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