2
* fs/eventpoll.c (Efficient event retrieval implementation)
3
* Copyright (C) 2001,...,2009 Davide Libenzi
5
* This program is free software; you can redistribute it and/or modify
6
* it under the terms of the GNU General Public License as published by
7
* the Free Software Foundation; either version 2 of the License, or
8
* (at your option) any later version.
10
* Davide Libenzi <davidel@xmailserver.org>
14
#include <linux/init.h>
15
#include <linux/kernel.h>
16
#include <linux/sched.h>
18
#include <linux/file.h>
19
#include <linux/signal.h>
20
#include <linux/errno.h>
22
#include <linux/slab.h>
23
#include <linux/poll.h>
24
#include <linux/string.h>
25
#include <linux/list.h>
26
#include <linux/hash.h>
27
#include <linux/spinlock.h>
28
#include <linux/syscalls.h>
29
#include <linux/rbtree.h>
30
#include <linux/wait.h>
31
#include <linux/eventpoll.h>
32
#include <linux/mount.h>
33
#include <linux/bitops.h>
34
#include <linux/mutex.h>
35
#include <linux/anon_inodes.h>
36
#include <asm/uaccess.h>
37
#include <asm/system.h>
40
#include <linux/atomic.h>
44
* There are three level of locking required by epoll :
48
* 3) ep->lock (spinlock)
50
* The acquire order is the one listed above, from 1 to 3.
51
* We need a spinlock (ep->lock) because we manipulate objects
52
* from inside the poll callback, that might be triggered from
53
* a wake_up() that in turn might be called from IRQ context.
54
* So we can't sleep inside the poll callback and hence we need
55
* a spinlock. During the event transfer loop (from kernel to
56
* user space) we could end up sleeping due a copy_to_user(), so
57
* we need a lock that will allow us to sleep. This lock is a
58
* mutex (ep->mtx). It is acquired during the event transfer loop,
59
* during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
60
* Then we also need a global mutex to serialize eventpoll_release_file()
62
* This mutex is acquired by ep_free() during the epoll file
63
* cleanup path and it is also acquired by eventpoll_release_file()
64
* if a file has been pushed inside an epoll set and it is then
65
* close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
66
* It is also acquired when inserting an epoll fd onto another epoll
67
* fd. We do this so that we walk the epoll tree and ensure that this
68
* insertion does not create a cycle of epoll file descriptors, which
69
* could lead to deadlock. We need a global mutex to prevent two
70
* simultaneous inserts (A into B and B into A) from racing and
71
* constructing a cycle without either insert observing that it is
73
* It is necessary to acquire multiple "ep->mtx"es at once in the
74
* case when one epoll fd is added to another. In this case, we
75
* always acquire the locks in the order of nesting (i.e. after
76
* epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
77
* before e2->mtx). Since we disallow cycles of epoll file
78
* descriptors, this ensures that the mutexes are well-ordered. In
79
* order to communicate this nesting to lockdep, when walking a tree
80
* of epoll file descriptors, we use the current recursion depth as
82
* It is possible to drop the "ep->mtx" and to use the global
83
* mutex "epmutex" (together with "ep->lock") to have it working,
84
* but having "ep->mtx" will make the interface more scalable.
85
* Events that require holding "epmutex" are very rare, while for
86
* normal operations the epoll private "ep->mtx" will guarantee
87
* a better scalability.
90
/* Epoll private bits inside the event mask */
91
#define EP_PRIVATE_BITS (EPOLLONESHOT | EPOLLET)
93
/* Maximum number of nesting allowed inside epoll sets */
94
#define EP_MAX_NESTS 4
96
#define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
98
#define EP_UNACTIVE_PTR ((void *) -1L)
100
#define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
102
struct epoll_filefd {
108
* Structure used to track possible nested calls, for too deep recursions
111
struct nested_call_node {
112
struct list_head llink;
118
* This structure is used as collector for nested calls, to check for
119
* maximum recursion dept and loop cycles.
121
struct nested_calls {
122
struct list_head tasks_call_list;
127
* Each file descriptor added to the eventpoll interface will
128
* have an entry of this type linked to the "rbr" RB tree.
131
/* RB tree node used to link this structure to the eventpoll RB tree */
134
/* List header used to link this structure to the eventpoll ready list */
135
struct list_head rdllink;
138
* Works together "struct eventpoll"->ovflist in keeping the
139
* single linked chain of items.
143
/* The file descriptor information this item refers to */
144
struct epoll_filefd ffd;
146
/* Number of active wait queue attached to poll operations */
149
/* List containing poll wait queues */
150
struct list_head pwqlist;
152
/* The "container" of this item */
153
struct eventpoll *ep;
155
/* List header used to link this item to the "struct file" items list */
156
struct list_head fllink;
158
/* The structure that describe the interested events and the source fd */
159
struct epoll_event event;
163
* This structure is stored inside the "private_data" member of the file
164
* structure and represents the main data structure for the eventpoll
168
/* Protect the access to this structure */
172
* This mutex is used to ensure that files are not removed
173
* while epoll is using them. This is held during the event
174
* collection loop, the file cleanup path, the epoll file exit
175
* code and the ctl operations.
