xref: /openbmc/linux/fs/eventpoll.c (revision f20c7d91)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  *  fs/eventpoll.c (Efficient event retrieval implementation)
4  *  Copyright (C) 2001,...,2009	 Davide Libenzi
5  *
6  *  Davide Libenzi <davidel@xmailserver.org>
7  */
8 
9 #include <linux/init.h>
10 #include <linux/kernel.h>
11 #include <linux/sched/signal.h>
12 #include <linux/fs.h>
13 #include <linux/file.h>
14 #include <linux/signal.h>
15 #include <linux/errno.h>
16 #include <linux/mm.h>
17 #include <linux/slab.h>
18 #include <linux/poll.h>
19 #include <linux/string.h>
20 #include <linux/list.h>
21 #include <linux/hash.h>
22 #include <linux/spinlock.h>
23 #include <linux/syscalls.h>
24 #include <linux/rbtree.h>
25 #include <linux/wait.h>
26 #include <linux/eventpoll.h>
27 #include <linux/mount.h>
28 #include <linux/bitops.h>
29 #include <linux/mutex.h>
30 #include <linux/anon_inodes.h>
31 #include <linux/device.h>
32 #include <linux/uaccess.h>
33 #include <asm/io.h>
34 #include <asm/mman.h>
35 #include <linux/atomic.h>
36 #include <linux/proc_fs.h>
37 #include <linux/seq_file.h>
38 #include <linux/compat.h>
39 #include <linux/rculist.h>
40 #include <net/busy_poll.h>
41 
42 /*
43  * LOCKING:
44  * There are three level of locking required by epoll :
45  *
46  * 1) epmutex (mutex)
47  * 2) ep->mtx (mutex)
48  * 3) ep->lock (rwlock)
49  *
50  * The acquire order is the one listed above, from 1 to 3.
51  * We need a rwlock (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()
61  * and ep_free().
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
72  * going to.
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
81  * the lockdep subkey.
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.
88  */
89 
90 /* Epoll private bits inside the event mask */
91 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET | EPOLLEXCLUSIVE)
92 
93 #define EPOLLINOUT_BITS (EPOLLIN | EPOLLOUT)
94 
95 #define EPOLLEXCLUSIVE_OK_BITS (EPOLLINOUT_BITS | EPOLLERR | EPOLLHUP | \
96 				EPOLLWAKEUP | EPOLLET | EPOLLEXCLUSIVE)
97 
98 /* Maximum number of nesting allowed inside epoll sets */
99 #define EP_MAX_NESTS 4
100 
101 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
102 
103 #define EP_UNACTIVE_PTR ((void *) -1L)
104 
105 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
106 
107 struct epoll_filefd {
108 	struct file *file;
109 	int fd;
110 } __packed;
111 
112 /*
113  * Structure used to track possible nested calls, for too deep recursions
114  * and loop cycles.
115  */
116 struct nested_call_node {
117 	struct list_head llink;
118 	void *cookie;
119 	void *ctx;
120 };
121 
122 /*
123  * This structure is used as collector for nested calls, to check for
124  * maximum recursion dept and loop cycles.
125  */
126 struct nested_calls {
127 	struct list_head tasks_call_list;
128 	spinlock_t lock;
129 };
130 
131 /*
132  * Each file descriptor added to the eventpoll interface will
133  * have an entry of this type linked to the "rbr" RB tree.
134  * Avoid increasing the size of this struct, there can be many thousands
135  * of these on a server and we do not want this to take another cache line.
136  */
137 struct epitem {
138 	union {
139 		/* RB tree node links this structure to the eventpoll RB tree */
140 		struct rb_node rbn;
141 		/* Used to free the struct epitem */
142 		struct rcu_head rcu;
143 	};
144 
145 	/* List header used to link this structure to the eventpoll ready list */
146 	struct list_head rdllink;
147 
148 	/*
149 	 * Works together "struct eventpoll"->ovflist in keeping the
150 	 * single linked chain of items.
151 	 */
152 	struct epitem *next;
153 
154 	/* The file descriptor information this item refers to */
155 	struct epoll_filefd ffd;
156 
157 	/* Number of active wait queue attached to poll operations */
158 	int nwait;
159 
160 	/* List containing poll wait queues */
161 	struct list_head pwqlist;
162 
163 	/* The "container" of this item */
164 	struct eventpoll *ep;
165 
166 	/* List header used to link this item to the "struct file" items list */
167 	struct list_head fllink;
168 
169 	/* wakeup_source used when EPOLLWAKEUP is set */
170 	struct wakeup_source __rcu *ws;
171 
172 	/* The structure that describe the interested events and the source fd */
173 	struct epoll_event event;
174 };
175 
176 /*
177  * This structure is stored inside the "private_data" member of the file
178  * structure and represents the main data structure for the eventpoll
179  * interface.
180  */
181 struct eventpoll {
182 	/*
183 	 * This mutex is used to ensure that files are not removed
184 	 * while epoll is using them. This is held during the event
185 	 * collection loop, the file cleanup path, the epoll file exit
186 	 * code and the ctl operations.
187 	 */
188 	struct mutex mtx;
189 
190 	/* Wait queue used by sys_epoll_wait() */
191 	wait_queue_head_t wq;
192 
193 	/* Wait queue used by file->poll() */
194 	wait_queue_head_t poll_wait;
195 
196 	/* List of ready file descriptors */
197 	struct list_head rdllist;
198 
199 	/* Lock which protects rdllist and ovflist */
200 	rwlock_t lock;
201 
202 	/* RB tree root used to store monitored fd structs */
203 	struct rb_root_cached rbr;
204 
205 	/*
206 	 * This is a single linked list that chains all the "struct epitem" that
207 	 * happened while transferring ready events to userspace w/out
208 	 * holding ->lock.
209 	 */
210 	struct epitem *ovflist;
211 
212 	/* wakeup_source used when ep_scan_ready_list is running */
213 	struct wakeup_source *ws;
214 
215 	/* The user that created the eventpoll descriptor */
216 	struct user_struct *user;
217 
218 	struct file *file;
219 
220 	/* used to optimize loop detection check */
221 	struct list_head visited_list_link;
222 	int visited;
223 
224 #ifdef CONFIG_NET_RX_BUSY_POLL
225 	/* used to track busy poll napi_id */
226 	unsigned int napi_id;
227 #endif
228 
229 #ifdef CONFIG_DEBUG_LOCK_ALLOC
230 	/* tracks wakeup nests for lockdep validation */
231 	u8 nests;
232 #endif
233 };
234 
235 /* Wait structure used by the poll hooks */
236 struct eppoll_entry {
237 	/* List header used to link this structure to the "struct epitem" */
238 	struct list_head llink;
239 
240 	/* The "base" pointer is set to the container "struct epitem" */
241 	struct epitem *base;
242 
243 	/*
244 	 * Wait queue item that will be linked to the target file wait
245 	 * queue head.
246 	 */
247 	wait_queue_entry_t wait;
248 
249 	/* The wait queue head that linked the "wait" wait queue item */
250 	wait_queue_head_t *whead;
251 };
252 
253 /* Wrapper struct used by poll queueing */
254 struct ep_pqueue {
255 	poll_table pt;
256 	struct epitem *epi;
257 };
258 
259 /* Used by the ep_send_events() function as callback private data */
260 struct ep_send_events_data {
261 	int maxevents;
262 	struct epoll_event __user *events;
263 	int res;
264 };
265 
266 /*
267  * Configuration options available inside /proc/sys/fs/epoll/
268  */
269 /* Maximum number of epoll watched descriptors, per user */
270 static long max_user_watches __read_mostly;
271 
272 /*
273  * This mutex is used to serialize ep_free() and eventpoll_release_file().
274  */
275 static DEFINE_MUTEX(epmutex);
276 
277 /* Used to check for epoll file descriptor inclusion loops */
278 static struct nested_calls poll_loop_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 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		= &long_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 epitem *epi)
341 {
342 	return !list_empty(&epi->rdllink);
343 }
344 
345 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_entry_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_entry_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 /* Initialize the poll safe wake up structure */
363 static void ep_nested_calls_init(struct nested_calls *ncalls)
364 {
365 	INIT_LIST_HEAD(&ncalls->tasks_call_list);
366 	spin_lock_init(&ncalls->lock);
367 }
368 
369 /**
370  * ep_events_available - Checks if ready events might be available.
371  *
372  * @ep: Pointer to the eventpoll context.
373  *
374  * Returns: Returns a value different than zero if ready events are available,
375  *          or zero otherwise.
376  */
377 static inline int ep_events_available(struct eventpoll *ep)
378 {
379 	return !list_empty_careful(&ep->rdllist) ||
380 		READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR;
381 }
382 
383 #ifdef CONFIG_NET_RX_BUSY_POLL
384 static bool ep_busy_loop_end(void *p, unsigned long start_time)
385 {
386 	struct eventpoll *ep = p;
387 
388 	return ep_events_available(ep) || busy_loop_timeout(start_time);
389 }
390 
391 /*
392  * Busy poll if globally on and supporting sockets found && no events,
393  * busy loop will return if need_resched or ep_events_available.
394  *
395  * we must do our busy polling with irqs enabled
396  */
397 static void ep_busy_loop(struct eventpoll *ep, int nonblock)
398 {
399 	unsigned int napi_id = READ_ONCE(ep->napi_id);
400 
401 	if ((napi_id >= MIN_NAPI_ID) && net_busy_loop_on())
402 		napi_busy_loop(napi_id, nonblock ? NULL : ep_busy_loop_end, ep);
403 }
404 
405 static inline void ep_reset_busy_poll_napi_id(struct eventpoll *ep)
406 {
407 	if (ep->napi_id)
408 		ep->napi_id = 0;
409 }
410 
411 /*
412  * Set epoll busy poll NAPI ID from sk.
