lib/tevent/tevent_epoll.c

   1 /*
   2    Unix SMB/CIFS implementation.
   3
   4    main select loop and event handling - epoll implementation
   5
   6    Copyright (C) Andrew Tridgell        2003-2005
   7    Copyright (C) Stefan Metzmacher      2005-2013
   8    Copyright (C) Jeremy Allison         2013
   9
  10      ** NOTE! The following LGPL license applies to the tevent
  11      ** library. This does NOT imply that all of Samba is released
  12      ** under the LGPL
  13
  14    This library is free software; you can redistribute it and/or
  15    modify it under the terms of the GNU Lesser General Public
  16    License as published by the Free Software Foundation; either
  17    version 3 of the License, or (at your option) any later version.
  18
  19    This library is distributed in the hope that it will be useful,
  20    but WITHOUT ANY WARRANTY; without even the implied warranty of
  21    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  22    Lesser General Public License for more details.
  23
  24    You should have received a copy of the GNU Lesser General Public
  25    License along with this library; if not, see <http://www.gnu.org/licenses/>.
  26 */
  27
  28 #include "replace.h"
  29 #include "system/filesys.h"
  30 #include "system/select.h"
  31 #include "tevent.h"
  32 #include "tevent_internal.h"
  33 #include "tevent_util.h"
  34
  35 struct epoll_event_context {
  36         /* a pointer back to the generic event_context */
  37         struct tevent_context *ev;
  38
  39         /* when using epoll this is the handle from epoll_create */
  40         int epoll_fd;
  41
  42         pid_t pid;
  43
  44         bool panic_force_replay;
  45         bool *panic_state;
  46         bool (*panic_fallback)(struct tevent_context *ev, bool replay);
  47 };
  48
  49 #define EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT      (1<<0)
  50 #define EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR   (1<<1)
  51 #define EPOLL_ADDITIONAL_FD_FLAG_GOT_ERROR      (1<<2)
  52 #define EPOLL_ADDITIONAL_FD_FLAG_HAS_MPX        (1<<3)
  53
  54 #ifdef TEST_PANIC_FALLBACK
  55
  56 static int epoll_create_panic_fallback(struct epoll_event_context *epoll_ev,
  57                                        int size)
  58 {
  59         if (epoll_ev->panic_fallback == NULL) {
  60                 return epoll_create(size);
  61         }
  62
  63         /* 50% of the time, fail... */
  64         if ((random() % 2) == 0) {
  65                 errno = EINVAL;
  66                 return -1;
  67         }
  68
  69         return epoll_create(size);
  70 }
  71
  72 static int epoll_ctl_panic_fallback(struct epoll_event_context *epoll_ev,
  73                                     int epfd, int op, int fd,
  74                                     struct epoll_event *event)
  75 {
  76         if (epoll_ev->panic_fallback == NULL) {
  77                 return epoll_ctl(epfd, op, fd, event);
  78         }
  79
  80         /* 50% of the time, fail... */
  81         if ((random() % 2) == 0) {
  82                 errno = EINVAL;
  83                 return -1;
  84         }
  85
  86         return epoll_ctl(epfd, op, fd, event);
  87 }
  88
  89 static int epoll_wait_panic_fallback(struct epoll_event_context *epoll_ev,
  90                                      int epfd,
  91                                      struct epoll_event *events,
  92                                      int maxevents,
  93                                      int timeout)
  94 {
  95         if (epoll_ev->panic_fallback == NULL) {
  96                 return epoll_wait(epfd, events, maxevents, timeout);
  97         }
  98
  99         /* 50% of the time, fail... */
 100         if ((random() % 2) == 0) {
 101                 errno = EINVAL;
 102                 return -1;
 103         }
 104
 105         return epoll_wait(epfd, events, maxevents, timeout);
 106 }
 107
 108 #define epoll_create(_size) \
 109         epoll_create_panic_fallback(epoll_ev, _size)
 110 #define epoll_ctl(_epfd, _op, _fd, _event) \
 111         epoll_ctl_panic_fallback(epoll_ev,_epfd, _op, _fd, _event)
 112 #define epoll_wait(_epfd, _events, _maxevents, _timeout) \
 113         epoll_wait_panic_fallback(epoll_ev, _epfd, _events, _maxevents, _timeout)
 114 #endif
 115
 116 /*
 117   called to set the panic fallback function.
