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_fallback)(struct tevent_context *ev, bool replay);
  44 };
  45
  46 #ifdef TEST_PANIC_FALLBACK
  47 static int epoll_wait_panic_fallback(int epfd,
  48                                 struct epoll_event *events,
  49                                 int maxevents,
  50                                 int timeout)
  51 {
  52         /* 50% of the time, fail... */
  53         if ((random() % 2) == 0) {
  54                 errno = EINVAL;
  55                 return -1;
  56         }
  57
  58         return epoll_wait(epfd, events, maxevents, timeout);
  59 }
  60
  61 #define epoll_wait epoll_wait_panic_fallback
  62 #endif
  63
  64 /*
  65   called to set the panic fallback function.
  66 */
  67 _PRIVATE_ bool tevent_epoll_set_panic_fallback(struct tevent_context *ev,
  68                                 bool (*panic_fallback)(struct tevent_context *ev,
  69                                                        bool replay))
  70 {
  71         struct epoll_event_context *epoll_ev;
  72
  73         if (ev->additional_data == NULL) {
  74                 return false;
  75         }
  76
  77         epoll_ev = talloc_get_type(ev->additional_data,
  78                                 struct epoll_event_context);
  79         if (epoll_ev == NULL) {
  80                 return false;
  81         }
  82         epoll_ev->panic_fallback = panic_fallback;
  83         return true;
  84 }
  85
  86 /*
  87   called when a epoll call fails
  88 */
  89 static void epoll_panic(struct epoll_event_context *epoll_ev,
  90                         const char *reason, bool replay)
  91 {
  92         struct tevent_context *ev = epoll_ev->ev;
  93
  94         if (epoll_ev->panic_fallback == NULL) {
  95                 tevent_debug(ev, TEVENT_DEBUG_FATAL,
  96                         "%s (%s) replay[%u] - calling abort()\n",
  97                         reason, strerror(errno), (unsigned)replay);
  98                 abort();
  99         }
 100
 101         tevent_debug(ev, TEVENT_DEBUG_WARNING,
 102                      "%s (%s) replay[%u] - calling panic_fallback\n",
 103                      reason, strerror(errno), (unsigned)replay);
 104
 105         if (!epoll_ev->panic_fallback(ev, replay)) {
 106                 /* Fallback failed. */
 107                 tevent_debug(ev, TEVENT_DEBUG_FATAL,
 108                         "%s (%s) replay[%u] - calling abort()\n",
 109                         reason, strerror(errno), (unsigned)replay);
 110                 abort();
 111         }
 112 }
 113
 114 /*
 115   map from TEVENT_FD_* to EPOLLIN/EPOLLOUT
 116 */
 117 static uint32_t epoll_map_flags(uint16_t flags)
 118 {
 119         uint32_t ret = 0;
 120         if (flags & TEVENT_FD_READ) ret |= (EPOLLIN | EPOLLERR | EPOLLHUP);
 121         if (flags & TEVENT_FD_WRITE) ret |= (EPOLLOUT | EPOLLERR | EPOLLHUP);
 122         return ret;
 123 }
 124
 125 /*
 126  free the epoll fd
 127 */
 128 static int epoll_ctx_destructor(struct epoll_event_context *epoll_ev)
 129 {
 130         close(epoll_ev->epoll_fd);
 131         epoll_ev->epoll_fd = -1;
 132         return 0;
 133 }
 134
 135 /*
 136  init the epoll fd
 137 */
 138 static int epoll_init_ctx(struct epoll_event_context *epoll_ev)
 139 {
 140         epoll_ev->epoll_fd = epoll_create(64);
 141         if (epoll_ev->epoll_fd == -1) {
 142                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_FATAL,
 143                              "Failed to create epoll handle.\n");
 144                 return -1;
 145         }
 146
 147         if (!ev_set_close_on_exec(epoll_ev->epoll_fd)) {
 148                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_WARNING,
 149                              "Failed to set close-on-exec, file descriptor may be leaked to children.\n");
 150         }
 151
 152         epoll_ev->pid = getpid();
 153         talloc_set_destructor(epoll_ev, epoll_ctx_destructor);
 154
 155         return 0;
 156 }
 157
 158 static void epoll_add_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde);
 159
 160 /*
 161   reopen the epoll handle when our pid changes
 162   see http://junkcode.samba.org/ftp/unpacked/junkcode/epoll_fork.c for an
 163   demonstration of why this is needed
 164  */
 165 static void epoll_check_reopen(struct epoll_event_context *epoll_ev)
 166 {
 167         struct tevent_fd *fde;
 168
 169         if (epoll_ev->pid == getpid()) {
 170                 return;
 171         }
 172
 173         close(epoll_ev->epoll_fd);
 174         epoll_ev->epoll_fd = epoll_create(64);
 175         if (epoll_ev->epoll_fd == -1) {
 176                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_FATAL,
