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 #ifdef UTCP_DEBUG
 286 #warning debugging
 287         c->snd.iss = 0;
 288 #else
 289         c->snd.iss = rand();
 290 #endif
 291         c->snd.una = c->snd.iss;
 292         c->snd.nxt = c->snd.iss + 1;
 293         c->rcv.wnd = utcp->mtu;
 294         c->snd.last = c->snd.nxt;
 295         c->snd.cwnd = utcp->mtu;
 296         c->utcp = utcp;
 297
 298         // Add it to the sorted list of connections
 299
 300         utcp->connections[utcp->nconnections++] = c;
 301         qsort(utcp->connections, utcp->nconnections, sizeof *utcp->connections, compare);
 302
 303         return c;
 304 }
 305
 306 struct utcp_connection *utcp_connect(struct utcp *utcp, uint16_t dst, utcp_recv_t recv, void *priv) {
 307         struct utcp_connection *c = allocate_connection(utcp, 0, dst);
 308         if(!c)
 309                 return NULL;
 310
 311         c->recv = recv;
 312         c->priv = priv;
 313
 314         struct hdr hdr;
 315
 316         hdr.src = c->src;
 317         hdr.dst = c->dst;
 318         hdr.seq = c->snd.iss;
 319         hdr.ack = 0;
 320         hdr.wnd = c->rcv.wnd;
 321         hdr.ctl = SYN;
 322         hdr.aux = 0;
 323
 324         set_state(c, SYN_SENT);
 325
 326         print_packet(utcp, "send", &hdr, sizeof hdr);
 327         utcp->send(utcp, &hdr, sizeof hdr);
 328
 329         gettimeofday(&c->conn_timeout, NULL);
 330         c->conn_timeout.tv_sec += utcp->timeout;
 331
 332         return c;
 333 }
 334
 335 void utcp_accept(struct utcp_connection *c, utcp_recv_t recv, void *priv) {
 336         if(c->reapable || c->state != SYN_RECEIVED) {
 337                 debug("Error: accept() called on invalid connection %p in state %s\n", c, strstate[c->state]);
 338                 return;
 339         }
 340
 341         debug("%p accepted, %p %p\n", c, recv, priv);
 342         c->recv = recv;
 343         c->priv = priv;
 344         set_state(c, ESTABLISHED);
 345 }
 346
 347 static void ack(struct utcp_connection *c, bool sendatleastone) {
 348         int32_t left = seqdiff(c->snd.last, c->snd.nxt);
 349         int32_t cwndleft = c->snd.cwnd - seqdiff(c->snd.nxt, c->snd.una);
 350
 351         assert(left >= 0);
 352
 353         if(cwndleft <= 0)
 354                 cwndleft = 0;
 355
 356         if(cwndleft < left)
 357                 left = cwndleft;
 358
 359         if(!left && !sendatleastone)
 360                 return;
 361
 362         struct {
 363                 struct hdr hdr;
 364                 char data[];
 365         } *pkt;
 366
 367         pkt = malloc(sizeof pkt->hdr + c->utcp->mtu);
 368         if(!pkt)
 369                 return;
 370
 371         pkt->hdr.src = c->src;
 372         pkt->hdr.dst = c->dst;
 373         pkt->hdr.ack = c->rcv.nxt;
 374         pkt->hdr.wnd = c->snd.wnd;
 375         pkt->hdr.ctl = ACK;
 376         pkt->hdr.aux = 0;
 377
 378         do {
 379                 uint32_t seglen = left > c->utcp->mtu ? c->utcp->mtu : left;
 380                 pkt->hdr.seq = c->snd.nxt;
 381
 382                 buffer_copy(&c->sndbuf, pkt->data, seqdiff(c->snd.nxt, c->snd.una), seglen);
 383
 384                 c->snd.nxt += seglen;
 385                 left -= seglen;
 386
 387                 if(seglen && fin_wanted(c, c->snd.nxt)) {
 388                         seglen--;
 389                         pkt->hdr.ctl |= FIN;
 390                 }
 391
 392                 print_packet(c->utcp, "send", pkt, sizeof pkt->hdr + seglen);
 393                 c->utcp->send(c->utcp, pkt, sizeof pkt->hdr + seglen);
 394         } while(left);
 395
 396         free(pkt);
 397 }
 398
 399 ssize_t utcp_send(struct utcp_connection *c, const void *data, size_t len) {
 400         if(c->reapable) {
 401                 debug("Error: send() called on closed connection %p\n", c);
 402                 errno = EBADF;
 403                 return -1;
 404         }
 405
 406         switch(c->state) {
 407         case CLOSED:
 408         case LISTEN:
 409         case SYN_SENT:
