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         ret = epoll_ctl(epoll_ev->epoll_fd, EPOLL_CTL_DEL, fde->fd, &event);
 308         if (ret != 0 && errno == ENOENT) {
 309                 /*
 310                  * This can happen after a epoll_check_reopen
 311                  * within epoll_event_fd_destructor.
 312                  */
 313                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_TRACE,
 314                              "EPOLL_CTL_DEL ignoring ENOENT for fd[%d]\n",
 315                              fde->fd);
 316         } else if (ret != 0) {
 317                 epoll_panic(epoll_ev, "EPOLL_CTL_DEL failed", false);
 318                 return;
 319         }
 320         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 321 }
 322
 323 /*
 324  change the epoll event to the given fd_event
 325 */
 326 static void epoll_mod_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde)
 327 {
 328         struct epoll_event event;
 329         int ret;
 330
 331         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 332
 333         ZERO_STRUCT(event);
 334         event.events = epoll_map_flags(fde->flags);
 335         event.data.ptr = fde;
 336         ret = epoll_ctl(epoll_ev->epoll_fd, EPOLL_CTL_MOD, fde->fd, &event);
 337         if (ret != 0) {
 338                 epoll_panic(epoll_ev, "EPOLL_CTL_MOD failed", false);
 339                 return;
 340         }
 341
 342         /* only if we want to read we want to tell the event handler about errors */
 343         if (fde->flags & TEVENT_FD_READ) {
 344                 fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 345         }
 346 }
 347
 348 static void epoll_update_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde)
 349 {
 350         bool got_error = (fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_GOT_ERROR);
 351         bool want_read = (fde->flags & TEVENT_FD_READ);
 352         bool want_write= (fde->flags & TEVENT_FD_WRITE);
 353
 354         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 355
 356         /* there's already an event */
 357         if (fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT) {
 358                 if (want_read || (want_write && !got_error)) {
 359                         epoll_mod_event(epoll_ev, fde);
 360                         return;
 361                 }
 362                 /*
 363                  * if we want to match the select behavior, we need to remove the epoll_event
 364                  * when the caller isn't interested in events.
 365                  *
 366                  * this is because epoll reports EPOLLERR and EPOLLHUP, even without asking for them
 367                  */
 368                 epoll_del_event(epoll_ev, fde);
 369                 return;
 370         }
 371
 372         /* there's no epoll_event attached to the fde */
 373         if (want_read || (want_write && !got_error)) {
 374                 epoll_add_event(epoll_ev, fde);
 375                 return;
 376         }
 377 }
 378
 379 /*
 380   event loop handling using epoll
 381 */
 382 static int epoll_event_loop(struct epoll_event_context *epoll_ev, struct timeval *tvalp)
 383 {
 384         int ret, i;
 385 #define MAXEVENTS 1
 386         struct epoll_event events[MAXEVENTS];
 387         int timeout = -1;
 388         int wait_errno;
 389
 390         if (tvalp) {
 391                 /* it's better to trigger timed events a bit later than too early */
 392                 timeout = ((tvalp->tv_usec+999) / 1000) + (tvalp->tv_sec*1000);
 393         }
 394
 395         if (epoll_ev->ev->signal_events &&
 396             tevent_common_check_signal(epoll_ev->ev)) {
 397                 return 0;
 398         }
 399
 400         tevent_trace_point_callback(epoll_ev->ev, TEVENT_TRACE_BEFORE_WAIT);
 401         ret = epoll_wait(epoll_ev->epoll_fd, events, MAXEVENTS, timeout);
 402         wait_errno = errno;
 403         tevent_trace_point_callback(epoll_ev->ev, TEVENT_TRACE_AFTER_WAIT);
 404
 405         if (ret == -1 && wait_errno == EINTR && epoll_ev->ev->signal_events) {
 406                 if (tevent_common_check_signal(epoll_ev->ev)) {
 407                         return 0;
 408                 }
 409         }
 410
