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                         c->snd.nxt = c->snd.una + len;
 504                         buffer_copy(&c->sndbuf, pkt->data, 0, len);
 505                         print_packet(c->utcp, "rtrx", pkt, sizeof pkt->hdr + len);
 506                         utcp->send(utcp, pkt, sizeof pkt->hdr + len);
 507                         break;
 508
 509                 case CLOSED:
 510                 case LISTEN:
 511                 case TIME_WAIT:
 512                 case FIN_WAIT_2:
 513                         // We shouldn't need to retransmit anything in this state.
 514 #ifdef UTCP_DEBUG
 515                         abort();
 516 #endif
 517                         timerclear(&c->rtrx_timeout);
 518                         break;
 519         }
 520
 521         free(pkt);
 522 }
 523
 524
 525 ssize_t utcp_recv(struct utcp *utcp, const void *data, size_t len) {
 526         if(!utcp) {
 527                 errno = EFAULT;
 528                 return -1;
 529         }
 530
 531         if(!len)
 532                 return 0;
 533
 534         if(!data) {
 535                 errno = EFAULT;
 536                 return -1;
 537         }
 538
 539         print_packet(utcp, "recv", data, len);
 540
 541         // Drop packets smaller than the header
 542
 543         struct hdr hdr;
 544         if(len < sizeof hdr) {
 545                 errno = EBADMSG;
 546                 return -1;
 547         }
 548
 549         // Make a copy from the potentially unaligned data to a struct hdr
 550
 551         memcpy(&hdr, data, sizeof hdr);
 552         data += sizeof hdr;
 553         len -= sizeof hdr;
 554
 555         // Drop packets with an unknown CTL flag
 556
 557         if(hdr.ctl & ~(SYN | ACK | RST | FIN)) {
 558                 errno = EBADMSG;
 559                 return -1;
 560         }
 561
 562         // Try to match the packet to an existing connection
 563
 564         struct utcp_connection *c = find_connection(utcp, hdr.dst, hdr.src);
 565
 566         // Is it for a new connection?
 567
 568         if(!c) {
 569                 // Ignore RST packets
 570
 571                 if(hdr.ctl & RST)
 572                         return 0;
 573
 574                 // Is it a SYN packet and are we LISTENing?
 575
 576                 if(hdr.ctl & SYN && !(hdr.ctl & ACK) && utcp->accept) {
 577                         // If we don't want to accept it, send a RST back
 578                         if((utcp->pre_accept && !utcp->pre_accept(utcp, hdr.dst))) {
 579                                 len = 1;
 580                                 goto reset;
 581                         }
 582
 583                         // Try to allocate memory, otherwise send a RST back
 584                         c = allocate_connection(utcp, hdr.dst, hdr.src);
 585                         if(!c) {
 586                                 len = 1;
 587                                 goto reset;
 588                         }
 589
 590                         // Return SYN+ACK, go to SYN_RECEIVED state
 591                         c->snd.wnd = hdr.wnd;
 592                         c->rcv.irs = hdr.seq;
 593                         c->rcv.nxt = c->rcv.irs + 1;
 594                         set_state(c, SYN_RECEIVED);
 595
 596                         hdr.dst = c->dst;
 597                         hdr.src = c->src;
 598                         hdr.ack = c->rcv.irs + 1;
 599                         hdr.seq = c->snd.iss;
 600                         hdr.ctl = SYN | ACK;
 601                         print_packet(c->utcp, "send", &hdr, sizeof hdr);
 602                         utcp->send(utcp, &hdr, sizeof hdr);
 603                 } else {
 604                         // No, we don't want your packets, send a RST back
 605                         len = 1;
 606                         goto reset;
 607                 }
 608
 609                 return 0;
 610         }
 611
 612         debug("%p state %s\n", c->utcp, strstate[c->state]);
 613
 614         // In case this is for a CLOSED connection, ignore the packet.
 615         // TODO: make it so incoming packets can never match a CLOSED connection.
