git.meshlink.io Git - utcp/blob - utcp.c

   1 /*
   2     utcp.c -- Userspace TCP
   3     Copyright (C) 2014 Guus Sliepen <guus@tinc-vpn.org>
   4
   5     This program is free software; you can redistribute it and/or modify
   6     it under the terms of the GNU General Public License as published by
   7     the Free Software Foundation; either version 2 of the License, or
   8     (at your option) any later version.
   9
  10     This program is distributed in the hope that it will be useful,
  11     but WITHOUT ANY WARRANTY; without even the implied warranty of
  12     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13     GNU General Public License for more details.
  14
  15     You should have received a copy of the GNU General Public License along
  16     with this program; if not, write to the Free Software Foundation, Inc.,
  17     51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
  18 */
  19
  20 #define _GNU_SOURCE
  21
  22 #include <assert.h>
  23 #include <errno.h>
  24 #include <stdio.h>
  25 #include <stdlib.h>
  26 #include <stdint.h>
  27 #include <stdbool.h>
  28 #include <string.h>
  29 #include <unistd.h>
  30 #include <sys/time.h>
  31 #include <sys/socket.h>
  32
  33 #include "utcp_priv.h"
  34
  35 #ifndef EBADMSG
  36 #define EBADMSG         104
  37 #endif
  38
  39 #ifndef SHUT_RDWR
  40 #define SHUT_RDWR 2
  41 #endif
  42
  43 #ifdef poll
  44 #undef poll
  45 #endif
  46
  47 #ifndef timersub
  48 #define timersub(a, b, r) do {\
  49         (r)->tv_sec = (a)->tv_sec - (b)->tv_sec;\
  50         (r)->tv_usec = (a)->tv_usec - (b)->tv_usec;\
  51         if((r)->tv_usec < 0)\
  52                 (r)->tv_sec--, (r)->tv_usec += 1000000;\
  53 } while (0)
  54 #endif
  55
  56 #ifdef UTCP_DEBUG
  57 #include <stdarg.h>
  58
  59 static void debug(const char *format, ...) {
  60         va_list ap;
  61         va_start(ap, format);
  62         vfprintf(stderr, format, ap);
  63         va_end(ap);
  64 }
  65
  66 static void print_packet(struct utcp *utcp, const char *dir, const void *pkt, size_t len) {
  67         struct hdr hdr;
  68         if(len < sizeof hdr) {
  69                 debug("%p %s: short packet (%zu bytes)\n", utcp, dir, len);
  70                 return;
  71         }
  72
  73         memcpy(&hdr, pkt, sizeof hdr);
  74         fprintf (stderr, "%p %s: len=%zu, src=%u dst=%u seq=%u ack=%u wnd=%u ctl=", utcp, dir, len, hdr.src, hdr.dst, hdr.seq, hdr.ack, hdr.wnd);
  75         if(hdr.ctl & SYN)
  76                 debug("SYN");
  77         if(hdr.ctl & RST)
  78                 debug("RST");
  79         if(hdr.ctl & FIN)
  80                 debug("FIN");
  81         if(hdr.ctl & ACK)
  82                 debug("ACK");
  83
  84         if(len > sizeof hdr) {
  85                 debug(" data=");
  86                 for(int i = sizeof hdr; i < len; i++) {
  87                         const char *data = pkt;
  88                         debug("%c", data[i] >= 32 ? data[i] : '.');
  89                 }
  90         }
  91
  92         debug("\n");
  93 }
  94 #else
  95 #define debug(...)
  96 #define print_packet(...)
  97 #endif
  98
  99 static void set_state(struct utcp_connection *c, enum state state) {
 100         c->state = state;
 101         if(state == ESTABLISHED)
 102                 timerclear(&c->conn_timeout);
 103         debug("%p new state: %s\n", c->utcp, strstate[state]);
 104 }
 105
 106 static inline void list_connections(struct utcp *utcp) {
 107         debug("%p has %d connections:\n", utcp, utcp->nconnections);
 108         for(int i = 0; i < utcp->nconnections; i++)
 109                 debug("  %u -> %u state %s\n", utcp->connections[i]->src, utcp->connections[i]->dst, strstate[utcp->connections[i]->state]);
 110 }
 111
 112 static int32_t seqdiff(uint32_t a, uint32_t b) {
 113         return a - b;
 114 }
 115
 116 // Buffer functions
 117 // TODO: convert to ringbuffers to avoid memmove() operations.
 118
 119 // Store data into the buffer
 120 static ssize_t buffer_put(struct buffer *buf, const void *data, size_t len) {
 121         if(buf->maxsize <= buf->used)
 122                 return 0;
 123         if(len > buf->maxsize - buf->used)
 124                 len = buf->maxsize - buf->used;
 125         if(len > buf->size - buf->used) {
 126                 size_t newsize = buf->size;
 127                 do {
 128                         newsize *= 2;
 129                 } while(newsize < buf->used + len);
 130                 if(newsize > buf->maxsize)
 131                         newsize = buf->maxsize;
 132                 char *newdata = realloc(buf->data, newsize);
 133                 fprintf(stderr, "%p = realloc(%p, %zu)\n", newdata, buf->data, newsize);
 134                 if(!newdata)
 135                         return -1;
 136                 buf->data = newdata;
 137                 buf->size = newsize;
 138         }
 139         fprintf(stderr, "memcpy(%p, %p, %zu)\n", buf->data + buf->used, data, len);
 140         memcpy(buf->data + buf->used, data, len);
 141         buf->used += len;
 142         return len;
 143 }
 144
 145 // Get data from the buffer. data can be NULL.
