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 #ifdef TEST_PANIC_FALLBACK
  49
  50 static int epoll_create_panic_fallback(struct epoll_event_context *epoll_ev,
  51                                        int size)
  52 {
  53         if (epoll_ev->panic_fallback == NULL) {
  54                 return epoll_create(size);
  55         }
  56
  57         /* 50% of the time, fail... */
  58         if ((random() % 2) == 0) {
  59                 errno = EINVAL;
  60                 return -1;
  61         }
  62
  63         return epoll_create(size);
  64 }
  65
  66 static int epoll_ctl_panic_fallback(struct epoll_event_context *epoll_ev,
  67                                     int epfd, int op, int fd,
  68                                     struct epoll_event *event)
  69 {
  70         if (epoll_ev->panic_fallback == NULL) {
  71                 return epoll_ctl(epfd, op, fd, event);
  72         }
  73
  74         /* 50% of the time, fail... */
  75         if ((random() % 2) == 0) {
  76                 errno = EINVAL;
  77                 return -1;
  78         }
  79
  80         return epoll_ctl(epfd, op, fd, event);
  81 }
  82
  83 static int epoll_wait_panic_fallback(struct epoll_event_context *epoll_ev,
  84                                      int epfd,
  85                                      struct epoll_event *events,
  86                                      int maxevents,
  87                                      int timeout)
  88 {
  89         if (epoll_ev->panic_fallback == NULL) {
  90                 return epoll_wait(epfd, events, maxevents, timeout);
  91         }
  92
  93         /* 50% of the time, fail... */
  94         if ((random() % 2) == 0) {
  95                 errno = EINVAL;
  96                 return -1;
  97         }
  98
  99         return epoll_wait(epfd, events, maxevents, timeout);
 100 }
 101
 102 #define epoll_create(_size) \
 103         epoll_create_panic_fallback(epoll_ev, _size)
 104 #define epoll_ctl(_epfd, _op, _fd, _event) \
 105         epoll_ctl_panic_fallback(epoll_ev,_epfd, _op, _fd, _event)
 106 #define epoll_wait(_epfd, _events, _maxevents, _timeout) \
 107         epoll_wait_panic_fallback(epoll_ev, _epfd, _events, _maxevents, _timeout)
 108 #endif
 109
 110 /*
 111   called to set the panic fallback function.
 112 */
 113 _PRIVATE_ bool tevent_epoll_set_panic_fallback(struct tevent_context *ev,
 114                                 bool (*panic_fallback)(struct tevent_context *ev,
 115                                                        bool replay))
 116 {
 117         struct epoll_event_context *epoll_ev;
 118
 119         if (ev->additional_data == NULL) {
 120                 return false;
 121         }
 122
 123         epoll_ev = talloc_get_type(ev->additional_data,
 124                                 struct epoll_event_context);
 125         if (epoll_ev == NULL) {
 126                 return false;
 127         }
 128         epoll_ev->panic_fallback = panic_fallback;
 129         return true;
 130 }
 131
 132 /*
 133   called when a epoll call fails
 134 */
 135 static void epoll_panic(struct epoll_event_context *epoll_ev,
 136                         const char *reason, bool replay)
 137 {
 138         struct tevent_context *ev = epoll_ev->ev;
 139         bool (*panic_fallback)(struct tevent_context *ev, bool replay);
 140
 141         panic_fallback = epoll_ev->panic_fallback;
 142
 143         if (epoll_ev->panic_state != NULL) {
 144                 *epoll_ev->panic_state = true;
 145         }
 146
 147         if (epoll_ev->panic_force_replay) {
 148                 replay = true;
 149         }
 150
 151         TALLOC_FREE(ev->additional_data);
 152
 153         if (panic_fallback == NULL) {
 154                 tevent_debug(ev, TEVENT_DEBUG_FATAL,
 155                         "%s (%s) replay[%u] - calling abort()\n",
 156                         reason, strerror(errno), (unsigned)replay);
 157                 abort();
 158         }
 159
 160         tevent_debug(ev, TEVENT_DEBUG_WARNING,
 161                      "%s (%s) replay[%u] - calling panic_fallback\n",
 162                      reason, strerror(errno), (unsigned)replay);
 163
 164         if (!panic_fallback(ev, replay)) {
 165                 /* Fallback failed. */
 166                 tevent_debug(ev, TEVENT_DEBUG_FATAL,
 167                         "%s (%s) replay[%u] - calling abort()\n",
 168                         reason, strerror(errno), (unsigned)replay);
 169                 abort();
 170         }
 171 }
 172
 173 /*
 174   map from TEVENT_FD_* to EPOLLIN/EPOLLOUT
 175 */
 176 static uint32_t epoll_map_flags(uint16_t flags)
 177 {
 178         uint32_t ret = 0;
 179         if (flags & TEVENT_FD_READ) ret |= (EPOLLIN | EPOLLERR | EPOLLHUP);
 180         if (flags & TEVENT_FD_WRITE) ret |= (EPOLLOUT | EPOLLERR | EPOLLHUP);
 181         return ret;
 182 }
 183
 184 /*
 185  free the epoll fd
 186 */
 187 static int epoll_ctx_destructor(struct epoll_event_context *epoll_ev)
 188 {