 616
 617         if(c->state == CLOSED)
 618                 return 0;
 619
 620         // It is for an existing connection.
 621
 622         uint32_t prevrcvnxt = c->rcv.nxt;
 623
 624         // 1. Drop invalid packets.
 625
 626         // 1a. Drop packets that should not happen in our current state.
 627
 628         switch(c->state) {
 629         case SYN_SENT:
 630         case SYN_RECEIVED:
 631         case ESTABLISHED:
 632         case FIN_WAIT_1:
 633         case FIN_WAIT_2:
 634         case CLOSE_WAIT:
 635         case CLOSING:
 636         case LAST_ACK:
 637         case TIME_WAIT:
 638                 break;
 639         default:
 640 #ifdef UTCP_DEBUG
 641                 abort();
 642 #endif
 643                 break;
 644         }
 645
 646         // 1b. Drop packets with a sequence number not in our receive window.
 647
 648         bool acceptable;
 649
 650         if(c->state == SYN_SENT)
 651                 acceptable = true;
 652
 653         // TODO: handle packets overlapping c->rcv.nxt.
 654 #if 0
 655         // Only use this when accepting out-of-order packets.
 656         else if(len == 0)
 657                 if(c->rcv.wnd == 0)
 658                         acceptable = hdr.seq == c->rcv.nxt;
 659                 else
 660                         acceptable = (seqdiff(hdr.seq, c->rcv.nxt) >= 0 && seqdiff(hdr.seq, c->rcv.nxt + c->rcv.wnd) < 0);
 661         else
 662                 if(c->rcv.wnd == 0)
 663                         // We don't accept data when the receive window is zero.
 664                         acceptable = false;
 665                 else
 666                         // Both start and end of packet must be within the receive window
 667                         acceptable = (seqdiff(hdr.seq, c->rcv.nxt) >= 0 && seqdiff(hdr.seq, c->rcv.nxt + c->rcv.wnd) < 0)
 668                                 || (seqdiff(hdr.seq + len + 1, c->rcv.nxt) >= 0 && seqdiff(hdr.seq + len - 1, c->rcv.nxt + c->rcv.wnd) < 0);
 669 #else
 670         if(c->state != SYN_SENT)
 671                 acceptable = hdr.seq == c->rcv.nxt;
 672 #endif
 673
 674         if(!acceptable) {
 675                 debug("Packet not acceptable, %u <= %u + %zu < %u\n", c->rcv.nxt, hdr.seq, len, c->rcv.nxt + c->rcv.wnd);
 676                 // Ignore unacceptable RST packets.
 677                 if(hdr.ctl & RST)
 678                         return 0;
 679                 // Otherwise, send an ACK back in the hope things improve.
 680                 ack(c, true);
 681                 return 0;
 682         }
 683
 684         c->snd.wnd = hdr.wnd; // TODO: move below
 685
 686         // 1c. Drop packets with an invalid ACK.
 687         // ackno should not roll back, and it should also not be bigger than snd.nxt.
 688
 689         if(hdr.ctl & ACK && (seqdiff(hdr.ack, c->snd.nxt) > 0 || seqdiff(hdr.ack, c->snd.una) < 0)) {
 690                 debug("Packet ack seqno out of range, %u %u %u\n", hdr.ack, c->snd.una, c->snd.nxt);
 691                 // Ignore unacceptable RST packets.
 692                 if(hdr.ctl & RST)
 693                         return 0;
 694                 goto reset;
 695         }
 696
 697         // 2. Handle RST packets
 698
 699         if(hdr.ctl & RST) {
 700                 switch(c->state) {
 701                 case SYN_SENT:
 702                         if(!(hdr.ctl & ACK))
 703                                 return 0;
 704                         // The peer has refused our connection.
 705                         set_state(c, CLOSED);
 706                         errno = ECONNREFUSED;
 707                         if(c->recv)
 708                                 c->recv(c, NULL, 0);
 709                         return 0;
 710                 case SYN_RECEIVED:
 711                         if(hdr.ctl & ACK)
 712                                 return 0;
 713                         // We haven't told the application about this connection yet. Silently delete.
