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      2005-2009
   8
   9      ** NOTE! The following LGPL license applies to the tevent
  10      ** library. This does NOT imply that all of Samba is released
  11      ** under the LGPL
  12
  13    This library is free software; you can redistribute it and/or
  14    modify it under the terms of the GNU Lesser General Public
  15    License as published by the Free Software Foundation; either
  16    version 3 of the License, or (at your option) any later version.
  17
  18    This library is distributed in the hope that it will be useful,
  19    but WITHOUT ANY WARRANTY; without even the implied warranty of
  20    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  21    Lesser General Public License for more details.
  22
  23    You should have received a copy of the GNU Lesser General Public
  24    License along with this library; if not, see <http://www.gnu.org/licenses/>.
  25 */
  26
  27 #include "replace.h"
  28 #include "system/filesys.h"
  29 #include "system/select.h"
  30 #include "tevent.h"
  31 #include "tevent_internal.h"
  32 #include "tevent_util.h"
  33
  34 struct epoll_event_context {
  35         /* a pointer back to the generic event_context */
  36         struct tevent_context *ev;
  37
  38         /* when using epoll this is the handle from epoll_create */
  39         int epoll_fd;
  40
  41         pid_t pid;
  42
  43         bool panic_force_replay;
  44         bool *panic_state;
  45         bool (*panic_fallback)(struct tevent_context *ev, bool replay);
  46 };
  47
  48 #define EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT      (1<<0)
  49 #define EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR   (1<<1)
  50 #define EPOLL_ADDITIONAL_FD_FLAG_GOT_ERROR      (1<<2)
  51
  52 #ifdef TEST_PANIC_FALLBACK
  53
  54 static int epoll_create_panic_fallback(struct epoll_event_context *epoll_ev,
  55                                        int size)
  56 {
  57         if (epoll_ev->panic_fallback == NULL) {
  58                 return epoll_create(size);
  59         }
  60
  61         /* 50% of the time, fail... */
  62         if ((random() % 2) == 0) {
  63                 errno = EINVAL;
  64                 return -1;
  65         }
  66
  67         return epoll_create(size);
  68 }
  69
  70 static int epoll_ctl_panic_fallback(struct epoll_event_context *epoll_ev,
  71                                     int epfd, int op, int fd,
  72                                     struct epoll_event *event)
  73 {
  74         if (epoll_ev->panic_fallback == NULL) {
  75                 return epoll_ctl(epfd, op, fd, event);
  76         }
  77
  78         /* 50% of the time, fail... */
  79         if ((random() % 2) == 0) {
  80                 errno = EINVAL;
  81                 return -1;
  82         }
  83
  84         return epoll_ctl(epfd, op, fd, event);
  85 }
  86
  87 static int epoll_wait_panic_fallback(struct epoll_event_context *epoll_ev,
  88                                      int epfd,
  89                                      struct epoll_event *events,
  90                                      int maxevents,
  91                                      int timeout)
  92 {
  93         if (epoll_ev->panic_fallback == NULL) {
  94                 return epoll_wait(epfd, events, maxevents, timeout);
  95         }
  96
  97         /* 50% of the time, fail... */
  98         if ((random() % 2) == 0) {
  99                 errno = EINVAL;
 100                 return -1;
 101         }
 102
 103         return epoll_wait(epfd, events, maxevents, timeout);
 104 }
 105
 106 #define epoll_create(_size) \
 107         epoll_create_panic_fallback(epoll_ev, _size)
 108 #define epoll_ctl(_epfd, _op, _fd, _event) \
 109         epoll_ctl_panic_fallback(epoll_ev,_epfd, _op, _fd, _event)
 110 #define epoll_wait(_epfd, _events, _maxevents, _timeout) \
 111         epoll_wait_panic_fallback(epoll_ev, _epfd, _events, _maxevents, _timeout)
 112 #endif
 113
 114 /*
 115   called to set the panic fallback function.