 410         case SYN_RECEIVED:
 411                 debug("Error: send() called on unconnected connection %p\n", c);
 412                 errno = ENOTCONN;
 413                 return -1;
 414         case ESTABLISHED:
 415         case CLOSE_WAIT:
 416                 break;
 417         case FIN_WAIT_1:
 418         case FIN_WAIT_2:
 419         case CLOSING:
 420         case LAST_ACK:
 421         case TIME_WAIT:
 422                 debug("Error: send() called on closing connection %p\n", c);
 423                 errno = EPIPE;
 424                 return -1;
 425         }
 426
 427         // Add data to send buffer
 428
 429         if(!len)
 430                 return 0;
 431
 432         if(!data) {
 433                 errno = EFAULT;
 434                 return -1;
 435         }
 436
 437         len = buffer_put(&c->sndbuf, data, len);
 438         if(len <= 0) {
 439                 errno = EWOULDBLOCK;
 440                 return 0;
 441         }
 442
 443         c->snd.last += len;
 444         ack(c, false);
 445         return len;
 446 }
 447
 448 static void swap_ports(struct hdr *hdr) {
 449         uint16_t tmp = hdr->src;
 450         hdr->src = hdr->dst;
 451         hdr->dst = tmp;
 452 }
 453
 454 static void retransmit(struct utcp_connection *c) {
 455         if(c->state == CLOSED || c->snd.nxt == c->snd.una)
 456                 return;
 457
 458         struct utcp *utcp = c->utcp;
 459
 460         struct {
 461                 struct hdr hdr;
 462                 char data[];
 463         } *pkt;
 464
 465         pkt = malloc(sizeof pkt->hdr + c->utcp->mtu);
 466         if(!pkt)
 467                 return;
 468
 469         pkt->hdr.src = c->src;
 470         pkt->hdr.dst = c->dst;
 471
 472         switch(c->state) {
 473                 case SYN_SENT:
 474                         // Send our SYN again
 475                         pkt->hdr.seq = c->snd.iss;
 476                         pkt->hdr.ack = 0;
 477                         pkt->hdr.wnd = c->rcv.wnd;
 478                         pkt->hdr.ctl = SYN;
 479                         print_packet(c->utcp, "rtrx", pkt, sizeof pkt->hdr);
 480                         utcp->send(utcp, pkt, sizeof pkt->hdr);
 481                         break;
 482
 483                 case SYN_RECEIVED:
 484                         // Send SYNACK again
 485                         pkt->hdr.seq = c->snd.nxt;
 486                         pkt->hdr.ack = c->rcv.nxt;
 487                         pkt->hdr.ctl = SYN | ACK;
 488                         print_packet(c->utcp, "rtrx", pkt, sizeof pkt->hdr);
 489                         utcp->send(utcp, pkt, sizeof pkt->hdr);
 490                         break;
 491
 492                 case ESTABLISHED:
 493                 case FIN_WAIT_1:
 494                 case CLOSE_WAIT:
 495                 case CLOSING:
 496                 case LAST_ACK:
 497                         // Send unacked data again.
 498                         pkt->hdr.seq = c->snd.una;
 499                         pkt->hdr.ack = c->rcv.nxt;
 500                         pkt->hdr.ctl = ACK;
 501                         uint32_t len = seqdiff(c->snd.last, c->snd.una);
 502                         if(len > utcp->mtu)
 503                                 len = utcp->mtu;
 504                         if(fin_wanted(c, c->snd.una + len)) {
 505                                 len--;
 506                                 pkt->hdr.ctl |= FIN;
 507                         }
 508                         c->snd.nxt = c->snd.una + len;
 509                         buffer_copy(&c->sndbuf, pkt->data, 0, len);
 510                         print_packet(c->utcp, "rtrx", pkt, sizeof pkt->hdr + len);
 511                         utcp->send(utcp, pkt, sizeof pkt->hdr + len);
 512                         break;
 513
 514                 case CLOSED:
 515                 case LISTEN:
 516                 case TIME_WAIT:
 517                 case FIN_WAIT_2:
 518                         // We shouldn't need to retransmit anything in this state.
