src/fdevent.c

   1 #include "first.h"
   2
   3 #include "fdevent_impl.h"
   4 #include "fdevent.h"
   5 #include "base.h"
   6 #include "buffer.h"
   7 #include "log.h"
   8
   9 #include <sys/types.h>
  10 #include <sys/wait.h>
  11 #include "sys-socket.h"
  12
  13 #include <unistd.h>
  14 #include <stdlib.h>
  15 #include <string.h>
  16 #include <errno.h>
  17 #include <fcntl.h>
  18 #include <time.h>
  19
  20 #ifdef SOCK_CLOEXEC
  21 static int use_sock_cloexec;
  22 #endif
  23 #ifdef SOCK_NONBLOCK
  24 static int use_sock_nonblock;
  25 #endif
  26
  27 int fdevent_config(server *srv) {
  28         static const struct ev_map { fdevent_handler_t et; const char *name; } event_handlers[] =
  29         {
  30                 /* - epoll is most reliable
  31                  * - select works everywhere
  32                  */
  33 #ifdef FDEVENT_USE_LINUX_EPOLL
  34                 { FDEVENT_HANDLER_LINUX_SYSEPOLL, "linux-sysepoll" },
  35                 { FDEVENT_HANDLER_LINUX_SYSEPOLL, "epoll" },
  36 #endif
  37 #ifdef FDEVENT_USE_SOLARIS_PORT
  38                 { FDEVENT_HANDLER_SOLARIS_PORT,   "solaris-eventports" },
  39 #endif
  40 #ifdef FDEVENT_USE_SOLARIS_DEVPOLL
  41                 { FDEVENT_HANDLER_SOLARIS_DEVPOLL,"solaris-devpoll" },
  42 #endif
  43 #ifdef FDEVENT_USE_FREEBSD_KQUEUE
  44                 { FDEVENT_HANDLER_FREEBSD_KQUEUE, "freebsd-kqueue" },
  45                 { FDEVENT_HANDLER_FREEBSD_KQUEUE, "kqueue" },
  46 #endif
  47 #ifdef FDEVENT_USE_POLL
  48                 { FDEVENT_HANDLER_POLL,           "poll" },
  49 #endif
  50 #ifdef FDEVENT_USE_SELECT
  51                 { FDEVENT_HANDLER_SELECT,         "select" },
  52 #endif
  53 #ifdef FDEVENT_USE_LIBEV
  54                 { FDEVENT_HANDLER_LIBEV,          "libev" },
  55 #endif
  56                 { FDEVENT_HANDLER_UNSET,          NULL }
  57         };
  58
  59         if (buffer_string_is_empty(srv->srvconf.event_handler)) {
  60                 /* choose a good default
  61                  *
  62                  * the event_handler list is sorted by 'goodness'
  63                  * taking the first available should be the best solution
  64                  */
  65                 srv->event_handler = event_handlers[0].et;
  66
  67                 if (FDEVENT_HANDLER_UNSET == srv->event_handler) {
  68                         log_error_write(srv, __FILE__, __LINE__, "s",
  69                                         "sorry, there is no event handler for this system");
  70
  71                         return -1;
  72                 }
  73
  74                 buffer_copy_string(srv->srvconf.event_handler, event_handlers[0].name);
  75         } else {
  76                 /*
  77                  * User override
  78                  */
  79
  80                 for (size_t i = 0; event_handlers[i].name; i++) {
  81                         if (0 == strcmp(event_handlers[i].name, srv->srvconf.event_handler->ptr)) {
  82                                 srv->event_handler = event_handlers[i].et;
  83                                 break;
  84                         }
  85                 }
  86
  87                 if (FDEVENT_HANDLER_UNSET == srv->event_handler) {
  88                         log_error_write(srv, __FILE__, __LINE__, "sb",
  89                                         "the selected event-handler in unknown or not supported:",
  90                                         srv->srvconf.event_handler );
  91
  92                         return -1;
  93                 }
  94         }
  95
  96       #ifdef FDEVENT_USE_SELECT
  97         if (srv->event_handler == FDEVENT_HANDLER_SELECT) {
  98                 /* select limits itself
  99                  *
 100                  * as it is a hard limit and will lead to a segfault we add some safety
 101                  * */
 102                 srv->max_fds = FD_SETSIZE - 200;
 103         }
 104         else
 105       #endif
 106         {
 107                 srv->max_fds = 4096;
 108         }
 109
 110         return 0;
 111 }
 112
 113 const char * fdevent_show_event_handlers(void) {
 114     return
 115       "\nEvent Handlers:\n\n"
 116 #ifdef FDEVENT_USE_SELECT
 117       "\t+ select (generic)\n"
 118 #else
 119       "\t- select (generic)\n"
 120 #endif
 121 #ifdef FDEVENT_USE_POLL
 122       "\t+ poll (Unix)\n"
 123 #else
 124       "\t- poll (Unix)\n"
 125 #endif
 126 #ifdef FDEVENT_USE_LINUX_EPOLL
 127       "\t+ epoll (Linux)\n"
 128 #else
 129       "\t- epoll (Linux)\n"
 130 #endif
 131 #ifdef FDEVENT_USE_SOLARIS_DEVPOLL
 132       "\t+ /dev/poll (Solaris)\n"
 133 #else
 134       "\t- /dev/poll (Solaris)\n"
 135 #endif
 136 #ifdef FDEVENT_USE_SOLARIS_PORT
 137       "\t+ eventports (Solaris)\n"
 138 #else
 139       "\t- eventports (Solaris)\n"
 140 #endif
 141 #ifdef FDEVENT_USE_FREEBSD_KQUEUE
 142       "\t+ kqueue (FreeBSD)\n"
 143 #else
 144       "\t- kqueue (FreeBSD)\n"
 145 #endif
 146 #ifdef FDEVENT_USE_LIBEV
 147       "\t+ libev (generic)\n"
 148 #else
 149       "\t- libev (generic)\n"
 150 #endif
 151       ;
 152 }
 153
 154 fdevents *fdevent_init(server *srv) {
 155         fdevents *ev;
 156         int type = srv->event_handler;
 157         size_t maxfds;
 158
 159       #ifdef SOCK_CLOEXEC
 160         /* Test if SOCK_CLOEXEC is supported by kernel.