 714                         free_connection(c);
 715                         return 0;
 716                 case ESTABLISHED:
 717                 case FIN_WAIT_1:
 718                 case FIN_WAIT_2:
 719                 case CLOSE_WAIT:
 720                         if(hdr.ctl & ACK)
 721                                 return 0;
 722                         // The peer has aborted our connection.
 723                         set_state(c, CLOSED);
 724                         errno = ECONNRESET;
 725                         if(c->recv)
 726                                 c->recv(c, NULL, 0);
 727                         return 0;
 728                 case CLOSING:
 729                 case LAST_ACK:
 730                 case TIME_WAIT:
 731                         if(hdr.ctl & ACK)
 732                                 return 0;
 733                         // As far as the application is concerned, the connection has already been closed.
 734                         // If it has called utcp_close() already, we can immediately free this connection.
 735                         if(c->reapable) {
 736                                 free_connection(c);
 737                                 return 0;
 738                         }
 739                         // Otherwise, immediately move to the CLOSED state.
 740                         set_state(c, CLOSED);
 741                         return 0;
 742                 default:
 743 #ifdef UTCP_DEBUG
 744                         abort();
 745 #endif
 746                         break;
 747                 }
 748         }
 749
 750         // 3. Advance snd.una
 751
 752         uint32_t advanced = seqdiff(hdr.ack, c->snd.una);
 753         prevrcvnxt = c->rcv.nxt;
 754
 755         if(advanced) {
 756                 int32_t data_acked = advanced;
 757
 758                 switch(c->state) {
 759                         case SYN_SENT:
 760                         case SYN_RECEIVED:
 761                                 data_acked--;
 762                                 break;
 763                         // TODO: handle FIN as well.
 764                         default:
 765                                 break;
 766                 }
 767
 768                 assert(data_acked >= 0);
 769
 770                 int32_t bufused = seqdiff(c->snd.last, c->snd.una);
 771                 assert(data_acked <= bufused);
 772
 773                 if(data_acked)
 774                         buffer_get(&c->sndbuf, NULL, data_acked);
 775
 776                 c->snd.una = hdr.ack;
 777
 778                 c->dupack = 0;
 779                 c->snd.cwnd += utcp->mtu;
 780                 if(c->snd.cwnd > c->sndbuf.maxsize)
 781                         c->snd.cwnd = c->sndbuf.maxsize;
 782
 783                 // Check if we have sent a FIN that is now ACKed.
 784                 switch(c->state) {
 785                 case FIN_WAIT_1:
 786                         if(c->snd.una == c->snd.last)
 787                                 set_state(c, FIN_WAIT_2);
 788                         break;
 789                 case CLOSING:
 790                         if(c->snd.una == c->snd.last) {
 791                                 gettimeofday(&c->conn_timeout, NULL);
 792                                 c->conn_timeout.tv_sec += 60;
 793                                 set_state(c, TIME_WAIT);
 794                         }
 795                         break;
 796                 default:
 797                         break;
 798                 }
 799         } else {
 800                 if(!len) {
 801                         c->dupack++;
 802                         if(c->dupack == 3) {
 803                                 debug("Triplicate ACK\n");
 804                                 //TODO: Resend one packet and go to fast recovery mode. See RFC 6582.
 805                                 //We do a very simple variant here; reset the nxt pointer to the last acknowledged packet from the peer.
 806                                 //This will cause us to start retransmitting, but at the same speed as the incoming ACKs arrive,
 807                                 //thus preventing a drop in speed.
 808                                 c->snd.nxt = c->snd.una;
 809                         }
 810                 }
 811         }
 812
 813         // 4. Update timers
 814
 815         if(advanced) {
 816                 timerclear(&c->conn_timeout); // It will be set anew in utcp_timeout() if c->snd.una != c->snd.nxt.