 116 */
 117 _PRIVATE_ bool tevent_epoll_set_panic_fallback(struct tevent_context *ev,
 118                                 bool (*panic_fallback)(struct tevent_context *ev,
 119                                                        bool replay))
 120 {
 121         struct epoll_event_context *epoll_ev;
 122
 123         if (ev->additional_data == NULL) {
 124                 return false;
 125         }
 126
 127         epoll_ev = talloc_get_type(ev->additional_data,
 128                                 struct epoll_event_context);
 129         if (epoll_ev == NULL) {
 130                 return false;
 131         }
 132         epoll_ev->panic_fallback = panic_fallback;
 133         return true;
 134 }
 135
 136 /*
 137   called when a epoll call fails
 138 */
 139 static void epoll_panic(struct epoll_event_context *epoll_ev,
 140                         const char *reason, bool replay)
 141 {
 142         struct tevent_context *ev = epoll_ev->ev;
 143         bool (*panic_fallback)(struct tevent_context *ev, bool replay);
 144
 145         panic_fallback = epoll_ev->panic_fallback;
 146
 147         if (epoll_ev->panic_state != NULL) {
 148                 *epoll_ev->panic_state = true;
 149         }
 150
 151         if (epoll_ev->panic_force_replay) {
 152                 replay = true;
 153         }
 154
 155         TALLOC_FREE(ev->additional_data);
 156
 157         if (panic_fallback == NULL) {
 158                 tevent_debug(ev, TEVENT_DEBUG_FATAL,
 159                         "%s (%s) replay[%u] - calling abort()\n",
 160                         reason, strerror(errno), (unsigned)replay);
 161                 abort();
 162         }
 163
 164         tevent_debug(ev, TEVENT_DEBUG_WARNING,
 165                      "%s (%s) replay[%u] - calling panic_fallback\n",
 166                      reason, strerror(errno), (unsigned)replay);
 167
 168         if (!panic_fallback(ev, replay)) {
 169                 /* Fallback failed. */
 170                 tevent_debug(ev, TEVENT_DEBUG_FATAL,
 171                         "%s (%s) replay[%u] - calling abort()\n",
 172                         reason, strerror(errno), (unsigned)replay);
 173                 abort();
 174         }
 175 }
 176
 177 /*
 178   map from TEVENT_FD_* to EPOLLIN/EPOLLOUT
 179 */
 180 static uint32_t epoll_map_flags(uint16_t flags)
 181 {
 182         uint32_t ret = 0;
 183         if (flags & TEVENT_FD_READ) ret |= (EPOLLIN | EPOLLERR | EPOLLHUP);
 184         if (flags & TEVENT_FD_WRITE) ret |= (EPOLLOUT | EPOLLERR | EPOLLHUP);
 185         return ret;
 186 }
 187
 188 /*
 189  free the epoll fd
 190 */
 191 static int epoll_ctx_destructor(struct epoll_event_context *epoll_ev)
 192 {
 193         close(epoll_ev->epoll_fd);
 194         epoll_ev->epoll_fd = -1;
 195         return 0;
 196 }
 197
 198 /*
 199  init the epoll fd
 200 */
 201 static int epoll_init_ctx(struct epoll_event_context *epoll_ev)
 202 {
 203         epoll_ev->epoll_fd = epoll_create(64);
 204         if (epoll_ev->epoll_fd == -1) {
 205                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_FATAL,
 206                              "Failed to create epoll handle.\n");
 207                 return -1;
 208         }
 209
 210         if (!ev_set_close_on_exec(epoll_ev->epoll_fd)) {
 211                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_WARNING,
 212                              "Failed to set close-on-exec, file descriptor may be leaked to children.\n");
 213         }
 214
 215         epoll_ev->pid = getpid();
 216         talloc_set_destructor(epoll_ev, epoll_ctx_destructor);
 217
 218         return 0;
 219 }
 220
 221 static void epoll_update_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde);
 222
 223 /*
 224   reopen the epoll handle when our pid changes
 225   see http://junkcode.samba.org/ftp/unpacked/junkcode/epoll_fork.c for an
 226   demonstration of why this is needed
 227  */
 228 static void epoll_check_reopen(struct epoll_event_context *epoll_ev)
 229 {
 230         struct tevent_fd *fde;
 231         bool *caller_panic_state = epoll_ev->panic_state;
 232         bool panic_triggered = false;
 233
 234         if (epoll_ev->pid == getpid()) {
 235                 return;
 236         }
 237
 238         close(epoll_ev->epoll_fd);
 239         epoll_ev->epoll_fd = epoll_create(64);
 240         if (epoll_ev->epoll_fd == -1) {
 241                 epoll_panic(epoll_ev, "epoll_create() failed", false);
 242                 return;
 243         }
 244
 245         if (!ev_set_close_on_exec(epoll_ev->epoll_fd)) {
 246                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_WARNING,
 247                              "Failed to set close-on-exec, file descriptor may be leaked to children.\n");