 161          * Linux kernels < 2.6.27 might return EINVAL if SOCK_CLOEXEC used
 162          * https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=529929
 163          * http://www.linksysinfo.org/index.php?threads/lighttpd-no-longer-starts-toastman-1-28-0510-7.73132/
 164          * Test if SOCK_NONBLOCK is ignored by kernel on sockets.
 165          * (reported on Android running a custom ROM)
 166          * https://redmine.lighttpd.net/issues/2883
 167          */
 168        #ifdef SOCK_NONBLOCK
 169         int fd = socket(AF_INET, SOCK_STREAM | SOCK_CLOEXEC | SOCK_NONBLOCK, 0);
 170        #else
 171         int fd = socket(AF_INET, SOCK_STREAM | SOCK_CLOEXEC, 0);
 172        #endif
 173         if (fd >= 0) {
 174                 int flags = fcntl(fd, F_GETFL, 0);
 175               #ifdef SOCK_NONBLOCK
 176                 use_sock_nonblock = (-1 != flags && (flags & O_NONBLOCK));
 177               #endif
 178                 use_sock_cloexec = 1;
 179                 close(fd);
 180         }
 181       #endif
 182
 183       #ifdef FDEVENT_USE_SELECT
 184         if (type == FDEVENT_HANDLER_SELECT) {
 185                 if (srv->max_fds > (int)FD_SETSIZE - 200) {
 186                         srv->max_fds = (int)FD_SETSIZE - 200;
 187                 }
 188         }
 189       #endif
 190         maxfds = srv->max_fds + 1; /*(+1 for event-handler fd)*/
 191
 192         ev = calloc(1, sizeof(*ev));
 193         force_assert(NULL != ev);
 194         ev->srv = srv;
 195         ev->fdarray = calloc(maxfds, sizeof(*ev->fdarray));
 196         if (NULL == ev->fdarray) {
 197                 log_error_write(srv, __FILE__, __LINE__, "SDS",
 198                                 "server.max-fds too large? (", maxfds-1, ")");
 199                 free(ev);
 200                 return NULL;
 201         }
 202         ev->maxfds = maxfds;
 203
 204         switch(type) {
 205         #ifdef FDEVENT_USE_POLL
 206         case FDEVENT_HANDLER_POLL:
 207                 if (0 == fdevent_poll_init(ev)) return ev;
 208                 break;
 209         #endif
 210         #ifdef FDEVENT_USE_SELECT
 211         case FDEVENT_HANDLER_SELECT:
 212                 if (0 == fdevent_select_init(ev)) return ev;
 213                 break;
 214         #endif
 215         #ifdef FDEVENT_USE_LINUX_EPOLL
 216         case FDEVENT_HANDLER_LINUX_SYSEPOLL:
 217                 if (0 == fdevent_linux_sysepoll_init(ev)) return ev;
 218                 break;
 219         #endif
 220         #ifdef FDEVENT_USE_SOLARIS_DEVPOLL
 221         case FDEVENT_HANDLER_SOLARIS_DEVPOLL:
 222                 if (0 == fdevent_solaris_devpoll_init(ev)) return ev;
 223                 break;
 224         #endif
 225         #ifdef FDEVENT_USE_SOLARIS_PORT
 226         case FDEVENT_HANDLER_SOLARIS_PORT:
 227                 if (0 == fdevent_solaris_port_init(ev)) return ev;
 228                 break;
 229         #endif
 230         #ifdef FDEVENT_USE_FREEBSD_KQUEUE
 231         case FDEVENT_HANDLER_FREEBSD_KQUEUE:
 232                 if (0 == fdevent_freebsd_kqueue_init(ev)) return ev;
 233                 break;
 234         #endif
 235         #ifdef FDEVENT_USE_LIBEV
 236         case FDEVENT_HANDLER_LIBEV:
 237                 if (0 == fdevent_libev_init(ev)) return ev;
 238                 break;
 239         #endif
 240         case FDEVENT_HANDLER_UNSET:
 241         default:
 242                 break;
 243         }
 244
 245         free(ev->fdarray);
 246         free(ev);
 247
 248         log_error_write(srv, __FILE__, __LINE__, "sBS",
 249                 "event-handler failed:", srv->srvconf.event_handler, "; try to set server.event-handler = \"poll\" or \"select\"");
 250         return NULL;
 251 }
 252
 253 void fdevent_free(fdevents *ev) {
 254         size_t i;
 255         if (!ev) return;
 256
 257         if (ev->free) ev->free(ev);
 258
 259         for (i = 0; i < ev->maxfds; i++) {
 260                 /* (fdevent_sched_run() should already have been run,
 261                  *  but take reasonable precautions anyway) */
 262                 if (ev->fdarray[i])
 263                         free((fdnode *)((uintptr_t)ev->fdarray[i] & ~0x3));
 264         }
 265
 266         free(ev->fdarray);
 267         free(ev);
 268 }
 269
 270 int fdevent_reset(fdevents *ev) {
 271         int rc = (NULL != ev->reset) ? ev->reset(ev) : 0;
 272         if (-1 == rc) {
 273                 log_error_write(ev->srv, __FILE__, __LINE__, "sBS",
 274                         "event-handler failed:", ev->srv->srvconf.event_handler, "; try to set server.event-handler = \"poll\" or \"select\"");
 275         }
 276         return rc;
 277 }
 278
 279 static fdnode *fdnode_init(void) {
 280         return calloc(1, sizeof(fdnode));
 281 }
 282