 189         close(epoll_ev->epoll_fd);
 190         epoll_ev->epoll_fd = -1;
 191         return 0;
 192 }
 193
 194 /*
 195  init the epoll fd
 196 */
 197 static int epoll_init_ctx(struct epoll_event_context *epoll_ev)
 198 {
 199         epoll_ev->epoll_fd = epoll_create(64);
 200         if (epoll_ev->epoll_fd == -1) {
 201                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_FATAL,
 202                              "Failed to create epoll handle.\n");
 203                 return -1;
 204         }
 205
 206         if (!ev_set_close_on_exec(epoll_ev->epoll_fd)) {
 207                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_WARNING,
 208                              "Failed to set close-on-exec, file descriptor may be leaked to children.\n");
 209         }
 210
 211         epoll_ev->pid = getpid();
 212         talloc_set_destructor(epoll_ev, epoll_ctx_destructor);
 213
 214         return 0;
 215 }
 216
 217 static void epoll_add_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde);
 218
 219 /*
 220   reopen the epoll handle when our pid changes
 221   see http://junkcode.samba.org/ftp/unpacked/junkcode/epoll_fork.c for an
 222   demonstration of why this is needed
 223  */
 224 static void epoll_check_reopen(struct epoll_event_context *epoll_ev)
 225 {
 226         struct tevent_fd *fde;
 227         bool *caller_panic_state = epoll_ev->panic_state;
 228         bool panic_triggered = false;
 229
 230         if (epoll_ev->pid == getpid()) {
 231                 return;
 232         }
 233
 234         close(epoll_ev->epoll_fd);
 235         epoll_ev->epoll_fd = epoll_create(64);
 236         if (epoll_ev->epoll_fd == -1) {
 237                 epoll_panic(epoll_ev, "epoll_create() failed", false);
 238                 return;
 239         }
 240
 241         if (!ev_set_close_on_exec(epoll_ev->epoll_fd)) {
 242                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_WARNING,
 243                              "Failed to set close-on-exec, file descriptor may be leaked to children.\n");
 244         }
 245
 246         epoll_ev->pid = getpid();
 247         epoll_ev->panic_state = &panic_triggered;
 248         for (fde=epoll_ev->ev->fd_events;fde;fde=fde->next) {
 249                 epoll_add_event(epoll_ev, fde);
 250                 if (panic_triggered) {
 251                         if (caller_panic_state != NULL) {
 252                                 *caller_panic_state = true;
 253                         }
 254                         return;
 255                 }
 256         }
 257         epoll_ev->panic_state = NULL;
 258 }
 259
 260 #define EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT      (1<<0)
 261 #define EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR   (1<<1)
 262 #define EPOLL_ADDITIONAL_FD_FLAG_GOT_ERROR      (1<<2)
 263
 264 /*
 265  add the epoll event to the given fd_event
 266 */
 267 static void epoll_add_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde)
 268 {
 269         struct epoll_event event;
 270         int ret;
 271
 272         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 273
 274         /* if we don't want events yet, don't add an epoll_event */
 275         if (fde->flags == 0) return;
 276
 277         ZERO_STRUCT(event);
 278         event.events = epoll_map_flags(fde->flags);
 279         event.data.ptr = fde;
 280         ret = epoll_ctl(epoll_ev->epoll_fd, EPOLL_CTL_ADD, fde->fd, &event);
 281         if (ret != 0) {
 282                 epoll_panic(epoll_ev, "EPOLL_CTL_ADD failed", false);
 283                 return;
 284         }
 285         fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 286
 287         /* only if we want to read we want to tell the event handler about errors */
 288         if (fde->flags & TEVENT_FD_READ) {
 289                 fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 290         }
 291 }
 292
 293 /*
 294  delete the epoll event for given fd_event
 295 */
 296 static void epoll_del_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde)
 297 {
 298         struct epoll_event event;
 299         int ret;
 300
 301         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 302
 303         /* if there's no epoll_event, we don't need to delete it */
 304         if (!(fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT)) return;
 305
 306         ZERO_STRUCT(event);
 307         event.events = epoll_map_flags(fde->flags);
 308         event.data.ptr = fde;
 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_change_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_del_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
 496         if (ev == NULL) {
 497                 return tevent_common_fd_destructor(fde);
 498         }
 499
 500         epoll_ev = talloc_get_type_abort(ev->additional_data,
 501                                          struct epoll_event_context);
 502
 503         /*
 504          * we must remove the event from the list