 248         }
 249
 250         epoll_ev->pid = getpid();
 251         epoll_ev->panic_state = &panic_triggered;
 252         for (fde=epoll_ev->ev->fd_events;fde;fde=fde->next) {
 253                 fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 254                 epoll_update_event(epoll_ev, fde);
 255
 256                 if (panic_triggered) {
 257                         if (caller_panic_state != NULL) {
 258                                 *caller_panic_state = true;
 259                         }
 260                         return;
 261                 }
 262         }
 263         epoll_ev->panic_state = NULL;
 264 }
 265
 266 /*
 267  add the epoll event to the given fd_event
 268 */
 269 static void epoll_add_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde)
 270 {
 271         struct epoll_event event;
 272         int ret;
 273
 274         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 275
 276         /* if we don't want events yet, don't add an epoll_event */
 277         if (fde->flags == 0) return;
 278
 279         ZERO_STRUCT(event);
 280         event.events = epoll_map_flags(fde->flags);
 281         event.data.ptr = fde;
 282         ret = epoll_ctl(epoll_ev->epoll_fd, EPOLL_CTL_ADD, fde->fd, &event);
 283         if (ret != 0) {
 284                 epoll_panic(epoll_ev, "EPOLL_CTL_ADD failed", false);
 285                 return;
 286         }
 287         fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 288
 289         /* only if we want to read we want to tell the event handler about errors */
 290         if (fde->flags & TEVENT_FD_READ) {
 291                 fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 292         }
 293 }
 294
 295 /*
 296  delete the epoll event for given fd_event
 297 */
 298 static void epoll_del_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde)
 299 {
 300         struct epoll_event event;
 301         int ret;
 302
 303         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 304
 305         /* if there's no epoll_event, we don't need to delete it */
 306         if (!(fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT)) return;
 307
 308         ZERO_STRUCT(event);
 309         ret = epoll_ctl(epoll_ev->epoll_fd, EPOLL_CTL_DEL, fde->fd, &event);
 310         if (ret != 0 && errno == ENOENT) {
 311                 /*
 312                  * This can happen after a epoll_check_reopen
 313                  * within epoll_event_fd_destructor.
 314                  */
 315                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_TRACE,
 316                              "EPOLL_CTL_DEL ignoring ENOENT for fd[%d]\n",
 317                              fde->fd);
 318         } else if (ret != 0) {
 319                 epoll_panic(epoll_ev, "EPOLL_CTL_DEL failed", false);
 320                 return;
 321         }
 322         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 323 }
 324
 325 /*
 326  change the epoll event to the given fd_event
 327 */
 328 static void epoll_mod_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde)
 329 {
 330         struct epoll_event event;
 331         int ret;
 332
 333         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 334
 335         ZERO_STRUCT(event);
 336         event.events = epoll_map_flags(fde->flags);
 337         event.data.ptr = fde;
 338         ret = epoll_ctl(epoll_ev->epoll_fd, EPOLL_CTL_MOD, fde->fd, &event);
 339         if (ret != 0) {
 340                 epoll_panic(epoll_ev, "EPOLL_CTL_MOD failed", false);
 341                 return;
 342         }
 343
 344         /* only if we want to read we want to tell the event handler about errors */
 345         if (fde->flags & TEVENT_FD_READ) {
 346                 fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 347         }
 348 }
 349
 350 static void epoll_update_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde)
 351 {
 352         bool got_error = (fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_GOT_ERROR);
 353         bool want_read = (fde->flags & TEVENT_FD_READ);
 354         bool want_write= (fde->flags & TEVENT_FD_WRITE);
 355
 356         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 357
 358         /* there's already an event */
 359         if (fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT) {
 360                 if (want_read || (want_write && !got_error)) {
 361                         epoll_mod_event(epoll_ev, fde);
 362                         return;
 363                 }
 364                 /*
 365                  * if we want to match the select behavior, we need to remove the epoll_event
 366                  * when the caller isn't interested in events.