 817                 if(c->snd.una == c->snd.nxt)
 818                         timerclear(&c->rtrx_timeout);
 819         }
 820
 821         // 5. Process SYN stuff
 822
 823         if(hdr.ctl & SYN) {
 824                 switch(c->state) {
 825                 case SYN_SENT:
 826                         // This is a SYNACK. It should always have ACKed the SYN.
 827                         if(!advanced)
 828                                 goto reset;
 829                         c->rcv.irs = hdr.seq;
 830                         c->rcv.nxt = hdr.seq;
 831                         set_state(c, ESTABLISHED);
 832                         // TODO: notify application of this somehow.
 833                         break;
 834                 case SYN_RECEIVED:
 835                 case ESTABLISHED:
 836                 case FIN_WAIT_1:
 837                 case FIN_WAIT_2:
 838                 case CLOSE_WAIT:
 839                 case CLOSING:
 840                 case LAST_ACK:
 841                 case TIME_WAIT:
 842                         // Ehm, no. We should never receive a second SYN.
 843                         goto reset;
 844                 default:
 845 #ifdef UTCP_DEBUG
 846                         abort();
 847 #endif
 848                         return 0;
 849                 }
 850
 851                 // SYN counts as one sequence number
 852                 c->rcv.nxt++;
 853         }
 854
 855         // 6. Process new data
 856
 857         if(c->state == SYN_RECEIVED) {
 858                 // This is the ACK after the SYNACK. It should always have ACKed the SYNACK.
 859                 if(!advanced)
 860                         goto reset;
 861
 862                 // Are we still LISTENing?
 863                 if(utcp->accept)
 864                         utcp->accept(c, c->src);
 865
 866                 if(c->state != ESTABLISHED) {
 867                         set_state(c, CLOSED);
 868                         c->reapable = true;
 869                         goto reset;
 870                 }
 871         }
 872
 873         if(len) {
 874                 switch(c->state) {
 875                 case SYN_SENT:
 876                 case SYN_RECEIVED:
 877                         // This should never happen.
 878 #ifdef UTCP_DEBUG
 879                         abort();
 880 #endif
 881                         return 0;
 882                 case ESTABLISHED:
 883                 case FIN_WAIT_1:
 884                 case FIN_WAIT_2:
 885                         break;
 886                 case CLOSE_WAIT:
 887                 case CLOSING:
 888                 case LAST_ACK:
 889                 case TIME_WAIT:
 890                         // Ehm no, We should never receive more data after a FIN.
 891                         goto reset;
 892                 default:
 893 #ifdef UTCP_DEBUG
 894                         abort();
 895 #endif
 896                         return 0;
 897                 }
 898
 899                 ssize_t rxd;
 900
 901                 if(c->recv) {
 902                         rxd = c->recv(c, data, len);
 903                         if(rxd != len) {
 904                                 // TODO: once we have a receive buffer, handle the application not accepting all data.
 905                                 abort();
 906                         }
 907                         if(rxd < 0)
 908                                 rxd = 0;
 909                         else if(rxd > len)
 910                                 rxd = len; // Bad application, bad!
 911                 } else {
 912                         rxd = len;
 913                 }
 914
 915                 c->rcv.nxt += len;
 916         }
 917
 918         // 7. Process FIN stuff
 919
 920         if(hdr.ctl & FIN) {
 921                 switch(c->state) {
 922                 case SYN_SENT:
 923                 case SYN_RECEIVED:
 924                         // This should never happen.
 925 #ifdef UTCP_DEBUG
 926                         abort();
 927 #endif
 928                         break;
 929                 case ESTABLISHED:
 930                         set_state(c, CLOSE_WAIT);
 931                         break;
 932                 case FIN_WAIT_1:
 933                         set_state(c, CLOSING);
 934                         break;
 935                 case FIN_WAIT_2:
 936                         gettimeofday(&c->conn_timeout, NULL);
 937                         c->conn_timeout.tv_sec += 60;
 938                         set_state(c, TIME_WAIT);
 939                         break;
 940                 case CLOSE_WAIT:
 941                 case CLOSING:
 942                 case LAST_ACK:
 943                 case TIME_WAIT:
 944                         // Ehm, no. We should never receive a second FIN.
