path: root/fs/eventpoll.c (plain)
blob: e76d0c3fa2c7d1c27b2279604fa919b76735b445

1	/*
2	* fs/eventpoll.c (Efficient event retrieval implementation)
3	* Copyright (C) 2001,...,2009 Davide Libenzi
4	*
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.
9	*
10	* Davide Libenzi <davidel@xmailserver.org>
11	*
12	*/
13
14	#include <linux/init.h>
15	#include <linux/kernel.h>
16	#include <linux/sched.h>
17	#include <linux/fs.h>
18	#include <linux/file.h>
19	#include <linux/signal.h>
20	#include <linux/errno.h>
21	#include <linux/mm.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 <linux/device.h>
37	#include <linux/freezer.h>
38	#include <asm/uaccess.h>
39	#include <asm/io.h>
40	#include <asm/mman.h>
41	#include <linux/atomic.h>
42	#include <linux/proc_fs.h>
43	#include <linux/seq_file.h>
44	#include <linux/compat.h>
45	#include <linux/rculist.h>
46
47	/*
48	* LOCKING:
49	* There are three level of locking required by epoll :
50	*
51	* 1) epmutex (mutex)
52	* 2) ep->mtx (mutex)
53	* 3) ep->lock (spinlock)
54	*
55	* The acquire order is the one listed above, from 1 to 3.
56	* We need a spinlock (ep->lock) because we manipulate objects
57	* from inside the poll callback, that might be triggered from
58	* a wake_up() that in turn might be called from IRQ context.
59	* So we can't sleep inside the poll callback and hence we need
60	* a spinlock. During the event transfer loop (from kernel to
61	* user space) we could end up sleeping due a copy_to_user(), so
62	* we need a lock that will allow us to sleep. This lock is a
63	* mutex (ep->mtx). It is acquired during the event transfer loop,
64	* during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
65	* Then we also need a global mutex to serialize eventpoll_release_file()
66	* and ep_free().
67	* This mutex is acquired by ep_free() during the epoll file
68	* cleanup path and it is also acquired by eventpoll_release_file()
69	* if a file has been pushed inside an epoll set and it is then
70	* close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
71	* It is also acquired when inserting an epoll fd onto another epoll
72	* fd. We do this so that we walk the epoll tree and ensure that this
73	* insertion does not create a cycle of epoll file descriptors, which
74	* could lead to deadlock. We need a global mutex to prevent two
75	* simultaneous inserts (A into B and B into A) from racing and
76	* constructing a cycle without either insert observing that it is
77	* going to.
78	* It is necessary to acquire multiple "ep->mtx"es at once in the
79	* case when one epoll fd is added to another. In this case, we
80	* always acquire the locks in the order of nesting (i.e. after
81	* epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
82	* before e2->mtx). Since we disallow cycles of epoll file
83	* descriptors, this ensures that the mutexes are well-ordered. In
84	* order to communicate this nesting to lockdep, when walking a tree
85	* of epoll file descriptors, we use the current recursion depth as
86	* the lockdep subkey.
87	* It is possible to drop the "ep->mtx" and to use the global
88	* mutex "epmutex" (together with "ep->lock") to have it working,
89	* but having "ep->mtx" will make the interface more scalable.
90	* Events that require holding "epmutex" are very rare, while for
91	* normal operations the epoll private "ep->mtx" will guarantee
92	* a better scalability.
93	*/
94
95	/* Epoll private bits inside the event mask */
96	#define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET | EPOLLEXCLUSIVE)
97
98	#define EPOLLINOUT_BITS (POLLIN | POLLOUT)
99
100	#define EPOLLEXCLUSIVE_OK_BITS (EPOLLINOUT_BITS | POLLERR | POLLHUP | \
101	EPOLLWAKEUP | EPOLLET | EPOLLEXCLUSIVE)
102
103	/* Maximum number of nesting allowed inside epoll sets */
104	#define EP_MAX_NESTS 4
105
106	#define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
107
108	#define EP_UNACTIVE_PTR ((void *) -1L)
109
110	#define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
111
112	struct epoll_filefd {
113	struct file *file;
114	int fd;
115	} __packed;
116
117	/*
118	* Structure used to track possible nested calls, for too deep recursions
119	* and loop cycles.
120	*/
121	struct nested_call_node {
122	struct list_head llink;
123	void *cookie;
124	void *ctx;
125	};
126
127	/*
128	* This structure is used as collector for nested calls, to check for
129	* maximum recursion dept and loop cycles.
130	*/
131	struct nested_calls {
132	struct list_head tasks_call_list;
133	spinlock_t lock;
134	};
135
136	/*
137	* Each file descriptor added to the eventpoll interface will
138	* have an entry of this type linked to the "rbr" RB tree.
139	* Avoid increasing the size of this struct, there can be many thousands
140	* of these on a server and we do not want this to take another cache line.
141	*/
142	struct epitem {
143	union {
144	/* RB tree node links this structure to the eventpoll RB tree */
145	struct rb_node rbn;
146	/* Used to free the struct epitem */
147	struct rcu_head rcu;
148	};
149
150	/* List header used to link this structure to the eventpoll ready list */
151	struct list_head rdllink;
152
153	/*
154	* Works together "struct eventpoll"->ovflist in keeping the
155	* single linked chain of items.
156	*/
157	struct epitem *next;
158
159	/* The file descriptor information this item refers to */
160	struct epoll_filefd ffd;
161
162	/* Number of active wait queue attached to poll operations */
163	int nwait;
164
165	/* List containing poll wait queues */
166	struct list_head pwqlist;
167
168	/* The "container" of this item */
169	struct eventpoll *ep;
170
171	/* List header used to link this item to the "struct file" items list */
172	struct list_head fllink;
173
174	/* wakeup_source used when EPOLLWAKEUP is set */
175	struct wakeup_source __rcu *ws;
176
177	/* The structure that describe the interested events and the source fd */
178	struct epoll_event event;
179	};
180
181	/*
182	* This structure is stored inside the "private_data" member of the file
183	* structure and represents the main data structure for the eventpoll
184	* interface.
185	*/
186	struct eventpoll {
187	/* Protect the access to this structure */
188	spinlock_t lock;
189
190	/*
191	* This mutex is used to ensure that files are not removed
192	* while epoll is using them. This is held during the event
193	* collection loop, the file cleanup path, the epoll file exit
194	* code and the ctl operations.
195	*/
196	struct mutex mtx;
197
198	/* Wait queue used by sys_epoll_wait() */
199	wait_queue_head_t wq;
200
201	/* Wait queue used by file->poll() */
202	wait_queue_head_t poll_wait;
203
204	/* List of ready file descriptors */
205	struct list_head rdllist;
206
207	/* RB tree root used to store monitored fd structs */
208	struct rb_root rbr;
209
210	/*
211	* This is a single linked list that chains all the "struct epitem" that
212	* happened while transferring ready events to userspace w/out
213	* holding ->lock.
214	*/
215	struct epitem *ovflist;
216
217	/* wakeup_source used when ep_scan_ready_list is running */
218	struct wakeup_source *ws;
219
220	/* The user that created the eventpoll descriptor */
221	struct user_struct *user;
222
223	struct file *file;
224
225	/* used to optimize loop detection check */
226	int visited;
227	struct list_head visited_list_link;
228	};
229
230	/* Wait structure used by the poll hooks */
231	struct eppoll_entry {
232	/* List header used to link this structure to the "struct epitem" */
233	struct list_head llink;
234
235	/* The "base" pointer is set to the container "struct epitem" */
236	struct epitem *base;
237
238	/*
239	* Wait queue item that will be linked to the target file wait
240	* queue head.
241	*/
242	wait_queue_t wait;
243
244	/* The wait queue head that linked the "wait" wait queue item */
245	wait_queue_head_t *whead;
246	};
247
248	/* Wrapper struct used by poll queueing */
249	struct ep_pqueue {
250	poll_table pt;
251	struct epitem *epi;
252	};
253
254	/* Used by the ep_send_events() function as callback private data */
255	struct ep_send_events_data {
256	int maxevents;
257	struct epoll_event __user *events;
258	};
259
260	/*
261	* Configuration options available inside /proc/sys/fs/epoll/
262	*/
263	/* Maximum number of epoll watched descriptors, per user */
264	static long max_user_watches __read_mostly;
265
266	/*
267	* This mutex is used to serialize ep_free() and eventpoll_release_file().
268	*/
269	static DEFINE_MUTEX(epmutex);
270
271	/* Used to check for epoll file descriptor inclusion loops */
272	static struct nested_calls poll_loop_ncalls;
273
274	/* Used for safe wake up implementation */
275	static struct nested_calls poll_safewake_ncalls;
276
277	/* Used to call file's f_op->poll() under the nested calls boundaries */
278	static struct nested_calls poll_readywalk_ncalls;
279
280	/* Slab cache used to allocate "struct epitem" */
281	static struct kmem_cache *epi_cache __read_mostly;
282
283	/* Slab cache used to allocate "struct eppoll_entry" */
284	static struct kmem_cache *pwq_cache __read_mostly;
285
286	/* Visited nodes during ep_loop_check(), so we can unset them when we finish */
287	static LIST_HEAD(visited_list);
288
289	/*
290	* List of files with newly added links, where we may need to limit the number
291	* of emanating paths. Protected by the epmutex.
