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 #ifndef max
  57 #define max(a, b) ((a) > (b) ? (a) : (b))
  58 #endif
  59
  60 #ifdef UTCP_DEBUG
  61 #include <stdarg.h>
  62
  63 static void debug(const char *format, ...) {
  64         va_list ap;
  65         va_start(ap, format);
  66         vfprintf(stderr, format, ap);
  67         va_end(ap);
  68 }
  69
  70 static void print_packet(struct utcp *utcp, const char *dir, const void *pkt, size_t len) {
  71         struct hdr hdr;
  72         if(len < sizeof hdr) {
  73                 debug("%p %s: short packet (%zu bytes)\n", utcp, dir, len);
  74                 return;
  75         }
  76
  77         memcpy(&hdr, pkt, sizeof hdr);
  78         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);
  79         if(hdr.ctl & SYN)
  80                 debug("SYN");
  81         if(hdr.ctl & RST)
  82                 debug("RST");
  83         if(hdr.ctl & FIN)
  84                 debug("FIN");
  85         if(hdr.ctl & ACK)
  86                 debug("ACK");
  87
  88         if(len > sizeof hdr) {
  89                 uint32_t datalen = len - sizeof hdr;
  90                 uint8_t *str = malloc((datalen << 1) + 7);
  91                 memcpy(str, " data=", 6);
  92                 uint8_t *strptr = str + 6;
  93                 const uint8_t *data = pkt;
  94                 const uint8_t *dataend = data + datalen;
  95
  96                 while(data != dataend) {
  97                         *strptr = (*data >> 4) > 9? (*data >> 4) + 55 : (*data >> 4) + 48;
  98                         ++strptr;
  99                         *strptr = (*data & 0xf) > 9? (*data & 0xf) + 55 : (*data & 0xf) + 48;
 100                         ++strptr;
 101                         ++data;
 102                 }
 103                 *strptr = 0;
 104
 105                 debug(str);
 106                 free(str);
 107         }
 108
 109         debug("\n");
 110 }
 111 #else
 112 #define debug(...)
 113 #define print_packet(...)
 114 #endif
 115
 116 static void set_state(struct utcp_connection *c, enum state state) {
 117         c->state = state;
 118         if(state == ESTABLISHED)
 119                 timerclear(&c->conn_timeout);
 120         debug("%p new state: %s\n", c->utcp, strstate[state]);
 121 }
 122
 123 static bool fin_wanted(struct utcp_connection *c, uint32_t seq) {
 124         if(seq != c->snd.last)
 125                 return false;
 126         switch(c->state) {
 127         case FIN_WAIT_1:
 128         case CLOSING:
 129         case LAST_ACK:
 130                 return true;
 131         default:
 132                 return false;
 133         }
 134 }
 135
 136 static inline void list_connections(struct utcp *utcp) {
 137         debug("%p has %d connections:\n", utcp, utcp->nconnections);
 138         for(int i = 0; i < utcp->nconnections; i++)
 139                 debug("  %u -> %u state %s\n", utcp->connections[i]->src, utcp->connections[i]->dst, strstate[utcp->connections[i]->state]);
 140 }
 141
 142 static int32_t seqdiff(uint32_t a, uint32_t b) {
 143         return a - b;
 144 }
 145
 146 // Buffer functions
 147 // TODO: convert to ringbuffers to avoid memmove() operations.
 148
 149 // Store data into the buffer
 150 static ssize_t buffer_put_at(struct buffer *buf, size_t offset, const void *data, size_t len) {
 151         if(buf->maxsize <= buf->used)
 152                 return 0;
 153
 154         debug("buffer_put_at %zu %zu %zu\n", buf->used, offset, len);
 155
 156         size_t required = offset + len;
 157         if(required > buf->maxsize) {
 158                 if(offset >= buf->maxsize)
 159                         return 0;
 160                 abort();
 161                 len = buf->maxsize - offset;
 162                 required = buf->maxsize;
 163         }
 164
 165         if(required > buf->size) {
 166                 size_t newsize = buf->size;
 167                 if(!newsize) {
 168                         newsize = required;
 169                 } else {
 170                         do {
 171                                 newsize *= 2;
 172                         } while(newsize < buf->used + len);
 173                 }
 174                 if(newsize > buf->maxsize)
 175                         newsize = buf->maxsize;
 176                 char *newdata = realloc(buf->data, newsize);
 177                 if(!newdata)
 178                         return -1;
 179                 buf->data = newdata;
 180                 buf->size = newsize;
 181         }
 182
 183         memcpy(buf->data + offset, data, len);
 184         if(required > buf->used)
 185                 buf->used = required;
 186         return len;
 187 }
 188
 189 static ssize_t buffer_put(struct buffer *buf, const void *data, size_t len) {
 190         return buffer_put_at(buf, buf->used, data, len);
 191 }
 192
 193 // Get data from the buffer. data can be NULL.
 194 static ssize_t buffer_get(struct buffer *buf, void *data, size_t len) {
 195         if(len > buf->used)
 196                 len = buf->used;
 197         if(data)
 198                 memcpy(data, buf->data, len);
 199         if(len < buf->used)
 200                 memmove(buf->data, buf->data + len, buf->used - len);
 201         buf->used -= len;
 202         return len;
 203 }
 204
 205 // Copy data from the buffer without removing it.
 206 static ssize_t buffer_copy(struct buffer *buf, void *data, size_t offset, size_t len) {
 207         if(offset >= buf->used)
 208                 return 0;
 209         if(offset + len > buf->used)
 210                 len = buf->used - offset;
 211         memcpy(data, buf->data + offset, len);
 212         return len;
 213 }
 214
 215 static bool buffer_init(struct buffer *buf, uint32_t len, uint32_t maxlen) {
 216         memset(buf, 0, sizeof *buf);
 217         if(len) {
 218                 buf->data = malloc(len);
 219                 if(!buf->data)
 220                         return false;
 221         }
 222         buf->size = len;
 223         buf->maxsize = maxlen;
 224         return true;
 225 }
 226
 227 static void buffer_exit(struct buffer *buf) {
 228         free(buf->data);
 229         memset(buf, 0, sizeof *buf);
 230 }
 231
 232 static uint32_t buffer_free(const struct buffer *buf) {
 233         return buf->maxsize - buf->used;
 234 }
 235
 236 // Connections are stored in a sorted list.
 237 // This gives O(log(N)) lookup time, O(N log(N)) insertion time and O(N) deletion time.
