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_at(struct buffer *buf, size_t offset, const void *data, size_t len) {
 134         if(buf->maxsize <= buf->used)
 135                 return 0;
 136
 137         debug("buffer_put_at %zu %zu %zu\n", buf->used, offset, len);
 138
 139         size_t required = offset + len;
 140         if(required > buf->maxsize) {
 141                 if(offset >= buf->maxsize)
 142                         return 0;
 143                 abort();
 144                 len = buf->maxsize - offset;
 145                 required = buf->maxsize;
 146         }
 147
 148         if(required > buf->size) {
 149                 size_t newsize = buf->size;
 150                 if(!newsize) {
 151                         newsize = required;
 152                 } else {
 153                         do {
 154                                 newsize *= 2;
 155                         } while(newsize < buf->used + len);
 156                 }
 157                 if(newsize > buf->maxsize)
 158                         newsize = buf->maxsize;
 159                 char *newdata = realloc(buf->data, newsize);
 160                 if(!newdata)
 161                         return -1;
 162                 buf->data = newdata;
 163                 buf->size = newsize;
 164         }
 165
 166         memcpy(buf->data + offset, data, len);
 167         if(required > buf->used)
 168                 buf->used = required;
 169         return len;
 170 }
 171
 172 static ssize_t buffer_put(struct buffer *buf, const void *data, size_t len) {
 173         return buffer_put_at(buf, buf->used, data, len);
 174 }
 175
 176 // Get data from the buffer. data can be NULL.
 177 static ssize_t buffer_get(struct buffer *buf, void *data, size_t len) {
 178         if(len > buf->used)
 179                 len = buf->used;
 180         if(data)
 181                 memcpy(data, buf->data, len);
 182         if(len < buf->used)
 183                 memmove(buf->data, buf->data + len, buf->used - len);
 184         buf->used -= len;
 185         return len;
 186 }
 187
 188 // Copy data from the buffer without removing it.
 189 static ssize_t buffer_copy(struct buffer *buf, void *data, size_t offset, size_t len) {
 190         if(offset >= buf->used)
 191                 return 0;
 192         if(offset + len > buf->used)
 193                 len = buf->used - offset;
 194         memcpy(data, buf->data + offset, len);
 195         return len;
 196 }
 197
 198 static bool buffer_init(struct buffer *buf, uint32_t len, uint32_t maxlen) {
 199         memset(buf, 0, sizeof *buf);
 200         buf->data = malloc(len);
 201         if(!len)
 202                 return false;
 203         buf->size = len;
 204         buf->maxsize = maxlen;
 205         return true;
 206 }
 207
 208 static void buffer_exit(struct buffer *buf) {
 209         free(buf->data);
 210         memset(buf, 0, sizeof *buf);
 211 }
 212
 213 static uint32_t buffer_free(const struct buffer *buf) {
 214         return buf->maxsize - buf->used;
 215 }
 216
 217 // Connections are stored in a sorted list.
 218 // This gives O(log(N)) lookup time, O(N log(N)) insertion time and O(N) deletion time.
 219
 220 static int compare(const void *va, const void *vb) {
 221         assert(va && vb);
 222
 223         const struct utcp_connection *a = *(struct utcp_connection **)va;
 224         const struct utcp_connection *b = *(struct utcp_connection **)vb;
 225
 226         assert(a && b);
 227         assert(a->src && b->src);
 228
 229         int c = (int)a->src - (int)b->src;
 230         if(c)
 231                 return c;
 232         c = (int)a->dst - (int)b->dst;
 233         return c;
 234 }
 235
 236 static struct utcp_connection *find_connection(const struct utcp *utcp, uint16_t src, uint16_t dst) {
 237         if(!utcp->nconnections)
 238                 return NULL;
 239         struct utcp_connection key = {
 240                 .src = src,
 241                 .dst = dst,
 242         }, *keyp = &key;
 243         struct utcp_connection **match = bsearch(&keyp, utcp->connections, utcp->nconnections, sizeof *utcp->connections, compare);
 244         return match ? *match : NULL;
 245 }
 246
 247 static void free_connection(struct utcp_connection *c) {
 248         struct utcp *utcp = c->utcp;
 249         struct utcp_connection **cp = bsearch(&c, utcp->connections, utcp->nconnections, sizeof *utcp->connections, compare);
 250
 251         assert(cp);
 252
 253         int i = cp - utcp->connections;
 254         memmove(cp, cp + 1, (utcp->nconnections - i - 1) * sizeof *cp);
 255         utcp->nconnections--;
 256
 257         buffer_exit(&c->sndbuf);
 258         free(c);
 259 }
 260
 261 static struct utcp_connection *allocate_connection(struct utcp *utcp, uint16_t src, uint16_t dst) {
 262         // Check whether this combination of src and dst is free
 263
 264         if(src) {
 265                 if(find_connection(utcp, src, dst)) {
 266                         errno = EADDRINUSE;
 267                         return NULL;
 268                 }
 269         } else { // If src == 0, generate a random port number with the high bit set
 270                 if(utcp->nconnections >= 32767) {
 271                         errno = ENOMEM;
 272                         return NULL;