413  */
414 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
415 {
416 	struct eventpoll *ep;
417 	unsigned int napi_id;
418 	struct socket *sock;
419 	struct sock *sk;
420 	int err;
421 
422 	if (!net_busy_loop_on())
423 		return;
424 
425 	sock = sock_from_file(epi->ffd.file, &err);
426 	if (!sock)
427 		return;
428 
429 	sk = sock->sk;
430 	if (!sk)
431 		return;
432 
433 	napi_id = READ_ONCE(sk->sk_napi_id);
434 	ep = epi->ep;
435 
436 	/* Non-NAPI IDs can be rejected
437 	 *	or
438 	 * Nothing to do if we already have this ID
439 	 */
440 	if (napi_id < MIN_NAPI_ID || napi_id == ep->napi_id)
441 		return;
442 
443 	/* record NAPI ID for use in next busy poll */
444 	ep->napi_id = napi_id;
445 }
446 
447 #else
448 
449 static inline void ep_busy_loop(struct eventpoll *ep, int nonblock)
450 {
451 }
452 
453 static inline void ep_reset_busy_poll_napi_id(struct eventpoll *ep)
454 {
455 }
456 
457 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
458 {
459 }
460 
461 #endif /* CONFIG_NET_RX_BUSY_POLL */
462 
463 /**
464  * ep_call_nested - Perform a bound (possibly) nested call, by checking
465  *                  that the recursion limit is not exceeded, and that
466  *                  the same nested call (by the meaning of same cookie) is
467  *                  no re-entered.
468  *
469  * @ncalls: Pointer to the nested_calls structure to be used for this call.
470  * @nproc: Nested call core function pointer.
471  * @priv: Opaque data to be passed to the @nproc callback.
472  * @cookie: Cookie to be used to identify this nested call.
473  * @ctx: This instance context.
474  *
475  * Returns: Returns the code returned by the @nproc callback, or -1 if
476  *          the maximum recursion limit has been exceeded.
477  */
478 static int ep_call_nested(struct nested_calls *ncalls,
479 			  int (*nproc)(void *, void *, int), void *priv,
480 			  void *cookie, void *ctx)
481 {
482 	int error, call_nests = 0;
483 	unsigned long flags;
484 	struct list_head *lsthead = &ncalls->tasks_call_list;
485 	struct nested_call_node *tncur;
486 	struct nested_call_node tnode;
487 
488 	spin_lock_irqsave(&ncalls->lock, flags);
489 
490 	/*
491 	 * Try to see if the current task is already inside this wakeup call.
492 	 * We use a list here, since the population inside this set is always
493 	 * very much limited.
494 	 */
495 	list_for_each_entry(tncur, lsthead, llink) {
496 		if (tncur->ctx == ctx &&
497 		    (tncur->cookie == cookie || ++call_nests > EP_MAX_NESTS)) {
498 			/*
499 			 * Ops ... loop detected or maximum nest level reached.
500 			 * We abort this wake by breaking the cycle itself.
501 			 */
502 			error = -1;
503 			goto out_unlock;
504 		}
505 	}
506 
507 	/* Add the current task and cookie to the list */
508 	tnode.ctx = ctx;
509 	tnode.cookie = cookie;
510 	list_add(&tnode.llink, lsthead);
511 
512 	spin_unlock_irqrestore(&ncalls->lock, flags);
513 
514 	/* Call the nested function */
515 	error = (*nproc)(priv, cookie, call_nests);
516 
517 	/* Remove the current task from the list */
518 	spin_lock_irqsave(&ncalls->lock, flags);
519 	list_del(&tnode.llink);
520 out_unlock:
521 	spin_unlock_irqrestore(&ncalls->lock, flags);
522 
523 	return error;
524 }
525 
526 /*
527  * As described in commit 0ccf831cb lockdep: annotate epoll
528  * the use of wait queues used by epoll is done in a very controlled
529  * manner. Wake ups can nest inside each other, but are never done
530  * with the same locking. For example:
531  *
532  *   dfd = socket(...);
533  *   efd1 = epoll_create();
534  *   efd2 = epoll_create();
535  *   epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
536  *   epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
537  *
538  * When a packet arrives to the device underneath "dfd", the net code will
539  * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
540  * callback wakeup entry on that queue, and the wake_up() performed by the
541  * "dfd" net code will end up in ep_poll_callback(). At this point epoll
542  * (efd1) notices that it may have some event ready, so it needs to wake up
543  * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
544  * that ends up in another wake_up(), after having checked about the
545  * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
546  * avoid stack blasting.
547  *
548  * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
549  * this special case of epoll.
550  */
551 #ifdef CONFIG_DEBUG_LOCK_ALLOC
552 
553 static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi)
554 {
555 	struct eventpoll *ep_src;
556 	unsigned long flags;
557 	u8 nests = 0;
558 
559 	/*
560 	 * To set the subclass or nesting level for spin_lock_irqsave_nested()
561 	 * it might be natural to create a per-cpu nest count. However, since
562 	 * we can recurse on ep->poll_wait.lock, and a non-raw spinlock can
563 	 * schedule() in the -rt kernel, the per-cpu variable are no longer
564 	 * protected. Thus, we are introducing a per eventpoll nest field.
565 	 * If we are not being call from ep_poll_callback(), epi is NULL and
566 	 * we are at the first level of nesting, 0. Otherwise, we are being
567 	 * called from ep_poll_callback() and if a previous wakeup source is
568 	 * not an epoll file itself, we are at depth 1 since the wakeup source
569 	 * is depth 0. If the wakeup source is a previous epoll file in the
570 	 * wakeup chain then we use its nests value and record ours as
571 	 * nests + 1. The previous epoll file nests value is stable since its
572 	 * already holding its own poll_wait.lock.
573 	 */
574 	if (epi) {
575 		if ((is_file_epoll(epi->ffd.file))) {
576 			ep_src = epi->ffd.file->private_data;
577 			nests = ep_src->nests;
578 		} else {
579 			nests = 1;
580 		}
581 	}
582 	spin_lock_irqsave_nested(&ep->poll_wait.lock, flags, nests);
583 	ep->nests = nests + 1;
584 	wake_up_locked_poll(&ep->poll_wait, EPOLLIN);
585 	ep->nests = 0;
586 	spin_unlock_irqrestore(&ep->poll_wait.lock, flags);
587 }
588 
589 #else
590 
591 static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi)
592 {
593 	wake_up_poll(&ep->poll_wait, EPOLLIN);
594 }
595 
596 #endif
597 
598 static void ep_remove_wait_queue(struct eppoll_entry *pwq)
599 {
600 	wait_queue_head_t *whead;
601 
602 	rcu_read_lock();
603 	/*
604 	 * If it is cleared by POLLFREE, it should be rcu-safe.
605 	 * If we read NULL we need a barrier paired with
606 	 * smp_store_release() in ep_poll_callback(), otherwise
607 	 * we rely on whead->lock.
608 	 */
609 	whead = smp_load_acquire(&pwq->whead);
610 	if (whead)
611 		remove_wait_queue(whead, &pwq->wait);
612 	rcu_read_unlock();
613 }
614 
615 /*
616  * This function unregisters poll callbacks from the associated file
617  * descriptor.  Must be called with "mtx" held (or "epmutex" if called from
618  * ep_free).
619  */
620 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
621 {
622 	struct list_head *lsthead = &epi->pwqlist;
623 	struct eppoll_entry *pwq;
624 
625 	while (!list_empty(lsthead)) {
626 		pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
627 
628 		list_del(&pwq->llink);
629 		ep_remove_wait_queue(pwq);
630 		kmem_cache_free(pwq_cache, pwq);
631 	}
632 }
633 
634 /* call only when ep->mtx is held */
635 static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
636 {
637 	return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
638 }
639 
640 /* call only when ep->mtx is held */
641 static inline void ep_pm_stay_awake(struct epitem *epi)
642 {
643 	struct wakeup_source *ws = ep_wakeup_source(epi);
644 
645 	if (ws)
646 		__pm_stay_awake(ws);
647 }
648 
649 static inline bool ep_has_wakeup_source(struct epitem *epi)
650 {
651 	return rcu_access_pointer(epi->ws) ? true : false;
652 }
653 
654 /* call when ep->mtx cannot be held (ep_poll_callback) */
655 static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
656 {
657 	struct wakeup_source *ws;
658 
659 	rcu_read_lock();
660 	ws = rcu_dereference(epi->ws);
661 	if (ws)
662 		__pm_stay_awake(ws);
663 	rcu_read_unlock();
664 }
665 
666 /**
667  * ep_scan_ready_list - Scans the ready list in a way that makes possible for
668  *                      the scan code, to call f_op->poll(). Also allows for
669  *                      O(NumReady) performance.
670  *
671  * @ep: Pointer to the epoll private data structure.
672  * @sproc: Pointer to the scan callback.
673  * @priv: Private opaque data passed to the @sproc callback.
674  * @depth: The current depth of recursive f_op->poll calls.
675  * @ep_locked: caller already holds ep->mtx
676  *
677  * Returns: The same integer error code returned by the @sproc callback.
678  */
679 static __poll_t ep_scan_ready_list(struct eventpoll *ep,
680 			      __poll_t (*sproc)(struct eventpoll *,
681 					   struct list_head *, void *),
682 			      void *priv, int depth, bool ep_locked)
683 {
684 	__poll_t res;
685 	struct epitem *epi, *nepi;
686 	LIST_HEAD(txlist);
687 
688 	lockdep_assert_irqs_enabled();
689 
690 	/*
691 	 * We need to lock this because we could be hit by
692 	 * eventpoll_release_file() and epoll_ctl().
693 	 */
694 
695 	if (!ep_locked)
696 		mutex_lock_nested(&ep->mtx, depth);
697 
698 	/*
699 	 * Steal the ready list, and re-init the original one to the
700 	 * empty list. Also, set ep->ovflist to NULL so that events
701 	 * happening while looping w/out locks, are not lost. We cannot
702 	 * have the poll callback to queue directly on ep->rdllist,
703 	 * because we want the "sproc" callback to be able to do it
704 	 * in a lockless way.
705 	 */
706 	write_lock_irq(&ep->lock);
707 	list_splice_init(&ep->rdllist, &txlist);
708 	WRITE_ONCE(ep->ovflist, NULL);
709 	write_unlock_irq(&ep->lock);
710 
711 	/*
712 	 * Now call the callback function.
713 	 */
714 	res = (*sproc)(ep, &txlist, priv);
715 
716 	write_lock_irq(&ep->lock);
717 	/*
718 	 * During the time we spent inside the "sproc" callback, some
719 	 * other events might have been queued by the poll callback.
720 	 * We re-insert them inside the main ready-list here.
