src/ct_server.c

   1 /*
   2  * curvetun - the cipherspace wormhole creator
   3  * Part of the netsniff-ng project
   4  * By Daniel Borkmann <daniel@netsniff-ng.org>
   5  * Copyright 2011 Daniel Borkmann <daniel@netsniff-ng.org>,
   6  * Subject to the GPL.
   7  */
   8
   9 #define _GNU_SOURCE
  10 #include <stdio.h>
  11 #include <stdlib.h>
  12 #include <fcntl.h>
  13 #include <errno.h>
  14 #include <string.h>
  15 #include <unistd.h>
  16 #include <pthread.h>
  17 #include <syslog.h>
  18 #include <signal.h>
  19 #include <netdb.h>
  20 #include <stdint.h>
  21 #include <netinet/in.h>
  22 #include <netinet/tcp.h>
  23 #include <netinet/udp.h>
  24 #include <sys/poll.h>
  25 #include <sys/types.h>
  26 #include <sys/socket.h>
  27 #include <sys/wait.h>
  28 #include <sys/epoll.h>
  29 #include <arpa/inet.h>
  30
  31 #include "die.h"
  32 #include "netdev.h"
  33 #include "write_or_die.h"
  34 #include "psched.h"
  35 #include "xmalloc.h"
  36 #include "curvetun.h"
  37 #include "compiler.h"
  38 #include "deflate.h"
  39 #include "cpusched.h"
  40 #include "trie.h"
  41
  42 struct parent_info {
  43         int efd;
  44         int refd;
  45         int tunfd;
  46         int ipv4;
  47         int udp;
  48 };
  49
  50 struct worker_struct {
  51         pthread_t trid;
  52         int efd[2];
  53         unsigned int cpu;
  54         struct parent_info parent;
  55         int (*handler)(int fd, const struct worker_struct *ws,
  56                        char *buff, size_t len);
  57 };
  58
  59 static struct worker_struct *threadpool = NULL;
  60
  61 extern sig_atomic_t sigint;
  62
  63 static int handler_udp_tun_to_net(int fd, const struct worker_struct *ws,
  64                                   char *buff, size_t len) __pure;
  65 static int handler_udp_net_to_tun(int fd, const struct worker_struct *ws,
  66                                   char *buff, size_t len) __pure;
  67 static int handler_udp(int fd, const struct worker_struct *ws,
  68                        char *buff, size_t len) __pure;
  69 static int handler_tcp_tun_to_net(int fd, const struct worker_struct *ws,
  70                                   char *buff, size_t len) __pure;
  71 static int handler_tcp_net_to_tun(int fd, const struct worker_struct *ws,
  72                                   char *buff, size_t len) __pure;
  73 static int handler_tcp(int fd, const struct worker_struct *ws,
  74                        char *buff, size_t len) __pure;
  75 static void *worker(void *self) __pure;
  76
  77 static int handler_udp_tun_to_net(int fd, const struct worker_struct *ws,
  78                                   char *buff, size_t len)
  79 {
  80         int dfd, state, keep = 1;
  81         char *pbuff;
  82         ssize_t rlen, err, plen;
  83         struct ct_proto *hdr;
  84         struct sockaddr_storage naddr;
  85         socklen_t nlen;
  86
  87         errno = 0;
  88         while ((rlen = read(fd, buff + sizeof(struct ct_proto),
  89                             len - sizeof(struct ct_proto))) > 0) {
  90                 dfd = -1;
  91                 nlen = 0;
  92                 memset(&naddr, 0, sizeof(naddr));
  93
  94                 hdr = (struct ct_proto *) buff;
  95                 hdr->canary = htons(CANARY);
  96                 hdr->flags = 0;
  97
  98                 trie_addr_lookup(buff + sizeof(struct ct_proto),
  99                                  rlen - sizeof(struct ct_proto),
 100                                  ws->parent.ipv4, &dfd, &naddr,
 101                                  (size_t *) &nlen);
 102
 103                 if (dfd < 0 || nlen == 0) {
 104                         syslog(LOG_INFO, "CPU%u: UDP tunnel lookup failed: "
 105                                "unknown destination\n", ws->cpu);
 106                         continue;
 107                 }
 108
 109                 plen = z_deflate(buff + sizeof(struct ct_proto),
 110                                  rlen - sizeof(struct ct_proto), &pbuff);
 111                 if (plen < 0) {
 112                         syslog(LOG_ERR, "CPU%u: UDP tunnel deflate error: %s\n",
 113                                ws->cpu, strerror(errno));