 146 static ssize_t buffer_get(struct buffer *buf, void *data, size_t len) {
 147         if(len > buf->used)
 148                 len = buf->used;
 149         if(data)
 150                 memcpy(data, buf->data, len);
 151         if(len < buf->used)
 152                 memmove(buf->data, buf->data + len, buf->used - len);
 153         buf->used -= len;
 154         return len;
 155 }
 156
 157 // Copy data from the buffer without removing it.
 158 static ssize_t buffer_copy(struct buffer *buf, void *data, size_t offset, size_t len) {
 159         if(offset >= buf->used)
 160                 return 0;
 161         if(offset + len > buf->used)
 162                 len = buf->used - offset;
 163         memcpy(data, buf->data + offset, len);
 164         return len;
 165 }
 166
 167 static bool buffer_init(struct buffer *buf, uint32_t len, uint32_t maxlen) {
 168         memset(buf, 0, sizeof *buf);
 169         buf->data = malloc(len);
 170         if(!len)
 171                 return false;
 172         buf->size = len;
 173         buf->maxsize = maxlen;
 174         return true;
 175 }
 176
 177 static void buffer_exit(struct buffer *buf) {
 178         free(buf->data);
 179         memset(buf, 0, sizeof *buf);
 180 }
 181
 182 static uint32_t buffer_free(const struct buffer *buf) {
 183         return buf->maxsize - buf->used;
 184 }
 185
 186 // Connections are stored in a sorted list.
 187 // This gives O(log(N)) lookup time, O(N log(N)) insertion time and O(N) deletion time.
 188
 189 static int compare(const void *va, const void *vb) {
 190         assert(va && vb);
 191
 192         const struct utcp_connection *a = *(struct utcp_connection **)va;
 193         const struct utcp_connection *b = *(struct utcp_connection **)vb;
 194
 195         assert(a && b);
 196         assert(a->src && b->src);
 197
 198         int c = (int)a->src - (int)b->src;
 199         if(c)
 200                 return c;
 201         c = (int)a->dst - (int)b->dst;
 202         return c;
 203 }
 204
 205 static struct utcp_connection *find_connection(const struct utcp *utcp, uint16_t src, uint16_t dst) {
 206         if(!utcp->nconnections)
 207                 return NULL;
 208         struct utcp_connection key = {
 209                 .src = src,
 210                 .dst = dst,
 211         }, *keyp = &key;
 212         struct utcp_connection **match = bsearch(&keyp, utcp->connections, utcp->nconnections, sizeof *utcp->connections, compare);
 213         return match ? *match : NULL;
 214 }
 215
 216 static void free_connection(struct utcp_connection *c) {
 217         struct utcp *utcp = c->utcp;
 218         struct utcp_connection **cp = bsearch(&c, utcp->connections, utcp->nconnections, sizeof *utcp->connections, compare);
 219
 220         assert(cp);
 221
 222         int i = cp - utcp->connections;
 223         memmove(cp, cp + 1, (utcp->nconnections - i - 1) * sizeof *cp);
 224         utcp->nconnections--;
 225
 226         buffer_exit(&c->sndbuf);
 227         free(c);
 228 }
 229
 230 static struct utcp_connection *allocate_connection(struct utcp *utcp, uint16_t src, uint16_t dst) {
 231         // Check whether this combination of src and dst is free
 232
 233         if(src) {
 234                 if(find_connection(utcp, src, dst)) {
 235                         errno = EADDRINUSE;
 236                         return NULL;
 237                 }
 238         } else { // If src == 0, generate a random port number with the high bit set
 239                 if(utcp->nconnections >= 32767) {
 240                         errno = ENOMEM;
 241                         return NULL;
 242                 }
 243                 src = rand() | 0x8000;
 244                 while(find_connection(utcp, src, dst))
 245                         src++;
 246         }
 247
 248         // Allocate memory for the new connection
 249
 250         if(utcp->nconnections >= utcp->nallocated) {
 251                 if(!utcp->nallocated)
 252                         utcp->nallocated = 4;
 253                 else
 254                         utcp->nallocated *= 2;
 255                 struct utcp_connection **new_array = realloc(utcp->connections, utcp->nallocated * sizeof *utcp->connections);
 256                 if(!new_array)
 257                         return NULL;
 258                 utcp->connections = new_array;
 259         }
 260
 261         struct utcp_connection *c = calloc(1, sizeof *c);
 262         if(!c)
 263                 return NULL;
 264
 265         if(!buffer_init(&c->sndbuf, DEFAULT_SNDBUFSIZE, DEFAULT_MAXSNDBUFSIZE)) {
 266                 free(c);
 267                 return NULL;
 268         }
 269
 270         // Fill in the details
 271
 272         c->src = src;
 273         c->dst = dst;
 274         c->snd.iss = rand();
 275         c->snd.una = c->snd.iss;
 276         c->snd.nxt = c->snd.iss + 1;
 277         c->rcv.wnd = utcp->mtu;
 278         c->snd.last = c->snd.nxt;
 279         c->snd.cwnd = utcp->mtu;
 280         c->utcp = utcp;
 281
 282         // Add it to the sorted list of connections
 283
 284         utcp->connections[utcp->nconnections++] = c;
 285         qsort(utcp->connections, utcp->nconnections, sizeof *utcp->connections, compare);
 286
 287         return c;
 288 }
 289
 290 struct utcp_connection *utcp_connect(struct utcp *utcp, uint16_t dst, utcp_recv_t recv, void *priv) {
 291         struct utcp_connection *c = allocate_connection(utcp, 0, dst);
 292         if(!c)
 293                 return NULL;
 294
 295         c->recv = recv;
 296         c->priv = priv;
 297
 298         struct hdr hdr;
 299
 300         hdr.src = c->src;
 301         hdr.dst = c->dst;
 302         hdr.seq = c->snd.iss;
 303         hdr.ack = 0;