 945                         goto reset;
 946                 default:
 947 #ifdef UTCP_DEBUG
 948                         abort();
 949 #endif
 950                         break;
 951                 }
 952
 953                 // FIN counts as one sequence number
 954                 c->rcv.nxt++;
 955                 len++;
 956
 957                 // Inform the application that the peer closed the connection.
 958                 if(c->recv) {
 959                         errno = 0;
 960                         c->recv(c, NULL, 0);
 961                 }
 962         }
 963
 964         // Now we send something back if:
 965         // - we advanced rcv.nxt (ie, we got some data that needs to be ACKed)
 966         //   -> sendatleastone = true
 967         // - or we got an ack, so we should maybe send a bit more data
 968         //   -> sendatleastone = false
 969
 970 ack:
 971         ack(c, prevrcvnxt != c->rcv.nxt);
 972         return 0;
 973
 974 reset:
 975         swap_ports(&hdr);
 976         hdr.wnd = 0;
 977         if(hdr.ctl & ACK) {
 978                 hdr.seq = hdr.ack;
 979                 hdr.ctl = RST;
 980         } else {
 981                 hdr.ack = hdr.seq + len;
 982                 hdr.seq = 0;
 983                 hdr.ctl = RST | ACK;
 984         }
 985         print_packet(utcp, "send", &hdr, sizeof hdr);
 986         utcp->send(utcp, &hdr, sizeof hdr);
 987         return 0;
 988
 989 }
 990
 991 int utcp_shutdown(struct utcp_connection *c, int dir) {
 992         debug("%p shutdown %d at %u\n", c ? c->utcp : NULL, dir, c ? c->snd.last : 0);
 993         if(!c) {
 994                 errno = EFAULT;
 995                 return -1;
 996         }
 997
 998         if(c->reapable) {
 999                 debug("Error: shutdown() called on closed connection %p\n", c);
1000                 errno = EBADF;
1001                 return -1;
1002         }
1003
1004         if(!(dir == UTCP_SHUT_RD || dir == UTCP_SHUT_WR || dir == UTCP_SHUT_RDWR)) {
1005                 errno = EINVAL;
1006                 return -1;
1007         }
1008
1009         // TCP does not have a provision for stopping incoming packets.
1010         // The best we can do is to just ignore them.
1011         if(dir == UTCP_SHUT_RD || dir == UTCP_SHUT_RDWR)
1012                 c->recv = NULL;
1013
1014         // The rest of the code deals with shutting down writes.
1015         if(dir == UTCP_SHUT_RD)
1016                 return 0;
1017
1018         switch(c->state) {
1019         case CLOSED:
1020         case LISTEN:
1021                 errno = ENOTCONN;
1022                 return -1;
1023
1024         case SYN_SENT:
1025                 set_state(c, CLOSED);
1026                 return 0;
1027
1028         case SYN_RECEIVED:
1029         case ESTABLISHED:
1030                 set_state(c, FIN_WAIT_1);
1031                 break;
1032         case FIN_WAIT_1:
1033         case FIN_WAIT_2:
1034                 return 0;
1035         case CLOSE_WAIT:
1036                 set_state(c, CLOSING);
1037                 break;
1038
1039         case CLOSING:
1040         case LAST_ACK:
1041         case TIME_WAIT:
1042                 return 0;
1043         }
1044
1045         c->snd.last++;
1046
1047         ack(c, false);
1048         return 0;
1049 }
1050
1051 int utcp_close(struct utcp_connection *c) {
1052         if(utcp_shutdown(c, SHUT_RDWR))
1053                 return -1;
1054         c->recv = NULL;
1055         c->poll = NULL;
1056         c->reapable = true;
1057         return 0;
1058 }
1059
1060 int utcp_abort(struct utcp_connection *c) {
1061         if(!c) {
1062                 errno = EFAULT;
1063                 return -1;
1064         }
1065
1066         if(c->reapable) {