292	*/
293	static LIST_HEAD(tfile_check_list);
294
295	#ifdef CONFIG_SYSCTL
296
297	#include <linux/sysctl.h>
298
299	static long zero;
300	static long long_max = LONG_MAX;
301
302	struct ctl_table epoll_table[] = {
303	{
304	.procname = "max_user_watches",
305	.data = &max_user_watches,
306	.maxlen = sizeof(max_user_watches),
307	.mode = 0644,
308	.proc_handler = proc_doulongvec_minmax,
309	.extra1 = &zero,
310	.extra2 = &long_max,
311	},
312	{ }
313	};
314	#endif /* CONFIG_SYSCTL */
315
316	static const struct file_operations eventpoll_fops;
317
318	static inline int is_file_epoll(struct file *f)
319	{
320	return f->f_op == &eventpoll_fops;
321	}
322
323	/* Setup the structure that is used as key for the RB tree */
324	static inline void ep_set_ffd(struct epoll_filefd *ffd,
325	struct file *file, int fd)
326	{
327	ffd->file = file;
328	ffd->fd = fd;
329	}
330
331	/* Compare RB tree keys */
332	static inline int ep_cmp_ffd(struct epoll_filefd *p1,
333	struct epoll_filefd *p2)
334	{
335	return (p1->file > p2->file ? +1:
336	(p1->file < p2->file ? -1 : p1->fd - p2->fd));
337	}
338
339	/* Tells us if the item is currently linked */
340	static inline int ep_is_linked(struct list_head *p)
341	{
342	return !list_empty(p);
343	}
344
345	static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_t *p)
346	{
347	return container_of(p, struct eppoll_entry, wait);
348	}
349
350	/* Get the "struct epitem" from a wait queue pointer */
351	static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
352	{
353	return container_of(p, struct eppoll_entry, wait)->base;
354	}
355
356	/* Get the "struct epitem" from an epoll queue wrapper */
357	static inline struct epitem *ep_item_from_epqueue(poll_table *p)
358	{
359	return container_of(p, struct ep_pqueue, pt)->epi;
360	}
361
362	/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
363	static inline int ep_op_has_event(int op)
364	{
365	return op != EPOLL_CTL_DEL;
366	}
367
368	/* Initialize the poll safe wake up structure */
369	static void ep_nested_calls_init(struct nested_calls *ncalls)
370	{
371	INIT_LIST_HEAD(&ncalls->tasks_call_list);
372	spin_lock_init(&ncalls->lock);
373	}
374
375	/**
376	* ep_events_available - Checks if ready events might be available.
377	*
378	* @ep: Pointer to the eventpoll context.
379	*
380	* Returns: Returns a value different than zero if ready events are available,
381	* or zero otherwise.
382	*/
383	static inline int ep_events_available(struct eventpoll *ep)
384	{
385	return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
386	}
387
388	/**
389	* ep_call_nested - Perform a bound (possibly) nested call, by checking
390	* that the recursion limit is not exceeded, and that
391	* the same nested call (by the meaning of same cookie) is
392	* no re-entered.
393	*
394	* @ncalls: Pointer to the nested_calls structure to be used for this call.
395	* @max_nests: Maximum number of allowed nesting calls.
396	* @nproc: Nested call core function pointer.
397	* @priv: Opaque data to be passed to the @nproc callback.
398	* @cookie: Cookie to be used to identify this nested call.
399	* @ctx: This instance context.
400	*
401	* Returns: Returns the code returned by the @nproc callback, or -1 if
402	* the maximum recursion limit has been exceeded.
403	*/
404	static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
405	int (*nproc)(void *, void *, int), void *priv,
406	void *cookie, void *ctx)
407	{
408	int error, call_nests = 0;
409	unsigned long flags;
410	struct list_head *lsthead = &ncalls->tasks_call_list;
411	struct nested_call_node *tncur;
412	struct nested_call_node tnode;
413
414	spin_lock_irqsave(&ncalls->lock, flags);
415
416	/*
417	* Try to see if the current task is already inside this wakeup call.
418	* We use a list here, since the population inside this set is always
419	* very much limited.
420	*/
421	list_for_each_entry(tncur, lsthead, llink) {
422	if (tncur->ctx == ctx &&
423	(tncur->cookie == cookie || ++call_nests > max_nests)) {
424	/*
425	* Ops ... loop detected or maximum nest level reached.
426	* We abort this wake by breaking the cycle itself.
427	*/
428	error = -1;
429	goto out_unlock;
430	}
431	}
432
433	/* Add the current task and cookie to the list */
434	tnode.ctx = ctx;
435	tnode.cookie = cookie;
436	list_add(&tnode.llink, lsthead);
437
438	spin_unlock_irqrestore(&ncalls->lock, flags);
439
440	/* Call the nested function */
441	error = (*nproc)(priv, cookie, call_nests);
442
443	/* Remove the current task from the list */
444	spin_lock_irqsave(&ncalls->lock, flags);
445	list_del(&tnode.llink);
446	out_unlock:
447	spin_unlock_irqrestore(&ncalls->lock, flags);
448
449	return error;
450	}
451
452	/*
453	* As described in commit 0ccf831cb lockdep: annotate epoll
454	* the use of wait queues used by epoll is done in a very controlled
455	* manner. Wake ups can nest inside each other, but are never done
456	* with the same locking. For example:
457	*
458	* dfd = socket(...);
459	* efd1 = epoll_create();
460	* efd2 = epoll_create();
461	* epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
462	* epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
463	*
464	* When a packet arrives to the device underneath "dfd", the net code will
465	* issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
466	* callback wakeup entry on that queue, and the wake_up() performed by the
467	* "dfd" net code will end up in ep_poll_callback(). At this point epoll
468	* (efd1) notices that it may have some event ready, so it needs to wake up
469	* the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
470	* that ends up in another wake_up(), after having checked about the
471	* recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
472	* avoid stack blasting.
473	*
474	* When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
475	* this special case of epoll.
476	*/
477	#ifdef CONFIG_DEBUG_LOCK_ALLOC
478	static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
479	unsigned long events, int subclass)
480	{
481	unsigned long flags;
482
483	spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
484	wake_up_locked_poll(wqueue, events);
485	spin_unlock_irqrestore(&wqueue->lock, flags);
486	}
487	#else
488	static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
489	unsigned long events, int subclass)
490	{
491	wake_up_poll(wqueue, events);
492	}
493	#endif
494
495	static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
496	{
497	ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
498	1 + call_nests);
499	return 0;
500	}
501
502	/*
503	* Perform a safe wake up of the poll wait list. The problem is that
504	* with the new callback'd wake up system, it is possible that the
505	* poll callback is reentered from inside the call to wake_up() done
506	* on the poll wait queue head. The rule is that we cannot reenter the
507	* wake up code from the same task more than EP_MAX_NESTS times,
508	* and we cannot reenter the same wait queue head at all. This will
509	* enable to have a hierarchy of epoll file descriptor of no more than
510	* EP_MAX_NESTS deep.
511	*/
512	static void ep_poll_safewake(wait_queue_head_t *wq)
513	{
514	int this_cpu = get_cpu();
515
516	ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
517	ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
518
519	put_cpu();
520	}
521
522	static void ep_remove_wait_queue(struct eppoll_entry *pwq)
523	{
524	wait_queue_head_t *whead;
525
526	rcu_read_lock();
527	/*
528	* If it is cleared by POLLFREE, it should be rcu-safe.
529	* If we read NULL we need a barrier paired with
530	* smp_store_release() in ep_poll_callback(), otherwise
531	* we rely on whead->lock.
532	*/
533	whead = smp_load_acquire(&pwq->whead);
534	if (whead)
535	remove_wait_queue(whead, &pwq->wait);
536	rcu_read_unlock();
537	}
538
539	/*
540	* This function unregisters poll callbacks from the associated file
541	* descriptor. Must be called with "mtx" held (or "epmutex" if called from
542	* ep_free).
543	*/
544	static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
545	{
546	struct list_head *lsthead = &epi->pwqlist;
547	struct eppoll_entry *pwq;
548
549	while (!list_empty(lsthead)) {
550	pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
551
552	list_del(&pwq->llink);
553	ep_remove_wait_queue(pwq);
554	kmem_cache_free(pwq_cache, pwq);
555	}
556	}
557
558	/* call only when ep->mtx is held */
559	static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
560	{
561	return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
562	}
563
564	/* call only when ep->mtx is held */
565	static inline void ep_pm_stay_awake(struct epitem *epi)
566	{
567	struct wakeup_source *ws = ep_wakeup_source(epi);
568
569	if (ws)
570	__pm_stay_awake(ws);
571	}
572
573	static inline bool ep_has_wakeup_source(struct epitem *epi)
574	{
575	return rcu_access_pointer(epi->ws) ? true : false;
576	}
577
578	/* call when ep->mtx cannot be held (ep_poll_callback) */
579	static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
580	{
581	struct wakeup_source *ws;
582
583	rcu_read_lock();
584	ws = rcu_dereference(epi->ws);
585	if (ws)
586	__pm_stay_awake(ws);
587	rcu_read_unlock();
588	}
589
590	/**
591	* ep_scan_ready_list - Scans the ready list in a way that makes possible for
592	* the scan code, to call f_op->poll(). Also allows for
593	* O(NumReady) performance.