 114                         continue;
 115                 }
 116
 117                 hdr->payload = htons((uint16_t) plen);
 118
 119                 state = 1;
 120                 setsockopt(dfd, IPPROTO_UDP, UDP_CORK, &state, sizeof(state));
 121
 122                 err = sendto(dfd, hdr, sizeof(struct ct_proto), 0,
 123                              (struct sockaddr *) &naddr, nlen);
 124                 if (err < 0)
 125                         syslog(LOG_ERR, "CPU%u: UDP tunnel write error: %s\n",
 126                                ws->cpu, strerror(errno));
 127
 128                 err = sendto(dfd, pbuff, plen, 0, (struct sockaddr *) &naddr,
 129                              nlen);
 130                 if (err < 0)
 131                         syslog(LOG_ERR, "CPU%u: UDP tunnel write error: %s\n",
 132                                ws->cpu, strerror(errno));
 133
 134                 state = 0;
 135                 setsockopt(dfd, IPPROTO_UDP, UDP_CORK, &state, sizeof(state));
 136
 137                 errno = 0;
 138         }
 139
 140         if (rlen < 0 && errno != EAGAIN)
 141                 syslog(LOG_ERR, "CPU%u: UDP tunnel read error: %s\n",
 142                        ws->cpu, strerror(errno));
 143
 144         return keep;
 145 }
 146
 147 static void handler_udp_notify_close(int fd, struct sockaddr_storage *addr,
 148                                      socklen_t len)
 149 {
 150         ssize_t err;
 151         struct ct_proto hdr;
 152
 153         memset(&hdr, 0, sizeof(hdr));
 154         hdr.flags |= PROTO_FLAG_EXIT;
 155         hdr.payload = 0;
 156         hdr.canary = htons(CANARY);
 157
 158         err = sendto(fd, &hdr, sizeof(hdr), 0, (struct sockaddr *) addr, len);
 159 }
 160
 161 static int handler_udp_net_to_tun(int fd, const struct worker_struct *ws,
 162                                   char *buff, size_t len)
 163 {
 164         int keep = 1;
 165         char *pbuff;
 166         ssize_t rlen, err, plen;
 167         struct ct_proto *hdr;
 168         struct sockaddr_storage naddr;
 169         socklen_t nlen;
 170
 171         nlen = sizeof(naddr);
 172         memset(&naddr, 0, sizeof(naddr));
 173
 174         errno = 0;
 175         while ((rlen = recvfrom(fd, buff, len, 0, (struct sockaddr *) &naddr,
 176                                 &nlen)) > 0) {
 177                 hdr = (struct ct_proto *) buff;
 178
 179                 if (unlikely(rlen < sizeof(struct ct_proto)))
 180                         goto close;
 181                 if (unlikely(rlen - sizeof(*hdr) != ntohs(hdr->payload)))
 182                         goto close;
 183                 if (unlikely(ntohs(hdr->canary) != CANARY))
 184                         goto close;
 185                 if (unlikely(ntohs(hdr->payload) == 0))
 186                         goto close;
 187                 if (hdr->flags & PROTO_FLAG_EXIT) {
 188 close:
 189                         trie_addr_remove_addr(&naddr, nlen);
 190                         handler_udp_notify_close(fd, &naddr, nlen);
 191                         nlen = sizeof(naddr);
 192                         memset(&naddr, 0, sizeof(naddr));
 193                         continue;
 194                 }
 195
 196                 plen = z_inflate(buff + sizeof(struct ct_proto),
 197                                  rlen - sizeof(struct ct_proto), &pbuff);
 198                 if (plen < 0) {
 199                         syslog(LOG_ERR, "CPU%u: UDP net inflate error: %s\n",
 200                                ws->cpu, strerror(errno));
 201                         goto next;
 202                 }
 203
 204                 err = trie_addr_maybe_update(pbuff, plen, ws->parent.ipv4,
 205                                              fd, &naddr, nlen);
 206                 if (err) {
 207                         syslog(LOG_INFO, "CPU%u: Malicious packet dropped "
 208                                "from id %d\n", ws->cpu, fd);
 209                         continue;
 210                 }
 211
 212                 err = write(ws->parent.tunfd, pbuff, plen);
 213                 if (err < 0)
 214                         syslog(LOG_ERR, "CPU%u: UDP net write error: %s\n",