 519 #ifdef UTCP_DEBUG
 520                         abort();
 521 #endif
 522                         timerclear(&c->rtrx_timeout);
 523                         break;
 524         }
 525
 526         free(pkt);
 527 }
 528
 529
 530 ssize_t utcp_recv(struct utcp *utcp, const void *data, size_t len) {
 531         if(!utcp) {
 532                 errno = EFAULT;
 533                 return -1;
 534         }
 535
 536         if(!len)
 537                 return 0;
 538
 539         if(!data) {
 540                 errno = EFAULT;
 541                 return -1;
 542         }
 543
 544         print_packet(utcp, "recv", data, len);
 545
 546         // Drop packets smaller than the header
 547
 548         struct hdr hdr;
 549         if(len < sizeof hdr) {
 550                 errno = EBADMSG;
 551                 return -1;
 552         }
 553
 554         // Make a copy from the potentially unaligned data to a struct hdr
 555
 556         memcpy(&hdr, data, sizeof hdr);
 557         data += sizeof hdr;
 558         len -= sizeof hdr;
 559
 560         // Drop packets with an unknown CTL flag
 561
 562         if(hdr.ctl & ~(SYN | ACK | RST | FIN)) {
 563                 errno = EBADMSG;
 564                 return -1;
 565         }
 566
 567         // Try to match the packet to an existing connection
 568
 569         struct utcp_connection *c = find_connection(utcp, hdr.dst, hdr.src);
 570
 571         // Is it for a new connection?
 572
 573         if(!c) {
 574                 // Ignore RST packets
 575
 576                 if(hdr.ctl & RST)
 577                         return 0;
 578
 579                 // Is it a SYN packet and are we LISTENing?
 580
 581                 if(hdr.ctl & SYN && !(hdr.ctl & ACK) && utcp->accept) {
 582                         // If we don't want to accept it, send a RST back
 583                         if((utcp->pre_accept && !utcp->pre_accept(utcp, hdr.dst))) {
 584                                 len = 1;
 585                                 goto reset;
 586                         }
 587
 588                         // Try to allocate memory, otherwise send a RST back
 589                         c = allocate_connection(utcp, hdr.dst, hdr.src);
 590                         if(!c) {
 591                                 len = 1;
 592                                 goto reset;
 593                         }
 594
 595                         // Return SYN+ACK, go to SYN_RECEIVED state
 596                         c->snd.wnd = hdr.wnd;
 597                         c->rcv.irs = hdr.seq;
 598                         c->rcv.nxt = c->rcv.irs + 1;
 599                         set_state(c, SYN_RECEIVED);
 600
 601                         hdr.dst = c->dst;
 602                         hdr.src = c->src;
 603                         hdr.ack = c->rcv.irs + 1;
 604                         hdr.seq = c->snd.iss;
 605                         hdr.ctl = SYN | ACK;
 606                         print_packet(c->utcp, "send", &hdr, sizeof hdr);
 607                         utcp->send(utcp, &hdr, sizeof hdr);
 608                 } else {
 609                         // No, we don't want your packets, send a RST back
 610                         len = 1;
 611                         goto reset;
 612                 }
 613
 614                 return 0;
 615         }
 616
 617         debug("%p state %s\n", c->utcp, strstate[c->state]);
 618
 619         // In case this is for a CLOSED connection, ignore the packet.
 620         // TODO: make it so incoming packets can never match a CLOSED connection.
 621
 622         if(c->state == CLOSED)
 623                 return 0;
 624
 625         // It is for an existing connection.
 626
 627         uint32_t prevrcvnxt = c->rcv.nxt;
 628
 629         // 1. Drop invalid packets.
 630
 631         // 1a. Drop packets that should not happen in our current state.
 632
 633         switch(c->state) {
 634         case SYN_SENT:
 635         case SYN_RECEIVED:
 636         case ESTABLISHED:
 637         case FIN_WAIT_1:
 638         case FIN_WAIT_2:
 639         case CLOSE_WAIT:
 640         case CLOSING:
 641         case LAST_ACK:
 642         case TIME_WAIT:
 643                 break;
 644         default:
 645 #ifdef UTCP_DEBUG
 646                 abort();
 647 #endif
 648                 break;
 649         }
 650
 651         // 1b. Drop packets with a sequence number not in our receive window.
 652
 653         bool acceptable;
 654
 655         if(c->state == SYN_SENT)
 656                 acceptable = true;
 657
 658         // TODO: handle packets overlapping c->rcv.nxt.
 659 #if 0
 660         // Only use this when accepting out-of-order packets.
 661         else if(len == 0)
 662                 if(c->rcv.wnd == 0)
 663                         acceptable = hdr.seq == c->rcv.nxt;
 664                 else
 665                         acceptable = (seqdiff(hdr.seq, c->rcv.nxt) >= 0 && seqdiff(hdr.seq, c->rcv.nxt + c->rcv.wnd) < 0);
 666         else
 667                 if(c->rcv.wnd == 0)
 668                         // We don't accept data when the receive window is zero.
