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 bool fin_wanted(struct utcp_connection *c, uint32_t seq) {
 107         if(seq != c->snd.last)
 108                 return false;
 109         switch(c->state) {
 110         case FIN_WAIT_1:
 111         case CLOSING:
 112         case LAST_ACK:
 113                 return true;
 114         default:
 115                 return false;
 116         }
 117 }
 118
 119 static inline void list_connections(struct utcp *utcp) {
 120         debug("%p has %d connections:\n", utcp, utcp->nconnections);
 121         for(int i = 0; i < utcp->nconnections; i++)
 122                 debug("  %u -> %u state %s\n", utcp->connections[i]->src, utcp->connections[i]->dst, strstate[utcp->connections[i]->state]);
 123 }
 124
 125 static int32_t seqdiff(uint32_t a, uint32_t b) {
 126         return a - b;
 127 }
 128
 129 // Buffer functions
 130 // TODO: convert to ringbuffers to avoid memmove() operations.
 131
 132 // Store data into the buffer
 133 static ssize_t buffer_put(struct buffer *buf, const void *data, size_t len) {
 134         if(buf->maxsize <= buf->used)
 135                 return 0;
 136         if(len > buf->maxsize - buf->used)
 137                 len = buf->maxsize - buf->used;
 138         if(len > buf->size - buf->used) {
 139                 size_t newsize = buf->size;
 140                 do {
 141                         newsize *= 2;
 142                 } while(newsize < buf->used + len);
 143                 if(newsize > buf->maxsize)
 144                         newsize = buf->maxsize;
 145                 char *newdata = realloc(buf->data, newsize);
 146                 if(!newdata)
 147                         return -1;
 148                 buf->data = newdata;
 149                 buf->size = newsize;
 150         }
 151         memcpy(buf->data + buf->used, data, len);
 152         buf->used += len;
 153         return len;
 154 }
 155
 156 // Get data from the buffer. data can be NULL.
 157 static ssize_t buffer_get(struct buffer *buf, void *data, size_t len) {
 158         if(len > buf->used)
 159                 len = buf->used;
 160         if(data)
 161                 memcpy(data, buf->data, len);
 162         if(len < buf->used)
 163                 memmove(buf->data, buf->data + len, buf->used - len);
 164         buf->used -= len;
 165         return len;
 166 }
 167
 168 // Copy data from the buffer without removing it.
 169 static ssize_t buffer_copy(struct buffer *buf, void *data, size_t offset, size_t len) {
 170         if(offset >= buf->used)
 171                 return 0;
 172         if(offset + len > buf->used)
 173                 len = buf->used - offset;
 174         memcpy(data, buf->data + offset, len);
 175         return len;
 176 }
 177
 178 static bool buffer_init(struct buffer *buf, uint32_t len, uint32_t maxlen) {
 179         memset(buf, 0, sizeof *buf);
 180         buf->data = malloc(len);
 181         if(!len)
 182                 return false;
 183         buf->size = len;
 184         buf->maxsize = maxlen;
 185         return true;
 186 }
 187
 188 static void buffer_exit(struct buffer *buf) {
 189         free(buf->data);
 190         memset(buf, 0, sizeof *buf);
 191 }
 192
 193 static uint32_t buffer_free(const struct buffer *buf) {
 194         return buf->maxsize - buf->used;
 195 }
 196
 197 // Connections are stored in a sorted list.
 198 // This gives O(log(N)) lookup time, O(N log(N)) insertion time and O(N) deletion time.
 199
 200 static int compare(const void *va, const void *vb) {
 201         assert(va && vb);
 202
 203         const struct utcp_connection *a = *(struct utcp_connection **)va;
 204         const struct utcp_connection *b = *(struct utcp_connection **)vb;
 205
 206         assert(a && b);
 207         assert(a->src && b->src);
 208
 209         int c = (int)a->src - (int)b->src;
 210         if(c)
 211                 return c;
 212         c = (int)a->dst - (int)b->dst;
 213         return c;
 214 }
 215
 216 static struct utcp_connection *find_connection(const struct utcp *utcp, uint16_t src, uint16_t dst) {
 217         if(!utcp->nconnections)
 218                 return NULL;
 219         struct utcp_connection key = {
 220                 .src = src,
 221                 .dst = dst,
 222         }, *keyp = &key;
 223         struct utcp_connection **match = bsearch(&keyp, utcp->connections, utcp->nconnections, sizeof *utcp->connections, compare);
 224         return match ? *match : NULL;
 225 }
 226
 227 static void free_connection(struct utcp_connection *c) {
 228         struct utcp *utcp = c->utcp;
 229         struct utcp_connection **cp = bsearch(&c, utcp->connections, utcp->nconnections, sizeof *utcp->connections, compare);
 230
 231         assert(cp);
 232
 233         int i = cp - utcp->connections;
 234         memmove(cp, cp + 1, (utcp->nconnections - i - 1) * sizeof *cp);
 235         utcp->nconnections--;
 236
 237         buffer_exit(&c->sndbuf);
 238         free(c);
 239 }
 240
 241 static struct utcp_connection *allocate_connection(struct utcp *utcp, uint16_t src, uint16_t dst) {
 242         // Check whether this combination of src and dst is free
 243
 244         if(src) {
 245                 if(find_connection(utcp, src, dst)) {
 246                         errno = EADDRINUSE;
 247                         return NULL;
 248                 }