 215                                ws->cpu, strerror(errno));
 216
 217 next:
 218                 nlen = sizeof(naddr);
 219                 memset(&naddr, 0, sizeof(naddr));
 220                 errno = 0;
 221         }
 222
 223         if (rlen < 0 && errno != EAGAIN)
 224                 syslog(LOG_ERR, "CPU%u: UDP net read error: %s\n",
 225                        ws->cpu, strerror(errno));
 226
 227         return keep;
 228 }
 229
 230 static int handler_udp(int fd, const struct worker_struct *ws,
 231                        char *buff, size_t len)
 232 {
 233         int ret = 0;
 234         if (fd == ws->parent.tunfd)
 235                 ret = handler_udp_tun_to_net(fd, ws, buff, len);
 236         else
 237                 ret = handler_udp_net_to_tun(fd, ws, buff, len);
 238         return ret;
 239 }
 240
 241 static int handler_tcp_tun_to_net(int fd, const struct worker_struct *ws,
 242                                   char *buff, size_t len)
 243 {
 244         int dfd, state, keep = 1;
 245         char *pbuff;
 246         ssize_t rlen, err, plen;
 247         struct ct_proto *hdr;
 248         socklen_t nlen;
 249
 250         errno = 0;
 251         while ((rlen = read(fd, buff + sizeof(struct ct_proto),
 252                             len - sizeof(struct ct_proto))) > 0) {
 253                 dfd = -1;
 254
 255                 hdr = (struct ct_proto *) buff;
 256                 hdr->canary = htons(CANARY);
 257                 hdr->flags = 0;
 258
 259                 trie_addr_lookup(buff + sizeof(struct ct_proto),
 260                                  rlen - sizeof(struct ct_proto),
 261                                  ws->parent.ipv4, &dfd, NULL,
 262                                  (size_t *) &nlen);
 263
 264                 if (dfd < 0) {
 265                         syslog(LOG_INFO, "CPU%u: TCP tunnel lookup failed: "
 266                                "unknown destination\n", ws->cpu);
 267                         continue;
 268                 }
 269
 270                 plen = z_deflate(buff + sizeof(struct ct_proto),
 271                                  rlen - sizeof(struct ct_proto), &pbuff);
 272                 if (plen < 0) {
 273                         syslog(LOG_ERR, "CPU%u: TCP tunnel deflate error: %s\n",
 274                                ws->cpu, strerror(errno));
 275                         continue;
 276                 }
 277
 278                 hdr->payload = htons((uint16_t) plen);
 279
 280                 state = 1;
 281                 setsockopt(dfd, IPPROTO_TCP, TCP_CORK, &state, sizeof(state));
 282
 283                 err = write_exact(dfd, hdr, sizeof(struct ct_proto), 0);
 284                 if (err < 0)
 285                         syslog(LOG_ERR, "CPU%u: TCP tunnel write error: %s\n",
 286                                ws->cpu, strerror(errno));
 287
 288                 err = write_exact(dfd, pbuff, plen, 0);
 289                 if (err < 0)
 290                         syslog(LOG_ERR, "CPU%u: TCP tunnel write error: %s\n",
 291                                ws->cpu, strerror(errno));
 292
 293                 state = 0;
 294                 setsockopt(dfd, IPPROTO_TCP, TCP_CORK, &state, sizeof(state));
 295
 296                 errno = 0;
 297         }
 298
 299         if (rlen < 0 && errno != EAGAIN)
 300                 syslog(LOG_ERR, "CPU%u: TCP tunnel read error: %s\n",
 301                        ws->cpu, strerror(errno));
 302
 303         return keep;
 304 }
 305
 306 ssize_t handler_tcp_read(int fd, char *buff, size_t len)
 307 {
 308         ssize_t rlen;
 309         struct ct_proto *hdr = (struct ct_proto *) buff;
 310
 311         /* May exit on EAGAIN if 0 Byte read */
 312         rlen = read_exact(fd, buff, sizeof(struct ct_proto), 1);
 313         if (rlen < 0)
 314                 return rlen;
 315
 316         /* May not exit on EAGAIN if 0 Byte read */
 317         rlen = read_exact(fd, buff + sizeof(struct ct_proto),
 318                           ntohs(hdr->payload), 0);
 319         if (rlen < 0)
 320                 return rlen;
 321
 322         return sizeof(struct ct_proto) + rlen;
 323 }
 324
