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