 249         } else { // If src == 0, generate a random port number with the high bit set
 250                 if(utcp->nconnections >= 32767) {
 251                         errno = ENOMEM;
 252                         return NULL;
 253                 }
 254                 src = rand() | 0x8000;
 255                 while(find_connection(utcp, src, dst))
 256                         src++;
 257         }
 258
 259         // Allocate memory for the new connection
 260
 261         if(utcp->nconnections >= utcp->nallocated) {
 262                 if(!utcp->nallocated)
 263                         utcp->nallocated = 4;
 264                 else
 265                         utcp->nallocated *= 2;
 266                 struct utcp_connection **new_array = realloc(utcp->connections, utcp->nallocated * sizeof *utcp->connections);
 267                 if(!new_array)
 268                         return NULL;
 269                 utcp->connections = new_array;
 270         }
 271
 272         struct utcp_connection *c = calloc(1, sizeof *c);
 273         if(!c)
 274                 return NULL;
 275
 276         if(!buffer_init(&c->sndbuf, DEFAULT_SNDBUFSIZE, DEFAULT_MAXSNDBUFSIZE)) {
 277                 free(c);
 278                 return NULL;
 279         }
 280
 281         // Fill in the details
 282
 283         c->src = src;
 284         c->dst = dst;
 285         c->snd.iss = rand();
 286         c->snd.una = c->snd.iss;
 287         c->snd.nxt = c->snd.iss + 1;
 288         c->rcv.wnd = utcp->mtu;
 289         c->snd.last = c->snd.nxt;
 290         c->snd.cwnd = utcp->mtu;
 291         c->utcp = utcp;
 292
 293         // Add it to the sorted list of connections
 294
 295         utcp->connections[utcp->nconnections++] = c;
 296         qsort(utcp->connections, utcp->nconnections, sizeof *utcp->connections, compare);
 297
 298         return c;
 299 }
 300
 301 struct utcp_connection *utcp_connect(struct utcp *utcp, uint16_t dst, utcp_recv_t recv, void *priv) {
 302         struct utcp_connection *c = allocate_connection(utcp, 0, dst);
 303         if(!c)
 304                 return NULL;
 305
 306         c->recv = recv;
 307         c->priv = priv;
 308
 309         struct hdr hdr;
 310
 311         hdr.src = c->src;
 312         hdr.dst = c->dst;
 313         hdr.seq = c->snd.iss;
 314         hdr.ack = 0;
 315         hdr.wnd = c->rcv.wnd;
 316         hdr.ctl = SYN;
 317         hdr.aux = 0;
 318
 319         set_state(c, SYN_SENT);
 320
 321         print_packet(utcp, "send", &hdr, sizeof hdr);
 322         utcp->send(utcp, &hdr, sizeof hdr);
 323
 324         gettimeofday(&c->conn_timeout, NULL);
 325         c->conn_timeout.tv_sec += utcp->timeout;
 326
 327         return c;
 328 }
 329
 330 void utcp_accept(struct utcp_connection *c, utcp_recv_t recv, void *priv) {
 331         if(c->reapable || c->state != SYN_RECEIVED) {
 332                 debug("Error: accept() called on invalid connection %p in state %s\n", c, strstate[c->state]);
 333                 return;
 334         }
 335
 336         debug("%p accepted, %p %p\n", c, recv, priv);
 337         c->recv = recv;
 338         c->priv = priv;
 339         set_state(c, ESTABLISHED);
 340 }
 341
 342 static void ack(struct utcp_connection *c, bool sendatleastone) {
 343         int32_t left = seqdiff(c->snd.last, c->snd.nxt);
 344         int32_t cwndleft = c->snd.cwnd - seqdiff(c->snd.nxt, c->snd.una);
 345
 346         assert(left >= 0);
 347
 348         if(cwndleft <= 0)
 349                 cwndleft = 0;
 350
 351         if(cwndleft < left)
 352                 left = cwndleft;
 353
 354         if(!left && !sendatleastone)
 355                 return;
 356
 357         struct {
 358                 struct hdr hdr;
 359                 char data[];
 360         } *pkt;
 361
 362         pkt = malloc(sizeof pkt->hdr + c->utcp->mtu);
 363         if(!pkt->data)
 364                 return;
 365
 366         pkt->hdr.src = c->src;
 367         pkt->hdr.dst = c->dst;
 368         pkt->hdr.ack = c->rcv.nxt;
 369         pkt->hdr.wnd = c->snd.wnd;
 370         pkt->hdr.ctl = ACK;
 371         pkt->hdr.aux = 0;
 372
 373         do {
 374                 uint32_t seglen = left > c->utcp->mtu ? c->utcp->mtu : left;
 375                 pkt->hdr.seq = c->snd.nxt;
 376
 377                 buffer_copy(&c->sndbuf, pkt->data, seqdiff(c->snd.nxt, c->snd.una), seglen);
 378
 379                 c->snd.nxt += seglen;
 380                 left -= seglen;
 381
 382                 if(seglen && fin_wanted(c, c->snd.nxt)) {
 383                         seglen--;
 384                         pkt->hdr.ctl |= FIN;
 385                 }
 386
 387                 print_packet(c->utcp, "send", pkt, sizeof pkt->hdr + seglen);
 388                 c->utcp->send(c->utcp, pkt, sizeof pkt->hdr + seglen);
 389         } while(left);
 390
 391         free(pkt);
 392 }
 393
 394 ssize_t utcp_send(struct utcp_connection *c, const void *data, size_t len) {
 395         if(c->reapable) {
 396                 debug("Error: send() called on closed connection %p\n", c);
 397                 errno = EBADF;
 398                 return -1;
 399         }
 400
 401         switch(c->state) {
 402         case CLOSED:
 403         case LISTEN:
 404         case SYN_SENT:
 405         case SYN_RECEIVED:
 406                 debug("Error: send() called on unconnected connection %p\n", c);