 325 static void handler_tcp_notify_close(int fd)
 326 {
 327         ssize_t err;
 328         struct ct_proto hdr;
 329
 330         memset(&hdr, 0, sizeof(hdr));
 331         hdr.flags |= PROTO_FLAG_EXIT;
 332         hdr.payload = 0;
 333         hdr.canary = htons(CANARY);
 334
 335         err = write(fd, &hdr, sizeof(hdr));
 336 }
 337
 338 static int handler_tcp_net_to_tun(int fd, const struct worker_struct *ws,
 339                                   char *buff, size_t len)
 340 {
 341         int keep = 1, count = 0;
 342         char *pbuff;
 343         ssize_t rlen, err, plen;
 344         struct ct_proto *hdr;
 345         uint64_t fd64;
 346
 347         errno = 0;
 348         while ((rlen = handler_tcp_read(fd, buff, len)) > 0) {
 349                 hdr = (struct ct_proto *) buff;
 350
 351                 if (unlikely(rlen < sizeof(struct ct_proto)))
 352                         goto close;
 353                 if (unlikely(rlen - sizeof(*hdr) != ntohs(hdr->payload)))
 354                         goto close;
 355                 if (unlikely(ntohs(hdr->canary) != CANARY))
 356                         goto close;
 357                 if (unlikely(ntohs(hdr->payload) == 0))
 358                         goto close;
 359                 if (hdr->flags & PROTO_FLAG_EXIT) {
 360 close:
 361                         fd64 = fd;
 362                         trie_addr_remove(fd);
 363                         handler_tcp_notify_close(fd);
 364                         rlen = write(ws->parent.efd, &fd64, sizeof(fd64));
 365                         if (rlen != sizeof(fd64))
 366                                 syslog(LOG_ERR, "CPU%u: TCP event write error: %s\n",
 367                                        ws->cpu, strerror(errno));
 368                         keep = 0;
 369                         return keep;
 370                 }
 371
 372                 plen = z_inflate(buff + sizeof(struct ct_proto),
 373                                  rlen - sizeof(struct ct_proto), &pbuff);
 374                 if (plen < 0) {
 375                         syslog(LOG_ERR, "CPU%u: TCP net inflate error: %s\n",
 376                                ws->cpu, strerror(errno));
 377                         continue;
 378                 }
 379
 380                 err = trie_addr_maybe_update(pbuff, plen, ws->parent.ipv4,
 381                                              fd, NULL, 0);
 382                 if (err) {
 383                         syslog(LOG_INFO, "CPU%u: Malicious packet dropped "
 384                                "from id %d\n", ws->cpu, fd);
 385                         continue;
 386                 }
 387
 388                 err = write(ws->parent.tunfd, pbuff, plen);
 389                 if (err < 0)
 390                         syslog(LOG_ERR, "CPU%u: TCP net write error: %s\n",
 391                                ws->cpu, strerror(errno));
 392
 393                 count++;
 394                 if (count == 10) {
 395                         err = write_exact(ws->efd[1], &fd, sizeof(fd), 1);
 396                         if (err != sizeof(fd))
 397                                 syslog(LOG_ERR, "CPU%u: TCP net put fd back in "
 398                                        "pipe error: %s\n", ws->cpu, strerror(errno));
 399                         return keep;
 400                 }
 401
 402                 errno = 0;
 403         }
 404
 405         if (rlen < 0 && errno != EAGAIN && errno != EBADF)
 406                 syslog(LOG_ERR, "CPU%u: TCP net read error: %s\n",
 407                        ws->cpu, strerror(errno));
 408
 409         return keep;
 410 }
 411
 412 static int handler_tcp(int fd, const struct worker_struct *ws,
 413                        char *buff, size_t len)
 414 {
 415         int ret = 0;
 416         if (fd == ws->parent.tunfd)
 417                 ret = handler_tcp_tun_to_net(fd, ws, buff, len);
 418         else
 419                 ret = handler_tcp_net_to_tun(fd, ws, buff, len);
 420         return ret;
 421 }
 422
 423 static void *worker(void *self)
 424 {
 425         int fd, old_state;
 426         ssize_t ret;
 427         size_t blen = TUNBUFF_SIZ; //FIXME
 428         const struct worker_struct *ws = self;
