xref: /openbmc/linux/fs/eventpoll.c (revision 981ab3f1)
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/signal.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/uaccess.h>
38 #include <asm/io.h>
39 #include <asm/mman.h>
40 #include <linux/atomic.h>
41 #include <linux/proc_fs.h>
42 #include <linux/seq_file.h>
43 #include <linux/compat.h>
44 #include <linux/rculist.h>
45 #include <net/busy_poll.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 #ifdef CONFIG_NET_RX_BUSY_POLL
230 	/* used to track busy poll napi_id */
231 	unsigned int napi_id;
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 };
264 
265 /*
266  * Configuration options available inside /proc/sys/fs/epoll/
267  */
268 /* Maximum number of epoll watched descriptors, per user */
269 static long max_user_watches __read_mostly;
270 
271 /*
272  * This mutex is used to serialize ep_free() and eventpoll_release_file().
273  */
274 static DEFINE_MUTEX(epmutex);
275 
276 /* Used to check for epoll file descriptor inclusion loops */
277 static struct nested_calls poll_loop_ncalls;
278 
279 /* Used for safe wake up implementation */
280 static struct nested_calls poll_safewake_ncalls;
281 
282 /* Used to call file's f_op->poll() under the nested calls boundaries */
283 static struct nested_calls poll_readywalk_ncalls;
284 
285 /* Slab cache used to allocate "struct epitem" */
286 static struct kmem_cache *epi_cache __read_mostly;
287 
288 /* Slab cache used to allocate "struct eppoll_entry" */
289 static struct kmem_cache *pwq_cache __read_mostly;
290 
291 /* Visited nodes during ep_loop_check(), so we can unset them when we finish */
292 static LIST_HEAD(visited_list);
293 
294 /*
295  * List of files with newly added links, where we may need to limit the number
296  * of emanating paths. Protected by the epmutex.
297  */
298 static LIST_HEAD(tfile_check_list);
299 
300 #ifdef CONFIG_SYSCTL
301 
302 #include <linux/sysctl.h>
303 
304 static long zero;
305 static long long_max = LONG_MAX;
306 
307 struct ctl_table epoll_table[] = {
308 	{
309 		.procname	= "max_user_watches",
310 		.data		= &max_user_watches,
311 		.maxlen		= sizeof(max_user_watches),
312 		.mode		= 0644,
313 		.proc_handler	= proc_doulongvec_minmax,
314 		.extra1		= &zero,
315 		.extra2		= &long_max,
316 	},
317 	{ }
318 };
319 #endif /* CONFIG_SYSCTL */
320 
321 static const struct file_operations eventpoll_fops;
322 
323 static inline int is_file_epoll(struct file *f)
324 {
325 	return f->f_op == &eventpoll_fops;
326 }
327 
328 /* Setup the structure that is used as key for the RB tree */
329 static inline void ep_set_ffd(struct epoll_filefd *ffd,
330 			      struct file *file, int fd)
331 {
332 	ffd->file = file;
333 	ffd->fd = fd;
334 }
335 
336 /* Compare RB tree keys */
337 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
338 			     struct epoll_filefd *p2)
339 {
340 	return (p1->file > p2->file ? +1:
341 	        (p1->file < p2->file ? -1 : p1->fd - p2->fd));
342 }
343 
344 /* Tells us if the item is currently linked */
345 static inline int ep_is_linked(struct list_head *p)
346 {
347 	return !list_empty(p);
348 }
349 
350 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_entry_t *p)
351 {
352 	return container_of(p, struct eppoll_entry, wait);
353 }
354 
355 /* Get the "struct epitem" from a wait queue pointer */
356 static inline struct epitem *ep_item_from_wait(wait_queue_entry_t *p)
357 {
358 	return container_of(p, struct eppoll_entry, wait)->base;
359 }
360 
361 /* Get the "struct epitem" from an epoll queue wrapper */
362 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
363 {
364 	return container_of(p, struct ep_pqueue, pt)->epi;
365 }
366 
367 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
368 static inline int ep_op_has_event(int op)
369 {
370 	return op != EPOLL_CTL_DEL;
371 }
372 
373 /* Initialize the poll safe wake up structure */
374 static void ep_nested_calls_init(struct nested_calls *ncalls)
375 {
376 	INIT_LIST_HEAD(&ncalls->tasks_call_list);
377 	spin_lock_init(&ncalls->lock);
378 }
379 
380 /**
381  * ep_events_available - Checks if ready events might be available.
382  *
383  * @ep: Pointer to the eventpoll context.
384  *
385  * Returns: Returns a value different than zero if ready events are available,
386  *          or zero otherwise.
387  */
388 static inline int ep_events_available(struct eventpoll *ep)
389 {
390 	return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
391 }
392 
393 #ifdef CONFIG_NET_RX_BUSY_POLL
394 static bool ep_busy_loop_end(void *p, unsigned long start_time)
395 {
396 	struct eventpoll *ep = p;
397 
398 	return ep_events_available(ep) || busy_loop_timeout(start_time);
399 }
400 #endif /* CONFIG_NET_RX_BUSY_POLL */
401 
402 /*
403  * Busy poll if globally on and supporting sockets found && no events,
404  * busy loop will return if need_resched or ep_events_available.
405  *
406  * we must do our busy polling with irqs enabled
407  */
408 static void ep_busy_loop(struct eventpoll *ep, int nonblock)
409 {
410 #ifdef CONFIG_NET_RX_BUSY_POLL
411 	unsigned int napi_id = READ_ONCE(ep->napi_id);
412 
413 	if ((napi_id >= MIN_NAPI_ID) && net_busy_loop_on())
414 		napi_busy_loop(napi_id, nonblock ? NULL : ep_busy_loop_end, ep);
415 #endif
416 }
417 
418 static inline void ep_reset_busy_poll_napi_id(struct eventpoll *ep)
419 {
420 #ifdef CONFIG_NET_RX_BUSY_POLL
421 	if (ep->napi_id)
422 		ep->napi_id = 0;
423 #endif
424 }
425 
426 /*
427  * Set epoll busy poll NAPI ID from sk.