 283 static void fdnode_free(fdnode *fdn) {
 284         free(fdn);
 285 }
 286
 287 fdnode * fdevent_register(fdevents *ev, int fd, fdevent_handler handler, void *ctx) {
 288         fdnode *fdn  = ev->fdarray[fd] = fdnode_init();
 289         force_assert(NULL != fdn);
 290         fdn->handler = handler;
 291         fdn->fd      = fd;
 292         fdn->ctx     = ctx;
 293         fdn->events  = 0;
 294         fdn->fde_ndx = -1;
 295       #ifdef FDEVENT_USE_LIBEV
 296         fdn->handler_ctx = NULL;
 297       #endif
 298         return fdn;
 299 }
 300
 301 void fdevent_unregister(fdevents *ev, int fd) {
 302         fdnode *fdn = ev->fdarray[fd];
 303         if ((uintptr_t)fdn & 0x3) return; /*(should not happen)*/
 304         ev->fdarray[fd] = NULL;
 305         fdnode_free(fdn);
 306 }
 307
 308 void fdevent_sched_close(fdevents *ev, int fd, int issock) {
 309         fdnode *fdn = ev->fdarray[fd];
 310         if ((uintptr_t)fdn & 0x3) return;
 311         ev->fdarray[fd] = (fdnode *)((uintptr_t)fdn | (issock ? 0x1 : 0x2));
 312         fdn->handler = (fdevent_handler)NULL;
 313         fdn->ctx = ev->pendclose;
 314         ev->pendclose = fdn;
 315 }
 316
 317 static void fdevent_sched_run(fdevents *ev) {
 318         server *srv = ev->srv;
 319         for (fdnode *fdn = ev->pendclose; fdn; ) {
 320                 int fd, rc;
 321                 fdnode *fdn_tmp;
 322               #ifdef _WIN32
 323                 rc = (uintptr_t)fdn & 0x3;
 324               #endif
 325                 fdn = (fdnode *)((uintptr_t)fdn & ~0x3);
 326                 fd = fdn->fd;
 327               #ifdef _WIN32
 328                 if (rc == 0x1) {
 329                         rc = closesocket(fd);
 330                 }
 331                 else if (rc == 0x2) {
 332                         rc = close(fd);
 333                 }
 334               #else
 335                 rc = close(fd);
 336               #endif
 337
 338                 if (0 != rc) {
 339                         log_error_write(srv, __FILE__, __LINE__, "sds", "close failed ", fd, strerror(errno));
 340                 }
 341                 else {
 342                         --srv->cur_fds;
 343                 }
 344
 345                 fdn_tmp = fdn;
 346                 fdn = (fdnode *)fdn->ctx; /* next */
 347                 /*(fdevent_unregister)*/
 348                 fdnode_free(fdn_tmp);
 349                 ev->fdarray[fd] = NULL;
 350         }
 351         ev->pendclose = NULL;
 352 }
 353
 354 __attribute_cold__
 355 __attribute_noinline__
 356 static int fdevent_fdnode_event_unsetter_retry(fdevents *ev, fdnode *fdn) {
 357     do {
 358         switch (errno) {
 359          #ifdef EWOULDBLOCK
 360          #if EAGAIN != EWOULDBLOCK
 361           case EWOULDBLOCK:
 362          #endif
 363          #endif
 364           case EAGAIN:
 365           case EINTR:
 366             /* temporary error; retry */
 367             break;
 368           /*case ENOMEM:*/
 369           default:
 370             /* unrecoverable error; might leak fd */
 371             log_error_write(ev->srv, __FILE__, __LINE__, "sDsS",
 372                             "fdevent event_del failed on fd", fdn->fd, ":",
 373                             strerror(errno));
 374             return 0;
 375         }
 376     } while (0 != ev->event_del(ev, fdn));
 377     return 1;
 378 }
 379
 380 static void fdevent_fdnode_event_unsetter(fdevents *ev, fdnode *fdn) {
 381     if (-1 == fdn->fde_ndx) return;
 382     if (0 != ev->event_del(ev, fdn))
 383         fdevent_fdnode_event_unsetter_retry(ev, fdn);
 384     fdn->fde_ndx = -1;
 385     fdn->events = 0;
 386 }
 387
 388 __attribute_cold__
 389 __attribute_noinline__
 390 static int fdevent_fdnode_event_setter_retry(fdevents *ev, fdnode *fdn, int events) {
 391     do {
 392         switch (errno) {
 393          #ifdef EWOULDBLOCK
 394          #if EAGAIN != EWOULDBLOCK
 395           case EWOULDBLOCK:
 396          #endif
 397          #endif
 398           case EAGAIN:
 399           case EINTR:
 400             /* temporary error; retry */
 401             break;
 402           /*case ENOMEM:*/
 403           default:
 404             /* unrecoverable error */
 405             log_error_write(ev->srv, __FILE__, __LINE__, "sDsS",
 406                             "fdevent event_set failed on fd", fdn->fd, ":",
 407                             strerror(errno));
 408             return 0;
 409         }
 410     } while (0 != ev->event_set(ev, fdn, events));
 411     return 1;
 412 }
 413
 414 static void fdevent_fdnode_event_setter(fdevents *ev, fdnode *fdn, int events) {
 415     /*(Note: skips registering with kernel if initial events is 0,
 416      * so caller should pass non-zero events for initial registration.