594	*
595	* @ep: Pointer to the epoll private data structure.
596	* @sproc: Pointer to the scan callback.
597	* @priv: Private opaque data passed to the @sproc callback.
598	* @depth: The current depth of recursive f_op->poll calls.
599	* @ep_locked: caller already holds ep->mtx
600	*
601	* Returns: The same integer error code returned by the @sproc callback.
602	*/
603	static int ep_scan_ready_list(struct eventpoll *ep,
604	int (*sproc)(struct eventpoll *,
605	struct list_head *, void *),
606	void *priv, int depth, bool ep_locked)
607	{
608	int error, pwake = 0;
609	unsigned long flags;
610	struct epitem *epi, *nepi;
611	LIST_HEAD(txlist);
612
613	/*
614	* We need to lock this because we could be hit by
615	* eventpoll_release_file() and epoll_ctl().
616	*/
617
618	if (!ep_locked)
619	mutex_lock_nested(&ep->mtx, depth);
620
621	/*
622	* Steal the ready list, and re-init the original one to the
623	* empty list. Also, set ep->ovflist to NULL so that events
624	* happening while looping w/out locks, are not lost. We cannot
625	* have the poll callback to queue directly on ep->rdllist,
626	* because we want the "sproc" callback to be able to do it
627	* in a lockless way.
628	*/
629	spin_lock_irqsave(&ep->lock, flags);
630	list_splice_init(&ep->rdllist, &txlist);
631	ep->ovflist = NULL;
632	spin_unlock_irqrestore(&ep->lock, flags);
633
634	/*
635	* Now call the callback function.
636	*/
637	error = (*sproc)(ep, &txlist, priv);
638
639	spin_lock_irqsave(&ep->lock, flags);
640	/*
641	* During the time we spent inside the "sproc" callback, some
642	* other events might have been queued by the poll callback.
643	* We re-insert them inside the main ready-list here.
644	*/
645	for (nepi = ep->ovflist; (epi = nepi) != NULL;
646	nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
647	/*
648	* We need to check if the item is already in the list.
649	* During the "sproc" callback execution time, items are
650	* queued into ->ovflist but the "txlist" might already
651	* contain them, and the list_splice() below takes care of them.
652	*/
653	if (!ep_is_linked(&epi->rdllink)) {
654	list_add_tail(&epi->rdllink, &ep->rdllist);
655	ep_pm_stay_awake(epi);
656	}
657	}
658	/*
659	* We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
660	* releasing the lock, events will be queued in the normal way inside
661	* ep->rdllist.
662	*/
663	ep->ovflist = EP_UNACTIVE_PTR;
664
665	/*
666	* Quickly re-inject items left on "txlist".
667	*/
668	list_splice(&txlist, &ep->rdllist);
669	__pm_relax(ep->ws);
670
671	if (!list_empty(&ep->rdllist)) {
672	/*
673	* Wake up (if active) both the eventpoll wait list and
674	* the ->poll() wait list (delayed after we release the lock).
675	*/
676	if (waitqueue_active(&ep->wq))
677	wake_up_locked(&ep->wq);
678	if (waitqueue_active(&ep->poll_wait))
679	pwake++;
680	}
681	spin_unlock_irqrestore(&ep->lock, flags);
682
683	if (!ep_locked)
684	mutex_unlock(&ep->mtx);
685
686	/* We have to call this outside the lock */
687	if (pwake)
688	ep_poll_safewake(&ep->poll_wait);
689
690	return error;
691	}
692
693	static void epi_rcu_free(struct rcu_head *head)
694	{
695	struct epitem *epi = container_of(head, struct epitem, rcu);
696	kmem_cache_free(epi_cache, epi);
697	}
698
699	/*
700	* Removes a "struct epitem" from the eventpoll RB tree and deallocates
701	* all the associated resources. Must be called with "mtx" held.
702	*/
703	static int ep_remove(struct eventpoll *ep, struct epitem *epi)
704	{
705	unsigned long flags;
706	struct file *file = epi->ffd.file;
707
708	/*
709	* Removes poll wait queue hooks. We _have_ to do this without holding
710	* the "ep->lock" otherwise a deadlock might occur. This because of the
711	* sequence of the lock acquisition. Here we do "ep->lock" then the wait
712	* queue head lock when unregistering the wait queue. The wakeup callback
713	* will run by holding the wait queue head lock and will call our callback
714	* that will try to get "ep->lock".
715	*/
716	ep_unregister_pollwait(ep, epi);
717
718	/* Remove the current item from the list of epoll hooks */
719	spin_lock(&file->f_lock);
720	list_del_rcu(&epi->fllink);
721	spin_unlock(&file->f_lock);
722
723	rb_erase(&epi->rbn, &ep->rbr);
724
725	spin_lock_irqsave(&ep->lock, flags);
726	if (ep_is_linked(&epi->rdllink))
727	list_del_init(&epi->rdllink);
728	spin_unlock_irqrestore(&ep->lock, flags);
729
730	wakeup_source_unregister(ep_wakeup_source(epi));
731	/*
732	* At this point it is safe to free the eventpoll item. Use the union
733	* field epi->rcu, since we are trying to minimize the size of
734	* 'struct epitem'. The 'rbn' field is no longer in use. Protected by
735	* ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
736	* use of the rbn field.
737	*/
738	call_rcu(&epi->rcu, epi_rcu_free);
739
740	atomic_long_dec(&ep->user->epoll_watches);
741
742	return 0;
743	}
744
745	static void ep_free(struct eventpoll *ep)
746	{
747	struct rb_node *rbp;
748	struct epitem *epi;
749
750	/* We need to release all tasks waiting for these file */
751	if (waitqueue_active(&ep->poll_wait))
752	ep_poll_safewake(&ep->poll_wait);
753
754	/*
755	* We need to lock this because we could be hit by
756	* eventpoll_release_file() while we're freeing the "struct eventpoll".
757	* We do not need to hold "ep->mtx" here because the epoll file
758	* is on the way to be removed and no one has references to it
759	* anymore. The only hit might come from eventpoll_release_file() but
760	* holding "epmutex" is sufficient here.
761	*/
762	mutex_lock(&epmutex);
763
764	/*
765	* Walks through the whole tree by unregistering poll callbacks.
766	*/
767	for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
768	epi = rb_entry(rbp, struct epitem, rbn);
769
770	ep_unregister_pollwait(ep, epi);
771	cond_resched();
772	}
773
774	/*
775	* Walks through the whole tree by freeing each "struct epitem". At this
776	* point we are sure no poll callbacks will be lingering around, and also by
777	* holding "epmutex" we can be sure that no file cleanup code will hit
778	* us during this operation. So we can avoid the lock on "ep->lock".
779	* We do not need to lock ep->mtx, either, we only do it to prevent
780	* a lockdep warning.
781	*/
782	mutex_lock(&ep->mtx);
783	while ((rbp = rb_first(&ep->rbr)) != NULL) {
784	epi = rb_entry(rbp, struct epitem, rbn);
785	ep_remove(ep, epi);
786	cond_resched();
787	}
788	mutex_unlock(&ep->mtx);
789
790	mutex_unlock(&epmutex);
791	mutex_destroy(&ep->mtx);
792	free_uid(ep->user);
793	wakeup_source_unregister(ep->ws);
794	kfree(ep);
795	}
796
797	static int ep_eventpoll_release(struct inode *inode, struct file *file)
798	{
799	struct eventpoll *ep = file->private_data;
800
801	if (ep)
802	ep_free(ep);
803
804	return 0;
805	}
806
807	static inline unsigned int ep_item_poll(struct epitem *epi, poll_table *pt)
808	{
809	pt->_key = epi->event.events;
810
811	return epi->ffd.file->f_op->poll(epi->ffd.file, pt) & epi->event.events;
812	}
813
814	static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
815	void *priv)
816	{
817	struct epitem *epi, *tmp;
818	poll_table pt;
819
820	init_poll_funcptr(&pt, NULL);
821
822	list_for_each_entry_safe(epi, tmp, head, rdllink) {
823	if (ep_item_poll(epi, &pt))
824	return POLLIN | POLLRDNORM;
825	else {
826	/*
827	* Item has been dropped into the ready list by the poll
828	* callback, but it's not actually ready, as far as
829	* caller requested events goes. We can remove it here.
