| blob | c10e73a7cb8b588f2e2181f002adfb49918e31ca |
1 .\" Copyright (C) 2003 Davide Libenzi
2 .\"
3 .\" %%%LICENSE_START(GPLv2+_SW_3_PARA)
4 .\" This program is free software; you can redistribute it and/or modify
5 .\" it under the terms of the GNU General Public License as published by
6 .\" the Free Software Foundation; either version 2 of the License, or
7 .\" (at your option) any later version.
8 .\"
9 .\" This program is distributed in the hope that it will be useful,
10 .\" but WITHOUT ANY WARRANTY; without even the implied warranty of
11 .\" MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 .\" GNU General Public License for more details.
13 .\"
14 .\" You should have received a copy of the GNU General Public
15 .\" License along with this manual; if not, see
16 .\" <http://www.gnu.org/licenses/>.
17 .\" %%%LICENSE_END
18 .\"
19 .\" Davide Libenzi <davidel@xmailserver.org>
20 .\"
21 .TH EPOLL 7 2021-03-22 "Linux" "Linux Programmer's Manual"
22 .SH NAME
23 epoll \- I/O event notification facility
24 .SH SYNOPSIS
25 .nf
26 .B #include <sys/epoll.h>
27 .fi
28 .SH DESCRIPTION
29 The
30 .B epoll
31 API performs a similar task to
32 .BR poll (2):
33 monitoring multiple file descriptors to see if I/O is possible on any of them.
34 The
35 .B epoll
36 API can be used either as an edge-triggered or a level-triggered
37 interface and scales well to large numbers of watched file descriptors.
38 .PP
39 The central concept of the
40 .B epoll
41 API is the
42 .B epoll
43 .IR instance ,
44 an in-kernel data structure which, from a user-space perspective,
45 can be considered as a container for two lists:
46 .IP \(bu 2
47 The
48 .I interest
49 list (sometimes also called the
50 .B epoll
51 set): the set of file descriptors that the process has registered
52 an interest in monitoring.
53 .IP \(bu
54 The
55 .I ready
56 list: the set of file descriptors that are "ready" for I/O.
57 The ready list is a subset of
58 (or, more precisely, a set of references to)
59 the file descriptors in the interest list.
60 The ready list is dynamically populated
61 by the kernel as a result of I/O activity on those file descriptors.
62 .PP
63 The following system calls are provided to
64 create and manage an
65 .B epoll
66 instance:
67 .IP \(bu 2
68 .BR epoll_create (2)
69 creates a new
70 .B epoll
71 instance and returns a file descriptor referring to that instance.
72 (The more recent
73 .BR epoll_create1 (2)
74 extends the functionality of
75 .BR epoll_create (2).)
76 .IP \(bu
77 Interest in particular file descriptors is then registered via
78 .BR epoll_ctl (2),
79 which adds items to the interest list of the
80 .B epoll
81 instance.
82 .IP \(bu
83 .BR epoll_wait (2)
84 waits for I/O events,
85 blocking the calling thread if no events are currently available.
86 (This system call can be thought of as fetching items from
87 the ready list of the
88 .B epoll
89 instance.)
90 .\"
91 .SS Level-triggered and edge-triggered
92 The
93 .B epoll
94 event distribution interface is able to behave both as edge-triggered
95 (ET) and as level-triggered (LT).
96 The difference between the two mechanisms
97 can be described as follows.
98 Suppose that
99 this scenario happens:
100 .IP 1. 3
101 The file descriptor that represents the read side of a pipe
102 .RI ( rfd )
103 is registered on the
104 .B epoll
105 instance.
106 .IP 2.
107 A pipe writer writes 2\ kB of data on the write side of the pipe.
108 .IP 3.
109 A call to
110 .BR epoll_wait (2)
111 is done that will return
112 .I rfd
113 as a ready file descriptor.
114 .IP 4.
115 The pipe reader reads 1\ kB of data from
116 .IR rfd .
117 .IP 5.
118 A call to
119 .BR epoll_wait (2)
120 is done.