 417      * If never registered due to never being called with non-zero events,
 418      * then FDEVENT_HUP or FDEVENT_ERR will never be returned.) */
 419     if (fdn->events == events) return;/*(no change; nothing to do)*/
 420
 421     if (0 == ev->event_set(ev, fdn, events)
 422         || fdevent_fdnode_event_setter_retry(ev, fdn, events))
 423         fdn->events = events;
 424 }
 425
 426 void fdevent_fdnode_event_del(fdevents *ev, fdnode *fdn) {
 427     if (NULL != fdn) fdevent_fdnode_event_unsetter(ev, fdn);
 428 }
 429
 430 void fdevent_fdnode_event_set(fdevents *ev, fdnode *fdn, int events) {
 431     if (NULL != fdn) fdevent_fdnode_event_setter(ev, fdn, events);
 432 }
 433
 434 void fdevent_fdnode_event_add(fdevents *ev, fdnode *fdn, int event) {
 435     if (NULL != fdn) fdevent_fdnode_event_setter(ev, fdn, (fdn->events|event));
 436 }
 437
 438 void fdevent_fdnode_event_clr(fdevents *ev, fdnode *fdn, int event) {
 439     if (NULL != fdn) fdevent_fdnode_event_setter(ev, fdn, (fdn->events&~event));
 440 }
 441
 442 int fdevent_poll(fdevents *ev, int timeout_ms) {
 443     int n = ev->poll(ev, timeout_ms);
 444     if (n >= 0)
 445         fdevent_sched_run(ev);
 446     else if (errno != EINTR)
 447         log_error_write(ev->srv, __FILE__, __LINE__, "SS",
 448                         "fdevent_poll failed: ", strerror(errno));
 449     return n;
 450 }
 451
 452 void fdevent_setfd_cloexec(int fd) {
 453 #ifdef FD_CLOEXEC
 454         if (fd < 0) return;
 455         force_assert(-1 != fcntl(fd, F_SETFD, FD_CLOEXEC));
 456 #else
 457         UNUSED(fd);
 458 #endif
 459 }
 460
 461 void fdevent_clrfd_cloexec(int fd) {
 462 #ifdef FD_CLOEXEC
 463         if (fd >= 0) force_assert(-1 != fcntl(fd, F_SETFD, 0));
 464 #else
 465         UNUSED(fd);
 466 #endif
 467 }
 468
 469 int fdevent_fcntl_set_nb(fdevents *ev, int fd) {
 470         UNUSED(ev);
 471 #ifdef O_NONBLOCK
 472         return fcntl(fd, F_SETFL, O_NONBLOCK | O_RDWR);
 473 #else
 474         UNUSED(fd);
 475         return 0;
 476 #endif
 477 }
 478
 479 int fdevent_fcntl_set_nb_cloexec(fdevents *ev, int fd) {
 480         fdevent_setfd_cloexec(fd);
 481         return fdevent_fcntl_set_nb(ev, fd);
 482 }
 483
 484 int fdevent_fcntl_set_nb_cloexec_sock(fdevents *ev, int fd) {
 485 #if defined(SOCK_CLOEXEC) && defined(SOCK_NONBLOCK)
 486         if (use_sock_cloexec && use_sock_nonblock)
 487                 return 0;
 488 #endif
 489         return fdevent_fcntl_set_nb_cloexec(ev, fd);
 490 }
 491
 492 int fdevent_socket_cloexec(int domain, int type, int protocol) {
 493         int fd;
 494 #ifdef SOCK_CLOEXEC
 495         if (use_sock_cloexec)
 496                 return socket(domain, type | SOCK_CLOEXEC, protocol);
 497 #endif
 498         if (-1 != (fd = socket(domain, type, protocol))) {
 499 #ifdef FD_CLOEXEC
 500                 force_assert(-1 != fcntl(fd, F_SETFD, FD_CLOEXEC));
 501 #endif
 502         }
 503         return fd;
 504 }
 505
 506 int fdevent_socket_nb_cloexec(int domain, int type, int protocol) {
 507         int fd;
 508 #ifdef SOCK_CLOEXEC
 509        #ifdef SOCK_NONBLOCK
 510         if (use_sock_cloexec && use_sock_nonblock)
 511                 return socket(domain, type | SOCK_CLOEXEC | SOCK_NONBLOCK, protocol);
 512        #else
 513         if (use_sock_cloexec) {
 514                 fd = socket(domain, type | SOCK_CLOEXEC, protocol);
 515               #ifdef O_NONBLOCK
 516                 if (-1 != fd) force_assert(-1 != fcntl(fd,F_SETFL,O_NONBLOCK|O_RDWR));
 517               #endif
 518                 return fd;
 519         }
 520        #endif
 521 #endif
 522         if (-1 != (fd = socket(domain, type, protocol))) {
 523 #ifdef FD_CLOEXEC
 524                 force_assert(-1 != fcntl(fd, F_SETFD, FD_CLOEXEC));
 525 #endif
 526 #ifdef O_NONBLOCK
 527                 force_assert(-1 != fcntl(fd, F_SETFL, O_NONBLOCK | O_RDWR));
 528 #endif
 529         }
 530         return fd;
 531 }
 532
 533 #ifndef O_BINARY
 534 #define O_BINARY 0
 535 #endif
 536 #ifndef O_LARGEFILE
 537 #define O_LARGEFILE 0
 538 #endif
 539 #ifndef O_NOCTTY
 540 #define O_NOCTTY 0
 541 #endif
 542 #ifndef O_NOFOLLOW
 543 #define O_NOFOLLOW 0
 544 #endif
 545
 546 /*(O_NOFOLLOW is not handled here)*/
 547 /*(Note: O_NOFOLLOW affects only the final path segment, the target file,