 238
 239 static int compare(const void *va, const void *vb) {
 240         assert(va && vb);
 241
 242         const struct utcp_connection *a = *(struct utcp_connection **)va;
 243         const struct utcp_connection *b = *(struct utcp_connection **)vb;
 244
 245         assert(a && b);
 246         assert(a->src && b->src);
 247
 248         int c = (int)a->src - (int)b->src;
 249         if(c)
 250                 return c;
 251         c = (int)a->dst - (int)b->dst;
 252         return c;
 253 }
 254
 255 static struct utcp_connection *find_connection(const struct utcp *utcp, uint16_t src, uint16_t dst) {
 256         if(!utcp->nconnections)
 257                 return NULL;
 258         struct utcp_connection key = {
 259                 .src = src,
 260                 .dst = dst,
 261         }, *keyp = &key;
 262         struct utcp_connection **match = bsearch(&keyp, utcp->connections, utcp->nconnections, sizeof *utcp->connections, compare);
 263         return match ? *match : NULL;
 264 }
 265
 266 static void free_connection(struct utcp_connection *c) {
 267         struct utcp *utcp = c->utcp;
 268         struct utcp_connection **cp = bsearch(&c, utcp->connections, utcp->nconnections, sizeof *utcp->connections, compare);
 269
 270         assert(cp);
 271
 272         int i = cp - utcp->connections;
 273         memmove(cp, cp + 1, (utcp->nconnections - i - 1) * sizeof *cp);
 274         utcp->nconnections--;
 275
 276         buffer_exit(&c->rcvbuf);
 277         buffer_exit(&c->sndbuf);
 278         free(c);
 279 }
 280
 281 static struct utcp_connection *allocate_connection(struct utcp *utcp, uint16_t src, uint16_t dst) {
 282         // Check whether this combination of src and dst is free
 283
 284         if(src) {
 285                 if(find_connection(utcp, src, dst)) {
 286                         errno = EADDRINUSE;
 287                         return NULL;
 288                 }
 289         } else { // If src == 0, generate a random port number with the high bit set
 290                 if(utcp->nconnections >= 32767) {
 291                         errno = ENOMEM;
 292                         return NULL;
 293                 }
 294                 src = rand() | 0x8000;
 295                 while(find_connection(utcp, src, dst))
 296                         src++;
 297         }
 298
 299         // Allocate memory for the new connection
 300
 301         if(utcp->nconnections >= utcp->nallocated) {
 302                 if(!utcp->nallocated)
 303                         utcp->nallocated = 4;
 304                 else
 305                         utcp->nallocated *= 2;
 306                 struct utcp_connection **new_array = realloc(utcp->connections, utcp->nallocated * sizeof *utcp->connections);
 307                 if(!new_array)
 308                         return NULL;
 309                 utcp->connections = new_array;
 310         }
 311
 312         struct utcp_connection *c = calloc(1, sizeof *c);
 313         if(!c)
 314                 return NULL;
 315
 316         if(!buffer_init(&c->sndbuf, DEFAULT_SNDBUFSIZE, DEFAULT_MAXSNDBUFSIZE)) {
 317                 free(c);
 318                 return NULL;
 319         }
 320
 321         if(!buffer_init(&c->rcvbuf, DEFAULT_RCVBUFSIZE, DEFAULT_MAXRCVBUFSIZE)) {
 322                 buffer_exit(&c->sndbuf);
 323                 free(c);
 324                 return NULL;
 325         }
 326
 327         // Fill in the details
 328
 329         c->src = src;
 330         c->dst = dst;
 331 #ifdef UTCP_DEBUG
 332         c->snd.iss = 0;
 333 #else
 334         c->snd.iss = rand();
 335 #endif
 336         c->snd.una = c->snd.iss;
 337         c->snd.nxt = c->snd.iss + 1;
 338         c->rcv.wnd = utcp->mtu;
 339         c->snd.last = c->snd.nxt;
 340         c->snd.cwnd = utcp->mtu;
 341         c->utcp = utcp;
 342
 343         // Add it to the sorted list of connections
 344
 345         utcp->connections[utcp->nconnections++] = c;
 346         qsort(utcp->connections, utcp->nconnections, sizeof *utcp->connections, compare);
 347
 348         return c;
 349 }
 350
 351 // Update RTT variables. See RFC 6298.
 352 static void update_rtt(struct utcp_connection *c, uint32_t rtt) {
 353         if(!rtt) {
 354                 debug("invalid rtt\n");
 355                 return;
 356         }
 357
 358         struct utcp *utcp = c->utcp;
 359
 360         if(!utcp->srtt) {
 361                 utcp->srtt = rtt;
 362                 utcp->rttvar = rtt / 2;
 363                 utcp->rto = rtt + max(2 * rtt, CLOCK_GRANULARITY);
 364         } else {
 365                 utcp->rttvar = (utcp->rttvar * 3 + abs(utcp->srtt - rtt)) / 4;
 366                 utcp->srtt = (utcp->srtt * 7 + rtt) / 8;
 367                 utcp->rto = utcp->srtt + max(utcp->rttvar, CLOCK_GRANULARITY);
 368         }
 369
 370         if(utcp->rto > MAX_RTO)
 371                 utcp->rto = MAX_RTO;
 372
 373         debug("rtt %u srtt %u rttvar %u rto %u\n", rtt, utcp->srtt, utcp->rttvar, utcp->rto);
 374 }
 375
 376 static void start_retransmit_timer(struct utcp_connection *c) {
 377         gettimeofday(&c->rtrx_timeout, NULL);
 378         c->rtrx_timeout.tv_usec += c->utcp->rto;
 379         while(c->rtrx_timeout.tv_usec >= 1000000) {
 380                 c->rtrx_timeout.tv_usec -= 1000000;
 381                 c->rtrx_timeout.tv_sec++;
 382         }
 383         debug("timeout set to %lu.%06lu (%u)\n", c->rtrx_timeout.tv_sec, c->rtrx_timeout.tv_usec, c->utcp->rto);
 384 }
 385
 386 static void stop_retransmit_timer(struct utcp_connection *c) {
 387         timerclear(&c->rtrx_timeout);
 388         debug("timeout cleared\n");
 389 }
 390
 391 struct utcp_connection *utcp_connect(struct utcp *utcp, uint16_t dst, utcp_recv_t recv, void *priv) {
 392         struct utcp_connection *c = allocate_connection(utcp, 0, dst);
 393         if(!c)
 394                 return NULL;
 395
 396         c->recv = recv;
 397         c->priv = priv;
 398
 399         struct hdr hdr;
 400
 401         hdr.src = c->src;
 402         hdr.dst = c->dst;
 403         hdr.seq = c->snd.iss;
 404         hdr.ack = 0;
 405         hdr.wnd = c->rcv.wnd;
 406         hdr.ctl = SYN;
 407         hdr.aux = 0;
 408
 409         set_state(c, SYN_SENT);