 669                         acceptable = false;
 670                 else
 671                         // Both start and end of packet must be within the receive window
 672                         acceptable = (seqdiff(hdr.seq, c->rcv.nxt) >= 0 && seqdiff(hdr.seq, c->rcv.nxt + c->rcv.wnd) < 0)
 673                                 || (seqdiff(hdr.seq + len + 1, c->rcv.nxt) >= 0 && seqdiff(hdr.seq + len - 1, c->rcv.nxt + c->rcv.wnd) < 0);
 674 #else
 675         if(c->state != SYN_SENT)
 676                 acceptable = hdr.seq == c->rcv.nxt;
 677 #endif
 678
 679         if(!acceptable) {
 680                 debug("Packet not acceptable, %u <= %u + %zu < %u\n", c->rcv.nxt, hdr.seq, len, c->rcv.nxt + c->rcv.wnd);
 681                 // Ignore unacceptable RST packets.
 682                 if(hdr.ctl & RST)
 683                         return 0;
 684                 // Otherwise, send an ACK back in the hope things improve.
 685                 ack(c, true);
 686                 return 0;
 687         }
 688
 689         c->snd.wnd = hdr.wnd; // TODO: move below
 690
 691         // 1c. Drop packets with an invalid ACK.
 692         // ackno should not roll back, and it should also not be bigger than snd.nxt.
 693
 694         if(hdr.ctl & ACK && (seqdiff(hdr.ack, c->snd.nxt) > 0 || seqdiff(hdr.ack, c->snd.una) < 0)) {
 695                 debug("Packet ack seqno out of range, %u %u %u\n", hdr.ack, c->snd.una, c->snd.nxt);
 696                 // Ignore unacceptable RST packets.
 697                 if(hdr.ctl & RST)
 698                         return 0;
 699                 goto reset;
 700         }
 701
 702         // 2. Handle RST packets
 703
 704         if(hdr.ctl & RST) {
 705                 switch(c->state) {
 706                 case SYN_SENT:
 707                         if(!(hdr.ctl & ACK))
 708                                 return 0;
 709                         // The peer has refused our connection.
 710                         set_state(c, CLOSED);
 711                         errno = ECONNREFUSED;
 712                         if(c->recv)
 713                                 c->recv(c, NULL, 0);
 714                         return 0;
 715                 case SYN_RECEIVED:
 716                         if(hdr.ctl & ACK)
 717                                 return 0;
 718                         // We haven't told the application about this connection yet. Silently delete.
 719                         free_connection(c);
 720                         return 0;
 721                 case ESTABLISHED:
 722                 case FIN_WAIT_1:
 723                 case FIN_WAIT_2:
 724                 case CLOSE_WAIT:
 725                         if(hdr.ctl & ACK)
 726                                 return 0;
 727                         // The peer has aborted our connection.
 728                         set_state(c, CLOSED);
 729                         errno = ECONNRESET;
 730                         if(c->recv)
 731                                 c->recv(c, NULL, 0);
 732                         return 0;
 733                 case CLOSING:
 734                 case LAST_ACK:
 735                 case TIME_WAIT:
 736                         if(hdr.ctl & ACK)
 737                                 return 0;
 738                         // As far as the application is concerned, the connection has already been closed.
 739                         // If it has called utcp_close() already, we can immediately free this connection.
 740                         if(c->reapable) {
 741                                 free_connection(c);
 742                                 return 0;
 743                         }
 744                         // Otherwise, immediately move to the CLOSED state.
 745                         set_state(c, CLOSED);
 746                         return 0;
 747                 default:
 748 #ifdef UTCP_DEBUG
 749                         abort();
 750 #endif
 751                         break;
 752                 }
 753         }
 754
 755         // 3. Advance snd.una
 756
 757         uint32_t advanced = seqdiff(hdr.ack, c->snd.una);
 758         prevrcvnxt = c->rcv.nxt;
 759
 760         if(advanced) {
 761                 int32_t data_acked = advanced;
 762
 763                 switch(c->state) {
 764                         case SYN_SENT:
 765                         case SYN_RECEIVED:
 766                                 data_acked--;
 767                                 break;
 768                         // TODO: handle FIN as well.