 407                 errno = ENOTCONN;
 408                 return -1;
 409         case ESTABLISHED:
 410         case CLOSE_WAIT:
 411                 break;
 412         case FIN_WAIT_1:
 413         case FIN_WAIT_2:
 414         case CLOSING:
 415         case LAST_ACK:
 416         case TIME_WAIT:
 417                 debug("Error: send() called on closing connection %p\n", c);
 418                 errno = EPIPE;
 419                 return -1;
 420         }
 421
 422         // Add data to send buffer
 423
 424         if(!len)
 425                 return 0;
 426
 427         if(!data) {
 428                 errno = EFAULT;
 429                 return -1;
 430         }
 431
 432         len = buffer_put(&c->sndbuf, data, len);
 433         if(len <= 0) {
 434                 errno = EWOULDBLOCK;
 435                 return 0;
 436         }
 437
 438         c->snd.last += len;
 439         ack(c, false);
 440         return len;
 441 }
 442
 443 static void swap_ports(struct hdr *hdr) {
 444         uint16_t tmp = hdr->src;
 445         hdr->src = hdr->dst;
 446         hdr->dst = tmp;
 447 }
 448
 449 static void retransmit(struct utcp_connection *c) {
 450         if(c->state == CLOSED || c->snd.nxt == c->snd.una)
 451                 return;
 452
 453         struct utcp *utcp = c->utcp;
 454
 455         struct {
 456                 struct hdr hdr;
 457                 char data[];
 458         } *pkt;
 459
 460         pkt = malloc(sizeof pkt->hdr + c->utcp->mtu);
 461         if(!pkt)
 462                 return;
 463
 464         pkt->hdr.src = c->src;
 465         pkt->hdr.dst = c->dst;
 466
 467         switch(c->state) {
 468                 case SYN_SENT:
 469                         // Send our SYN again
 470                         pkt->hdr.seq = c->snd.iss;
 471                         pkt->hdr.ack = 0;
 472                         pkt->hdr.wnd = c->rcv.wnd;
 473                         pkt->hdr.ctl = SYN;
 474                         print_packet(c->utcp, "rtrx", pkt, sizeof pkt->hdr);
 475                         utcp->send(utcp, pkt, sizeof pkt->hdr);
 476                         break;
 477
 478                 case SYN_RECEIVED:
 479                         // Send SYNACK again
 480                         pkt->hdr.seq = c->snd.nxt;
 481                         pkt->hdr.ack = c->rcv.nxt;
 482                         pkt->hdr.ctl = SYN | ACK;
 483                         print_packet(c->utcp, "rtrx", pkt, sizeof pkt->hdr);
 484                         utcp->send(utcp, pkt, sizeof pkt->hdr);
 485                         break;
 486
 487                 case ESTABLISHED:
 488                 case FIN_WAIT_1:
 489                 case CLOSE_WAIT:
 490                 case CLOSING:
 491                 case LAST_ACK:
 492                         // Send unacked data again.
 493                         pkt->hdr.seq = c->snd.una;
 494                         pkt->hdr.ack = c->rcv.nxt;
 495                         pkt->hdr.ctl = ACK;
 496                         uint32_t len = seqdiff(c->snd.last, c->snd.una);
 497                         if(len > utcp->mtu)
 498                                 len = utcp->mtu;
 499                         if(fin_wanted(c, c->snd.una + len)) {
 500                                 len--;
 501                                 pkt->hdr.ctl |= FIN;
 502                         }
 503                         buffer_copy(&c->sndbuf, pkt->data, 0, len);
 504                         print_packet(c->utcp, "rtrx", pkt, sizeof pkt->hdr + len);
 505                         utcp->send(utcp, pkt, sizeof pkt->hdr + len);
 506                         break;
 507
 508                 case CLOSED:
 509                 case LISTEN:
 510                 case TIME_WAIT:
 511                 case FIN_WAIT_2:
 512                         // We shouldn't need to retransmit anything in this state.
 513 #ifdef UTCP_DEBUG
 514                         abort();
 515 #endif
 516                         timerclear(&c->rtrx_timeout);
 517                         break;
 518         }
 519
 520         free(pkt);
 521 }
 522
 523
 524 ssize_t utcp_recv(struct utcp *utcp, const void *data, size_t len) {
 525         if(!utcp) {
 526                 errno = EFAULT;
 527                 return -1;
 528         }
 529
 530         if(!len)
 531                 return 0;
 532
 533         if(!data) {
 534                 errno = EFAULT;
 535                 return -1;
 536         }
 537
 538         print_packet(utcp, "recv", data, len);
 539
 540         // Drop packets smaller than the header
 541
 542         struct hdr hdr;
 543         if(len < sizeof hdr) {
 544                 errno = EBADMSG;
 545                 return -1;
 546         }
 547
 548         // Make a copy from the potentially unaligned data to a struct hdr
 549
 550         memcpy(&hdr, data, sizeof hdr);
 551         data += sizeof hdr;
 552         len -= sizeof hdr;
 553
 554         // Drop packets with an unknown CTL flag
 555
 556         if(hdr.ctl & ~(SYN | ACK | RST | FIN)) {
 557                 errno = EBADMSG;
 558                 return -1;
 559         }
 560
 561         // Try to match the packet to an existing connection
 562
 563         struct utcp_connection *c = find_connection(utcp, hdr.dst, hdr.src);
 564
 565         // Is it for a new connection?