 410
 411         print_packet(utcp, "send", &hdr, sizeof hdr);
 412         utcp->send(utcp, &hdr, sizeof hdr);
 413
 414         gettimeofday(&c->conn_timeout, NULL);
 415         c->conn_timeout.tv_sec += utcp->timeout;
 416
 417         return c;
 418 }
 419
 420 void utcp_accept(struct utcp_connection *c, utcp_recv_t recv, void *priv) {
 421         if(c->reapable || c->state != SYN_RECEIVED) {
 422                 debug("Error: accept() called on invalid connection %p in state %s\n", c, strstate[c->state]);
 423                 return;
 424         }
 425
 426         debug("%p accepted, %p %p\n", c, recv, priv);
 427         c->recv = recv;
 428         c->priv = priv;
 429         set_state(c, ESTABLISHED);
 430 }
 431
 432 static void ack(struct utcp_connection *c, bool sendatleastone) {
 433         int32_t left = seqdiff(c->snd.last, c->snd.nxt);
 434         int32_t cwndleft = c->snd.cwnd - seqdiff(c->snd.nxt, c->snd.una);
 435         debug("cwndleft = %d\n", cwndleft);
 436
 437         assert(left >= 0);
 438
 439         if(cwndleft <= 0)
 440                 cwndleft = 0;
 441
 442         if(cwndleft < left)
 443                 left = cwndleft;
 444
 445         if(!left && !sendatleastone)
 446                 return;
 447
 448         struct {
 449                 struct hdr hdr;
 450                 char data[];
 451         } *pkt;
 452
 453         pkt = malloc(sizeof pkt->hdr + c->utcp->mtu);
 454         if(!pkt)
 455                 return;
 456
 457         pkt->hdr.src = c->src;
 458         pkt->hdr.dst = c->dst;
 459         pkt->hdr.ack = c->rcv.nxt;
 460         pkt->hdr.wnd = c->snd.wnd;
 461         pkt->hdr.ctl = ACK;
 462         pkt->hdr.aux = 0;
 463
 464         do {
 465                 uint32_t seglen = left > c->utcp->mtu ? c->utcp->mtu : left;
 466                 pkt->hdr.seq = c->snd.nxt;
 467
 468                 buffer_copy(&c->sndbuf, pkt->data, seqdiff(c->snd.nxt, c->snd.una), seglen);
 469
 470                 c->snd.nxt += seglen;
 471                 left -= seglen;
 472
 473                 if(seglen && fin_wanted(c, c->snd.nxt)) {
 474                         seglen--;
 475                         pkt->hdr.ctl |= FIN;
 476                 }
 477
 478                 if(!c->rtt_start.tv_sec) {
 479                         // Start RTT measurement
 480                         gettimeofday(&c->rtt_start, NULL);
 481                         c->rtt_seq = pkt->hdr.seq + seglen;
 482                         debug("Starting RTT measurement, expecting ack %u\n", c->rtt_seq);
 483                 }
 484
 485                 print_packet(c->utcp, "send", pkt, sizeof pkt->hdr + seglen);
 486                 c->utcp->send(c->utcp, pkt, sizeof pkt->hdr + seglen);
 487         } while(left);
 488
 489         free(pkt);
 490 }
 491
 492 ssize_t utcp_send(struct utcp_connection *c, const void *data, size_t len) {
 493         if(c->reapable) {
 494                 debug("Error: send() called on closed connection %p\n", c);
 495                 errno = EBADF;
 496                 return -1;
 497         }
 498
 499         switch(c->state) {
 500         case CLOSED:
 501         case LISTEN:
 502         case SYN_SENT:
 503         case SYN_RECEIVED:
 504                 debug("Error: send() called on unconnected connection %p\n", c);
 505                 errno = ENOTCONN;
 506                 return -1;
 507         case ESTABLISHED:
 508         case CLOSE_WAIT:
 509                 break;
 510         case FIN_WAIT_1:
 511         case FIN_WAIT_2:
 512         case CLOSING:
 513         case LAST_ACK:
 514         case TIME_WAIT:
 515                 debug("Error: send() called on closing connection %p\n", c);
 516                 errno = EPIPE;
 517                 return -1;
 518         }
 519
 520         // Add data to send buffer
 521
 522         if(!len)
 523                 return 0;
 524
 525         if(!data) {
 526                 errno = EFAULT;
 527                 return -1;
 528         }
 529
 530         len = buffer_put(&c->sndbuf, data, len);
 531         if(len <= 0) {
 532                 errno = EWOULDBLOCK;
 533                 return 0;
 534         }
 535
 536         c->snd.last += len;
 537         ack(c, false);
 538         if(!timerisset(&c->rtrx_timeout))
 539                 start_retransmit_timer(c);
 540         return len;
 541 }
 542
 543 static void swap_ports(struct hdr *hdr) {
 544         uint16_t tmp = hdr->src;
 545         hdr->src = hdr->dst;
 546         hdr->dst = tmp;
 547 }
 548
 549 static void retransmit(struct utcp_connection *c) {
 550         if(c->state == CLOSED || c->snd.nxt == c->snd.una)
 551                 return;
 552
 553         struct utcp *utcp = c->utcp;
 554
 555         struct {
 556                 struct hdr hdr;
 557                 char data[];
 558         } *pkt;
 559
 560         pkt = malloc(sizeof pkt->hdr + c->utcp->mtu);
 561         if(!pkt)
 562                 return;
 563
 564         pkt->hdr.src = c->src;
 565         pkt->hdr.dst = c->dst;
 566         pkt->hdr.wnd = c->rcv.wnd;
 567         pkt->hdr.aux = 0;
 568
 569         switch(c->state) {
 570                 case SYN_SENT:
 571                         // Send our SYN again
 572                         pkt->hdr.seq = c->snd.iss;
 573                         pkt->hdr.ack = 0;
 574                         pkt->hdr.ctl = SYN;
 575                         print_packet(c->utcp, "rtrx", pkt, sizeof pkt->hdr);
 576                         utcp->send(utcp, pkt, sizeof pkt->hdr);
 577                         break;
 578
 579                 case SYN_RECEIVED:
 580                         // Send SYNACK again
 581                         pkt->hdr.seq = c->snd.nxt;
 582                         pkt->hdr.ack = c->rcv.nxt;
 583                         pkt->hdr.ctl = SYN | ACK;
 584                         print_packet(c->utcp, "rtrx", pkt, sizeof pkt->hdr);
 585                         utcp->send(utcp, pkt, sizeof pkt->hdr);
 586                         break;
 587
 588                 case ESTABLISHED:
 589                 case FIN_WAIT_1:
 590                 case CLOSE_WAIT:
 591                 case CLOSING:
 592                 case LAST_ACK:
 593                         // Send unacked data again.