721 	 */
722 	for (nepi = READ_ONCE(ep->ovflist); (epi = nepi) != NULL;
723 	     nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
724 		/*
725 		 * We need to check if the item is already in the list.
726 		 * During the "sproc" callback execution time, items are
727 		 * queued into ->ovflist but the "txlist" might already
728 		 * contain them, and the list_splice() below takes care of them.
729 		 */
730 		if (!ep_is_linked(epi)) {
731 			/*
732 			 * ->ovflist is LIFO, so we have to reverse it in order
733 			 * to keep in FIFO.
734 			 */
735 			list_add(&epi->rdllink, &ep->rdllist);
736 			ep_pm_stay_awake(epi);
737 		}
738 	}
739 	/*
740 	 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
741 	 * releasing the lock, events will be queued in the normal way inside
742 	 * ep->rdllist.
743 	 */
744 	WRITE_ONCE(ep->ovflist, EP_UNACTIVE_PTR);
745 
746 	/*
747 	 * Quickly re-inject items left on "txlist".
748 	 */
749 	list_splice(&txlist, &ep->rdllist);
750 	__pm_relax(ep->ws);
751 	write_unlock_irq(&ep->lock);
752 
753 	if (!ep_locked)
754 		mutex_unlock(&ep->mtx);
755 
756 	return res;
757 }
758 
759 static void epi_rcu_free(struct rcu_head *head)
760 {
761 	struct epitem *epi = container_of(head, struct epitem, rcu);
762 	kmem_cache_free(epi_cache, epi);
763 }
764 
765 /*
766  * Removes a "struct epitem" from the eventpoll RB tree and deallocates
767  * all the associated resources. Must be called with "mtx" held.
768  */
769 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
770 {
771 	struct file *file = epi->ffd.file;
772 
773 	lockdep_assert_irqs_enabled();
774 
775 	/*
776 	 * Removes poll wait queue hooks.
777 	 */
778 	ep_unregister_pollwait(ep, epi);
779 
780 	/* Remove the current item from the list of epoll hooks */
781 	spin_lock(&file->f_lock);
782 	list_del_rcu(&epi->fllink);
783 	spin_unlock(&file->f_lock);
784 
785 	rb_erase_cached(&epi->rbn, &ep->rbr);
786 
787 	write_lock_irq(&ep->lock);
788 	if (ep_is_linked(epi))
789 		list_del_init(&epi->rdllink);
790 	write_unlock_irq(&ep->lock);
791 
792 	wakeup_source_unregister(ep_wakeup_source(epi));
793 	/*
794 	 * At this point it is safe to free the eventpoll item. Use the union
795 	 * field epi->rcu, since we are trying to minimize the size of
796 	 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
797 	 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
798 	 * use of the rbn field.
799 	 */
800 	call_rcu(&epi->rcu, epi_rcu_free);
801 
802 	atomic_long_dec(&ep->user->epoll_watches);
803 
804 	return 0;
805 }
806 
807 static void ep_free(struct eventpoll *ep)
808 {
809 	struct rb_node *rbp;
810 	struct epitem *epi;
811 
812 	/* We need to release all tasks waiting for these file */
813 	if (waitqueue_active(&ep->poll_wait))
814 		ep_poll_safewake(ep, NULL);
815 
816 	/*
817 	 * We need to lock this because we could be hit by
818 	 * eventpoll_release_file() while we're freeing the "struct eventpoll".
819 	 * We do not need to hold "ep->mtx" here because the epoll file
820 	 * is on the way to be removed and no one has references to it
821 	 * anymore. The only hit might come from eventpoll_release_file() but
822 	 * holding "epmutex" is sufficient here.
823 	 */
824 	mutex_lock(&epmutex);
825 
826 	/*
827 	 * Walks through the whole tree by unregistering poll callbacks.
828 	 */
829 	for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
830 		epi = rb_entry(rbp, struct epitem, rbn);
831 
832 		ep_unregister_pollwait(ep, epi);
833 		cond_resched();
834 	}
835 
836 	/*
837 	 * Walks through the whole tree by freeing each "struct epitem". At this
838 	 * point we are sure no poll callbacks will be lingering around, and also by
839 	 * holding "epmutex" we can be sure that no file cleanup code will hit
840 	 * us during this operation. So we can avoid the lock on "ep->lock".
841 	 * We do not need to lock ep->mtx, either, we only do it to prevent
842 	 * a lockdep warning.
843 	 */
844 	mutex_lock(&ep->mtx);
845 	while ((rbp = rb_first_cached(&ep->rbr)) != NULL) {
846 		epi = rb_entry(rbp, struct epitem, rbn);
847 		ep_remove(ep, epi);
848 		cond_resched();
849 	}
850 	mutex_unlock(&ep->mtx);
851 
852 	mutex_unlock(&epmutex);
853 	mutex_destroy(&ep->mtx);
854 	free_uid(ep->user);
855 	wakeup_source_unregister(ep->ws);
856 	kfree(ep);
857 }
858 
859 static int ep_eventpoll_release(struct inode *inode, struct file *file)
860 {
861 	struct eventpoll *ep = file->private_data;
862 
863 	if (ep)
864 		ep_free(ep);
865 
866 	return 0;
867 }
868 
869 static __poll_t ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
870 			       void *priv);
871 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
872 				 poll_table *pt);
873 
874 /*
875  * Differs from ep_eventpoll_poll() in that internal callers already have
876  * the ep->mtx so we need to start from depth=1, such that mutex_lock_nested()
877  * is correctly annotated.
878  */
879 static __poll_t ep_item_poll(const struct epitem *epi, poll_table *pt,
880 				 int depth)
881 {
882 	struct eventpoll *ep;
883 	bool locked;
884 
885 	pt->_key = epi->event.events;
886 	if (!is_file_epoll(epi->ffd.file))
887 		return vfs_poll(epi->ffd.file, pt) & epi->event.events;
888 
889 	ep = epi->ffd.file->private_data;
890 	poll_wait(epi->ffd.file, &ep->poll_wait, pt);
891 	locked = pt && (pt->_qproc == ep_ptable_queue_proc);
892 
893 	return ep_scan_ready_list(epi->ffd.file->private_data,
894 				  ep_read_events_proc, &depth, depth,
895 				  locked) & epi->event.events;
896 }
897 
898 static __poll_t ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
899 			       void *priv)
900 {
901 	struct epitem *epi, *tmp;
902 	poll_table pt;
903 	int depth = *(int *)priv;
904 
905 	init_poll_funcptr(&pt, NULL);
906 	depth++;
907 
908 	list_for_each_entry_safe(epi, tmp, head, rdllink) {
909 		if (ep_item_poll(epi, &pt, depth)) {
910 			return EPOLLIN | EPOLLRDNORM;
911 		} else {
912 			/*
913 			 * Item has been dropped into the ready list by the poll
914 			 * callback, but it's not actually ready, as far as
915 			 * caller requested events goes. We can remove it here.
916 			 */
917 			__pm_relax(ep_wakeup_source(epi));
918 			list_del_init(&epi->rdllink);
919 		}
920 	}
921 
922 	return 0;
923 }
924 
925 static __poll_t ep_eventpoll_poll(struct file *file, poll_table *wait)
926 {
927 	struct eventpoll *ep = file->private_data;
928 	int depth = 0;
929 
930 	/* Insert inside our poll wait queue */
931 	poll_wait(file, &ep->poll_wait, wait);
932 
933 	/*
934 	 * Proceed to find out if wanted events are really available inside
935 	 * the ready list.
936 	 */
937 	return ep_scan_ready_list(ep, ep_read_events_proc,
938 				  &depth, depth, false);
939 }
940 
941 #ifdef CONFIG_PROC_FS
942 static void ep_show_fdinfo(struct seq_file *m, struct file *f)
943 {
944 	struct eventpoll *ep = f->private_data;
945 	struct rb_node *rbp;
946 
947 	mutex_lock(&ep->mtx);
948 	for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
949 		struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
950 		struct inode *inode = file_inode(epi->ffd.file);
951 
952 		seq_printf(m, "tfd: %8d events: %8x data: %16llx "
953 			   " pos:%lli ino:%lx sdev:%x\n",
954 			   epi->ffd.fd, epi->event.events,
955 			   (long long)epi->event.data,
956 			   (long long)epi->ffd.file->f_pos,
957 			   inode->i_ino, inode->i_sb->s_dev);
958 		if (seq_has_overflowed(m))
959 			break;
960 	}
961 	mutex_unlock(&ep->mtx);
962 }
963 #endif
964 
965 /* File callbacks that implement the eventpoll file behaviour */
966 static const struct file_operations eventpoll_fops = {
967 #ifdef CONFIG_PROC_FS
968 	.show_fdinfo	= ep_show_fdinfo,
969 #endif
970 	.release	= ep_eventpoll_release,
971 	.poll		= ep_eventpoll_poll,
972 	.llseek		= noop_llseek,
973 };
974 
975 /*
976  * This is called from eventpoll_release() to unlink files from the eventpoll
977  * interface. We need to have this facility to cleanup correctly files that are
978  * closed without being removed from the eventpoll interface.
979  */
980 void eventpoll_release_file(struct file *file)
981 {
982 	struct eventpoll *ep;
983 	struct epitem *epi, *next;
984 
985 	/*
986 	 * We don't want to get "file->f_lock" because it is not
987 	 * necessary. It is not necessary because we're in the "struct file"
988 	 * cleanup path, and this means that no one is using this file anymore.
989 	 * So, for example, epoll_ctl() cannot hit here since if we reach this
990 	 * point, the file counter already went to zero and fget() would fail.
991 	 * The only hit might come from ep_free() but by holding the mutex
992 	 * will correctly serialize the operation. We do need to acquire
993 	 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
994 	 * from anywhere but ep_free().
995 	 *
996 	 * Besides, ep_remove() acquires the lock, so we can't hold it here.
