK; 508 break; 509 case TCP_ST_LAST_ACK: 510 case TCP_ST_FIN_WAIT_1: 511 /* if it is on ack_list, send it after sending ack */ 512 if (sndvar->on_send_list || sndvar->on_ack_list) 513 return (-1); 514 flag = TCP_FLAG_FIN | TCP_FLAG_ACK; /* Send FIN/ACK here */ 515 break; 516 case TCP_ST_CLOSING: 517 if (sndvar->is_fin_sent) { 518 /* if the sequence is for FIN, send FIN */ 519 flag = (cur_stream->snd_nxt == sndvar->fss) ? 520 (TCP_FLAG_FIN | TCP_FLAG_ACK) : TCP_FLAG_ACK; 521 } else { 522 /* if FIN is not sent, send fin with ack */ 523 flag = TCP_FLAG_FIN | TCP_FLAG_ACK; 524 } 525 case TCP_ST_CLOSED_RSVD: /* Send RST here */ 526 TRACE_DBG("Stream %d: Try sending RST (TCP_ST_CLOSED_RSVD)\n", 527 cur_stream->id); 528 /* first flush the data and ack */ 529 if (sndvar->on_send_list || sndvar->on_ack_list) 530 return (-1); 531 ret = SendTCPPacket(mtcp, cur_stream, cur_ts, TCP_FLAG_RST, NULL, 0); 532 if (ret >= 0) 533 DestroyTCPStream(mtcp, cur_stream); 534 return (ret); 535 default: 536 TRACE_ERROR("Stream %d: shouldn't send a control packet\n", 537 cur_stream->id); 538 assert(0); /* can't reach here! */ 539 return (0); 540 } 541 542 return SendTCPPacket(mtcp, cur_stream, cur_ts, flag, NULL, 0); 543 } 544 /*----------------------------------------------------------------------------*/ 545 inline int 546 WriteTCPControlList(mtcp_manager_t mtcp, 547 struct mtcp_sender *sender, uint32_t cur_ts, int thresh) 548 { 549 tcp_stream *cur_stream; 550 tcp_stream *next, *last; 551 int cnt = 0; 552 int ret; 553 554 thresh = MIN(thresh, sender->control_list_cnt); 555 556 /* Send TCP control messages */ 557 cnt = 0; 558 cur_stream = TAILQ_FIRST(&sender->control_list); 559 last = TAILQ_LAST(&sender->control_list, control_head); 560 while (cur_stream) { 561 if (++cnt > thresh) 562 break; 563 564 TRACE_LOOP("Inside control loop. cnt: %u, stream: %d\n", 565 cnt, cur_stream->id); 566 next = TAILQ_NEXT(cur_stream, sndvar->control_link); 567 568 TAILQ_REMOVE(&sender->control_list, cur_stream, sndvar->control_link); 569 sender->control_list_cnt--; 570 571 if (cur_stream->sndvar->on_control_list) { 572 cur_stream->sndvar->on_control_list = FALSE; 573 //TRACE_DBG("Stream %u: Sending control packet\n", cur_stream->id); 574 ret = SendControlPacket(mtcp, cur_stream, cur_ts); 575 if (ret < 0) { 576 TAILQ_INSERT_HEAD(&sender->control_list, 577 cur_stream, sndvar->control_link); 578 cur_stream->sndvar->on_control_list = TRUE; 579 sender->control_list_cnt++; 580 /* since there is no available write buffer, break */ 581 break; 582 } 583 } else { 584 TRACE_ERROR("Stream %d: not on control list.