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         struct tevent_fd *mpx_fde = NULL;
 434
 435         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 436         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 437
 438         if (fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_HAS_MPX) {
 439                 /*
 440                  * This is a multiplexed fde, we need to modify both events.
 441                  */
 442                 mpx_fde = talloc_get_type_abort(fde->additional_data,
 443                                                 struct tevent_fd);
 444
 445                 mpx_fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 446                 mpx_fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 447         }
 448
 449         ZERO_STRUCT(event);
 450         ret = epoll_ctl(epoll_ev->epoll_fd, EPOLL_CTL_DEL, fde->fd, &event);
 451         if (ret != 0 && errno == ENOENT) {
 452                 /*
 453                  * This can happen after a epoll_check_reopen
 454                  * within epoll_event_fd_destructor.
 455                  */
 456                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_TRACE,
 457                              "EPOLL_CTL_DEL ignoring ENOENT for fd[%d]\n",
 458                              fde->fd);
 459                 return;
 460         } else if (ret != 0 && errno == EBADF) {
 461                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_WARNING,
 462                              "EPOLL_CTL_DEL EBADF for "
 463                              "fde[%p] mpx_fde[%p] fd[%d] - disabling\n",
 464                              fde, mpx_fde, fde->fd);
 465                 DLIST_REMOVE(epoll_ev->ev->fd_events, fde);
 466                 fde->event_ctx = NULL;
 467                 if (mpx_fde != NULL) {
 468                         DLIST_REMOVE(epoll_ev->ev->fd_events, mpx_fde);
 469                         mpx_fde->event_ctx = NULL;
 470                 }
 471                 return;
 472         } else if (ret != 0) {
 473                 epoll_panic(epoll_ev, "EPOLL_CTL_DEL failed", false);
 474                 return;
 475         }
 476 }
 477
 478 /*
 479  change the epoll event to the given fd_event
 480 */
 481 static void epoll_mod_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde)
 482 {
 483         struct tevent_fd *mpx_fde = NULL;
 484         struct epoll_event event;
 485         int ret;
 486
 487         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 488         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 489
 490         if (fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_HAS_MPX) {
 491                 /*
 492                  * This is a multiplexed fde, we need to include both
 493                  * flags in the modified event.
 494                  */
 495                 mpx_fde = talloc_get_type_abort(fde->additional_data,
 496                                                 struct tevent_fd);
 497
 498                 mpx_fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 499                 mpx_fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 500         }
 501
 502         ZERO_STRUCT(event);
 503         event.events = epoll_map_flags(fde->flags);
 504         if (mpx_fde != NULL) {
 505                 event.events |= epoll_map_flags(mpx_fde->flags);
 506         }
 507         event.data.ptr = fde;
 508         ret = epoll_ctl(epoll_ev->epoll_fd, EPOLL_CTL_MOD, fde->fd, &event);
 509         if (ret != 0 && errno == EBADF) {
 510                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_ERROR,
 511                              "EPOLL_CTL_MOD EBADF for "
 512                              "fde[%p] mpx_fde[%p] fd[%d] - disabling\n",
 513                              fde, mpx_fde, fde->fd);
 514                 DLIST_REMOVE(epoll_ev->ev->fd_events, fde);
 515                 fde->event_ctx = NULL;
 516                 if (mpx_fde != NULL) {
 517                         DLIST_REMOVE(epoll_ev->ev->fd_events, mpx_fde);
 518                         mpx_fde->event_ctx = NULL;
 519                 }
 520                 return;
 521         } else if (ret != 0) {
 522                 epoll_panic(epoll_ev, "EPOLL_CTL_MOD failed", false);
 523                 return;
 524         }
 525
 526         fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 527         /* only if we want to read we want to tell the event handler about errors */
 528         if (fde->flags & TEVENT_FD_READ) {
 529                 fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 530         }
 531
 532         if (mpx_fde == NULL) {
 533                 return;
 534         }
 535
 536         mpx_fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 537         /* only if we want to read we want to tell the event handler about errors */
 538         if (mpx_fde->flags & TEVENT_FD_READ) {
 539                 mpx_fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 540         }
 541 }
 542
 543 static void epoll_update_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde)
 544 {
 545         bool got_error = (fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_GOT_ERROR);
 546         bool want_read = (fde->flags & TEVENT_FD_READ);
 547         bool want_write= (fde->flags & TEVENT_FD_WRITE);
 548         struct tevent_fd *mpx_fde = NULL;
 549
 550         if (fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_HAS_MPX) {
 551                 /*
 552                  * work out what the multiplexed fde wants.