 429         struct pollfd fds;
 430         char *buff;
 431
 432         fds.fd = ws->efd[0];
 433         fds.events = POLLIN;
 434
 435         buff = xmalloc(blen);
 436         syslog(LOG_INFO, "curvetun thread on CPU%u up!\n", ws->cpu);
 437         pthread_cleanup_push(xfree, buff);
 438
 439         while (likely(!sigint)) {
 440                 poll(&fds, 1, -1);
 441                 if ((fds.revents & POLLIN) != POLLIN)
 442                         continue;
 443                 pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &old_state);
 444                 while ((ret = read_exact(ws->efd[0], &fd, sizeof(fd), 1)) > 0) {
 445                         if (ret != sizeof(fd)) {
 446                                 syslog(LOG_ERR, "CPU%u: Thread could not read "
 447                                        "event descriptor!\n", ws->cpu);
 448                                 sched_yield();
 449                                 continue;
 450                         }
 451
 452                         ret = ws->handler(fd, ws, buff, blen);
 453                         if (ret) {
 454                                 ret = write_exact(ws->parent.refd, &fd, sizeof(fd), 1);
 455                                 if (ret != sizeof(fd))
 456                                         syslog(LOG_ERR, "CPU%u: Retriggering failed: "
 457                                                "%s\n", ws->cpu, strerror(errno));
 458                         }
 459                 }
 460                 pthread_setcancelstate(old_state, NULL);
 461         }
 462
 463         syslog(LOG_INFO, "curvetun thread on CPU%u down!\n", ws->cpu);
 464         pthread_cleanup_pop(1);
 465         pthread_exit((void *) ((long) ws->cpu));
 466 }
 467
 468 static void thread_spawn_or_panic(unsigned int cpus, int efd, int refd,
 469                                   int tunfd, int ipv4, int udp)
 470 {
 471         int i, ret;
 472         cpu_set_t cpuset;
 473         unsigned int threads;
 474
 475         threads = cpus * THREADS_PER_CPU;
 476         for (i = 0; i < threads; ++i) {
 477                 CPU_ZERO(&cpuset);
 478                 threadpool[i].cpu = i % cpus;
 479                 CPU_SET(threadpool[i].cpu, &cpuset);
 480
 481                 ret = pipe2(threadpool[i].efd, O_NONBLOCK);
 482                 if (ret < 0)
 483                         panic("Cannot create event socket!\n");
 484
 485                 threadpool[i].parent.efd = efd;
 486                 threadpool[i].parent.refd = refd;
 487                 threadpool[i].parent.tunfd = tunfd;
 488                 threadpool[i].parent.ipv4 = ipv4;
 489                 threadpool[i].parent.udp = udp;
 490                 threadpool[i].handler = udp ? handler_udp : handler_tcp;
 491
 492                 ret = pthread_create(&threadpool[i].trid, NULL,
 493                                      worker, &threadpool[i]);
 494                 if (ret < 0)
 495                         panic("Thread creation failed!\n");
 496
 497                 ret = pthread_setaffinity_np(threadpool[i].trid,
 498                                              sizeof(cpu_set_t), &cpuset);
 499                 if (ret < 0)
 500                         panic("Thread CPU migration failed!\n");
 501
 502                 pthread_detach(threadpool[i].trid);
 503         }
 504
 505         sleep(1);
 506 }
 507
 508 static void thread_finish(unsigned int cpus)
 509 {
 510         int i, ret;
 511         unsigned int threads;
 512
 513         threads = cpus * THREADS_PER_CPU;
 514         for (i = 0; i < threads; ++i) {
 515                 ret = pthread_join(threadpool[i].trid, NULL);
 516                 if (ret < 0)
 517                         continue;
 518                 close(threadpool[i].efd[0]);
 519                 close(threadpool[i].efd[1]);
 520         }
 521 }
 522
 523 int server_main(int port, int udp, int lnum)
 524 {
 525         int lfd = -1, kdpfd, nfds, nfd, curfds, efd[2], refd[2], tunfd;
 526         int ipv4 = 0, i;
 527         unsigned int cpus = 0, threads;
 528         ssize_t ret;
 529         struct epoll_event ev, *events;
 530         struct addrinfo hints, *ahead, *ai;