\n", cur_stream->id); 585 } 586 587 if (cur_stream == last) 588 break; 589 cur_stream = next; 590 } 591 592 return cnt; 593 } 594 /*----------------------------------------------------------------------------*/ 595 inline int 596 WriteTCPDataList(mtcp_manager_t mtcp, 597 struct mtcp_sender *sender, uint32_t cur_ts, int thresh) 598 { 599 tcp_stream *cur_stream; 600 tcp_stream *next, *last; 601 int cnt = 0; 602 int ret; 603 604 /* Send data */ 605 cnt = 0; 606 cur_stream = TAILQ_FIRST(&sender->send_list); 607 last = TAILQ_LAST(&sender->send_list, send_head); 608 while (cur_stream) { 609 if (++cnt > thresh) 610 break; 611 612 TRACE_LOOP("Inside send loop. cnt: %u, stream: %d\n", 613 cnt, cur_stream->id); 614 next = TAILQ_NEXT(cur_stream, sndvar->send_link); 615 616 TAILQ_REMOVE(&sender->send_list, cur_stream, sndvar->send_link); 617 if (cur_stream->sndvar->on_send_list) { 618 ret = 0; 619 620 /* Send data here */ 621 /* Only can send data when ESTABLISHED or CLOSE_WAIT */ 622 if (cur_stream->state == TCP_ST_ESTABLISHED) { 623 if (cur_stream->sndvar->on_control_list) { 624 /* delay sending data after until on_control_list becomes off */ 625 //TRACE_DBG("Stream %u: delay sending data.\n", cur_stream->id); 626 ret = -1; 627 } else { 628 ret = FlushTCPSendingBuffer(mtcp, cur_stream, cur_ts); 629 } 630 } else if (cur_stream->state == TCP_ST_CLOSE_WAIT || 631 cur_stream->state == TCP_ST_FIN_WAIT_1 || 632 cur_stream->state == TCP_ST_LAST_ACK) { 633 ret = FlushTCPSendingBuffer(mtcp, cur_stream, cur_ts); 634 } else { 635 TRACE_DBG("Stream %d: on_send_list at state %s\n", 636 cur_stream->id, TCPStateToString(cur_stream)); 637 #if DUMP_STREAM 638 DumpStream(mtcp, cur_stream); 639 #endif 640 } 641 642 if (ret < 0) { 643 TAILQ_INSERT_TAIL(&sender->send_list, cur_stream, sndvar->send_link); 644 /* since there is no available write buffer, break */ 645 break; 646 647 } else { 648 cur_stream->sndvar->on_send_list = FALSE; 649 sender->send_list_cnt--; 650 /* the ret value is the number of packets sent. */ 651 /* decrease ack_cnt for the piggybacked acks */ 652 #if ACK_PIGGYBACK 653 if (cur_stream->sndvar->ack_cnt > 0) { 654 if (cur_stream->sndvar->ack_cnt > ret) { 655 cur_stream->sndvar->ack_cnt -= ret; 656 } else { 657 cur_stream->sndvar->ack_cnt = 0; 658 } 659 } 660 #endif 661 #if 1 662 if (cur_stream->control_list_waiting) { 663 if (!cur_stream->sndvar->on_ack_list) { 664 cur_stream->control_list_waiting = FALSE; 665 AddtoControlList(mtcp, cur_stream, cur_ts); 666 } 667 } 668 #endif 669 } 670 } else { 671 TRACE_ERROR("Stream %d: not on send list.\n", cur_stream->id); 672 #ifdef DUMP_STREAM 673 DumpStream(mtcp, cur_stream); 674 #endif 675 } 676 677 if (cur_stream == last) 678 break; 679 cur_stream = next; 680 } 681 682 return cnt; 683 } 684 /*----------------------------------------------------------------------------*/ 685 inline int 686 WriteTCPACKList(mtcp_manager_t mtcp, 687 struct mtcp_sender *sender, uint32_t cur_ts, int thresh) 688 { 689 tcp_stream *cur_stream; 690 tcp_stream *next, *last; 691 int to_ack; 692 int cnt = 0; 693 int ret; 694 695 /* Send aggregated acks */ 696 cnt = 0; 697 cur_stream = TAILQ_FIRST(&sender->ack_list); 698 last = TAILQ_LAST(&sender->ack_list, ack_head); 699 while (cur_stream) { 700 if (++cnt > thresh) 701 break; 702 703 