git.meshlink.io Git - utcp/blob - utcp.c

   1 /*
   2     utcp.c -- Userspace TCP
   3     Copyright (C) 2014 Guus Sliepen <guus@tinc-vpn.org>
   4
   5     This program is free software; you can redistribute it and/or modify
   6     it under the terms of the GNU General Public License as published by
   7     the Free Software Foundation; either version 2 of the License, or
   8     (at your option) any later version.
   9
  10     This program is distributed in the hope that it will be useful,
  11     but WITHOUT ANY WARRANTY; without even the implied warranty of
  12     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13     GNU General Public License for more details.
  14
  15     You should have received a copy of the GNU General Public License along
  16     with this program; if not, write to the Free Software Foundation, Inc.,
  17     51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
  18 */
  19
  20 #define _GNU_SOURCE
  21
  22 #include <assert.h>
  23 #include <errno.h>
  24 #include <stdio.h>
  25 #include <stdlib.h>
  26 #include <stdint.h>
  27 #include <stdbool.h>
  28 #include <string.h>
  29 #include <unistd.h>
  30 #include <sys/time.h>
  31 #include <sys/socket.h>
  32
  33 #include "utcp_priv.h"
  34
  35 #ifndef EBADMSG
  36 #define EBADMSG         104
  37 #endif
  38
  39 #ifndef SHUT_RDWR
  40 #define SHUT_RDWR 2
  41 #endif
  42
  43 #ifdef poll
  44 #undef poll
  45 #endif
  46
  47 #ifndef timersub
  48 #define timersub(a, b, r) do {\
  49         (r)->tv_sec = (a)->tv_sec - (b)->tv_sec;\
  50         (r)->tv_usec = (a)->tv_usec - (b)->tv_usec;\
  51         if((r)->tv_usec < 0)\
  52                 (r)->tv_sec--, (r)->tv_usec += 1000000;\
  53 } while (0)
  54 #endif
  55
  56 #ifdef UTCP_DEBUG
  57 #include <stdarg.h>
  58
  59 static void debug(const char *format, ...) {
  60         va_list ap;
  61         va_start(ap, format);
  62         vfprintf(stderr, format, ap);
  63         va_end(ap);
  64 }
  65
  66 static void print_packet(struct utcp *utcp, const char *dir, const void *pkt, size_t len) {
  67         struct hdr hdr;
  68         if(len < sizeof hdr) {
  69                 debug("%p %s: short packet (%zu bytes)\n", utcp, dir, len);
  70                 return;
  71         }
  72
  73         memcpy(&hdr, pkt, sizeof hdr);
  74         fprintf (stderr, "%p %s: len=%zu, src=%u dst=%u seq=%u ack=%u wnd=%u ctl=", utcp, dir, len, hdr.src, hdr.dst, hdr.seq, hdr.ack, hdr.wnd);
  75         if(hdr.ctl & SYN)
  76                 debug("SYN");
  77         if(hdr.ctl & RST)
  78                 debug("RST");
  79         if(hdr.ctl & FIN)
  80                 debug("FIN");
  81         if(hdr.ctl & ACK)
  82                 debug("ACK");
  83
  84         if(len > sizeof hdr) {
  85                 debug(" data=");
  86                 for(int i = sizeof hdr; i < len; i++) {
  87                         const char *data = pkt;
  88                         debug("%c", data[i] >= 32 ? data[i] : '.');
  89                 }
  90         }
  91
  92         debug("\n");
  93 }
  94 #else
  95 #define debug(...)
  96 #define print_packet(...)
  97 #endif
  98
  99 static void set_state(struct utcp_connection *c, enum state state) {
 100         c->state = state;
 101         if(state == ESTABLISHED)
 102                 timerclear(&c->conn_timeout);
 103         debug("%p new state: %s\n", c->utcp, strstate[state]);
 104 }
 105
 106 static inline void list_connections(struct utcp *utcp) {
 107         debug("%p has %d connections:\n", utcp, utcp->nconnections);
 108         for(int i = 0; i < utcp->nconnections; i++)
 109                 debug("  %u -> %u state %s\n", utcp->connections[i]->src, utcp->connections[i]->dst, strstate[utcp->connections[i]->state]);
 110 }
 111
 112 static int32_t seqdiff(uint32_t a, uint32_t b) {
 113         return a - b;
 114 }
 115
 116 // Buffer functions
 117 // TODO: convert to ringbuffers to avoid memmove() operations.
 118
 119 // Store data into the buffer
 120 static ssize_t buffer_put(struct buffer *buf, const void *data, size_t len) {
 121         if(buf->maxsize <= buf->used)
 122                 return 0;
 123         if(len > buf->maxsize - buf->used)
 124                 len = buf->maxsize - buf->used;
 125         if(len > buf->size - buf->used) {
 126                 size_t newsize = buf->size;
 127                 do {
 128                         newsize *= 2;
 129                 } while(newsize < buf->used + len);
 130                 if(newsize > buf->maxsize)
 131                         newsize = buf->maxsize;
 132                 char *newdata = realloc(buf->data, newsize);
 133                 if(!newdata)
 134                         return -1;
 135                 buf->data = newdata;
 136                 buf->size = newsize;
 137         }
 138         memcpy(buf->data + buf->used, data, len);
 139         buf->used += len;
 140         return len;
 141 }
 142
 143 // Get data from the buffer. data can be NULL.