830	*/
831	__pm_relax(ep_wakeup_source(epi));
832	list_del_init(&epi->rdllink);
833	}
834	}
835
836	return 0;
837	}
838
839	static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
840	poll_table *pt);
841
842	struct readyevents_arg {
843	struct eventpoll *ep;
844	bool locked;
845	};
846
847	static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
848	{
849	struct readyevents_arg *arg = priv;
850
851	return ep_scan_ready_list(arg->ep, ep_read_events_proc, NULL,
852	call_nests + 1, arg->locked);
853	}
854
855	static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
856	{
857	int pollflags;
858	struct eventpoll *ep = file->private_data;
859	struct readyevents_arg arg;
860
861	/*
862	* During ep_insert() we already hold the ep->mtx for the tfile.
863	* Prevent re-aquisition.
864	*/
865	arg.locked = wait && (wait->_qproc == ep_ptable_queue_proc);
866	arg.ep = ep;
867
868	/* Insert inside our poll wait queue */
869	poll_wait(file, &ep->poll_wait, wait);
870
871	/*
872	* Proceed to find out if wanted events are really available inside
873	* the ready list. This need to be done under ep_call_nested()
874	* supervision, since the call to f_op->poll() done on listed files
875	* could re-enter here.
876	*/
877	pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
878	ep_poll_readyevents_proc, &arg, ep, current);
879
880	return pollflags != -1 ? pollflags : 0;
881	}
882
883	#ifdef CONFIG_PROC_FS
884	static void ep_show_fdinfo(struct seq_file *m, struct file *f)
885	{
886	struct eventpoll *ep = f->private_data;
887	struct rb_node *rbp;
888
889	mutex_lock(&ep->mtx);
890	for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
891	struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
892
893	seq_printf(m, "tfd: %8d events: %8x data: %16llx\n",
894	epi->ffd.fd, epi->event.events,
895	(long long)epi->event.data);
896	if (seq_has_overflowed(m))
897	break;
898	}
899	mutex_unlock(&ep->mtx);
900	}
901	#endif
902
903	/* File callbacks that implement the eventpoll file behaviour */
904	static const struct file_operations eventpoll_fops = {
905	#ifdef CONFIG_PROC_FS
906	.show_fdinfo = ep_show_fdinfo,
907	#endif
908	.release = ep_eventpoll_release,
909	.poll = ep_eventpoll_poll,
910	.llseek = noop_llseek,
911	};
912
913	/*
914	* This is called from eventpoll_release() to unlink files from the eventpoll
915	* interface. We need to have this facility to cleanup correctly files that are
916	* closed without being removed from the eventpoll interface.
917	*/
918	void eventpoll_release_file(struct file *file)
919	{
920	struct eventpoll *ep;
921	struct epitem *epi, *next;
922
923	/*
924	* We don't want to get "file->f_lock" because it is not
925	* necessary. It is not necessary because we're in the "struct file"
926	* cleanup path, and this means that no one is using this file anymore.
927	* So, for example, epoll_ctl() cannot hit here since if we reach this
928	* point, the file counter already went to zero and fget() would fail.
929	* The only hit might come from ep_free() but by holding the mutex
930	* will correctly serialize the operation. We do need to acquire
931	* "ep->mtx" after "epmutex" because ep_remove() requires it when called
932	* from anywhere but ep_free().
933	*
934	* Besides, ep_remove() acquires the lock, so we can't hold it here.
935	*/
936	mutex_lock(&epmutex);
937	list_for_each_entry_safe(epi, next, &file->f_ep_links, fllink) {
938	ep = epi->ep;
939	mutex_lock_nested(&ep->mtx, 0);
940	ep_remove(ep, epi);
941	mutex_unlock(&ep->mtx);
942	}
943	mutex_unlock(&epmutex);
944	}
945
946	static int ep_alloc(struct eventpoll **pep)
947	{
948	int error;
949	struct user_struct *user;
950	struct eventpoll *ep;
951
952	user = get_current_user();
953	error = -ENOMEM;
954	ep = kzalloc(sizeof(*ep), GFP_KERNEL);
955	if (unlikely(!ep))
956	goto free_uid;
957
958	spin_lock_init(&ep->lock);
959	mutex_init(&ep->mtx);
960	init_waitqueue_head(&ep->wq);
961	init_waitqueue_head(&ep->poll_wait);
962	INIT_LIST_HEAD(&ep->rdllist);
963	ep->rbr = RB_ROOT;
964	ep->ovflist = EP_UNACTIVE_PTR;
965	ep->user = user;
966
967	*pep = ep;
968
969	return 0;
970
971	free_uid:
972	free_uid(user);
973	return error;
974	}
975
976	/*
977	* Search the file inside the eventpoll tree. The RB tree operations
978	* are protected by the "mtx" mutex, and ep_find() must be called with
979	* "mtx" held.
980	*/
981	static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
982	{
983	int kcmp;
984	struct rb_node *rbp;
985	struct epitem *epi, *epir = NULL;
986	struct epoll_filefd ffd;
987
988	ep_set_ffd(&ffd, file, fd);
989	for (rbp = ep->rbr.rb_node; rbp; ) {
990	epi = rb_entry(rbp, struct epitem, rbn);
991	kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
992	if (kcmp > 0)
993	rbp = rbp->rb_right;
994	else if (kcmp < 0)
995	rbp = rbp->rb_left;
996	else {
997	epir = epi;
998	break;
999	}
1000	}
1001
1002	return epir;
1003	}
1004
1005	/*
1006	* This is the callback that is passed to the wait queue wakeup
1007	* mechanism. It is called by the stored file descriptors when they
1008	* have events to report.
1009	*/
1010	static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
1011	{
1012	int pwake = 0;
1013	unsigned long flags;
1014	struct epitem *epi = ep_item_from_wait(wait);
1015	struct eventpoll *ep = epi->ep;
1016	int ewake = 0;
1017
1018	spin_lock_irqsave(&ep->lock, flags);
1019
1020	/*
1021	* If the event mask does not contain any poll(2) event, we consider the
1022	* descriptor to be disabled. This condition is likely the effect of the
1023	* EPOLLONESHOT bit that disables the descriptor when an event is received,
1024	* until the next EPOLL_CTL_MOD will be issued.
1025	*/
1026	if (!(epi->event.events & ~EP_PRIVATE_BITS))
1027	goto out_unlock;
1028
1029	/*
1030	* Check the events coming with the callback. At this stage, not
1031	* every device reports the events in the "key" parameter of the
1032	* callback. We need to be able to handle both cases here, hence the
1033	* test for "key" != NULL before the event match test.
1034	*/
1035	if (key && !((unsigned long) key & epi->event.events))
1036	goto out_unlock;
1037
1038	/*
1039	* If we are transferring events to userspace, we can hold no locks
1040	* (because we're accessing user memory, and because of linux f_op->poll()
1041	* semantics). All the events that happen during that period of time are
1042	* chained in ep->ovflist and requeued later on.
1043	*/
1044	if (ep->ovflist != EP_UNACTIVE_PTR) {
1045	if (epi->next == EP_UNACTIVE_PTR) {
1046	epi->next = ep->ovflist;
1047	ep->ovflist = epi;
1048	if (epi->ws) {
1049	/*
1050	* Activate ep->ws since epi->ws may get
1051	* deactivated at any time.
1052	*/
1053	__pm_stay_awake(ep->ws);
1054	}
1055
1056	}
1057	goto out_unlock;
1058	}
1059
1060	/* If this file is already in the ready list we exit soon */
1061	if (!ep_is_linked(&epi->rdllink)) {
1062	list_add_tail(&epi->rdllink, &ep->rdllist);
1063	ep_pm_stay_awake_rcu(epi);
1064	}
1065
1066	/*
1067	* Wake up ( if active ) both the eventpoll wait list and the ->poll()
1068	* wait list.
1069	*/
1070	if (waitqueue_active(&ep->wq)) {
1071	if ((epi->event.events & EPOLLEXCLUSIVE) &&
1072	!((unsigned long)key & POLLFREE)) {
1073	switch ((unsigned long)key & EPOLLINOUT_BITS) {
1074	case POLLIN:
1075	if (epi->event.events & POLLIN)
1076	ewake = 1;
1077	break;
1078	case POLLOUT:
1079	if (epi->event.events & POLLOUT)
1080	ewake = 1;
1081	break;
1082	case 0:
1083	ewake = 1;
1084	break;
1085	}
1086	}
1087	wake_up_locked(&ep->wq);
1088	}
1089	if (waitqueue_active(&ep->poll_wait))
1090	pwake++;
1091
1092	out_unlock:
1093	spin_unlock_irqrestore(&ep->lock, flags);
1094
1095	/* We have to call this outside the lock */
1096	if (pwake)
1097	ep_poll_safewake(&ep->poll_wait);
1098
1099	if (!(epi->event.events & EPOLLEXCLUSIVE))
1100	ewake = 1;
1101
1102	if ((unsigned long)key & POLLFREE) {
1103	/*
1104	* If we race with ep_remove_wait_queue() it can miss
1105	* ->whead = NULL and do another remove_wait_queue() after
1106	* us, so we can't use __remove_wait_queue().
1107	*/
1108	list_del_init(&wait->task_list);
1109	/*
1110	* ->whead != NULL protects us from the race with ep_free()
1111	* or ep_remove(), ep_remove_wait_queue() takes whead->lock
1112	* held by the caller. Once we nullify it, nothing protects
1113	* ep/epi or even wait.