179
/* Wait queue used by sys_epoll_wait() */
180
wait_queue_head_t wq;
182
/* Wait queue used by file->poll() */
183
wait_queue_head_t poll_wait;
185
/* List of ready file descriptors */
186
struct list_head rdllist;
188
/* RB tree root used to store monitored fd structs */
192
* This is a single linked list that chains all the "struct epitem" that
193
* happened while transferring ready events to userspace w/out
196
struct epitem *ovflist;
198
/* The user that created the eventpoll descriptor */
199
struct user_struct *user;
202
/* Wait structure used by the poll hooks */
203
struct eppoll_entry {
204
/* List header used to link this structure to the "struct epitem" */
205
struct list_head llink;
207
/* The "base" pointer is set to the container "struct epitem" */
211
* Wait queue item that will be linked to the target file wait
216
/* The wait queue head that linked the "wait" wait queue item */
217
wait_queue_head_t *whead;
220
/* Wrapper struct used by poll queueing */
226
/* Used by the ep_send_events() function as callback private data */
227
struct ep_send_events_data {
229
struct epoll_event __user *events;
233
* Configuration options available inside /proc/sys/fs/epoll/
235
/* Maximum number of epoll watched descriptors, per user */
236
static long max_user_watches __read_mostly;
239
* This mutex is used to serialize ep_free() and eventpoll_release_file().
241
static DEFINE_MUTEX(epmutex);
243
/* Used to check for epoll file descriptor inclusion loops */
244
static struct nested_calls poll_loop_ncalls;
246
/* Used for safe wake up implementation */
247
static struct nested_calls poll_safewake_ncalls;
249
/* Used to call file's f_op->poll() under the nested calls boundaries */
250
static struct nested_calls poll_readywalk_ncalls;
252
/* Slab cache used to allocate "struct epitem" */
253
static struct kmem_cache *epi_cache __read_mostly;
255
/* Slab cache used to allocate "struct eppoll_entry" */
256
static struct kmem_cache *pwq_cache __read_mostly;
260
#include <linux/sysctl.h>
263
static long long_max = LONG_MAX;
265
ctl_table epoll_table[] = {
267
.procname = "max_user_watches",
268
.data = &max_user_watches,
269
.maxlen = sizeof(max_user_watches),
271
.proc_handler = proc_doulongvec_minmax,
277
#endif /* CONFIG_SYSCTL */
280
/* Setup the structure that is used as key for the RB tree */
281
static inline void ep_set_ffd(struct epoll_filefd *ffd,
282
struct file *file, int fd)
288
/* Compare RB tree keys */
289
static inline int ep_cmp_ffd(struct epoll_filefd *p1,
290
struct epoll_filefd *p2)
292
return (p1->file > p2->file ? +1:
293
(p1->file < p2->file ? -1 : p1->fd - p2->fd));
296
/* Tells us if the item is currently linked */
297
static inline int ep_is_linked(struct list_head *p)
299
return !list_empty(p);
302
/* Get the "struct epitem" from a wait queue pointer */
303
static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
305
return container_of(p, struct eppoll_entry, wait)->base;
308
/* Get the "struct epitem" from an epoll queue wrapper */
309
static inline struct epitem *ep_item_from_epqueue(poll_table *p)
311
return container_of(p, struct ep_pqueue, pt)->epi;
314
/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
315
static inline int ep_op_has_event(int op)
317
return op != EPOLL_CTL_DEL;
320
/* Initialize the poll safe wake up structure */
321
static void ep_nested_calls_init(struct nested_calls *ncalls)
323
INIT_LIST_HEAD(&ncalls->tasks_call_list);
324
spin_lock_init(&ncalls->lock);
328
* ep_events_available - Checks if ready events might be available.
330
* @ep: Pointer to the eventpoll context.
332
* Returns: Returns a value different than zero if ready events are available,
335
static inline int ep_events_available(struct eventpoll *ep)
337
return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
341
* ep_call_nested - Perform a bound (possibly) nested call, by checking
342
* that the recursion limit is not exceeded, and that
343
* the same nested call (by the meaning of same cookie) is
346
* @ncalls: Pointer to the nested_calls structure to be used for this call.
347
* @max_nests: Maximum number of allowed nesting calls.
348
* @nproc: Nested call core function pointer.
349
* @priv: Opaque data to be passed to the @nproc callback.
350
* @cookie: Cookie to be used to identify this nested call.
351
* @ctx: This instance context.
353
* Returns: Returns the code returned by the @nproc callback, or -1 if
354
* the maximum recursion limit has been exceeded.
356
static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
357
int (*nproc)(void *, void *, int), void *priv,
358
void *cookie, void *ctx)
360
int error, call_nests = 0;
362
struct list_head *lsthead = &ncalls->tasks_call_list;
363
struct nested_call_node *tncur;
364
struct nested_call_node tnode;
366
spin_lock_irqsave(&ncalls->lock, flags);
369
* Try to see if the current task is already inside this wakeup call.