 177                              "Failed to recreate epoll handle after fork\n");
 178                 return;
 179         }
 180
 181         if (!ev_set_close_on_exec(epoll_ev->epoll_fd)) {
 182                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_WARNING,
 183                              "Failed to set close-on-exec, file descriptor may be leaked to children.\n");
 184         }
 185
 186         epoll_ev->pid = getpid();
 187         for (fde=epoll_ev->ev->fd_events;fde;fde=fde->next) {
 188                 epoll_add_event(epoll_ev, fde);
 189         }
 190 }
 191
 192 #define EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT      (1<<0)
 193 #define EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR   (1<<1)
 194 #define EPOLL_ADDITIONAL_FD_FLAG_GOT_ERROR      (1<<2)
 195
 196 /*
 197  add the epoll event to the given fd_event
 198 */
 199 static void epoll_add_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde)
 200 {
 201         struct epoll_event event;
 202
 203         if (epoll_ev->epoll_fd == -1) return;
 204
 205         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 206
 207         /* if we don't want events yet, don't add an epoll_event */
 208         if (fde->flags == 0) return;
 209
 210         ZERO_STRUCT(event);
 211         event.events = epoll_map_flags(fde->flags);
 212         event.data.ptr = fde;
 213         if (epoll_ctl(epoll_ev->epoll_fd, EPOLL_CTL_ADD, fde->fd, &event) != 0) {
 214                 epoll_panic(epoll_ev, "EPOLL_CTL_ADD failed", false);
 215                 return;
 216         }
 217         fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 218
 219         /* only if we want to read we want to tell the event handler about errors */
 220         if (fde->flags & TEVENT_FD_READ) {
 221                 fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 222         }
 223 }
 224
 225 /*
 226  delete the epoll event for given fd_event
 227 */
 228 static void epoll_del_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde)
 229 {
 230         struct epoll_event event;
 231
 232         if (epoll_ev->epoll_fd == -1) return;
 233
 234         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 235
 236         /* if there's no epoll_event, we don't need to delete it */
 237         if (!(fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT)) return;
 238
 239         ZERO_STRUCT(event);
 240         event.events = epoll_map_flags(fde->flags);
 241         event.data.ptr = fde;
 242         if (epoll_ctl(epoll_ev->epoll_fd, EPOLL_CTL_DEL, fde->fd, &event) != 0) {
 243                 tevent_debug(epoll_ev->ev, TEVENT_DEBUG_FATAL,
 244                              "epoll_del_event failed! probable early close bug (%s)\n",
 245                              strerror(errno));
 246         }
 247         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT;
 248 }
 249
 250 /*
 251  change the epoll event to the given fd_event
 252 */
 253 static void epoll_mod_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde)
 254 {
 255         struct epoll_event event;
 256         if (epoll_ev->epoll_fd == -1) return;
 257
 258         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 259
 260         ZERO_STRUCT(event);
 261         event.events = epoll_map_flags(fde->flags);
 262         event.data.ptr = fde;
 263         if (epoll_ctl(epoll_ev->epoll_fd, EPOLL_CTL_MOD, fde->fd, &event) != 0) {
 264                 epoll_panic(epoll_ev, "EPOLL_CTL_MOD failed", false);
 265                 return;
 266         }
 267
 268         /* only if we want to read we want to tell the event handler about errors */
 269         if (fde->flags & TEVENT_FD_READ) {
 270                 fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 271         }
 272 }
 273
 274 static void epoll_change_event(struct epoll_event_context *epoll_ev, struct tevent_fd *fde)
 275 {
 276         bool got_error = (fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_GOT_ERROR);
 277         bool want_read = (fde->flags & TEVENT_FD_READ);
 278         bool want_write= (fde->flags & TEVENT_FD_WRITE);
 279
 280         if (epoll_ev->epoll_fd == -1) return;
 281
 282         fde->additional_flags &= ~EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR;
 283
 284         /* there's already an event */
 285         if (fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_HAS_EVENT) {
 286                 if (want_read || (want_write && !got_error)) {
 287                         epoll_mod_event(epoll_ev, fde);
 288                         return;
 289                 }
 290                 /*
 291                  * if we want to match the select behavior, we need to remove the epoll_event
 292                  * when the caller isn't interested in events.