1114	*/
1115	smp_store_release(&ep_pwq_from_wait(wait)->whead, NULL);
1116	}
1117
1118	return ewake;
1119	}
1120
1121	/*
1122	* This is the callback that is used to add our wait queue to the
1123	* target file wakeup lists.
1124	*/
1125	static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1126	poll_table *pt)
1127	{
1128	struct epitem *epi = ep_item_from_epqueue(pt);
1129	struct eppoll_entry *pwq;
1130
1131	if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
1132	init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1133	pwq->whead = whead;
1134	pwq->base = epi;
1135	if (epi->event.events & EPOLLEXCLUSIVE)
1136	add_wait_queue_exclusive(whead, &pwq->wait);
1137	else
1138	add_wait_queue(whead, &pwq->wait);
1139	list_add_tail(&pwq->llink, &epi->pwqlist);
1140	epi->nwait++;
1141	} else {
1142	/* We have to signal that an error occurred */
1143	epi->nwait = -1;
1144	}
1145	}
1146
1147	static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1148	{
1149	int kcmp;
1150	struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
1151	struct epitem *epic;
1152
1153	while (*p) {
1154	parent = *p;
1155	epic = rb_entry(parent, struct epitem, rbn);
1156	kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1157	if (kcmp > 0)
1158	p = &parent->rb_right;
1159	else
1160	p = &parent->rb_left;
1161	}
1162	rb_link_node(&epi->rbn, parent, p);
1163	rb_insert_color(&epi->rbn, &ep->rbr);
1164	}
1165
1166
1167
1168	#define PATH_ARR_SIZE 5
1169	/*
1170	* These are the number paths of length 1 to 5, that we are allowing to emanate
1171	* from a single file of interest. For example, we allow 1000 paths of length
1172	* 1, to emanate from each file of interest. This essentially represents the
1173	* potential wakeup paths, which need to be limited in order to avoid massive
1174	* uncontrolled wakeup storms. The common use case should be a single ep which
1175	* is connected to n file sources. In this case each file source has 1 path
1176	* of length 1. Thus, the numbers below should be more than sufficient. These
1177	* path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1178	* and delete can't add additional paths. Protected by the epmutex.
1179	*/
1180	static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1181	static int path_count[PATH_ARR_SIZE];
1182
1183	static int path_count_inc(int nests)
1184	{
1185	/* Allow an arbitrary number of depth 1 paths */
1186	if (nests == 0)
1187	return 0;
1188
1189	if (++path_count[nests] > path_limits[nests])
1190	return -1;
1191	return 0;
1192	}
1193
1194	static void path_count_init(void)
1195	{
1196	int i;
1197
1198	for (i = 0; i < PATH_ARR_SIZE; i++)
1199	path_count[i] = 0;
1200	}
1201
1202	static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1203	{
1204	int error = 0;
1205	struct file *file = priv;
1206	struct file *child_file;
1207	struct epitem *epi;
1208
1209	/* CTL_DEL can remove links here, but that can't increase our count */
1210	rcu_read_lock();
1211	list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) {
1212	child_file = epi->ep->file;
1213	if (is_file_epoll(child_file)) {
1214	if (list_empty(&child_file->f_ep_links)) {
1215	if (path_count_inc(call_nests)) {
1216	error = -1;
1217	break;
1218	}
1219	} else {
1220	error = ep_call_nested(&poll_loop_ncalls,
1221	EP_MAX_NESTS,
1222	reverse_path_check_proc,
1223	child_file, child_file,
1224	current);
1225	}
1226	if (error != 0)
1227	break;
1228	} else {
1229	printk(KERN_ERR "reverse_path_check_proc: "
1230	"file is not an ep!\n");
1231	}
1232	}
1233	rcu_read_unlock();
1234	return error;
1235	}
1236
1237	/**
1238	* reverse_path_check - The tfile_check_list is list of file *, which have
1239	* links that are proposed to be newly added. We need to
1240	* make sure that those added links don't add too many
1241	* paths such that we will spend all our time waking up
1242	* eventpoll objects.
1243	*
1244	* Returns: Returns zero if the proposed links don't create too many paths,
1245	* -1 otherwise.
1246	*/
1247	static int reverse_path_check(void)
1248	{
1249	int error = 0;
1250	struct file *current_file;
1251
1252	/* let's call this for all tfiles */
1253	list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
1254	path_count_init();
1255	error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1256	reverse_path_check_proc, current_file,
1257	current_file, current);
1258	if (error)
1259	break;
1260	}
1261	return error;
1262	}
1263
1264	static int ep_create_wakeup_source(struct epitem *epi)
1265	{
1266	const char *name;
1267	struct wakeup_source *ws;
1268
1269	if (!epi->ep->ws) {
1270	epi->ep->ws = wakeup_source_register("eventpoll");
1271	if (!epi->ep->ws)
1272	return -ENOMEM;
1273	}
1274
1275	name = epi->ffd.file->f_path.dentry->d_name.name;
1276	ws = wakeup_source_register(name);
1277
1278	if (!ws)
1279	return -ENOMEM;
1280	rcu_assign_pointer(epi->ws, ws);
1281
1282	return 0;
1283	}
1284
1285	/* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1286	static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1287	{
1288	struct wakeup_source *ws = ep_wakeup_source(epi);
1289
1290	RCU_INIT_POINTER(epi->ws, NULL);
1291
1292	/*
1293	* wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1294	* used internally by wakeup_source_remove, too (called by
1295	* wakeup_source_unregister), so we cannot use call_rcu
1296	*/
1297	synchronize_rcu();
1298	wakeup_source_unregister(ws);
1299	}
1300
1301	/*
1302	* Must be called with "mtx" held.
1303	*/
1304	static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
1305	struct file *tfile, int fd, int full_check)
1306	{
1307	int error, revents, pwake = 0;
1308	unsigned long flags;
1309	long user_watches;
1310	struct epitem *epi;
1311	struct ep_pqueue epq;
1312
1313	user_watches = atomic_long_read(&ep->user->epoll_watches);
1314	if (unlikely(user_watches >= max_user_watches))
1315	return -ENOSPC;
1316	if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1317	return -ENOMEM;
1318
1319	/* Item initialization follow here ... */
1320	INIT_LIST_HEAD(&epi->rdllink);
1321	INIT_LIST_HEAD(&epi->fllink);
1322	INIT_LIST_HEAD(&epi->pwqlist);
1323	epi->ep = ep;
1324	ep_set_ffd(&epi->ffd, tfile, fd);
1325	epi->event = *event;
1326	epi->nwait = 0;
1327	epi->next = EP_UNACTIVE_PTR;
1328	if (epi->event.events & EPOLLWAKEUP) {
1329	error = ep_create_wakeup_source(epi);
1330	if (error)
1331	goto error_create_wakeup_source;
1332	} else {
1333	RCU_INIT_POINTER(epi->ws, NULL);
1334	}
1335
1336	/* Initialize the poll table using the queue callback */
1337	epq.epi = epi;
1338	init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1339
1340	/*
1341	* Attach the item to the poll hooks and get current event bits.
1342	* We can safely use the file* here because its usage count has
1343	* been increased by the caller of this function. Note that after
1344	* this operation completes, the poll callback can start hitting
1345	* the new item.
1346	*/
1347	revents = ep_item_poll(epi, &epq.pt);
1348
1349	/*
1350	* We have to check if something went wrong during the poll wait queue
1351	* install process. Namely an allocation for a wait queue failed due
1352	* high memory pressure.
1353	*/
1354	error = -ENOMEM;
1355	if (epi->nwait < 0)
1356	goto error_unregister;
1357
1358	/* Add the current item to the list of active epoll hook for this file */
1359	spin_lock(&tfile->f_lock);
1360	list_add_tail_rcu(&epi->fllink, &tfile->f_ep_links);
1361	spin_unlock(&tfile->f_lock);
1362
1363	/*
1364	* Add the current item to the RB tree. All RB tree operations are
1365	* protected by "mtx", and ep_insert() is called with "mtx" held.
1366	*/
1367	ep_rbtree_insert(ep, epi);
1368
1369	/* now check if we've created too many backpaths */
1370	error = -EINVAL;
1371	if (full_check && reverse_path_check())
1372	goto error_remove_epi;
1373
1374	/* We have to drop the new item inside our item list to keep track of it */
1375	spin_lock_irqsave(&ep->lock, flags);
1376
1377	/* If the file is already "ready" we drop it inside the ready list */
1378	if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
1379	list_add_tail(&epi->rdllink, &ep->rdllist);
1380	ep_pm_stay_awake(epi);
1381
1382	/* Notify waiting tasks that events are available */
1383	if (waitqueue_active(&ep->wq))
1384	wake_up_locked(&ep->wq);
1385	if (waitqueue_active(&ep->poll_wait))
1386	pwake++;
1387	}
1388
1389	spin_unlock_irqrestore(&ep->lock, flags);
1390
1391	atomic_long_inc(&ep->user->epoll_watches);
1392
1393	/* We have to call this outside the lock */
1394	if (pwake)
1395	ep_poll_safewake(&ep->poll_wait);
1396
1397	return 0;
1398
1399	error_remove_epi:
1400	spin_lock(&tfile->f_lock);
1401	list_del_rcu(&epi->fllink);
1402	spin_unlock(&tfile->f_lock);
1403
1404	rb_erase(&epi->rbn, &ep->rbr);
1405
1406	error_unregister:
1407	ep_unregister_pollwait(ep, epi);
1408
1409	/*
1410	* We need to do this because an event could have been arrived on some
1411	* allocated wait queue. Note that we don't care about the ep->ovflist
1412	* list, since that is used/cleaned only inside a section bound by "mtx".