 304         hdr.wnd = c->rcv.wnd;
 305         hdr.ctl = SYN;
 306         hdr.aux = 0;
 307
 308         set_state(c, SYN_SENT);
 309
 310         print_packet(utcp, "send", &hdr, sizeof hdr);
 311         utcp->send(utcp, &hdr, sizeof hdr);
 312
 313         gettimeofday(&c->conn_timeout, NULL);
 314         c->conn_timeout.tv_sec += utcp->timeout;
 315
 316         return c;
 317 }
 318
 319 void utcp_accept(struct utcp_connection *c, utcp_recv_t recv, void *priv) {
 320         if(c->reapable || c->state != SYN_RECEIVED) {
 321                 debug("Error: accept() called on invalid connection %p in state %s\n", c, strstate[c->state]);
 322                 return;
 323         }
 324
 325         debug("%p accepted, %p %p\n", c, recv, priv);
 326         c->recv = recv;
 327         c->priv = priv;
 328         set_state(c, ESTABLISHED);
 329 }
 330
 331 static void ack(struct utcp_connection *c, bool sendatleastone) {
 332         int32_t left = seqdiff(c->snd.last, c->snd.nxt);
 333         int32_t cwndleft = c->snd.cwnd - seqdiff(c->snd.nxt, c->snd.una);
 334
 335         assert(left >= 0);
 336
 337         if(cwndleft <= 0)
 338                 cwndleft = 0;
 339
 340         if(cwndleft < left)
 341                 left = cwndleft;
 342
 343         if(!left && !sendatleastone)
 344                 return;
 345
 346         struct {
 347                 struct hdr hdr;
 348                 char data[];
 349         } *pkt;
 350
 351         pkt = malloc(sizeof pkt->hdr + c->utcp->mtu);
 352         if(!pkt->data)
 353                 return;
 354
 355         pkt->hdr.src = c->src;
 356         pkt->hdr.dst = c->dst;
 357         pkt->hdr.ack = c->rcv.nxt;
 358         pkt->hdr.wnd = c->snd.wnd;
 359         pkt->hdr.ctl = ACK;
 360         pkt->hdr.aux = 0;
 361
 362         do {
 363                 uint32_t seglen = left > c->utcp->mtu ? c->utcp->mtu : left;
 364                 pkt->hdr.seq = c->snd.nxt;
 365
 366                 buffer_copy(&c->sndbuf, pkt->data, seqdiff(c->snd.nxt, c->snd.una), seglen);
 367
 368                 c->snd.nxt += seglen;
 369                 left -= seglen;
 370
 371                 if(c->state != ESTABLISHED && !left && seglen) {
 372                         switch(c->state) {
 373                         case FIN_WAIT_1:
 374                         case CLOSING:
 375                                 seglen--;
 376                                 pkt->hdr.ctl |= FIN;
 377                                 break;
 378                         default:
 379                                 break;
 380                         }
 381                 }
 382
 383                 print_packet(c->utcp, "send", pkt, sizeof pkt->hdr + seglen);
 384                 c->utcp->send(c->utcp, pkt, sizeof pkt->hdr + seglen);
 385         } while(left);
 386
 387         free(pkt);
 388 }
 389
 390 ssize_t utcp_send(struct utcp_connection *c, const void *data, size_t len) {
 391         if(c->reapable) {
 392                 debug("Error: send() called on closed connection %p\n", c);
 393                 errno = EBADF;
 394                 return -1;
 395         }
 396
 397         switch(c->state) {
 398         case CLOSED:
 399         case LISTEN:
 400         case SYN_SENT:
 401         case SYN_RECEIVED:
 402                 debug("Error: send() called on unconnected connection %p\n", c);
 403                 errno = ENOTCONN;
 404                 return -1;
 405         case ESTABLISHED:
 406         case CLOSE_WAIT:
 407                 break;
 408         case FIN_WAIT_1:
 409         case FIN_WAIT_2:
 410         case CLOSING:
 411         case LAST_ACK:
 412         case TIME_WAIT:
 413                 debug("Error: send() called on closing connection %p\n", c);
 414                 errno = EPIPE;
 415                 return -1;
 416         }
 417
 418         // Add data to send buffer
 419
 420         if(!len)
 421                 return 0;
 422
 423         if(!data) {
 424                 errno = EFAULT;
 425                 return -1;
 426         }
 427
 428         len = buffer_put(&c->sndbuf, data, len);
 429         if(len <= 0) {
 430                 errno = EWOULDBLOCK;
 431                 return 0;
 432         }
 433
 434         c->snd.last += len;
 435         ack(c, false);
 436         return len;
 437 }
 438
 439 static void swap_ports(struct hdr *hdr) {
 440         uint16_t tmp = hdr->src;
 441         hdr->src = hdr->dst;
 442         hdr->dst = tmp;
 443 }
 444
 445 ssize_t utcp_recv(struct utcp *utcp, const void *data, size_t len) {
 446         if(!utcp) {
 447                 errno = EFAULT;
 448                 return -1;
 449         }
 450
 451         if(!len)
 452                 return 0;
 453
 454         if(!data) {
 455                 errno = EFAULT;
 456                 return -1;
 457         }
 458
 459         print_packet(utcp, "recv", data, len);
 460
 461         // Drop packets smaller than the header
 462
 463         struct hdr hdr;
 464         if(len < sizeof hdr) {
 465                 errno = EBADMSG;
 466                 return -1;
 467         }
 468
 469         // Make a copy from the potentially unaligned data to a struct hdr
 470
 471         memcpy(&hdr, data, sizeof hdr);
 472         data += sizeof hdr;
 473         len -= sizeof hdr;
 474
 475         // Drop packets with an unknown CTL flag
 476
 477         if(hdr.ctl & ~(SYN | ACK | RST | FIN)) {
 478                 errno = EBADMSG;
 479                 return -1;
 480         }
 481
 482         // Try to match the packet to an existing connection
 483
 484         struct utcp_connection *c = find_connection(utcp, hdr.dst, hdr.src);
 485
 486         // Is it for a new connection?