 144 static ssize_t buffer_get(struct buffer *buf, void *data, size_t len) {
 145         if(len > buf->used)
 146                 len = buf->used;
 147         if(data)
 148                 memcpy(data, buf->data, len);
 149         if(len < buf->used)
 150                 memmove(buf->data, buf->data + len, buf->used - len);
 151         buf->used -= len;
 152         return len;
 153 }
 154
 155 // Copy data from the buffer without removing it.
 156 static ssize_t buffer_copy(struct buffer *buf, void *data, size_t offset, size_t len) {
 157         if(offset >= buf->used)
 158                 return 0;
 159         if(offset + len > buf->used)
 160                 len = buf->used - offset;
 161         memcpy(data, buf->data + offset, len);
 162         return len;
 163 }
 164
 165 static bool buffer_init(struct buffer *buf, uint32_t len, uint32_t maxlen) {
 166         memset(buf, 0, sizeof *buf);
 167         buf->data = malloc(len);
 168         if(!len)
 169                 return false;
 170         buf->size = len;
 171         buf->maxsize = maxlen;
 172         return true;
 173 }
 174
 175 static void buffer_exit(struct buffer *buf) {
 176         free(buf->data);
 177         memset(buf, 0, sizeof *buf);
 178 }
 179
 180 static uint32_t buffer_free(const struct buffer *buf) {
 181         return buf->maxsize - buf->used;
 182 }
 183
 184 // Connections are stored in a sorted list.
 185 // This gives O(log(N)) lookup time, O(N log(N)) insertion time and O(N) deletion time.
 186
 187 static int compare(const void *va, const void *vb) {
 188         assert(va && vb);
 189
 190         const struct utcp_connection *a = *(struct utcp_connection **)va;
 191         const struct utcp_connection *b = *(struct utcp_connection **)vb;
 192
 193         assert(a && b);
 194         assert(a->src && b->src);
 195
 196         int c = (int)a->src - (int)b->src;
 197         if(c)
 198                 return c;
 199         c = (int)a->dst - (int)b->dst;
 200         return c;
 201 }
 202
 203 static struct utcp_connection *find_connection(const struct utcp *utcp, uint16_t src, uint16_t dst) {
 204         if(!utcp->nconnections)
 205                 return NULL;
 206         struct utcp_connection key = {
 207                 .src = src,
 208                 .dst = dst,
 209         }, *keyp = &key;
 210         struct utcp_connection **match = bsearch(&keyp, utcp->connections, utcp->nconnections, sizeof *utcp->connections, compare);
 211         return match ? *match : NULL;
 212 }
 213
 214 static void free_connection(struct utcp_connection *c) {
 215         struct utcp *utcp = c->utcp;
 216         struct utcp_connection **cp = bsearch(&c, utcp->connections, utcp->nconnections, sizeof *utcp->connections, compare);
 217
 218         assert(cp);
 219
 220         int i = cp - utcp->connections;
 221         memmove(cp, cp + 1, (utcp->nconnections - i - 1) * sizeof *cp);
 222         utcp->nconnections--;
 223
 224         buffer_exit(&c->sndbuf);
 225         free(c);
 226 }
 227
 228 static struct utcp_connection *allocate_connection(struct utcp *utcp, uint16_t src, uint16_t dst) {
 229         // Check whether this combination of src and dst is free
 230
 231         if(src) {
 232                 if(find_connection(utcp, src, dst)) {
 233                         errno = EADDRINUSE;
 234                         return NULL;
 235                 }
 236         } else { // If src == 0, generate a random port number with the high bit set
 237                 if(utcp->nconnections >= 32767) {
 238                         errno = ENOMEM;
 239                         return NULL;
 240                 }
 241                 src = rand() | 0x8000;
 242                 while(find_connection(utcp, src, dst))
 243                         src++;
 244         }
 245
 246         // Allocate memory for the new connection
 247
 248         if(utcp->nconnections >= utcp->nallocated) {
 249                 if(!utcp->nallocated)
 250                         utcp->nallocated = 4;
 251                 else
 252                         utcp->nallocated *= 2;
 253                 struct utcp_connection **new_array = realloc(utcp->connections, utcp->nallocated * sizeof *utcp->connections);
 254                 if(!new_array)
 255                         return NULL;
 256                 utcp->connections = new_array;
 257         }
 258
 259         struct utcp_connection *c = calloc(1, sizeof *c);
 260         if(!c)
 261                 return NULL;
 262
 263         if(!buffer_init(&c->sndbuf, DEFAULT_SNDBUFSIZE, DEFAULT_MAXSNDBUFSIZE)) {
 264                 free(c);
 265                 return NULL;
 266         }
 267
 268         // Fill in the details
 269
 270         c->src = src;
 271         c->dst = dst;
 272         c->snd.iss = rand();
 273         c->snd.una = c->snd.iss;
 274         c->snd.nxt = c->snd.iss + 1;
 275         c->rcv.wnd = utcp->mtu;
 276         c->snd.last = c->snd.nxt;
 277         c->snd.cwnd = utcp->mtu;
 278         c->utcp = utcp;
 279
 280         // Add it to the sorted list of connections
 281
 282         utcp->connections[utcp->nconnections++] = c;
 283         qsort(utcp->connections, utcp->nconnections, sizeof *utcp->connections, compare);
 284
 285         return c;
 286 }
 287
 288 struct utcp_connection *utcp_connect(struct utcp *utcp, uint16_t dst, utcp_recv_t recv, void *priv) {