 594                         pkt->hdr.seq = c->snd.una;
 595                         pkt->hdr.ack = c->rcv.nxt;
 596                         pkt->hdr.ctl = ACK;
 597                         uint32_t len = seqdiff(c->snd.last, c->snd.una);
 598                         if(len > utcp->mtu)
 599                                 len = utcp->mtu;
 600                         if(fin_wanted(c, c->snd.una + len)) {
 601                                 len--;
 602                                 pkt->hdr.ctl |= FIN;
 603                         }
 604                         c->snd.nxt = c->snd.una + len;
 605                         c->snd.cwnd = utcp->mtu; // reduce cwnd on retransmit
 606                         buffer_copy(&c->sndbuf, pkt->data, 0, len);
 607                         print_packet(c->utcp, "rtrx", pkt, sizeof pkt->hdr + len);
 608                         utcp->send(utcp, pkt, sizeof pkt->hdr + len);
 609                         break;
 610
 611                 case CLOSED:
 612                 case LISTEN:
 613                 case TIME_WAIT:
 614                 case FIN_WAIT_2:
 615                         // We shouldn't need to retransmit anything in this state.
 616 #ifdef UTCP_DEBUG
 617                         abort();
 618 #endif
 619                         stop_retransmit_timer(c);
 620                         goto cleanup;
 621         }
 622
 623         start_retransmit_timer(c);
 624         utcp->rto *= 2;
 625         if(utcp->rto > MAX_RTO)
 626                 utcp->rto = MAX_RTO;
 627         c->rtt_start.tv_sec = 0; // invalidate RTT timer
 628
 629 cleanup:
 630         free(pkt);
 631 }
 632
 633 // Update receive buffer and SACK entries after consuming data.
 634 static void sack_consume(struct utcp_connection *c, size_t len) {
 635         debug("sack_consume %zu\n", len);
 636         if(len > c->rcvbuf.used)
 637                 abort();
 638
 639         buffer_get(&c->rcvbuf, NULL, len);
 640
 641         for(int i = 0; i < NSACKS && c->sacks[i].len; ) {
 642                 if(len < c->sacks[i].offset) {
 643                         c->sacks[i].offset -= len;
 644                         i++;
 645                 } else if(len < c->sacks[i].offset + c->sacks[i].len) {
 646                         c->sacks[i].offset = 0;
 647                         c->sacks[i].len -= len - c->sacks[i].offset;
 648                         i++;
 649                 } else {
 650                         if(i < NSACKS - 1) {
 651                                 memmove(&c->sacks[i], &c->sacks[i + 1], (NSACKS - 1 - i) * sizeof c->sacks[i]);
 652                                 c->sacks[i + 1].len = 0;
 653                         } else {
 654                                 c->sacks[i].len = 0;
 655                                 break;
 656                         }
 657                 }
 658         }
 659
 660         for(int i = 0; i < NSACKS && c->sacks[i].len; i++)
 661                 debug("SACK[%d] offset %u len %u\n", i, c->sacks[i].offset, c->sacks[i].len);
 662 }
 663
 664 static void handle_out_of_order(struct utcp_connection *c, uint32_t offset, const void *data, size_t len) {
 665         debug("out of order packet, offset %u\n", offset);
 666         // Packet loss or reordering occured. Store the data in the buffer.
 667         ssize_t rxd = buffer_put_at(&c->rcvbuf, offset, data, len);
 668         if(rxd < len)
 669                 abort();
 670
 671         // Make note of where we put it.
 672         for(int i = 0; i < NSACKS; i++) {
 673                 if(!c->sacks[i].len) { // nothing to merge, add new entry
 674                         debug("New SACK entry %d\n", i);
 675                         c->sacks[i].offset = offset;
 676                         c->sacks[i].len = rxd;
 677                         break;
 678                 } else if(offset < c->sacks[i].offset) {
 679                         if(offset + rxd < c->sacks[i].offset) { // insert before
 680                                 if(!c->sacks[NSACKS - 1].len) { // only if room left
 681                                         debug("Insert SACK entry at %d\n", i);
 682                                         memmove(&c->sacks[i + 1], &c->sacks[i], (NSACKS - i - 1) * sizeof c->sacks[i]);
 683                                         c->sacks[i].offset = offset;
 684                                         c->sacks[i].len = rxd;
 685                                 }
 686                                 break;
 687                         } else { // merge
 688                                 debug("Merge with start of SACK entry at %d\n", i);
 689                                 c->sacks[i].offset = offset;
 690                                 break;
 691                         }
 692                 } else if(offset <= c->sacks[i].offset + c->sacks[i].len) {
 693                         if(offset + rxd > c->sacks[i].offset + c->sacks[i].len) { // merge
 694                                 debug("Merge with end of SACK entry at %d\n", i);
 695                                 c->sacks[i].len = offset + rxd - c->sacks[i].offset;
 696                                 // TODO: handle potential merge with next entry
 697                         }
 698                         break;
 699                 }
 700         }
 701
 702         for(int i = 0; i < NSACKS && c->sacks[i].len; i++)
 703                 debug("SACK[%d] offset %u len %u\n", i, c->sacks[i].offset, c->sacks[i].len);
 704 }
 705
 706 static void handle_in_order(struct utcp_connection *c, const void *data, size_t len) {
 707         // Check if we can process out-of-order data now.
 708         if(c->sacks[0].len && len >= c->sacks[0].offset) { // TODO: handle overlap with second SACK
 709                 debug("incoming packet len %zu connected with SACK at %u\n", len, c->sacks[0].offset);
 710                 buffer_put_at(&c->rcvbuf, 0, data, len); // TODO: handle return value
 711                 len = max(len, c->sacks[0].offset + c->sacks[0].len);
 712                 data = c->rcvbuf.data;
 713         }
 714
 715         if(c->recv) {
 716                 ssize_t rxd = c->recv(c, data, len);
 717                 if(rxd != len) {
 718                         // TODO: handle the application not accepting all data.
 719                         abort();
 720                 }
 721         }
 722
 723         if(c->rcvbuf.used)
 724                 sack_consume(c, len);
 725
 726         c->rcv.nxt += len;
 727 }
 728
 729
 730 static void handle_incoming_data(struct utcp_connection *c, uint32_t seq, const void *data, size_t len) {
 731         uint32_t offset = seqdiff(seq, c->rcv.nxt);
 732         if(offset + len > c->rcvbuf.maxsize)
 733                 abort();
 734
 735         if(offset)
 736                 handle_out_of_order(c, offset, data, len);
 737         else
 738                 handle_in_order(c, data, len);
 739 }
 740
 741
 742 ssize_t utcp_recv(struct utcp *utcp, const void *data, size_t len) {
 743         if(!utcp) {
 744                 errno = EFAULT;
 745                 return -1;
 746         }
 747
 748         if(!len)
 749                 return 0;
 750
 751         if(!data) {
 752                 errno = EFAULT;
 753                 return -1;
 754         }
 755
 756         print_packet(utcp, "recv", data, len);
 757
 758         // Drop packets smaller than the header
 759
 760         struct hdr hdr;
 761         if(len < sizeof hdr) {
 762                 errno = EBADMSG;
 763                 return -1;
 764         }
 765
 766         // Make a copy from the potentially unaligned data to a struct hdr
 767
 768         memcpy(&hdr, data, sizeof hdr);
 769         data += sizeof hdr;
 770         len -= sizeof hdr;
 771
 772         // Drop packets with an unknown CTL flag
 773
 774         if(hdr.ctl & ~(SYN | ACK | RST | FIN)) {
 775                 errno = EBADMSG;
 776                 return -1;
 777         }
 778
 779         // Try to match the packet to an existing connection
 780
 781         struct utcp_connection *c = find_connection(utcp, hdr.dst, hdr.src);
 782
 783         // Is it for a new connection?