 487
 488         if(!c) {
 489                 // Ignore RST packets
 490
 491                 if(hdr.ctl & RST)
 492                         return 0;
 493
 494                 // Is it a SYN packet and are we LISTENing?
 495
 496                 if(hdr.ctl & SYN && !(hdr.ctl & ACK) && utcp->accept) {
 497                         // If we don't want to accept it, send a RST back
 498                         if((utcp->pre_accept && !utcp->pre_accept(utcp, hdr.dst))) {
 499                                 len = 1;
 500                                 goto reset;
 501                         }
 502
 503                         // Try to allocate memory, otherwise send a RST back
 504                         c = allocate_connection(utcp, hdr.dst, hdr.src);
 505                         if(!c) {
 506                                 len = 1;
 507                                 goto reset;
 508                         }
 509
 510                         // Return SYN+ACK, go to SYN_RECEIVED state
 511                         c->snd.wnd = hdr.wnd;
 512                         c->rcv.irs = hdr.seq;
 513                         c->rcv.nxt = c->rcv.irs + 1;
 514                         set_state(c, SYN_RECEIVED);
 515
 516                         hdr.dst = c->dst;
 517                         hdr.src = c->src;
 518                         hdr.ack = c->rcv.irs + 1;
 519                         hdr.seq = c->snd.iss;
 520                         hdr.ctl = SYN | ACK;
 521                         print_packet(c->utcp, "send", &hdr, sizeof hdr);
 522                         utcp->send(utcp, &hdr, sizeof hdr);
 523                 } else {
 524                         // No, we don't want your packets, send a RST back
 525                         len = 1;
 526                         goto reset;
 527                 }
 528
 529                 return 0;
 530         }
 531
 532         debug("%p state %s\n", c->utcp, strstate[c->state]);
 533
 534         // In case this is for a CLOSED connection, ignore the packet.
 535         // TODO: make it so incoming packets can never match a CLOSED connection.
 536
 537         if(c->state == CLOSED)
 538                 return 0;
 539
 540         // It is for an existing connection.
 541
 542         uint32_t prevrcvnxt = c->rcv.nxt;
 543
 544         // 1. Drop invalid packets.
 545
 546         // 1a. Drop packets that should not happen in our current state.
 547
 548         switch(c->state) {
 549         case SYN_SENT:
 550         case SYN_RECEIVED:
 551         case ESTABLISHED:
 552         case FIN_WAIT_1:
 553         case FIN_WAIT_2:
 554         case CLOSE_WAIT:
 555         case CLOSING:
 556         case LAST_ACK:
 557         case TIME_WAIT:
 558                 break;
 559         default:
 560                 abort();
 561         }
 562
 563         // 1b. Drop packets with a sequence number not in our receive window.
 564
 565         bool acceptable;
 566
 567         if(c->state == SYN_SENT)
 568                 acceptable = true;
 569
 570         // TODO: handle packets overlapping c->rcv.nxt.
 571 #if 0
 572         // Only use this when accepting out-of-order packets.
 573         else if(len == 0)
 574                 if(c->rcv.wnd == 0)
 575                         acceptable = hdr.seq == c->rcv.nxt;
 576                 else
 577                         acceptable = (seqdiff(hdr.seq, c->rcv.nxt) >= 0 && seqdiff(hdr.seq, c->rcv.nxt + c->rcv.wnd) < 0);
 578         else
 579                 if(c->rcv.wnd == 0)
 580                         // We don't accept data when the receive window is zero.
 581                         acceptable = false;
 582                 else
 583                         // Both start and end of packet must be within the receive window
 584                         acceptable = (seqdiff(hdr.seq, c->rcv.nxt) >= 0 && seqdiff(hdr.seq, c->rcv.nxt + c->rcv.wnd) < 0)
 585                                 || (seqdiff(hdr.seq + len + 1, c->rcv.nxt) >= 0 && seqdiff(hdr.seq + len - 1, c->rcv.nxt + c->rcv.wnd) < 0);
 586 #else
 587         if(c->state != SYN_SENT)
 588                 acceptable = hdr.seq == c->rcv.nxt;
 589 #endif
 590
 591         if(!acceptable) {
 592                 debug("Packet not acceptable, %u  <= %u + %zu < %u\n", c->rcv.nxt, hdr.seq, len, c->rcv.nxt + c->rcv.wnd);
 593                 // Ignore unacceptable RST packets.
