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