 566
 567         if(!c) {
 568                 // Ignore RST packets
 569
 570                 if(hdr.ctl & RST)
 571                         return 0;
 572
 573                 // Is it a SYN packet and are we LISTENing?
 574
 575                 if(hdr.ctl & SYN && !(hdr.ctl & ACK) && utcp->accept) {
 576                         // If we don't want to accept it, send a RST back
 577                         if((utcp->pre_accept && !utcp->pre_accept(utcp, hdr.dst))) {
 578                                 len = 1;
 579                                 goto reset;
 580                         }
 581
 582                         // Try to allocate memory, otherwise send a RST back
 583                         c = allocate_connection(utcp, hdr.dst, hdr.src);
 584                         if(!c) {
 585                                 len = 1;
 586                                 goto reset;
 587                         }
 588
 589                         // Return SYN+ACK, go to SYN_RECEIVED state
 590                         c->snd.wnd = hdr.wnd;
 591                         c->rcv.irs = hdr.seq;
 592                         c->rcv.nxt = c->rcv.irs + 1;
 593                         set_state(c, SYN_RECEIVED);
 594
 595                         hdr.dst = c->dst;
 596                         hdr.src = c->src;
 597                         hdr.ack = c->rcv.irs + 1;
 598                         hdr.seq = c->snd.iss;
 599                         hdr.ctl = SYN | ACK;
 600                         print_packet(c->utcp, "send", &hdr, sizeof hdr);
 601                         utcp->send(utcp, &hdr, sizeof hdr);
 602                 } else {
 603                         // No, we don't want your packets, send a RST back
 604                         len = 1;
 605                         goto reset;
 606                 }
 607
 608                 return 0;
 609         }
 610
 611         debug("%p state %s\n", c->utcp, strstate[c->state]);
 612
 613         // In case this is for a CLOSED connection, ignore the packet.
 614         // TODO: make it so incoming packets can never match a CLOSED connection.
 615
 616         if(c->state == CLOSED)
 617                 return 0;
 618
 619         // It is for an existing connection.
 620
 621         uint32_t prevrcvnxt = c->rcv.nxt;
 622
 623         // 1. Drop invalid packets.
 624
 625         // 1a. Drop packets that should not happen in our current state.
 626
 627         switch(c->state) {
 628         case SYN_SENT:
 629         case SYN_RECEIVED:
 630         case ESTABLISHED:
 631         case FIN_WAIT_1:
 632         case FIN_WAIT_2:
 633         case CLOSE_WAIT:
 634         case CLOSING:
 635         case LAST_ACK:
 636         case TIME_WAIT:
 637                 break;
 638         default:
 639 #ifdef UTCP_DEBUG
 640                 abort();
 641 #endif
 642                 break;
 643         }
 644
 645         // 1b. Drop packets with a sequence number not in our receive window.
 646
 647         bool acceptable;
 648
 649         if(c->state == SYN_SENT)
 650                 acceptable = true;
 651
 652         // TODO: handle packets overlapping c->rcv.nxt.
 653 #if 0
 654         // Only use this when accepting out-of-order packets.
 655         else if(len == 0)
 656                 if(c->rcv.wnd == 0)
 657                         acceptable = hdr.seq == c->rcv.nxt;
 658                 else
 659                         acceptable = (seqdiff(hdr.seq, c->rcv.nxt) >= 0 && seqdiff(hdr.seq, c->rcv.nxt + c->rcv.wnd) < 0);
 660         else
 661                 if(c->rcv.wnd == 0)
 662                         // We don't accept data when the receive window is zero.
 663                         acceptable = false;
 664                 else
 665                         // Both start and end of packet must be within the receive window
 666                         acceptable = (seqdiff(hdr.seq, c->rcv.nxt) >= 0 && seqdiff(hdr.seq, c->rcv.nxt + c->rcv.wnd) < 0)
 667                                 || (seqdiff(hdr.seq + len + 1, c->rcv.nxt) >= 0 && seqdiff(hdr.seq + len - 1, c->rcv.nxt + c->rcv.wnd) < 0);
 668 #else
 669         if(c->state != SYN_SENT)
 670                 acceptable = hdr.seq == c->rcv.nxt;
 671 #endif
 672
 673         if(!acceptable) {
 674                 debug("Packet not acceptable, %u <= %u + %zu < %u\n", c->rcv.nxt, hdr.seq, len, c->rcv.nxt + c->rcv.wnd);
 675                 // Ignore unacceptable RST packets.
 676                 if(hdr.ctl & RST)
 677                         return 0;
 678                 // Otherwise, send an ACK back in the hope things improve.