 784
 785         if(!c) {
 786                 // Ignore RST packets
 787
 788                 if(hdr.ctl & RST)
 789                         return 0;
 790
 791                 // Is it a SYN packet and are we LISTENing?
 792
 793                 if(hdr.ctl & SYN && !(hdr.ctl & ACK) && utcp->accept) {
 794                         // If we don't want to accept it, send a RST back
 795                         if((utcp->pre_accept && !utcp->pre_accept(utcp, hdr.dst))) {
 796                                 len = 1;
 797                                 goto reset;
 798                         }
 799
 800                         // Try to allocate memory, otherwise send a RST back
 801                         c = allocate_connection(utcp, hdr.dst, hdr.src);
 802                         if(!c) {
 803                                 len = 1;
 804                                 goto reset;
 805                         }
 806
 807                         // Return SYN+ACK, go to SYN_RECEIVED state
 808                         c->snd.wnd = hdr.wnd;
 809                         c->rcv.irs = hdr.seq;
 810                         c->rcv.nxt = c->rcv.irs + 1;
 811                         set_state(c, SYN_RECEIVED);
 812
 813                         hdr.dst = c->dst;
 814                         hdr.src = c->src;
 815                         hdr.ack = c->rcv.irs + 1;
 816                         hdr.seq = c->snd.iss;
 817                         hdr.ctl = SYN | ACK;
 818                         print_packet(c->utcp, "send", &hdr, sizeof hdr);
 819                         utcp->send(utcp, &hdr, sizeof hdr);
 820                 } else {
 821                         // No, we don't want your packets, send a RST back
 822                         len = 1;
 823                         goto reset;
 824                 }
 825
 826                 return 0;
 827         }
 828
 829         debug("%p state %s\n", c->utcp, strstate[c->state]);
 830
 831         // In case this is for a CLOSED connection, ignore the packet.
 832         // TODO: make it so incoming packets can never match a CLOSED connection.
 833
 834         if(c->state == CLOSED)
 835                 return 0;
 836
 837         // It is for an existing connection.
 838
 839         uint32_t prevrcvnxt = c->rcv.nxt;
 840
 841         // 1. Drop invalid packets.
 842
 843         // 1a. Drop packets that should not happen in our current state.
 844
 845         switch(c->state) {
 846         case SYN_SENT:
 847         case SYN_RECEIVED:
 848         case ESTABLISHED:
 849         case FIN_WAIT_1:
 850         case FIN_WAIT_2:
 851         case CLOSE_WAIT:
 852         case CLOSING:
 853         case LAST_ACK:
 854         case TIME_WAIT:
 855                 break;
 856         default:
 857 #ifdef UTCP_DEBUG
 858                 abort();
 859 #endif
 860                 break;
 861         }
 862
 863         // 1b. Drop packets with a sequence number not in our receive window.
 864
 865         bool acceptable;
 866
 867         if(c->state == SYN_SENT)
 868                 acceptable = true;
 869         else if(len == 0)
 870                 acceptable = seqdiff(hdr.seq, c->rcv.nxt) >= 0;
 871         else {
 872                 int32_t rcv_offset = seqdiff(hdr.seq, c->rcv.nxt);
 873
 874                 // cut already accepted front overlapping
 875                 if(rcv_offset < 0) {
 876                         acceptable = rcv_offset + len >= 0;
 877                         if(acceptable) {
 878                                 data -= rcv_offset;
 879                                 len += rcv_offset;
 880                         }
 881                 }
 882
 883                 acceptable = seqdiff(hdr.seq, c->rcv.nxt) >= 0 && seqdiff(hdr.seq, c->rcv.nxt) + len <= c->rcvbuf.maxsize;
 884         }
 885
 886         if(!acceptable) {
 887                 debug("Packet not acceptable, %u <= %u + %zu < %u\n", c->rcv.nxt, hdr.seq, len, c->rcv.nxt + c->rcvbuf.maxsize);
 888                 // Ignore unacceptable RST packets.
 889                 if(hdr.ctl & RST)
 890                         return 0;
 891                 // Otherwise, send an ACK back in the hope things improve.
 892                 ack(c, true);
 893                 return 0;
 894         }
 895
 896         c->snd.wnd = hdr.wnd; // TODO: move below
 897
 898         // 1c. Drop packets with an invalid ACK.
 899         // ackno should not roll back, and it should also not be bigger than what we ever could have sent
 900         // (= snd.una + c->sndbuf.used).
 901
 902         if(hdr.ctl & ACK && (seqdiff(hdr.ack, c->snd.last) > 0 || seqdiff(hdr.ack, c->snd.una) < 0)) {
 903                 debug("Packet ack seqno out of range, %u <= %u < %u\n", c->snd.una, hdr.ack, c->snd.una + c->sndbuf.used);
 904                 // Ignore unacceptable RST packets.
 905                 if(hdr.ctl & RST)
 906                         return 0;
 907                 goto reset;
 908         }
 909
 910         // 2. Handle RST packets
 911
 912         if(hdr.ctl & RST) {
 913                 switch(c->state) {
 914                 case SYN_SENT:
 915                         if(!(hdr.ctl & ACK))
 916                                 return 0;
 917                         // The peer has refused our connection.
 918                         set_state(c, CLOSED);
 919                         errno = ECONNREFUSED;
 920                         if(c->recv)
 921                                 c->recv(c, NULL, 0);
 922                         return 0;
 923                 case SYN_RECEIVED:
 924                         if(hdr.ctl & ACK)
 925                                 return 0;
 926                         // We haven't told the application about this connection yet. Silently delete.
 927                         free_connection(c);
 928                         return 0;
 929                 case ESTABLISHED:
 930                 case FIN_WAIT_1:
 931                 case FIN_WAIT_2:
 932                 case CLOSE_WAIT:
 933                         if(hdr.ctl & ACK)
 934                                 return 0;
 935                         // The peer has aborted our connection.
 936                         set_state(c, CLOSED);
 937                         errno = ECONNRESET;
 938                         if(c->recv)
 939                                 c->recv(c, NULL, 0);
 940                         return 0;
 941                 case CLOSING:
 942                 case LAST_ACK:
 943                 case TIME_WAIT:
 944                         if(hdr.ctl & ACK)
 945                                 return 0;
 946                         // As far as the application is concerned, the connection has already been closed.