 548  * not any intermediate symlinks along the path)*/
 549
 550 /* O_CLOEXEC handled further below, if defined) */
 551 #ifdef O_NONBLOCK
 552 #define FDEVENT_O_FLAGS \
 553         (O_BINARY | O_LARGEFILE | O_NOCTTY | O_NONBLOCK)
 554 #else
 555 #define FDEVENT_O_FLAGS \
 556         (O_BINARY | O_LARGEFILE | O_NOCTTY )
 557 #endif
 558
 559 int fdevent_open_cloexec(const char *pathname, int symlinks, int flags, mode_t mode) {
 560         if (!symlinks) flags |= O_NOFOLLOW;
 561 #ifdef O_CLOEXEC
 562         return open(pathname, flags | O_CLOEXEC | FDEVENT_O_FLAGS, mode);
 563 #else
 564         int fd = open(pathname, flags | FDEVENT_O_FLAGS, mode);
 565 #ifdef FD_CLOEXEC
 566         if (fd != -1)
 567                 force_assert(-1 != fcntl(fd, F_SETFD, FD_CLOEXEC));
 568 #endif
 569         return fd;
 570 #endif
 571 }
 572
 573
 574 int fdevent_open_devnull(void) {
 575   #if defined(_WIN32)
 576     return fdevent_open_cloexec("nul", 0, O_RDWR, 0);
 577   #else
 578     return fdevent_open_cloexec("/dev/null", 0, O_RDWR, 0);
 579   #endif
 580 }
 581
 582
 583 int fdevent_open_dirname(char *path, int symlinks) {
 584     /*(handle special cases of no dirname or dirname is root directory)*/
 585     char * const c = strrchr(path, '/');
 586     const char * const dname = (NULL != c ? c == path ? "/" : path : ".");
 587     int dfd;
 588     int flags = O_RDONLY;
 589   #ifdef O_DIRECTORY
 590     flags |= O_DIRECTORY;
 591   #endif
 592     if (NULL != c) *c = '\0';
 593     dfd = fdevent_open_cloexec(dname, symlinks, flags, 0);
 594     if (NULL != c) *c = '/';
 595     return dfd;
 596 }
 597
 598
 599 int fdevent_mkstemp_append(char *path) {
 600   #ifdef __COVERITY__
 601     /* POSIX-2008 requires mkstemp create file with 0600 perms */
 602     umask(0600);
 603   #endif
 604     /* coverity[secure_temp : FALSE] */
 605     const int fd = mkstemp(path);
 606     if (fd < 0) return fd;
 607
 608     if (0 != fcntl(fd, F_SETFL, fcntl(fd, F_GETFL, 0) | O_APPEND)) {
 609         /* (should not happen; fd is regular file) */
 610         int errnum = errno;
 611         close(fd);
 612         errno = errnum;
 613         return -1;
 614     }
 615
 616     fdevent_setfd_cloexec(fd);
 617     return fd;
 618 }
 619
 620
 621 int fdevent_accept_listenfd(int listenfd, struct sockaddr *addr, size_t *addrlen) {
 622         int fd;
 623         socklen_t len = (socklen_t) *addrlen;
 624
 625       #if defined(SOCK_CLOEXEC) && defined(SOCK_NONBLOCK)
 626        #if defined(__NetBSD__)
 627         const int sock_cloexec = 1;
 628         fd = paccept(listenfd, addr, &len, NULL, SOCK_CLOEXEC | SOCK_NONBLOCK);
 629        #else
 630         int sock_cloexec = use_sock_cloexec;
 631         if (sock_cloexec) {
 632                 fd = accept4(listenfd, addr, &len, SOCK_CLOEXEC | SOCK_NONBLOCK);
 633                 if (fd >= 0) {
 634                         if (!use_sock_nonblock) {
 635                                 if (0 != fdevent_fcntl_set_nb(NULL, fd)) {
 636                                         close(fd);
 637                                         fd = -1;
 638                                 }
 639                         }
 640                 } else if (errno == ENOSYS || errno == ENOTSUP) {
 641                         fd = accept(listenfd, addr, &len);
 642                         sock_cloexec = 0;
 643                 }
 644         }
 645         else {
 646                 fd = accept(listenfd, addr, &len);
 647         }
 648        #endif
 649       #else
 650         const int sock_cloexec = 0;
 651         fd = accept(listenfd, addr, &len);
 652       #endif
 653
 654         if (fd >= 0) {
 655                 *addrlen = (size_t)len;
 656                 if (!sock_cloexec && 0 != fdevent_fcntl_set_nb_cloexec(NULL, fd)) {
 657                         close(fd);
 658                         fd = -1;
 659                 }
 660         }
 661         return fd;
 662 }
 663
 664
 665 #ifdef __APPLE__
 666 #include <crt_externs.h>
 667 #define environ (* _NSGetEnviron())
 668 #else
 669 extern char **environ;
 670 #endif
 671 char ** fdevent_environ (void) { return environ; }
 672
 673
 674 #ifdef FD_CLOEXEC
 675 static int fdevent_dup2_close_clrfd_cloexec(int oldfd, int newfd) {