 367                  *
 368                  * this is because epoll reports EPOLLERR and EPOLLHUP, even without asking for them
 369                  */
 370                 epoll_del_event(epoll_ev, fde);
 371                 return;
 372         }
 373
 374         /* there's no epoll_event attached to the fde */
 375         if (want_read || (want_write && !got_error)) {
 376                 epoll_add_event(epoll_ev, fde);
 377                 return;
 378         }
 379 }
 380
 381 /*
 382   event loop handling using epoll
 383 */
 384 static int epoll_event_loop(struct epoll_event_context *epoll_ev, struct timeval *tvalp)
 385 {
 386         int ret, i;
 387 #define MAXEVENTS 1
 388         struct epoll_event events[MAXEVENTS];
 389         int timeout = -1;
 390         int wait_errno;
 391
 392         if (tvalp) {
 393                 /* it's better to trigger timed events a bit later than too early */
 394                 timeout = ((tvalp->tv_usec+999) / 1000) + (tvalp->tv_sec*1000);
 395         }
 396
 397         if (epoll_ev->ev->signal_events &&
 398             tevent_common_check_signal(epoll_ev->ev)) {
 399                 return 0;
 400         }
 401
 402         tevent_trace_point_callback(epoll_ev->ev, TEVENT_TRACE_BEFORE_WAIT);
 403         ret = epoll_wait(epoll_ev->epoll_fd, events, MAXEVENTS, timeout);
 404         wait_errno = errno;
 405         tevent_trace_point_callback(epoll_ev->ev, TEVENT_TRACE_AFTER_WAIT);
 406
 407         if (ret == -1 && wait_errno == EINTR && epoll_ev->ev->signal_events) {
 408                 if (tevent_common_check_signal(epoll_ev->ev)) {
 409                         return 0;
 410                 }
 411         }
 412
 413         if (ret == -1 && wait_errno != EINTR) {
 414                 epoll_panic(epoll_ev, "epoll_wait() failed", true);
 415                 return -1;
 416         }
 417
 418         if (ret == 0 && tvalp) {
 419                 /* we don't care about a possible delay here */
 420                 tevent_common_loop_timer_delay(epoll_ev->ev);
 421                 return 0;
 422         }
 423
 424         for (i=0;i<ret;i++) {
 425                 struct tevent_fd *fde = talloc_get_type(events[i].data.ptr,
 426                                                        struct tevent_fd);
 427                 uint16_t flags = 0;
 428
 429                 if (fde == NULL) {
 430                         epoll_panic(epoll_ev, "epoll_wait() gave bad data", true);
 431                         return -1;
 432                 }
 433                 if (events[i].events & (EPOLLHUP|EPOLLERR)) {
 434                         fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_GOT_ERROR;
 435                         /*
 436                          * if we only wait for TEVENT_FD_WRITE, we should not tell the
 437                          * event handler about it, and remove the epoll_event,
 438                          * as we only report errors when waiting for read events,
 439                          * to match the select() behavior
 440                          */
 441                         if (!(fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR)) {
 442                                 epoll_update_event(epoll_ev, fde);
 443                                 continue;
 444                         }
 445                         flags |= TEVENT_FD_READ;
 446                 }
 447                 if (events[i].events & EPOLLIN) flags |= TEVENT_FD_READ;
 448                 if (events[i].events & EPOLLOUT) flags |= TEVENT_FD_WRITE;
 449                 if (flags) {
 450                         fde->handler(epoll_ev->ev, fde, flags, fde->private_data);
 451                         break;
 452                 }
 453         }
 454
 455         return 0;
 456 }
 457
 458 /*
 459   create a epoll_event_context structure.
 460 */
 461 static int epoll_event_context_init(struct tevent_context *ev)
 462 {
 463         int ret;
 464         struct epoll_event_context *epoll_ev;
 465
 466         /*
 467          * We might be called during tevent_re_initialise()
 468          * which means we need to free our old additional_data.
 469          */
 470         TALLOC_FREE(ev->additional_data);
 471
 472         epoll_ev = talloc_zero(ev, struct epoll_event_context);
 473         if (!epoll_ev) return -1;
 474         epoll_ev->ev = ev;
 475         epoll_ev->epoll_fd = -1;
 476
 477         ret = epoll_init_ctx(epoll_ev);
 478         if (ret != 0) {
 479                 talloc_free(epoll_ev);
 480                 return ret;
 481         }
 482
 483         ev->additional_data = epoll_ev;
 484         return 0;
 485 }
 486
 487 /*
 488   destroy an fd_event
 489 */
 490 static int epoll_event_fd_destructor(struct tevent_fd *fde)
 491 {
 492         struct tevent_context *ev = fde->event_ctx;
 493         struct epoll_event_context *epoll_ev = NULL;
 494         bool panic_triggered = false;
 495         int flags = fde->flags;
 496
 497         if (ev == NULL) {
 498                 return tevent_common_fd_destructor(fde);
 499         }
 500
 501         epoll_ev = talloc_get_type_abort(ev->additional_data,
 502                                          struct epoll_event_context);