1067                 debug("Error: abort() called on closed connection %p\n", c);
1068                 errno = EBADF;
1069                 return -1;
1070         }
1071
1072         c->recv = NULL;
1073         c->poll = NULL;
1074         c->reapable = true;
1075
1076         switch(c->state) {
1077         case CLOSED:
1078                 return 0;
1079         case LISTEN:
1080         case SYN_SENT:
1081         case CLOSING:
1082         case LAST_ACK:
1083         case TIME_WAIT:
1084                 set_state(c, CLOSED);
1085                 return 0;
1086
1087         case SYN_RECEIVED:
1088         case ESTABLISHED:
1089         case FIN_WAIT_1:
1090         case FIN_WAIT_2:
1091         case CLOSE_WAIT:
1092                 set_state(c, CLOSED);
1093                 break;
1094         }
1095
1096         // Send RST
1097
1098         struct hdr hdr;
1099
1100         hdr.src = c->src;
1101         hdr.dst = c->dst;
1102         hdr.seq = c->snd.nxt;
1103         hdr.ack = 0;
1104         hdr.wnd = 0;
1105         hdr.ctl = RST;
1106
1107         print_packet(c->utcp, "send", &hdr, sizeof hdr);
1108         c->utcp->send(c->utcp, &hdr, sizeof hdr);
1109         return 0;
1110 }
1111
1112 /* Handle timeouts.
1113  * One call to this function will loop through all connections,
1114  * checking if something needs to be resent or not.
1115  * The return value is the time to the next timeout in milliseconds,
1116  * or maybe a negative value if the timeout is infinite.
1117  */
1118 struct timeval utcp_timeout(struct utcp *utcp) {
1119         struct timeval now;
1120         gettimeofday(&now, NULL);
1121         struct timeval next = {now.tv_sec + 3600, now.tv_usec};
1122
1123         for(int i = 0; i < utcp->nconnections; i++) {
1124                 struct utcp_connection *c = utcp->connections[i];
1125                 if(!c)
1126                         continue;
1127
1128                 if(c->state == CLOSED) {
1129                         if(c->reapable) {
1130                                 debug("Reaping %p\n", c);
1131                                 free_connection(c);
1132                                 i--;
1133                         }
1134                         continue;
1135                 }
1136
1137                 if(timerisset(&c->conn_timeout) && timercmp(&c->conn_timeout, &now, <)) {
1138                         errno = ETIMEDOUT;
1139                         c->state = CLOSED;
1140                         if(c->recv)
1141                                 c->recv(c, NULL, 0);
1142                         continue;
1143                 }
1144
1145                 if(timerisset(&c->rtrx_timeout) && timercmp(&c->rtrx_timeout, &now, <)) {
1146                         retransmit(c);
1147                 }
1148
1149                 if(c->poll && buffer_free(&c->sndbuf) && (c->state == ESTABLISHED || c->state == CLOSE_WAIT))
1150                         c->poll(c, buffer_free(&c->sndbuf));
1151
1152                 if(timerisset(&c->conn_timeout) && timercmp(&c->conn_timeout, &next, <))
1153                         next = c->conn_timeout;
1154
1155                 if(c->snd.nxt != c->snd.una) {
1156                         c->rtrx_timeout = now;
1157                         c->rtrx_timeout.tv_sec++;
1158                 } else {
1159                         timerclear(&c->rtrx_timeout);
1160                 }
1161
1162                 if(timerisset(&c->rtrx_timeout) && timercmp(&c->rtrx_timeout, &next, <))
1163                         next = c->rtrx_timeout;
1164         }
1165
1166         struct timeval diff;
1167         timersub(&next, &now, &diff);
1168         return diff;
1169 }
1170
1171 bool utcp_is_active(struct utcp *utcp) {