 769                         default:
 770                                 break;
 771                 }
 772
 773                 assert(data_acked >= 0);
 774
 775                 int32_t bufused = seqdiff(c->snd.last, c->snd.una);
 776                 assert(data_acked <= bufused);
 777
 778                 if(data_acked)
 779                         buffer_get(&c->sndbuf, NULL, data_acked);
 780
 781                 c->snd.una = hdr.ack;
 782
 783                 c->dupack = 0;
 784                 c->snd.cwnd += utcp->mtu;
 785                 if(c->snd.cwnd > c->sndbuf.maxsize)
 786                         c->snd.cwnd = c->sndbuf.maxsize;
 787
 788                 // Check if we have sent a FIN that is now ACKed.
 789                 switch(c->state) {
 790                 case FIN_WAIT_1:
 791                         if(c->snd.una == c->snd.last)
 792                                 set_state(c, FIN_WAIT_2);
 793                         break;
 794                 case CLOSING:
 795                         if(c->snd.una == c->snd.last) {
 796                                 gettimeofday(&c->conn_timeout, NULL);
 797                                 c->conn_timeout.tv_sec += 60;
 798                                 set_state(c, TIME_WAIT);
 799                         }
 800                         break;
 801                 default:
 802                         break;
 803                 }
 804         } else {
 805                 if(!len) {
 806                         c->dupack++;
 807                         if(c->dupack == 3) {
 808                                 debug("Triplicate ACK\n");
 809                                 //TODO: Resend one packet and go to fast recovery mode. See RFC 6582.
 810                                 //We do a very simple variant here; reset the nxt pointer to the last acknowledged packet from the peer.
 811                                 //This will cause us to start retransmitting, but at the same speed as the incoming ACKs arrive,
 812                                 //thus preventing a drop in speed.
 813                                 c->snd.nxt = c->snd.una;
 814                         }
 815                 }
 816         }
 817
 818         // 4. Update timers
 819
 820         if(advanced) {
 821                 timerclear(&c->conn_timeout); // It will be set anew in utcp_timeout() if c->snd.una != c->snd.nxt.
 822                 if(c->snd.una == c->snd.nxt)
 823                         timerclear(&c->rtrx_timeout);
 824         }
 825
 826         // 5. Process SYN stuff
 827
 828         if(hdr.ctl & SYN) {
 829                 switch(c->state) {
 830                 case SYN_SENT:
 831                         // This is a SYNACK. It should always have ACKed the SYN.
 832                         if(!advanced)
 833                                 goto reset;
 834                         c->rcv.irs = hdr.seq;
 835                         c->rcv.nxt = hdr.seq;
 836                         set_state(c, ESTABLISHED);
 837                         // TODO: notify application of this somehow.
 838                         break;
 839                 case SYN_RECEIVED:
 840                 case ESTABLISHED:
 841                 case FIN_WAIT_1:
 842                 case FIN_WAIT_2:
 843                 case CLOSE_WAIT:
 844                 case CLOSING:
 845                 case LAST_ACK:
 846                 case TIME_WAIT:
 847                         // Ehm, no. We should never receive a second SYN.
 848                         goto reset;
 849                 default:
 850 #ifdef UTCP_DEBUG
 851                         abort();
 852 #endif
 853                         return 0;
 854                 }
 855
 856                 // SYN counts as one sequence number
 857                 c->rcv.nxt++;
 858         }
 859
 860         // 6. Process new data
 861
 862         if(c->state == SYN_RECEIVED) {
 863                 // This is the ACK after the SYNACK. It should always have ACKed the SYNACK.
 864                 if(!advanced)
 865                         goto reset;
 866
 867                 // Are we still LISTENing?
 868                 if(utcp->accept)
 869                         utcp->accept(c, c->src);
 870
 871                 if(c->state != ESTABLISHED) {
 872                         set_state(c, CLOSED);
 873                         c->reapable = true;
 874                         goto reset;
 875                 }
 876         }
 877
 878         if(len) {
 879                 switch(c->state) {
 880                 case SYN_SENT:
 881                 case SYN_RECEIVED:
 882                         // This should never happen.
 883 #ifdef UTCP_DEBUG
 884                         abort();
 885 #endif
 886                         return 0;
 887                 case ESTABLISHED:
 888                 case FIN_WAIT_1:
 889                 case FIN_WAIT_2:
 890                         break;
 891                 case CLOSE_WAIT:
 892                 case CLOSING:
 893                 case LAST_ACK:
 894                 case TIME_WAIT:
 895                         // Ehm no, We should never receive more data after a FIN.