 594                 if(hdr.ctl & RST)
 595                         return 0;
 596                 // Otherwise, send an ACK back in the hope things improve.
 597                 goto ack;
 598         }
 599
 600         c->snd.wnd = hdr.wnd; // TODO: move below
 601
 602         // 1c. Drop packets with an invalid ACK.
 603         // ackno should not roll back, and it should also not be bigger than snd.nxt.
 604
 605         if(hdr.ctl & ACK && (seqdiff(hdr.ack, c->snd.nxt) > 0 || seqdiff(hdr.ack, c->snd.una) < 0)) {
 606                 debug("Packet ack seqno out of range, %u %u %u\n", hdr.ack, c->snd.una, c->snd.nxt);
 607                 // Ignore unacceptable RST packets.
 608                 if(hdr.ctl & RST)
 609                         return 0;
 610                 goto reset;
 611         }
 612
 613         // 2. Handle RST packets
 614
 615         if(hdr.ctl & RST) {
 616                 switch(c->state) {
 617                 case SYN_SENT:
 618                         if(!(hdr.ctl & ACK))
 619                                 return 0;
 620                         // The peer has refused our connection.
 621                         set_state(c, CLOSED);
 622                         errno = ECONNREFUSED;
 623                         if(c->recv)
 624                                 c->recv(c, NULL, 0);
 625                         return 0;
 626                 case SYN_RECEIVED:
 627                         if(hdr.ctl & ACK)
 628                                 return 0;
 629                         // We haven't told the application about this connection yet. Silently delete.
 630                         free_connection(c);
 631                         return 0;
 632                 case ESTABLISHED:
 633                 case FIN_WAIT_1:
 634                 case FIN_WAIT_2:
 635                 case CLOSE_WAIT:
 636                         if(hdr.ctl & ACK)
 637                                 return 0;
 638                         // The peer has aborted our connection.
 639                         set_state(c, CLOSED);
 640                         errno = ECONNRESET;
 641                         if(c->recv)
 642                                 c->recv(c, NULL, 0);
 643                         return 0;
 644                 case CLOSING:
 645                 case LAST_ACK:
 646                 case TIME_WAIT:
 647                         if(hdr.ctl & ACK)
 648                                 return 0;
 649                         // As far as the application is concerned, the connection has already been closed.
 650                         // If it has called utcp_close() already, we can immediately free this connection.
 651                         if(c->reapable) {
 652                                 free_connection(c);
 653                                 return 0;
 654                         }
 655                         // Otherwise, immediately move to the CLOSED state.
 656                         set_state(c, CLOSED);
 657                         return 0;
 658                 default:
 659                         abort();
 660                 }
 661         }
 662
 663         // 3. Advance snd.una
 664
 665         uint32_t advanced = seqdiff(hdr.ack, c->snd.una);
 666         prevrcvnxt = c->rcv.nxt;
 667
 668         if(advanced) {
 669                 int32_t data_acked = advanced;
 670
 671                 switch(c->state) {
 672                         case SYN_SENT:
 673                         case SYN_RECEIVED:
 674                                 data_acked--;
 675                                 break;
 676                         // TODO: handle FIN as well.
 677                         default:
 678                                 break;
 679                 }
 680
 681                 assert(data_acked >= 0);
 682
 683                 int32_t bufused = seqdiff(c->snd.last, c->snd.una);
 684                 assert(data_acked <= bufused);
 685
 686                 if(data_acked)
 687                         buffer_get(&c->sndbuf, NULL, data_acked);
 688
 689                 c->snd.una = hdr.ack;
 690
 691                 c->dupack = 0;
 692                 c->snd.cwnd += utcp->mtu;
 693                 if(c->snd.cwnd > c->sndbuf.maxsize)
 694                         c->snd.cwnd = c->sndbuf.maxsize;
 695
 696                 // Check if we have sent a FIN that is now ACKed.
 697                 switch(c->state) {
 698                 case FIN_WAIT_1:
 699                         if(c->snd.una == c->snd.last)
 700                                 set_state(c, FIN_WAIT_2);
 701                         break;
 702                 case CLOSING:
 703                         if(c->snd.una == c->snd.last) {
 704                                 gettimeofday(&c->conn_timeout, NULL);
 705                                 c->conn_timeout.tv_sec += 60;
 706                                 set_state(c, TIME_WAIT);
 707                         }
 708                         break;
 709                 default:
 710                         break;
 711                 }
 712         } else {
 713                 if(!len) {
 714                         c->dupack++;
 715                         if(c->dupack >= 3) {
 716                                 debug("Triplicate ACK\n");
 717                                 //TODO: Resend one packet and go to fast recovery mode. See RFC 6582.
 718                                 //abort();
 719                         }
 720                 }
 721         }
 722
 723         // 4. Update timers
 724
 725         if(advanced) {
 726                 timerclear(&c->conn_timeout); // It will be set anew in utcp_timeout() if c->snd.una != c->snd.nxt.
 727                 if(c->snd.una == c->snd.nxt)
 728                         timerclear(&c->rtrx_timeout);
 729         }
 730
 731         // 5. Process SYN stuff
 732
 733         if(hdr.ctl & SYN) {
 734                 switch(c->state) {
 735                 case SYN_SENT:
 736                         // This is a SYNACK. It should always have ACKed the SYN.