 676     if (oldfd >= 0) {
 677         if (oldfd != newfd) {
 678             force_assert(oldfd > STDERR_FILENO);
 679             if (newfd != dup2(oldfd, newfd)) return -1;
 680         }
 681         else {
 682             fdevent_clrfd_cloexec(newfd);
 683         }
 684     }
 685     return newfd;
 686 }
 687 #else
 688 static int fdevent_dup2_close_clrfd_cloexec(int oldfd, int newfd, int reuse) {
 689     if (oldfd >= 0) {
 690         if (oldfd != newfd) {
 691             force_assert(oldfd > STDERR_FILENO);
 692             if (newfd != dup2(oldfd, newfd)) return -1;
 693             if (!reuse) close(oldfd);
 694         }
 695     }
 696     return newfd;
 697 }
 698 #endif
 699
 700
 701 int fdevent_set_stdin_stdout_stderr(int fdin, int fdout, int fderr) {
 702   #ifdef FD_CLOEXEC
 703     if (STDIN_FILENO != fdevent_dup2_close_clrfd_cloexec(fdin, STDIN_FILENO))
 704         return -1;
 705     if (STDOUT_FILENO != fdevent_dup2_close_clrfd_cloexec(fdout, STDOUT_FILENO))
 706         return -1;
 707     if (STDERR_FILENO != fdevent_dup2_close_clrfd_cloexec(fderr, STDERR_FILENO))
 708         return -1;
 709   #else
 710     if (STDIN_FILENO != fdevent_dup2_close_clrfd_cloexec(fdin, STDIN_FILENO,
 711                                                          fdin == fdout
 712                                                          || fdin == fderr))
 713         return -1;
 714     if (STDOUT_FILENO != fdevent_dup2_close_clrfd_cloexec(fdout, STDOUT_FILENO,
 715                                                           fdout == fderr))
 716         return -1;
 717     if (STDERR_FILENO != fdevent_dup2_close_clrfd_cloexec(fderr, STDERR_FILENO,
 718                                                           0))
 719         return -1;
 720   #endif
 721
 722     return 0;
 723 }
 724
 725
 726 #include <stdio.h>      /* perror() */
 727 #include <signal.h>     /* signal() */
 728
 729 pid_t fdevent_fork_execve(const char *name, char *argv[], char *envp[], int fdin, int fdout, int fderr, int dfd) {
 730  #ifdef HAVE_FORK
 731
 732     pid_t pid = fork();
 733     if (0 != pid) return pid; /* parent (pid > 0) or fork() error (-1 == pid) */
 734
 735     /* child (0 == pid) */
 736
 737     if (-1 != dfd) {
 738         if (0 != fchdir(dfd))
 739             _exit(errno);
 740         close(dfd);
 741     }
 742
 743     if (0 != fdevent_set_stdin_stdout_stderr(fdin, fdout, fderr)) _exit(errno);
 744   #ifdef FD_CLOEXEC
 745     /*(might not be sufficient for open fds, but modern OS have FD_CLOEXEC)*/
 746     for (int i = 3; i < 256; ++i) close(i);
 747   #endif
 748
 749     /* reset_signals which may have been ignored (SIG_IGN) */
 750   #ifdef SIGTTOU
 751     signal(SIGTTOU, SIG_DFL);
 752   #endif
 753   #ifdef SIGTTIN
 754     signal(SIGTTIN, SIG_DFL);
 755   #endif
 756   #ifdef SIGTSTP
 757     signal(SIGTSTP, SIG_DFL);
 758   #endif
 759     signal(SIGPIPE, SIG_DFL);
 760
 761     execve(name, argv, envp ? envp : environ);
 762
 763     if (0 == memcmp(argv[0], "/bin/sh", sizeof("/bin/sh")-1)
 764         && argv[1] && 0 == memcmp(argv[1], "-c", sizeof("-c")-1))
 765         perror(argv[2]);
 766     else
 767         perror(argv[0]);
 768     _exit(errno);
 769
 770  #else
 771
 772     UNUSED(name);
 773     UNUSED(argv);
 774     UNUSED(envp);
 775     UNUSED(fdin);
 776     UNUSED(fdout);
 777     UNUSED(fderr);
 778     UNUSED(dfd);
 779     return (pid_t)-1;
 780
 781  #endif
 782 }
 783
 784
 785 typedef struct fdevent_cmd_pipe {
 786     pid_t pid;
 787     int fds[2];
 788     const char *cmd;
 789     time_t start;
 790 } fdevent_cmd_pipe;
 791
 792 typedef struct fdevent_cmd_pipes {
 793     fdevent_cmd_pipe *ptr;
 794     size_t used;
 795     size_t size;
 796 } fdevent_cmd_pipes;
 797
 798 static fdevent_cmd_pipes cmd_pipes;
 799
 800
 801 static pid_t fdevent_open_logger_pipe_spawn(const char *logger, int rfd) {
 802     char *args[4];
 803     int devnull = fdevent_open_devnull();
 804     pid_t pid;
 805
 806     if (-1 == devnull) {
 807         return -1;
 808     }
 809
 810     *(const char **)&args[0] = "/bin/sh";
 811     *(const char **)&args[1] = "-c";
 812     *(const char **)&args[2] = logger;
 813     args[3] = NULL;
 814
 815     pid = fdevent_fork_execve(args[0], args, NULL, rfd, devnull, devnull, -1);