 273                 }
 274                 src = rand() | 0x8000;
 275                 while(find_connection(utcp, src, dst))
 276                         src++;
 277         }
 278
 279         // Allocate memory for the new connection
 280
 281         if(utcp->nconnections >= utcp->nallocated) {
 282                 if(!utcp->nallocated)
 283                         utcp->nallocated = 4;
 284                 else
 285                         utcp->nallocated *= 2;
 286                 struct utcp_connection **new_array = realloc(utcp->connections, utcp->nallocated * sizeof *utcp->connections);
 287                 if(!new_array)
 288                         return NULL;
 289                 utcp->connections = new_array;
 290         }
 291
 292         struct utcp_connection *c = calloc(1, sizeof *c);
 293         if(!c)
 294                 return NULL;
 295
 296         if(!buffer_init(&c->sndbuf, DEFAULT_SNDBUFSIZE, DEFAULT_MAXSNDBUFSIZE)) {
 297                 free(c);
 298                 return NULL;
 299         }
 300
 301         // Fill in the details
 302
 303         c->src = src;
 304         c->dst = dst;
 305 #ifdef UTCP_DEBUG
 306 #warning debugging
 307         c->snd.iss = 0;
 308 #else
 309         c->snd.iss = rand();
 310 #endif
 311         c->snd.una = c->snd.iss;
 312         c->snd.nxt = c->snd.iss + 1;
 313         c->rcv.wnd = utcp->mtu;
 314         c->snd.last = c->snd.nxt;
 315         c->snd.cwnd = utcp->mtu;
 316         c->utcp = utcp;
 317
 318         // Add it to the sorted list of connections
 319
 320         utcp->connections[utcp->nconnections++] = c;
 321         qsort(utcp->connections, utcp->nconnections, sizeof *utcp->connections, compare);
 322
 323         return c;
 324 }
 325
 326 struct utcp_connection *utcp_connect(struct utcp *utcp, uint16_t dst, utcp_recv_t recv, void *priv) {
 327         struct utcp_connection *c = allocate_connection(utcp, 0, dst);
 328         if(!c)
 329                 return NULL;
 330
 331         c->recv = recv;
 332         c->priv = priv;
 333
 334         struct hdr hdr;
 335
 336         hdr.src = c->src;
 337         hdr.dst = c->dst;
 338         hdr.seq = c->snd.iss;
 339         hdr.ack = 0;
 340         hdr.wnd = c->rcv.wnd;
 341         hdr.ctl = SYN;
 342         hdr.aux = 0;
 343
 344         set_state(c, SYN_SENT);
 345
 346         print_packet(utcp, "send", &hdr, sizeof hdr);
 347         utcp->send(utcp, &hdr, sizeof hdr);
 348
 349         gettimeofday(&c->conn_timeout, NULL);
 350         c->conn_timeout.tv_sec += utcp->timeout;
 351
 352         return c;
 353 }
 354
 355 void utcp_accept(struct utcp_connection *c, utcp_recv_t recv, void *priv) {
 356         if(c->reapable || c->state != SYN_RECEIVED) {
 357                 debug("Error: accept() called on invalid connection %p in state %s\n", c, strstate[c->state]);
 358                 return;
 359         }
 360
 361         debug("%p accepted, %p %p\n", c, recv, priv);
 362         c->recv = recv;
 363         c->priv = priv;
 364         set_state(c, ESTABLISHED);
 365 }
 366
 367 static void ack(struct utcp_connection *c, bool sendatleastone) {
 368         int32_t left = seqdiff(c->snd.last, c->snd.nxt);
 369         int32_t cwndleft = c->snd.cwnd - seqdiff(c->snd.nxt, c->snd.una);
 370
 371         assert(left >= 0);
 372
 373         if(cwndleft <= 0)
 374                 cwndleft = 0;
 375
 376         if(cwndleft < left)
 377                 left = cwndleft;
 378
 379         if(!left && !sendatleastone)
 380                 return;
 381
 382         struct {
 383                 struct hdr hdr;
 384                 char data[];
 385         } *pkt;
 386
 387         pkt = malloc(sizeof pkt->hdr + c->utcp->mtu);
 388         if(!pkt)
 389                 return;
 390
 391         pkt->hdr.src = c->src;
 392         pkt->hdr.dst = c->dst;
 393         pkt->hdr.ack = c->rcv.nxt;
 394         pkt->hdr.wnd = c->snd.wnd;
 395         pkt->hdr.ctl = ACK;
 396         pkt->hdr.aux = 0;
 397
 398         do {
 399                 uint32_t seglen = left > c->utcp->mtu ? c->utcp->mtu : left;
 400                 pkt->hdr.seq = c->snd.nxt;
 401
 402                 buffer_copy(&c->sndbuf, pkt->data, seqdiff(c->snd.nxt, c->snd.una), seglen);
 403
 404                 c->snd.nxt += seglen;
 405                 left -= seglen;
 406
 407                 if(seglen && fin_wanted(c, c->snd.nxt)) {
 408                         seglen--;
 409                         pkt->hdr.ctl |= FIN;
 410                 }
 411
 412                 print_packet(c->utcp, "send", pkt, sizeof pkt->hdr + seglen);
 413                 c->utcp->send(c->utcp, pkt, sizeof pkt->hdr + seglen);
 414         } while(left);
 415
 416         free(pkt);
 417 }
 418
 419 ssize_t utcp_send(struct utcp_connection *c, const void *data, size_t len) {
 420         if(c->reapable) {
 421                 debug("Error: send() called on closed connection %p\n", c);
 422                 errno = EBADF;
 423                 return -1;
 424         }
 425
 426         switch(c->state) {
 427         case CLOSED:
 428         case LISTEN:
 429         case SYN_SENT:
 430         case SYN_RECEIVED:
 431                 debug("Error: send() called on unconnected connection %p\n", c);
 432                 errno = ENOTCONN;
 433                 return -1;