 289         struct utcp_connection *c = allocate_connection(utcp, 0, dst);
 290         if(!c)
 291                 return NULL;
 292
 293         c->recv = recv;
 294         c->priv = priv;
 295
 296         struct hdr hdr;
 297
 298         hdr.src = c->src;
 299         hdr.dst = c->dst;
 300         hdr.seq = c->snd.iss;
 301         hdr.ack = 0;
 302         hdr.wnd = c->rcv.wnd;
 303         hdr.ctl = SYN;
 304         hdr.aux = 0;
 305
 306         set_state(c, SYN_SENT);
 307
 308         print_packet(utcp, "send", &hdr, sizeof hdr);
 309         utcp->send(utcp, &hdr, sizeof hdr);
 310
 311         gettimeofday(&c->conn_timeout, NULL);
 312         c->conn_timeout.tv_sec += utcp->timeout;
 313
 314         return c;
 315 }
 316
 317 void utcp_accept(struct utcp_connection *c, utcp_recv_t recv, void *priv) {
 318         if(c->reapable || c->state != SYN_RECEIVED) {
 319                 debug("Error: accept() called on invalid connection %p in state %s\n", c, strstate[c->state]);
 320                 return;
 321         }
 322
 323         debug("%p accepted, %p %p\n", c, recv, priv);
 324         c->recv = recv;
 325         c->priv = priv;
 326         set_state(c, ESTABLISHED);
 327 }
 328
 329 static void ack(struct utcp_connection *c, bool sendatleastone) {
 330         int32_t left = seqdiff(c->snd.last, c->snd.nxt);
 331         int32_t cwndleft = c->snd.cwnd - seqdiff(c->snd.nxt, c->snd.una);
 332
 333         assert(left >= 0);
 334
 335         if(cwndleft <= 0)
 336                 cwndleft = 0;
 337
 338         if(cwndleft < left)
 339                 left = cwndleft;
 340
 341         if(!left && !sendatleastone)
 342                 return;
 343
 344         struct {
 345                 struct hdr hdr;
 346                 char data[];
 347         } *pkt;
 348
 349         pkt = malloc(sizeof pkt->hdr + c->utcp->mtu);
 350         if(!pkt->data)
 351                 return;
 352
 353         pkt->hdr.src = c->src;
 354         pkt->hdr.dst = c->dst;
 355         pkt->hdr.ack = c->rcv.nxt;
 356         pkt->hdr.wnd = c->snd.wnd;
 357         pkt->hdr.ctl = ACK;
 358         pkt->hdr.aux = 0;
 359
 360         do {
 361                 uint32_t seglen = left > c->utcp->mtu ? c->utcp->mtu : left;
 362                 pkt->hdr.seq = c->snd.nxt;
 363
 364                 buffer_copy(&c->sndbuf, pkt->data, seqdiff(c->snd.nxt, c->snd.una), seglen);
 365
 366                 c->snd.nxt += seglen;
 367                 left -= seglen;
 368
 369                 if(c->state != ESTABLISHED && !left && seglen) {
 370                         switch(c->state) {
 371                         case FIN_WAIT_1:
 372                         case CLOSING:
 373                                 seglen--;
 374                                 pkt->hdr.ctl |= FIN;
 375                                 break;
 376                         default:
 377                                 break;
 378                         }
 379                 }
 380
 381                 print_packet(c->utcp, "send", pkt, sizeof pkt->hdr + seglen);
 382                 c->utcp->send(c->utcp, pkt, sizeof pkt->hdr + seglen);
 383         } while(left);
 384
 385         free(pkt);
 386 }
 387
 388 ssize_t utcp_send(struct utcp_connection *c, const void *data, size_t len) {
 389         if(c->reapable) {
 390                 debug("Error: send() called on closed connection %p\n", c);
 391                 errno = EBADF;
 392                 return -1;
 393         }
 394
 395         switch(c->state) {
 396         case CLOSED:
 397         case LISTEN:
 398         case SYN_SENT:
 399         case SYN_RECEIVED:
 400                 debug("Error: send() called on unconnected connection %p\n", c);
 401                 errno = ENOTCONN;
 402                 return -1;
 403         case ESTABLISHED:
 404         case CLOSE_WAIT:
 405                 break;
 406         case FIN_WAIT_1:
 407         case FIN_WAIT_2:
 408         case CLOSING:
 409         case LAST_ACK:
 410         case TIME_WAIT:
 411                 debug("Error: send() called on closing connection %p\n", c);
 412                 errno = EPIPE;
 413                 return -1;
 414         }
 415
 416         // Add data to send buffer
 417
 418         if(!len)
 419                 return 0;
 420
 421         if(!data) {
 422                 errno = EFAULT;
 423                 return -1;
 424         }
 425
 426         len = buffer_put(&c->sndbuf, data, len);
 427         if(len <= 0) {
 428                 errno = EWOULDBLOCK;
 429                 return 0;
 430         }
 431
 432         c->snd.last += len;
 433         ack(c, false);
 434         return len;
 435 }
 436
 437 static void swap_ports(struct hdr *hdr) {
 438         uint16_t tmp = hdr->src;
 439         hdr->src = hdr->dst;
 440         hdr->dst = tmp;
 441 }
 442
 443 static void retransmit(struct utcp_connection *c) {
 444         if(c->state == CLOSED || c->snd.nxt == c->snd.una)
 445                 return;
 446
 447         struct utcp *utcp = c->utcp;
 448
 449         struct {
 450                 struct hdr hdr;
 451                 char data[];
 452         } *pkt;
 453
 454         pkt = malloc(sizeof pkt->hdr + c->utcp->mtu);
 455         if(!pkt)
 456                 return;
 457
 458         pkt->hdr.src = c->src;
 459         pkt->hdr.dst = c->dst;
 460
 461         switch(c->state) {