 896                         goto reset;
 897                 default:
 898 #ifdef UTCP_DEBUG
 899                         abort();
 900 #endif
 901                         return 0;
 902                 }
 903
 904                 ssize_t rxd;
 905
 906                 if(c->recv) {
 907                         rxd = c->recv(c, data, len);
 908                         if(rxd != len) {
 909                                 // TODO: once we have a receive buffer, handle the application not accepting all data.
 910                                 abort();
 911                         }
 912                         if(rxd < 0)
 913                                 rxd = 0;
 914                         else if(rxd > len)
 915                                 rxd = len; // Bad application, bad!
 916                 } else {
 917                         rxd = len;
 918                 }
 919
 920                 c->rcv.nxt += len;
 921         }
 922
 923         // 7. Process FIN stuff
 924
 925         if(hdr.ctl & FIN) {
 926                 switch(c->state) {
 927                 case SYN_SENT:
 928                 case SYN_RECEIVED:
 929                         // This should never happen.
 930 #ifdef UTCP_DEBUG
 931                         abort();
 932 #endif
 933                         break;
 934                 case ESTABLISHED:
 935                         set_state(c, CLOSE_WAIT);
 936                         break;
 937                 case FIN_WAIT_1:
 938                         set_state(c, CLOSING);
 939                         break;
 940                 case FIN_WAIT_2:
 941                         gettimeofday(&c->conn_timeout, NULL);
 942                         c->conn_timeout.tv_sec += 60;
 943                         set_state(c, TIME_WAIT);
 944                         break;
 945                 case CLOSE_WAIT:
 946                 case CLOSING:
 947                 case LAST_ACK:
 948                 case TIME_WAIT:
 949                         // Ehm, no. We should never receive a second FIN.
 950                         goto reset;
 951                 default:
 952 #ifdef UTCP_DEBUG
 953                         abort();
 954 #endif
 955                         break;
 956                 }
 957
 958                 // FIN counts as one sequence number
 959                 c->rcv.nxt++;
 960                 len++;
 961
 962                 // Inform the application that the peer closed the connection.
 963                 if(c->recv) {
 964                         errno = 0;
 965                         c->recv(c, NULL, 0);
 966                 }
 967         }
 968
 969         // Now we send something back if:
 970         // - we advanced rcv.nxt (ie, we got some data that needs to be ACKed)
 971         //   -> sendatleastone = true
 972         // - or we got an ack, so we should maybe send a bit more data
 973         //   -> sendatleastone = false
 974
 975 ack:
 976         ack(c, prevrcvnxt != c->rcv.nxt);
 977         return 0;
 978
 979 reset:
 980         swap_ports(&hdr);
 981         hdr.wnd = 0;
 982         if(hdr.ctl & ACK) {
 983                 hdr.seq = hdr.ack;
 984                 hdr.ctl = RST;
 985         } else {
 986                 hdr.ack = hdr.seq + len;
 987                 hdr.seq = 0;
 988                 hdr.ctl = RST | ACK;
 989         }
 990         print_packet(utcp, "send", &hdr, sizeof hdr);
 991         utcp->send(utcp, &hdr, sizeof hdr);
 992         return 0;
 993
 994 }
 995
 996 int utcp_shutdown(struct utcp_connection *c, int dir) {
 997         debug("%p shutdown %d at %u\n", c ? c->utcp : NULL, dir, c ? c->snd.last : 0);
 998         if(!c) {
 999                 errno = EFAULT;
1000                 return -1;
1001         }
1002
1003         if(c->reapable) {
1004                 debug("Error: shutdown() called on closed connection %p\n", c);
1005                 errno = EBADF;
1006                 return -1;
1007         }
1008
1009         if(!(dir == UTCP_SHUT_RD || dir == UTCP_SHUT_WR || dir == UTCP_SHUT_RDWR)) {
1010                 errno = EINVAL;
1011                 return -1;
1012         }
1013
1014         // TCP does not have a provision for stopping incoming packets.
1015         // The best we can do is to just ignore them.
1016         if(dir == UTCP_SHUT_RD || dir == UTCP_SHUT_RDWR)
1017                 c->recv = NULL;
1018
1019         // The rest of the code deals with shutting down writes.