 553                  */
 554                 mpx_fde = talloc_get_type_abort(fde->additional_data,
 555                                                 struct tevent_fd);
 556
 557                 if (mpx_fde->flags & TEVENT_FD_READ) {
 558                         want_read = true;
 559                 }
 560
 561                 if (mpx_fde->flags & TEVENT_FD_WRITE) {
 562                         want_write = true;
 563                 }
 564         }
 565
 566         /* there's already an event */
 567         if (fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT) {
 568                 if (want_read || (want_write && !got_error)) {
 569                         epoll_mod_event(epoll_ev, fde);
 570                         return;
 571                 }
 572                 /*
 573                  * if we want to match the select behavior, we need to remove the epoll_event
 574                  * when the caller isn't interested in events.
 575                  *
 576                  * this is because epoll reports EPOLLERR and EPOLLHUP, even without asking for them
 577                  */
 578                 epoll_del_event(epoll_ev, fde);
 579                 return;
 580         }
 581
 582         /* there's no epoll_event attached to the fde */
 583         if (want_read || (want_write && !got_error)) {
 584                 epoll_add_event(epoll_ev, fde);
 585                 return;
 586         }
 587 }
 588
 589 /*
 590   Cope with epoll returning EPOLLHUP|EPOLLERR on an event.
 591   Return true if there's nothing else to do, false if
 592   this event needs further handling.
 593 */
 594 static bool epoll_handle_hup_or_err(struct epoll_event_context *epoll_ev,
 595                                 struct tevent_fd *fde)
 596 {
 597         if (fde == NULL) {
 598                 /* Nothing to do if no event. */
 599                 return true;
 600         }
 601
 602         fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_GOT_ERROR;
 603         /*
 604          * if we only wait for TEVENT_FD_WRITE, we should not tell the
 605          * event handler about it, and remove the epoll_event,
 606          * as we only report errors when waiting for read events,
 607          * to match the select() behavior
 608          */
 609         if (!(fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR)) {
 610                 /*
 611                  * Do the same as the poll backend and
 612                  * remove the writeable flag.
 613                  */
 614                 fde->flags &= ~TEVENT_FD_WRITE;
 615                 return true;
 616         }
 617         /* This has TEVENT_FD_READ set, we're not finished. */
 618         return false;
 619 }
 620
 621 /*
 622   event loop handling using epoll
 623 */
 624 static int epoll_event_loop(struct epoll_event_context *epoll_ev, struct timeval *tvalp)
 625 {
 626         int ret, i;
 627 #define MAXEVENTS 1
 628         struct epoll_event events[MAXEVENTS];
 629         int timeout = -1;
 630         int wait_errno;
 631
 632         if (tvalp) {
 633                 /* it's better to trigger timed events a bit later than too early */
 634                 timeout = ((tvalp->tv_usec+999) / 1000) + (tvalp->tv_sec*1000);
 635         }
 636
 637         if (epoll_ev->ev->signal_events &&
 638             tevent_common_check_signal(epoll_ev->ev)) {
 639                 return 0;
 640         }
 641
 642         tevent_trace_point_callback(epoll_ev->ev, TEVENT_TRACE_BEFORE_WAIT);
 643         ret = epoll_wait(epoll_ev->epoll_fd, events, MAXEVENTS, timeout);
 644         wait_errno = errno;
 645         tevent_trace_point_callback(epoll_ev->ev, TEVENT_TRACE_AFTER_WAIT);
 646
 647         if (ret == -1 && wait_errno == EINTR && epoll_ev->ev->signal_events) {
 648                 if (tevent_common_check_signal(epoll_ev->ev)) {
 649                         return 0;
 650                 }
 651         }
 652
 653         if (ret == -1 && wait_errno != EINTR) {
 654                 epoll_panic(epoll_ev, "epoll_wait() failed", true);
 655                 return -1;
 656         }
 657
 658         if (ret == 0 && tvalp) {
 659                 /* we don't care about a possible delay here */
 660                 tevent_common_loop_timer_delay(epoll_ev->ev);
 661                 return 0;
 662         }
 663
 664         for (i=0;i<ret;i++) {
 665                 struct tevent_fd *fde = talloc_get_type(events[i].data.ptr,
 666                                                        struct tevent_fd);
 667                 uint16_t flags = 0;
 668                 struct tevent_fd *mpx_fde = NULL;
 669
 670                 if (fde == NULL) {
 671                         epoll_panic(epoll_ev, "epoll_wait() gave bad data", true);
 672                         return -1;
 673                 }
 674                 if (fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_HAS_MPX) {
 675                         /*
 676                          * Save off the multiplexed event in case we need
 677                          * to use it to call the handler function.