 462                 case LISTEN:
 463                         // TODO: this should not happen
 464                         break;
 465
 466                 case SYN_SENT:
 467                         pkt->hdr.seq = c->snd.iss;
 468                         pkt->hdr.ack = 0;
 469                         pkt->hdr.wnd = c->rcv.wnd;
 470                         pkt->hdr.ctl = SYN;
 471                         print_packet(c->utcp, "rtrx", pkt, sizeof pkt->hdr);
 472                         utcp->send(utcp, pkt, sizeof pkt->hdr);
 473                         break;
 474
 475                 case SYN_RECEIVED:
 476                         pkt->hdr.seq = c->snd.nxt;
 477                         pkt->hdr.ack = c->rcv.nxt;
 478                         pkt->hdr.ctl = SYN | ACK;
 479                         print_packet(c->utcp, "rtrx", pkt, sizeof pkt->hdr);
 480                         utcp->send(utcp, pkt, sizeof pkt->hdr);
 481                         break;
 482
 483                 case ESTABLISHED:
 484                 case FIN_WAIT_1:
 485                         pkt->hdr.seq = c->snd.una;
 486                         pkt->hdr.ack = c->rcv.nxt;
 487                         pkt->hdr.ctl = ACK;
 488                         uint32_t len = seqdiff(c->snd.nxt, c->snd.una);
 489                         if(c->state == FIN_WAIT_1)
 490                                 len--;
 491                         if(len > utcp->mtu)
 492                                 len = utcp->mtu;
 493                         else {
 494                                 if(c->state == FIN_WAIT_1)
 495                                         pkt->hdr.ctl |= FIN;
 496                         }
 497                         buffer_copy(&c->sndbuf, pkt->data, 0, len);
 498                         print_packet(c->utcp, "rtrx", pkt, sizeof pkt->hdr + len);
 499                         utcp->send(utcp, pkt, sizeof pkt->hdr + len);
 500                         break;
 501
 502                 default:
 503                         // TODO: implement
 504                         abort();
 505         }
 506
 507         free(pkt);
 508 }
 509
 510
 511 ssize_t utcp_recv(struct utcp *utcp, const void *data, size_t len) {
 512         if(!utcp) {
 513                 errno = EFAULT;
 514                 return -1;
 515         }
 516
 517         if(!len)
 518                 return 0;
 519
 520         if(!data) {
 521                 errno = EFAULT;
 522                 return -1;
 523         }
 524
 525         print_packet(utcp, "recv", data, len);
 526
 527         // Drop packets smaller than the header
 528
 529         struct hdr hdr;
 530         if(len < sizeof hdr) {
 531                 errno = EBADMSG;
 532                 return -1;
 533         }
 534
 535         // Make a copy from the potentially unaligned data to a struct hdr
 536
 537         memcpy(&hdr, data, sizeof hdr);
 538         data += sizeof hdr;
 539         len -= sizeof hdr;
 540
 541         // Drop packets with an unknown CTL flag
 542
 543         if(hdr.ctl & ~(SYN | ACK | RST | FIN)) {
 544                 errno = EBADMSG;
 545                 return -1;
 546         }
 547
 548         // Try to match the packet to an existing connection
 549
 550         struct utcp_connection *c = find_connection(utcp, hdr.dst, hdr.src);
 551
 552         // Is it for a new connection?
 553
 554         if(!c) {
 555                 // Ignore RST packets
 556
 557                 if(hdr.ctl & RST)
 558                         return 0;
 559
 560                 // Is it a SYN packet and are we LISTENing?
 561
 562                 if(hdr.ctl & SYN && !(hdr.ctl & ACK) && utcp->accept) {
 563                         // If we don't want to accept it, send a RST back
 564                         if((utcp->pre_accept && !utcp->pre_accept(utcp, hdr.dst))) {
 565                                 len = 1;
 566                                 goto reset;
 567                         }
 568
 569                         // Try to allocate memory, otherwise send a RST back
 570                         c = allocate_connection(utcp, hdr.dst, hdr.src);
 571                         if(!c) {
 572                                 len = 1;
 573                                 goto reset;
 574                         }
 575
 576                         // Return SYN+ACK, go to SYN_RECEIVED state
 577                         c->snd.wnd = hdr.wnd;
 578                         c->rcv.irs = hdr.seq;
 579                         c->rcv.nxt = c->rcv.irs + 1;
 580                         set_state(c, SYN_RECEIVED);
 581
 582                         hdr.dst = c->dst;
 583                         hdr.src = c->src;
 584                         hdr.ack = c->rcv.irs + 1;
 585                         hdr.seq = c->snd.iss;
 586                         hdr.ctl = SYN | ACK;
 587                         print_packet(c->utcp, "send", &hdr, sizeof hdr);
 588                         utcp->send(utcp, &hdr, sizeof hdr);
 589                 } else {
 590                         // No, we don't want your packets, send a RST back
 591                         len = 1;
 592                         goto reset;
 593                 }
 594
 595                 return 0;
 596         }
 597
 598         debug("%p state %s\n", c->utcp, strstate[c->state]);
 599
 600         // In case this is for a CLOSED connection, ignore the packet.