 118 */
 119 _PRIVATE_ bool tevent_epoll_set_panic_fallback(struct tevent_context *ev,
 120                                 bool (*panic_fallback)(struct tevent_context *ev,
 121                                                        bool replay))
 122 {
 123         struct epoll_event_context *epoll_ev;
 124
 125         if (ev->additional_data == NULL) {
 126                 return false;
 127         }
 128
 129         epoll_ev = talloc_get_type(ev->additional_data,
 130                                 struct epoll_event_context);
 131         if (epoll_ev == NULL) {
 132                 return false;
 133         }
 134         epoll_ev->panic_fallback = panic_fallback;
 135         return true;
 136 }
 137
 138 /*
 139   called when a epoll call fails
 140 */
 141 static void epoll_panic(struct epoll_event_context *epoll_ev,
 142                         const char *reason, bool replay)
 143 {
 144         struct tevent_context *ev = epoll_ev->ev;
 145         bool (*panic_fallback)(struct tevent_context *ev, bool replay);
 146
 147         panic_fallback = epoll_ev->panic_fallback;
 148
 149         if (epoll_ev->panic_state != NULL) {
 150                 *epoll_ev->panic_state = true;
 151         }
 152
 153         if (epoll_ev->panic_force_replay) {
 154                 replay = true;
 155         }
 156
 157         TALLOC_FREE(ev->additional_data);
 158
 159         if (panic_fallback == NULL) {
 160                 tevent_debug(ev, TEVENT_DEBUG_FATAL,
 161                         "%s (%s) replay[%u] - calling abort()\n",
 162                         reason, strerror(errno), (unsigned)replay);
 163                 abort();
 164         }
 165
 166         tevent_debug(ev, TEVENT_DEBUG_WARNING,
 167                      "%s (%s) replay[%u] - calling panic_fallback\n",
 168                      reason, strerror(errno), (unsigned)replay);
 169
 170         if (!panic_fallback(ev, replay)) {
 171                 /* Fallback failed. */
 172                 tevent_debug(ev, TEVENT_DEBUG_FATAL,
 173                         "%s (%s) replay[%u] - calling abort()\n",
 174                         reason, strerror(errno), (unsigned)replay);
 175                 abort();
 176         }
 177 }
 178
 179 /*
 180   map from TEVENT_FD_* to EPOLLIN/EPOLLOUT
 181 */
 182 static uint32_t epoll_map_flags(uint16_t flags)
 183 {
 184         uint32_t ret = 0;
 185         if (flags & TEVENT_FD_READ) ret |= (EPOLLIN | EPOLLERR | EPOLLHUP);
 186         if (flags & TEVENT_FD_WRITE) ret |= (EPOLLOUT | EPOLLERR | EPOLLHUP);
 187         return ret;
 188 }
 189
 190 /*
 191  free the epoll fd
 192 */
 193 static int epoll_ctx_destructor(struct epoll_event_context *epoll_ev)
 194 {
 195         close(epoll_ev->epoll_fd);
 196         epoll_ev->epoll_fd = -1;
 197         return 0;
 198 }
 199
 200 /*
 201  init the epoll fd
 202 */
 203 static int epoll_init_ctx(struct epoll_event_context *epoll_ev)
 204 {
 205         epoll_ev->epoll_fd = epoll_create(64);
 206         if (epoll_ev->epoll_fd == -1) {
 207                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_FATAL,
 208                              "Failed to create epoll handle.\n");
 209                 return -1;
 210         }
 211
 212         if (!ev_set_close_on_exec(epoll_ev->epoll_fd)) {
 213                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_WARNING,
 214                              "Failed to set close-on-exec, file descriptor may be leaked to children.\n");
 215         }
 216
 217         epoll_ev->pid = getpid();
 218         talloc_set_destructor(epoll_ev, epoll_ctx_destructor);
 219
 220         return 0;
 221 }
 222
 223 static void epoll_update_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde);
 224
 225 /*
 226   reopen the epoll handle when our pid changes
 227   see http://junkcode.samba.org/ftp/unpacked/junkcode/epoll_fork.c for an
 228   demonstration of why this is needed
 229  */
 230 static void epoll_check_reopen(struct epoll_event_context *epoll_ev)
 231 {
 232         struct tevent_fd *fde;
 233         bool *caller_panic_state = epoll_ev->panic_state;
 234         bool panic_triggered = false;
 235
 236         if (epoll_ev->pid == getpid()) {
 237                 return;
 238         }
 239
 240         close(epoll_ev->epoll_fd);
 241         epoll_ev->epoll_fd = epoll_create(64);
 242         if (epoll_ev->epoll_fd == -1) {
 243                 epoll_panic(epoll_ev, "epoll_create() failed", false);