 679                 ack(c, true);
 680                 return 0;
 681         }
 682
 683         c->snd.wnd = hdr.wnd; // TODO: move below
 684
 685         // 1c. Drop packets with an invalid ACK.
 686         // ackno should not roll back, and it should also not be bigger than snd.nxt.
 687
 688         if(hdr.ctl & ACK && (seqdiff(hdr.ack, c->snd.nxt) > 0 || seqdiff(hdr.ack, c->snd.una) < 0)) {
 689                 debug("Packet ack seqno out of range, %u %u %u\n", hdr.ack, c->snd.una, c->snd.nxt);
 690                 // Ignore unacceptable RST packets.
 691                 if(hdr.ctl & RST)
 692                         return 0;
 693                 goto reset;
 694         }
 695
 696         // 2. Handle RST packets
 697
 698         if(hdr.ctl & RST) {
 699                 switch(c->state) {
 700                 case SYN_SENT:
 701                         if(!(hdr.ctl & ACK))
 702                                 return 0;
 703                         // The peer has refused our connection.
 704                         set_state(c, CLOSED);
 705                         errno = ECONNREFUSED;
 706                         if(c->recv)
 707                                 c->recv(c, NULL, 0);
 708                         return 0;
 709                 case SYN_RECEIVED:
 710                         if(hdr.ctl & ACK)
 711                                 return 0;
 712                         // We haven't told the application about this connection yet. Silently delete.
 713                         free_connection(c);
 714                         return 0;
 715                 case ESTABLISHED:
 716                 case FIN_WAIT_1:
 717                 case FIN_WAIT_2:
 718                 case CLOSE_WAIT:
 719                         if(hdr.ctl & ACK)
 720                                 return 0;
 721                         // The peer has aborted our connection.
 722                         set_state(c, CLOSED);
 723                         errno = ECONNRESET;
 724                         if(c->recv)
 725                                 c->recv(c, NULL, 0);
 726                         return 0;
 727                 case CLOSING:
 728                 case LAST_ACK:
 729                 case TIME_WAIT:
 730                         if(hdr.ctl & ACK)
 731                                 return 0;
 732                         // As far as the application is concerned, the connection has already been closed.
 733                         // If it has called utcp_close() already, we can immediately free this connection.
 734                         if(c->reapable) {
 735                                 free_connection(c);
 736                                 return 0;
 737                         }
 738                         // Otherwise, immediately move to the CLOSED state.
 739                         set_state(c, CLOSED);
 740                         return 0;
 741                 default:
 742 #ifdef UTCP_DEBUG
 743                         abort();
 744 #endif
 745                         break;
 746                 }
 747         }
 748
 749         // 3. Advance snd.una
 750
 751         uint32_t advanced = seqdiff(hdr.ack, c->snd.una);
 752         prevrcvnxt = c->rcv.nxt;
 753
 754         if(advanced) {
 755                 int32_t data_acked = advanced;
 756
 757                 switch(c->state) {
 758                         case SYN_SENT:
 759                         case SYN_RECEIVED:
 760                                 data_acked--;
 761                                 break;
 762                         // TODO: handle FIN as well.
 763                         default:
 764                                 break;
 765                 }
 766
 767                 assert(data_acked >= 0);
 768
 769                 int32_t bufused = seqdiff(c->snd.last, c->snd.una);
 770                 assert(data_acked <= bufused);
 771
 772                 if(data_acked)
 773                         buffer_get(&c->sndbuf, NULL, data_acked);
 774
 775                 c->snd.una = hdr.ack;
 776
 777                 c->dupack = 0;
 778                 c->snd.cwnd += utcp->mtu;
 779                 if(c->snd.cwnd > c->sndbuf.maxsize)
 780                         c->snd.cwnd = c->sndbuf.maxsize;
 781
 782                 // Check if we have sent a FIN that is now ACKed.
 783                 switch(c->state) {
 784                 case FIN_WAIT_1:
 785                         if(c->snd.una == c->snd.last)
 786                                 set_state(c, FIN_WAIT_2);
 787                         break;
 788                 case CLOSING:
 789                         if(c->snd.una == c->snd.last) {
 790                                 gettimeofday(&c->conn_timeout, NULL);
 791                                 c->conn_timeout.tv_sec += 60;
 792                                 set_state(c, TIME_WAIT);
 793                         }
 794                         break;
 795                 default:
 796                         break;
 797                 }
 798         } else {
 799                 if(!len) {
 800                         c->dupack++;
 801                         if(c->dupack == 3) {
 802                                 debug("Triplicate ACK\n");
 803                                 //TODO: Resend one packet and go to fast recovery mode. See RFC 6582.
 804                                 //We do a very simple variant here; reset the nxt pointer to the last acknowledged packet from the peer.
 805                                 //This will cause us to start retransmitting, but at the same speed as the incoming ACKs arrive,
 806                                 //thus preventing a drop in speed.
 807                                 c->snd.nxt = c->snd.una;
 808                         }
 809                 }
 810         }
 811
 812         // 4. Update timers
 813
 814         if(advanced) {
 815                 timerclear(&c->conn_timeout); // It will be set anew in utcp_timeout() if c->snd.una != c->snd.nxt.