 434         case ESTABLISHED:
 435         case CLOSE_WAIT:
 436                 break;
 437         case FIN_WAIT_1:
 438         case FIN_WAIT_2:
 439         case CLOSING:
 440         case LAST_ACK:
 441         case TIME_WAIT:
 442                 debug("Error: send() called on closing connection %p\n", c);
 443                 errno = EPIPE;
 444                 return -1;
 445         }
 446
 447         // Add data to send buffer
 448
 449         if(!len)
 450                 return 0;
 451
 452         if(!data) {
 453                 errno = EFAULT;
 454                 return -1;
 455         }
 456
 457         len = buffer_put(&c->sndbuf, data, len);
 458         if(len <= 0) {
 459                 errno = EWOULDBLOCK;
 460                 return 0;
 461         }
 462
 463         c->snd.last += len;
 464         ack(c, false);
 465         return len;
 466 }
 467
 468 static void swap_ports(struct hdr *hdr) {
 469         uint16_t tmp = hdr->src;
 470         hdr->src = hdr->dst;
 471         hdr->dst = tmp;
 472 }
 473
 474 static void retransmit(struct utcp_connection *c) {
 475         if(c->state == CLOSED || c->snd.nxt == c->snd.una)
 476                 return;
 477
 478         struct utcp *utcp = c->utcp;
 479
 480         struct {
 481                 struct hdr hdr;
 482                 char data[];
 483         } *pkt;
 484
 485         pkt = malloc(sizeof pkt->hdr + c->utcp->mtu);
 486         if(!pkt)
 487                 return;
 488
 489         pkt->hdr.src = c->src;
 490         pkt->hdr.dst = c->dst;
 491
 492         switch(c->state) {
 493                 case SYN_SENT:
 494                         // Send our SYN again
 495                         pkt->hdr.seq = c->snd.iss;
 496                         pkt->hdr.ack = 0;
 497                         pkt->hdr.wnd = c->rcv.wnd;
 498                         pkt->hdr.ctl = SYN;
 499                         print_packet(c->utcp, "rtrx", pkt, sizeof pkt->hdr);
 500                         utcp->send(utcp, pkt, sizeof pkt->hdr);
 501                         break;
 502
 503                 case SYN_RECEIVED:
 504                         // Send SYNACK again
 505                         pkt->hdr.seq = c->snd.nxt;
 506                         pkt->hdr.ack = c->rcv.nxt;
 507                         pkt->hdr.ctl = SYN | ACK;
 508                         print_packet(c->utcp, "rtrx", pkt, sizeof pkt->hdr);
 509                         utcp->send(utcp, pkt, sizeof pkt->hdr);
 510                         break;
 511
 512                 case ESTABLISHED:
 513                 case FIN_WAIT_1:
 514                 case CLOSE_WAIT:
 515                 case CLOSING:
 516                 case LAST_ACK:
 517                         // Send unacked data again.
 518                         pkt->hdr.seq = c->snd.una;
 519                         pkt->hdr.ack = c->rcv.nxt;
 520                         pkt->hdr.ctl = ACK;
 521                         uint32_t len = seqdiff(c->snd.last, c->snd.una);
 522                         if(len > utcp->mtu)
 523                                 len = utcp->mtu;
 524                         if(fin_wanted(c, c->snd.una + len)) {
 525                                 len--;
 526                                 pkt->hdr.ctl |= FIN;
 527                         }
 528                         c->snd.nxt = c->snd.una + len;
 529                         buffer_copy(&c->sndbuf, pkt->data, 0, len);
 530                         print_packet(c->utcp, "rtrx", pkt, sizeof pkt->hdr + len);
 531                         utcp->send(utcp, pkt, sizeof pkt->hdr + len);
 532                         break;
 533
 534                 case CLOSED:
 535                 case LISTEN:
 536                 case TIME_WAIT:
 537                 case FIN_WAIT_2:
 538                         // We shouldn't need to retransmit anything in this state.
 539 #ifdef UTCP_DEBUG
 540                         abort();
 541 #endif
 542                         timerclear(&c->rtrx_timeout);
 543                         break;
 544         }
 545
 546         free(pkt);
 547 }
 548
 549
 550 ssize_t utcp_recv(struct utcp *utcp, const void *data, size_t len) {
 551         if(!utcp) {
 552                 errno = EFAULT;
 553                 return -1;
 554         }
 555
 556         if(!len)
 557                 return 0;
 558
 559         if(!data) {
 560                 errno = EFAULT;
 561                 return -1;
 562         }
 563
 564         print_packet(utcp, "recv", data, len);
 565
 566         // Drop packets smaller than the header
 567
 568         struct hdr hdr;
 569         if(len < sizeof hdr) {
 570                 errno = EBADMSG;
 571                 return -1;
 572         }
 573
 574         // Make a copy from the potentially unaligned data to a struct hdr
 575
 576         memcpy(&hdr, data, sizeof hdr);
 577         data += sizeof hdr;
 578         len -= sizeof hdr;
 579
 580         // Drop packets with an unknown CTL flag
 581
 582         if(hdr.ctl & ~(SYN | ACK | RST | FIN)) {
 583                 errno = EBADMSG;
 584                 return -1;
 585         }
 586
 587         // Try to match the packet to an existing connection
 588
 589         struct utcp_connection *c = find_connection(utcp, hdr.dst, hdr.src);
 590
 591         // Is it for a new connection?