 531
 532         openlog("curvetun", LOG_PID | LOG_CONS | LOG_NDELAY, LOG_DAEMON);
 533         syslog(LOG_INFO, "curvetun server booting!\n");
 534
 535         ret = z_alloc_or_maybe_die(Z_DEFAULT_COMPRESSION);
 536         if (ret < 0)
 537                 panic("Cannot init zLib!\n");
 538
 539         memset(&hints, 0, sizeof(hints));
 540         hints.ai_family = PF_UNSPEC;
 541         hints.ai_socktype = udp ? SOCK_DGRAM : SOCK_STREAM;
 542         hints.ai_protocol = udp ? IPPROTO_UDP : IPPROTO_TCP;
 543         hints.ai_flags = AI_PASSIVE;
 544
 545         ret = getaddrinfo(NULL, "6666", &hints, &ahead);
 546         if (ret < 0)
 547                 panic("Cannot get address info!\n");
 548
 549         for (ai = ahead; ai != NULL && lfd < 0; ai = ai->ai_next) {
 550                 lfd = socket(ai->ai_family, ai->ai_socktype, ai->ai_protocol);
 551                 if (lfd < 0)
 552                         continue;
 553                 if (ai->ai_family == AF_INET6) {
 554                         int one = 1;
 555 #ifdef IPV6_V6ONLY
 556                         ret = setsockopt(lfd, IPPROTO_IPV6, IPV6_V6ONLY,
 557                                          &one, sizeof(one));
 558                         if (ret < 0) {
 559                                 close(lfd);
 560                                 lfd = -1;
 561                                 continue;
 562                         }
 563 #else
 564                         close(lfd);
 565                         lfd = -1;
 566                         continue;
 567 #endif /* IPV6_V6ONLY */
 568                 }
 569                 set_reuseaddr(lfd);
 570                 ret = bind(lfd, ai->ai_addr, ai->ai_addrlen);
 571                 if (ret < 0) {
 572                         close(lfd);
 573                         lfd = -1;
 574                         continue;
 575                 }
 576                 if (!udp) {
 577                         ret = listen(lfd, 5);
 578                         if (ret < 0) {
 579                                 close(lfd);
 580                                 lfd = -1;
 581                                 continue;
 582                         }
 583                 }
 584                 ipv4 = (ai->ai_family == AF_INET6 ? 0 :
 585                         (ai->ai_family == AF_INET ? 1 : -1));
 586                 syslog(LOG_INFO, "curvetun on IPv%d via %s!\n",
 587                        ipv4 ? 4 : 6, udp ? "UDP" : "TCP");
 588         }
 589
 590         freeaddrinfo(ahead);
 591         if (lfd < 0 || ipv4 < 0)
 592                 panic("Cannot create socket!\n");
 593
 594         tunfd = tun_open_or_die(DEVNAME_SERVER);
 595
 596         ret = pipe2(efd, O_NONBLOCK);
 597         if (ret < 0)
 598                 panic("Cannot create parent event fd!\n");
 599
 600         ret = pipe2(refd, O_NONBLOCK);
 601         if (ret < 0)
 602                 panic("Cannot create parent (r)event fd!\n");
 603
 604         set_nonblocking(lfd);
 605
 606         events = xzmalloc(MAX_EPOLL_SIZE * sizeof(*events));
 607         for (i = 0; i < MAX_EPOLL_SIZE; ++i)
 608                 events[i].data.fd = -1;
 609
 610         kdpfd = epoll_create(MAX_EPOLL_SIZE);
 611         if (kdpfd < 0)
 612                 panic("Cannot create socket!\n");
 613
 614         memset(&ev, 0, sizeof(ev));
 615         ev.events = udp ? EPOLLIN | EPOLLET | EPOLLONESHOT : EPOLLIN;
 616         ev.data.fd = lfd;
 617         ret = epoll_ctl(kdpfd, EPOLL_CTL_ADD, lfd, &ev);
 618         if (ret < 0)
 619                 panic("Cannot add socket for epoll!\n");
 620
 621         memset(&ev, 0, sizeof(ev));
 622         ev.events = EPOLLIN;
 623         ev.data.fd = efd[0];
 624         ret = epoll_ctl(kdpfd, EPOLL_CTL_ADD, efd[0], &ev);
 625         if (ret < 0)
 626                 panic("Cannot add socket for events!\n");
 627
 628         memset(&ev, 0, sizeof(ev));
 629         ev.events = EPOLLIN;
 630         ev.data.fd = refd[0];
 631         ret = epoll_ctl(kdpfd, EPOLL_CTL_ADD, refd[0], &ev);
 632         if (ret < 0)
 633                 panic("Cannot add socket for (r)events!\n");
 634
 635         memset(&ev, 0, sizeof(ev));
