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
 206         struct hdr hdr;
 207
 208         hdr.src = c->src;
 209         hdr.dst = c->dst;
 210         hdr.seq = c->snd.iss;
 211         hdr.ack = 0;
 212         hdr.wnd = c->rcv.wnd;
 213         hdr.ctl = SYN;
 214         hdr.aux = 0;
 215
 216         set_state(c, SYN_SENT);
 217
 218         print_packet(utcp, "send", &hdr, sizeof hdr);
 219         utcp->send(utcp, &hdr, sizeof hdr);
 220
 221         gettimeofday(&c->conn_timeout, NULL);
 222         c->conn_timeout.tv_sec += utcp->timeout;
 223
 224         return c;
 225 }
 226
 227 void utcp_accept(struct utcp_connection *c, utcp_recv_t recv, void *priv) {
 228         if(c->reapable || c->state != SYN_RECEIVED) {
 229                 debug("Error: accept() called on invalid connection %p in state %s\n", c, strstate[c->state]);
 230                 return;
 231         }
 232
 233         debug("%p accepted, %p %p\n", c, recv, priv);
 234         c->recv = recv;
 235         c->priv = priv;
 236         set_state(c, ESTABLISHED);
 237 }
 238
 239 static void ack(struct utcp_connection *c, bool sendatleastone) {
 240         int32_t left = seqdiff(c->snd.last, c->snd.nxt);
 241         int32_t cwndleft = c->snd.cwnd - seqdiff(c->snd.nxt, c->snd.una);
 242         char *data = c->sndbuf + seqdiff(c->snd.nxt, c->snd.una);
 243
 244         assert(left >= 0);
 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 ssize_t 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         uint32_t prevrcvnxt = c->rcv.nxt;
 600
 601         if(advanced) {
 602                 int32_t data_acked = advanced;
 603
 604                 switch(c->state) {
 605                         case SYN_SENT:
 606                         case SYN_RECEIVED:
 607                                 data_acked--;
 608                                 break;
 609                         // TODO: handle FIN as well.
 610                         default:
 611                                 break;
 612                 }
 613
 614                 assert(data_acked >= 0);
 615
 616                 int32_t bufused = seqdiff(c->snd.last, c->snd.una);
 617                 assert(data_acked <= bufused);
 618
 619                 // Make room in the send buffer.
 620                 // TODO: try to avoid memmoving too much. Circular buffer?
 621                 uint32_t left = bufused - data_acked;
 622                 if(data_acked && left)
 623                         memmove(c->sndbuf, c->sndbuf + data_acked, left);
 624
 625                 c->snd.una = hdr.ack;
 626
 627                 c->dupack = 0;
 628                 c->snd.cwnd += utcp->mtu;
 629                 if(c->snd.cwnd > c->maxsndbufsize)
 630                         c->snd.cwnd = c->maxsndbufsize;
 631
 632                 // Check if we have sent a FIN that is now ACKed.
 633                 switch(c->state) {
 634                 case FIN_WAIT_1:
 635                         if(c->snd.una == c->snd.last)
 636                                 set_state(c, FIN_WAIT_2);
 637                         break;
 638                 case CLOSING:
 639                         if(c->snd.una == c->snd.last) {
 640                                 gettimeofday(&c->conn_timeout, NULL);
 641                                 c->conn_timeout.tv_sec += 60;
 642                                 set_state(c, TIME_WAIT);
 643                         }
 644                         break;
 645                 default:
 646                         break;
 647                 }
 648         } else {
 649                 if(!len) {
 650                         c->dupack++;
 651                         if(c->dupack >= 3) {
 652                                 debug("Triplicate ACK\n");
 653                                 //TODO: Resend one packet and go to fast recovery mode. See RFC 6582.
 654                                 //abort();
 655                         }
 656                 }
 657         }
 658
 659         // 4. Update timers
 660
 661         if(advanced) {
 662                 timerclear(&c->conn_timeout); // It will be set anew in utcp_timeout() if c->snd.una != c->snd.nxt.