 816
 817     if (pid > 0) {
 818         close(devnull);
 819     }
 820     else {
 821         int errnum = errno;
 822         close(devnull);
 823         errno = errnum;
 824     }
 825     return pid;
 826 }
 827
 828
 829 static void fdevent_restart_logger_pipe(fdevent_cmd_pipe *fcp, time_t ts) {
 830     if (fcp->pid > 0) return;  /* assert */
 831     if (fcp->start + 5 < ts) { /* limit restart to once every 5 sec */
 832         /* restart child process using existing pipe fds */
 833         fcp->start = ts;
 834         fcp->pid = fdevent_open_logger_pipe_spawn(fcp->cmd, fcp->fds[0]);
 835     }
 836 }
 837
 838
 839 void fdevent_restart_logger_pipes(time_t ts) {
 840     for (size_t i = 0; i < cmd_pipes.used; ++i) {
 841         fdevent_cmd_pipe * const fcp = cmd_pipes.ptr+i;
 842         if (fcp->pid > 0) continue;
 843         fdevent_restart_logger_pipe(fcp, ts);
 844     }
 845 }
 846
 847
 848 int fdevent_waitpid_logger_pipe_pid(pid_t pid, time_t ts) {
 849     for (size_t i = 0; i < cmd_pipes.used; ++i) {
 850         fdevent_cmd_pipe * const fcp = cmd_pipes.ptr+i;
 851         if (pid != fcp->pid) continue;
 852         fcp->pid = -1;
 853         fdevent_restart_logger_pipe(fcp, ts);
 854         return 1;
 855     }
 856     return 0;
 857 }
 858
 859
 860 void fdevent_clr_logger_pipe_pids(void) {
 861     for (size_t i = 0; i < cmd_pipes.used; ++i) {
 862         fdevent_cmd_pipe *fcp = cmd_pipes.ptr+i;
 863         fcp->pid = -1;
 864     }
 865 }
 866
 867
 868 int fdevent_reaped_logger_pipe(pid_t pid) {
 869     for (size_t i = 0; i < cmd_pipes.used; ++i) {
 870         fdevent_cmd_pipe *fcp = cmd_pipes.ptr+i;
 871         if (fcp->pid == pid) {
 872             time_t ts = time(NULL);
 873             if (fcp->start + 5 < ts) { /* limit restart to once every 5 sec */
 874                 fcp->start = ts;
 875                 fcp->pid = fdevent_open_logger_pipe_spawn(fcp->cmd,fcp->fds[0]);
 876                 return 1;
 877             }
 878             else {
 879                 fcp->pid = -1;
 880                 return -1;
 881             }
 882         }
 883     }
 884     return 0;
 885 }
 886
 887
 888 void fdevent_close_logger_pipes(void) {
 889     for (size_t i = 0; i < cmd_pipes.used; ++i) {
 890         fdevent_cmd_pipe *fcp = cmd_pipes.ptr+i;
 891         close(fcp->fds[0]);
 892         if (fcp->fds[1] != STDERR_FILENO) close(fcp->fds[1]);
 893     }
 894     free(cmd_pipes.ptr);
 895     cmd_pipes.ptr = NULL;
 896     cmd_pipes.used = 0;
 897     cmd_pipes.size = 0;
 898 }
 899
 900
 901 void fdevent_breakagelog_logger_pipe(int fd) {
 902     for (size_t i = 0; i < cmd_pipes.used; ++i) {
 903         fdevent_cmd_pipe *fcp = cmd_pipes.ptr+i;
 904         if (fcp->fds[1] != fd) continue;
 905         fcp->fds[1] = STDERR_FILENO;
 906         break;
 907     }
 908 }
 909
 910
 911 static void fdevent_init_logger_pipe(const char *cmd, int fds[2], pid_t pid) {
 912     fdevent_cmd_pipe *fcp;
 913     if (cmd_pipes.used == cmd_pipes.size) {
 914         cmd_pipes.size += 4;
 915         cmd_pipes.ptr =
 916           realloc(cmd_pipes.ptr, cmd_pipes.size * sizeof(fdevent_cmd_pipe));
 917         force_assert(cmd_pipes.ptr);
 918     }
 919     fcp = cmd_pipes.ptr + cmd_pipes.used++;
 920     fcp->cmd = cmd; /* note: cmd must persist in memory (or else copy here) */
 921     fcp->fds[0] = fds[0];
 922     fcp->fds[1] = fds[1];
 923     fcp->pid = pid;
 924     fcp->start = time(NULL);
 925 }
 926
 927
 928 static int fdevent_open_logger_pipe(const char *logger) {
 929     int fds[2];
 930     pid_t pid;
 931     if (pipe(fds)) {
 932         return -1;
 933     }
 934     fdevent_setfd_cloexec(fds[0]);
 935     fdevent_setfd_cloexec(fds[1]);
 936     /*(nonblocking write() from lighttpd)*/
 937     if (0 != fdevent_fcntl_set_nb(NULL, fds[1])) { /*(ignore)*/ }
 938
 939     pid = fdevent_open_logger_pipe_spawn(logger, fds[0]);
 940
 941     if (pid > 0) {
 942         fdevent_init_logger_pipe(logger, fds, pid);
 943         return fds[1];
 944     }
 945     else {
 946         int errnum = errno;
 947         close(fds[0]);
 948         close(fds[1]);
 949         errno = errnum;
 950         return -1;
 951     }
 952 }
 953
 954
 955 int fdevent_open_logger(const char *logger) {