370
* We use a list here, since the population inside this set is always
373
list_for_each_entry(tncur, lsthead, llink) {
374
if (tncur->ctx == ctx &&
375
(tncur->cookie == cookie || ++call_nests > max_nests)) {
377
* Ops ... loop detected or maximum nest level reached.
378
* We abort this wake by breaking the cycle itself.
385
/* Add the current task and cookie to the list */
387
tnode.cookie = cookie;
388
list_add(&tnode.llink, lsthead);
390
spin_unlock_irqrestore(&ncalls->lock, flags);
392
/* Call the nested function */
393
error = (*nproc)(priv, cookie, call_nests);
395
/* Remove the current task from the list */
396
spin_lock_irqsave(&ncalls->lock, flags);
397
list_del(&tnode.llink);
399
spin_unlock_irqrestore(&ncalls->lock, flags);
404
#ifdef CONFIG_DEBUG_LOCK_ALLOC
405
static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
406
unsigned long events, int subclass)
410
spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
411
wake_up_locked_poll(wqueue, events);
412
spin_unlock_irqrestore(&wqueue->lock, flags);
415
static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
416
unsigned long events, int subclass)
418
wake_up_poll(wqueue, events);
422
static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
424
ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
430
* Perform a safe wake up of the poll wait list. The problem is that
431
* with the new callback'd wake up system, it is possible that the
432
* poll callback is reentered from inside the call to wake_up() done
433
* on the poll wait queue head. The rule is that we cannot reenter the
434
* wake up code from the same task more than EP_MAX_NESTS times,
435
* and we cannot reenter the same wait queue head at all. This will
436
* enable to have a hierarchy of epoll file descriptor of no more than
439
static void ep_poll_safewake(wait_queue_head_t *wq)
441
int this_cpu = get_cpu();
443
ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
444
ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
450
* This function unregisters poll callbacks from the associated file
451
* descriptor. Must be called with "mtx" held (or "epmutex" if called from
454
static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
456
struct list_head *lsthead = &epi->pwqlist;
457
struct eppoll_entry *pwq;
459
while (!list_empty(lsthead)) {
460
pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
462
list_del(&pwq->llink);
463
remove_wait_queue(pwq->whead, &pwq->wait);
464
kmem_cache_free(pwq_cache, pwq);
469
* ep_scan_ready_list - Scans the ready list in a way that makes possible for
470
* the scan code, to call f_op->poll(). Also allows for
471
* O(NumReady) performance.
473
* @ep: Pointer to the epoll private data structure.
474
* @sproc: Pointer to the scan callback.
475
* @priv: Private opaque data passed to the @sproc callback.
476
* @depth: The current depth of recursive f_op->poll calls.
478
* Returns: The same integer error code returned by the @sproc callback.
480
static int ep_scan_ready_list(struct eventpoll *ep,
481
int (*sproc)(struct eventpoll *,
482
struct list_head *, void *),
486
int error, pwake = 0;
488
struct epitem *epi, *nepi;
492
* We need to lock this because we could be hit by
493
* eventpoll_release_file() and epoll_ctl().
495
mutex_lock_nested(&ep->mtx, depth);
498
* Steal the ready list, and re-init the original one to the
499
* empty list. Also, set ep->ovflist to NULL so that events
500
* happening while looping w/out locks, are not lost. We cannot
501
* have the poll callback to queue directly on ep->rdllist,
502
* because we want the "sproc" callback to be able to do it
505
spin_lock_irqsave(&ep->lock, flags);
506
list_splice_init(&ep->rdllist, &txlist);
508
spin_unlock_irqrestore(&ep->lock, flags);
511
* Now call the callback function.
513
error = (*sproc)(ep, &txlist, priv);
515
spin_lock_irqsave(&ep->lock, flags);
517
* During the time we spent inside the "sproc" callback, some
518
* other events might have been queued by the poll callback.
519
* We re-insert them inside the main ready-list here.
521
for (nepi = ep->ovflist; (epi = nepi) != NULL;
522
nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
524
* We need to check if the item is already in the list.
525
* During the "sproc" callback execution time, items are
526
* queued into ->ovflist but the "txlist" might already
527
* contain them, and the list_splice() below takes care of them.
529
if (!ep_is_linked(&epi->rdllink))
530
list_add_tail(&epi->rdllink, &ep->rdllist);
533
* We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
534
* releasing the lock, events will be queued in the normal way inside
537
ep->ovflist = EP_UNACTIVE_PTR;
540
* Quickly re-inject items left on "txlist".
542
list_splice(&txlist, &ep->rdllist);
544
if (!list_empty(&ep->rdllist)) {
546
* Wake up (if active) both the eventpoll wait list and
547
* the ->poll() wait list (delayed after we release the lock).
549
if (waitqueue_active(&ep->wq))
550
wake_up_locked(&ep->wq);
551
if (waitqueue_active(&ep->poll_wait))
554
spin_unlock_irqrestore(&ep->lock, flags);
556
mutex_unlock(&ep->mtx);
558
/* We have to call this outside the lock */
560
ep_poll_safewake(&ep->poll_wait);
566
* Removes a "struct epitem" from the eventpoll RB tree and deallocates
567
* all the associated resources. Must be called with "mtx" held.