997 	 */
998 	mutex_lock(&epmutex);
999 	list_for_each_entry_safe(epi, next, &file->f_ep_links, fllink) {
1000 		ep = epi->ep;
1001 		mutex_lock_nested(&ep->mtx, 0);
1002 		ep_remove(ep, epi);
1003 		mutex_unlock(&ep->mtx);
1004 	}
1005 	mutex_unlock(&epmutex);
1006 }
1007 
1008 static int ep_alloc(struct eventpoll **pep)
1009 {
1010 	int error;
1011 	struct user_struct *user;
1012 	struct eventpoll *ep;
1013 
1014 	user = get_current_user();
1015 	error = -ENOMEM;
1016 	ep = kzalloc(sizeof(*ep), GFP_KERNEL);
1017 	if (unlikely(!ep))
1018 		goto free_uid;
1019 
1020 	mutex_init(&ep->mtx);
1021 	rwlock_init(&ep->lock);
1022 	init_waitqueue_head(&ep->wq);
1023 	init_waitqueue_head(&ep->poll_wait);
1024 	INIT_LIST_HEAD(&ep->rdllist);
1025 	ep->rbr = RB_ROOT_CACHED;
1026 	ep->ovflist = EP_UNACTIVE_PTR;
1027 	ep->user = user;
1028 
1029 	*pep = ep;
1030 
1031 	return 0;
1032 
1033 free_uid:
1034 	free_uid(user);
1035 	return error;
1036 }
1037 
1038 /*
1039  * Search the file inside the eventpoll tree. The RB tree operations
1040  * are protected by the "mtx" mutex, and ep_find() must be called with
1041  * "mtx" held.
1042  */
1043 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
1044 {
1045 	int kcmp;
1046 	struct rb_node *rbp;
1047 	struct epitem *epi, *epir = NULL;
1048 	struct epoll_filefd ffd;
1049 
1050 	ep_set_ffd(&ffd, file, fd);
1051 	for (rbp = ep->rbr.rb_root.rb_node; rbp; ) {
1052 		epi = rb_entry(rbp, struct epitem, rbn);
1053 		kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
1054 		if (kcmp > 0)
1055 			rbp = rbp->rb_right;
1056 		else if (kcmp < 0)
1057 			rbp = rbp->rb_left;
1058 		else {
1059 			epir = epi;
1060 			break;
1061 		}
1062 	}
1063 
1064 	return epir;
1065 }
1066 
1067 #ifdef CONFIG_CHECKPOINT_RESTORE
1068 static struct epitem *ep_find_tfd(struct eventpoll *ep, int tfd, unsigned long toff)
1069 {
1070 	struct rb_node *rbp;
1071 	struct epitem *epi;
1072 
1073 	for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1074 		epi = rb_entry(rbp, struct epitem, rbn);
1075 		if (epi->ffd.fd == tfd) {
1076 			if (toff == 0)
1077 				return epi;
1078 			else
1079 				toff--;
1080 		}
1081 		cond_resched();
1082 	}
1083 
1084 	return NULL;
1085 }
1086 
1087 struct file *get_epoll_tfile_raw_ptr(struct file *file, int tfd,
1088 				     unsigned long toff)
1089 {
1090 	struct file *file_raw;
1091 	struct eventpoll *ep;
1092 	struct epitem *epi;
1093 
1094 	if (!is_file_epoll(file))
1095 		return ERR_PTR(-EINVAL);
1096 
1097 	ep = file->private_data;
1098 
1099 	mutex_lock(&ep->mtx);
1100 	epi = ep_find_tfd(ep, tfd, toff);
1101 	if (epi)
1102 		file_raw = epi->ffd.file;
1103 	else
1104 		file_raw = ERR_PTR(-ENOENT);
1105 	mutex_unlock(&ep->mtx);
1106 
1107 	return file_raw;
1108 }
1109 #endif /* CONFIG_CHECKPOINT_RESTORE */
1110 
1111 /**
1112  * Adds a new entry to the tail of the list in a lockless way, i.e.
1113  * multiple CPUs are allowed to call this function concurrently.
1114  *
1115  * Beware: it is necessary to prevent any other modifications of the
1116  *         existing list until all changes are completed, in other words
1117  *         concurrent list_add_tail_lockless() calls should be protected
1118  *         with a read lock, where write lock acts as a barrier which
1119  *         makes sure all list_add_tail_lockless() calls are fully
1120  *         completed.
1121  *
1122  *        Also an element can be locklessly added to the list only in one
1123  *        direction i.e. either to the tail either to the head, otherwise
1124  *        concurrent access will corrupt the list.
1125  *
1126  * Returns %false if element has been already added to the list, %true
1127  * otherwise.
1128  */
1129 static inline bool list_add_tail_lockless(struct list_head *new,
1130 					  struct list_head *head)
1131 {
1132 	struct list_head *prev;
1133 
1134 	/*
1135 	 * This is simple 'new->next = head' operation, but cmpxchg()
1136 	 * is used in order to detect that same element has been just
1137 	 * added to the list from another CPU: the winner observes
1138 	 * new->next == new.
1139 	 */
1140 	if (cmpxchg(&new->next, new, head) != new)
1141 		return false;
1142 
1143 	/*
1144 	 * Initially ->next of a new element must be updated with the head
1145 	 * (we are inserting to the tail) and only then pointers are atomically
1146 	 * exchanged.  XCHG guarantees memory ordering, thus ->next should be
1147 	 * updated before pointers are actually swapped and pointers are
1148 	 * swapped before prev->next is updated.
1149 	 */
1150 
1151 	prev = xchg(&head->prev, new);
1152 
1153 	/*
1154 	 * It is safe to modify prev->next and new->prev, because a new element
1155 	 * is added only to the tail and new->next is updated before XCHG.
1156 	 */
1157 
1158 	prev->next = new;
1159 	new->prev = prev;
1160 
1161 	return true;
1162 }
1163 
1164 /**
1165  * Chains a new epi entry to the tail of the ep->ovflist in a lockless way,
1166  * i.e. multiple CPUs are allowed to call this function concurrently.
1167  *
1168  * Returns %false if epi element has been already chained, %true otherwise.
1169  */
1170 static inline bool chain_epi_lockless(struct epitem *epi)
1171 {
1172 	struct eventpoll *ep = epi->ep;
1173 
1174 	/* Fast preliminary check */
1175 	if (epi->next != EP_UNACTIVE_PTR)
1176 		return false;
1177 
1178 	/* Check that the same epi has not been just chained from another CPU */
1179 	if (cmpxchg(&epi->next, EP_UNACTIVE_PTR, NULL) != EP_UNACTIVE_PTR)
1180 		return false;
1181 
1182 	/* Atomically exchange tail */
1183 	epi->next = xchg(&ep->ovflist, epi);
1184 
1185 	return true;
1186 }
1187 
1188 /*
1189  * This is the callback that is passed to the wait queue wakeup
1190  * mechanism. It is called by the stored file descriptors when they
1191  * have events to report.
1192  *
1193  * This callback takes a read lock in order not to content with concurrent
1194  * events from another file descriptors, thus all modifications to ->rdllist
1195  * or ->ovflist are lockless.  Read lock is paired with the write lock from
1196  * ep_scan_ready_list(), which stops all list modifications and guarantees
1197  * that lists state is seen correctly.
1198  *
1199  * Another thing worth to mention is that ep_poll_callback() can be called
1200  * concurrently for the same @epi from different CPUs if poll table was inited
1201  * with several wait queues entries.  Plural wakeup from different CPUs of a
1202  * single wait queue is serialized by wq.lock, but the case when multiple wait
1203  * queues are used should be detected accordingly.  This is detected using
1204  * cmpxchg() operation.
1205  */
1206 static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1207 {
1208 	int pwake = 0;
1209 	struct epitem *epi = ep_item_from_wait(wait);
1210 	struct eventpoll *ep = epi->ep;
1211 	__poll_t pollflags = key_to_poll(key);
1212 	unsigned long flags;
1213 	int ewake = 0;
1214 
1215 	read_lock_irqsave(&ep->lock, flags);
1216 
1217 	ep_set_busy_poll_napi_id(epi);
1218 
1219 	/*
1220 	 * If the event mask does not contain any poll(2) event, we consider the
1221 	 * descriptor to be disabled. This condition is likely the effect of the
1222 	 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1223 	 * until the next EPOLL_CTL_MOD will be issued.
1224 	 */
1225 	if (!(epi->event.events & ~EP_PRIVATE_BITS))
1226 		goto out_unlock;
1227 
1228 	/*
1229 	 * Check the events coming with the callback. At this stage, not
1230 	 * every device reports the events in the "key" parameter of the
1231 	 * callback. We need to be able to handle both cases here, hence the
1232 	 * test for "key" != NULL before the event match test.
1233 	 */
1234 	if (pollflags && !(pollflags & epi->event.events))
1235 		goto out_unlock;
1236 
1237 	/*
1238 	 * If we are transferring events to userspace, we can hold no locks
1239 	 * (because we're accessing user memory, and because of linux f_op->poll()
1240 	 * semantics). All the events that happen during that period of time are
1241 	 * chained in ep->ovflist and requeued later on.
1242 	 */
1243 	if (READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR) {
1244 		if (chain_epi_lockless(epi))
1245 			ep_pm_stay_awake_rcu(epi);
1246 	} else if (!ep_is_linked(epi)) {
1247 		/* In the usual case, add event to ready list. */
1248 		if (list_add_tail_lockless(&epi->rdllink, &ep->rdllist))
1249 			ep_pm_stay_awake_rcu(epi);
1250 	}
1251 
1252 	/*
1253 	 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1254 	 * wait list.
1255 	 */
1256 	if (waitqueue_active(&ep->wq)) {
1257 		if ((epi->event.events & EPOLLEXCLUSIVE) &&
1258 					!(pollflags & POLLFREE)) {
1259 			switch (pollflags & EPOLLINOUT_BITS) {
1260 			case EPOLLIN:
1261 				if (epi->event.events & EPOLLIN)
1262 					ewake = 1;
1263 				break;
1264 			case EPOLLOUT:
1265 				if (epi->event.events & EPOLLOUT)
1266 					ewake = 1;
1267 				break;
1268 			case 0:
1269 				ewake = 1;
1270 				break;
1271 			}
1272 		}
1273 		wake_up(&ep->wq);
1274 	}
1275 	if (waitqueue_active(&ep->poll_wait))
1276 		pwake++;
1277 
1278 out_unlock:
1279 	read_unlock_irqrestore(&ep->lock, flags);
1280 
1281 	/* We have to call this outside the lock */
1282 	if (pwake)
1283 		ep_poll_safewake(ep, epi);
1284 
1285 	if (!(epi->event.events & EPOLLEXCLUSIVE))
1286 		ewake = 1;
1287 
1288 	if (pollflags & POLLFREE) {
1289 		/*
1290 		 * If we race with ep_remove_wait_queue() it can miss
1291 		 * ->whead = NULL and do another remove_wait_queue() after
1292 		 * us, so we can't use __remove_wait_queue().