 816                 if(c->snd.una == c->snd.nxt)
 817                         timerclear(&c->rtrx_timeout);
 818         }
 819
 820         // 5. Process SYN stuff
 821
 822         if(hdr.ctl & SYN) {
 823                 switch(c->state) {
 824                 case SYN_SENT:
 825                         // This is a SYNACK. It should always have ACKed the SYN.
 826                         if(!advanced)
 827                                 goto reset;
 828                         c->rcv.irs = hdr.seq;
 829                         c->rcv.nxt = hdr.seq;
 830                         set_state(c, ESTABLISHED);
 831                         // TODO: notify application of this somehow.
 832                         break;
 833                 case SYN_RECEIVED:
 834                 case ESTABLISHED:
 835                 case FIN_WAIT_1:
 836                 case FIN_WAIT_2:
 837                 case CLOSE_WAIT:
 838                 case CLOSING:
 839                 case LAST_ACK:
 840                 case TIME_WAIT:
 841                         // Ehm, no. We should never receive a second SYN.
 842                         goto reset;
 843                 default:
 844 #ifdef UTCP_DEBUG
 845                         abort();
 846 #endif
 847                         return 0;
 848                 }
 849
 850                 // SYN counts as one sequence number
 851                 c->rcv.nxt++;
 852         }
 853
 854         // 6. Process new data
 855
 856         if(c->state == SYN_RECEIVED) {
 857                 // This is the ACK after the SYNACK. It should always have ACKed the SYNACK.
 858                 if(!advanced)
 859                         goto reset;
 860
 861                 // Are we still LISTENing?
 862                 if(utcp->accept)
 863                         utcp->accept(c, c->src);
 864
 865                 if(c->state != ESTABLISHED) {
 866                         set_state(c, CLOSED);
 867                         c->reapable = true;
 868                         goto reset;
 869                 }
 870         }
 871
 872         if(len) {
 873                 switch(c->state) {
 874                 case SYN_SENT:
 875                 case SYN_RECEIVED:
 876                         // This should never happen.
 877 #ifdef UTCP_DEBUG
 878                         abort();
 879 #endif
 880                         return 0;
 881                 case ESTABLISHED:
 882                 case FIN_WAIT_1:
 883                 case FIN_WAIT_2:
 884                         break;
 885                 case CLOSE_WAIT:
 886                 case CLOSING:
 887                 case LAST_ACK:
 888                 case TIME_WAIT:
 889                         // Ehm no, We should never receive more data after a FIN.
 890                         goto reset;
 891                 default:
 892 #ifdef UTCP_DEBUG
 893                         abort();
 894 #endif
 895                         return 0;
 896                 }
 897
 898                 ssize_t rxd;
 899
 900                 if(c->recv) {
 901                         rxd = c->recv(c, data, len);
 902                         if(rxd != len) {
 903                                 // TODO: once we have a receive buffer, handle the application not accepting all data.
 904                                 abort();
 905                         }
 906                         if(rxd < 0)
 907                                 rxd = 0;
 908                         else if(rxd > len)
 909                                 rxd = len; // Bad application, bad!
 910                 } else {
 911                         rxd = len;
 912                 }
 913
 914                 c->rcv.nxt += len;
 915         }
 916
 917         // 7. Process FIN stuff
 918
 919         if(hdr.ctl & FIN) {
 920                 switch(c->state) {
 921                 case SYN_SENT:
 922                 case SYN_RECEIVED:
 923                         // This should never happen.
 924 #ifdef UTCP_DEBUG
 925                         abort();
 926 #endif
 927                         break;
 928                 case ESTABLISHED:
 929                         set_state(c, CLOSE_WAIT);
 930                         break;
 931                 case FIN_WAIT_1:
 932                         set_state(c, CLOSING);
 933                         break;
 934                 case FIN_WAIT_2:
 935                         gettimeofday(&c->conn_timeout, NULL);
 936                         c->conn_timeout.tv_sec += 60;
 937                         set_state(c, TIME_WAIT);
 938                         break;
 939                 case CLOSE_WAIT:
 940                 case CLOSING:
 941                 case LAST_ACK:
 942                 case TIME_WAIT:
 943                         // Ehm, no. We should never receive a second FIN.
 944                         goto reset;
 945                 default:
 946 #ifdef UTCP_DEBUG
 947                         abort();
 948 #endif
 949                         break;
 950                 }
 951
 952                 // FIN counts as one sequence number
 953                 c->rcv.nxt++;
 954                 len++;
 955
 956                 // Inform the application that the peer closed the connection.