 592
 593         if(!c) {
 594                 // Ignore RST packets
 595
 596                 if(hdr.ctl & RST)
 597                         return 0;
 598
 599                 // Is it a SYN packet and are we LISTENing?
 600
 601                 if(hdr.ctl & SYN && !(hdr.ctl & ACK) && utcp->accept) {
 602                         // If we don't want to accept it, send a RST back
 603                         if((utcp->pre_accept && !utcp->pre_accept(utcp, hdr.dst))) {
 604                                 len = 1;
 605                                 goto reset;
 606                         }
 607
 608                         // Try to allocate memory, otherwise send a RST back
 609                         c = allocate_connection(utcp, hdr.dst, hdr.src);
 610                         if(!c) {
 611                                 len = 1;
 612                                 goto reset;
 613                         }
 614
 615                         // Return SYN+ACK, go to SYN_RECEIVED state
 616                         c->snd.wnd = hdr.wnd;
 617                         c->rcv.irs = hdr.seq;
 618                         c->rcv.nxt = c->rcv.irs + 1;
 619                         set_state(c, SYN_RECEIVED);
 620
 621                         hdr.dst = c->dst;
 622                         hdr.src = c->src;
 623                         hdr.ack = c->rcv.irs + 1;
 624                         hdr.seq = c->snd.iss;
 625                         hdr.ctl = SYN | ACK;
 626                         print_packet(c->utcp, "send", &hdr, sizeof hdr);
 627                         utcp->send(utcp, &hdr, sizeof hdr);
 628                 } else {
 629                         // No, we don't want your packets, send a RST back
 630                         len = 1;
 631                         goto reset;
 632                 }
 633
 634                 return 0;
 635         }
 636
 637         debug("%p state %s\n", c->utcp, strstate[c->state]);
 638
 639         // In case this is for a CLOSED connection, ignore the packet.
 640         // TODO: make it so incoming packets can never match a CLOSED connection.
 641
 642         if(c->state == CLOSED)
 643                 return 0;
 644
 645         // It is for an existing connection.
 646
 647         uint32_t prevrcvnxt = c->rcv.nxt;
 648
 649         // 1. Drop invalid packets.
 650
 651         // 1a. Drop packets that should not happen in our current state.
 652
 653         switch(c->state) {
 654         case SYN_SENT:
 655         case SYN_RECEIVED:
 656         case ESTABLISHED:
 657         case FIN_WAIT_1:
 658         case FIN_WAIT_2:
 659         case CLOSE_WAIT:
 660         case CLOSING:
 661         case LAST_ACK:
 662         case TIME_WAIT:
 663                 break;
 664         default:
 665 #ifdef UTCP_DEBUG
 666                 abort();
 667 #endif
 668                 break;
 669         }
 670
 671         // 1b. Drop packets with a sequence number not in our receive window.
 672
 673         bool acceptable;
 674
 675         if(c->state == SYN_SENT)
 676                 acceptable = true;
 677
 678         // TODO: handle packets overlapping c->rcv.nxt.
 679 #if 0
 680         // Only use this when accepting out-of-order packets.
 681         else if(len == 0)
 682                 if(c->rcv.wnd == 0)
 683                         acceptable = hdr.seq == c->rcv.nxt;
 684                 else
 685                         acceptable = (seqdiff(hdr.seq, c->rcv.nxt) >= 0 && seqdiff(hdr.seq, c->rcv.nxt + c->rcv.wnd) < 0);
 686         else
 687                 if(c->rcv.wnd == 0)
 688                         // We don't accept data when the receive window is zero.
 689                         acceptable = false;
 690                 else
 691                         // Both start and end of packet must be within the receive window
 692                         acceptable = (seqdiff(hdr.seq, c->rcv.nxt) >= 0 && seqdiff(hdr.seq, c->rcv.nxt + c->rcv.wnd) < 0)
 693                                 || (seqdiff(hdr.seq + len + 1, c->rcv.nxt) >= 0 && seqdiff(hdr.seq + len - 1, c->rcv.nxt + c->rcv.wnd) < 0);
 694 #else
 695         if(c->state != SYN_SENT)
 696                 acceptable = hdr.seq == c->rcv.nxt;
 697 #endif
 698
 699         if(!acceptable) {
 700                 debug("Packet not acceptable, %u <= %u + %zu < %u\n", c->rcv.nxt, hdr.seq, len, c->rcv.nxt + c->rcv.wnd);
 701                 // Ignore unacceptable RST packets.
 702                 if(hdr.ctl & RST)
 703                         return 0;
 704                 // Otherwise, send an ACK back in the hope things improve.
 705                 ack(c, true);
 706                 return 0;
 707         }
 708
 709         c->snd.wnd = hdr.wnd; // TODO: move below
 710
 711         // 1c. Drop packets with an invalid ACK.