 737                         if(!advanced)
 738                                 goto reset;
 739                         c->rcv.irs = hdr.seq;
 740                         c->rcv.nxt = hdr.seq;
 741                         set_state(c, ESTABLISHED);
 742                         // TODO: notify application of this somehow.
 743                         break;
 744                 case SYN_RECEIVED:
 745                 case ESTABLISHED:
 746                 case FIN_WAIT_1:
 747                 case FIN_WAIT_2:
 748                 case CLOSE_WAIT:
 749                 case CLOSING:
 750                 case LAST_ACK:
 751                 case TIME_WAIT:
 752                         // Ehm, no. We should never receive a second SYN.
 753                         goto reset;
 754                 default:
 755                         abort();
 756                 }
 757
 758                 // SYN counts as one sequence number
 759                 c->rcv.nxt++;
 760         }
 761
 762         // 6. Process new data
 763
 764         if(c->state == SYN_RECEIVED) {
 765                 // This is the ACK after the SYNACK. It should always have ACKed the SYNACK.
 766                 if(!advanced)
 767                         goto reset;
 768
 769                 // Are we still LISTENing?
 770                 if(utcp->accept)
 771                         utcp->accept(c, c->src);
 772
 773                 if(c->state != ESTABLISHED) {
 774                         set_state(c, CLOSED);
 775                         c->reapable = true;
 776                         goto reset;
 777                 }
 778         }
 779
 780         if(len) {
 781                 switch(c->state) {
 782                 case SYN_SENT:
 783                 case SYN_RECEIVED:
 784                         // This should never happen.
 785                         abort();
 786                 case ESTABLISHED:
 787                 case FIN_WAIT_1:
 788                 case FIN_WAIT_2:
 789                         break;
 790                 case CLOSE_WAIT:
 791                 case CLOSING:
 792                 case LAST_ACK:
 793                 case TIME_WAIT:
 794                         // Ehm no, We should never receive more data after a FIN.
 795                         goto reset;
 796                 default:
 797                         abort();
 798                 }
 799
 800                 ssize_t rxd;
 801
 802                 if(c->recv) {
 803                         rxd = c->recv(c, data, len);
 804                         if(rxd != len) {
 805                                 // TODO: once we have a receive buffer, handle the application not accepting all data.
 806                                 fprintf(stderr, "c->recv(%p, %p, %zu) returned %zd\n", c, data, len, rxd);
 807                                 abort();
 808                         }
 809                         if(rxd < 0)
 810                                 rxd = 0;
 811                         else if(rxd > len)
 812                                 rxd = len; // Bad application, bad!
 813                 } else {
 814                         rxd = len;
 815                 }
 816
 817                 c->rcv.nxt += len;
 818         }
 819
 820         // 7. Process FIN stuff
 821
 822         if(hdr.ctl & FIN) {
 823                 switch(c->state) {
 824                 case SYN_SENT:
 825                 case SYN_RECEIVED:
 826                         // This should never happen.
 827                         abort();
 828                 case ESTABLISHED:
 829                         set_state(c, CLOSE_WAIT);
 830                         break;
 831                 case FIN_WAIT_1:
 832                         set_state(c, CLOSING);
 833                         break;
 834                 case FIN_WAIT_2:
 835                         gettimeofday(&c->conn_timeout, NULL);
 836                         c->conn_timeout.tv_sec += 60;
 837                         set_state(c, TIME_WAIT);
 838                         break;
 839                 case CLOSE_WAIT:
 840                 case CLOSING:
 841                 case LAST_ACK:
 842                 case TIME_WAIT:
 843                         // Ehm, no. We should never receive a second FIN.
 844                         goto reset;
 845                 default:
 846                         abort();
 847                 }
 848
 849                 // FIN counts as one sequence number
 850                 c->rcv.nxt++;
 851                 len++;
 852
 853                 // Inform the application that the peer closed the connection.