 663                 if(c->snd.una == c->snd.nxt)
 664                         timerclear(&c->rtrx_timeout);
 665         }
 666
 667         // 5. Process SYN stuff
 668
 669         if(hdr.ctl & SYN) {
 670                 switch(c->state) {
 671                 case SYN_SENT:
 672                         // This is a SYNACK. It should always have ACKed the SYN.
 673                         if(!advanced)
 674                                 goto reset;
 675                         c->rcv.irs = hdr.seq;
 676                         c->rcv.nxt = hdr.seq;
 677                         set_state(c, ESTABLISHED);
 678                         // TODO: notify application of this somehow.
 679                         break;
 680                 case SYN_RECEIVED:
 681                 case ESTABLISHED:
 682                 case FIN_WAIT_1:
 683                 case FIN_WAIT_2:
 684                 case CLOSE_WAIT:
 685                 case CLOSING:
 686                 case LAST_ACK:
 687                 case TIME_WAIT:
 688                         // Ehm, no. We should never receive a second SYN.
 689                         goto reset;
 690                 default:
 691                         abort();
 692                 }
 693
 694                 // SYN counts as one sequence number
 695                 c->rcv.nxt++;
 696         }
 697
 698         // 6. Process new data
 699
 700         if(c->state == SYN_RECEIVED) {
 701                 // This is the ACK after the SYNACK. It should always have ACKed the SYNACK.
 702                 if(!advanced)
 703                         goto reset;
 704
 705                 // Are we still LISTENing?
 706                 if(utcp->accept)
 707                         utcp->accept(c, c->src);
 708
 709                 if(c->state != ESTABLISHED) {
 710                         set_state(c, CLOSED);
 711                         c->reapable = true;
 712                         goto reset;
 713                 }
 714         }
 715
 716         if(len) {
 717                 switch(c->state) {
 718                 case SYN_SENT:
 719                 case SYN_RECEIVED:
 720                         // This should never happen.
 721                         abort();
 722                 case ESTABLISHED:
 723                 case FIN_WAIT_1:
 724                 case FIN_WAIT_2:
 725                         break;
 726                 case CLOSE_WAIT:
 727                 case CLOSING:
 728                 case LAST_ACK:
 729                 case TIME_WAIT:
 730                         // Ehm no, We should never receive more data after a FIN.
 731                         goto reset;
 732                 default:
 733                         abort();
 734                 }
 735
 736                 ssize_t rxd;
 737
 738                 if(c->recv) {
 739                         rxd = c->recv(c, data, len);
 740                         if(rxd != len) {
 741                                 // TODO: once we have a receive buffer, handle the application not accepting all data.
 742                                 fprintf(stderr, "c->recv(%p, %p, %zu) returned %zd\n", c, data, len, rxd);
 743                                 abort();
 744                         }
 745                         if(rxd < 0)
 746                                 rxd = 0;
 747                         else if(rxd > len)
 748                                 rxd = len; // Bad application, bad!
 749                 } else {
 750                         rxd = len;
 751                 }
 752
 753                 c->rcv.nxt += len;
 754         }
 755
 756         // 7. Process FIN stuff
 757
 758         if(hdr.ctl & FIN) {
 759                 switch(c->state) {
 760                 case SYN_SENT:
 761                 case SYN_RECEIVED:
 762                         // This should never happen.
 763                         abort();
 764                 case ESTABLISHED:
 765                         set_state(c, CLOSE_WAIT);
 766                         break;
 767                 case FIN_WAIT_1:
 768                         set_state(c, CLOSING);
 769                         break;
 770                 case FIN_WAIT_2:
 771                         gettimeofday(&c->conn_timeout, NULL);
 772                         c->conn_timeout.tv_sec += 60;
 773                         set_state(c, TIME_WAIT);
 774                         break;
 775                 case CLOSE_WAIT:
 776                 case CLOSING:
 777                 case LAST_ACK:
 778                 case TIME_WAIT:
 779                         // Ehm, no. We should never receive a second FIN.