 293                  *
 294                  * this is because epoll reports EPOLLERR and EPOLLHUP, even without asking for them
 295                  */
 296                 epoll_del_event(epoll_ev, fde);
 297                 return;
 298         }
 299
 300         /* there's no epoll_event attached to the fde */
 301         if (want_read || (want_write && !got_error)) {
 302                 epoll_add_event(epoll_ev, fde);
 303                 return;
 304         }
 305 }
 306
 307 /*
 308   event loop handling using epoll
 309 */
 310 static int epoll_event_loop(struct epoll_event_context *epoll_ev, struct timeval *tvalp)
 311 {
 312         int ret, i;
 313 #define MAXEVENTS 1
 314         struct epoll_event events[MAXEVENTS];
 315         int timeout = -1;
 316
 317         if (epoll_ev->epoll_fd == -1) return -1;
 318
 319         if (tvalp) {
 320                 /* it's better to trigger timed events a bit later than to early */
 321                 timeout = ((tvalp->tv_usec+999) / 1000) + (tvalp->tv_sec*1000);
 322         }
 323
 324         if (epoll_ev->ev->signal_events &&
 325             tevent_common_check_signal(epoll_ev->ev)) {
 326                 return 0;
 327         }
 328
 329         tevent_trace_point_callback(epoll_ev->ev, TEVENT_TRACE_BEFORE_WAIT);
 330         ret = epoll_wait(epoll_ev->epoll_fd, events, MAXEVENTS, timeout);
 331         tevent_trace_point_callback(epoll_ev->ev, TEVENT_TRACE_AFTER_WAIT);
 332
 333         if (ret == -1 && errno == EINTR && epoll_ev->ev->signal_events) {
 334                 if (tevent_common_check_signal(epoll_ev->ev)) {
 335                         return 0;
 336                 }
 337         }
 338
 339         if (ret == -1 && errno != EINTR) {
 340                 epoll_panic(epoll_ev, "epoll_wait() failed", true);
 341                 return -1;
 342         }
 343
 344         if (ret == 0 && tvalp) {
 345                 /* we don't care about a possible delay here */
 346                 tevent_common_loop_timer_delay(epoll_ev->ev);
 347                 return 0;
 348         }
 349
 350         for (i=0;i<ret;i++) {
 351                 struct tevent_fd *fde = talloc_get_type(events[i].data.ptr,
 352                                                        struct tevent_fd);
 353                 uint16_t flags = 0;
 354
 355                 if (fde == NULL) {
 356                         epoll_panic(epoll_ev, "epoll_wait() gave bad data", true);
 357                         return -1;
 358                 }
 359                 if (events[i].events & (EPOLLHUP|EPOLLERR)) {
 360                         fde->additional_flags |= EPOLL_ADDITIONAL_FD_FLAG_GOT_ERROR;
 361                         /*
 362                          * if we only wait for TEVENT_FD_WRITE, we should not tell the
 363                          * event handler about it, and remove the epoll_event,
 364                          * as we only report errors when waiting for read events,
 365                          * to match the select() behavior
 366                          */
 367                         if (!(fde->additional_flags & EPOLL_ADDITIONAL_FD_FLAG_REPORT_ERROR)) {
 368                                 epoll_del_event(epoll_ev, fde);
 369                                 continue;
 370                         }
 371                         flags |= TEVENT_FD_READ;
 372                 }
 373                 if (events[i].events & EPOLLIN) flags |= TEVENT_FD_READ;
 374                 if (events[i].events & EPOLLOUT) flags |= TEVENT_FD_WRITE;
 375                 if (flags) {
 376                         fde->handler(epoll_ev->ev, fde, flags, fde->private_data);
 377                         break;
 378                 }
 379         }
 380
 381         return 0;
 382 }
 383
 384 /*
 385   create a epoll_event_context structure.
 386 */
 387 static int epoll_event_context_init(struct tevent_context *ev)
 388 {
 389         int ret;
 390         struct epoll_event_context *epoll_ev;
 391
 392         /*
 393          * We might be called during tevent_re_initialise()
 394          * which means we need to free our old additional_data.