1293 		 */
1294 		list_del_init(&wait->entry);
1295 		/*
1296 		 * ->whead != NULL protects us from the race with ep_free()
1297 		 * or ep_remove(), ep_remove_wait_queue() takes whead->lock
1298 		 * held by the caller. Once we nullify it, nothing protects
1299 		 * ep/epi or even wait.
1300 		 */
1301 		smp_store_release(&ep_pwq_from_wait(wait)->whead, NULL);
1302 	}
1303 
1304 	return ewake;
1305 }
1306 
1307 /*
1308  * This is the callback that is used to add our wait queue to the
1309  * target file wakeup lists.
1310  */
1311 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1312 				 poll_table *pt)
1313 {
1314 	struct epitem *epi = ep_item_from_epqueue(pt);
1315 	struct eppoll_entry *pwq;
1316 
1317 	if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
1318 		init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1319 		pwq->whead = whead;
1320 		pwq->base = epi;
1321 		if (epi->event.events & EPOLLEXCLUSIVE)
1322 			add_wait_queue_exclusive(whead, &pwq->wait);
1323 		else
1324 			add_wait_queue(whead, &pwq->wait);
1325 		list_add_tail(&pwq->llink, &epi->pwqlist);
1326 		epi->nwait++;
1327 	} else {
1328 		/* We have to signal that an error occurred */
1329 		epi->nwait = -1;
1330 	}
1331 }
1332 
1333 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1334 {
1335 	int kcmp;
1336 	struct rb_node **p = &ep->rbr.rb_root.rb_node, *parent = NULL;
1337 	struct epitem *epic;
1338 	bool leftmost = true;
1339 
1340 	while (*p) {
1341 		parent = *p;
1342 		epic = rb_entry(parent, struct epitem, rbn);
1343 		kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1344 		if (kcmp > 0) {
1345 			p = &parent->rb_right;
1346 			leftmost = false;
1347 		} else
1348 			p = &parent->rb_left;
1349 	}
1350 	rb_link_node(&epi->rbn, parent, p);
1351 	rb_insert_color_cached(&epi->rbn, &ep->rbr, leftmost);
1352 }
1353 
1354 
1355 
1356 #define PATH_ARR_SIZE 5
1357 /*
1358  * These are the number paths of length 1 to 5, that we are allowing to emanate
1359  * from a single file of interest. For example, we allow 1000 paths of length
1360  * 1, to emanate from each file of interest. This essentially represents the
1361  * potential wakeup paths, which need to be limited in order to avoid massive
1362  * uncontrolled wakeup storms. The common use case should be a single ep which
1363  * is connected to n file sources. In this case each file source has 1 path
1364  * of length 1. Thus, the numbers below should be more than sufficient. These
1365  * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1366  * and delete can't add additional paths. Protected by the epmutex.
1367  */
1368 static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1369 static int path_count[PATH_ARR_SIZE];
1370 
1371 static int path_count_inc(int nests)
1372 {
1373 	/* Allow an arbitrary number of depth 1 paths */
1374 	if (nests == 0)
1375 		return 0;
1376 
1377 	if (++path_count[nests] > path_limits[nests])
1378 		return -1;
1379 	return 0;
1380 }
1381 
1382 static void path_count_init(void)
1383 {
1384 	int i;
1385 
1386 	for (i = 0; i < PATH_ARR_SIZE; i++)
1387 		path_count[i] = 0;
1388 }
1389 
1390 static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1391 {
1392 	int error = 0;
1393 	struct file *file = priv;
1394 	struct file *child_file;
1395 	struct epitem *epi;
1396 
1397 	/* CTL_DEL can remove links here, but that can't increase our count */
1398 	rcu_read_lock();
1399 	list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) {
1400 		child_file = epi->ep->file;
1401 		if (is_file_epoll(child_file)) {
1402 			if (list_empty(&child_file->f_ep_links)) {
1403 				if (path_count_inc(call_nests)) {
1404 					error = -1;
1405 					break;
1406 				}
1407 			} else {
1408 				error = ep_call_nested(&poll_loop_ncalls,
1409 							reverse_path_check_proc,
1410 							child_file, child_file,
1411 							current);
1412 			}
1413 			if (error != 0)
1414 				break;
1415 		} else {
1416 			printk(KERN_ERR "reverse_path_check_proc: "
1417 				"file is not an ep!\n");
1418 		}
1419 	}
1420 	rcu_read_unlock();
1421 	return error;
1422 }
1423 
1424 /**
1425  * reverse_path_check - The tfile_check_list is list of file *, which have
1426  *                      links that are proposed to be newly added. We need to
1427  *                      make sure that those added links don't add too many
1428  *                      paths such that we will spend all our time waking up
1429  *                      eventpoll objects.
1430  *
1431  * Returns: Returns zero if the proposed links don't create too many paths,
1432  *	    -1 otherwise.
1433  */
1434 static int reverse_path_check(void)
1435 {
1436 	int error = 0;
1437 	struct file *current_file;
1438 
1439 	/* let's call this for all tfiles */
1440 	list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
1441 		path_count_init();
1442 		error = ep_call_nested(&poll_loop_ncalls,
1443 					reverse_path_check_proc, current_file,
1444 					current_file, current);
1445 		if (error)
1446 			break;
1447 	}
1448 	return error;
1449 }
1450 
1451 static int ep_create_wakeup_source(struct epitem *epi)
1452 {
1453 	const char *name;
1454 	struct wakeup_source *ws;
1455 
1456 	if (!epi->ep->ws) {
1457 		epi->ep->ws = wakeup_source_register(NULL, "eventpoll");
1458 		if (!epi->ep->ws)
1459 			return -ENOMEM;
1460 	}
1461 
1462 	name = epi->ffd.file->f_path.dentry->d_name.name;
1463 	ws = wakeup_source_register(NULL, name);
1464 
1465 	if (!ws)
1466 		return -ENOMEM;
1467 	rcu_assign_pointer(epi->ws, ws);
1468 
1469 	return 0;
1470 }
1471 
1472 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1473 static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1474 {
1475 	struct wakeup_source *ws = ep_wakeup_source(epi);
1476 
1477 	RCU_INIT_POINTER(epi->ws, NULL);
1478 
1479 	/*
1480 	 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1481 	 * used internally by wakeup_source_remove, too (called by
1482 	 * wakeup_source_unregister), so we cannot use call_rcu
1483 	 */
1484 	synchronize_rcu();
1485 	wakeup_source_unregister(ws);
1486 }
1487 
1488 /*
1489  * Must be called with "mtx" held.
1490  */
1491 static int ep_insert(struct eventpoll *ep, const struct epoll_event *event,
1492 		     struct file *tfile, int fd, int full_check)
1493 {
1494 	int error, pwake = 0;
1495 	__poll_t revents;
1496 	long user_watches;
1497 	struct epitem *epi;
1498 	struct ep_pqueue epq;
1499 
1500 	lockdep_assert_irqs_enabled();
1501 
1502 	user_watches = atomic_long_read(&ep->user->epoll_watches);
1503 	if (unlikely(user_watches >= max_user_watches))
1504 		return -ENOSPC;
1505 	if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1506 		return -ENOMEM;
1507 
1508 	/* Item initialization follow here ... */
1509 	INIT_LIST_HEAD(&epi->rdllink);
1510 	INIT_LIST_HEAD(&epi->fllink);
1511 	INIT_LIST_HEAD(&epi->pwqlist);
1512 	epi->ep = ep;
1513 	ep_set_ffd(&epi->ffd, tfile, fd);
1514 	epi->event = *event;
1515 	epi->nwait = 0;
1516 	epi->next = EP_UNACTIVE_PTR;
1517 	if (epi->event.events & EPOLLWAKEUP) {
1518 		error = ep_create_wakeup_source(epi);
1519 		if (error)
1520 			goto error_create_wakeup_source;
1521 	} else {
1522 		RCU_INIT_POINTER(epi->ws, NULL);
1523 	}
1524 
1525 	/* Initialize the poll table using the queue callback */
1526 	epq.epi = epi;
1527 	init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1528 
1529 	/*
1530 	 * Attach the item to the poll hooks and get current event bits.
1531 	 * We can safely use the file* here because its usage count has
1532 	 * been increased by the caller of this function. Note that after
1533 	 * this operation completes, the poll callback can start hitting
1534 	 * the new item.
1535 	 */
1536 	revents = ep_item_poll(epi, &epq.pt, 1);
1537 
1538 	/*
1539 	 * We have to check if something went wrong during the poll wait queue
1540 	 * install process. Namely an allocation for a wait queue failed due
1541 	 * high memory pressure.
1542 	 */
1543 	error = -ENOMEM;
1544 	if (epi->nwait < 0)
1545 		goto error_unregister;
1546 
1547 	/* Add the current item to the list of active epoll hook for this file */
1548 	spin_lock(&tfile->f_lock);
1549 	list_add_tail_rcu(&epi->fllink, &tfile->f_ep_links);
1550 	spin_unlock(&tfile->f_lock);
1551 
1552 	/*
1553 	 * Add the current item to the RB tree. All RB tree operations are
1554 	 * protected by "mtx", and ep_insert() is called with "mtx" held.