 601         // TODO: make it so incoming packets can never match a CLOSED connection.
 602
 603         if(c->state == CLOSED)
 604                 return 0;
 605
 606         // It is for an existing connection.
 607
 608         uint32_t prevrcvnxt = c->rcv.nxt;
 609
 610         // 1. Drop invalid packets.
 611
 612         // 1a. Drop packets that should not happen in our current state.
 613
 614         switch(c->state) {
 615         case SYN_SENT:
 616         case SYN_RECEIVED:
 617         case ESTABLISHED:
 618         case FIN_WAIT_1:
 619         case FIN_WAIT_2:
 620         case CLOSE_WAIT:
 621         case CLOSING:
 622         case LAST_ACK:
 623         case TIME_WAIT:
 624                 break;
 625         default:
 626                 abort();
 627         }
 628
 629         // 1b. Drop packets with a sequence number not in our receive window.
 630
 631         bool acceptable;
 632
 633         if(c->state == SYN_SENT)
 634                 acceptable = true;
 635
 636         // TODO: handle packets overlapping c->rcv.nxt.
 637 #if 0
 638         // Only use this when accepting out-of-order packets.
 639         else if(len == 0)
 640                 if(c->rcv.wnd == 0)
 641                         acceptable = hdr.seq == c->rcv.nxt;
 642                 else
 643                         acceptable = (seqdiff(hdr.seq, c->rcv.nxt) >= 0 && seqdiff(hdr.seq, c->rcv.nxt + c->rcv.wnd) < 0);
 644         else
 645                 if(c->rcv.wnd == 0)
 646                         // We don't accept data when the receive window is zero.
 647                         acceptable = false;
 648                 else
 649                         // Both start and end of packet must be within the receive window
 650                         acceptable = (seqdiff(hdr.seq, c->rcv.nxt) >= 0 && seqdiff(hdr.seq, c->rcv.nxt + c->rcv.wnd) < 0)
 651                                 || (seqdiff(hdr.seq + len + 1, c->rcv.nxt) >= 0 && seqdiff(hdr.seq + len - 1, c->rcv.nxt + c->rcv.wnd) < 0);
 652 #else
 653         if(c->state != SYN_SENT)
 654                 acceptable = hdr.seq == c->rcv.nxt;
 655 #endif
 656
 657         if(!acceptable) {
 658                 debug("Packet not acceptable, %u <= %u + %zu < %u\n", c->rcv.nxt, hdr.seq, len, c->rcv.nxt + c->rcv.wnd);
 659                 // Ignore unacceptable RST packets.
 660                 if(hdr.ctl & RST)
 661                         return 0;
 662                 // Otherwise, send an ACK back in the hope things improve.
 663                 ack(c, true);
 664                 return 0;
 665         }
 666
 667         c->snd.wnd = hdr.wnd; // TODO: move below
 668
 669         // 1c. Drop packets with an invalid ACK.
 670         // ackno should not roll back, and it should also not be bigger than snd.nxt.
 671
 672         if(hdr.ctl & ACK && (seqdiff(hdr.ack, c->snd.nxt) > 0 || seqdiff(hdr.ack, c->snd.una) < 0)) {
 673                 debug("Packet ack seqno out of range, %u %u %u\n", hdr.ack, c->snd.una, c->snd.nxt);
 674                 // Ignore unacceptable RST packets.
 675                 if(hdr.ctl & RST)
 676                         return 0;
 677                 goto reset;
 678         }
 679
 680         // 2. Handle RST packets
 681
 682         if(hdr.ctl & RST) {
 683                 switch(c->state) {
 684                 case SYN_SENT:
 685                         if(!(hdr.ctl & ACK))
 686                                 return 0;
 687                         // The peer has refused our connection.
 688                         set_state(c, CLOSED);
 689                         errno = ECONNREFUSED;
 690                         if(c->recv)
 691                                 c->recv(c, NULL, 0);
 692                         return 0;
 693                 case SYN_RECEIVED:
 694                         if(hdr.ctl & ACK)
 695                                 return 0;
 696                         // We haven't told the application about this connection yet. Silently delete.
 697                         free_connection(c);
 698                         return 0;
 699                 case ESTABLISHED:
 700                 case FIN_WAIT_1:
 701                 case FIN_WAIT_2:
 702                 case CLOSE_WAIT:
 703                         if(hdr.ctl & ACK)
 704                                 return 0;
 705                         // The peer has aborted our connection.
 706                         set_state(c, CLOSED);
 707                         errno = ECONNRESET;
 708                         if(c->recv)
 709                                 c->recv(c, NULL, 0);
 710                         return 0;
 711                 case CLOSING:
 712                 case LAST_ACK:
 713                 case TIME_WAIT:
 714                         if(hdr.ctl & ACK)
 715                                 return 0;
 716                         // As far as the application is concerned, the connection has already been closed.
 717                         // If it has called utcp_close() already, we can immediately free this connection.
 718                         if(c->reapable) {
 719                                 free_connection(c);
 720                                 return 0;
 721                         }
 722                         // Otherwise, immediately move to the CLOSED state.
 723                         set_state(c, CLOSED);
 724                         return 0;
 725                 default:
 726                         abort();
 727                 }
 728         }
 729
 730         // 3. Advance snd.una
 731
 732         uint32_t advanced = seqdiff(hdr.ack, c->snd.una);
 733         prevrcvnxt = c->rcv.nxt;
 734
 735         if(advanced) {
 736                 int32_t data_acked = advanced;
 737
 738                 switch(c->state) {
 739                         case SYN_SENT:
 740                         case SYN_RECEIVED:
 741                                 data_acked--;
 742                                 break;
 743                         // TODO: handle FIN as well.