 678                          */
 679                         mpx_fde = talloc_get_type_abort(fde->additional_data,
 680                                                         struct tevent_fd);
 681                 }
 682                 if (events[i].events & (EPOLLHUP|EPOLLERR)) {
 683                         bool handled_fde = epoll_handle_hup_or_err(epoll_ev, fde);
 684                         bool handled_mpx = epoll_handle_hup_or_err(epoll_ev, mpx_fde);
 685
 686                         if (handled_fde && handled_mpx) {
 687                                 epoll_update_event(epoll_ev, fde);
 688                                 continue;
 689                         }
 690
 691                         if (!handled_mpx) {
 692                                 /*
 693                                  * If the mpx event was the one that needs
 694                                  * further handling, it's the TEVENT_FD_READ
 695                                  * event so switch over and call that handler.
 696                                  */
 697                                 fde = mpx_fde;
 698                                 mpx_fde = NULL;
 699                         }
 700                         flags |= TEVENT_FD_READ;
 701                 }
 702                 if (events[i].events & EPOLLIN) flags |= TEVENT_FD_READ;
 703                 if (events[i].events & EPOLLOUT) flags |= TEVENT_FD_WRITE;
 704
 705                 if (flags & TEVENT_FD_WRITE) {
 706                         if (fde->flags & TEVENT_FD_WRITE) {
 707                                 mpx_fde = NULL;
 708                         }
 709                         if (mpx_fde && mpx_fde->flags & TEVENT_FD_WRITE) {
 710                                 fde = mpx_fde;
 711                                 mpx_fde = NULL;
 712                         }
 713                 }
 714
 715                 if (mpx_fde) {
 716                         /* Ensure we got the right fde. */
 717                         if ((flags & fde->flags) == 0) {
 718                                 fde = mpx_fde;
 719                                 mpx_fde = NULL;
 720                         }
 721                 }
 722
 723                 /*
 724                  * make sure we only pass the flags
 725                  * the handler is expecting.
 726                  */
 727                 flags &= fde->flags;
 728                 if (flags) {
 729                         fde->handler(epoll_ev->ev, fde, flags, fde->private_data);
 730                         break;
 731                 }
 732         }
 733
 734         return 0;
 735 }
 736
 737 /*
 738   create a epoll_event_context structure.
 739 */
 740 static int epoll_event_context_init(struct tevent_context *ev)
 741 {
 742         int ret;
 743         struct epoll_event_context *epoll_ev;
 744
 745         /*
 746          * We might be called during tevent_re_initialise()
 747          * which means we need to free our old additional_data.
 748          */
 749         TALLOC_FREE(ev->additional_data);
 750
 751         epoll_ev = talloc_zero(ev, struct epoll_event_context);
 752         if (!epoll_ev) return -1;
 753         epoll_ev->ev = ev;
 754         epoll_ev->epoll_fd = -1;
 755
 756         ret = epoll_init_ctx(epoll_ev);
 757         if (ret != 0) {
 758                 talloc_free(epoll_ev);
 759                 return ret;
 760         }
 761
 762         ev->additional_data = epoll_ev;
 763         return 0;
 764 }
 765
 766 /*
 767   destroy an fd_event
 768 */
 769 static int epoll_event_fd_destructor(struct tevent_fd *fde)
 770 {
 771         struct tevent_context *ev = fde->event_ctx;
 772         struct epoll_event_context *epoll_ev = NULL;
 773         bool panic_triggered = false;
 774         struct tevent_fd *mpx_fde = NULL;
 775         int flags = fde->flags;
 776
 777         if (ev == NULL) {
 778                 return tevent_common_fd_destructor(fde);
 779         }
 780
 781         epoll_ev = talloc_get_type_abort(ev->additional_data,
 782                                          struct epoll_event_context);
 783
 784         /*
 785          * we must remove the event from the list
 786          * otherwise a panic fallback handler may
 787          * reuse invalid memory
 788          */
 789         DLIST_REMOVE(ev->fd_events, fde);
 790
 791         if (fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_HAS_MPX) {
 792                 mpx_fde = talloc_get_type_abort(fde->additional_data,
 793                                                 struct tevent_fd);
 794
 795                 fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_HAS_MPX;
 796                 mpx_fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_HAS_MPX;
 797
 798                 fde->additional_data = NULL;
 799                 mpx_fde->additional_data = NULL;
 800
 801                 fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 802         }
 803
 804         epoll_ev->panic_state = &panic_triggered;
 805         epoll_check_reopen(epoll_ev);