1555 	 */
1556 	ep_rbtree_insert(ep, epi);
1557 
1558 	/* now check if we've created too many backpaths */
1559 	error = -EINVAL;
1560 	if (full_check && reverse_path_check())
1561 		goto error_remove_epi;
1562 
1563 	/* We have to drop the new item inside our item list to keep track of it */
1564 	write_lock_irq(&ep->lock);
1565 
1566 	/* record NAPI ID of new item if present */
1567 	ep_set_busy_poll_napi_id(epi);
1568 
1569 	/* If the file is already "ready" we drop it inside the ready list */
1570 	if (revents && !ep_is_linked(epi)) {
1571 		list_add_tail(&epi->rdllink, &ep->rdllist);
1572 		ep_pm_stay_awake(epi);
1573 
1574 		/* Notify waiting tasks that events are available */
1575 		if (waitqueue_active(&ep->wq))
1576 			wake_up(&ep->wq);
1577 		if (waitqueue_active(&ep->poll_wait))
1578 			pwake++;
1579 	}
1580 
1581 	write_unlock_irq(&ep->lock);
1582 
1583 	atomic_long_inc(&ep->user->epoll_watches);
1584 
1585 	/* We have to call this outside the lock */
1586 	if (pwake)
1587 		ep_poll_safewake(ep, NULL);
1588 
1589 	return 0;
1590 
1591 error_remove_epi:
1592 	spin_lock(&tfile->f_lock);
1593 	list_del_rcu(&epi->fllink);
1594 	spin_unlock(&tfile->f_lock);
1595 
1596 	rb_erase_cached(&epi->rbn, &ep->rbr);
1597 
1598 error_unregister:
1599 	ep_unregister_pollwait(ep, epi);
1600 
1601 	/*
1602 	 * We need to do this because an event could have been arrived on some
1603 	 * allocated wait queue. Note that we don't care about the ep->ovflist
1604 	 * list, since that is used/cleaned only inside a section bound by "mtx".
1605 	 * And ep_insert() is called with "mtx" held.
1606 	 */
1607 	write_lock_irq(&ep->lock);
1608 	if (ep_is_linked(epi))
1609 		list_del_init(&epi->rdllink);
1610 	write_unlock_irq(&ep->lock);
1611 
1612 	wakeup_source_unregister(ep_wakeup_source(epi));
1613 
1614 error_create_wakeup_source:
1615 	kmem_cache_free(epi_cache, epi);
1616 
1617 	return error;
1618 }
1619 
1620 /*
1621  * Modify the interest event mask by dropping an event if the new mask
1622  * has a match in the current file status. Must be called with "mtx" held.
1623  */
1624 static int ep_modify(struct eventpoll *ep, struct epitem *epi,
1625 		     const struct epoll_event *event)
1626 {
1627 	int pwake = 0;
1628 	poll_table pt;
1629 
1630 	lockdep_assert_irqs_enabled();
1631 
1632 	init_poll_funcptr(&pt, NULL);
1633 
1634 	/*
1635 	 * Set the new event interest mask before calling f_op->poll();
1636 	 * otherwise we might miss an event that happens between the
1637 	 * f_op->poll() call and the new event set registering.
1638 	 */
1639 	epi->event.events = event->events; /* need barrier below */
1640 	epi->event.data = event->data; /* protected by mtx */
1641 	if (epi->event.events & EPOLLWAKEUP) {
1642 		if (!ep_has_wakeup_source(epi))
1643 			ep_create_wakeup_source(epi);
1644 	} else if (ep_has_wakeup_source(epi)) {
1645 		ep_destroy_wakeup_source(epi);
1646 	}
1647 
1648 	/*
1649 	 * The following barrier has two effects:
1650 	 *
1651 	 * 1) Flush epi changes above to other CPUs.  This ensures
1652 	 *    we do not miss events from ep_poll_callback if an
1653 	 *    event occurs immediately after we call f_op->poll().
1654 	 *    We need this because we did not take ep->lock while
1655 	 *    changing epi above (but ep_poll_callback does take
1656 	 *    ep->lock).
1657 	 *
1658 	 * 2) We also need to ensure we do not miss _past_ events
1659 	 *    when calling f_op->poll().  This barrier also
1660 	 *    pairs with the barrier in wq_has_sleeper (see
1661 	 *    comments for wq_has_sleeper).
1662 	 *
1663 	 * This barrier will now guarantee ep_poll_callback or f_op->poll
1664 	 * (or both) will notice the readiness of an item.
1665 	 */
1666 	smp_mb();
1667 
1668 	/*
1669 	 * Get current event bits. We can safely use the file* here because
1670 	 * its usage count has been increased by the caller of this function.
1671 	 * If the item is "hot" and it is not registered inside the ready
1672 	 * list, push it inside.
1673 	 */
1674 	if (ep_item_poll(epi, &pt, 1)) {
1675 		write_lock_irq(&ep->lock);
1676 		if (!ep_is_linked(epi)) {
1677 			list_add_tail(&epi->rdllink, &ep->rdllist);
1678 			ep_pm_stay_awake(epi);
1679 
1680 			/* Notify waiting tasks that events are available */
1681 			if (waitqueue_active(&ep->wq))
1682 				wake_up(&ep->wq);
1683 			if (waitqueue_active(&ep->poll_wait))
1684 				pwake++;
1685 		}
1686 		write_unlock_irq(&ep->lock);
1687 	}
1688 
1689 	/* We have to call this outside the lock */
1690 	if (pwake)
1691 		ep_poll_safewake(ep, NULL);
1692 
1693 	return 0;
1694 }
1695 
1696 static __poll_t ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1697 			       void *priv)
1698 {
1699 	struct ep_send_events_data *esed = priv;
1700 	__poll_t revents;
1701 	struct epitem *epi, *tmp;
1702 	struct epoll_event __user *uevent = esed->events;
1703 	struct wakeup_source *ws;
1704 	poll_table pt;
1705 
1706 	init_poll_funcptr(&pt, NULL);
1707 	esed->res = 0;
1708 
1709 	/*
1710 	 * We can loop without lock because we are passed a task private list.
1711 	 * Items cannot vanish during the loop because ep_scan_ready_list() is
1712 	 * holding "mtx" during this call.
1713 	 */
1714 	lockdep_assert_held(&ep->mtx);
1715 
1716 	list_for_each_entry_safe(epi, tmp, head, rdllink) {
1717 		if (esed->res >= esed->maxevents)
1718 			break;
1719 
1720 		/*
1721 		 * Activate ep->ws before deactivating epi->ws to prevent
1722 		 * triggering auto-suspend here (in case we reactive epi->ws
1723 		 * below).
1724 		 *
1725 		 * This could be rearranged to delay the deactivation of epi->ws
1726 		 * instead, but then epi->ws would temporarily be out of sync
1727 		 * with ep_is_linked().
1728 		 */
1729 		ws = ep_wakeup_source(epi);
1730 		if (ws) {
1731 			if (ws->active)
1732 				__pm_stay_awake(ep->ws);
1733 			__pm_relax(ws);
1734 		}
1735 
1736 		list_del_init(&epi->rdllink);
1737 
1738 		/*
1739 		 * If the event mask intersect the caller-requested one,
1740 		 * deliver the event to userspace. Again, ep_scan_ready_list()
1741 		 * is holding ep->mtx, so no operations coming from userspace
1742 		 * can change the item.
1743 		 */
1744 		revents = ep_item_poll(epi, &pt, 1);
1745 		if (!revents)
1746 			continue;
1747 
1748 		if (__put_user(revents, &uevent->events) ||
1749 		    __put_user(epi->event.data, &uevent->data)) {
1750 			list_add(&epi->rdllink, head);
1751 			ep_pm_stay_awake(epi);
1752 			if (!esed->res)
1753 				esed->res = -EFAULT;
1754 			return 0;
1755 		}
1756 		esed->res++;
1757 		uevent++;
1758 		if (epi->event.events & EPOLLONESHOT)
1759 			epi->event.events &= EP_PRIVATE_BITS;
1760 		else if (!(epi->event.events & EPOLLET)) {
1761 			/*
1762 			 * If this file has been added with Level
1763 			 * Trigger mode, we need to insert back inside
1764 			 * the ready list, so that the next call to
1765 			 * epoll_wait() will check again the events
1766 			 * availability. At this point, no one can insert
1767 			 * into ep->rdllist besides us. The epoll_ctl()
1768 			 * callers are locked out by
1769 			 * ep_scan_ready_list() holding "mtx" and the
1770 			 * poll callback will queue them in ep->ovflist.
1771 			 */
1772 			list_add_tail(&epi->rdllink, &ep->rdllist);
1773 			ep_pm_stay_awake(epi);
1774 		}
1775 	}
1776 
1777 	return 0;
1778 }
1779 
1780 static int ep_send_events(struct eventpoll *ep,
1781 			  struct epoll_event __user *events, int maxevents)
1782 {
1783 	struct ep_send_events_data esed;
1784 
1785 	esed.maxevents = maxevents;
1786 	esed.events = events;
1787 
1788 	ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false);
1789 	return esed.res;
1790 }
1791 
1792 static inline struct timespec64 ep_set_mstimeout(long ms)
1793 {
1794 	struct timespec64 now, ts = {
1795 		.tv_sec = ms / MSEC_PER_SEC,
1796 		.tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1797 	};
1798 
1799 	ktime_get_ts64(&now);
1800 	return timespec64_add_safe(now, ts);
1801 }
1802 
1803 /**
1804  * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1805  *           event buffer.
1806  *
1807  * @ep: Pointer to the eventpoll context.
1808  * @events: Pointer to the userspace buffer where the ready events should be
1809  *          stored.
1810  * @maxevents: Size (in terms of number of events) of the caller event buffer.
1811  * @timeout: Maximum timeout for the ready events fetch operation, in
1812  *           milliseconds. If the @timeout is zero, the function will not block,
1813  *           while if the @timeout is less than zero, the function will block
1814  *           until at least one event has been retrieved (or an error
1815  *           occurred).
1816  *
1817  * Returns: Returns the number of ready events which have been fetched, or an
1818  *          error code, in case of error.
1819  */
1820 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1821 		   int maxevents, long timeout)
1822 {
1823 	int res = 0, eavail, timed_out = 0;
1824 	u64 slack = 0;
1825 	wait_queue_entry_t wait;
1826 	ktime_t expires, *to = NULL;
1827 
1828 	lockdep_assert_irqs_enabled();
1829 
1830 	if (timeout > 0) {
1831 		struct timespec64 end_time = ep_set_mstimeout(timeout);
1832 
1833 		slack = select_estimate_accuracy(&end_time);
1834 		to = &expires;
1835 		*to = timespec64_to_ktime(end_time);
1836 	} else if (timeout == 0) {
1837 		/*
1838 		 * Avoid the unnecessary trip to the wait queue loop, if the
1839 		 * caller specified a non blocking operation. We still need
1840 		 * lock because we could race and not see an epi being added
1841 		 * to the ready list while in irq callback. Thus incorrectly
1842 		 * returning 0 back to userspace.