569
static int ep_remove(struct eventpoll *ep, struct epitem *epi)
572
struct file *file = epi->ffd.file;
575
* Removes poll wait queue hooks. We _have_ to do this without holding
576
* the "ep->lock" otherwise a deadlock might occur. This because of the
577
* sequence of the lock acquisition. Here we do "ep->lock" then the wait
578
* queue head lock when unregistering the wait queue. The wakeup callback
579
* will run by holding the wait queue head lock and will call our callback
580
* that will try to get "ep->lock".
582
ep_unregister_pollwait(ep, epi);
584
/* Remove the current item from the list of epoll hooks */
585
spin_lock(&file->f_lock);
586
if (ep_is_linked(&epi->fllink))
587
list_del_init(&epi->fllink);
588
spin_unlock(&file->f_lock);
590
rb_erase(&epi->rbn, &ep->rbr);
592
spin_lock_irqsave(&ep->lock, flags);
593
if (ep_is_linked(&epi->rdllink))
594
list_del_init(&epi->rdllink);
595
spin_unlock_irqrestore(&ep->lock, flags);
597
/* At this point it is safe to free the eventpoll item */
598
kmem_cache_free(epi_cache, epi);
600
atomic_long_dec(&ep->user->epoll_watches);
605
static void ep_free(struct eventpoll *ep)
610
/* We need to release all tasks waiting for these file */
611
if (waitqueue_active(&ep->poll_wait))
612
ep_poll_safewake(&ep->poll_wait);
615
* We need to lock this because we could be hit by
616
* eventpoll_release_file() while we're freeing the "struct eventpoll".
617
* We do not need to hold "ep->mtx" here because the epoll file
618
* is on the way to be removed and no one has references to it
619
* anymore. The only hit might come from eventpoll_release_file() but
620
* holding "epmutex" is sufficient here.
622
mutex_lock(&epmutex);
625
* Walks through the whole tree by unregistering poll callbacks.
627
for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
628
epi = rb_entry(rbp, struct epitem, rbn);
630
ep_unregister_pollwait(ep, epi);
634
* Walks through the whole tree by freeing each "struct epitem". At this
635
* point we are sure no poll callbacks will be lingering around, and also by
636
* holding "epmutex" we can be sure that no file cleanup code will hit
637
* us during this operation. So we can avoid the lock on "ep->lock".
639
while ((rbp = rb_first(&ep->rbr)) != NULL) {
640
epi = rb_entry(rbp, struct epitem, rbn);
644
mutex_unlock(&epmutex);
645
mutex_destroy(&ep->mtx);
650
static int ep_eventpoll_release(struct inode *inode, struct file *file)
652
struct eventpoll *ep = file->private_data;
660
static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
663
struct epitem *epi, *tmp;
665
list_for_each_entry_safe(epi, tmp, head, rdllink) {
666
if (epi->ffd.file->f_op->poll(epi->ffd.file, NULL) &
668
return POLLIN | POLLRDNORM;
671
* Item has been dropped into the ready list by the poll
672
* callback, but it's not actually ready, as far as
673
* caller requested events goes. We can remove it here.
675
list_del_init(&epi->rdllink);
682
static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
684
return ep_scan_ready_list(priv, ep_read_events_proc, NULL, call_nests + 1);
687
static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
690
struct eventpoll *ep = file->private_data;
692
/* Insert inside our poll wait queue */
693
poll_wait(file, &ep->poll_wait, wait);
696
* Proceed to find out if wanted events are really available inside
697
* the ready list. This need to be done under ep_call_nested()
698
* supervision, since the call to f_op->poll() done on listed files
699
* could re-enter here.
701
pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
702
ep_poll_readyevents_proc, ep, ep, current);
704
return pollflags != -1 ? pollflags : 0;
707
/* File callbacks that implement the eventpoll file behaviour */
708
static const struct file_operations eventpoll_fops = {
709
.release = ep_eventpoll_release,
710
.poll = ep_eventpoll_poll,
711
.llseek = noop_llseek,
714
/* Fast test to see if the file is an eventpoll file */
715
static inline int is_file_epoll(struct file *f)
717
return f->f_op == &eventpoll_fops;
721
* This is called from eventpoll_release() to unlink files from the eventpoll
722
* interface. We need to have this facility to cleanup correctly files that are
723
* closed without being removed from the eventpoll interface.
725
void eventpoll_release_file(struct file *file)
727
struct list_head *lsthead = &file->f_ep_links;
728
struct eventpoll *ep;
732
* We don't want to get "file->f_lock" because it is not
733
* necessary. It is not necessary because we're in the "struct file"
734
* cleanup path, and this means that no one is using this file anymore.
735
* So, for example, epoll_ctl() cannot hit here since if we reach this
736
* point, the file counter already went to zero and fget() would fail.
737
* The only hit might come from ep_free() but by holding the mutex
738
* will correctly serialize the operation. We do need to acquire
739
* "ep->mtx" after "epmutex" because ep_remove() requires it when called
740
* from anywhere but ep_free().
742
* Besides, ep_remove() acquires the lock, so we can't hold it here.