 505          * otherwise a panic fallback handler may
 506          * reuse invalid memory
 507          */
 508         DLIST_REMOVE(ev->fd_events, fde);
 509
 510         epoll_ev->panic_state = &panic_triggered;
 511         epoll_check_reopen(epoll_ev);
 512         if (panic_triggered) {
 513                 return tevent_common_fd_destructor(fde);
 514         }
 515
 516         epoll_del_event(epoll_ev, fde);
 517         if (panic_triggered) {
 518                 return tevent_common_fd_destructor(fde);
 519         }
 520         epoll_ev->panic_state = NULL;
 521
 522         return tevent_common_fd_destructor(fde);
 523 }
 524
 525 /*
 526   add a fd based event
 527   return NULL on failure (memory allocation error)
 528 */
 529 static struct tevent_fd *epoll_event_add_fd(struct tevent_context *ev, TALLOC_CTX *mem_ctx,
 530                                             int fd, uint16_t flags,
 531                                             tevent_fd_handler_t handler,
 532                                             void *private_data,
 533                                             const char *handler_name,
 534                                             const char *location)
 535 {
 536         struct epoll_event_context *epoll_ev = talloc_get_type(ev->additional_data,
 537                                                            struct epoll_event_context);
 538         struct tevent_fd *fde;
 539         bool panic_triggered = false;
 540
 541         fde = tevent_common_add_fd(ev, mem_ctx, fd, flags,
 542                                    handler, private_data,
 543                                    handler_name, location);
 544         if (!fde) return NULL;
 545
 546         talloc_set_destructor(fde, epoll_event_fd_destructor);
 547
 548         epoll_ev->panic_state = &panic_triggered;
 549         epoll_check_reopen(epoll_ev);
 550         if (panic_triggered) {
 551                 return fde;
 552         }
 553         epoll_ev->panic_state = NULL;
 554
 555         epoll_add_event(epoll_ev, fde);
 556
 557         return fde;
 558 }
 559
 560 /*
 561   set the fd event flags
 562 */
 563 static void epoll_event_set_fd_flags(struct tevent_fd *fde, uint16_t flags)
 564 {
 565         struct tevent_context *ev;
 566         struct epoll_event_context *epoll_ev;
 567         bool panic_triggered = false;
 568
 569         if (fde->flags == flags) return;
 570
 571         ev = fde->event_ctx;
 572         epoll_ev = talloc_get_type(ev->additional_data, struct epoll_event_context);
 573
 574         fde->flags = flags;
 575
 576         epoll_ev->panic_state = &panic_triggered;
 577         epoll_check_reopen(epoll_ev);
 578         if (panic_triggered) {
 579                 return;
 580         }
 581         epoll_ev->panic_state = NULL;
 582
 583         epoll_change_event(epoll_ev, fde);
 584 }
 585
 586 /*
 587   do a single event loop using the events defined in ev
 588 */
 589 static int epoll_event_loop_once(struct tevent_context *ev, const char *location)
 590 {
 591         struct epoll_event_context *epoll_ev = talloc_get_type(ev->additional_data,
 592                                                            struct epoll_event_context);
 593         struct timeval tval;
 594         bool panic_triggered = false;
 595
 596         if (ev->signal_events &&
 597             tevent_common_check_signal(ev)) {
 598                 return 0;
 599         }
 600
 601         if (ev->immediate_events &&
 602             tevent_common_loop_immediate(ev)) {
 603                 return 0;
 604         }
 605
 606         tval = tevent_common_loop_timer_delay(ev);
 607         if (tevent_timeval_is_zero(&tval)) {
 608                 return 0;
 609         }
 610
 611         epoll_ev->panic_state = &panic_triggered;
 612         epoll_ev->panic_force_replay = true;
 613         epoll_check_reopen(epoll_ev);
 614         if (panic_triggered) {
 615                 errno = EINVAL;
 616                 return -1;
 617         }
 618         epoll_ev->panic_force_replay = false;
 619         epoll_ev->panic_state = NULL;
 620
 621         return epoll_event_loop(epoll_ev, &tval);
 622 }
 623
 624 static const struct tevent_ops epoll_event_ops = {
 625         .context_init           = epoll_event_context_init,
 626         .add_fd                 = epoll_event_add_fd,
 627         .set_fd_close_fn        = tevent_common_fd_set_close_fn,
 628         .get_fd_flags           = tevent_common_fd_get_flags,
 629         .set_fd_flags           = epoll_event_set_fd_flags,
 630         .add_timer              = tevent_common_add_timer,
 631         .schedule_immediate     = tevent_common_schedule_immediate,
 632         .add_signal             = tevent_common_add_signal,
 633         .loop_once              = epoll_event_loop_once,
 634         .loop_wait              = tevent_common_loop_wait,
 635 };
 636
 637 _PRIVATE_ bool tevent_epoll_init(void)
 638 {
 639         return tevent_register_backend("epoll", &epoll_event_ops);
 640 }