 956     if (logger[0] != '|') { /*(permit symlinks)*/
 957         int flags = O_APPEND | O_WRONLY | O_CREAT;
 958         return fdevent_open_cloexec(logger, 1, flags, 0644);
 959     }
 960     else {
 961         return fdevent_open_logger_pipe(logger+1); /*(skip the '|')*/
 962     }
 963 }
 964
 965 int fdevent_cycle_logger(const char *logger, int *curfd) {
 966     if (logger[0] != '|') {
 967         int fd = fdevent_open_logger(logger);
 968         if (-1 == fd) return -1; /*(error; leave *curfd as-is)*/
 969         if (-1 != *curfd) close(*curfd);
 970         *curfd = fd;
 971     }
 972     return *curfd;
 973 }
 974
 975
 976 #ifndef MSG_DONTWAIT
 977 #define MSG_DONTWAIT 0
 978 #endif
 979 #ifndef MSG_NOSIGNAL
 980 #define MSG_NOSIGNAL 0
 981 #endif
 982
 983
 984 ssize_t fdevent_socket_read_discard (int fd, char *buf, size_t sz, int family, int so_type) {
 985   #if defined(MSG_TRUNC) && defined(__linux__)
 986     if ((family == AF_INET || family == AF_INET6) && so_type == SOCK_STREAM) {
 987         ssize_t len = recv(fd, buf, sz, MSG_TRUNC|MSG_DONTWAIT|MSG_NOSIGNAL);
 988         if (len >= 0 || errno != EINVAL) return len;
 989     }
 990   #else
 991     UNUSED(family);
 992     UNUSED(so_type);
 993   #endif
 994     return read(fd, buf, sz);
 995 }
 996
 997
 998 #include <sys/ioctl.h>
 999 #ifdef HAVE_SYS_FILIO_H
1000 #include <sys/filio.h>  /* FIONREAD (for illumos (OpenIndiana)) */
1001 #endif
1002 #ifdef _WIN32
1003 #include <winsock2.h>
1004 #endif
1005 int fdevent_ioctl_fionread (int fd, int fdfmt, int *toread) {
1006   #ifdef _WIN32
1007     if (fdfmt != S_IFSOCK) { errno = ENOTSOCK; return -1; }
1008     return ioctlsocket(fd, FIONREAD, toread);
1009   #else
1010    #ifdef __CYGWIN__
1011     /*(cygwin supports FIONREAD on pipes, not sockets)*/
1012     if (fdfmt != S_IFIFO) { errno = EOPNOTSUPP; return -1; }
1013    #else
1014     UNUSED(fdfmt);
1015    #endif
1016     return ioctl(fd, FIONREAD, toread);
1017   #endif
1018 }
1019
1020
1021 int fdevent_connect_status(int fd) {
1022     /* try to finish the connect() */
1023     /*(should be called after connect() only when fd is writable (POLLOUT))*/
1024     int opt;
1025     socklen_t len = sizeof(opt);
1026     return (0 == getsockopt(fd,SOL_SOCKET,SO_ERROR,&opt,&len)) ? opt : errno;
1027 }
1028
1029
1030 #include <netinet/tcp.h>
1031 #if (defined(__APPLE__) && defined(__MACH__)) \
1032   || defined(__FreeBSD__) || defined(__NetBSD__) \
1033   || defined(__OpenBSD__) || defined(__DragonFly__)
1034 #include <netinet/tcp_fsm.h>
1035 #endif
1036
1037 /* fd must be TCP socket (AF_INET, AF_INET6), end-of-stream recv() 0 bytes */
1038 int fdevent_is_tcp_half_closed(int fd) {
1039   #ifdef TCP_CONNECTION_INFO     /* Darwin */
1040     struct tcp_connection_info tcpi;
1041     socklen_t tlen = sizeof(tcpi);
1042     return (0 == getsockopt(fd, IPPROTO_TCP, TCP_CONNECTION_INFO, &tcpi, &tlen)
1043             && tcpi.tcpi_state == TCPS_CLOSE_WAIT);
1044   #elif defined(TCP_INFO) && defined(TCPS_CLOSE_WAIT)
1045     /* FreeBSD, NetBSD (not present in OpenBSD or DragonFlyBSD) */
1046     struct tcp_info tcpi;
1047     socklen_t tlen = sizeof(tcpi);
1048     return (0 == getsockopt(fd, IPPROTO_TCP, TCP_INFO, &tcpi, &tlen)
1049             && tcpi.tcpi_state == TCPS_CLOSE_WAIT);
1050   #elif defined(TCP_INFO) && defined(__linux__)
1051     /* Linux (TCP_CLOSE_WAIT is enum, so can not #ifdef TCP_CLOSE_WAIT) */
1052     struct tcp_info tcpi;
1053     socklen_t tlen = sizeof(tcpi);/*SOL_TCP == IPPROTO_TCP*/
1054     return (0 == getsockopt(fd,     SOL_TCP, TCP_INFO, &tcpi, &tlen)
1055             && tcpi.tcpi_state == TCP_CLOSE_WAIT);
1056   #else
1057     UNUSED(fd);
1058     /*(0 != getpeername() error might indicate TCP RST, but success
1059      * would not differentiate between half-close and full-close)*/
1060     return 0; /* false (not half-closed) or TCP state unknown */
1061   #endif
1062 }
1063
1064
1065 int fdevent_set_tcp_nodelay (const int fd, const int opt)
1066 {
1067     return setsockopt(fd, IPPROTO_TCP, TCP_NODELAY, &opt, sizeof(opt));
1068 }
1069
1070
1071 int fdevent_set_so_reuseaddr (const int fd, const int opt)
1072 {
1073     return setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, &opt, sizeof(opt));
1074 }