 744                         default:
 745                                 break;
 746                 }
 747
 748                 assert(data_acked >= 0);
 749
 750                 int32_t bufused = seqdiff(c->snd.last, c->snd.una);
 751                 assert(data_acked <= bufused);
 752
 753                 if(data_acked)
 754                         buffer_get(&c->sndbuf, NULL, data_acked);
 755
 756                 c->snd.una = hdr.ack;
 757
 758                 c->dupack = 0;
 759                 c->snd.cwnd += utcp->mtu;
 760                 if(c->snd.cwnd > c->sndbuf.maxsize)
 761                         c->snd.cwnd = c->sndbuf.maxsize;
 762
 763                 // Check if we have sent a FIN that is now ACKed.
 764                 switch(c->state) {
 765                 case FIN_WAIT_1:
 766                         if(c->snd.una == c->snd.last)
 767                                 set_state(c, FIN_WAIT_2);
 768                         break;
 769                 case CLOSING:
 770                         if(c->snd.una == c->snd.last) {
 771                                 gettimeofday(&c->conn_timeout, NULL);
 772                                 c->conn_timeout.tv_sec += 60;
 773                                 set_state(c, TIME_WAIT);
 774                         }
 775                         break;
 776                 default:
 777                         break;
 778                 }
 779         } else {
 780                 if(!len) {
 781                         c->dupack++;
 782                         if(c->dupack == 3) {
 783                                 debug("Triplicate ACK\n");
 784                                 //TODO: Resend one packet and go to fast recovery mode. See RFC 6582.
 785                                 //We do a very simple variant here; reset the nxt pointer to the last acknowledged packet from the peer.
 786                                 //This will cause us to start retransmitting, but at the same speed as the incoming ACKs arrive,
 787                                 //thus preventing a drop in speed.
 788                                 c->snd.nxt = c->snd.una;
 789                         }
 790                 }
 791         }
 792
 793         // 4. Update timers
 794
 795         if(advanced) {
 796                 timerclear(&c->conn_timeout); // It will be set anew in utcp_timeout() if c->snd.una != c->snd.nxt.
 797                 if(c->snd.una == c->snd.nxt)
 798                         timerclear(&c->rtrx_timeout);
 799         }
 800
 801         // 5. Process SYN stuff
 802
 803         if(hdr.ctl & SYN) {
 804                 switch(c->state) {
 805                 case SYN_SENT:
 806                         // This is a SYNACK. It should always have ACKed the SYN.
 807                         if(!advanced)
 808                                 goto reset;
 809                         c->rcv.irs = hdr.seq;
 810                         c->rcv.nxt = hdr.seq;
 811                         set_state(c, ESTABLISHED);
 812                         // TODO: notify application of this somehow.
 813                         break;
 814                 case SYN_RECEIVED:
 815                 case ESTABLISHED:
 816                 case FIN_WAIT_1:
 817                 case FIN_WAIT_2:
 818                 case CLOSE_WAIT:
 819                 case CLOSING:
 820                 case LAST_ACK:
 821                 case TIME_WAIT:
 822                         // Ehm, no. We should never receive a second SYN.
 823                         goto reset;
 824                 default:
 825                         abort();
 826                 }
 827
 828                 // SYN counts as one sequence number
 829                 c->rcv.nxt++;
 830         }
 831
 832         // 6. Process new data
 833
 834         if(c->state == SYN_RECEIVED) {
 835                 // This is the ACK after the SYNACK. It should always have ACKed the SYNACK.
 836                 if(!advanced)
 837                         goto reset;
 838
 839                 // Are we still LISTENing?
 840                 if(utcp->accept)
 841                         utcp->accept(c, c->src);
 842
 843                 if(c->state != ESTABLISHED) {
 844                         set_state(c, CLOSED);
 845                         c->reapable = true;
 846                         goto reset;
 847                 }
 848         }
 849
 850         if(len) {
 851                 switch(c->state) {
 852                 case SYN_SENT:
 853                 case SYN_RECEIVED:
 854                         // This should never happen.
 855                         abort();
 856                 case ESTABLISHED:
 857                 case FIN_WAIT_1:
 858                 case FIN_WAIT_2:
 859                         break;
 860                 case CLOSE_WAIT:
 861                 case CLOSING:
 862                 case LAST_ACK:
 863                 case TIME_WAIT:
 864                         // Ehm no, We should never receive more data after a FIN.
 865                         goto reset;
 866                 default:
 867                         abort();
 868                 }
 869
 870                 ssize_t rxd;
 871
 872                 if(c->recv) {
 873                         rxd = c->recv(c, data, len);
 874                         if(rxd != len) {
 875                                 // TODO: once we have a receive buffer, handle the application not accepting all data.
 876                                 abort();
 877                         }
 878                         if(rxd < 0)
 879                                 rxd = 0;
 880                         else if(rxd > len)
 881                                 rxd = len; // Bad application, bad!
 882                 } else {
 883                         rxd = len;
 884                 }
 885
 886                 c->rcv.nxt += len;
 887         }
 888
 889         // 7. Process FIN stuff
 890
 891         if(hdr.ctl & FIN) {
 892                 switch(c->state) {
 893                 case SYN_SENT:
 894                 case SYN_RECEIVED:
 895                         // This should never happen.