1413	* And ep_insert() is called with "mtx" held.
1414	*/
1415	spin_lock_irqsave(&ep->lock, flags);
1416	if (ep_is_linked(&epi->rdllink))
1417	list_del_init(&epi->rdllink);
1418	spin_unlock_irqrestore(&ep->lock, flags);
1419
1420	wakeup_source_unregister(ep_wakeup_source(epi));
1421
1422	error_create_wakeup_source:
1423	kmem_cache_free(epi_cache, epi);
1424
1425	return error;
1426	}
1427
1428	/*
1429	* Modify the interest event mask by dropping an event if the new mask
1430	* has a match in the current file status. Must be called with "mtx" held.
1431	*/
1432	static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1433	{
1434	int pwake = 0;
1435	unsigned int revents;
1436	poll_table pt;
1437
1438	init_poll_funcptr(&pt, NULL);
1439
1440	/*
1441	* Set the new event interest mask before calling f_op->poll();
1442	* otherwise we might miss an event that happens between the
1443	* f_op->poll() call and the new event set registering.
1444	*/
1445	epi->event.events = event->events; /* need barrier below */
1446	epi->event.data = event->data; /* protected by mtx */
1447	if (epi->event.events & EPOLLWAKEUP) {
1448	if (!ep_has_wakeup_source(epi))
1449	ep_create_wakeup_source(epi);
1450	} else if (ep_has_wakeup_source(epi)) {
1451	ep_destroy_wakeup_source(epi);
1452	}
1453
1454	/*
1455	* The following barrier has two effects:
1456	*
1457	* 1) Flush epi changes above to other CPUs. This ensures
1458	* we do not miss events from ep_poll_callback if an
1459	* event occurs immediately after we call f_op->poll().
1460	* We need this because we did not take ep->lock while
1461	* changing epi above (but ep_poll_callback does take
1462	* ep->lock).
1463	*
1464	* 2) We also need to ensure we do not miss _past_ events
1465	* when calling f_op->poll(). This barrier also
1466	* pairs with the barrier in wq_has_sleeper (see
1467	* comments for wq_has_sleeper).
1468	*
1469	* This barrier will now guarantee ep_poll_callback or f_op->poll
1470	* (or both) will notice the readiness of an item.
1471	*/
1472	smp_mb();
1473
1474	/*
1475	* Get current event bits. We can safely use the file* here because
1476	* its usage count has been increased by the caller of this function.
1477	*/
1478	revents = ep_item_poll(epi, &pt);
1479
1480	/*
1481	* If the item is "hot" and it is not registered inside the ready
1482	* list, push it inside.
1483	*/
1484	if (revents & event->events) {
1485	spin_lock_irq(&ep->lock);
1486	if (!ep_is_linked(&epi->rdllink)) {
1487	list_add_tail(&epi->rdllink, &ep->rdllist);
1488	ep_pm_stay_awake(epi);
1489
1490	/* Notify waiting tasks that events are available */
1491	if (waitqueue_active(&ep->wq))
1492	wake_up_locked(&ep->wq);
1493	if (waitqueue_active(&ep->poll_wait))
1494	pwake++;
1495	}
1496	spin_unlock_irq(&ep->lock);
1497	}
1498
1499	/* We have to call this outside the lock */
1500	if (pwake)
1501	ep_poll_safewake(&ep->poll_wait);
1502
1503	return 0;
1504	}
1505
1506	static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1507	void *priv)
1508	{
1509	struct ep_send_events_data *esed = priv;
1510	int eventcnt;
1511	unsigned int revents;
1512	struct epitem *epi;
1513	struct epoll_event __user *uevent;
1514	struct wakeup_source *ws;
1515	poll_table pt;
1516
1517	init_poll_funcptr(&pt, NULL);
1518
1519	/*
1520	* We can loop without lock because we are passed a task private list.
1521	* Items cannot vanish during the loop because ep_scan_ready_list() is
1522	* holding "mtx" during this call.
1523	*/
1524	for (eventcnt = 0, uevent = esed->events;
1525	!list_empty(head) && eventcnt < esed->maxevents;) {
1526	epi = list_first_entry(head, struct epitem, rdllink);
1527
1528	/*
1529	* Activate ep->ws before deactivating epi->ws to prevent
1530	* triggering auto-suspend here (in case we reactive epi->ws
1531	* below).
1532	*
1533	* This could be rearranged to delay the deactivation of epi->ws
1534	* instead, but then epi->ws would temporarily be out of sync
1535	* with ep_is_linked().
1536	*/
1537	ws = ep_wakeup_source(epi);
1538	if (ws) {
1539	if (ws->active)
1540	__pm_stay_awake(ep->ws);
1541	__pm_relax(ws);
1542	}
1543
1544	list_del_init(&epi->rdllink);
1545
1546	revents = ep_item_poll(epi, &pt);
1547
1548	/*
1549	* If the event mask intersect the caller-requested one,
1550	* deliver the event to userspace. Again, ep_scan_ready_list()
1551	* is holding "mtx", so no operations coming from userspace
1552	* can change the item.
1553	*/
1554	if (revents) {
1555	if (__put_user(revents, &uevent->events) ||
1556	__put_user(epi->event.data, &uevent->data)) {
1557	list_add(&epi->rdllink, head);
1558	ep_pm_stay_awake(epi);
1559	return eventcnt ? eventcnt : -EFAULT;
1560	}
1561	eventcnt++;
1562	uevent++;
1563	if (epi->event.events & EPOLLONESHOT)
1564	epi->event.events &= EP_PRIVATE_BITS;
1565	else if (!(epi->event.events & EPOLLET)) {
1566	/*
1567	* If this file has been added with Level
1568	* Trigger mode, we need to insert back inside
1569	* the ready list, so that the next call to
1570	* epoll_wait() will check again the events
1571	* availability. At this point, no one can insert
1572	* into ep->rdllist besides us. The epoll_ctl()
1573	* callers are locked out by
1574	* ep_scan_ready_list() holding "mtx" and the
1575	* poll callback will queue them in ep->ovflist.
1576	*/
1577	list_add_tail(&epi->rdllink, &ep->rdllist);
1578	ep_pm_stay_awake(epi);
1579	}
1580	}
1581	}
1582
1583	return eventcnt;
1584	}
1585
1586	static int ep_send_events(struct eventpoll *ep,
1587	struct epoll_event __user *events, int maxevents)
1588	{
1589	struct ep_send_events_data esed;
1590
1591	esed.maxevents = maxevents;
1592	esed.events = events;
1593
1594	return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false);
1595	}
1596
1597	static inline struct timespec64 ep_set_mstimeout(long ms)
1598	{
1599	struct timespec64 now, ts = {
1600	.tv_sec = ms / MSEC_PER_SEC,
1601	.tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1602	};
1603
1604	ktime_get_ts64(&now);
1605	return timespec64_add_safe(now, ts);
1606	}
1607
1608	/**
1609	* ep_poll - Retrieves ready events, and delivers them to the caller supplied
1610	* event buffer.
1611	*
1612	* @ep: Pointer to the eventpoll context.
1613	* @events: Pointer to the userspace buffer where the ready events should be
1614	* stored.
1615	* @maxevents: Size (in terms of number of events) of the caller event buffer.
1616	* @timeout: Maximum timeout for the ready events fetch operation, in
1617	* milliseconds. If the @timeout is zero, the function will not block,
1618	* while if the @timeout is less than zero, the function will block
1619	* until at least one event has been retrieved (or an error
1620	* occurred).
1621	*
1622	* Returns: Returns the number of ready events which have been fetched, or an
1623	* error code, in case of error.
1624	*/
1625	static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1626	int maxevents, long timeout)
1627	{
1628	int res = 0, eavail, timed_out = 0;
1629	unsigned long flags;
1630	u64 slack = 0;
1631	wait_queue_t wait;
1632	ktime_t expires, *to = NULL;
1633
1634	if (timeout > 0) {
1635	struct timespec64 end_time = ep_set_mstimeout(timeout);
1636
1637	slack = select_estimate_accuracy(&end_time);
1638	to = &expires;
1639	*to = timespec64_to_ktime(end_time);
1640	} else if (timeout == 0) {
1641	/*
1642	* Avoid the unnecessary trip to the wait queue loop, if the
1643	* caller specified a non blocking operation.