TRACE_LOOP("Inside ack loop. cnt: %u\n", cnt); 704 next = TAILQ_NEXT(cur_stream, sndvar->ack_link); 705 706 if (cur_stream->sndvar->on_ack_list) { 707 /* this list is only to ack the data packets */ 708 /* if the ack is not data ack, then it will not process here */ 709 to_ack = FALSE; 710 if (cur_stream->state == TCP_ST_ESTABLISHED || 711 cur_stream->state == TCP_ST_CLOSE_WAIT || 712 cur_stream->state == TCP_ST_FIN_WAIT_1 || 713 cur_stream->state == TCP_ST_FIN_WAIT_2 || 714 cur_stream->state == TCP_ST_TIME_WAIT) { 715 /* TIMEWAIT is possible since the ack is queued 716 at FIN_WAIT_2 */ 717 #ifdef NEWRB 718 tcprb_t *rb; 719 if ((rb = cur_stream->rcvvar->rcvbuf) && 720 TCP_SEQ_LEQ(cur_stream->rcv_nxt, 721 (cur_stream->rcvvar->irs + 1) + rb->pile 722 + tcprb_cflen(rb))) { 723 to_ack = TRUE; 724 } 725 #else 726 if (cur_stream->rcvvar->rcvbuf) { 727 if (TCP_SEQ_LEQ(cur_stream->rcv_nxt, 728 cur_stream->rcvvar->rcvbuf->head_seq + 729 cur_stream->rcvvar->rcvbuf->merged_len)) { 730 to_ack = TRUE; 731 } 732 } 733 #endif 734 } else { 735 TRACE_DBG("Stream %u (%s): " 736 "Try sending ack at not proper state. " 737 "seq: %u, ack_seq: %u, on_control_list: %u\n", 738 cur_stream->id, TCPStateToString(cur_stream), 739 cur_stream->snd_nxt, cur_stream->rcv_nxt, 740 cur_stream->sndvar->on_control_list); 741 #ifdef DUMP_STREAM 742 DumpStream(mtcp, cur_stream); 743 #endif 744 } 745 746 if (to_ack) { 747 /* send the queued ack packets */ 748 while (cur_stream->sndvar->ack_cnt > 0) { 749 ret = SendTCPPacket(mtcp, cur_stream, 750 cur_ts, TCP_FLAG_ACK, NULL, 0); 751 if (ret < 0) { 752 /* since there is no available write buffer, break */ 753 break; 754 } 755 cur_stream->sndvar->ack_cnt--; 756 } 757 758 /* if is_wack is set, send packet to get window advertisement */ 759 if (cur_stream->sndvar->is_wack) { 760 cur_stream->sndvar->is_wack = FALSE; 761 ret = SendTCPPacket(mtcp, cur_stream, 762 cur_ts, TCP_FLAG_ACK | TCP_FLAG_WACK, NULL, 0); 763 if (ret < 0) { 764 /* since there is no available write buffer, break */ 765 cur_stream->sndvar->is_wack = TRUE; 766 } 767 } 768 769 if (!(cur_stream->sndvar->ack_cnt || cur_stream->sndvar->is_wack)) { 770 cur_stream->sndvar->on_ack_list = FALSE; 771 TAILQ_REMOVE(&sender->ack_list, cur_stream, sndvar->ack_link); 772 sender->ack_list_cnt--; 773 } 774 } else { 775 cur_stream->sndvar->on_ack_list = FALSE; 776 cur_stream->sndvar->ack_cnt = 0; 777 cur_stream->sndvar->is_wack = 0; 778 TAILQ_REMOVE(&sender->ack_list, cur_stream, sndvar->ack_link); 779 sender->ack_list_cnt--; 780 } 781 782 if (cur_stream->control_list_waiting) { 783 if (!cur_stream->sndvar->on_send_list) { 784 cur_stream->control_list_waiting = FALSE; 785 AddtoControlList(mtcp, cur_stream, cur_ts); 786 } 787 } 788 } else { 789 TRACE_ERROR("Stream %d: not on ack list.\n", cur_stream->id); 790 TAILQ_REMOVE(&sender->ack_list, cur_stream, sndvar->ack_link); 791 sender->ack_list_cnt--; 792 #ifdef DUMP_STREAM 793 thread_printf(mtcp, mtcp->log_fp, 794 "Stream %u: not on ack list.