 712         // ackno should not roll back, and it should also not be bigger than snd.nxt.
 713
 714         if(hdr.ctl & ACK && (seqdiff(hdr.ack, c->snd.nxt) > 0 || seqdiff(hdr.ack, c->snd.una) < 0)) {
 715                 debug("Packet ack seqno out of range, %u %u %u\n", hdr.ack, c->snd.una, c->snd.nxt);
 716                 // Ignore unacceptable RST packets.
 717                 if(hdr.ctl & RST)
 718                         return 0;
 719                 goto reset;
 720         }
 721
 722         // 2. Handle RST packets
 723
 724         if(hdr.ctl & RST) {
 725                 switch(c->state) {
 726                 case SYN_SENT:
 727                         if(!(hdr.ctl & ACK))
 728                                 return 0;
 729                         // The peer has refused our connection.
 730                         set_state(c, CLOSED);
 731                         errno = ECONNREFUSED;
 732                         if(c->recv)
 733                                 c->recv(c, NULL, 0);
 734                         return 0;
 735                 case SYN_RECEIVED:
 736                         if(hdr.ctl & ACK)
 737                                 return 0;
 738                         // We haven't told the application about this connection yet. Silently delete.
 739                         free_connection(c);
 740                         return 0;
 741                 case ESTABLISHED:
 742                 case FIN_WAIT_1:
 743                 case FIN_WAIT_2:
 744                 case CLOSE_WAIT:
 745                         if(hdr.ctl & ACK)
 746                                 return 0;
 747                         // The peer has aborted our connection.
 748                         set_state(c, CLOSED);
 749                         errno = ECONNRESET;
 750                         if(c->recv)
 751                                 c->recv(c, NULL, 0);
 752                         return 0;
 753                 case CLOSING:
 754                 case LAST_ACK:
 755                 case TIME_WAIT:
 756                         if(hdr.ctl & ACK)
 757                                 return 0;
 758                         // As far as the application is concerned, the connection has already been closed.
 759                         // If it has called utcp_close() already, we can immediately free this connection.
 760                         if(c->reapable) {
 761                                 free_connection(c);
 762                                 return 0;
 763                         }
 764                         // Otherwise, immediately move to the CLOSED state.
 765                         set_state(c, CLOSED);
 766                         return 0;
 767                 default:
 768 #ifdef UTCP_DEBUG
 769                         abort();
 770 #endif
 771                         break;
 772                 }
 773         }
 774
 775         // 3. Advance snd.una
 776
 777         uint32_t advanced = seqdiff(hdr.ack, c->snd.una);
 778         prevrcvnxt = c->rcv.nxt;
 779
 780         if(advanced) {
 781                 int32_t data_acked = advanced;
 782
 783                 switch(c->state) {
 784                         case SYN_SENT:
 785                         case SYN_RECEIVED:
 786                                 data_acked--;
 787                                 break;
 788                         // TODO: handle FIN as well.
 789                         default:
 790                                 break;
 791                 }
 792
 793                 assert(data_acked >= 0);
 794
 795                 int32_t bufused = seqdiff(c->snd.last, c->snd.una);
 796                 assert(data_acked <= bufused);
 797
 798                 if(data_acked)
 799                         buffer_get(&c->sndbuf, NULL, data_acked);
 800
 801                 c->snd.una = hdr.ack;
 802
 803                 c->dupack = 0;
 804                 c->snd.cwnd += utcp->mtu;
 805                 if(c->snd.cwnd > c->sndbuf.maxsize)
 806                         c->snd.cwnd = c->sndbuf.maxsize;
 807
 808                 // Check if we have sent a FIN that is now ACKed.
 809                 switch(c->state) {
 810                 case FIN_WAIT_1:
 811                         if(c->snd.una == c->snd.last)
 812                                 set_state(c, FIN_WAIT_2);
 813                         break;
 814                 case CLOSING:
 815                         if(c->snd.una == c->snd.last) {
 816                                 gettimeofday(&c->conn_timeout, NULL);
 817                                 c->conn_timeout.tv_sec += 60;
 818                                 set_state(c, TIME_WAIT);
 819                         }
 820                         break;
 821                 default:
 822                         break;
 823                 }
 824         } else {
 825                 if(!len) {
 826                         c->dupack++;
 827                         if(c->dupack == 3) {
 828                                 debug("Triplicate ACK\n");
 829                                 //TODO: Resend one packet and go to fast recovery mode. See RFC 6582.
 830                                 //We do a very simple variant here; reset the nxt pointer to the last acknowledged packet from the peer.
 831                                 //This will cause us to start retransmitting, but at the same speed as the incoming ACKs arrive,
 832                                 //thus preventing a drop in speed.
 833                                 c->snd.nxt = c->snd.una;
 834                         }
 835                 }
 836         }
 837
 838         // 4. Update timers
 839
 840         if(advanced) {
 841                 timerclear(&c->conn_timeout); // It will be set anew in utcp_timeout() if c->snd.una != c->snd.nxt.
 842                 if(c->snd.una == c->snd.nxt)
 843                         timerclear(&c->rtrx_timeout);
 844         }
 845
 846         // 5. Process SYN stuff
 847
 848         if(hdr.ctl & SYN) {
 849                 switch(c->state) {
 850                 case SYN_SENT:
 851                         // This is a SYNACK. It should always have ACKed the SYN.