 947                         // If it has called utcp_close() already, we can immediately free this connection.
 948                         if(c->reapable) {
 949                                 free_connection(c);
 950                                 return 0;
 951                         }
 952                         // Otherwise, immediately move to the CLOSED state.
 953                         set_state(c, CLOSED);
 954                         return 0;
 955                 default:
 956 #ifdef UTCP_DEBUG
 957                         abort();
 958 #endif
 959                         break;
 960                 }
 961         }
 962
 963         // 3. Advance snd.una
 964
 965         uint32_t advanced = seqdiff(hdr.ack, c->snd.una);
 966         prevrcvnxt = c->rcv.nxt;
 967
 968         if(advanced) {
 969                 // RTT measurement
 970                 if(c->rtt_start.tv_sec) {
 971                         if(c->rtt_seq == hdr.ack) {
 972                                 struct timeval now, diff;
 973                                 gettimeofday(&now, NULL);
 974                                 timersub(&now, &c->rtt_start, &diff);
 975                                 update_rtt(c, diff.tv_sec * 1000000 + diff.tv_usec);
 976                                 c->rtt_start.tv_sec = 0;
 977                         } else if(c->rtt_seq < hdr.ack) {
 978                                 debug("Cancelling RTT measurement: %u < %u\n", c->rtt_seq, hdr.ack);
 979                                 c->rtt_start.tv_sec = 0;
 980                         }
 981                 }
 982
 983                 int32_t data_acked = advanced;
 984
 985                 switch(c->state) {
 986                         case SYN_SENT:
 987                         case SYN_RECEIVED:
 988                                 data_acked--;
 989                                 break;
 990                         // TODO: handle FIN as well.
 991                         default:
 992                                 break;
 993                 }
 994
 995                 assert(data_acked >= 0);
 996
 997                 int32_t bufused = seqdiff(c->snd.last, c->snd.una);
 998                 assert(data_acked <= bufused);
 999
1000                 if(data_acked)
1001                         buffer_get(&c->sndbuf, NULL, data_acked);
1002
1003                 // Also advance snd.nxt if possible
1004                 if(seqdiff(c->snd.nxt, hdr.ack) < 0)
1005                         c->snd.nxt = hdr.ack;
1006
1007                 c->snd.una = hdr.ack;
1008
1009                 c->dupack = 0;
1010                 c->snd.cwnd += utcp->mtu;
1011                 if(c->snd.cwnd > c->sndbuf.maxsize)
1012                         c->snd.cwnd = c->sndbuf.maxsize;
1013
1014                 // Check if we have sent a FIN that is now ACKed.
1015                 switch(c->state) {
1016                 case FIN_WAIT_1:
1017                         if(c->snd.una == c->snd.last)
1018                                 set_state(c, FIN_WAIT_2);
1019                         break;
1020                 case CLOSING:
1021                         if(c->snd.una == c->snd.last) {
1022                                 gettimeofday(&c->conn_timeout, NULL);
1023                                 c->conn_timeout.tv_sec += 60;
1024                                 set_state(c, TIME_WAIT);
1025                         }
1026                         break;
1027                 default:
1028                         break;
1029                 }
1030         } else {
1031                 if(!len) {
1032                         c->dupack++;
1033                         if(c->dupack == 3) {
1034                                 debug("Triplicate ACK\n");
1035                                 //TODO: Resend one packet and go to fast recovery mode. See RFC 6582.
1036                                 //We do a very simple variant here; reset the nxt pointer to the last acknowledged packet from the peer.
1037                                 //Reset the congestion window so we wait for ACKs.
1038                                 c->snd.nxt = c->snd.una;
1039                                 c->snd.cwnd = utcp->mtu;
1040                                 start_retransmit_timer(c);
1041                         }
1042                 }
1043         }
1044
1045         // 4. Update timers
1046
1047         if(advanced) {
1048                 timerclear(&c->conn_timeout); // It will be set anew in utcp_timeout() if c->snd.una != c->snd.nxt.
1049                 if(c->snd.una == c->snd.last)
1050                         stop_retransmit_timer(c);
1051                 else
1052                         start_retransmit_timer(c);
1053         }
1054
1055         // 5. Process SYN stuff
1056
1057         if(hdr.ctl & SYN) {
1058                 switch(c->state) {
1059                 case SYN_SENT:
1060                         // This is a SYNACK. It should always have ACKed the SYN.
1061                         if(!advanced)
1062                                 goto reset;
1063                         c->rcv.irs = hdr.seq;
1064                         c->rcv.nxt = hdr.seq;
1065                         set_state(c, ESTABLISHED);
1066                         // TODO: notify application of this somehow.
1067                         break;
1068                 case SYN_RECEIVED:
1069                 case ESTABLISHED:
1070                 case FIN_WAIT_1:
1071                 case FIN_WAIT_2:
1072                 case CLOSE_WAIT:
1073                 case CLOSING:
1074                 case LAST_ACK:
1075                 case TIME_WAIT:
1076                         // Ehm, no. We should never receive a second SYN.
1077                         goto reset;
1078                 default:
1079 #ifdef UTCP_DEBUG
1080                         abort();
1081 #endif
1082                         return 0;
1083                 }
1084
1085                 // SYN counts as one sequence number
1086                 c->rcv.nxt++;
1087         }
1088
1089         // 6. Process new data
1090
1091         if(c->state == SYN_RECEIVED) {
1092                 // This is the ACK after the SYNACK. It should always have ACKed the SYNACK.
1093                 if(!advanced)
1094                         goto reset;
1095
1096                 // Are we still LISTENing?
1097                 if(utcp->accept)
1098                         utcp->accept(c, c->src);
1099
1100                 if(c->state != ESTABLISHED) {
1101                         set_state(c, CLOSED);
1102                         c->reapable = true;
1103                         goto reset;
1104                 }
1105         }
1106
1107         if(len) {
1108                 switch(c->state) {
1109                 case SYN_SENT:
1110                 case SYN_RECEIVED:
1111                         // This should never happen.
1112 #ifdef UTCP_DEBUG
1113                         abort();
1114 #endif
1115                         return 0;
1116                 case ESTABLISHED:
1117                 case FIN_WAIT_1:
1118                 case FIN_WAIT_2:
1119                         break;
1120                 case CLOSE_WAIT:
1121                 case CLOSING:
1122                 case LAST_ACK:
1123                 case TIME_WAIT:
1124                         // Ehm no, We should never receive more data after a FIN.
1125                         goto reset;
1126                 default:
1127 #ifdef UTCP_DEBUG
1128                         abort();
1129 #endif
1130                         return 0;
1131                 }
1132
1133                 handle_incoming_data(c, hdr.seq, data, len);
1134         }
1135
1136         // 7. Process FIN stuff
1137
1138         if((hdr.ctl & FIN) && hdr.seq + len == c->rcv.nxt) {
1139                 switch(c->state) {
1140                 case SYN_SENT:
1141                 case SYN_RECEIVED:
1142                         // This should never happen.