121 .PP
122 If the
123 .I rfd
124 file descriptor has been added to the
125 .B epoll
126 interface using the
127 .B EPOLLET
128 (edge-triggered)
129 flag, the call to
130 .BR epoll_wait (2)
131 done in step
132 .B 5
133 will probably hang despite the available data still present in the file
134 input buffer;
135 meanwhile the remote peer might be expecting a response based on the
136 data it already sent.
137 The reason for this is that edge-triggered mode
138 delivers events only when changes occur on the monitored file descriptor.
139 So, in step
140 .B 5
141 the caller might end up waiting for some data that is already present inside
142 the input buffer.
143 In the above example, an event on
144 .I rfd
145 will be generated because of the write done in
146 .B 2
147 and the event is consumed in
148 .BR 3 .
149 Since the read operation done in
150 .B 4
151 does not consume the whole buffer data, the call to
152 .BR epoll_wait (2)
153 done in step
154 .B 5
155 might block indefinitely.
156 .PP
157 An application that employs the
158 .B EPOLLET
159 flag should use nonblocking file descriptors to avoid having a blocking
160 read or write starve a task that is handling multiple file descriptors.
161 The suggested way to use
162 .B epoll
163 as an edge-triggered
164 .RB ( EPOLLET )
165 interface is as follows:
166 .IP a) 3
167 with nonblocking file descriptors; and
168 .IP b)
169 by waiting for an event only after
170 .BR read (2)
171 or
172 .BR write (2)
173 return
174 .BR EAGAIN .
175 .PP
176 By contrast, when used as a level-triggered interface
177 (the default, when
178 .B EPOLLET
179 is not specified),
180 .B epoll
181 is simply a faster
182 .BR poll (2),
183 and can be used wherever the latter is used since it shares the
184 same semantics.
185 .PP
186 Since even with edge-triggered
187 .BR epoll ,
188 multiple events can be generated upon receipt of multiple chunks of data,
189 the caller has the option to specify the
190 .B EPOLLONESHOT
191 flag, to tell
192 .B epoll
193 to disable the associated file descriptor after the receipt of an event with
194 .BR epoll_wait (2).
195 When the
196 .B EPOLLONESHOT
197 flag is specified,
198 it is the caller's responsibility to rearm the file descriptor using
199 .BR epoll_ctl (2)
200 with
201 .BR EPOLL_CTL_MOD .
202 .PP
203 If multiple threads
204 (or processes, if child processes have inherited the
205 .B epoll
206 file descriptor across
207 .BR fork (2))
208 are blocked in
209 .BR epoll_wait (2)
210 waiting on the same epoll file descriptor and a file descriptor
211 in the interest list that is marked for edge-triggered
212 .RB ( EPOLLET )
213 notification becomes ready,
214 just one of the threads (or processes) is awoken from
215 .BR epoll_wait (2).
216 This provides a useful optimization for avoiding "thundering herd" wake-ups
217 in some scenarios.
218 .\"
219 .SS Interaction with autosleep
220 If the system is in
221 .B autosleep
222 mode via
223 .I /sys/power/autosleep
224 and an event happens which wakes the device from sleep, the device
225 driver will keep the device awake only until that event is queued.
226 To keep the device awake until the event has been processed,
227 it is necessary to use the
228 .BR epoll_ctl (2)
229 .B EPOLLWAKEUP
230 flag.
231 .PP
232 When the
233 .B EPOLLWAKEUP
234 flag is set in the
235 .B events
236 field for a
237 .IR "struct epoll_event" ,
238 the system will be kept awake from the moment the event is queued,
239 through the
240 .BR epoll_wait (2)
241 call which returns the event until the subsequent
242 .BR epoll_wait (2)
243 call.
244 If the event should keep the system awake beyond that time,
245 then a separate
246 .I wake_lock
247 should be taken before the second
248 .BR epoll_wait (2)
249 call.
250 .SS /proc interfaces
251 The following interfaces can be used to limit the amount of
252 kernel memory consumed by epoll:
253 .\" Following was added in 2.6.28, but them removed in 2.6.29
254 .\" .TP
255 .\" .IR /proc/sys/fs/epoll/max_user_instances " (since Linux 2.6.28)"
256 .\" This specifies an upper limit on the number of epoll instances
257 .\" that can be created per real user ID.