\n", cur_stream->id); 795 DumpStream(mtcp, cur_stream); 796 #endif 797 } 798 799 if (cur_stream == last) 800 break; 801 cur_stream = next; 802 } 803 804 return cnt; 805 } 806 /*----------------------------------------------------------------------------*/ 807 inline struct mtcp_sender * 808 GetSender(mtcp_manager_t mtcp, tcp_stream *cur_stream) 809 { 810 if (cur_stream->sndvar->nif_out < 0) { 811 return mtcp->g_sender; 812 813 } else if (cur_stream->sndvar->nif_out >= g_config.mos->netdev_table->num) { 814 TRACE_ERROR("(NEVER HAPPEN) Failed to find appropriate sender.\n"); 815 return NULL; 816 817 } else { 818 return mtcp->n_sender[cur_stream->sndvar->nif_out]; 819 } 820 } 821 /*----------------------------------------------------------------------------*/ 822 inline void 823 AddtoControlList(mtcp_manager_t mtcp, tcp_stream *cur_stream, uint32_t cur_ts) 824 { 825 #if TRY_SEND_BEFORE_QUEUE 826 int ret; 827 struct mtcp_sender *sender = GetSender(mtcp, cur_stream); 828 assert(sender != NULL); 829 830 ret = SendControlPacket(mtcp, cur_stream, cur_ts); 831 if (ret < 0) { 832 #endif 833 if (!cur_stream->sndvar->on_control_list) { 834 struct mtcp_sender *sender = GetSender(mtcp, cur_stream); 835 assert(sender != NULL); 836 837 cur_stream->sndvar->on_control_list = TRUE; 838 TAILQ_INSERT_TAIL(&sender->control_list, cur_stream, sndvar->control_link); 839 sender->control_list_cnt++; 840 //TRACE_DBG("Stream %u: added to control list (cnt: %d)\n", 841 // cur_stream->id, sender->control_list_cnt); 842 } 843 #if TRY_SEND_BEFORE_QUEUE 844 } else { 845 if (cur_stream->sndvar->on_control_list) { 846 cur_stream->sndvar->on_control_list = FALSE; 847 TAILQ_REMOVE(&sender->control_list, cur_stream, sndvar->control_link); 848 sender->control_list_cnt--; 849 } 850 } 851 #endif 852 } 853 /*----------------------------------------------------------------------------*/ 854 inline void 855 AddtoSendList(mtcp_manager_t mtcp, tcp_stream *cur_stream) 856 { 857 struct mtcp_sender *sender = GetSender(mtcp, cur_stream); 858 assert(sender != NULL); 859 860 if(!cur_stream->sndvar->sndbuf) { 861 TRACE_ERROR("[%d] Stream %d: No send buffer available.\n", 862 mtcp->ctx->cpu, 863 cur_stream->id); 864 assert(0); 865 return; 866 } 867 868 if (!cur_stream->sndvar->on_send_list) { 869 cur_stream->sndvar->on_send_list = TRUE; 870 TAILQ_INSERT_TAIL(&sender->send_list, cur_stream, sndvar->send_link); 871 sender->send_list_cnt++; 872 } 873 } 874 /*----------------------------------------------------------------------------*/ 875 inline void 876 AddtoACKList(mtcp_manager_t mtcp, tcp_stream *cur_stream) 877 { 878 struct mtcp_sender *sender = GetSender(mtcp, cur_stream); 879 assert(sender != NULL); 880 881 if (!cur_stream->sndvar->on_ack_list) { 882 cur_stream->sndvar->on_ack_list = TRUE; 883 TAILQ_INSERT_TAIL(&sender->ack_list, cur_stream, sndvar->ack_link); 884 sender->ack_list_cnt++; 885 } 886 } 