 244                 return;
 245         }
 246
 247         if (!ev_set_close_on_exec(epoll_ev->epoll_fd)) {
 248                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_WARNING,
 249                              "Failed to set close-on-exec, file descriptor may be leaked to children.\n");
 250         }
 251
 252         epoll_ev->pid = getpid();
 253         epoll_ev->panic_state = &panic_triggered;
 254         for (fde=epoll_ev->ev->fd_events;fde;fde=fde->next) {
 255                 fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 256                 epoll_update_event(epoll_ev, fde);
 257
 258                 if (panic_triggered) {
 259                         if (caller_panic_state != NULL) {
 260                                 *caller_panic_state = true;
 261                         }
 262                         return;
 263                 }
 264         }
 265         epoll_ev->panic_state = NULL;
 266 }
 267
 268 /*
 269  epoll cannot add the same file descriptor twice, once
 270  with read, once with write which is allowed by the
 271  tevent backend. Multiplex the existing fde, flag it
 272  as such so we can search for the correct fde on
 273  event triggering.
 274 */
 275
 276 static int epoll_add_multiplex_fd(struct epoll_event_context *epoll_ev,
 277                                   struct tevent_fd *add_fde)
 278 {
 279         struct epoll_event event;
 280         struct tevent_fd *mpx_fde;
 281         int ret;
 282
 283         /* Find the existing fde that caused the EEXIST error. */
 284         for (mpx_fde = epoll_ev->ev->fd_events; mpx_fde; mpx_fde = mpx_fde->next) {
 285                 if (mpx_fde->fd != add_fde->fd) {
 286                         continue;
 287                 }
 288
 289                 if (mpx_fde == add_fde) {
 290                         continue;
 291                 }
 292
 293                 break;
 294         }
 295         if (mpx_fde == NULL) {
 296                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_FATAL,
 297                              "can't find multiplex fde for fd[%d]",
 298                              add_fde->fd);
 299                 return -1;
 300         }
 301
 302         if (mpx_fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_HAS_MPX) {
 303                 /* Logic error. Can't have more than 2 multiplexed fde's. */
 304                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_FATAL,
 305                              "multiplex fde for fd[%d] is already multiplexed\n",
 306                              mpx_fde->fd);
 307                 return -1;
 308         }
 309
 310         /*
 311          * The multiplex fde must have the same fd, and also
 312          * already have an epoll event attached.
 313          */
 314         if (!(mpx_fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT)) {
 315                 /* Logic error. Can't have more than 2 multiplexed fde's. */
 316                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_FATAL,
 317                              "multiplex fde for fd[%d] has no event\n",
 318                              mpx_fde->fd);
 319                 return -1;
 320         }
 321
 322         /* Modify the mpx_fde to add in the new flags. */
 323         ZERO_STRUCT(event);
 324         event.events = epoll_map_flags(mpx_fde->flags);
 325         event.events |= epoll_map_flags(add_fde->flags);
 326         event.data.ptr = mpx_fde;
 327         ret = epoll_ctl(epoll_ev->epoll_fd, EPOLL_CTL_MOD, mpx_fde->fd, &event);
 328         if (ret != 0 && errno == EBADF) {
 329                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_ERROR,
 330                              "EPOLL_CTL_MOD EBADF for "
 331                              "add_fde[%p] mpx_fde[%p] fd[%d] - disabling\n",
 332                              add_fde, mpx_fde, add_fde->fd);
 333                 DLIST_REMOVE(epoll_ev->ev->fd_events, mpx_fde);
 334                 mpx_fde->event_ctx = NULL;
 335                 DLIST_REMOVE(epoll_ev->ev->fd_events, add_fde);
 336                 add_fde->event_ctx = NULL;
 337                 return 0;
 338         } else if (ret != 0) {
 339                 return ret;
 340         }
 341
 342         /*
 343          * Make each fde->additional_data pointers point at each other
 344          * so we can look them up from each other. They are now paired.
