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

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