 780                         goto reset;
 781                 default:
 782                         abort();
 783                 }
 784
 785                 // FIN counts as one sequence number
 786                 c->rcv.nxt++;
 787                 len++;
 788
 789                 // Inform the application that the peer closed the connection.
 790                 if(c->recv) {
 791                         errno = 0;
 792                         c->recv(c, NULL, 0);
 793                 }
 794         }
 795
 796         // Now we send something back if:
 797         // - we advanced rcv.nxt (ie, we got some data that needs to be ACKed)
 798         //   -> sendatleastone = true
 799         // - or we got an ack, so we should maybe send a bit more data
 800         //   -> sendatleastone = false
 801
 802 ack:
 803         ack(c, prevrcvnxt != c->rcv.nxt);
 804         return 0;
 805
 806 reset:
 807         swap_ports(&hdr);
 808         hdr.wnd = 0;
 809         if(hdr.ctl & ACK) {
 810                 hdr.seq = hdr.ack;
 811                 hdr.ctl = RST;
 812         } else {
 813                 hdr.ack = hdr.seq + len;
 814                 hdr.seq = 0;
 815                 hdr.ctl = RST | ACK;
 816         }
 817         print_packet(utcp, "send", &hdr, sizeof hdr);
 818         utcp->send(utcp, &hdr, sizeof hdr);
 819         return 0;
 820
 821 }
 822
 823 int utcp_shutdown(struct utcp_connection *c, int dir) {
 824         debug("%p shutdown %d\n", c ? c->utcp : NULL, dir);
 825         if(!c) {
 826                 errno = EFAULT;
 827                 return -1;
 828         }
 829
 830         if(c->reapable) {
 831                 debug("Error: shutdown() called on closed connection %p\n", c);
 832                 errno = EBADF;
 833                 return -1;
 834         }
 835
 836         // TODO: handle dir
 837
 838         switch(c->state) {
 839         case CLOSED:
 840                 return 0;
 841         case LISTEN:
 842         case SYN_SENT:
 843                 set_state(c, CLOSED);
 844                 return 0;
 845
 846         case SYN_RECEIVED:
 847         case ESTABLISHED:
 848                 set_state(c, FIN_WAIT_1);
 849                 break;
 850         case FIN_WAIT_1:
 851         case FIN_WAIT_2:
 852                 return 0;
 853         case CLOSE_WAIT:
 854                 set_state(c, CLOSING);
 855                 break;
 856
 857         case CLOSING:
 858         case LAST_ACK:
 859         case TIME_WAIT:
 860                 return 0;
 861         }
 862
 863         c->snd.last++;
 864
 865         ack(c, false);
 866         return 0;
 867 }
 868
 869 int utcp_close(struct utcp_connection *c) {
 870         if(utcp_shutdown(c, SHUT_RDWR))
 871                 return -1;
 872         c->reapable = true;
 873         return 0;
 874 }
 875
 876 int utcp_abort(struct utcp_connection *c) {
 877         if(!c) {
 878                 errno = EFAULT;
 879                 return -1;
 880         }
 881
 882         if(c->reapable) {
 883                 debug("Error: abort() called on closed connection %p\n", c);
 884                 errno = EBADF;
 885                 return -1;
 886         }
 887
 888         c->reapable = true;
 889
 890         switch(c->state) {
 891         case CLOSED:
 892                 return 0;
 893         case LISTEN:
 894         case SYN_SENT:
 895         case CLOSING:
 896         case LAST_ACK:
 897         case TIME_WAIT:
 898                 set_state(c, CLOSED);