258 .TP
259 .IR /proc/sys/fs/epoll/max_user_watches " (since Linux 2.6.28)"
260 This specifies a limit on the total number of
261 file descriptors that a user can register across
262 all epoll instances on the system.
263 The limit is per real user ID.
264 Each registered file descriptor costs roughly 90 bytes on a 32-bit kernel,
265 and roughly 160 bytes on a 64-bit kernel.
266 Currently,
267 .\" 2.6.29 (in 2.6.28, the default was 1/32 of lowmem)
268 the default value for
269 .I max_user_watches
270 is 1/25 (4%) of the available low memory,
271 divided by the registration cost in bytes.
272 .SS Example for suggested usage
273 While the usage of
274 .B epoll
275 when employed as a level-triggered interface does have the same
276 semantics as
277 .BR poll (2),
278 the edge-triggered usage requires more clarification to avoid stalls
279 in the application event loop.
280 In this example, listener is a
281 nonblocking socket on which
282 .BR listen (2)
283 has been called.
284 The function
285 .I do_use_fd()
286 uses the new ready file descriptor until
287 .B EAGAIN
288 is returned by either
289 .BR read (2)
290 or
291 .BR write (2).
292 An event-driven state machine application should, after having received
293 .BR EAGAIN ,
294 record its current state so that at the next call to
295 .I do_use_fd()
296 it will continue to
297 .BR read (2)
298 or
299 .BR write (2)
300 from where it stopped before.
301 .PP
302 .in +4n
303 .EX
304 #define MAX_EVENTS 10
305 struct epoll_event ev, events[MAX_EVENTS];
306 int listen_sock, conn_sock, nfds, epollfd;
308 /* Code to set up listening socket, \(aqlisten_sock\(aq,
309 (socket(), bind(), listen()) omitted. */
311 epollfd = epoll_create1(0);
312 if (epollfd == \-1) {
313 perror("epoll_create1");
314 exit(EXIT_FAILURE);
315 }
317 ev.events = EPOLLIN;
318 ev.data.fd = listen_sock;
319 if (epoll_ctl(epollfd, EPOLL_CTL_ADD, listen_sock, &ev) == \-1) {
320 perror("epoll_ctl: listen_sock");
321 exit(EXIT_FAILURE);
322 }
324 for (;;) {
325 nfds = epoll_wait(epollfd, events, MAX_EVENTS, \-1);
326 if (nfds == \-1) {
327 perror("epoll_wait");
328 exit(EXIT_FAILURE);
329 }
331 for (n = 0; n < nfds; ++n) {
332 if (events[n].data.fd == listen_sock) {
333 conn_sock = accept(listen_sock,
334 (struct sockaddr *) &addr, &addrlen);
335 if (conn_sock == \-1) {
336 perror("accept");
337 exit(EXIT_FAILURE);
338 }
339 setnonblocking(conn_sock);
340 ev.events = EPOLLIN | EPOLLET;
341 ev.data.fd = conn_sock;
342 if (epoll_ctl(epollfd, EPOLL_CTL_ADD, conn_sock,
343 &ev) == \-1) {
344 perror("epoll_ctl: conn_sock");
345 exit(EXIT_FAILURE);
346 }
347 } else {
348 do_use_fd(events[n].data.fd);
349 }
350 }
351 }
352 .EE
353 .in
354 .PP
355 When used as an edge-triggered interface, for performance reasons, it is
356 possible to add the file descriptor inside the
357 .B epoll
358 interface
359 .RB ( EPOLL_CTL_ADD )
360 once by specifying
361 .RB ( EPOLLIN | EPOLLOUT ).
362 This allows you to avoid
363 continuously switching between
364 .B EPOLLIN
365 and
366 .B EPOLLOUT
367 calling
368 .BR epoll_ctl (2)
369 with
370 .BR EPOLL_CTL_MOD .
371 .SS Questions and answers
372 .IP 0. 4
373 What is the key used to distinguish the file descriptors registered in an
374 interest list?
375 .IP
376 The key is the combination of the file descriptor number and
377 the open file description
378 (also known as an "open file handle",
379 the kernel's internal representation of an open file).
380 .IP 1.
381 What happens if you register the same file descriptor on an
382 .B epoll
383 instance twice?