 806         if (panic_triggered) {
 807                 return tevent_common_fd_destructor(fde);
 808         }
 809
 810         if (mpx_fde != NULL) {
 811                 epoll_update_event(epoll_ev, mpx_fde);
 812                 if (panic_triggered) {
 813                         return tevent_common_fd_destructor(fde);
 814                 }
 815         }
 816
 817         fde->flags = 0;
 818         epoll_update_event(epoll_ev, fde);
 819         fde->flags = flags;
 820         if (panic_triggered) {
 821                 return tevent_common_fd_destructor(fde);
 822         }
 823         epoll_ev->panic_state = NULL;
 824
 825         return tevent_common_fd_destructor(fde);
 826 }
 827
 828 /*
 829   add a fd based event
 830   return NULL on failure (memory allocation error)
 831 */
 832 static struct tevent_fd *epoll_event_add_fd(struct tevent_context *ev, TALLOC_CTX *mem_ctx,
 833                                             int fd, uint16_t flags,
 834                                             tevent_fd_handler_t handler,
 835                                             void *private_data,
 836                                             const char *handler_name,
 837                                             const char *location)
 838 {
 839         struct epoll_event_context *epoll_ev = talloc_get_type(ev->additional_data,
 840                                                            struct epoll_event_context);
 841         struct tevent_fd *fde;
 842         bool panic_triggered = false;
 843
 844         fde = tevent_common_add_fd(ev, mem_ctx, fd, flags,
 845                                    handler, private_data,
 846                                    handler_name, location);
 847         if (!fde) return NULL;
 848
 849         talloc_set_destructor(fde, epoll_event_fd_destructor);
 850
 851         epoll_ev->panic_state = &panic_triggered;
 852         epoll_check_reopen(epoll_ev);
 853         if (panic_triggered) {
 854                 return fde;
 855         }
 856         epoll_ev->panic_state = NULL;
 857
 858         epoll_update_event(epoll_ev, fde);
 859
 860         return fde;
 861 }
 862
 863 /*
 864   set the fd event flags
 865 */
 866 static void epoll_event_set_fd_flags(struct tevent_fd *fde, uint16_t flags)
 867 {
 868         struct tevent_context *ev;
 869         struct epoll_event_context *epoll_ev;
 870         bool panic_triggered = false;
 871
 872         if (fde->flags == flags) return;
 873
 874         ev = fde->event_ctx;
 875         epoll_ev = talloc_get_type(ev->additional_data, struct epoll_event_context);
 876
 877         fde->flags = flags;
 878
 879         epoll_ev->panic_state = &panic_triggered;
 880         epoll_check_reopen(epoll_ev);
 881         if (panic_triggered) {
 882                 return;
 883         }
 884         epoll_ev->panic_state = NULL;
 885
 886         epoll_update_event(epoll_ev, fde);
 887 }
 888
 889 /*
 890   do a single event loop using the events defined in ev
 891 */
 892 static int epoll_event_loop_once(struct tevent_context *ev, const char *location)
 893 {
 894         struct epoll_event_context *epoll_ev = talloc_get_type(ev->additional_data,
 895                                                            struct epoll_event_context);
 896         struct timeval tval;
 897         bool panic_triggered = false;
 898
 899         if (ev->signal_events &&
 900             tevent_common_check_signal(ev)) {
 901                 return 0;
 902         }
 903
 904         if (ev->immediate_events &&
 905             tevent_common_loop_immediate(ev)) {
 906                 return 0;
 907         }
 908
 909         tval = tevent_common_loop_timer_delay(ev);
 910         if (tevent_timeval_is_zero(&tval)) {
 911                 return 0;
 912         }
 913
 914         epoll_ev->panic_state = &panic_triggered;
 915         epoll_ev->panic_force_replay = true;
 916         epoll_check_reopen(epoll_ev);
 917         if (panic_triggered) {
 918                 errno = EINVAL;
 919                 return -1;
 920         }
 921         epoll_ev->panic_force_replay = false;
 922         epoll_ev->panic_state = NULL;
 923
 924         return epoll_event_loop(epoll_ev, &tval);
 925 }
 926
 927 static const struct tevent_ops epoll_event_ops = {
 928         .context_init           = epoll_event_context_init,
 929         .add_fd                 = epoll_event_add_fd,
 930         .set_fd_close_fn        = tevent_common_fd_set_close_fn,
 931         .get_fd_flags           = tevent_common_fd_get_flags,
 932         .set_fd_flags           = epoll_event_set_fd_flags,
 933         .add_timer              = tevent_common_add_timer,
 934         .schedule_immediate     = tevent_common_schedule_immediate,
 935         .add_signal             = tevent_common_add_signal,
 936         .loop_once              = epoll_event_loop_once,
 937         .loop_wait              = tevent_common_loop_wait,
 938 };
 939
 940 _PRIVATE_ bool tevent_epoll_init(void)
 941 {
 942         return tevent_register_backend("epoll", &epoll_event_ops);
 943 }