 852                         if(!advanced)
 853                                 goto reset;
 854                         c->rcv.irs = hdr.seq;
 855                         c->rcv.nxt = hdr.seq;
 856                         set_state(c, ESTABLISHED);
 857                         // TODO: notify application of this somehow.
 858                         break;
 859                 case SYN_RECEIVED:
 860                 case ESTABLISHED:
 861                 case FIN_WAIT_1:
 862                 case FIN_WAIT_2:
 863                 case CLOSE_WAIT:
 864                 case CLOSING:
 865                 case LAST_ACK:
 866                 case TIME_WAIT:
 867                         // Ehm, no. We should never receive a second SYN.
 868                         goto reset;
 869                 default:
 870 #ifdef UTCP_DEBUG
 871                         abort();
 872 #endif
 873                         return 0;
 874                 }
 875
 876                 // SYN counts as one sequence number
 877                 c->rcv.nxt++;
 878         }
 879
 880         // 6. Process new data
 881
 882         if(c->state == SYN_RECEIVED) {
 883                 // This is the ACK after the SYNACK. It should always have ACKed the SYNACK.
 884                 if(!advanced)
 885                         goto reset;
 886
 887                 // Are we still LISTENing?
 888                 if(utcp->accept)
 889                         utcp->accept(c, c->src);
 890
 891                 if(c->state != ESTABLISHED) {
 892                         set_state(c, CLOSED);
 893                         c->reapable = true;
 894                         goto reset;
 895                 }
 896         }
 897
 898         if(len) {
 899                 switch(c->state) {
 900                 case SYN_SENT:
 901                 case SYN_RECEIVED:
 902                         // This should never happen.
 903 #ifdef UTCP_DEBUG
 904                         abort();
 905 #endif
 906                         return 0;
 907                 case ESTABLISHED:
 908                 case FIN_WAIT_1:
 909                 case FIN_WAIT_2:
 910                         break;
 911                 case CLOSE_WAIT:
 912                 case CLOSING:
 913                 case LAST_ACK:
 914                 case TIME_WAIT:
 915                         // Ehm no, We should never receive more data after a FIN.
 916                         goto reset;
 917                 default:
 918 #ifdef UTCP_DEBUG
 919                         abort();
 920 #endif
 921                         return 0;
 922                 }
 923
 924                 ssize_t rxd;
 925
 926                 if(c->recv) {
 927                         rxd = c->recv(c, data, len);
 928                         if(rxd != len) {
 929                                 // TODO: once we have a receive buffer, handle the application not accepting all data.
 930                                 abort();
 931                         }
 932                         if(rxd < 0)
 933                                 rxd = 0;
 934                         else if(rxd > len)
 935                                 rxd = len; // Bad application, bad!
 936                 } else {
 937                         rxd = len;
 938                 }
 939
 940                 c->rcv.nxt += len;
 941         }
 942
 943         // 7. Process FIN stuff
 944
 945         if(hdr.ctl & FIN) {
 946                 switch(c->state) {
 947                 case SYN_SENT:
 948                 case SYN_RECEIVED:
 949                         // This should never happen.
 950 #ifdef UTCP_DEBUG
 951                         abort();
 952 #endif
 953                         break;
 954                 case ESTABLISHED:
 955                         set_state(c, CLOSE_WAIT);
 956                         break;
 957                 case FIN_WAIT_1:
 958                         set_state(c, CLOSING);
 959                         break;
 960                 case FIN_WAIT_2:
 961                         gettimeofday(&c->conn_timeout, NULL);
 962                         c->conn_timeout.tv_sec += 60;
 963                         set_state(c, TIME_WAIT);
 964                         break;
 965                 case CLOSE_WAIT:
 966                 case CLOSING:
 967                 case LAST_ACK:
 968                 case TIME_WAIT:
 969                         // Ehm, no. We should never receive a second FIN.
 970                         goto reset;
 971                 default:
 972 #ifdef UTCP_DEBUG
 973                         abort();
 974 #endif
 975                         break;
 976                 }
 977
 978                 // FIN counts as one sequence number
 979                 c->rcv.nxt++;
 980                 len++;
 981
 982                 // Inform the application that the peer closed the connection.
 983                 if(c->recv) {
 984                         errno = 0;
 985                         c->recv(c, NULL, 0);
 986                 }
 987         }
 988
 989         // Now we send something back if:
 990         // - we advanced rcv.nxt (ie, we got some data that needs to be ACKed)
 991         //   -> sendatleastone = true
 992         // - or we got an ack, so we should maybe send a bit more data
 993         //   -> sendatleastone = false
 994
 995 ack:
 996         ack(c, prevrcvnxt != c->rcv.nxt);
 997         return 0;
 998
 999 reset:
1000         swap_ports(&hdr);
1001         hdr.wnd = 0;
1002         if(hdr.ctl & ACK) {
1003                 hdr.seq = hdr.ack;
1004                 hdr.ctl = RST;
1005         } else {
1006                 hdr.ack = hdr.seq + len;
1007                 hdr.seq = 0;
1008                 hdr.ctl = RST | ACK;
1009         }
1010         print_packet(utcp, "send", &hdr, sizeof hdr);
1011         utcp->send(utcp, &hdr, sizeof hdr);
1012         return 0;
1013
1014 }
1015
1016 int utcp_shutdown(struct utcp_connection *c, int dir) {
1017         debug("%p shutdown %d at %u\n", c ? c->utcp : NULL, dir, c ? c->snd.last : 0);
1018         if(!c) {
1019                 errno = EFAULT;
1020                 return -1;
1021         }
1022
1023         if(c->reapable) {
1024                 debug("Error: shutdown() called on closed connection %p\n", c);
1025                 errno = EBADF;
1026                 return -1;
1027         }
1028
1029         if(!(dir == UTCP_SHUT_RD || dir == UTCP_SHUT_WR || dir == UTCP_SHUT_RDWR)) {
1030                 errno = EINVAL;
1031                 return -1;
1032         }
1033
1034         // TCP does not have a provision for stopping incoming packets.