1020         if(dir == UTCP_SHUT_RD)
1021                 return 0;
1022
1023         switch(c->state) {
1024         case CLOSED:
1025         case LISTEN:
1026                 errno = ENOTCONN;
1027                 return -1;
1028
1029         case SYN_SENT:
1030                 set_state(c, CLOSED);
1031                 return 0;
1032
1033         case SYN_RECEIVED:
1034         case ESTABLISHED:
1035                 set_state(c, FIN_WAIT_1);
1036                 break;
1037         case FIN_WAIT_1:
1038         case FIN_WAIT_2:
1039                 return 0;
1040         case CLOSE_WAIT:
1041                 set_state(c, CLOSING);
1042                 break;
1043
1044         case CLOSING:
1045         case LAST_ACK:
1046         case TIME_WAIT:
1047                 return 0;
1048         }
1049
1050         c->snd.last++;
1051
1052         ack(c, false);
1053         return 0;
1054 }
1055
1056 int utcp_close(struct utcp_connection *c) {
1057         if(utcp_shutdown(c, SHUT_RDWR))
1058                 return -1;
1059         c->recv = NULL;
1060         c->poll = NULL;
1061         c->reapable = true;
1062         return 0;
1063 }
1064
1065 int utcp_abort(struct utcp_connection *c) {
1066         if(!c) {
1067                 errno = EFAULT;
1068                 return -1;
1069         }
1070
1071         if(c->reapable) {
1072                 debug("Error: abort() called on closed connection %p\n", c);
1073                 errno = EBADF;
1074                 return -1;
1075         }
1076
1077         c->recv = NULL;
1078         c->poll = NULL;
1079         c->reapable = true;
1080
1081         switch(c->state) {
1082         case CLOSED:
1083                 return 0;
1084         case LISTEN:
1085         case SYN_SENT:
1086         case CLOSING:
1087         case LAST_ACK:
1088         case TIME_WAIT:
1089                 set_state(c, CLOSED);
1090                 return 0;
1091
1092         case SYN_RECEIVED:
1093         case ESTABLISHED:
1094         case FIN_WAIT_1:
1095         case FIN_WAIT_2:
1096         case CLOSE_WAIT:
1097                 set_state(c, CLOSED);
1098                 break;
1099         }
1100
1101         // Send RST
1102
1103         struct hdr hdr;
1104
1105         hdr.src = c->src;
1106         hdr.dst = c->dst;
1107         hdr.seq = c->snd.nxt;
1108         hdr.ack = 0;
1109         hdr.wnd = 0;
1110         hdr.ctl = RST;
1111
1112         print_packet(c->utcp, "send", &hdr, sizeof hdr);
1113         c->utcp->send(c->utcp, &hdr, sizeof hdr);
1114         return 0;
1115 }
1116
1117 /* Handle timeouts.
1118  * One call to this function will loop through all connections,
1119  * checking if something needs to be resent or not.
1120  * The return value is the time to the next timeout in milliseconds,
1121  * or maybe a negative value if the timeout is infinite.
1122  */
1123 struct timeval utcp_timeout(struct utcp *utcp) {
1124         struct timeval now;
1125         gettimeofday(&now, NULL);
1126         struct timeval next = {now.tv_sec + 3600, now.tv_usec};
1127
1128         for(int i = 0; i < utcp->nconnections; i++) {
1129                 struct utcp_connection *c = utcp->connections[i];
1130                 if(!c)
1131                         continue;
1132
1133                 if(c->state == CLOSED) {
1134                         if(c->reapable) {
1135                                 debug("Reaping %p\n", c);
1136                                 free_connection(c);
1137                                 i--;
1138                         }
1139                         continue;
1140                 }
1141
1142                 if(timerisset(&c->conn_timeout) && timercmp(&c->conn_timeout, &now, <)) {
1143                         errno = ETIMEDOUT;
1144                         c->state = CLOSED;
1145                         if(c->recv)
1146                                 c->recv(c, NULL, 0);
1147                         continue;
1148                 }
1149
1150                 if(timerisset(&c->rtrx_timeout) && timercmp(&c->rtrx_timeout, &now, <)) {
1151                         retransmit(c);
1152                 }
1153
1154                 if(c->poll && buffer_free(&c->sndbuf) && (c->state == ESTABLISHED || c->state == CLOSE_WAIT))
1155                         c->poll(c, buffer_free(&c->sndbuf));
1156
1157                 if(timerisset(&c->conn_timeout) && timercmp(&c->conn_timeout, &next, <))