744
mutex_lock(&epmutex);
746
while (!list_empty(lsthead)) {
747
epi = list_first_entry(lsthead, struct epitem, fllink);
750
list_del_init(&epi->fllink);
751
mutex_lock_nested(&ep->mtx, 0);
753
mutex_unlock(&ep->mtx);
756
mutex_unlock(&epmutex);
759
static int ep_alloc(struct eventpoll **pep)
762
struct user_struct *user;
763
struct eventpoll *ep;
765
user = get_current_user();
767
ep = kzalloc(sizeof(*ep), GFP_KERNEL);
771
spin_lock_init(&ep->lock);
772
mutex_init(&ep->mtx);
773
init_waitqueue_head(&ep->wq);
774
init_waitqueue_head(&ep->poll_wait);
775
INIT_LIST_HEAD(&ep->rdllist);
777
ep->ovflist = EP_UNACTIVE_PTR;
790
* Search the file inside the eventpoll tree. The RB tree operations
791
* are protected by the "mtx" mutex, and ep_find() must be called with
794
static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
798
struct epitem *epi, *epir = NULL;
799
struct epoll_filefd ffd;
801
ep_set_ffd(&ffd, file, fd);
802
for (rbp = ep->rbr.rb_node; rbp; ) {
803
epi = rb_entry(rbp, struct epitem, rbn);
804
kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
819
* This is the callback that is passed to the wait queue wakeup
820
* mechanism. It is called by the stored file descriptors when they
821
* have events to report.
823
static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
827
struct epitem *epi = ep_item_from_wait(wait);
828
struct eventpoll *ep = epi->ep;
830
spin_lock_irqsave(&ep->lock, flags);
833
* If the event mask does not contain any poll(2) event, we consider the
834
* descriptor to be disabled. This condition is likely the effect of the
835
* EPOLLONESHOT bit that disables the descriptor when an event is received,
836
* until the next EPOLL_CTL_MOD will be issued.
838
if (!(epi->event.events & ~EP_PRIVATE_BITS))
842
* Check the events coming with the callback. At this stage, not
843
* every device reports the events in the "key" parameter of the
844
* callback. We need to be able to handle both cases here, hence the
845
* test for "key" != NULL before the event match test.
847
if (key && !((unsigned long) key & epi->event.events))
851
* If we are transferring events to userspace, we can hold no locks
852
* (because we're accessing user memory, and because of linux f_op->poll()
853
* semantics). All the events that happen during that period of time are
854
* chained in ep->ovflist and requeued later on.
856
if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
857
if (epi->next == EP_UNACTIVE_PTR) {
858
epi->next = ep->ovflist;
864
/* If this file is already in the ready list we exit soon */
865
if (!ep_is_linked(&epi->rdllink))
866
list_add_tail(&epi->rdllink, &ep->rdllist);
869
* Wake up ( if active ) both the eventpoll wait list and the ->poll()
872
if (waitqueue_active(&ep->wq))
873
wake_up_locked(&ep->wq);
874
if (waitqueue_active(&ep->poll_wait))
878
spin_unlock_irqrestore(&ep->lock, flags);
880
/* We have to call this outside the lock */
882
ep_poll_safewake(&ep->poll_wait);
888
* This is the callback that is used to add our wait queue to the
889
* target file wakeup lists.
891
static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
894
struct epitem *epi = ep_item_from_epqueue(pt);
895
struct eppoll_entry *pwq;
897
if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
898
init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
901
add_wait_queue(whead, &pwq->wait);
902
list_add_tail(&pwq->llink, &epi->pwqlist);
905
/* We have to signal that an error occurred */
910
static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
913
struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
918
epic = rb_entry(parent, struct epitem, rbn);
919
kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
921
p = &parent->rb_right;
923
p = &parent->rb_left;
925
rb_link_node(&epi->rbn, parent, p);
926
rb_insert_color(&epi->rbn, &ep->rbr);
930
* Must be called with "mtx" held.
932
static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
933
struct file *tfile, int fd)
935
int error, revents, pwake = 0;
939
struct ep_pqueue epq;
941
user_watches = atomic_long_read(&ep->user->epoll_watches);
942
if (unlikely(user_watches >= max_user_watches))
944
if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
947
/* Item initialization follow here ... */
948
INIT_LIST_HEAD(&epi->rdllink);
949
INIT_LIST_HEAD(&epi->fllink);
950
INIT_LIST_HEAD(&epi->pwqlist);
952
ep_set_ffd(&epi->ffd, tfile, fd);
955
epi->next = EP_UNACTIVE_PTR;
957
/* Initialize the poll table using the queue callback */
959
init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
962
* Attach the item to the poll hooks and get current event bits.
963
* We can safely use the file* here because its usage count has
964
* been increased by the caller of this function. Note that after
965
* this operation completes, the poll callback can start hitting
968
revents = tfile->f_op->poll(tfile, &epq.pt);
971
* We have to check if something went wrong during the poll wait queue
972
* install process. Namely an allocation for a wait queue failed due
973
* high memory pressure.