887 /*----------------------------------------------------------------------------*/ 888 inline void 889 RemoveFromControlList(mtcp_manager_t mtcp, tcp_stream *cur_stream) 890 { 891 struct mtcp_sender *sender = GetSender(mtcp, cur_stream); 892 assert(sender != NULL); 893 894 if (cur_stream->sndvar->on_control_list) { 895 cur_stream->sndvar->on_control_list = FALSE; 896 TAILQ_REMOVE(&sender->control_list, cur_stream, sndvar->control_link); 897 sender->control_list_cnt--; 898 //TRACE_DBG("Stream %u: Removed from control list (cnt: %d)\n", 899 // cur_stream->id, sender->control_list_cnt); 900 } 901 } 902 /*----------------------------------------------------------------------------*/ 903 inline void 904 RemoveFromSendList(mtcp_manager_t mtcp, tcp_stream *cur_stream) 905 { 906 struct mtcp_sender *sender = GetSender(mtcp, cur_stream); 907 assert(sender != NULL); 908 909 if (cur_stream->sndvar->on_send_list) { 910 cur_stream->sndvar->on_send_list = FALSE; 911 TAILQ_REMOVE(&sender->send_list, cur_stream, sndvar->send_link); 912 sender->send_list_cnt--; 913 } 914 } 915 /*----------------------------------------------------------------------------*/ 916 inline void 917 RemoveFromACKList(mtcp_manager_t mtcp, tcp_stream *cur_stream) 918 { 919 struct mtcp_sender *sender = GetSender(mtcp, cur_stream); 920 assert(sender != NULL); 921 922 if (cur_stream->sndvar->on_ack_list) { 923 cur_stream->sndvar->on_ack_list = FALSE; 924 TAILQ_REMOVE(&sender->ack_list, cur_stream, sndvar->ack_link); 925 sender->ack_list_cnt--; 926 } 927 } 928 /*----------------------------------------------------------------------------*/ 929 inline void 930 EnqueueACK(mtcp_manager_t mtcp, 931 tcp_stream *cur_stream, uint32_t cur_ts, uint8_t opt) 932 { 933 if (!(cur_stream->state == TCP_ST_ESTABLISHED || 934 cur_stream->state == TCP_ST_CLOSE_WAIT || 935 cur_stream->state == TCP_ST_FIN_WAIT_1 || 936 cur_stream->state == TCP_ST_FIN_WAIT_2)) { 937 TRACE_DBG("Stream %u: Enqueueing ack at state %s\n", 938 cur_stream->id, TCPStateToString(cur_stream)); 939 } 940 941 if (opt == ACK_OPT_NOW) { 942 if (cur_stream->sndvar->ack_cnt < cur_stream->sndvar->ack_cnt + 1) { 943 cur_stream->sndvar->ack_cnt++; 944 } 945 } else if (opt == ACK_OPT_AGGREGATE) { 946 if (cur_stream->sndvar->ack_cnt == 0) { 947 cur_stream->sndvar->ack_cnt = 1; 948 } 949 } else if (opt == ACK_OPT_WACK) { 950 cur_stream->sndvar->is_wack = TRUE; 951 } 952 AddtoACKList(mtcp, cur_stream); 953 } 954 /*----------------------------------------------------------------------------*/ 955 inline void 956 DumpControlList(mtcp_manager_t mtcp, struct mtcp_sender *sender) 957 { 958 tcp_stream *stream; 959 960 TRACE_DBG("Dumping control list (count: %d):\n", sender->control_list_cnt); 961 TAILQ_FOREACH(stream, &sender->control_list, sndvar->control_link) { 962 TRACE_DBG("Stream id: %u in control list\n", stream->id); 963 } 964 } 965 /*----------------------------------------------------------------------------*/ 