 636         ev.events = EPOLLIN | EPOLLET | EPOLLONESHOT;
 637         ev.data.fd = tunfd;
 638         ret = epoll_ctl(kdpfd, EPOLL_CTL_ADD, tunfd, &ev);
 639         if (ret < 0)
 640                 panic("Cannot add socket for tundev!\n");
 641
 642         curfds = 4;
 643
 644         trie_init();
 645
 646         cpus = get_number_cpus_online();
 647         threads = cpus * THREADS_PER_CPU;
 648         if (!((threads != 0) && ((threads & (threads - 1)) == 0)))
 649                 panic("thread number not power of two!\n");
 650         threadpool = xzmalloc(sizeof(*threadpool) * threads);
 651         thread_spawn_or_panic(cpus, efd[1], refd[1], tunfd, ipv4, udp);
 652
 653         init_cpusched(threads, MAX_EPOLL_SIZE);
 654         register_socket(tunfd);
 655         register_socket(lfd);
 656
 657         syslog(LOG_INFO, "tunnel id: %d, listener id: %d\n", tunfd, lfd);
 658         syslog(LOG_INFO, "curvetun up and running!\n");
 659
 660         while (likely(!sigint)) {
 661                 nfds = epoll_wait(kdpfd, events, curfds, -1);
 662                 if (nfds < 0) {
 663                         syslog(LOG_ERR, "epoll_wait error: %s\n",
 664                                strerror(errno));
 665                         break;
 666                 }
 667
 668                 for (i = 0; i < nfds; ++i) {
 669                         if (unlikely(events[i].data.fd < 0))
 670                                 continue;
 671                         if (events[i].data.fd == lfd && !udp) {
 672                                 int one, ncpu;
 673                                 char hbuff[256], sbuff[256];
 674                                 struct sockaddr_storage taddr;
 675                                 socklen_t tlen;
 676
 677                                 tlen = sizeof(taddr);
 678                                 nfd = accept(lfd, (struct sockaddr *) &taddr,
 679                                              &tlen);
 680                                 if (nfd < 0) {
 681                                         syslog(LOG_ERR, "accept error: %s\n",
 682                                                strerror(errno));
 683                                         continue;
 684                                 }
 685
 686                                 if (curfds + 1 > MAX_EPOLL_SIZE) {
 687                                         close(nfd);
 688                                         continue;
 689                                 }
 690
 691                                 curfds++;
 692                                 ncpu = register_socket(nfd);
 693
 694                                 memset(hbuff, 0, sizeof(hbuff));
 695                                 memset(sbuff, 0, sizeof(sbuff));
 696
 697                                 getnameinfo((struct sockaddr *) &taddr, tlen,
 698                                             hbuff, sizeof(hbuff),
 699                                             sbuff, sizeof(sbuff),
 700                                             NI_NUMERICHOST | NI_NUMERICSERV);
 701
 702                                 syslog(LOG_INFO, "New connection from %s:%s "
 703                                        "with id %d on CPU%d, %d active!\n",
 704                                        hbuff, sbuff, nfd, ncpu, curfds);
 705
 706                                 set_nonblocking(nfd);
 707
 708                                 one = 1;
 709                                 setsockopt(nfd, SOL_SOCKET, SO_KEEPALIVE,
 710                                            &one, sizeof(one));
 711                                 one = 1;
 712                                 setsockopt(nfd, IPPROTO_TCP, TCP_NODELAY,
 713                                            &one, sizeof(one));
 714
 715                                 memset(&ev, 0, sizeof(ev));
 716                                 ev.events = EPOLLIN | EPOLLET | EPOLLONESHOT;
 717                                 ev.data.fd = nfd;
 718                                 ret = epoll_ctl(kdpfd, EPOLL_CTL_ADD, nfd, &ev);
 719                                 if (ret < 0) {
 720                                         syslog(LOG_ERR, "Epoll ctl add error"
