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