966 static inline void 967 UpdatePassiveSendTCPContext_SynSent(struct tcp_stream *cur_stream, 968 struct pkt_ctx *pctx) 969 { 970 assert(cur_stream); 971 assert(pctx); 972 973 /* add event */ 974 if (cur_stream->state < TCP_ST_SYN_SENT) { 975 cur_stream->cb_events |= MOS_ON_TCP_STATE_CHANGE; 976 cur_stream->cb_events |= MOS_ON_CONN_START; 977 } 978 /* initialize TCP send variables of send-side stream */ 979 cur_stream->sndvar->cwnd = 1; 980 cur_stream->sndvar->ssthresh = cur_stream->sndvar->mss * 10; 981 cur_stream->sndvar->ip_id = htons(pctx->p.iph->id); 982 cur_stream->sndvar->iss = pctx->p.seq; 983 cur_stream->snd_nxt = pctx->p.seq + 1; 984 cur_stream->state = TCP_ST_SYN_SENT; 985 cur_stream->last_active_ts = pctx->p.cur_ts; 986 987 /* receive-side conn start event can also be tagged here */ 988 /* blocked since tcp_in.c takes care of this.. */ 989 /* cur_stream->pair_stream->cb_events |= MOS_ON_CONN_START; */ 990 } 991 /*----------------------------------------------------------------------------*/ 992 /** 993 * Called (when monitoring mode is enabled).. for every incoming packet from the 994 * NIC. 995 */ 996 void 997 UpdatePassiveSendTCPContext(mtcp_manager_t mtcp, struct tcp_stream *cur_stream, 998 struct pkt_ctx *pctx) 999 { 1000 struct tcphdr *tcph; 1001 1002 assert(cur_stream); 1003 tcph = pctx->p.tcph; 1004 1005 /* if it is a new TCP stream from client */ 1006 if (tcph->syn && !tcph->ack && cur_stream->state <= TCP_ST_SYN_SENT) { 1007 TRACE_STATE("Stream %d: %s\n", 1008 cur_stream->id, TCPStateToString(cur_stream)); 1009 UpdatePassiveSendTCPContext_SynSent(cur_stream, pctx); 1010 AddtoTimeoutList(mtcp, cur_stream); 1011 return; 1012 } 1013 1014 if (tcph->ack) { 1015 cur_stream->sndvar->ts_lastack_sent = pctx->p.cur_ts; 1016 cur_stream->last_active_ts = pctx->p.cur_ts; 1017 } 1018 1019 cur_stream->snd_nxt = pctx->p.seq + pctx->p.payloadlen; 1020 1021 /* test for reset packet */ 1022 if (tcph->rst) { 1023 cur_stream->have_reset = TRUE; 1024 /* test for reset packet */ 1025 cur_stream->state = TCP_ST_CLOSED_RSVD; 1026 cur_stream->cb_events |= MOS_ON_TCP_STATE_CHANGE; 1027 TRACE_STATE("Stream %d: %s\n", 1028 cur_stream->id, 1029 TCPStateToString(cur_stream)); 1030 return; 1031 } 1032 1033 /* 1034 * for all others, state transitioning is based on 1035 * current tcp_stream state 1036 */ 1037 switch (cur_stream->state) { 1038 case TCP_ST_SYN_SENT: 1039 /* control should not come here */ 1040 /* UpdatePassiveReceiveTCPContext() should take care of this */ 1041 #ifdef BE_RESILIENT_TO_PACKET_DROP 1042 if (tcph->ack && TCP_SEQ_GT(pctx->p.seq, cur_stream->sndvar->iss)) { 1043 cur_stream->state = TCP_ST_ESTABLISHED; 1044 cur_stream->cb_events |= MOS_ON_TCP_STATE_CHANGE; 1045 cur_stream->snd_nxt = pctx->p.seq; 1046 cur_stream->rcv_nxt = pctx->p.ack_seq; 1047 goto __Handle_TCP_ST_ESTABLISHED; 1048 } 1049 #endif 1050 break; 1051 case TCP_ST_SYN_RCVD: 1052 if (!tcph->ack) 1053 break; 1054 