 957                 if(c->recv) {
 958                         errno = 0;
 959                         c->recv(c, NULL, 0);
 960                 }
 961         }
 962
 963         // Now we send something back if:
 964         // - we advanced rcv.nxt (ie, we got some data that needs to be ACKed)
 965         //   -> sendatleastone = true
 966         // - or we got an ack, so we should maybe send a bit more data
 967         //   -> sendatleastone = false
 968
 969 ack:
 970         ack(c, prevrcvnxt != c->rcv.nxt);
 971         return 0;
 972
 973 reset:
 974         swap_ports(&hdr);
 975         hdr.wnd = 0;
 976         if(hdr.ctl & ACK) {
 977                 hdr.seq = hdr.ack;
 978                 hdr.ctl = RST;
 979         } else {
 980                 hdr.ack = hdr.seq + len;
 981                 hdr.seq = 0;
 982                 hdr.ctl = RST | ACK;
 983         }
 984         print_packet(utcp, "send", &hdr, sizeof hdr);
 985         utcp->send(utcp, &hdr, sizeof hdr);
 986         return 0;
 987
 988 }
 989
 990 int utcp_shutdown(struct utcp_connection *c, int dir) {
 991         debug("%p shutdown %d at %u\n", c ? c->utcp : NULL, dir, c->snd.last);
 992         if(!c) {
 993                 errno = EFAULT;
 994                 return -1;
 995         }
 996
 997         if(c->reapable) {
 998                 debug("Error: shutdown() called on closed connection %p\n", c);
 999                 errno = EBADF;
1000                 return -1;
1001         }
1002
1003         if(!(dir == UTCP_SHUT_RD || dir == UTCP_SHUT_WR || dir == UTCP_SHUT_RDWR)) {
1004                 errno = EINVAL;
1005                 return -1;
1006         }
1007
1008         // TCP does not have a provision for stopping incoming packets.
1009         // The best we can do is to just ignore them.
1010         if(dir == UTCP_SHUT_RD || dir == UTCP_SHUT_RDWR)
1011                 c->recv = NULL;
1012
1013         // The rest of the code deals with shutting down writes.
1014         if(dir == UTCP_SHUT_RD)
1015                 return 0;
1016
1017         switch(c->state) {
1018         case CLOSED:
1019         case LISTEN:
1020                 errno = ENOTCONN;
1021                 return -1;
1022
1023         case SYN_SENT:
1024                 set_state(c, CLOSED);
1025                 return 0;
1026
1027         case SYN_RECEIVED:
1028         case ESTABLISHED:
1029                 set_state(c, FIN_WAIT_1);
1030                 break;
1031         case FIN_WAIT_1:
1032         case FIN_WAIT_2:
1033                 return 0;
1034         case CLOSE_WAIT:
1035                 set_state(c, CLOSING);
1036                 break;
1037
1038         case CLOSING:
1039         case LAST_ACK:
1040         case TIME_WAIT:
1041                 return 0;
1042         }
1043
1044         c->snd.last++;
1045
1046         ack(c, false);
1047         return 0;
1048 }
1049
1050 int utcp_close(struct utcp_connection *c) {
1051         if(utcp_shutdown(c, SHUT_RDWR))
1052                 return -1;
1053         c->recv = NULL;
1054         c->poll = NULL;
1055         c->reapable = true;
1056         return 0;
1057 }
1058
1059 int utcp_abort(struct utcp_connection *c) {
1060         if(!c) {
1061                 errno = EFAULT;
1062                 return -1;
1063         }
1064
1065         if(c->reapable) {
1066                 debug("Error: abort() called on closed connection %p\n", c);
1067                 errno = EBADF;
1068                 return -1;
1069         }
1070
1071         c->recv = NULL;
1072         c->poll = NULL;
1073         c->reapable = true;
1074
1075         switch(c->state) {
1076         case CLOSED:
1077                 return 0;
1078         case LISTEN:
1079         case SYN_SENT:
1080         case CLOSING:
1081         case LAST_ACK:
1082         case TIME_WAIT:
1083                 set_state(c, CLOSED);
1084                 return 0;
1085
1086         case SYN_RECEIVED:
1087         case ESTABLISHED:
1088         case FIN_WAIT_1:
1089         case FIN_WAIT_2:
1090         case CLOSE_WAIT:
1091                 set_state(c, CLOSED);
1092                 break;
1093         }
1094
1095         // Send RST
1096
1097         struct hdr hdr;
1098
1099         hdr.src = c->src;
1100         hdr.dst = c->dst;
1101         hdr.seq = c->snd.nxt;
1102         hdr.ack = 0;
1103         hdr.wnd = 0;
1104         hdr.ctl = RST;
1105
1106         print_packet(c->utcp, "send", &hdr, sizeof hdr);
1107         c->utcp->send(c->utcp, &hdr, sizeof hdr);
1108         return 0;
1109 }
1110
1111 /* Handle timeouts.
1112  * One call to this function will loop through all connections,
1113  * checking if something needs to be resent or not.
1114  * The return value is the time to the next timeout in milliseconds,
1115  * or maybe a negative value if the timeout is infinite.