 503
 504         /*
 505          * we must remove the event from the list
 506          * otherwise a panic fallback handler may
 507          * reuse invalid memory
 508          */
 509         DLIST_REMOVE(ev->fd_events, fde);
 510
 511         epoll_ev->panic_state = &panic_triggered;
 512         epoll_check_reopen(epoll_ev);
 513         if (panic_triggered) {
 514                 return tevent_common_fd_destructor(fde);
 515         }
 516
 517         fde->flags = 0;
 518         epoll_update_event(epoll_ev, fde);
 519         fde->flags = flags;
 520         if (panic_triggered) {
 521                 return tevent_common_fd_destructor(fde);
 522         }
 523         epoll_ev->panic_state = NULL;
 524
 525         return tevent_common_fd_destructor(fde);
 526 }
 527
 528 /*
 529   add a fd based event
 530   return NULL on failure (memory allocation error)
 531 */
 532 static struct tevent_fd *epoll_event_add_fd(struct tevent_context *ev, TALLOC_CTX *mem_ctx,
 533                                             int fd, uint16_t flags,
 534                                             tevent_fd_handler_t handler,
 535                                             void *private_data,
 536                                             const char *handler_name,
 537                                             const char *location)
 538 {
 539         struct epoll_event_context *epoll_ev = talloc_get_type(ev->additional_data,
 540                                                            struct epoll_event_context);
 541         struct tevent_fd *fde;
 542         bool panic_triggered = false;
 543
 544         fde = tevent_common_add_fd(ev, mem_ctx, fd, flags,
 545                                    handler, private_data,
 546                                    handler_name, location);
 547         if (!fde) return NULL;
 548
 549         talloc_set_destructor(fde, epoll_event_fd_destructor);
 550
 551         epoll_ev->panic_state = &panic_triggered;
 552         epoll_check_reopen(epoll_ev);
 553         if (panic_triggered) {
 554                 return fde;
 555         }
 556         epoll_ev->panic_state = NULL;
 557
 558         epoll_update_event(epoll_ev, fde);
 559
 560         return fde;
 561 }
 562
 563 /*
 564   set the fd event flags
 565 */
 566 static void epoll_event_set_fd_flags(struct tevent_fd *fde, uint16_t flags)
 567 {
 568         struct tevent_context *ev;
 569         struct epoll_event_context *epoll_ev;
 570         bool panic_triggered = false;
 571
 572         if (fde->flags == flags) return;
 573
 574         ev = fde->event_ctx;
 575         epoll_ev = talloc_get_type(ev->additional_data, struct epoll_event_context);
 576
 577         fde->flags = flags;
 578
 579         epoll_ev->panic_state = &panic_triggered;
 580         epoll_check_reopen(epoll_ev);
 581         if (panic_triggered) {
 582                 return;
 583         }
 584         epoll_ev->panic_state = NULL;
 585
 586         epoll_update_event(epoll_ev, fde);
 587 }
 588
 589 /*
 590   do a single event loop using the events defined in ev
 591 */
 592 static int epoll_event_loop_once(struct tevent_context *ev, const char *location)
 593 {
 594         struct epoll_event_context *epoll_ev = talloc_get_type(ev->additional_data,
 595                                                            struct epoll_event_context);
 596         struct timeval tval;
 597         bool panic_triggered = false;
 598
 599         if (ev->signal_events &&
 600             tevent_common_check_signal(ev)) {
 601                 return 0;
 602         }
 603
 604         if (ev->immediate_events &&
 605             tevent_common_loop_immediate(ev)) {
 606                 return 0;
 607         }
 608
 609         tval = tevent_common_loop_timer_delay(ev);
 610         if (tevent_timeval_is_zero(&tval)) {
 611                 return 0;
 612         }
 613
 614         epoll_ev->panic_state = &panic_triggered;
 615         epoll_ev->panic_force_replay = true;
 616         epoll_check_reopen(epoll_ev);
 617         if (panic_triggered) {
 618                 errno = EINVAL;
 619                 return -1;
 620         }
 621         epoll_ev->panic_force_replay = false;
 622         epoll_ev->panic_state = NULL;
 623
 624         return epoll_event_loop(epoll_ev, &tval);
 625 }
 626
 627 static const struct tevent_ops epoll_event_ops = {
 628         .context_init           = epoll_event_context_init,
 629         .add_fd                 = epoll_event_add_fd,
 630         .set_fd_close_fn        = tevent_common_fd_set_close_fn,
 631         .get_fd_flags           = tevent_common_fd_get_flags,
 632         .set_fd_flags           = epoll_event_set_fd_flags,
 633         .add_timer              = tevent_common_add_timer,
 634         .schedule_immediate     = tevent_common_schedule_immediate,
 635         .add_signal             = tevent_common_add_signal,
 636         .loop_once              = epoll_event_loop_once,
 637         .loop_wait              = tevent_common_loop_wait,
 638 };
 639
 640 _PRIVATE_ bool tevent_epoll_init(void)
 641 {
 642         return tevent_register_backend("epoll", &epoll_event_ops);
 643 }