 721                                                "on id %d: %s\n", nfd,
 722                                                strerror(errno));
 723                                         close(nfd);
 724                                         curfds--;
 725                                         continue;
 726                                 }
 727                         } else if (events[i].data.fd == refd[0]) {
 728                                 int fd_one;
 729
 730                                 ret = read_exact(refd[0], &fd_one, sizeof(fd_one), 1);
 731                                 if (ret != sizeof(fd_one) || fd_one <= 0)
 732                                         continue;
 733
 734                                 memset(&ev, 0, sizeof(ev));
 735                                 ev.events = EPOLLIN | EPOLLET | EPOLLONESHOT;
 736                                 ev.data.fd = fd_one;
 737                                 ret = epoll_ctl(kdpfd, EPOLL_CTL_MOD, fd_one, &ev);
 738                                 if (ret < 0) {
 739                                         syslog(LOG_ERR, "Epoll ctl mod "
 740                                                "error on id %d: %s\n",
 741                                                fd_one, strerror(errno));
 742                                         close(fd_one);
 743                                         continue;
 744                                 }
 745                         } else if (events[i].data.fd == efd[0]) {
 746                                 int fd_del, test;
 747
 748                                 ret = read_exact(efd[0], &fd_del, sizeof(fd_del), 1);
 749                                 if (ret != sizeof(fd_del) || fd_del <= 0)
 750                                         continue;
 751
 752                                 ret = read(fd_del, &test, sizeof(test));
 753                                 if (ret < 0 && errno == EBADF)
 754                                         continue;
 755
 756                                 ret = epoll_ctl(kdpfd, EPOLL_CTL_DEL, fd_del, &ev);
 757                                 if (ret < 0) {
 758                                         syslog(LOG_ERR, "Epoll ctl del "
 759                                                "error on id %d: %s\n",
 760                                                fd_del, strerror(errno));
 761                                         close(fd_del);
 762                                         continue;
 763                                 }
 764                                 close(fd_del);
 765                                 curfds--;
 766                                 unregister_socket(fd_del);
 767
 768                                 syslog(LOG_INFO, "Closed connection with "
 769                                        "id %d, %d active!\n",
 770                                        fd_del, curfds);
 771                         } else {
 772                                 int cpu, fd_work = events[i].data.fd;
 773                                 cpu = socket_to_cpu(fd_work);
 774
 775                                 ret = write_exact(threadpool[cpu].efd[1],
 776                                                   &fd_work, sizeof(fd_work), 1);
 777                                 if (ret != sizeof(fd_work))
 778                                         syslog(LOG_ERR, "Write error on event "
 779                                                "dispatch: %s\n", strerror(errno));
 780                         }
 781                 }
 782         }
 783
 784         syslog(LOG_INFO, "curvetun prepare shut down!\n");
 785
 786         close(lfd);
 787         close(efd[0]);
 788         close(efd[1]);
 789         close(refd[0]);
 790         close(refd[1]);
 791         close(tunfd);
 792
 793         thread_finish(cpus);
 794         xfree(threadpool);
 795         xfree(events);
 796
 797         unregister_socket(lfd);
 798         unregister_socket(tunfd);
 799         destroy_cpusched();
 800         trie_cleanup();
 801         z_free();
 802
 803         syslog(LOG_INFO, "curvetun shut down!\n");
 804         closelog();
 805
 806         return 0;
 807 }
 808