1644	*/
1645	timed_out = 1;
1646	spin_lock_irqsave(&ep->lock, flags);
1647	goto check_events;
1648	}
1649
1650	fetch_events:
1651	spin_lock_irqsave(&ep->lock, flags);
1652
1653	if (!ep_events_available(ep)) {
1654	/*
1655	* We don't have any available event to return to the caller.
1656	* We need to sleep here, and we will be wake up by
1657	* ep_poll_callback() when events will become available.
1658	*/
1659	init_waitqueue_entry(&wait, current);
1660	__add_wait_queue_exclusive(&ep->wq, &wait);
1661
1662	for (;;) {
1663	/*
1664	* We don't want to sleep if the ep_poll_callback() sends us
1665	* a wakeup in between. That's why we set the task state
1666	* to TASK_INTERRUPTIBLE before doing the checks.
1667	*/
1668	set_current_state(TASK_INTERRUPTIBLE);
1669	if (ep_events_available(ep) || timed_out)
1670	break;
1671	if (signal_pending(current)) {
1672	res = -EINTR;
1673	break;
1674	}
1675
1676	spin_unlock_irqrestore(&ep->lock, flags);
1677	if (!freezable_schedule_hrtimeout_range(to, slack,
1678	HRTIMER_MODE_ABS))
1679	timed_out = 1;
1680
1681	spin_lock_irqsave(&ep->lock, flags);
1682	}
1683
1684	__remove_wait_queue(&ep->wq, &wait);
1685	__set_current_state(TASK_RUNNING);
1686	}
1687	check_events:
1688	/* Is it worth to try to dig for events ? */
1689	eavail = ep_events_available(ep);
1690
1691	spin_unlock_irqrestore(&ep->lock, flags);
1692
1693	/*
1694	* Try to transfer events to user space. In case we get 0 events and
1695	* there's still timeout left over, we go trying again in search of
1696	* more luck.
1697	*/
1698	if (!res && eavail &&
1699	!(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1700	goto fetch_events;
1701
1702	return res;
1703	}
1704
1705	/**
1706	* ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1707	* API, to verify that adding an epoll file inside another
1708	* epoll structure, does not violate the constraints, in
1709	* terms of closed loops, or too deep chains (which can
1710	* result in excessive stack usage).
1711	*
1712	* @priv: Pointer to the epoll file to be currently checked.
1713	* @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1714	* data structure pointer.
1715	* @call_nests: Current dept of the @ep_call_nested() call stack.
1716	*
1717	* Returns: Returns zero if adding the epoll @file inside current epoll
1718	* structure @ep does not violate the constraints, or -1 otherwise.
1719	*/
1720	static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1721	{
1722	int error = 0;
1723	struct file *file = priv;
1724	struct eventpoll *ep = file->private_data;
1725	struct eventpoll *ep_tovisit;
1726	struct rb_node *rbp;
1727	struct epitem *epi;
1728
1729	mutex_lock_nested(&ep->mtx, call_nests + 1);
1730	ep->visited = 1;
1731	list_add(&ep->visited_list_link, &visited_list);
1732	for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1733	epi = rb_entry(rbp, struct epitem, rbn);
1734	if (unlikely(is_file_epoll(epi->ffd.file))) {
1735	ep_tovisit = epi->ffd.file->private_data;
1736	if (ep_tovisit->visited)
1737	continue;
1738	error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1739	ep_loop_check_proc, epi->ffd.file,
1740	ep_tovisit, current);
1741	if (error != 0)
1742	break;
1743	} else {
1744	/*
1745	* If we've reached a file that is not associated with
1746	* an ep, then we need to check if the newly added
1747	* links are going to add too many wakeup paths. We do
1748	* this by adding it to the tfile_check_list, if it's
1749	* not already there, and calling reverse_path_check()
1750	* during ep_insert().
1751	*/
1752	if (list_empty(&epi->ffd.file->f_tfile_llink))
1753	list_add(&epi->ffd.file->f_tfile_llink,
1754	&tfile_check_list);
1755	}
1756	}
1757	mutex_unlock(&ep->mtx);
1758
1759	return error;
1760	}
1761
1762	/**
1763	* ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1764	* another epoll file (represented by @ep) does not create
1765	* closed loops or too deep chains.
1766	*
1767	* @ep: Pointer to the epoll private data structure.
1768	* @file: Pointer to the epoll file to be checked.
1769	*
1770	* Returns: Returns zero if adding the epoll @file inside current epoll
1771	* structure @ep does not violate the constraints, or -1 otherwise.
1772	*/
1773	static int ep_loop_check(struct eventpoll *ep, struct file *file)
1774	{
1775	int ret;
1776	struct eventpoll *ep_cur, *ep_next;
1777
1778	ret = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1779	ep_loop_check_proc, file, ep, current);
1780	/* clear visited list */
1781	list_for_each_entry_safe(ep_cur, ep_next, &visited_list,
1782	visited_list_link) {
1783	ep_cur->visited = 0;
1784	list_del(&ep_cur->visited_list_link);
1785	}
1786	return ret;
1787	}
1788
1789	static void clear_tfile_check_list(void)
1790	{
1791	struct file *file;
1792
1793	/* first clear the tfile_check_list */
1794	while (!list_empty(&tfile_check_list)) {
1795	file = list_first_entry(&tfile_check_list, struct file,
1796	f_tfile_llink);
1797	list_del_init(&file->f_tfile_llink);
1798	}
1799	INIT_LIST_HEAD(&tfile_check_list);
1800	}
1801
1802	/*
1803	* Open an eventpoll file descriptor.
1804	*/
1805	SYSCALL_DEFINE1(epoll_create1, int, flags)
1806	{
1807	int error, fd;
1808	struct eventpoll *ep = NULL;
1809	struct file *file;
1810
1811	/* Check the EPOLL_* constant for consistency. */
1812	BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1813
1814	if (flags & ~EPOLL_CLOEXEC)
1815	return -EINVAL;
1816	/*
1817	* Create the internal data structure ("struct eventpoll").
1818	*/
1819	error = ep_alloc(&ep);
1820	if (error < 0)
1821	return error;
1822	/*
1823	* Creates all the items needed to setup an eventpoll file. That is,
1824	* a file structure and a free file descriptor.
1825	*/
1826	fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1827	if (fd < 0) {
1828	error = fd;
1829	goto out_free_ep;
1830	}
1831	file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
1832	O_RDWR | (flags & O_CLOEXEC));
1833	if (IS_ERR(file)) {
1834	error = PTR_ERR(file);
1835	goto out_free_fd;
1836	}
1837	ep->file = file;
1838	fd_install(fd, file);
1839	return fd;
1840
1841	out_free_fd:
1842	put_unused_fd(fd);
1843	out_free_ep:
1844	ep_free(ep);
1845	return error;
1846	}
1847
1848	SYSCALL_DEFINE1(epoll_create, int, size)
1849	{
1850	if (size <= 0)
1851	return -EINVAL;
1852
1853	return sys_epoll_create1(0);
1854	}
1855
1856	/*
1857	* The following function implements the controller interface for
1858	* the eventpoll file that enables the insertion/removal/change of
1859	* file descriptors inside the interest set.
1860	*/
1861	SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1862	struct epoll_event __user *, event)
1863	{
1864	int error;
1865	int full_check = 0;
1866	struct fd f, tf;
1867	struct eventpoll *ep;
1868	struct epitem *epi;
1869	struct epoll_event epds;
1870	struct eventpoll *tep = NULL;
1871
1872	error = -EFAULT;
1873	if (ep_op_has_event(op) &&
1874	copy_from_user(&epds, event, sizeof(struct epoll_event)))
1875	goto error_return;
1876
1877	error = -EBADF;
1878	f = fdget(epfd);
1879	if (!f.file)
1880	goto error_return;
1881
1882	/* Get the "struct file *" for the target file */
1883	tf = fdget(fd);
1884	if (!tf.file)
1885	goto error_fput;
1886
1887	/* The target file descriptor must support poll */
1888	error = -EPERM;
1889	if (!tf.file->f_op->poll)
1890	goto error_tgt_fput;
1891
1892	/* Check if EPOLLWAKEUP is allowed */
1893	if (ep_op_has_event(op))
1894	ep_take_care_of_epollwakeup(&epds);
1895
1896	/*
1897	* We have to check that the file structure underneath the file descriptor
1898	* the user passed to us _is_ an eventpoll file. And also we do not permit
1899	* adding an epoll file descriptor inside itself.
1900	*/
1901	error = -EINVAL;
1902	if (f.file == tf.file || !is_file_epoll(f.file))
1903	goto error_tgt_fput;
1904
1905	/*
1906	* epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
1907	* so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
1908	* Also, we do not currently supported nested exclusive wakeups.
1909	*/
1910	if (epds.events & EPOLLEXCLUSIVE) {
1911	if (op == EPOLL_CTL_MOD)
1912	goto error_tgt_fput;
1913	if (op == EPOLL_CTL_ADD && (is_file_epoll(tf.file) ||
1914	(epds.events & ~EPOLLEXCLUSIVE_OK_BITS)))
1915	goto error_tgt_fput;
1916	}
1917
1918	/*
1919	* At this point it is safe to assume that the "private_data" contains
1920	* our own data structure.