1055 if (tcph->syn) { 1056 cur_stream->sndvar->iss = pctx->p.seq; 1057 cur_stream->snd_nxt = cur_stream->sndvar->iss + 1; 1058 TRACE_DBG("Stream %d (TCP_ST_SYN_RCVD): " 1059 "setting seq: %u = iss\n", 1060 cur_stream->id, pctx->p.seq); 1061 } 1062 #ifdef BE_RESILIENT_TO_PACKET_DROP 1063 else { 1064 cur_stream->state = TCP_ST_ESTABLISHED; 1065 cur_stream->cb_events |= MOS_ON_TCP_STATE_CHANGE; 1066 cur_stream->snd_nxt = pctx->p.seq; 1067 cur_stream->rcv_nxt = pctx->p.ack_seq; 1068 goto __Handle_TCP_ST_ESTABLISHED; 1069 } 1070 #endif 1071 TRACE_STATE("Stream %d: %s\n", 1072 cur_stream->id, 1073 TCPStateToString(cur_stream)); 1074 break; 1075 case TCP_ST_ESTABLISHED: 1076 #ifdef BE_RESILIENT_TO_PACKET_DROP 1077 __Handle_TCP_ST_ESTABLISHED: 1078 #endif 1079 /* if application decides to close, fin pkt is sent */ 1080 #ifdef BE_RESILIENT_TO_PACKET_DROP 1081 if (tcph->ack && TCP_SEQ_GT(ntohl(tcph->ack_seq), cur_stream->rcv_nxt)) 1082 { 1083 RAISE_DEBUG_EVENT(mtcp, cur_stream, 1084 "Move rcv_nxt from %u to %u.\n", 1085 cur_stream->rcv_nxt, ntohl(tcph->ack_seq)); 1086 cur_stream->rcv_nxt = ntohl(tcph->ack_seq); 1087 } 1088 #endif 1089 if (tcph->fin) { 1090 cur_stream->state = TCP_ST_FIN_WAIT_1; 1091 cur_stream->cb_events |= MOS_ON_TCP_STATE_CHANGE; 1092 cur_stream->sndvar->fss = pctx->p.seq + pctx->p.payloadlen; 1093 cur_stream->sndvar->is_fin_sent = TRUE; 1094 cur_stream->snd_nxt++; 1095 TRACE_STATE("Stream %d: %s\n", 1096 cur_stream->id, 1097 TCPStateToString(cur_stream)); 1098 } else { 1099 /* creating tcp send buffer still pending.. */ 1100 /* do we need peek for send buffer? */ 1101 } 1102 break; 1103 case TCP_ST_CLOSE_WAIT: 1104 /* if application decides to close, fin pkt is sent */ 1105 #ifdef BE_RESILIENT_TO_PACKET_DROP 1106 if (tcph->ack && TCP_SEQ_GT(ntohl(tcph->ack_seq), cur_stream->rcv_nxt)) 1107 { 1108 RAISE_DEBUG_EVENT(mtcp, cur_stream, 1109 "Move rcv_nxt from %u to %u.\n", 1110 cur_stream->rcv_nxt, ntohl(tcph->ack_seq)); 1111 cur_stream->rcv_nxt = ntohl(tcph->ack_seq); 1112 } 1113 #endif 1114 if (tcph->fin) { 1115 cur_stream->sndvar->fss = pctx->p.seq + pctx->p.payloadlen; 1116 cur_stream->sndvar->is_fin_sent = TRUE; 1117 cur_stream->snd_nxt++; 1118 cur_stream->state = TCP_ST_LAST_ACK; 1119 cur_stream->cb_events |= MOS_ON_TCP_STATE_CHANGE; 1120 TRACE_STATE("Stream %d: %s\n", 1121 cur_stream->id, 1122 TCPStateToString(cur_stream)); 1123 } else if (tcph->ack) { 1124 TRACE_STATE("Stream %d: %s\n", 1125 cur_stream->id, 1126 TCPStateToString(cur_stream)); 1127 } 1128 break; 1129 case TCP_ST_LAST_ACK: 1130 /* control should not come here */ 1131 /* UpdatePassiveReceiveTCPContext() should take care of this */ 1132 break; 1133 case TCP_ST_FIN_WAIT_1: 1134 /* control should not come here */ 1135 /* UpdatePassiveReceiveTCPContext() should take care of this */ 1136 break; 1137 case TCP_ST_FIN_WAIT_2: 1138 /* control should not come here */ 1139 /* UpdatePassiveReceiveTCPContext() should take care of