 345          */
 346         mpx_fde->additional_data = (struct tevent_fd *)add_fde;
 347         add_fde->additional_data = (struct tevent_fd *)mpx_fde;
 348
 349         /* Now flag both fde's as being multiplexed. */
 350         mpx_fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_HAS_MPX;
 351         add_fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_HAS_MPX;
 352
 353         /* we need to keep the GOT_ERROR flag */
 354         if (mpx_fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_GOT_ERROR) {
 355                 add_fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_GOT_ERROR;
 356         }
 357
 358         return 0;
 359 }
 360
 361 /*
 362  add the epoll event to the given fd_event
 363 */
 364 static void epoll_add_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde)
 365 {
 366         struct epoll_event event;
 367         int ret;
 368         struct tevent_fd *mpx_fde = NULL;
 369
 370         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 371         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 372
 373         if (fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_HAS_MPX) {
 374                 /*
 375                  * This is a multiplexed fde, we need to include both
 376                  * flags in the modified event.
 377                  */
 378                 mpx_fde = talloc_get_type_abort(fde->additional_data,
 379                                                 struct tevent_fd);
 380
 381                 mpx_fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 382                 mpx_fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 383         }
 384
 385         ZERO_STRUCT(event);
 386         event.events = epoll_map_flags(fde->flags);
 387         if (mpx_fde != NULL) {
 388                 event.events |= epoll_map_flags(mpx_fde->flags);
 389         }
 390         event.data.ptr = fde;
 391         ret = epoll_ctl(epoll_ev->epoll_fd, EPOLL_CTL_ADD, fde->fd, &event);
 392         if (ret != 0 && errno == EBADF) {
 393                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_ERROR,
 394                              "EPOLL_CTL_ADD EBADF for "
 395                              "fde[%p] mpx_fde[%p] fd[%d] - disabling\n",
 396                              fde, mpx_fde, fde->fd);
 397                 DLIST_REMOVE(epoll_ev->ev->fd_events, fde);
 398                 fde->event_ctx = NULL;
 399                 if (mpx_fde != NULL) {
 400                         DLIST_REMOVE(epoll_ev->ev->fd_events, mpx_fde);
 401                         mpx_fde->event_ctx = NULL;
 402                 }
 403                 return;
 404         } else if (ret != 0) {
 405                 epoll_panic(epoll_ev, "EPOLL_CTL_ADD failed", false);
 406                 return;
 407         }
 408
 409         fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 410         /* only if we want to read we want to tell the event handler about errors */
 411         if (fde->flags & TEVENT_FD_READ) {
 412                 fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 413         }
 414
 415         if (mpx_fde == NULL) {
 416                 return;
 417         }
 418
 419         mpx_fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 420         /* only if we want to read we want to tell the event handler about errors */
 421         if (mpx_fde->flags & TEVENT_FD_READ) {
 422                 mpx_fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 423         }
 424 }
 425
 426 /*
 427  delete the epoll event for given fd_event
 428 */
 429 static void epoll_del_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde)
 430 {
 431         struct epoll_event event;
 432         int ret;
 433
 434         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 435         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 436
 437         ZERO_STRUCT(event);
 438         ret = epoll_ctl(epoll_ev->epoll_fd, EPOLL_CTL_DEL, fde->fd, &event);
 439         if (ret != 0 && errno == ENOENT) {
 440                 /*
 441                  * This can happen after a epoll_check_reopen
 442                  * within epoll_event_fd_destructor.