 899                 return 0;
 900
 901         case SYN_RECEIVED:
 902         case ESTABLISHED:
 903         case FIN_WAIT_1:
 904         case FIN_WAIT_2:
 905         case CLOSE_WAIT:
 906                 set_state(c, CLOSED);
 907                 break;
 908         }
 909
 910         // Send RST
 911
 912         struct hdr hdr;
 913
 914         hdr.src = c->src;
 915         hdr.dst = c->dst;
 916         hdr.seq = c->snd.nxt;
 917         hdr.ack = 0;
 918         hdr.wnd = 0;
 919         hdr.ctl = RST;
 920
 921         print_packet(c->utcp, "send", &hdr, sizeof hdr);
 922         c->utcp->send(c->utcp, &hdr, sizeof hdr);
 923         return 0;
 924 }
 925
 926 static void retransmit(struct utcp_connection *c) {
 927         if(c->state == CLOSED || c->snd.nxt == c->snd.una)
 928                 return;
 929
 930         struct utcp *utcp = c->utcp;
 931
 932         struct {
 933                 struct hdr hdr;
 934                 char data[c->utcp->mtu];
 935         } pkt;
 936
 937         pkt.hdr.src = c->src;
 938         pkt.hdr.dst = c->dst;
 939
 940         switch(c->state) {
 941                 case LISTEN:
 942                         // TODO: this should not happen
 943                         break;
 944
 945                 case SYN_SENT:
 946                         pkt.hdr.seq = c->snd.iss;
 947                         pkt.hdr.ack = 0;
 948                         pkt.hdr.wnd = c->rcv.wnd;
 949                         pkt.hdr.ctl = SYN;
 950                         print_packet(c->utcp, "rtrx", &pkt, sizeof pkt.hdr);
 951                         utcp->send(utcp, &pkt, sizeof pkt.hdr);
 952                         break;
 953
 954                 case SYN_RECEIVED:
 955                         pkt.hdr.seq = c->snd.nxt;
 956                         pkt.hdr.ack = c->rcv.nxt;
 957                         pkt.hdr.ctl = SYN | ACK;
 958                         print_packet(c->utcp, "rtrx", &pkt, sizeof pkt.hdr);
 959                         utcp->send(utcp, &pkt, sizeof pkt.hdr);
 960                         break;
 961
 962                 case ESTABLISHED:
 963                 case FIN_WAIT_1:
 964                         pkt.hdr.seq = c->snd.una;
 965                         pkt.hdr.ack = c->rcv.nxt;
 966                         pkt.hdr.ctl = ACK;
 967                         uint32_t len = seqdiff(c->snd.nxt, c->snd.una);
 968                         if(c->state == FIN_WAIT_1)
 969                                 len--;
 970                         if(len > utcp->mtu)
 971                                 len = utcp->mtu;
 972                         else {
 973                                 if(c->state == FIN_WAIT_1)
 974                                         pkt.hdr.ctl |= FIN;
 975                         }
 976                         memcpy(pkt.data, c->sndbuf, len);
 977                         print_packet(c->utcp, "rtrx", &pkt, sizeof pkt.hdr + len);
 978                         utcp->send(utcp, &pkt, sizeof pkt.hdr + len);
 979                         break;
 980
 981                 default:
 982                         // TODO: implement
 983                         abort();
 984         }
 985 }
 986
 987 /* Handle timeouts.
 988  * One call to this function will loop through all connections,
 989  * checking if something needs to be resent or not.
 990  * The return value is the time to the next timeout in milliseconds,
 991  * or maybe a negative value if the timeout is infinite.