977
goto error_unregister;
979
/* Add the current item to the list of active epoll hook for this file */
980
spin_lock(&tfile->f_lock);
981
list_add_tail(&epi->fllink, &tfile->f_ep_links);
982
spin_unlock(&tfile->f_lock);
985
* Add the current item to the RB tree. All RB tree operations are
986
* protected by "mtx", and ep_insert() is called with "mtx" held.
988
ep_rbtree_insert(ep, epi);
990
/* We have to drop the new item inside our item list to keep track of it */
991
spin_lock_irqsave(&ep->lock, flags);
993
/* If the file is already "ready" we drop it inside the ready list */
994
if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
995
list_add_tail(&epi->rdllink, &ep->rdllist);
997
/* Notify waiting tasks that events are available */
998
if (waitqueue_active(&ep->wq))
999
wake_up_locked(&ep->wq);
1000
if (waitqueue_active(&ep->poll_wait))
1004
spin_unlock_irqrestore(&ep->lock, flags);
1006
atomic_long_inc(&ep->user->epoll_watches);
1008
/* We have to call this outside the lock */
1010
ep_poll_safewake(&ep->poll_wait);
1015
ep_unregister_pollwait(ep, epi);
1018
* We need to do this because an event could have been arrived on some
1019
* allocated wait queue. Note that we don't care about the ep->ovflist
1020
* list, since that is used/cleaned only inside a section bound by "mtx".
1021
* And ep_insert() is called with "mtx" held.
1023
spin_lock_irqsave(&ep->lock, flags);
1024
if (ep_is_linked(&epi->rdllink))
1025
list_del_init(&epi->rdllink);
1026
spin_unlock_irqrestore(&ep->lock, flags);
1028
kmem_cache_free(epi_cache, epi);
1034
* Modify the interest event mask by dropping an event if the new mask
1035
* has a match in the current file status. Must be called with "mtx" held.
1037
static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1040
unsigned int revents;
1043
* Set the new event interest mask before calling f_op->poll();
1044
* otherwise we might miss an event that happens between the
1045
* f_op->poll() call and the new event set registering.
1047
epi->event.events = event->events;
1048
epi->event.data = event->data; /* protected by mtx */
1051
* Get current event bits. We can safely use the file* here because
1052
* its usage count has been increased by the caller of this function.
1054
revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
1057
* If the item is "hot" and it is not registered inside the ready
1058
* list, push it inside.
1060
if (revents & event->events) {
1061
spin_lock_irq(&ep->lock);
1062
if (!ep_is_linked(&epi->rdllink)) {
1063
list_add_tail(&epi->rdllink, &ep->rdllist);
1065
/* Notify waiting tasks that events are available */
1066
if (waitqueue_active(&ep->wq))
1067
wake_up_locked(&ep->wq);
1068
if (waitqueue_active(&ep->poll_wait))
1071
spin_unlock_irq(&ep->lock);
1074
/* We have to call this outside the lock */
1076
ep_poll_safewake(&ep->poll_wait);
1081
static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1084
struct ep_send_events_data *esed = priv;
1086
unsigned int revents;
1088
struct epoll_event __user *uevent;
1091
* We can loop without lock because we are passed a task private list.
1092
* Items cannot vanish during the loop because ep_scan_ready_list() is
1093
* holding "mtx" during this call.
1095
for (eventcnt = 0, uevent = esed->events;
1096
!list_empty(head) && eventcnt < esed->maxevents;) {
1097
epi = list_first_entry(head, struct epitem, rdllink);
1099
list_del_init(&epi->rdllink);
1101
revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL) &
1105
* If the event mask intersect the caller-requested one,
1106
* deliver the event to userspace. Again, ep_scan_ready_list()
1107
* is holding "mtx", so no operations coming from userspace
1108
* can change the item.
1111
if (__put_user(revents, &uevent->events) ||
1112
__put_user(epi->event.data, &uevent->data)) {
1113
list_add(&epi->rdllink, head);
1114
return eventcnt ? eventcnt : -EFAULT;
1118
if (epi->event.events & EPOLLONESHOT)
1119
epi->event.events &= EP_PRIVATE_BITS;
1120
else if (!(epi->event.events & EPOLLET)) {
1122
* If this file has been added with Level
1123
* Trigger mode, we need to insert back inside
1124
* the ready list, so that the next call to
1125
* epoll_wait() will check again the events
1126
* availability. At this point, no one can insert
1127
* into ep->rdllist besides us. The epoll_ctl()
1128
* callers are locked out by
1129
* ep_scan_ready_list() holding "mtx" and the
1130
* poll callback will queue them in ep->ovflist.
1132
list_add_tail(&epi->rdllink, &ep->rdllist);
1140
static int ep_send_events(struct eventpoll *ep,
1141
struct epoll_event __user *events, int maxevents)
1143
struct ep_send_events_data esed;
1145
esed.maxevents = maxevents;
1146
esed.events = events;
1148
return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0);
1151
static inline struct timespec ep_set_mstimeout(long ms)
1153
struct timespec now, ts = {
1154
.tv_sec = ms / MSEC_PER_SEC,
1155
.tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1159
return timespec_add_safe(now, ts);
1163
* ep_poll - Retrieves ready events, and delivers them to the caller supplied
1166
* @ep: Pointer to the eventpoll context.