384 .IP
385 You will probably get
386 .BR EEXIST .
387 However, it is possible to add a duplicate
388 .RB ( dup (2),
389 .BR dup2 (2),
390 .BR fcntl (2)
391 .BR F_DUPFD )
392 file descriptor to the same
393 .B epoll
394 instance.
395 .\" But a file descriptor duplicated by fork(2) can't be added to the
396 .\" set, because the [file *, fd] pair is already in the epoll set.
397 .\" That is a somewhat ugly inconsistency. On the one hand, a child process
398 .\" cannot add the duplicate file descriptor to the epoll set. (In every
399 .\" other case that I can think of, file descriptors duplicated by fork have
400 .\" similar semantics to file descriptors duplicated by dup() and friends.) On
401 .\" the other hand, the very fact that the child has a duplicate of the
402 .\" file descriptor means that even if the parent closes its file descriptor,
403 .\" then epoll_wait() in the parent will continue to receive notifications for
404 .\" that file descriptor because of the duplicated file descriptor in the child.
405 .\"
406 .\" See http://thread.gmane.org/gmane.linux.kernel/596462/
407 .\" "epoll design problems with common fork/exec patterns"
408 .\"
409 .\" mtk, Feb 2008
410 This can be a useful technique for filtering events,
411 if the duplicate file descriptors are registered with different
412 .I events
413 masks.
414 .IP 2.
415 Can two
416 .B epoll
417 instances wait for the same file descriptor?
418 If so, are events reported to both
419 .B epoll
420 file descriptors?
421 .IP
422 Yes, and events would be reported to both.
423 However, careful programming may be needed to do this correctly.
424 .IP 3.
425 Is the
426 .B epoll
427 file descriptor itself poll/epoll/selectable?
428 .IP
429 Yes.
430 If an
431 .B epoll
432 file descriptor has events waiting, then it will
433 indicate as being readable.
434 .IP 4.
435 What happens if one attempts to put an
436 .B epoll
437 file descriptor into its own file descriptor set?
438 .IP
439 The
440 .BR epoll_ctl (2)
441 call fails
442 .RB ( EINVAL ).
443 However, you can add an
444 .B epoll
445 file descriptor inside another
446 .B epoll
447 file descriptor set.
448 .IP 5.
449 Can I send an
450 .B epoll
451 file descriptor over a UNIX domain socket to another process?
452 .IP
453 Yes, but it does not make sense to do this, since the receiving process
454 would not have copies of the file descriptors in the interest list.
455 .IP 6.
456 Will closing a file descriptor cause it to be removed from all
457 .B epoll
458 interest lists?
459 .IP
460 Yes, but be aware of the following point.
461 A file descriptor is a reference to an open file description (see
462 .BR open (2)).
463 Whenever a file descriptor is duplicated via
464 .BR dup (2),
465 .BR dup2 (2),
466 .BR fcntl (2)
467 .BR F_DUPFD ,
468 or
469 .BR fork (2),
470 a new file descriptor referring to the same open file description is
471 created.
472 An open file description continues to exist until all
473 file descriptors referring to it have been closed.
474 .IP
475 A file descriptor is removed from an
476 interest list only after all the file descriptors referring to the underlying
477 open file description have been closed.
478 This means that even after a file descriptor that is part of an
479 interest list has been closed,
480 events may be reported for that file descriptor if other file
481 descriptors referring to the same underlying file description remain open.
482 To prevent this happening,
483 the file descriptor must be explicitly removed from the interest list (using
484 .BR epoll_ctl (2)
485 .BR EPOLL_CTL_DEL )
486 before it is duplicated.
487 Alternatively,
488 the application must ensure that all file descriptors are closed
489 (which may be difficult if file descriptors were duplicated
490 behind the scenes by library functions that used
491 .BR dup (2)
492 or
493 .BR fork (2)).
494 .IP 7.
495 If more than one event occurs between
496 .BR epoll_wait (2)
497 calls, are they combined or reported separately?
498 .IP
499 They will be combined.
500 .IP 8.
501 Does an operation on a file descriptor affect the
502 already collected but not yet reported events?
503 .IP
504 You can do two operations on an existing file descriptor.
505 Remove would be meaningless for
506 this case.