 896                         abort();
 897                 case ESTABLISHED:
 898                         set_state(c, CLOSE_WAIT);
 899                         break;
 900                 case FIN_WAIT_1:
 901                         set_state(c, CLOSING);
 902                         break;
 903                 case FIN_WAIT_2:
 904                         gettimeofday(&c->conn_timeout, NULL);
 905                         c->conn_timeout.tv_sec += 60;
 906                         set_state(c, TIME_WAIT);
 907                         break;
 908                 case CLOSE_WAIT:
 909                 case CLOSING:
 910                 case LAST_ACK:
 911                 case TIME_WAIT:
 912                         // Ehm, no. We should never receive a second FIN.
 913                         goto reset;
 914                 default:
 915                         abort();
 916                 }
 917
 918                 // FIN counts as one sequence number
 919                 c->rcv.nxt++;
 920                 len++;
 921
 922                 // Inform the application that the peer closed the connection.
 923                 if(c->recv) {
 924                         errno = 0;
 925                         c->recv(c, NULL, 0);
 926                 }
 927         }
 928
 929         // Now we send something back if:
 930         // - we advanced rcv.nxt (ie, we got some data that needs to be ACKed)
 931         //   -> sendatleastone = true
 932         // - or we got an ack, so we should maybe send a bit more data
 933         //   -> sendatleastone = false
 934
 935 ack:
 936         ack(c, prevrcvnxt != c->rcv.nxt);
 937         return 0;
 938
 939 reset:
 940         swap_ports(&hdr);
 941         hdr.wnd = 0;
 942         if(hdr.ctl & ACK) {
 943                 hdr.seq = hdr.ack;
 944                 hdr.ctl = RST;
 945         } else {
 946                 hdr.ack = hdr.seq + len;
 947                 hdr.seq = 0;
 948                 hdr.ctl = RST | ACK;
 949         }
 950         print_packet(utcp, "send", &hdr, sizeof hdr);
 951         utcp->send(utcp, &hdr, sizeof hdr);
 952         return 0;
 953
 954 }
 955
 956 int utcp_shutdown(struct utcp_connection *c, int dir) {
 957         debug("%p shutdown %d\n", c ? c->utcp : NULL, dir);
 958         if(!c) {
 959                 errno = EFAULT;
 960                 return -1;
 961         }
 962
 963         if(c->reapable) {
 964                 debug("Error: shutdown() called on closed connection %p\n", c);
 965                 errno = EBADF;
 966                 return -1;
 967         }
 968
 969         // TODO: handle dir
 970
 971         switch(c->state) {
 972         case CLOSED:
 973                 return 0;
 974         case LISTEN:
 975         case SYN_SENT:
 976                 set_state(c, CLOSED);
 977                 return 0;
 978
 979         case SYN_RECEIVED:
 980         case ESTABLISHED:
 981                 set_state(c, FIN_WAIT_1);
 982                 break;
 983         case FIN_WAIT_1:
 984         case FIN_WAIT_2:
 985                 return 0;
 986         case CLOSE_WAIT:
 987                 set_state(c, CLOSING);
 988                 break;
 989
 990         case CLOSING:
 991         case LAST_ACK:
 992         case TIME_WAIT:
 993                 return 0;
 994         }
 995
 996         c->snd.last++;
 997
 998         ack(c, false);
 999         return 0;
1000 }
1001
1002 int utcp_close(struct utcp_connection *c) {
1003         if(utcp_shutdown(c, SHUT_RDWR))
1004                 return -1;
1005         c->recv = NULL;
1006         c->poll = NULL;
1007         c->reapable = true;
1008         return 0;
1009 }
1010
1011 int utcp_abort(struct utcp_connection *c) {
1012         if(!c) {
1013                 errno = EFAULT;
1014                 return -1;
1015         }
1016
1017         if(c->reapable) {
1018                 debug("Error: abort() called on closed connection %p\n", c);
1019                 errno = EBADF;
1020                 return -1;
1021         }
1022
1023         c->recv = NULL;
1024         c->poll = NULL;
1025         c->reapable = true;
1026
1027         switch(c->state) {
1028         case CLOSED:
1029                 return 0;
1030         case LISTEN:
1031         case SYN_SENT:
1032         case CLOSING:
1033         case LAST_ACK:
1034         case TIME_WAIT:
1035                 set_state(c, CLOSED);
1036                 return 0;
1037
1038         case SYN_RECEIVED:
1039         case ESTABLISHED:
1040         case FIN_WAIT_1:
1041         case FIN_WAIT_2:
1042         case CLOSE_WAIT:
1043                 set_state(c, CLOSED);
1044                 break;
1045         }
1046
1047         // Send RST
1048
1049         struct hdr hdr;
1050
1051         hdr.src = c->src;
1052         hdr.dst = c->dst;
1053         hdr.seq = c->snd.nxt;
1054         hdr.ack = 0;
1055         hdr.wnd = 0;
1056         hdr.ctl = RST;
1057
1058         print_packet(c->utcp, "send", &hdr, sizeof hdr);
1059         c->utcp->send(c->utcp, &hdr, sizeof hdr);
1060         return 0;
1061 }
1062
1063 /* Handle timeouts.
1064  * One call to this function will loop through all connections,
1065  * checking if something needs to be resent or not.