 395          */
 396         TALLOC_FREE(ev->additional_data);
 397
 398         epoll_ev = talloc_zero(ev, struct epoll_event_context);
 399         if (!epoll_ev) return -1;
 400         epoll_ev->ev = ev;
 401         epoll_ev->epoll_fd = -1;
 402
 403         ret = epoll_init_ctx(epoll_ev);
 404         if (ret != 0) {
 405                 talloc_free(epoll_ev);
 406                 return ret;
 407         }
 408
 409         ev->additional_data = epoll_ev;
 410         return 0;
 411 }
 412
 413 /*
 414   destroy an fd_event
 415 */
 416 static int epoll_event_fd_destructor(struct tevent_fd *fde)
 417 {
 418         struct tevent_context *ev = fde->event_ctx;
 419         struct epoll_event_context *epoll_ev = NULL;
 420
 421         if (ev) {
 422                 epoll_ev = talloc_get_type(ev->additional_data,
 423                                            struct epoll_event_context);
 424
 425                 epoll_check_reopen(epoll_ev);
 426
 427                 epoll_del_event(epoll_ev, fde);
 428         }
 429
 430         return tevent_common_fd_destructor(fde);
 431 }
 432
 433 /*
 434   add a fd based event
 435   return NULL on failure (memory allocation error)
 436 */
 437 static struct tevent_fd *epoll_event_add_fd(struct tevent_context *ev, TALLOC_CTX *mem_ctx,
 438                                             int fd, uint16_t flags,
 439                                             tevent_fd_handler_t handler,
 440                                             void *private_data,
 441                                             const char *handler_name,
 442                                             const char *location)
 443 {
 444         struct epoll_event_context *epoll_ev = talloc_get_type(ev->additional_data,
 445                                                            struct epoll_event_context);
 446         struct tevent_fd *fde;
 447
 448         epoll_check_reopen(epoll_ev);
 449
 450         fde = tevent_common_add_fd(ev, mem_ctx, fd, flags,
 451                                    handler, private_data,
 452                                    handler_name, location);
 453         if (!fde) return NULL;
 454
 455         talloc_set_destructor(fde, epoll_event_fd_destructor);
 456
 457         epoll_add_event(epoll_ev, fde);
 458
 459         return fde;
 460 }
 461
 462 /*
 463   set the fd event flags
 464 */
 465 static void epoll_event_set_fd_flags(struct tevent_fd *fde, uint16_t flags)
 466 {
 467         struct tevent_context *ev;
 468         struct epoll_event_context *epoll_ev;
 469
 470         if (fde->flags == flags) return;
 471
 472         ev = fde->event_ctx;
 473         epoll_ev = talloc_get_type(ev->additional_data, struct epoll_event_context);
 474
 475         fde->flags = flags;
 476
 477         epoll_check_reopen(epoll_ev);
 478
 479         epoll_change_event(epoll_ev, fde);
 480 }
 481
 482 /*
 483   do a single event loop using the events defined in ev
 484 */
 485 static int epoll_event_loop_once(struct tevent_context *ev, const char *location)
 486 {
 487         struct epoll_event_context *epoll_ev = talloc_get_type(ev->additional_data,
 488                                                            struct epoll_event_context);
 489         struct timeval tval;
 490
 491         if (ev->signal_events &&
 492             tevent_common_check_signal(ev)) {
 493                 return 0;
 494         }
 495
 496         if (ev->immediate_events &&
 497             tevent_common_loop_immediate(ev)) {
 498                 return 0;
 499         }
 500
 501         tval = tevent_common_loop_timer_delay(ev);
 502         if (tevent_timeval_is_zero(&tval)) {
 503                 return 0;
 504         }
 505
 506         epoll_check_reopen(epoll_ev);
 507
 508         return epoll_event_loop(epoll_ev, &tval);
 509 }
 510
 511 static const struct tevent_ops epoll_event_ops = {
 512         .context_init           = epoll_event_context_init,
 513         .add_fd                 = epoll_event_add_fd,
 514         .set_fd_close_fn        = tevent_common_fd_set_close_fn,
 515         .get_fd_flags           = tevent_common_fd_get_flags,
 516         .set_fd_flags           = epoll_event_set_fd_flags,
 517         .add_timer              = tevent_common_add_timer,
 518         .schedule_immediate     = tevent_common_schedule_immediate,
 519         .add_signal             = tevent_common_add_signal,
 520         .loop_once              = epoll_event_loop_once,
 521         .loop_wait              = tevent_common_loop_wait,
 522 };
 523
 524 _PRIVATE_ bool tevent_epoll_init(void)
 525 {
 526         return tevent_register_backend("epoll", &epoll_event_ops);
 527 }