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