 854                 if(c->recv) {
 855                         errno = 0;
 856                         c->recv(c, NULL, 0);
 857                 }
 858         }
 859
 860         // Now we send something back if:
 861         // - we advanced rcv.nxt (ie, we got some data that needs to be ACKed)
 862         //   -> sendatleastone = true
 863         // - or we got an ack, so we should maybe send a bit more data
 864         //   -> sendatleastone = false
 865
 866 ack:
 867         ack(c, prevrcvnxt != c->rcv.nxt);
 868         return 0;
 869
 870 reset:
 871         swap_ports(&hdr);
 872         hdr.wnd = 0;
 873         if(hdr.ctl & ACK) {
 874                 hdr.seq = hdr.ack;
 875                 hdr.ctl = RST;
 876         } else {
 877                 hdr.ack = hdr.seq + len;
 878                 hdr.seq = 0;
 879                 hdr.ctl = RST | ACK;
 880         }
 881         print_packet(utcp, "send", &hdr, sizeof hdr);
 882         utcp->send(utcp, &hdr, sizeof hdr);
 883         return 0;
 884
 885 }
 886
 887 int utcp_shutdown(struct utcp_connection *c, int dir) {
 888         debug("%p shutdown %d\n", c ? c->utcp : NULL, dir);
 889         if(!c) {
 890                 errno = EFAULT;
 891                 return -1;
 892         }
 893
 894         if(c->reapable) {
 895                 debug("Error: shutdown() called on closed connection %p\n", c);
 896                 errno = EBADF;
 897                 return -1;
 898         }
 899
 900         // TODO: handle dir
 901
 902         switch(c->state) {
 903         case CLOSED:
 904                 return 0;
 905         case LISTEN:
 906         case SYN_SENT:
 907                 set_state(c, CLOSED);
 908                 return 0;
 909
 910         case SYN_RECEIVED:
 911         case ESTABLISHED:
 912                 set_state(c, FIN_WAIT_1);
 913                 break;
 914         case FIN_WAIT_1:
 915         case FIN_WAIT_2:
 916                 return 0;
 917         case CLOSE_WAIT:
 918                 set_state(c, CLOSING);
 919                 break;
 920
 921         case CLOSING:
 922         case LAST_ACK:
 923         case TIME_WAIT:
 924                 return 0;
 925         }
 926
 927         c->snd.last++;
 928
 929         ack(c, false);
 930         return 0;
 931 }
 932
 933 int utcp_close(struct utcp_connection *c) {
 934         if(utcp_shutdown(c, SHUT_RDWR))
 935                 return -1;
 936         c->reapable = true;
 937         return 0;
 938 }
 939
 940 int utcp_abort(struct utcp_connection *c) {
 941         if(!c) {
 942                 errno = EFAULT;
 943                 return -1;
 944         }
 945
 946         if(c->reapable) {
 947                 debug("Error: abort() called on closed connection %p\n", c);
 948                 errno = EBADF;
 949                 return -1;
 950         }
 951
 952         c->reapable = true;
 953
 954         switch(c->state) {
 955         case CLOSED:
 956                 return 0;
 957         case LISTEN:
 958         case SYN_SENT:
 959         case CLOSING:
 960         case LAST_ACK:
 961         case TIME_WAIT:
 962                 set_state(c, CLOSED);
 963                 return 0;
 964
 965         case SYN_RECEIVED:
 966         case ESTABLISHED:
 967         case FIN_WAIT_1:
 968         case FIN_WAIT_2:
 969         case CLOSE_WAIT:
 970                 set_state(c, CLOSED);
 971                 break;
 972         }
 973
 974         // Send RST
 975
 976         struct hdr hdr;
 977
 978         hdr.src = c->src;
 979         hdr.dst = c->dst;
 980         hdr.seq = c->snd.nxt;
 981         hdr.ack = 0;
 982         hdr.wnd = 0;
 983         hdr.ctl = RST;
 984
 985         print_packet(c->utcp, "send", &hdr, sizeof hdr);
 986         c->utcp->send(c->utcp, &hdr, sizeof hdr);
 987         return 0;
 988 }
 989
 990 static void retransmit(struct utcp_connection *c) {
 991         if(c->state == CLOSED || c->snd.nxt == c->snd.una)
 992                 return;
 993
 994         struct utcp *utcp = c->utcp;
 995
 996         struct {
 997                 struct hdr hdr;
 998                 char data[];
 999         } *pkt;
1000
1001         pkt = malloc(sizeof pkt->hdr + c->utcp->mtu);
1002         if(!pkt)
1003                 return;
1004
1005         pkt->hdr.src = c->src;
1006         pkt->hdr.dst = c->dst;
1007
1008         switch(c->state) {
1009                 case LISTEN:
1010                         // TODO: this should not happen
1011                         break;
1012
1013                 case SYN_SENT:
1014                         pkt->hdr.seq = c->snd.iss;
1015                         pkt->hdr.ack = 0;
1016                         pkt->hdr.wnd = c->rcv.wnd;
1017                         pkt->hdr.ctl = SYN;
1018                         print_packet(c->utcp, "rtrx", pkt, sizeof pkt->hdr);
1019                         utcp->send(utcp, pkt, sizeof pkt->hdr);
1020                         break;
1021
1022                 case SYN_RECEIVED:
1023                         pkt->hdr.seq = c->snd.nxt;
1024                         pkt->hdr.ack = c->rcv.nxt;
1025                         pkt->hdr.ctl = SYN | ACK;
1026                         print_packet(c->utcp, "rtrx", pkt, sizeof pkt->hdr);
1027                         utcp->send(utcp, pkt, sizeof pkt->hdr);
1028                         break;
1029
1030                 case ESTABLISHED:
1031                 case FIN_WAIT_1:
1032                         pkt->hdr.seq = c->snd.una;
1033                         pkt->hdr.ack = c->rcv.nxt;
1034                         pkt->hdr.ctl = ACK;
1035                         uint32_t len = seqdiff(c->snd.nxt, c->snd.una);
1036                         if(c->state == FIN_WAIT_1)
1037                                 len--;
1038                         if(len > utcp->mtu)
1039                                 len = utcp->mtu;
1040                         else {
1041                                 if(c->state == FIN_WAIT_1)
1042                                         pkt->hdr.ctl |= FIN;
1043                         }
1044                         buffer_copy(&c->sndbuf, pkt->data, 0, len);
1045                         print_packet(c->utcp, "rtrx", pkt, sizeof pkt->hdr + len);
1046                         utcp->send(utcp, pkt, sizeof pkt->hdr + len);
1047                         break;
1048
1049                 default:
1050                         // TODO: implement
1051                         abort();
1052         }
1053
1054         free(pkt);
1055 }
1056
1057 /* Handle timeouts.
1058  * One call to this function will loop through all connections,
1059  * checking if something needs to be resent or not.
1060  * The return value is the time to the next timeout in milliseconds,
1061  * or maybe a negative value if the timeout is infinite.