1158                         next = c->conn_timeout;
1159
1160                 if(c->snd.nxt != c->snd.una) {
1161                         c->rtrx_timeout = now;
1162                         c->rtrx_timeout.tv_sec++;
1163                 } else {
1164                         timerclear(&c->rtrx_timeout);
1165                 }
1166
1167                 if(timerisset(&c->rtrx_timeout) && timercmp(&c->rtrx_timeout, &next, <))
1168                         next = c->rtrx_timeout;
1169         }
1170
1171         struct timeval diff;
1172         timersub(&next, &now, &diff);
1173         return diff;
1174 }
1175
1176 bool utcp_is_active(struct utcp *utcp) {
1177         if(!utcp)
1178                 return false;
1179
1180         for(int i = 0; i < utcp->nconnections; i++)
1181                 if(utcp->connections[i]->state != CLOSED && utcp->connections[i]->state != TIME_WAIT)
1182                         return true;
1183
1184         return false;
1185 }
1186
1187 struct utcp *utcp_init(utcp_accept_t accept, utcp_pre_accept_t pre_accept, utcp_send_t send, void *priv) {
1188         struct utcp *utcp = calloc(1, sizeof *utcp);
1189         if(!utcp)
1190                 return NULL;
1191
1192         if(!send) {
1193                 errno = EFAULT;
1194                 return NULL;
1195         }
1196
1197         utcp->accept = accept;
1198         utcp->pre_accept = pre_accept;
1199         utcp->send = send;
1200         utcp->priv = priv;
1201         utcp->mtu = 1000;
1202         utcp->timeout = 60;
1203
1204         return utcp;
1205 }
1206
1207 void utcp_exit(struct utcp *utcp) {
1208         if(!utcp)
1209                 return;
1210         for(int i = 0; i < utcp->nconnections; i++) {
1211                 if(!utcp->connections[i]->reapable)
1212                         debug("Warning, freeing unclosed connection %p\n", utcp->connections[i]);
1213                 buffer_exit(&utcp->connections[i]->sndbuf);
1214                 free(utcp->connections[i]);
1215         }
1216         free(utcp->connections);
1217         free(utcp);
1218 }
1219
1220 uint16_t utcp_get_mtu(struct utcp *utcp) {
1221         return utcp ? utcp->mtu : 0;
1222 }
1223
1224 void utcp_set_mtu(struct utcp *utcp, uint16_t mtu) {
1225         // TODO: handle overhead of the header
1226         if(utcp)
1227                 utcp->mtu = mtu;
1228 }
1229
1230 int utcp_get_user_timeout(struct utcp *u) {
1231         return u ? u->timeout : 0;
1232 }
1233
1234 void utcp_set_user_timeout(struct utcp *u, int timeout) {
1235         if(u)
1236                 u->timeout = timeout;
1237 }
1238
1239 size_t utcp_get_sndbuf(struct utcp_connection *c) {
1240         return c ? c->sndbuf.maxsize : 0;
1241 }
1242
1243 size_t utcp_get_sndbuf_free(struct utcp_connection *c) {
1244         if(c && (c->state == ESTABLISHED || c->state == CLOSE_WAIT))
1245                 return buffer_free(&c->sndbuf);
1246         else
1247                 return 0;
1248 }
1249
1250 void utcp_set_sndbuf(struct utcp_connection *c, size_t size) {
1251         if(!c)
1252                 return;
1253         c->sndbuf.maxsize = size;
1254         if(c->sndbuf.maxsize != size)
1255                 c->sndbuf.maxsize = -1;
1256 }
1257
1258 bool utcp_get_nodelay(struct utcp_connection *c) {
1259         return c ? c->nodelay : false;
1260 }
1261
1262 void utcp_set_nodelay(struct utcp_connection *c, bool nodelay) {
1263         if(c)
1264                 c->nodelay = nodelay;
1265 }
1266
1267 bool utcp_get_keepalive(struct utcp_connection *c) {
1268         return c ? c->keepalive : false;
1269 }
1270
1271 void utcp_set_keepalive(struct utcp_connection *c, bool keepalive) {
1272         if(c)
1273                 c->keepalive = keepalive;
1274 }
1275
1276 size_t utcp_get_outq(struct utcp_connection *c) {
1277         return c ? seqdiff(c->snd.nxt, c->snd.una) : 0;
1278 }
1279
1280 void utcp_set_recv_cb(struct utcp_connection *c, utcp_recv_t recv) {
1281         if(c)
1282                 c->recv = recv;
1283 }
1284
1285 void utcp_set_poll_cb(struct utcp_connection *c, utcp_poll_t poll) {
1286         if(c)
1287                 c->poll = poll;
1288 }
1289
1290 void utcp_set_accept_cb(struct utcp *utcp, utcp_accept_t accept, utcp_pre_accept_t pre_accept) {
1291         if(utcp) {
1292                 utcp->accept = accept;
1293                 utcp->pre_accept = pre_accept;
1294         }
1295 }