1143 #ifdef UTCP_DEBUG
1144                         abort();
1145 #endif
1146                         break;
1147                 case ESTABLISHED:
1148                         set_state(c, CLOSE_WAIT);
1149                         break;
1150                 case FIN_WAIT_1:
1151                         set_state(c, CLOSING);
1152                         break;
1153                 case FIN_WAIT_2:
1154                         gettimeofday(&c->conn_timeout, NULL);
1155                         c->conn_timeout.tv_sec += 60;
1156                         set_state(c, TIME_WAIT);
1157                         break;
1158                 case CLOSE_WAIT:
1159                 case CLOSING:
1160                 case LAST_ACK:
1161                 case TIME_WAIT:
1162                         // Ehm, no. We should never receive a second FIN.
1163                         goto reset;
1164                 default:
1165 #ifdef UTCP_DEBUG
1166                         abort();
1167 #endif
1168                         break;
1169                 }
1170
1171                 // FIN counts as one sequence number
1172                 c->rcv.nxt++;
1173                 len++;
1174
1175                 // Inform the application that the peer closed the connection.
1176                 if(c->recv) {
1177                         errno = 0;
1178                         c->recv(c, NULL, 0);
1179                 }
1180         }
1181
1182         // Now we send something back if:
1183         // - we advanced rcv.nxt (ie, we got some data that needs to be ACKed)
1184         //   -> sendatleastone = true
1185         // - or we got an ack, so we should maybe send a bit more data
1186         //   -> sendatleastone = false
1187
1188         ack(c, len || prevrcvnxt != c->rcv.nxt);
1189         return 0;
1190
1191 reset:
1192         swap_ports(&hdr);
1193         hdr.wnd = 0;
1194         if(hdr.ctl & ACK) {
1195                 hdr.seq = hdr.ack;
1196                 hdr.ctl = RST;
1197         } else {
1198                 hdr.ack = hdr.seq + len;
1199                 hdr.seq = 0;
1200                 hdr.ctl = RST | ACK;
1201         }
1202         print_packet(utcp, "send", &hdr, sizeof hdr);
1203         utcp->send(utcp, &hdr, sizeof hdr);
1204         return 0;
1205
1206 }
1207
1208 int utcp_shutdown(struct utcp_connection *c, int dir) {
1209         debug("%p shutdown %d at %u\n", c ? c->utcp : NULL, dir, c ? c->snd.last : 0);
1210         if(!c) {
1211                 errno = EFAULT;
1212                 return -1;
1213         }
1214
1215         if(c->reapable) {
1216                 debug("Error: shutdown() called on closed connection %p\n", c);
1217                 errno = EBADF;
1218                 return -1;
1219         }
1220
1221         if(!(dir == UTCP_SHUT_RD || dir == UTCP_SHUT_WR || dir == UTCP_SHUT_RDWR)) {
1222                 errno = EINVAL;
1223                 return -1;
1224         }
1225
1226         // TCP does not have a provision for stopping incoming packets.
1227         // The best we can do is to just ignore them.
1228         if(dir == UTCP_SHUT_RD || dir == UTCP_SHUT_RDWR)
1229                 c->recv = NULL;
1230
1231         // The rest of the code deals with shutting down writes.
1232         if(dir == UTCP_SHUT_RD)
1233                 return 0;
1234
1235         switch(c->state) {
1236         case CLOSED:
1237         case LISTEN:
1238                 errno = ENOTCONN;
1239                 return -1;
1240
1241         case SYN_SENT:
1242                 set_state(c, CLOSED);
1243                 return 0;
1244
1245         case SYN_RECEIVED:
1246         case ESTABLISHED:
1247                 set_state(c, FIN_WAIT_1);
1248                 break;
1249         case FIN_WAIT_1:
1250         case FIN_WAIT_2:
1251                 return 0;
1252         case CLOSE_WAIT:
1253                 set_state(c, CLOSING);
1254                 break;
1255
1256         case CLOSING:
1257         case LAST_ACK:
1258         case TIME_WAIT:
1259                 return 0;
1260         }
1261
1262         c->snd.last++;
1263
1264         ack(c, false);
1265         if(!timerisset(&c->rtrx_timeout))
1266                 start_retransmit_timer(c);
1267         return 0;
1268 }
1269
1270 int utcp_close(struct utcp_connection *c) {
1271         if(utcp_shutdown(c, SHUT_RDWR))
1272                 return -1;
1273         c->recv = NULL;
1274         c->poll = NULL;
1275         c->reapable = true;
1276         return 0;
1277 }
1278
1279 int utcp_abort(struct utcp_connection *c) {
1280         if(!c) {
1281                 errno = EFAULT;
1282                 return -1;
1283         }
1284
1285         if(c->reapable) {
1286                 debug("Error: abort() called on closed connection %p\n", c);
1287                 errno = EBADF;
1288                 return -1;
1289         }
1290
1291         c->recv = NULL;
1292         c->poll = NULL;
1293         c->reapable = true;
1294
1295         switch(c->state) {
1296         case CLOSED:
1297                 return 0;
1298         case LISTEN:
1299         case SYN_SENT:
1300         case CLOSING:
1301         case LAST_ACK:
1302         case TIME_WAIT:
1303                 set_state(c, CLOSED);
1304                 return 0;
1305
1306         case SYN_RECEIVED:
1307         case ESTABLISHED:
1308         case FIN_WAIT_1:
1309         case FIN_WAIT_2:
1310         case CLOSE_WAIT:
1311                 set_state(c, CLOSED);
1312                 break;
1313         }
1314
1315         // Send RST
1316
1317         struct hdr hdr;
1318
1319         hdr.src = c->src;
1320         hdr.dst = c->dst;
1321         hdr.seq = c->snd.nxt;
1322         hdr.ack = 0;
1323         hdr.wnd = 0;
1324         hdr.ctl = RST;
1325
1326         print_packet(c->utcp, "send", &hdr, sizeof hdr);
1327         c->utcp->send(c->utcp, &hdr, sizeof hdr);
1328         return 0;
1329 }
1330
1331 /* Handle timeouts.
1332  * One call to this function will loop through all connections,
1333  * checking if something needs to be resent or not.
1334  * The return value is the time to the next timeout in milliseconds,
1335  * or maybe a negative value if the timeout is infinite.
1336  */
1337 struct timeval utcp_timeout(struct utcp *utcp) {
1338         struct timeval now;
1339         gettimeofday(&now, NULL);
1340         struct timeval next = {now.tv_sec + 3600, now.tv_usec};
1341
1342         for(int i = 0; i < utcp->nconnections; i++) {
1343                 struct utcp_connection *c = utcp->connections[i];
1344                 if(!c)
1345                         continue;
1346
1347                 // delete connections that have been utcp_close()d.