 992  */
 993 int utcp_timeout(struct utcp *utcp) {
 994         struct timeval now;
 995         gettimeofday(&now, NULL);
 996         struct timeval next = {now.tv_sec + 3600, now.tv_usec};
 997
 998         for(int i = 0; i < utcp->nconnections; i++) {
 999                 struct utcp_connection *c = utcp->connections[i];
1000                 if(!c)
1001                         continue;
1002
1003                 if(c->state == CLOSED) {
1004                         if(c->reapable) {
1005                                 debug("Reaping %p\n", c);
1006                                 free_connection(c);
1007                                 i--;
1008                         }
1009                         continue;
1010                 }
1011
1012                 if(timerisset(&c->conn_timeout) && timercmp(&c->conn_timeout, &now, <)) {
1013                         errno = ETIMEDOUT;
1014                         c->state = CLOSED;
1015                         if(c->recv)
1016                                 c->recv(c, NULL, 0);
1017                         continue;
1018                 }
1019
1020                 if(timerisset(&c->rtrx_timeout) && timercmp(&c->rtrx_timeout, &now, <)) {
1021                         retransmit(c);
1022                 }
1023
1024                 if(timerisset(&c->conn_timeout) && timercmp(&c->conn_timeout, &next, <))
1025                         next = c->conn_timeout;
1026
1027                 if(c->snd.nxt != c->snd.una) {
1028                         c->rtrx_timeout = now;
1029                         c->rtrx_timeout.tv_sec++;
1030                 } else {
1031                         timerclear(&c->rtrx_timeout);
1032                 }
1033
1034                 if(timerisset(&c->rtrx_timeout) && timercmp(&c->rtrx_timeout, &next, <))
1035                         next = c->rtrx_timeout;
1036         }
1037
1038         struct timeval diff;
1039         timersub(&next, &now, &diff);
1040         if(diff.tv_sec < 0)
1041                 return 0;
1042         return diff.tv_sec * 1000 + diff.tv_usec / 1000;
1043 }
1044
1045 struct utcp *utcp_init(utcp_accept_t accept, utcp_pre_accept_t pre_accept, utcp_send_t send, void *priv) {
1046         struct utcp *utcp = calloc(1, sizeof *utcp);
1047         if(!utcp)
1048                 return NULL;
1049
1050         if(!send) {
1051                 errno = EFAULT;
1052                 return NULL;
1053         }
1054
1055         utcp->accept = accept;
1056         utcp->pre_accept = pre_accept;
1057         utcp->send = send;
1058         utcp->priv = priv;
1059         utcp->mtu = 1000;
1060         utcp->timeout = 60;
1061
1062         return utcp;
1063 }
1064
1065 void utcp_exit(struct utcp *utcp) {
1066         if(!utcp)
1067                 return;
1068         for(int i = 0; i < utcp->nconnections; i++) {
1069                 if(!utcp->connections[i]->reapable)
1070                         debug("Warning, freeing unclosed connection %p\n", utcp->connections[i]);
1071                 free(utcp->connections[i]->sndbuf);
1072                 free(utcp->connections[i]);
1073         }
1074         free(utcp->connections);
1075         free(utcp);
1076 }
1077
1078 uint16_t utcp_get_mtu(struct utcp *utcp) {
1079         return utcp->mtu;
1080 }
1081
1082 void utcp_set_mtu(struct utcp *utcp, uint16_t mtu) {
1083         // TODO: handle overhead of the header
1084         utcp->mtu = mtu;
1085 }
1086
1087 int utcp_get_user_timeout(struct utcp *u) {
1088         return u->timeout;
1089 }
1090
1091 void utcp_set_user_timeout(struct utcp *u, int timeout) {
1092         u->timeout = timeout;
1093 }
1094
1095 size_t utcp_get_sndbuf(struct utcp_connection *c) {
1096         return c->maxsndbufsize;
1097 }
1098
1099 void utcp_set_sndbuf(struct utcp_connection *c, size_t size) {
1100         c->maxsndbufsize = size;
1101         if(c->maxsndbufsize != size)
1102                 c->maxsndbufsize = -1;
1103 }
1104
1105 bool utcp_get_nodelay(struct utcp_connection *c) {
1106         return c->nodelay;
1107 }
1108
1109 void utcp_set_nodelay(struct utcp_connection *c, bool nodelay) {
1110         c->nodelay = nodelay;
1111 }
1112
1113 bool utcp_get_keepalive(struct utcp_connection *c) {
1114         return c->keepalive;
1115 }
1116
1117 void utcp_set_keepalive(struct utcp_connection *c, bool keepalive) {
1118         c->keepalive = keepalive;
1119 }
1120
1121 size_t utcp_get_outq(struct utcp_connection *c) {
1122         return seqdiff(c->snd.nxt, c->snd.una);
1123 }