1035         // The best we can do is to just ignore them.
1036         if(dir == UTCP_SHUT_RD || dir == UTCP_SHUT_RDWR)
1037                 c->recv = NULL;
1038
1039         // The rest of the code deals with shutting down writes.
1040         if(dir == UTCP_SHUT_RD)
1041                 return 0;
1042
1043         switch(c->state) {
1044         case CLOSED:
1045         case LISTEN:
1046                 errno = ENOTCONN;
1047                 return -1;
1048
1049         case SYN_SENT:
1050                 set_state(c, CLOSED);
1051                 return 0;
1052
1053         case SYN_RECEIVED:
1054         case ESTABLISHED:
1055                 set_state(c, FIN_WAIT_1);
1056                 break;
1057         case FIN_WAIT_1:
1058         case FIN_WAIT_2:
1059                 return 0;
1060         case CLOSE_WAIT:
1061                 set_state(c, CLOSING);
1062                 break;
1063
1064         case CLOSING:
1065         case LAST_ACK:
1066         case TIME_WAIT:
1067                 return 0;
1068         }
1069
1070         c->snd.last++;
1071
1072         ack(c, false);
1073         return 0;
1074 }
1075
1076 int utcp_close(struct utcp_connection *c) {
1077         if(utcp_shutdown(c, SHUT_RDWR))
1078                 return -1;
1079         c->recv = NULL;
1080         c->poll = NULL;
1081         c->reapable = true;
1082         return 0;
1083 }
1084
1085 int utcp_abort(struct utcp_connection *c) {
1086         if(!c) {
1087                 errno = EFAULT;
1088                 return -1;
1089         }
1090
1091         if(c->reapable) {
1092                 debug("Error: abort() called on closed connection %p\n", c);
1093                 errno = EBADF;
1094                 return -1;
1095         }
1096
1097         c->recv = NULL;
1098         c->poll = NULL;
1099         c->reapable = true;
1100
1101         switch(c->state) {
1102         case CLOSED:
1103                 return 0;
1104         case LISTEN:
1105         case SYN_SENT:
1106         case CLOSING:
1107         case LAST_ACK:
1108         case TIME_WAIT:
1109                 set_state(c, CLOSED);
1110                 return 0;
1111
1112         case SYN_RECEIVED:
1113         case ESTABLISHED:
1114         case FIN_WAIT_1:
1115         case FIN_WAIT_2:
1116         case CLOSE_WAIT:
1117                 set_state(c, CLOSED);
1118                 break;
1119         }
1120
1121         // Send RST
1122
1123         struct hdr hdr;
1124
1125         hdr.src = c->src;
1126         hdr.dst = c->dst;
1127         hdr.seq = c->snd.nxt;
1128         hdr.ack = 0;
1129         hdr.wnd = 0;
1130         hdr.ctl = RST;
1131
1132         print_packet(c->utcp, "send", &hdr, sizeof hdr);
1133         c->utcp->send(c->utcp, &hdr, sizeof hdr);
1134         return 0;
1135 }
1136
1137 /* Handle timeouts.
1138  * One call to this function will loop through all connections,
1139  * checking if something needs to be resent or not.
1140  * The return value is the time to the next timeout in milliseconds,
1141  * or maybe a negative value if the timeout is infinite.