1172         if(!utcp)
1173                 return false;
1174
1175         for(int i = 0; i < utcp->nconnections; i++)
1176                 if(utcp->connections[i]->state != CLOSED && utcp->connections[i]->state != TIME_WAIT)
1177                         return true;
1178
1179         return false;
1180 }
1181
1182 struct utcp *utcp_init(utcp_accept_t accept, utcp_pre_accept_t pre_accept, utcp_send_t send, void *priv) {
1183         struct utcp *utcp = calloc(1, sizeof *utcp);
1184         if(!utcp)
1185                 return NULL;
1186
1187         if(!send) {
1188                 errno = EFAULT;
1189                 return NULL;
1190         }
1191
1192         utcp->accept = accept;
1193         utcp->pre_accept = pre_accept;
1194         utcp->send = send;
1195         utcp->priv = priv;
1196         utcp->mtu = 1000;
1197         utcp->timeout = 60;
1198
1199         return utcp;
1200 }
1201
1202 void utcp_exit(struct utcp *utcp) {
1203         if(!utcp)
1204                 return;
1205         for(int i = 0; i < utcp->nconnections; i++) {
1206                 if(!utcp->connections[i]->reapable)
1207                         debug("Warning, freeing unclosed connection %p\n", utcp->connections[i]);
1208                 buffer_exit(&utcp->connections[i]->sndbuf);
1209                 free(utcp->connections[i]);
1210         }
1211         free(utcp->connections);
1212         free(utcp);
1213 }
1214
1215 uint16_t utcp_get_mtu(struct utcp *utcp) {
1216         return utcp ? utcp->mtu : 0;
1217 }
1218
1219 void utcp_set_mtu(struct utcp *utcp, uint16_t mtu) {
1220         // TODO: handle overhead of the header
1221         if(utcp)
1222                 utcp->mtu = mtu;
1223 }
1224
1225 int utcp_get_user_timeout(struct utcp *u) {
1226         return u ? u->timeout : 0;
1227 }
1228
1229 void utcp_set_user_timeout(struct utcp *u, int timeout) {
1230         if(u)
1231                 u->timeout = timeout;
1232 }
1233
1234 size_t utcp_get_sndbuf(struct utcp_connection *c) {
1235         return c ? c->sndbuf.maxsize : 0;
1236 }
1237
1238 size_t utcp_get_sndbuf_free(struct utcp_connection *c) {
1239         if(c && (c->state == ESTABLISHED || c->state == CLOSE_WAIT))
1240                 return buffer_free(&c->sndbuf);
1241         else
1242                 return 0;
1243 }
1244
1245 void utcp_set_sndbuf(struct utcp_connection *c, size_t size) {
1246         if(!c)
1247                 return;
1248         c->sndbuf.maxsize = size;
1249         if(c->sndbuf.maxsize != size)
1250                 c->sndbuf.maxsize = -1;
1251 }
1252
1253 bool utcp_get_nodelay(struct utcp_connection *c) {
1254         return c ? c->nodelay : false;
1255 }
1256
1257 void utcp_set_nodelay(struct utcp_connection *c, bool nodelay) {
1258         if(c)
1259                 c->nodelay = nodelay;
1260 }
1261
1262 bool utcp_get_keepalive(struct utcp_connection *c) {
1263         return c ? c->keepalive : false;
1264 }
1265
1266 void utcp_set_keepalive(struct utcp_connection *c, bool keepalive) {
1267         if(c)
1268                 c->keepalive = keepalive;
1269 }
1270
1271 size_t utcp_get_outq(struct utcp_connection *c) {
1272         return c ? seqdiff(c->snd.nxt, c->snd.una) : 0;
1273 }
1274
1275 void utcp_set_recv_cb(struct utcp_connection *c, utcp_recv_t recv) {
1276         if(c)
1277                 c->recv = recv;
1278 }
1279
1280 void utcp_set_poll_cb(struct utcp_connection *c, utcp_poll_t poll) {
1281         if(c)
1282                 c->poll = poll;
1283 }
1284
1285 void utcp_set_accept_cb(struct utcp *utcp, utcp_accept_t accept, utcp_pre_accept_t pre_accept) {
1286         if(utcp) {
1287                 utcp->accept = accept;
1288                 utcp->pre_accept = pre_accept;
1289         }
1290 }