1116  */
1117 struct timeval utcp_timeout(struct utcp *utcp) {
1118         struct timeval now;
1119         gettimeofday(&now, NULL);
1120         struct timeval next = {now.tv_sec + 3600, now.tv_usec};
1121
1122         for(int i = 0; i < utcp->nconnections; i++) {
1123                 struct utcp_connection *c = utcp->connections[i];
1124                 if(!c)
1125                         continue;
1126
1127                 if(c->state == CLOSED) {
1128                         if(c->reapable) {
1129                                 debug("Reaping %p\n", c);
1130                                 free_connection(c);
1131                                 i--;
1132                         }
1133                         continue;
1134                 }
1135
1136                 if(timerisset(&c->conn_timeout) && timercmp(&c->conn_timeout, &now, <)) {
1137                         errno = ETIMEDOUT;
1138                         c->state = CLOSED;
1139                         if(c->recv)
1140                                 c->recv(c, NULL, 0);
1141                         continue;
1142                 }
1143
1144                 if(timerisset(&c->rtrx_timeout) && timercmp(&c->rtrx_timeout, &now, <)) {
1145                         retransmit(c);
1146                 }
1147
1148                 if(c->poll && buffer_free(&c->sndbuf) && (c->state == ESTABLISHED || c->state == CLOSE_WAIT))
1149                         c->poll(c, buffer_free(&c->sndbuf));
1150
1151                 if(timerisset(&c->conn_timeout) && timercmp(&c->conn_timeout, &next, <))
1152                         next = c->conn_timeout;
1153
1154                 if(c->snd.nxt != c->snd.una) {
1155                         c->rtrx_timeout = now;
1156                         c->rtrx_timeout.tv_sec++;
1157                 } else {
1158                         timerclear(&c->rtrx_timeout);
1159                 }
1160
1161                 if(timerisset(&c->rtrx_timeout) && timercmp(&c->rtrx_timeout, &next, <))
1162                         next = c->rtrx_timeout;
1163         }
1164
1165         struct timeval diff;
1166         timersub(&next, &now, &diff);
1167         return diff;
1168 }
1169
1170 bool utcp_is_active(struct utcp *utcp) {
1171         if(!utcp)
1172                 return false;
1173
1174         for(int i = 0; i < utcp->nconnections; i++)
1175                 if(utcp->connections[i]->state != CLOSED && utcp->connections[i]->state != TIME_WAIT)
1176                         return true;
1177
1178         return false;
1179 }
1180
1181 struct utcp *utcp_init(utcp_accept_t accept, utcp_pre_accept_t pre_accept, utcp_send_t send, void *priv) {
1182         struct utcp *utcp = calloc(1, sizeof *utcp);
1183         if(!utcp)
1184                 return NULL;
1185
1186         if(!send) {
1187                 errno = EFAULT;
1188                 return NULL;
1189         }
1190
1191         utcp->accept = accept;
1192         utcp->pre_accept = pre_accept;
1193         utcp->send = send;
1194         utcp->priv = priv;
1195         utcp->mtu = 1000;
1196         utcp->timeout = 60;
1197
1198         return utcp;
1199 }
1200
1201 void utcp_exit(struct utcp *utcp) {
1202         if(!utcp)
1203                 return;
1204         for(int i = 0; i < utcp->nconnections; i++) {
1205                 if(!utcp->connections[i]->reapable)
1206                         debug("Warning, freeing unclosed connection %p\n", utcp->connections[i]);
1207                 buffer_exit(&utcp->connections[i]->sndbuf);
1208                 free(utcp->connections[i]);
1209         }
1210         free(utcp->connections);
1211         free(utcp);
1212 }
1213
1214 uint16_t utcp_get_mtu(struct utcp *utcp) {
1215         return utcp ? utcp->mtu : 0;
1216 }
1217
1218 void utcp_set_mtu(struct utcp *utcp, uint16_t mtu) {
1219         // TODO: handle overhead of the header
1220         if(utcp)
1221                 utcp->mtu = mtu;
1222 }
1223
1224 int utcp_get_user_timeout(struct utcp *u) {
1225         return u ? u->timeout : 0;
1226 }
1227
1228 void utcp_set_user_timeout(struct utcp *u, int timeout) {
1229         if(u)
1230                 u->timeout = timeout;
1231 }
1232
1233 size_t utcp_get_sndbuf(struct utcp_connection *c) {
1234         return c ? c->sndbuf.maxsize : 0;
1235 }
1236
1237 size_t utcp_get_sndbuf_free(struct utcp_connection *c) {
1238         if(c && (c->state == ESTABLISHED || c->state == CLOSE_WAIT))
1239                 return buffer_free(&c->sndbuf);
1240         else
1241                 return 0;
1242 }
1243
1244 void utcp_set_sndbuf(struct utcp_connection *c, size_t size) {
1245         if(!c)
1246                 return;
1247         c->sndbuf.maxsize = size;
1248         if(c->sndbuf.maxsize != size)
1249                 c->sndbuf.maxsize = -1;
1250 }
1251
1252 bool utcp_get_nodelay(struct utcp_connection *c) {
1253         return c ? c->nodelay : false;
1254 }
1255
1256 void utcp_set_nodelay(struct utcp_connection *c, bool nodelay) {
1257         if(c)
1258                 c->nodelay = nodelay;
1259 }
1260
1261 bool utcp_get_keepalive(struct utcp_connection *c) {
1262         return c ? c->keepalive : false;
1263 }
1264
1265 void utcp_set_keepalive(struct utcp_connection *c, bool keepalive) {
1266         if(c)
1267                 c->keepalive = keepalive;
1268 }
1269
1270 size_t utcp_get_outq(struct utcp_connection *c) {
1271         return c ? seqdiff(c->snd.nxt, c->snd.una) : 0;
1272 }
1273
1274 void utcp_set_recv_cb(struct utcp_connection *c, utcp_recv_t recv) {
1275         if(c)
1276                 c->recv = recv;
1277 }
1278
1279 void utcp_set_poll_cb(struct utcp_connection *c, utcp_poll_t poll) {
1280         if(c)
1281                 c->poll = poll;
1282 }
1283
1284 void utcp_set_accept_cb(struct utcp *utcp, utcp_accept_t accept, utcp_pre_accept_t pre_accept) {
1285         if(utcp) {
1286                 utcp->accept = accept;
1287                 utcp->pre_accept = pre_accept;
1288         }
1289 }