 411         if (ret == -1 && wait_errno != EINTR) {
 412                 epoll_panic(epoll_ev, "epoll_wait() failed", true);
 413                 return -1;
 414         }
 415
 416         if (ret == 0 && tvalp) {
 417                 /* we don't care about a possible delay here */
 418                 tevent_common_loop_timer_delay(epoll_ev->ev);
 419                 return 0;
 420         }
 421
 422         for (i=0;i<ret;i++) {
 423                 struct tevent_fd *fde = talloc_get_type(events[i].data.ptr,
 424                                                        struct tevent_fd);
 425                 uint16_t flags = 0;
 426
 427                 if (fde == NULL) {
 428                         epoll_panic(epoll_ev, "epoll_wait() gave bad data", true);
 429                         return -1;
 430                 }
 431                 if (events[i].events & (EPOLLHUP|EPOLLERR)) {
 432                         fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_GOT_ERROR;
 433                         /*
 434                          * if we only wait for TEVENT_FD_WRITE, we should not tell the
 435                          * event handler about it, and remove the epoll_event,
 436                          * as we only report errors when waiting for read events,
 437                          * to match the select() behavior
 438                          */
 439                         if (!(fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR)) {
 440                                 epoll_del_event(epoll_ev, fde);
 441                                 continue;
 442                         }
 443                         flags |= TEVENT_FD_READ;
 444                 }
 445                 if (events[i].events & EPOLLIN) flags |= TEVENT_FD_READ;
 446                 if (events[i].events & EPOLLOUT) flags |= TEVENT_FD_WRITE;
 447                 if (flags) {
 448                         fde->handler(epoll_ev->ev, fde, flags, fde->private_data);
 449                         break;
 450                 }
 451         }
 452
 453         return 0;
 454 }
 455
 456 /*
 457   create a epoll_event_context structure.
 458 */
 459 static int epoll_event_context_init(struct tevent_context *ev)
 460 {
 461         int ret;
 462         struct epoll_event_context *epoll_ev;
 463
 464         /*
 465          * We might be called during tevent_re_initialise()
 466          * which means we need to free our old additional_data.
 467          */
 468         TALLOC_FREE(ev->additional_data);
 469
 470         epoll_ev = talloc_zero(ev, struct epoll_event_context);
 471         if (!epoll_ev) return -1;
 472         epoll_ev->ev = ev;
 473         epoll_ev->epoll_fd = -1;
 474
 475         ret = epoll_init_ctx(epoll_ev);
 476         if (ret != 0) {
 477                 talloc_free(epoll_ev);
 478                 return ret;
 479         }
 480
 481         ev->additional_data = epoll_ev;
 482         return 0;
 483 }
 484
 485 /*
 486   destroy an fd_event
 487 */
 488 static int epoll_event_fd_destructor(struct tevent_fd *fde)
 489 {
 490         struct tevent_context *ev = fde->event_ctx;
 491         struct epoll_event_context *epoll_ev = NULL;
 492         bool panic_triggered = false;
 493
 494         if (ev == NULL) {
 495                 return tevent_common_fd_destructor(fde);
 496         }
 497
 498         epoll_ev = talloc_get_type_abort(ev->additional_data,
 499                                          struct epoll_event_context);
 500
 501         /*
 502          * we must remove the event from the list
 503          * otherwise a panic fallback handler may
 504          * reuse invalid memory
 505          */
 506         DLIST_REMOVE(ev->fd_events, fde);
 507
 508         epoll_ev->panic_state = &panic_triggered;
 509         epoll_check_reopen(epoll_ev);
 510         if (panic_triggered) {
 511                 return tevent_common_fd_destructor(fde);
 512         }
 513
 514         epoll_del_event(epoll_ev, fde);
 515         if (panic_triggered) {
 516                 return tevent_common_fd_destructor(fde);
 517         }
 518         epoll_ev->panic_state = NULL;
 519
 520         return tevent_common_fd_destructor(fde);
 521 }
 522
 523 /*
 524   add a fd based event
 525   return NULL on failure (memory allocation error)