1066  * The return value is the time to the next timeout in milliseconds,
1067  * or maybe a negative value if the timeout is infinite.
1068  */
1069 struct timeval utcp_timeout(struct utcp *utcp) {
1070         struct timeval now;
1071         gettimeofday(&now, NULL);
1072         struct timeval next = {now.tv_sec + 3600, now.tv_usec};
1073
1074         for(int i = 0; i < utcp->nconnections; i++) {
1075                 struct utcp_connection *c = utcp->connections[i];
1076                 if(!c)
1077                         continue;
1078
1079                 if(c->state == CLOSED) {
1080                         if(c->reapable) {
1081                                 debug("Reaping %p\n", c);
1082                                 free_connection(c);
1083                                 i--;
1084                         }
1085                         continue;
1086                 }
1087
1088                 if(timerisset(&c->conn_timeout) && timercmp(&c->conn_timeout, &now, <)) {
1089                         errno = ETIMEDOUT;
1090                         c->state = CLOSED;
1091                         if(c->recv)
1092                                 c->recv(c, NULL, 0);
1093                         continue;
1094                 }
1095
1096                 if(timerisset(&c->rtrx_timeout) && timercmp(&c->rtrx_timeout, &now, <)) {
1097                         retransmit(c);
1098                 }
1099
1100                 if(c->poll && buffer_free(&c->sndbuf) && (c->state == ESTABLISHED || c->state == CLOSE_WAIT))
1101                         c->poll(c, buffer_free(&c->sndbuf));
1102
1103                 if(timerisset(&c->conn_timeout) && timercmp(&c->conn_timeout, &next, <))
1104                         next = c->conn_timeout;
1105
1106                 if(c->snd.nxt != c->snd.una) {
1107                         c->rtrx_timeout = now;
1108                         c->rtrx_timeout.tv_sec++;
1109                 } else {
1110                         timerclear(&c->rtrx_timeout);
1111                 }
1112
1113                 if(timerisset(&c->rtrx_timeout) && timercmp(&c->rtrx_timeout, &next, <))
1114                         next = c->rtrx_timeout;
1115         }
1116
1117         struct timeval diff;
1118         timersub(&next, &now, &diff);
1119         return diff;
1120 }
1121
1122 struct utcp *utcp_init(utcp_accept_t accept, utcp_pre_accept_t pre_accept, utcp_send_t send, void *priv) {
1123         struct utcp *utcp = calloc(1, sizeof *utcp);
1124         if(!utcp)
1125                 return NULL;
1126
1127         if(!send) {
1128                 errno = EFAULT;
1129                 return NULL;
1130         }
1131
1132         utcp->accept = accept;
1133         utcp->pre_accept = pre_accept;
1134         utcp->send = send;
1135         utcp->priv = priv;
1136         utcp->mtu = 1000;
1137         utcp->timeout = 60;
1138
1139         return utcp;
1140 }
1141
1142 void utcp_exit(struct utcp *utcp) {
1143         if(!utcp)
1144                 return;
1145         for(int i = 0; i < utcp->nconnections; i++) {
1146                 if(!utcp->connections[i]->reapable)
1147                         debug("Warning, freeing unclosed connection %p\n", utcp->connections[i]);
1148                 buffer_exit(&utcp->connections[i]->sndbuf);
1149                 free(utcp->connections[i]);
1150         }
1151         free(utcp->connections);
1152         free(utcp);
1153 }
1154
1155 uint16_t utcp_get_mtu(struct utcp *utcp) {
1156         return utcp->mtu;
1157 }
1158
1159 void utcp_set_mtu(struct utcp *utcp, uint16_t mtu) {
1160         // TODO: handle overhead of the header
1161         utcp->mtu = mtu;
1162 }
1163
1164 int utcp_get_user_timeout(struct utcp *u) {
1165         return u->timeout;
1166 }
1167
1168 void utcp_set_user_timeout(struct utcp *u, int timeout) {
1169         u->timeout = timeout;
1170 }
1171
1172 size_t utcp_get_sndbuf(struct utcp_connection *c) {
1173         return c->sndbuf.maxsize;
1174 }
1175
1176 size_t utcp_get_sndbuf_free(struct utcp_connection *c) {
1177         return buffer_free(&c->sndbuf);
1178 }
1179
1180 void utcp_set_sndbuf(struct utcp_connection *c, size_t size) {
1181         c->sndbuf.maxsize = size;
1182         if(c->sndbuf.maxsize != size)
1183                 c->sndbuf.maxsize = -1;
1184 }
1185
1186 bool utcp_get_nodelay(struct utcp_connection *c) {
1187         return c->nodelay;
1188 }
1189
1190 void utcp_set_nodelay(struct utcp_connection *c, bool nodelay) {
1191         c->nodelay = nodelay;
1192 }
1193
1194 bool utcp_get_keepalive(struct utcp_connection *c) {
1195         return c->keepalive;
1196 }
1197
1198 void utcp_set_keepalive(struct utcp_connection *c, bool keepalive) {
1199         c->keepalive = keepalive;
1200 }
1201
1202 size_t utcp_get_outq(struct utcp_connection *c) {
1203         return seqdiff(c->snd.nxt, c->snd.una);
1204 }
1205
1206 void utcp_set_recv_cb(struct utcp_connection *c, utcp_recv_t recv) {
1207         c->recv = recv;
1208 }
1209
1210 void utcp_set_poll_cb(struct utcp_connection *c, utcp_poll_t poll) {
1211         c->poll = poll;
1212 }