1142  */
1143 struct timeval utcp_timeout(struct utcp *utcp) {
1144         struct timeval now;
1145         gettimeofday(&now, NULL);
1146         struct timeval next = {now.tv_sec + 3600, now.tv_usec};
1147
1148         for(int i = 0; i < utcp->nconnections; i++) {
1149                 struct utcp_connection *c = utcp->connections[i];
1150                 if(!c)
1151                         continue;
1152
1153                 if(c->state == CLOSED) {
1154                         if(c->reapable) {
1155                                 debug("Reaping %p\n", c);
1156                                 free_connection(c);
1157                                 i--;
1158                         }
1159                         continue;
1160                 }
1161
1162                 if(timerisset(&c->conn_timeout) && timercmp(&c->conn_timeout, &now, <)) {
1163                         errno = ETIMEDOUT;
1164                         c->state = CLOSED;
1165                         if(c->recv)
1166                                 c->recv(c, NULL, 0);
1167                         continue;
1168                 }
1169
1170                 if(timerisset(&c->rtrx_timeout) && timercmp(&c->rtrx_timeout, &now, <)) {
1171                         retransmit(c);
1172                 }
1173
1174                 if(c->poll && buffer_free(&c->sndbuf) && (c->state == ESTABLISHED || c->state == CLOSE_WAIT))
1175                         c->poll(c, buffer_free(&c->sndbuf));
1176
1177                 if(timerisset(&c->conn_timeout) && timercmp(&c->conn_timeout, &next, <))
1178                         next = c->conn_timeout;
1179
1180                 if(c->snd.nxt != c->snd.una) {
1181                         c->rtrx_timeout = now;
1182                         c->rtrx_timeout.tv_sec++;
1183                 } else {
1184                         timerclear(&c->rtrx_timeout);
1185                 }
1186
1187                 if(timerisset(&c->rtrx_timeout) && timercmp(&c->rtrx_timeout, &next, <))
1188                         next = c->rtrx_timeout;
1189         }
1190
1191         struct timeval diff;
1192         timersub(&next, &now, &diff);
1193         return diff;
1194 }
1195
1196 bool utcp_is_active(struct utcp *utcp) {
1197         if(!utcp)
1198                 return false;
1199
1200         for(int i = 0; i < utcp->nconnections; i++)
1201                 if(utcp->connections[i]->state != CLOSED && utcp->connections[i]->state != TIME_WAIT)
1202                         return true;
1203
1204         return false;
1205 }
1206
1207 struct utcp *utcp_init(utcp_accept_t accept, utcp_pre_accept_t pre_accept, utcp_send_t send, void *priv) {
1208         struct utcp *utcp = calloc(1, sizeof *utcp);
1209         if(!utcp)
1210                 return NULL;
1211
1212         if(!send) {
1213                 errno = EFAULT;
1214                 return NULL;
1215         }
1216
1217         utcp->accept = accept;
1218         utcp->pre_accept = pre_accept;
1219         utcp->send = send;
1220         utcp->priv = priv;
1221         utcp->mtu = 1000;
1222         utcp->timeout = 60;
1223
1224         return utcp;
1225 }
1226
1227 void utcp_exit(struct utcp *utcp) {
1228         if(!utcp)
1229                 return;
1230         for(int i = 0; i < utcp->nconnections; i++) {
1231                 if(!utcp->connections[i]->reapable)
1232                         debug("Warning, freeing unclosed connection %p\n", utcp->connections[i]);
1233                 buffer_exit(&utcp->connections[i]->sndbuf);
1234                 free(utcp->connections[i]);
1235         }
1236         free(utcp->connections);
1237         free(utcp);
1238 }
1239
1240 uint16_t utcp_get_mtu(struct utcp *utcp) {
1241         return utcp ? utcp->mtu : 0;
1242 }
1243
1244 void utcp_set_mtu(struct utcp *utcp, uint16_t mtu) {
1245         // TODO: handle overhead of the header
1246         if(utcp)
1247                 utcp->mtu = mtu;
1248 }
1249
1250 int utcp_get_user_timeout(struct utcp *u) {
1251         return u ? u->timeout : 0;
1252 }
1253
1254 void utcp_set_user_timeout(struct utcp *u, int timeout) {
1255         if(u)
1256                 u->timeout = timeout;
1257 }
1258
1259 size_t utcp_get_sndbuf(struct utcp_connection *c) {
1260         return c ? c->sndbuf.maxsize : 0;
1261 }
1262
1263 size_t utcp_get_sndbuf_free(struct utcp_connection *c) {
1264         if(c && (c->state == ESTABLISHED || c->state == CLOSE_WAIT))
1265                 return buffer_free(&c->sndbuf);
1266         else
1267                 return 0;
1268 }
1269
1270 void utcp_set_sndbuf(struct utcp_connection *c, size_t size) {
1271         if(!c)
1272                 return;
1273         c->sndbuf.maxsize = size;
1274         if(c->sndbuf.maxsize != size)
1275                 c->sndbuf.maxsize = -1;
1276 }
1277
1278 bool utcp_get_nodelay(struct utcp_connection *c) {
1279         return c ? c->nodelay : false;
1280 }
1281
1282 void utcp_set_nodelay(struct utcp_connection *c, bool nodelay) {
1283         if(c)
1284                 c->nodelay = nodelay;
1285 }
1286
1287 bool utcp_get_keepalive(struct utcp_connection *c) {
1288         return c ? c->keepalive : false;
1289 }
1290
1291 void utcp_set_keepalive(struct utcp_connection *c, bool keepalive) {
1292         if(c)
1293                 c->keepalive = keepalive;
1294 }
1295
1296 size_t utcp_get_outq(struct utcp_connection *c) {
1297         return c ? seqdiff(c->snd.nxt, c->snd.una) : 0;
1298 }
1299
1300 void utcp_set_recv_cb(struct utcp_connection *c, utcp_recv_t recv) {
1301         if(c)
1302                 c->recv = recv;
1303 }
1304
1305 void utcp_set_poll_cb(struct utcp_connection *c, utcp_poll_t poll) {
1306         if(c)
1307                 c->poll = poll;
1308 }
1309
1310 void utcp_set_accept_cb(struct utcp *utcp, utcp_accept_t accept, utcp_pre_accept_t pre_accept) {
1311         if(utcp) {
1312                 utcp->accept = accept;
1313                 utcp->pre_accept = pre_accept;
1314         }
1315 }