 526 */
 527 static struct tevent_fd *epoll_event_add_fd(struct tevent_context *ev, TALLOC_CTX *mem_ctx,
 528                                             int fd, uint16_t flags,
 529                                             tevent_fd_handler_t handler,
 530                                             void *private_data,
 531                                             const char *handler_name,
 532                                             const char *location)
 533 {
 534         struct epoll_event_context *epoll_ev = talloc_get_type(ev->additional_data,
 535                                                            struct epoll_event_context);
 536         struct tevent_fd *fde;
 537         bool panic_triggered = false;
 538
 539         fde = tevent_common_add_fd(ev, mem_ctx, fd, flags,
 540                                    handler, private_data,
 541                                    handler_name, location);
 542         if (!fde) return NULL;
 543
 544         talloc_set_destructor(fde, epoll_event_fd_destructor);
 545
 546         epoll_ev->panic_state = &panic_triggered;
 547         epoll_check_reopen(epoll_ev);
 548         if (panic_triggered) {
 549                 return fde;
 550         }
 551         epoll_ev->panic_state = NULL;
 552
 553         epoll_add_event(epoll_ev, fde);
 554
 555         return fde;
 556 }
 557
 558 /*
 559   set the fd event flags
 560 */
 561 static void epoll_event_set_fd_flags(struct tevent_fd *fde, uint16_t flags)
 562 {
 563         struct tevent_context *ev;
 564         struct epoll_event_context *epoll_ev;
 565         bool panic_triggered = false;
 566
 567         if (fde->flags == flags) return;
 568
 569         ev = fde->event_ctx;
 570         epoll_ev = talloc_get_type(ev->additional_data, struct epoll_event_context);
 571
 572         fde->flags = flags;
 573
 574         epoll_ev->panic_state = &panic_triggered;
 575         epoll_check_reopen(epoll_ev);
 576         if (panic_triggered) {
 577                 return;
 578         }
 579         epoll_ev->panic_state = NULL;
 580
 581         epoll_update_event(epoll_ev, fde);
 582 }
 583
 584 /*
 585   do a single event loop using the events defined in ev
 586 */
 587 static int epoll_event_loop_once(struct tevent_context *ev, const char *location)
 588 {
 589         struct epoll_event_context *epoll_ev = talloc_get_type(ev->additional_data,
 590                                                            struct epoll_event_context);
 591         struct timeval tval;
 592         bool panic_triggered = false;
 593
 594         if (ev->signal_events &&
 595             tevent_common_check_signal(ev)) {
 596                 return 0;
 597         }
 598
 599         if (ev->immediate_events &&
 600             tevent_common_loop_immediate(ev)) {
 601                 return 0;
 602         }
 603
 604         tval = tevent_common_loop_timer_delay(ev);
 605         if (tevent_timeval_is_zero(&tval)) {
 606                 return 0;
 607         }
 608
 609         epoll_ev->panic_state = &panic_triggered;
 610         epoll_ev->panic_force_replay = true;
 611         epoll_check_reopen(epoll_ev);
 612         if (panic_triggered) {
 613                 errno = EINVAL;
 614                 return -1;
 615         }
 616         epoll_ev->panic_force_replay = false;
 617         epoll_ev->panic_state = NULL;
 618
 619         return epoll_event_loop(epoll_ev, &tval);
 620 }
 621
 622 static const struct tevent_ops epoll_event_ops = {
 623         .context_init           = epoll_event_context_init,
 624         .add_fd                 = epoll_event_add_fd,
 625         .set_fd_close_fn        = tevent_common_fd_set_close_fn,
 626         .get_fd_flags           = tevent_common_fd_get_flags,
 627         .set_fd_flags           = epoll_event_set_fd_flags,
 628         .add_timer              = tevent_common_add_timer,
 629         .schedule_immediate     = tevent_common_schedule_immediate,
 630         .add_signal             = tevent_common_add_signal,
 631         .loop_once              = epoll_event_loop_once,
 632         .loop_wait              = tevent_common_loop_wait,
 633 };
 634
 635 _PRIVATE_ bool tevent_epoll_init(void)
 636 {
 637         return tevent_register_backend("epoll", &epoll_event_ops);
 638 }