1062  */
1063 int utcp_timeout(struct utcp *utcp) {
1064         struct timeval now;
1065         gettimeofday(&now, NULL);
1066         struct timeval next = {now.tv_sec + 3600, now.tv_usec};
1067
1068         for(int i = 0; i < utcp->nconnections; i++) {
1069                 struct utcp_connection *c = utcp->connections[i];
1070                 if(!c)
1071                         continue;
1072
1073                 if(c->state == CLOSED) {
1074                         if(c->reapable) {
1075                                 debug("Reaping %p\n", c);
1076                                 free_connection(c);
1077                                 i--;
1078                         }
1079                         continue;
1080                 }
1081
1082                 if(timerisset(&c->conn_timeout) && timercmp(&c->conn_timeout, &now, <)) {
1083                         errno = ETIMEDOUT;
1084                         c->state = CLOSED;
1085                         if(c->recv)
1086                                 c->recv(c, NULL, 0);
1087                         continue;
1088                 }
1089
1090                 if(timerisset(&c->rtrx_timeout) && timercmp(&c->rtrx_timeout, &now, <)) {
1091                         retransmit(c);
1092                 }
1093
1094                 if(c->poll && buffer_free(&c->sndbuf) && (c->state == ESTABLISHED || c->state == CLOSE_WAIT))
1095                         c->poll(c, buffer_free(&c->sndbuf));
1096
1097                 if(timerisset(&c->conn_timeout) && timercmp(&c->conn_timeout, &next, <))
1098                         next = c->conn_timeout;
1099
1100                 if(c->snd.nxt != c->snd.una) {
1101                         c->rtrx_timeout = now;
1102                         c->rtrx_timeout.tv_sec++;
1103                 } else {
1104                         timerclear(&c->rtrx_timeout);
1105                 }
1106
1107                 if(timerisset(&c->rtrx_timeout) && timercmp(&c->rtrx_timeout, &next, <))
1108                         next = c->rtrx_timeout;
1109         }
1110
1111         struct timeval diff;
1112         timersub(&next, &now, &diff);
1113         if(diff.tv_sec < 0)
1114                 return 0;
1115         return diff.tv_sec * 1000 + diff.tv_usec / 1000;
1116 }
1117
1118 struct utcp *utcp_init(utcp_accept_t accept, utcp_pre_accept_t pre_accept, utcp_send_t send, void *priv) {
1119         struct utcp *utcp = calloc(1, sizeof *utcp);
1120         if(!utcp)
1121                 return NULL;
1122
1123         if(!send) {
1124                 errno = EFAULT;
1125                 return NULL;
1126         }
1127
1128         utcp->accept = accept;
1129         utcp->pre_accept = pre_accept;
1130         utcp->send = send;
1131         utcp->priv = priv;
1132         utcp->mtu = 1000;
1133         utcp->timeout = 60;
1134
1135         return utcp;
1136 }
1137
1138 void utcp_exit(struct utcp *utcp) {
1139         if(!utcp)
1140                 return;
1141         for(int i = 0; i < utcp->nconnections; i++) {
1142                 if(!utcp->connections[i]->reapable)
1143                         debug("Warning, freeing unclosed connection %p\n", utcp->connections[i]);
1144                 buffer_exit(&utcp->connections[i]->sndbuf);
1145                 free(utcp->connections[i]);
1146         }
1147         free(utcp->connections);
1148         free(utcp);
1149 }
1150
1151 uint16_t utcp_get_mtu(struct utcp *utcp) {
1152         return utcp->mtu;
1153 }
1154
1155 void utcp_set_mtu(struct utcp *utcp, uint16_t mtu) {
1156         // TODO: handle overhead of the header
1157         utcp->mtu = mtu;
1158 }
1159
1160 int utcp_get_user_timeout(struct utcp *u) {
1161         return u->timeout;
1162 }
1163
1164 void utcp_set_user_timeout(struct utcp *u, int timeout) {
1165         u->timeout = timeout;
1166 }
1167
1168 size_t utcp_get_sndbuf(struct utcp_connection *c) {
1169         return c->sndbuf.maxsize;
1170 }
1171
1172 size_t utcp_get_sndbuf_free(struct utcp_connection *c) {
1173         return buffer_free(&c->sndbuf);
1174 }
1175
1176 void utcp_set_sndbuf(struct utcp_connection *c, size_t size) {
1177         c->sndbuf.maxsize = size;
1178         if(c->sndbuf.maxsize != size)
1179                 c->sndbuf.maxsize = -1;
1180 }
1181
1182 bool utcp_get_nodelay(struct utcp_connection *c) {
1183         return c->nodelay;
1184 }
1185
1186 void utcp_set_nodelay(struct utcp_connection *c, bool nodelay) {
1187         c->nodelay = nodelay;
1188 }
1189
1190 bool utcp_get_keepalive(struct utcp_connection *c) {
1191         return c->keepalive;
1192 }
1193
1194 void utcp_set_keepalive(struct utcp_connection *c, bool keepalive) {
1195         c->keepalive = keepalive;
1196 }
1197
1198 size_t utcp_get_outq(struct utcp_connection *c) {
1199         return seqdiff(c->snd.nxt, c->snd.una);
1200 }
1201
1202 void utcp_set_recv_cb(struct utcp_connection *c, utcp_recv_t recv) {
1203         c->recv = recv;
1204 }
1205
1206 void utcp_set_poll_cb(struct utcp_connection *c, utcp_poll_t poll) {
1207         c->poll = poll;
1208 }