1167
* @events: Pointer to the userspace buffer where the ready events should be
1169
* @maxevents: Size (in terms of number of events) of the caller event buffer.
1170
* @timeout: Maximum timeout for the ready events fetch operation, in
1171
* milliseconds. If the @timeout is zero, the function will not block,
1172
* while if the @timeout is less than zero, the function will block
1173
* until at least one event has been retrieved (or an error
1176
* Returns: Returns the number of ready events which have been fetched, or an
1177
* error code, in case of error.
1179
static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1180
int maxevents, long timeout)
1182
int res = 0, eavail, timed_out = 0;
1183
unsigned long flags;
1186
ktime_t expires, *to = NULL;
1189
struct timespec end_time = ep_set_mstimeout(timeout);
1191
slack = select_estimate_accuracy(&end_time);
1193
*to = timespec_to_ktime(end_time);
1194
} else if (timeout == 0) {
1196
* Avoid the unnecessary trip to the wait queue loop, if the
1197
* caller specified a non blocking operation.
1200
spin_lock_irqsave(&ep->lock, flags);
1205
spin_lock_irqsave(&ep->lock, flags);
1207
if (!ep_events_available(ep)) {
1209
* We don't have any available event to return to the caller.
1210
* We need to sleep here, and we will be wake up by
1211
* ep_poll_callback() when events will become available.
1213
init_waitqueue_entry(&wait, current);
1214
__add_wait_queue_exclusive(&ep->wq, &wait);
1218
* We don't want to sleep if the ep_poll_callback() sends us
1219
* a wakeup in between. That's why we set the task state
1220
* to TASK_INTERRUPTIBLE before doing the checks.
1222
set_current_state(TASK_INTERRUPTIBLE);
1223
if (ep_events_available(ep) || timed_out)
1225
if (signal_pending(current)) {
1230
spin_unlock_irqrestore(&ep->lock, flags);
1231
if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
1234
spin_lock_irqsave(&ep->lock, flags);
1236
__remove_wait_queue(&ep->wq, &wait);
1238
set_current_state(TASK_RUNNING);
1241
/* Is it worth to try to dig for events ? */
1242
eavail = ep_events_available(ep);
1244
spin_unlock_irqrestore(&ep->lock, flags);
1247
* Try to transfer events to user space. In case we get 0 events and
1248
* there's still timeout left over, we go trying again in search of
1251
if (!res && eavail &&
1252
!(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1259
* ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1260
* API, to verify that adding an epoll file inside another
1261
* epoll structure, does not violate the constraints, in
1262
* terms of closed loops, or too deep chains (which can
1263
* result in excessive stack usage).
1265
* @priv: Pointer to the epoll file to be currently checked.
1266
* @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1267
* data structure pointer.
1268
* @call_nests: Current dept of the @ep_call_nested() call stack.
1270
* Returns: Returns zero if adding the epoll @file inside current epoll
1271
* structure @ep does not violate the constraints, or -1 otherwise.
1273
static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1276
struct file *file = priv;
1277
struct eventpoll *ep = file->private_data;
1278
struct rb_node *rbp;
1281
mutex_lock_nested(&ep->mtx, call_nests + 1);
1282
for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1283
epi = rb_entry(rbp, struct epitem, rbn);
1284
if (unlikely(is_file_epoll(epi->ffd.file))) {
1285
error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1286
ep_loop_check_proc, epi->ffd.file,
1287
epi->ffd.file->private_data, current);
1292
mutex_unlock(&ep->mtx);
1298
* ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1299
* another epoll file (represented by @ep) does not create
1300
* closed loops or too deep chains.
1302
* @ep: Pointer to the epoll private data structure.
1303
* @file: Pointer to the epoll file to be checked.
1305
* Returns: Returns zero if adding the epoll @file inside current epoll
1306
* structure @ep does not violate the constraints, or -1 otherwise.
1308
static int ep_loop_check(struct eventpoll *ep, struct file *file)
1310
return ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1311
ep_loop_check_proc, file, ep, current);
1315
* Open an eventpoll file descriptor.
1317
SYSCALL_DEFINE1(epoll_create1, int, flags)
1320
struct eventpoll *ep = NULL;
1322
/* Check the EPOLL_* constant for consistency. */
1323
BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1325
if (flags & ~EPOLL_CLOEXEC)
1328
* Create the internal data structure ("struct eventpoll").
1330
error = ep_alloc(&ep);
1334
* Creates all the items needed to setup an eventpoll file. That is,
1335
* a file structure and a free file descriptor.
1337
error = anon_inode_getfd("[eventpoll]", &eventpoll_fops, ep,
1338
O_RDWR | (flags & O_CLOEXEC));
1345
SYSCALL_DEFINE1(epoll_create, int, size)
1350
return sys_epoll_create1(0);
1354
* The following function implements the controller interface for
1355
* the eventpoll file that enables the insertion/removal/change of
1356
* file descriptors inside the interest set.