1843 		 */
1844 		timed_out = 1;
1845 
1846 		write_lock_irq(&ep->lock);
1847 		eavail = ep_events_available(ep);
1848 		write_unlock_irq(&ep->lock);
1849 
1850 		goto send_events;
1851 	}
1852 
1853 fetch_events:
1854 
1855 	if (!ep_events_available(ep))
1856 		ep_busy_loop(ep, timed_out);
1857 
1858 	eavail = ep_events_available(ep);
1859 	if (eavail)
1860 		goto send_events;
1861 
1862 	/*
1863 	 * Busy poll timed out.  Drop NAPI ID for now, we can add
1864 	 * it back in when we have moved a socket with a valid NAPI
1865 	 * ID onto the ready list.
1866 	 */
1867 	ep_reset_busy_poll_napi_id(ep);
1868 
1869 	do {
1870 		/*
1871 		 * Internally init_wait() uses autoremove_wake_function(),
1872 		 * thus wait entry is removed from the wait queue on each
1873 		 * wakeup. Why it is important? In case of several waiters
1874 		 * each new wakeup will hit the next waiter, giving it the
1875 		 * chance to harvest new event. Otherwise wakeup can be
1876 		 * lost. This is also good performance-wise, because on
1877 		 * normal wakeup path no need to call __remove_wait_queue()
1878 		 * explicitly, thus ep->lock is not taken, which halts the
1879 		 * event delivery.
1880 		 */
1881 		init_wait(&wait);
1882 
1883 		write_lock_irq(&ep->lock);
1884 		/*
1885 		 * Barrierless variant, waitqueue_active() is called under
1886 		 * the same lock on wakeup ep_poll_callback() side, so it
1887 		 * is safe to avoid an explicit barrier.
1888 		 */
1889 		__set_current_state(TASK_INTERRUPTIBLE);
1890 
1891 		/*
1892 		 * Do the final check under the lock. ep_scan_ready_list()
1893 		 * plays with two lists (->rdllist and ->ovflist) and there
1894 		 * is always a race when both lists are empty for short
1895 		 * period of time although events are pending, so lock is
1896 		 * important.
1897 		 */
1898 		eavail = ep_events_available(ep);
1899 		if (!eavail) {
1900 			if (signal_pending(current))
1901 				res = -EINTR;
1902 			else
1903 				__add_wait_queue_exclusive(&ep->wq, &wait);
1904 		}
1905 		write_unlock_irq(&ep->lock);
1906 
1907 		if (eavail || res)
1908 			break;
1909 
1910 		if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS)) {
1911 			timed_out = 1;
1912 			break;
1913 		}
1914 
1915 		/* We were woken up, thus go and try to harvest some events */
1916 		eavail = 1;
1917 
1918 	} while (0);
1919 
1920 	__set_current_state(TASK_RUNNING);
1921 
1922 	if (!list_empty_careful(&wait.entry)) {
1923 		write_lock_irq(&ep->lock);
1924 		__remove_wait_queue(&ep->wq, &wait);
1925 		write_unlock_irq(&ep->lock);
1926 	}
1927 
1928 send_events:
1929 	if (fatal_signal_pending(current)) {
1930 		/*
1931 		 * Always short-circuit for fatal signals to allow
1932 		 * threads to make a timely exit without the chance of
1933 		 * finding more events available and fetching
1934 		 * repeatedly.
1935 		 */
1936 		res = -EINTR;
1937 	}
1938 	/*
1939 	 * Try to transfer events to user space. In case we get 0 events and
1940 	 * there's still timeout left over, we go trying again in search of
1941 	 * more luck.
1942 	 */
1943 	if (!res && eavail &&
1944 	    !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1945 		goto fetch_events;
1946 
1947 	return res;
1948 }
1949 
1950 /**
1951  * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1952  *                      API, to verify that adding an epoll file inside another
1953  *                      epoll structure, does not violate the constraints, in
1954  *                      terms of closed loops, or too deep chains (which can
1955  *                      result in excessive stack usage).
1956  *
1957  * @priv: Pointer to the epoll file to be currently checked.
1958  * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1959  *          data structure pointer.
1960  * @call_nests: Current dept of the @ep_call_nested() call stack.
1961  *
1962  * Returns: Returns zero if adding the epoll @file inside current epoll
1963  *          structure @ep does not violate the constraints, or -1 otherwise.
1964  */
1965 static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1966 {
1967 	int error = 0;
1968 	struct file *file = priv;
1969 	struct eventpoll *ep = file->private_data;
1970 	struct eventpoll *ep_tovisit;
1971 	struct rb_node *rbp;
1972 	struct epitem *epi;
1973 
1974 	mutex_lock_nested(&ep->mtx, call_nests + 1);
1975 	ep->visited = 1;
1976 	list_add(&ep->visited_list_link, &visited_list);
1977 	for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1978 		epi = rb_entry(rbp, struct epitem, rbn);
1979 		if (unlikely(is_file_epoll(epi->ffd.file))) {
1980 			ep_tovisit = epi->ffd.file->private_data;
1981 			if (ep_tovisit->visited)
1982 				continue;
1983 			error = ep_call_nested(&poll_loop_ncalls,
1984 					ep_loop_check_proc, epi->ffd.file,
1985 					ep_tovisit, current);
1986 			if (error != 0)
1987 				break;
1988 		} else {
1989 			/*
1990 			 * If we've reached a file that is not associated with
1991 			 * an ep, then we need to check if the newly added
1992 			 * links are going to add too many wakeup paths. We do
1993 			 * this by adding it to the tfile_check_list, if it's
1994 			 * not already there, and calling reverse_path_check()
1995 			 * during ep_insert().
1996 			 */
1997 			if (list_empty(&epi->ffd.file->f_tfile_llink))
1998 				list_add(&epi->ffd.file->f_tfile_llink,
1999 					 &tfile_check_list);
2000 		}
2001 	}
2002 	mutex_unlock(&ep->mtx);
2003 
2004 	return error;
2005 }
2006 
2007 /**
2008  * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
2009  *                 another epoll file (represented by @ep) does not create
2010  *                 closed loops or too deep chains.
2011  *
2012  * @ep: Pointer to the epoll private data structure.
2013  * @file: Pointer to the epoll file to be checked.
2014  *
2015  * Returns: Returns zero if adding the epoll @file inside current epoll
2016  *          structure @ep does not violate the constraints, or -1 otherwise.
2017  */
2018 static int ep_loop_check(struct eventpoll *ep, struct file *file)
2019 {
2020 	int ret;
2021 	struct eventpoll *ep_cur, *ep_next;
2022 
2023 	ret = ep_call_nested(&poll_loop_ncalls,
2024 			      ep_loop_check_proc, file, ep, current);
2025 	/* clear visited list */
2026 	list_for_each_entry_safe(ep_cur, ep_next, &visited_list,
2027 							visited_list_link) {
2028 		ep_cur->visited = 0;
2029 		list_del(&ep_cur->visited_list_link);
2030 	}
2031 	return ret;
2032 }
2033 
2034 static void clear_tfile_check_list(void)
2035 {
2036 	struct file *file;
2037 
2038 	/* first clear the tfile_check_list */
2039 	while (!list_empty(&tfile_check_list)) {
2040 		file = list_first_entry(&tfile_check_list, struct file,
2041 					f_tfile_llink);
2042 		list_del_init(&file->f_tfile_llink);
2043 	}
2044 	INIT_LIST_HEAD(&tfile_check_list);
2045 }
2046 
2047 /*
2048  * Open an eventpoll file descriptor.
2049  */
2050 static int do_epoll_create(int flags)
2051 {
2052 	int error, fd;
2053 	struct eventpoll *ep = NULL;
2054 	struct file *file;
2055 
2056 	/* Check the EPOLL_* constant for consistency.  */
2057 	BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
2058 
2059 	if (flags & ~EPOLL_CLOEXEC)
2060 		return -EINVAL;
2061 	/*
2062 	 * Create the internal data structure ("struct eventpoll").
2063 	 */
2064 	error = ep_alloc(&ep);
2065 	if (error < 0)
2066 		return error;
2067 	/*
2068 	 * Creates all the items needed to setup an eventpoll file. That is,
2069 	 * a file structure and a free file descriptor.
2070 	 */
2071 	fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
2072 	if (fd < 0) {
2073 		error = fd;
2074 		goto out_free_ep;
2075 	}
2076 	file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
2077 				 O_RDWR | (flags & O_CLOEXEC));
2078 	if (IS_ERR(file)) {
2079 		error = PTR_ERR(file);
2080 		goto out_free_fd;
2081 	}
2082 	ep->file = file;
2083 	fd_install(fd, file);
2084 	return fd;
2085 
2086 out_free_fd:
2087 	put_unused_fd(fd);
2088 out_free_ep:
2089 	ep_free(ep);
2090 	return error;
2091 }
2092 
2093 SYSCALL_DEFINE1(epoll_create1, int, flags)
2094 {
2095 	return do_epoll_create(flags);
2096 }
2097 
2098 SYSCALL_DEFINE1(epoll_create, int, size)
2099 {
2100 	if (size <= 0)
2101 		return -EINVAL;
2102 
2103 	return do_epoll_create(0);
2104 }
2105 
2106 static inline int epoll_mutex_lock(struct mutex *mutex, int depth,
2107 				   bool nonblock)
2108 {
2109 	if (!nonblock) {
2110 		mutex_lock_nested(mutex, depth);
2111 		return 0;
2112 	}
2113 	if (mutex_trylock(mutex))
2114 		return 0;
2115 	return -EAGAIN;
2116 }
2117 
2118 int do_epoll_ctl(int epfd, int op, int fd, struct epoll_event *epds,
2119 		 bool nonblock)
2120 {
2121 	int error;
2122 	int full_check = 0;
2123 	struct fd f, tf;
2124 	struct eventpoll *ep;
2125 	struct epitem *epi;
2126 	struct eventpoll *tep = NULL;
2127 
2128 	error = -EBADF;
2129 	f = fdget(epfd);
2130 	if (!f.file)
2131 		goto error_return;
2132 
2133 	/* Get the "struct file *" for the target file */
2134 	tf = fdget(fd);
2135 	if (!tf.file)
2136 		goto error_fput;
2137 
2138 	/* The target file descriptor must support poll */
2139 	error = -EPERM;
2140 	if (!file_can_poll(tf.file))
2141 		goto error_tgt_fput;
2142 
2143 	/* Check if EPOLLWAKEUP is allowed */
2144 	if (ep_op_has_event(op))
2145 		ep_take_care_of_epollwakeup(epds);
2146 
2147 	/*
2148 	 * We have to check that the file structure underneath the file descriptor
2149 	 * the user passed to us _is_ an eventpoll file. And also we do not permit
2150 	 * adding an epoll file descriptor inside itself.
