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