1358
SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1359
struct epoll_event __user *, event)
1362
int did_lock_epmutex = 0;
1363
struct file *file, *tfile;
1364
struct eventpoll *ep;
1366
struct epoll_event epds;
1369
if (ep_op_has_event(op) &&
1370
copy_from_user(&epds, event, sizeof(struct epoll_event)))
1373
/* Get the "struct file *" for the eventpoll file */
1379
/* Get the "struct file *" for the target file */
1384
/* The target file descriptor must support poll */
1386
if (!tfile->f_op || !tfile->f_op->poll)
1387
goto error_tgt_fput;
1390
* We have to check that the file structure underneath the file descriptor
1391
* the user passed to us _is_ an eventpoll file. And also we do not permit
1392
* adding an epoll file descriptor inside itself.
1395
if (file == tfile || !is_file_epoll(file))
1396
goto error_tgt_fput;
1399
* At this point it is safe to assume that the "private_data" contains
1400
* our own data structure.
1402
ep = file->private_data;
1405
* When we insert an epoll file descriptor, inside another epoll file
1406
* descriptor, there is the change of creating closed loops, which are
1407
* better be handled here, than in more critical paths.
1409
* We hold epmutex across the loop check and the insert in this case, in
1410
* order to prevent two separate inserts from racing and each doing the
1411
* insert "at the same time" such that ep_loop_check passes on both
1412
* before either one does the insert, thereby creating a cycle.
1414
if (unlikely(is_file_epoll(tfile) && op == EPOLL_CTL_ADD)) {
1415
mutex_lock(&epmutex);
1416
did_lock_epmutex = 1;
1418
if (ep_loop_check(ep, tfile) != 0)
1419
goto error_tgt_fput;
1423
mutex_lock_nested(&ep->mtx, 0);
1426
* Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1427
* above, we can be sure to be able to use the item looked up by
1428
* ep_find() till we release the mutex.
1430
epi = ep_find(ep, tfile, fd);
1436
epds.events |= POLLERR | POLLHUP;
1437
error = ep_insert(ep, &epds, tfile, fd);
1443
error = ep_remove(ep, epi);
1449
epds.events |= POLLERR | POLLHUP;
1450
error = ep_modify(ep, epi, &epds);
1455
mutex_unlock(&ep->mtx);
1458
if (unlikely(did_lock_epmutex))
1459
mutex_unlock(&epmutex);
1470
* Implement the event wait interface for the eventpoll file. It is the kernel
1471
* part of the user space epoll_wait(2).
1473
SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1474
int, maxevents, int, timeout)
1478
struct eventpoll *ep;
1480
/* The maximum number of event must be greater than zero */
1481
if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1484
/* Verify that the area passed by the user is writeable */
1485
if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
1490
/* Get the "struct file *" for the eventpoll file */
1497
* We have to check that the file structure underneath the fd
1498
* the user passed to us _is_ an eventpoll file.
1501
if (!is_file_epoll(file))
1505
* At this point it is safe to assume that the "private_data" contains
1506
* our own data structure.
1508
ep = file->private_data;
1510
/* Time to fish for events ... */
1511
error = ep_poll(ep, events, maxevents, timeout);
1520
#ifdef HAVE_SET_RESTORE_SIGMASK
1523
* Implement the event wait interface for the eventpoll file. It is the kernel
1524
* part of the user space epoll_pwait(2).
1526
SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
1527
int, maxevents, int, timeout, const sigset_t __user *, sigmask,
1531
sigset_t ksigmask, sigsaved;
1534
* If the caller wants a certain signal mask to be set during the wait,
1538
if (sigsetsize != sizeof(sigset_t))
1540
if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1542
sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1543
sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1546
error = sys_epoll_wait(epfd, events, maxevents, timeout);
1549
* If we changed the signal mask, we need to restore the original one.
1550
* In case we've got a signal while waiting, we do not restore the
1551
* signal mask yet, and we allow do_signal() to deliver the signal on
1552
* the way back to userspace, before the signal mask is restored.
1555
if (error == -EINTR) {
1556
memcpy(¤t->saved_sigmask, &sigsaved,
1558
set_restore_sigmask();
1560
sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1566
#endif /* HAVE_SET_RESTORE_SIGMASK */
1568
static int __init eventpoll_init(void)
1574
* Allows top 4% of lomem to be allocated for epoll watches (per user).
1576
max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
1578
BUG_ON(max_user_watches < 0);
1581
* Initialize the structure used to perform epoll file descriptor
1582
* inclusion loops checks.
1584
ep_nested_calls_init(&poll_loop_ncalls);
1586
/* Initialize the structure used to perform safe poll wait head wake ups */
1587
ep_nested_calls_init(&poll_safewake_ncalls);
1589
/* Initialize the structure used to perform file's f_op->poll() calls */
1590
ep_nested_calls_init(&poll_readywalk_ncalls);
1592
/* Allocates slab cache used to allocate "struct epitem" items */
1593
epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1594
0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
1596
/* Allocates slab cache used to allocate "struct eppoll_entry" */
1597
pwq_cache = kmem_cache_create("eventpoll_pwq",
1598
sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
1602
fs_initcall(eventpoll_init);