 443                  */
 444                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_TRACE,
 445                              "EPOLL_CTL_DEL ignoring ENOENT for fd[%d]\n",
 446                              fde->fd);
 447                 return;
 448         } else if (ret != 0 && errno == EBADF) {
 449                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_WARNING,
 450                              "EPOLL_CTL_DEL EBADF for "
 451                              "fde[%p] fd[%d] - disabling\n",
 452                              fde, fde->fd);
 453                 DLIST_REMOVE(epoll_ev->ev->fd_events, fde);
 454                 fde->event_ctx = NULL;
 455                 return;
 456         } else if (ret != 0) {
 457                 epoll_panic(epoll_ev, "EPOLL_CTL_DEL failed", false);
 458                 return;
 459         }
 460 }
 461
 462 /*
 463  change the epoll event to the given fd_event
 464 */
 465 static void epoll_mod_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde)
 466 {
 467         struct epoll_event event;
 468         int ret;
 469
 470         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 471         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 472
 473         ZERO_STRUCT(event);
 474         event.events = epoll_map_flags(fde->flags);
 475         event.data.ptr = fde;
 476         ret = epoll_ctl(epoll_ev->epoll_fd, EPOLL_CTL_MOD, fde->fd, &event);
 477         if (ret != 0 && errno == EBADF) {
 478                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_ERROR,
 479                              "EPOLL_CTL_MOD EBADF for "
 480                              "fde[%p] fd[%d] - disabling\n",
 481                              fde, fde->fd);
 482                 DLIST_REMOVE(epoll_ev->ev->fd_events, fde);
 483                 fde->event_ctx = NULL;
 484                 return;
 485         } else if (ret != 0) {
 486                 epoll_panic(epoll_ev, "EPOLL_CTL_MOD failed", false);
 487                 return;
 488         }
 489
 490         fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 491         /* only if we want to read we want to tell the event handler about errors */
 492         if (fde->flags & TEVENT_FD_READ) {
 493                 fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 494         }
 495 }
 496
 497 static void epoll_update_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde)
 498 {
 499         bool got_error = (fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_GOT_ERROR);
 500         bool want_read = (fde->flags & TEVENT_FD_READ);
 501         bool want_write= (fde->flags & TEVENT_FD_WRITE);
 502
 503         /* there's already an event */
 504         if (fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT) {
 505                 if (want_read || (want_write && !got_error)) {
 506                         epoll_mod_event(epoll_ev, fde);
 507                         return;
 508                 }
 509                 /*
 510                  * if we want to match the select behavior, we need to remove the epoll_event
 511                  * when the caller isn't interested in events.
 512                  *
 513                  * this is because epoll reports EPOLLERR and EPOLLHUP, even without asking for them
 514                  */
 515                 epoll_del_event(epoll_ev, fde);
 516                 return;
 517         }
 518
 519         /* there's no epoll_event attached to the fde */
 520         if (want_read || (want_write && !got_error)) {
 521                 epoll_add_event(epoll_ev, fde);
 522                 return;
 523         }
 524 }
 525
 526 /*
 527   event loop handling using epoll
 528 */
 529 static int epoll_event_loop(struct epoll_event_context *epoll_ev, struct timeval *tvalp)
 530 {
 531         int ret, i;
 532 #define MAXEVENTS 1
 533         struct epoll_event events[MAXEVENTS];
 534         int timeout = -1;
 535         int wait_errno;
 536
 537         if (tvalp) {
 538                 /* it's better to trigger timed events a bit later than too early */
 539                 timeout = ((tvalp->tv_usec+999) / 1000) + (tvalp->tv_sec*1000);
 540         }
 541
 542         if (epoll_ev->ev->signal_events &&
 543             tevent_common_check_signal(epoll_ev->ev)) {
 544                 return 0;
 545         }
 546
 547         tevent_trace_point_callback(epoll_ev->ev, TEVENT_TRACE_BEFORE_WAIT);
 548         ret = epoll_wait(epoll_ev->epoll_fd, events, MAXEVENTS, timeout);
