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