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