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

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