 549         wait_errno = errno;
 550         tevent_trace_point_callback(epoll_ev->ev, TEVENT_TRACE_AFTER_WAIT);
 551
 552         if (ret == -1 && wait_errno == EINTR && epoll_ev->ev->signal_events) {
 553                 if (tevent_common_check_signal(epoll_ev->ev)) {
 554                         return 0;
 555                 }
 556         }
 557
 558         if (ret == -1 && wait_errno != EINTR) {
 559                 epoll_panic(epoll_ev, "epoll_wait() failed", true);
 560                 return -1;
 561         }
 562
 563         if (ret == 0 && tvalp) {
 564                 /* we don't care about a possible delay here */
 565                 tevent_common_loop_timer_delay(epoll_ev->ev);
 566                 return 0;
 567         }
 568
 569         for (i=0;i<ret;i++) {
 570                 struct tevent_fd *fde = talloc_get_type(events[i].data.ptr,
 571                                                        struct tevent_fd);
 572                 uint16_t flags = 0;
 573
 574                 if (fde == NULL) {
 575                         epoll_panic(epoll_ev, "epoll_wait() gave bad data", true);
 576                         return -1;
 577                 }
 578                 if (events[i].events & (EPOLLHUP|EPOLLERR)) {
 579                         fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_GOT_ERROR;
 580                         /*
 581                          * if we only wait for TEVENT_FD_WRITE, we should not tell the
 582                          * event handler about it, and remove the epoll_event,
 583                          * as we only report errors when waiting for read events,
 584                          * to match the select() behavior
 585                          */
 586                         if (!(fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR)) {
 587                                 epoll_update_event(epoll_ev, fde);
 588                                 continue;
 589                         }
 590                         flags |= TEVENT_FD_READ;
 591                 }
 592                 if (events[i].events & EPOLLIN) flags |= TEVENT_FD_READ;
 593                 if (events[i].events & EPOLLOUT) flags |= TEVENT_FD_WRITE;
 594                 if (flags) {
 595                         fde->handler(epoll_ev->ev, fde, flags, fde->private_data);
 596                         break;
 597                 }
 598         }
 599
 600         return 0;
 601 }
 602
 603 /*
 604   create a epoll_event_context structure.
 605 */
 606 static int epoll_event_context_init(struct tevent_context *ev)
 607 {
 608         int ret;
 609         struct epoll_event_context *epoll_ev;
 610
 611         /*
 612          * We might be called during tevent_re_initialise()
 613          * which means we need to free our old additional_data.
 614          */
 615         TALLOC_FREE(ev->additional_data);
 616
 617         epoll_ev = talloc_zero(ev, struct epoll_event_context);
 618         if (!epoll_ev) return -1;
 619         epoll_ev->ev = ev;
 620         epoll_ev->epoll_fd = -1;
 621
 622         ret = epoll_init_ctx(epoll_ev);
 623         if (ret != 0) {
 624                 talloc_free(epoll_ev);
 625                 return ret;
 626         }
 627
 628         ev->additional_data = epoll_ev;
 629         return 0;
 630 }
 631
 632 /*
 633   destroy an fd_event
 634 */
 635 static int epoll_event_fd_destructor(struct tevent_fd *fde)
 636 {
 637         struct tevent_context *ev = fde->event_ctx;
 638         struct epoll_event_context *epoll_ev = NULL;
 639         bool panic_triggered = false;
 640         int flags = fde->flags;
 641
 642         if (ev == NULL) {
 643                 return tevent_common_fd_destructor(fde);
 644         }
 645
 646         epoll_ev = talloc_get_type_abort(ev->additional_data,
 647                                          struct epoll_event_context);
 648
 649         /*
 650          * we must remove the event from the list
 651          * otherwise a panic fallback handler may
 652          * reuse invalid memory
 653          */
 654         DLIST_REMOVE(ev->fd_events, fde);
 655
 656         epoll_ev->panic_state = &panic_triggered;
 657         epoll_check_reopen(epoll_ev);
 658         if (panic_triggered) {
 659                 return tevent_common_fd_destructor(fde);
 660         }
 661
 662         fde->flags = 0;
 663         epoll_update_event(epoll_ev, fde);
 664         fde->flags = flags;
 665         if (panic_triggered) {
 666                 return tevent_common_fd_destructor(fde);