507 Modify will reread available I/O.
508 .IP 9.
509 Do I need to continuously read/write a file descriptor
510 until
511 .B EAGAIN
512 when using the
513 .B EPOLLET
514 flag (edge-triggered behavior)?
515 .IP
516 Receiving an event from
517 .BR epoll_wait (2)
518 should suggest to you that such
519 file descriptor is ready for the requested I/O operation.
520 You must consider it ready until the next (nonblocking)
521 read/write yields
522 .BR EAGAIN .
523 When and how you will use the file descriptor is entirely up to you.
524 .IP
525 For packet/token-oriented files (e.g., datagram socket,
526 terminal in canonical mode),
527 the only way to detect the end of the read/write I/O space
528 is to continue to read/write until
529 .BR EAGAIN .
530 .IP
531 For stream-oriented files (e.g., pipe, FIFO, stream socket), the
532 condition that the read/write I/O space is exhausted can also be detected by
533 checking the amount of data read from / written to the target file
534 descriptor.
535 For example, if you call
536 .BR read (2)
537 by asking to read a certain amount of data and
538 .BR read (2)
539 returns a lower number of bytes, you
540 can be sure of having exhausted the read I/O space for the file
541 descriptor.
542 The same is true when writing using
543 .BR write (2).
544 (Avoid this latter technique if you cannot guarantee that
545 the monitored file descriptor always refers to a stream-oriented file.)
546 .SS Possible pitfalls and ways to avoid them
547 .TP
548 .B o Starvation (edge-triggered)
549 .PP
550 If there is a large amount of I/O space,
551 it is possible that by trying to drain
552 it the other files will not get processed causing starvation.
553 (This problem is not specific to
554 .BR epoll .)
555 .PP
556 The solution is to maintain a ready list
557 and mark the file descriptor as ready
558 in its associated data structure, thereby allowing the application to
559 remember which files need to be processed but still round robin amongst
560 all the ready files.
561 This also supports ignoring subsequent events you
562 receive for file descriptors that are already ready.
563 .TP
564 .B o If using an event cache...
565 .PP
566 If you use an event cache or store all the file descriptors returned from
567 .BR epoll_wait (2),
568 then make sure to provide a way to mark
569 its closure dynamically (i.e., caused by
570 a previous event's processing).
571 Suppose you receive 100 events from
572 .BR epoll_wait (2),
573 and in event #47 a condition causes event #13 to be closed.
574 If you remove the structure and
575 .BR close (2)
576 the file descriptor for event #13, then your
577 event cache might still say there are events waiting for that
578 file descriptor causing confusion.
579 .PP
580 One solution for this is to call, during the processing of event 47,
581 .BR epoll_ctl ( EPOLL_CTL_DEL )
582 to delete file descriptor 13 and
583 .BR close (2),
584 then mark its associated
585 data structure as removed and link it to a cleanup list.
586 If you find another
587 event for file descriptor 13 in your batch processing,
588 you will discover the file descriptor had been
589 previously removed and there will be no confusion.
590 .SH VERSIONS
591 The
592 .B epoll
593 API was introduced in Linux kernel 2.5.44.
594 .\" Its interface should be finalized in Linux kernel 2.5.66.
595 Support was added to glibc in version 2.3.2.
596 .SH CONFORMING TO
597 The
598 .B epoll
599 API is Linux-specific.
600 Some other systems provide similar
601 mechanisms, for example, FreeBSD has
602 .IR kqueue ,
603 and Solaris has
604 .IR /dev/poll .
605 .SH NOTES
606 The set of file descriptors that is being monitored via
607 an epoll file descriptor can be viewed via the entry for
608 the epoll file descriptor in the process's
609 .IR /proc/[pid]/fdinfo
610 directory.
611 See
612 .BR proc (5)
613 for further details.
614 .PP
615 The
616 .BR kcmp (2)
617 .B KCMP_EPOLL_TFD
618 operation can be used to test whether a file descriptor
619 is present in an epoll instance.
620 .SH SEE ALSO
621 .BR epoll_create (2),
622 .BR epoll_create1 (2),
623 .BR epoll_ctl (2),
624 .BR epoll_wait (2),
625 .BR poll (2),
626 .BR select (2)