1921	*/
1922	ep = f.file->private_data;
1923
1924	/*
1925	* When we insert an epoll file descriptor, inside another epoll file
1926	* descriptor, there is the change of creating closed loops, which are
1927	* better be handled here, than in more critical paths. While we are
1928	* checking for loops we also determine the list of files reachable
1929	* and hang them on the tfile_check_list, so we can check that we
1930	* haven't created too many possible wakeup paths.
1931	*
1932	* We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
1933	* the epoll file descriptor is attaching directly to a wakeup source,
1934	* unless the epoll file descriptor is nested. The purpose of taking the
1935	* 'epmutex' on add is to prevent complex toplogies such as loops and
1936	* deep wakeup paths from forming in parallel through multiple
1937	* EPOLL_CTL_ADD operations.
1938	*/
1939	mutex_lock_nested(&ep->mtx, 0);
1940	if (op == EPOLL_CTL_ADD) {
1941	if (!list_empty(&f.file->f_ep_links) ||
1942	is_file_epoll(tf.file)) {
1943	full_check = 1;
1944	mutex_unlock(&ep->mtx);
1945	mutex_lock(&epmutex);
1946	if (is_file_epoll(tf.file)) {
1947	error = -ELOOP;
1948	if (ep_loop_check(ep, tf.file) != 0) {
1949	clear_tfile_check_list();
1950	goto error_tgt_fput;
1951	}
1952	} else
1953	list_add(&tf.file->f_tfile_llink,
1954	&tfile_check_list);
1955	mutex_lock_nested(&ep->mtx, 0);
1956	if (is_file_epoll(tf.file)) {
1957	tep = tf.file->private_data;
1958	mutex_lock_nested(&tep->mtx, 1);
1959	}
1960	}
1961	}
1962
1963	/*
1964	* Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1965	* above, we can be sure to be able to use the item looked up by
1966	* ep_find() till we release the mutex.
1967	*/
1968	epi = ep_find(ep, tf.file, fd);
1969
1970	error = -EINVAL;
1971	switch (op) {
1972	case EPOLL_CTL_ADD:
1973	if (!epi) {
1974	epds.events |= POLLERR | POLLHUP;
1975	error = ep_insert(ep, &epds, tf.file, fd, full_check);
1976	} else
1977	error = -EEXIST;
1978	if (full_check)
1979	clear_tfile_check_list();
1980	break;
1981	case EPOLL_CTL_DEL:
1982	if (epi)
1983	error = ep_remove(ep, epi);
1984	else
1985	error = -ENOENT;
1986	break;
1987	case EPOLL_CTL_MOD:
1988	if (epi) {
1989	if (!(epi->event.events & EPOLLEXCLUSIVE)) {
1990	epds.events |= POLLERR | POLLHUP;
1991	error = ep_modify(ep, epi, &epds);
1992	}
1993	} else
1994	error = -ENOENT;
1995	break;
1996	}
1997	if (tep != NULL)
1998	mutex_unlock(&tep->mtx);
1999	mutex_unlock(&ep->mtx);
2000
2001	error_tgt_fput:
2002	if (full_check)
2003	mutex_unlock(&epmutex);
2004
2005	fdput(tf);
2006	error_fput:
2007	fdput(f);
2008	error_return:
2009
2010	return error;
2011	}
2012
2013	/*
2014	* Implement the event wait interface for the eventpoll file. It is the kernel
2015	* part of the user space epoll_wait(2).
2016	*/
2017	SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
2018	int, maxevents, int, timeout)
2019	{
2020	int error;
2021	struct fd f;
2022	struct eventpoll *ep;
2023
2024	/* The maximum number of event must be greater than zero */
2025	if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
2026	return -EINVAL;
2027
2028	/* Verify that the area passed by the user is writeable */
2029	if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event)))
2030	return -EFAULT;
2031
2032	/* Get the "struct file *" for the eventpoll file */
2033	f = fdget(epfd);
2034	if (!f.file)
2035	return -EBADF;
2036
2037	/*
2038	* We have to check that the file structure underneath the fd
2039	* the user passed to us _is_ an eventpoll file.
2040	*/
2041	error = -EINVAL;
2042	if (!is_file_epoll(f.file))
2043	goto error_fput;
2044
2045	/*
2046	* At this point it is safe to assume that the "private_data" contains
2047	* our own data structure.
2048	*/
2049	ep = f.file->private_data;
2050
2051	/* Time to fish for events ... */
2052	error = ep_poll(ep, events, maxevents, timeout);
2053
2054	error_fput:
2055	fdput(f);
2056	return error;
2057	}
2058
2059	/*
2060	* Implement the event wait interface for the eventpoll file. It is the kernel
2061	* part of the user space epoll_pwait(2).
2062	*/
2063	SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2064	int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2065	size_t, sigsetsize)
2066	{
2067	int error;
2068	sigset_t ksigmask, sigsaved;
2069
2070	/*
2071	* If the caller wants a certain signal mask to be set during the wait,
2072	* we apply it here.
2073	*/
2074	if (sigmask) {
2075	if (sigsetsize != sizeof(sigset_t))
2076	return -EINVAL;
2077	if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
2078	return -EFAULT;
2079	sigsaved = current->blocked;
2080	set_current_blocked(&ksigmask);
2081	}
2082
2083	error = sys_epoll_wait(epfd, events, maxevents, timeout);
2084
2085	/*
2086	* If we changed the signal mask, we need to restore the original one.
2087	* In case we've got a signal while waiting, we do not restore the
2088	* signal mask yet, and we allow do_signal() to deliver the signal on
2089	* the way back to userspace, before the signal mask is restored.
2090	*/
2091	if (sigmask) {
2092	if (error == -EINTR) {
2093	memcpy(&current->saved_sigmask, &sigsaved,
2094	sizeof(sigsaved));
2095	set_restore_sigmask();
2096	} else
2097	set_current_blocked(&sigsaved);
2098	}
2099
2100	return error;
2101	}
2102
2103	#ifdef CONFIG_COMPAT
2104	COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2105	struct epoll_event __user *, events,
2106	int, maxevents, int, timeout,
2107	const compat_sigset_t __user *, sigmask,
2108	compat_size_t, sigsetsize)
2109	{
2110	long err;
2111	compat_sigset_t csigmask;
2112	sigset_t ksigmask, sigsaved;
2113
2114	/*
2115	* If the caller wants a certain signal mask to be set during the wait,
2116	* we apply it here.
2117	*/
2118	if (sigmask) {
2119	if (sigsetsize != sizeof(compat_sigset_t))
2120	return -EINVAL;
2121	if (copy_from_user(&csigmask, sigmask, sizeof(csigmask)))
2122	return -EFAULT;
2123	sigset_from_compat(&ksigmask, &csigmask);
2124	sigsaved = current->blocked;
2125	set_current_blocked(&ksigmask);
2126	}
2127
2128	err = sys_epoll_wait(epfd, events, maxevents, timeout);
2129
2130	/*
2131	* If we changed the signal mask, we need to restore the original one.
2132	* In case we've got a signal while waiting, we do not restore the
2133	* signal mask yet, and we allow do_signal() to deliver the signal on
2134	* the way back to userspace, before the signal mask is restored.
2135	*/
2136	if (sigmask) {
2137	if (err == -EINTR) {
2138	memcpy(&current->saved_sigmask, &sigsaved,
2139	sizeof(sigsaved));
2140	set_restore_sigmask();
2141	} else
2142	set_current_blocked(&sigsaved);
2143	}
2144
2145	return err;
2146	}
2147	#endif
2148
2149	static int __init eventpoll_init(void)
2150	{
2151	struct sysinfo si;
2152
2153	si_meminfo(&si);
2154	/*
2155	* Allows top 4% of lomem to be allocated for epoll watches (per user).
2156	*/
2157	max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2158	EP_ITEM_COST;
2159	BUG_ON(max_user_watches < 0);
2160
2161	/*
2162	* Initialize the structure used to perform epoll file descriptor
2163	* inclusion loops checks.
2164	*/
2165	ep_nested_calls_init(&poll_loop_ncalls);
2166
2167	/* Initialize the structure used to perform safe poll wait head wake ups */
2168	ep_nested_calls_init(&poll_safewake_ncalls);
2169
2170	/* Initialize the structure used to perform file's f_op->poll() calls */
2171	ep_nested_calls_init(&poll_readywalk_ncalls);
2172
2173	/*
2174	* We can have many thousands of epitems, so prevent this from
2175	* using an extra cache line on 64-bit (and smaller) CPUs
2176	*/
2177	BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2178
2179	/* Allocates slab cache used to allocate "struct epitem" items */
2180	epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2181	0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2182
2183	/* Allocates slab cache used to allocate "struct eppoll_entry" */
2184	pwq_cache = kmem_cache_create("eventpoll_pwq",
2185	sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
2186
2187	return 0;
2188	}
2189	fs_initcall(eventpoll_init);
2190