 667         }
 668         epoll_ev->panic_state = NULL;
 669
 670         return tevent_common_fd_destructor(fde);
 671 }
 672
 673 /*
 674   add a fd based event
 675   return NULL on failure (memory allocation error)
 676 */
 677 static struct tevent_fd *epoll_event_add_fd(struct tevent_context *ev, TALLOC_CTX *mem_ctx,
 678                                             int fd, uint16_t flags,
 679                                             tevent_fd_handler_t handler,
 680                                             void *private_data,
 681                                             const char *handler_name,
 682                                             const char *location)
 683 {
 684         struct epoll_event_context *epoll_ev = talloc_get_type(ev->additional_data,
 685                                                            struct epoll_event_context);
 686         struct tevent_fd *fde;
 687         bool panic_triggered = false;
 688
 689         fde = tevent_common_add_fd(ev, mem_ctx, fd, flags,
 690                                    handler, private_data,
 691                                    handler_name, location);
 692         if (!fde) return NULL;
 693
 694         talloc_set_destructor(fde, epoll_event_fd_destructor);
 695
 696         epoll_ev->panic_state = &panic_triggered;
 697         epoll_check_reopen(epoll_ev);
 698         if (panic_triggered) {
 699                 return fde;
 700         }
 701         epoll_ev->panic_state = NULL;
 702
 703         epoll_update_event(epoll_ev, fde);
 704
 705         return fde;
 706 }
 707
 708 /*
 709   set the fd event flags
 710 */
 711 static void epoll_event_set_fd_flags(struct tevent_fd *fde, uint16_t flags)
 712 {
 713         struct tevent_context *ev;
 714         struct epoll_event_context *epoll_ev;
 715         bool panic_triggered = false;
 716
 717         if (fde->flags == flags) return;
 718
 719         ev = fde->event_ctx;
 720         epoll_ev = talloc_get_type(ev->additional_data, struct epoll_event_context);
 721
 722         fde->flags = flags;
 723
 724         epoll_ev->panic_state = &panic_triggered;
 725         epoll_check_reopen(epoll_ev);
 726         if (panic_triggered) {
 727                 return;
 728         }
 729         epoll_ev->panic_state = NULL;
 730
 731         epoll_update_event(epoll_ev, fde);
 732 }
 733
 734 /*
 735   do a single event loop using the events defined in ev
 736 */
 737 static int epoll_event_loop_once(struct tevent_context *ev, const char *location)
 738 {
 739         struct epoll_event_context *epoll_ev = talloc_get_type(ev->additional_data,
 740                                                            struct epoll_event_context);
 741         struct timeval tval;
 742         bool panic_triggered = false;
 743
 744         if (ev->signal_events &&
 745             tevent_common_check_signal(ev)) {
 746                 return 0;
 747         }
 748
 749         if (ev->immediate_events &&
 750             tevent_common_loop_immediate(ev)) {
 751                 return 0;
 752         }
 753
 754         tval = tevent_common_loop_timer_delay(ev);
 755         if (tevent_timeval_is_zero(&tval)) {
 756                 return 0;
 757         }
 758
 759         epoll_ev->panic_state = &panic_triggered;
 760         epoll_ev->panic_force_replay = true;
 761         epoll_check_reopen(epoll_ev);
 762         if (panic_triggered) {
 763                 errno = EINVAL;
 764                 return -1;
 765         }
 766         epoll_ev->panic_force_replay = false;
 767         epoll_ev->panic_state = NULL;
 768
 769         return epoll_event_loop(epoll_ev, &tval);
 770 }
 771
 772 static const struct tevent_ops epoll_event_ops = {
 773         .context_init           = epoll_event_context_init,
 774         .add_fd                 = epoll_event_add_fd,
 775         .set_fd_close_fn        = tevent_common_fd_set_close_fn,
 776         .get_fd_flags           = tevent_common_fd_get_flags,
 777         .set_fd_flags           = epoll_event_set_fd_flags,
 778         .add_timer              = tevent_common_add_timer,
 779         .schedule_immediate     = tevent_common_schedule_immediate,
 780         .add_signal             = tevent_common_add_signal,
 781         .loop_once              = epoll_event_loop_once,
 782         .loop_wait              = tevent_common_loop_wait,
 783 };
 784
 785 _PRIVATE_ bool tevent_epoll_init(void)
 786 {
 787         return tevent_register_backend("epoll", &epoll_event_ops);
 788 }