2151 	 */
2152 	error = -EINVAL;
2153 	if (f.file == tf.file || !is_file_epoll(f.file))
2154 		goto error_tgt_fput;
2155 
2156 	/*
2157 	 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
2158 	 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
2159 	 * Also, we do not currently supported nested exclusive wakeups.
2160 	 */
2161 	if (ep_op_has_event(op) && (epds->events & EPOLLEXCLUSIVE)) {
2162 		if (op == EPOLL_CTL_MOD)
2163 			goto error_tgt_fput;
2164 		if (op == EPOLL_CTL_ADD && (is_file_epoll(tf.file) ||
2165 				(epds->events & ~EPOLLEXCLUSIVE_OK_BITS)))
2166 			goto error_tgt_fput;
2167 	}
2168 
2169 	/*
2170 	 * At this point it is safe to assume that the "private_data" contains
2171 	 * our own data structure.
2172 	 */
2173 	ep = f.file->private_data;
2174 
2175 	/*
2176 	 * When we insert an epoll file descriptor, inside another epoll file
2177 	 * descriptor, there is the change of creating closed loops, which are
2178 	 * better be handled here, than in more critical paths. While we are
2179 	 * checking for loops we also determine the list of files reachable
2180 	 * and hang them on the tfile_check_list, so we can check that we
2181 	 * haven't created too many possible wakeup paths.
2182 	 *
2183 	 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
2184 	 * the epoll file descriptor is attaching directly to a wakeup source,
2185 	 * unless the epoll file descriptor is nested. The purpose of taking the
2186 	 * 'epmutex' on add is to prevent complex toplogies such as loops and
2187 	 * deep wakeup paths from forming in parallel through multiple
2188 	 * EPOLL_CTL_ADD operations.
2189 	 */
2190 	error = epoll_mutex_lock(&ep->mtx, 0, nonblock);
2191 	if (error)
2192 		goto error_tgt_fput;
2193 	if (op == EPOLL_CTL_ADD) {
2194 		if (!list_empty(&f.file->f_ep_links) ||
2195 						is_file_epoll(tf.file)) {
2196 			mutex_unlock(&ep->mtx);
2197 			error = epoll_mutex_lock(&epmutex, 0, nonblock);
2198 			if (error)
2199 				goto error_tgt_fput;
2200 			full_check = 1;
2201 			if (is_file_epoll(tf.file)) {
2202 				error = -ELOOP;
2203 				if (ep_loop_check(ep, tf.file) != 0) {
2204 					clear_tfile_check_list();
2205 					goto error_tgt_fput;
2206 				}
2207 			} else
2208 				list_add(&tf.file->f_tfile_llink,
2209 							&tfile_check_list);
2210 			error = epoll_mutex_lock(&ep->mtx, 0, nonblock);
2211 			if (error) {
2212 out_del:
2213 				list_del(&tf.file->f_tfile_llink);
2214 				goto error_tgt_fput;
2215 			}
2216 			if (is_file_epoll(tf.file)) {
2217 				tep = tf.file->private_data;
2218 				error = epoll_mutex_lock(&tep->mtx, 1, nonblock);
2219 				if (error) {
2220 					mutex_unlock(&ep->mtx);
2221 					goto out_del;
2222 				}
2223 			}
2224 		}
2225 	}
2226 
2227 	/*
2228 	 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
2229 	 * above, we can be sure to be able to use the item looked up by
2230 	 * ep_find() till we release the mutex.
2231 	 */
2232 	epi = ep_find(ep, tf.file, fd);
2233 
2234 	error = -EINVAL;
2235 	switch (op) {
2236 	case EPOLL_CTL_ADD:
2237 		if (!epi) {
2238 			epds->events |= EPOLLERR | EPOLLHUP;
2239 			error = ep_insert(ep, epds, tf.file, fd, full_check);
2240 		} else
2241 			error = -EEXIST;
2242 		if (full_check)
2243 			clear_tfile_check_list();
2244 		break;
2245 	case EPOLL_CTL_DEL:
2246 		if (epi)
2247 			error = ep_remove(ep, epi);
2248 		else
2249 			error = -ENOENT;
2250 		break;
2251 	case EPOLL_CTL_MOD:
2252 		if (epi) {
2253 			if (!(epi->event.events & EPOLLEXCLUSIVE)) {
2254 				epds->events |= EPOLLERR | EPOLLHUP;
2255 				error = ep_modify(ep, epi, epds);
2256 			}
2257 		} else
2258 			error = -ENOENT;
2259 		break;
2260 	}
2261 	if (tep != NULL)
2262 		mutex_unlock(&tep->mtx);
2263 	mutex_unlock(&ep->mtx);
2264 
2265 error_tgt_fput:
2266 	if (full_check)
2267 		mutex_unlock(&epmutex);
2268 
2269 	fdput(tf);
2270 error_fput:
2271 	fdput(f);
2272 error_return:
2273 
2274 	return error;
2275 }
2276 
2277 /*
2278  * The following function implements the controller interface for
2279  * the eventpoll file that enables the insertion/removal/change of
2280  * file descriptors inside the interest set.
2281  */
2282 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
2283 		struct epoll_event __user *, event)
2284 {
2285 	struct epoll_event epds;
2286 
2287 	if (ep_op_has_event(op) &&
2288 	    copy_from_user(&epds, event, sizeof(struct epoll_event)))
2289 		return -EFAULT;
2290 
2291 	return do_epoll_ctl(epfd, op, fd, &epds, false);
2292 }
2293 
2294 /*
2295  * Implement the event wait interface for the eventpoll file. It is the kernel
2296  * part of the user space epoll_wait(2).
2297  */
2298 static int do_epoll_wait(int epfd, struct epoll_event __user *events,
2299 			 int maxevents, int timeout)
2300 {
2301 	int error;
2302 	struct fd f;
2303 	struct eventpoll *ep;
2304 
2305 	/* The maximum number of event must be greater than zero */
2306 	if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
2307 		return -EINVAL;
2308 
2309 	/* Verify that the area passed by the user is writeable */
2310 	if (!access_ok(events, maxevents * sizeof(struct epoll_event)))
2311 		return -EFAULT;
2312 
2313 	/* Get the "struct file *" for the eventpoll file */
2314 	f = fdget(epfd);
2315 	if (!f.file)
2316 		return -EBADF;
2317 
2318 	/*
2319 	 * We have to check that the file structure underneath the fd
2320 	 * the user passed to us _is_ an eventpoll file.
2321 	 */
2322 	error = -EINVAL;
2323 	if (!is_file_epoll(f.file))
2324 		goto error_fput;
2325 
2326 	/*
2327 	 * At this point it is safe to assume that the "private_data" contains
2328 	 * our own data structure.
2329 	 */
2330 	ep = f.file->private_data;
2331 
2332 	/* Time to fish for events ... */
2333 	error = ep_poll(ep, events, maxevents, timeout);
2334 
2335 error_fput:
2336 	fdput(f);
2337 	return error;
2338 }
2339 
2340 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
2341 		int, maxevents, int, timeout)
2342 {
2343 	return do_epoll_wait(epfd, events, maxevents, timeout);
2344 }
2345 
2346 /*
2347  * Implement the event wait interface for the eventpoll file. It is the kernel
2348  * part of the user space epoll_pwait(2).
2349  */
2350 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2351 		int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2352 		size_t, sigsetsize)
2353 {
2354 	int error;
2355 
2356 	/*
2357 	 * If the caller wants a certain signal mask to be set during the wait,
2358 	 * we apply it here.
2359 	 */
2360 	error = set_user_sigmask(sigmask, sigsetsize);
2361 	if (error)
2362 		return error;
2363 
2364 	error = do_epoll_wait(epfd, events, maxevents, timeout);
2365 	restore_saved_sigmask_unless(error == -EINTR);
2366 
2367 	return error;
2368 }
2369 
2370 #ifdef CONFIG_COMPAT
2371 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2372 			struct epoll_event __user *, events,
2373 			int, maxevents, int, timeout,
2374 			const compat_sigset_t __user *, sigmask,
2375 			compat_size_t, sigsetsize)
2376 {
2377 	long err;
2378 
2379 	/*
2380 	 * If the caller wants a certain signal mask to be set during the wait,
2381 	 * we apply it here.
2382 	 */
2383 	err = set_compat_user_sigmask(sigmask, sigsetsize);
2384 	if (err)
2385 		return err;
2386 
2387 	err = do_epoll_wait(epfd, events, maxevents, timeout);
2388 	restore_saved_sigmask_unless(err == -EINTR);
2389 
2390 	return err;
2391 }
2392 #endif
2393 
2394 static int __init eventpoll_init(void)
2395 {
2396 	struct sysinfo si;
2397 
2398 	si_meminfo(&si);
2399 	/*
2400 	 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2401 	 */
2402 	max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2403 		EP_ITEM_COST;
2404 	BUG_ON(max_user_watches < 0);
2405 
2406 	/*
2407 	 * Initialize the structure used to perform epoll file descriptor
2408 	 * inclusion loops checks.
2409 	 */
2410 	ep_nested_calls_init(&poll_loop_ncalls);
2411 
2412 	/*
2413 	 * We can have many thousands of epitems, so prevent this from
2414 	 * using an extra cache line on 64-bit (and smaller) CPUs
2415 	 */
2416 	BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2417 
2418 	/* Allocates slab cache used to allocate "struct epitem" items */
2419 	epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2420 			0, SLAB_HWCACHE_ALIGN|SLAB_PANIC|SLAB_ACCOUNT, NULL);
2421 
2422 	/* Allocates slab cache used to allocate "struct eppoll_entry" */
2423 	pwq_cache = kmem_cache_create("eventpoll_pwq",
2424 		sizeof(struct eppoll_entry), 0, SLAB_PANIC|SLAB_ACCOUNT, NULL);
2425 
2426 	return 0;
2427 }
2428 fs_initcall(eventpoll_init);
2429