1348                 if(c->state == CLOSED) {
1349                         if(c->reapable) {
1350                                 debug("Reaping %p\n", c);
1351                                 free_connection(c);
1352                                 i--;
1353                         }
1354                         continue;
1355                 }
1356
1357                 if(timerisset(&c->conn_timeout) && timercmp(&c->conn_timeout, &now, <)) {
1358                         errno = ETIMEDOUT;
1359                         c->state = CLOSED;
1360                         if(c->recv)
1361                                 c->recv(c, NULL, 0);
1362                         continue;
1363                 }
1364
1365                 if(timerisset(&c->rtrx_timeout) && timercmp(&c->rtrx_timeout, &now, <)) {
1366                         debug("retransmit()\n");
1367                         retransmit(c);
1368                 }
1369
1370                 if(c->poll && buffer_free(&c->sndbuf) && (c->state == ESTABLISHED || c->state == CLOSE_WAIT))
1371                         c->poll(c, buffer_free(&c->sndbuf));
1372
1373                 if(timerisset(&c->conn_timeout) && timercmp(&c->conn_timeout, &next, <))
1374                         next = c->conn_timeout;
1375
1376                 if(timerisset(&c->rtrx_timeout) && timercmp(&c->rtrx_timeout, &next, <))
1377                         next = c->rtrx_timeout;
1378         }
1379
1380         struct timeval diff;
1381         timersub(&next, &now, &diff);
1382         return diff;
1383 }
1384
1385 bool utcp_is_active(struct utcp *utcp) {
1386         if(!utcp)
1387                 return false;
1388
1389         for(int i = 0; i < utcp->nconnections; i++)
1390                 if(utcp->connections[i]->state != CLOSED && utcp->connections[i]->state != TIME_WAIT)
1391                         return true;
1392
1393         return false;
1394 }
1395
1396 struct utcp *utcp_init(utcp_accept_t accept, utcp_pre_accept_t pre_accept, utcp_send_t send, void *priv) {
1397         if(!send) {
1398                 errno = EFAULT;
1399                 return NULL;
1400         }
1401
1402         struct utcp *utcp = calloc(1, sizeof *utcp);
1403         if(!utcp)
1404                 return NULL;
1405
1406         utcp->accept = accept;
1407         utcp->pre_accept = pre_accept;
1408         utcp->send = send;
1409         utcp->priv = priv;
1410         utcp->mtu = DEFAULT_MTU;
1411         utcp->timeout = DEFAULT_USER_TIMEOUT; // s
1412         utcp->rto = START_RTO; // us
1413
1414         return utcp;
1415 }
1416
1417 void utcp_exit(struct utcp *utcp) {
1418         if(!utcp)
1419                 return;
1420         for(int i = 0; i < utcp->nconnections; i++) {
1421                 if(!utcp->connections[i]->reapable)
1422                         debug("Warning, freeing unclosed connection %p\n", utcp->connections[i]);
1423                 buffer_exit(&utcp->connections[i]->rcvbuf);
1424                 buffer_exit(&utcp->connections[i]->sndbuf);
1425                 free(utcp->connections[i]);
1426         }
1427         free(utcp->connections);
1428         free(utcp);
1429 }
1430
1431 uint16_t utcp_get_mtu(struct utcp *utcp) {
1432         return utcp ? utcp->mtu : 0;
1433 }
1434
1435 void utcp_set_mtu(struct utcp *utcp, uint16_t mtu) {
1436         // TODO: handle overhead of the header
1437         if(utcp)
1438                 utcp->mtu = mtu;
1439 }
1440
1441 int utcp_get_user_timeout(struct utcp *u) {
1442         return u ? u->timeout : 0;
1443 }
1444
1445 void utcp_set_user_timeout(struct utcp *u, int timeout) {
1446         if(u)
1447                 u->timeout = timeout;
1448 }
1449
1450 size_t utcp_get_sndbuf(struct utcp_connection *c) {
1451         return c ? c->sndbuf.maxsize : 0;
1452 }
1453
1454 size_t utcp_get_sndbuf_free(struct utcp_connection *c) {
1455         if(c && (c->state == ESTABLISHED || c->state == CLOSE_WAIT))
1456                 return buffer_free(&c->sndbuf);
1457         else
1458                 return 0;
1459 }
1460
1461 void utcp_set_sndbuf(struct utcp_connection *c, size_t size) {
1462         if(!c)
1463                 return;
1464         c->sndbuf.maxsize = size;
1465         if(c->sndbuf.maxsize != size)
1466                 c->sndbuf.maxsize = -1;
1467 }
1468
1469 size_t utcp_get_rcvbuf(struct utcp_connection *c) {
1470         return c ? c->rcvbuf.maxsize : 0;
1471 }
1472
1473 size_t utcp_get_rcvbuf_free(struct utcp_connection *c) {
1474         if(c && (c->state == ESTABLISHED || c->state == CLOSE_WAIT))
1475                 return buffer_free(&c->rcvbuf);
1476         else
1477                 return 0;
1478 }
1479
1480 void utcp_set_rcvbuf(struct utcp_connection *c, size_t size) {
1481         if(!c)
1482                 return;
1483         c->rcvbuf.maxsize = size;
1484         if(c->rcvbuf.maxsize != size)
1485                 c->rcvbuf.maxsize = -1;
1486 }
1487
1488 bool utcp_get_nodelay(struct utcp_connection *c) {
1489         return c ? c->nodelay : false;
1490 }
1491
1492 void utcp_set_nodelay(struct utcp_connection *c, bool nodelay) {
1493         if(c)
1494                 c->nodelay = nodelay;
1495 }
1496
1497 bool utcp_get_keepalive(struct utcp_connection *c) {
1498         return c ? c->keepalive : false;
1499 }
1500
1501 void utcp_set_keepalive(struct utcp_connection *c, bool keepalive) {
1502         if(c)
1503                 c->keepalive = keepalive;
1504 }
1505
1506 size_t utcp_get_outq(struct utcp_connection *c) {
1507         return c ? seqdiff(c->snd.nxt, c->snd.una) : 0;
1508 }
1509
1510 void utcp_set_recv_cb(struct utcp_connection *c, utcp_recv_t recv) {
1511         if(c)
1512                 c->recv = recv;
1513 }
1514
1515 void utcp_set_poll_cb(struct utcp_connection *c, utcp_poll_t poll) {
1516         if(c)
1517                 c->poll = poll;
1518 }
1519
1520 void utcp_set_accept_cb(struct utcp *utcp, utcp_accept_t accept, utcp_pre_accept_t pre_accept) {
1521         if(utcp) {
1522                 utcp->accept = accept;
1523                 utcp->pre_accept = pre_accept;
1524         }
1525 }