this */ 1140 break; 1141 case TCP_ST_CLOSING: 1142 /* control should not come here */ 1143 /* UpdatePassiveReceiveTCPContext() should take care of this */ 1144 break; 1145 case TCP_ST_TIME_WAIT: 1146 /* control may come here but... */ 1147 /* UpdatePassiveReceiveTCPContext() should take care of this */ 1148 if (tcph->ack) { 1149 TRACE_STATE("Stream %d: %s\n", 1150 cur_stream->id, 1151 TCPStateToString(cur_stream)); 1152 } 1153 break; 1154 case TCP_ST_CLOSED: 1155 case TCP_ST_CLOSED_RSVD: 1156 /* Waiting to be destroyed */ 1157 break; 1158 default: 1159 TRACE_DBG("This should not happen.. Error state: %s reached!\n" 1160 "tcph->syn: %d, tcph->ack: %d\n", 1161 TCPStateToString(cur_stream), pctx->p.tcph->syn, 1162 pctx->p.tcph->ack); 1163 assert(0); 1164 /* This will be enabled once passiverecvcontext is completed */ 1165 /*exit(EXIT_FAILURE);*/ 1166 } 1167 1168 UNUSED(mtcp); 1169 return; 1170 } 1171 /*----------------------------------------------------------------------------*/ 1172 void 1173 PostSendTCPAction(mtcp_manager_t mtcp, struct pkt_ctx *pctx, 1174 struct tcp_stream *recvside_stream, 1175 struct tcp_stream *sendside_stream) 1176 { 1177 uint32_t snd_seq_drift, rcv_seq_drift; 1178 1179 snd_seq_drift = FetchSeqDrift(sendside_stream, pctx->p.seq); 1180 rcv_seq_drift = FetchSeqDrift(recvside_stream, pctx->p.ack_seq); 1181 1182 if (snd_seq_drift != 0 || rcv_seq_drift != 0) { 1183 pctx->p.tcph->seq = htonl(pctx->p.seq + 1184 snd_seq_drift); 1185 pctx->p.tcph->ack_seq = htonl(pctx->p.ack_seq - 1186 rcv_seq_drift); 1187 pctx->p.seq += snd_seq_drift; 1188 pctx->p.ack_seq -= rcv_seq_drift; 1189 1190 /* Recompute checksums */ 1191 pctx->p.iph->check = 0; 1192 pctx->p.iph->check = ip_fast_csum(pctx->p.iph, pctx->p.iph->ihl); 1193 1194 pctx->p.tcph->check = 0; 1195 pctx->p.tcph->check = TCPCalcChecksum((uint16_t *)pctx->p.tcph, 1196 ntohs(pctx->p.iph->tot_len) - (pctx->p.iph->ihl<<2), 1197 pctx->p.iph->saddr, pctx->p.iph->daddr); 1198 } 1199 #if 0 1200 /* This block of code will go away in future revisions */ 1201 /* update sequence no. if seq_drift > 0 */ 1202 if (sendside_stream->sndvar->seq_drift != 0 || 1203 recvside_stream->sndvar->seq_drift != 0) { 1204 1205 pctx->p.tcph->seq = htonl(pctx->p.seq + 1206 sendside_stream->sndvar->seq_drift); 1207 pctx->p.tcph->ack_seq = htonl(pctx->p.ack_seq - 1208 recvside_stream->sndvar->seq_drift); 1209 pctx->p.seq += sendside_stream->sndvar->seq_drift; 1210 pctx->p.ack_seq -= recvside_stream->sndvar->seq_drift; 1211 1212 /* Recompute checksums */ 1213 pctx->p.iph->check = 0; 1214 pctx->p.iph->check = ip_fast_csum(pctx->p.iph, pctx->p.iph->ihl); 1215 1216 pctx->p.tcph->check = 0; 1217 pctx->p.tcph->check = TCPCalcChecksum((uint16_t *)pctx->p.tcph, 1218 ntohs(pctx->p.iph->tot_len) - (pctx->p.iph->ihl<<2), 1219 pctx->p.iph->saddr, pctx->p.iph->daddr); 1220 } 1221 #endif 1222 } 1223 /*----------------------------------------------------------------------------*/ 1224