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