428  */
429 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
430 {
431 #ifdef CONFIG_NET_RX_BUSY_POLL
432 	struct eventpoll *ep;
433 	unsigned int napi_id;
434 	struct socket *sock;
435 	struct sock *sk;
436 	int err;
437 
438 	if (!net_busy_loop_on())
439 		return;
440 
441 	sock = sock_from_file(epi->ffd.file, &err);
442 	if (!sock)
443 		return;
444 
445 	sk = sock->sk;
446 	if (!sk)
447 		return;
448 
449 	napi_id = READ_ONCE(sk->sk_napi_id);
450 	ep = epi->ep;
451 
452 	/* Non-NAPI IDs can be rejected
453 	 *	or
454 	 * Nothing to do if we already have this ID
455 	 */
456 	if (napi_id < MIN_NAPI_ID || napi_id == ep->napi_id)
457 		return;
458 
459 	/* record NAPI ID for use in next busy poll */
460 	ep->napi_id = napi_id;
461 #endif
462 }
463 
464 /**
465  * ep_call_nested - Perform a bound (possibly) nested call, by checking
466  *                  that the recursion limit is not exceeded, and that
467  *                  the same nested call (by the meaning of same cookie) is
468  *                  no re-entered.
469  *
470  * @ncalls: Pointer to the nested_calls structure to be used for this call.
471  * @max_nests: Maximum number of allowed nesting calls.
472  * @nproc: Nested call core function pointer.
473  * @priv: Opaque data to be passed to the @nproc callback.
474  * @cookie: Cookie to be used to identify this nested call.
475  * @ctx: This instance context.
476  *
477  * Returns: Returns the code returned by the @nproc callback, or -1 if
478  *          the maximum recursion limit has been exceeded.
479  */
480 static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
481 			  int (*nproc)(void *, void *, int), void *priv,
482 			  void *cookie, void *ctx)
483 {
484 	int error, call_nests = 0;
485 	unsigned long flags;
486 	struct list_head *lsthead = &ncalls->tasks_call_list;
487 	struct nested_call_node *tncur;
488 	struct nested_call_node tnode;
489 
490 	spin_lock_irqsave(&ncalls->lock, flags);
491 
492 	/*
493 	 * Try to see if the current task is already inside this wakeup call.
494 	 * We use a list here, since the population inside this set is always
495 	 * very much limited.
496 	 */
497 	list_for_each_entry(tncur, lsthead, llink) {
498 		if (tncur->ctx == ctx &&
499 		    (tncur->cookie == cookie || ++call_nests > max_nests)) {
500 			/*
501 			 * Ops ... loop detected or maximum nest level reached.
502 			 * We abort this wake by breaking the cycle itself.
503 			 */
504 			error = -1;
505 			goto out_unlock;
506 		}
507 	}
508 
509 	/* Add the current task and cookie to the list */
510 	tnode.ctx = ctx;
511 	tnode.cookie = cookie;
512 	list_add(&tnode.llink, lsthead);
513 
514 	spin_unlock_irqrestore(&ncalls->lock, flags);
515 
516 	/* Call the nested function */
517 	error = (*nproc)(priv, cookie, call_nests);
518 
519 	/* Remove the current task from the list */
520 	spin_lock_irqsave(&ncalls->lock, flags);
521 	list_del(&tnode.llink);
522 out_unlock:
523 	spin_unlock_irqrestore(&ncalls->lock, flags);
524 
525 	return error;
526 }
527 
528 /*
529  * As described in commit 0ccf831cb lockdep: annotate epoll
530  * the use of wait queues used by epoll is done in a very controlled
531  * manner. Wake ups can nest inside each other, but are never done
532  * with the same locking. For example:
533  *
534  *   dfd = socket(...);
535  *   efd1 = epoll_create();
536  *   efd2 = epoll_create();
537  *   epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
538  *   epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
539  *
540  * When a packet arrives to the device underneath "dfd", the net code will
541  * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
542  * callback wakeup entry on that queue, and the wake_up() performed by the
543  * "dfd" net code will end up in ep_poll_callback(). At this point epoll
544  * (efd1) notices that it may have some event ready, so it needs to wake up
545  * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
546  * that ends up in another wake_up(), after having checked about the
547  * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
548  * avoid stack blasting.
549  *
550  * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
551  * this special case of epoll.
552  */
553 #ifdef CONFIG_DEBUG_LOCK_ALLOC
554 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
555 				     unsigned long events, int subclass)
556 {
557 	unsigned long flags;
558 
559 	spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
560 	wake_up_locked_poll(wqueue, events);
561 	spin_unlock_irqrestore(&wqueue->lock, flags);
562 }
563 #else
564 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
565 				     unsigned long events, int subclass)
566 {
567 	wake_up_poll(wqueue, events);
568 }
569 #endif
570 
571 static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
572 {
573 	ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
574 			  1 + call_nests);
575 	return 0;
576 }
577 
578 /*
579  * Perform a safe wake up of the poll wait list. The problem is that
580  * with the new callback'd wake up system, it is possible that the
581  * poll callback is reentered from inside the call to wake_up() done
582  * on the poll wait queue head. The rule is that we cannot reenter the
583  * wake up code from the same task more than EP_MAX_NESTS times,
584  * and we cannot reenter the same wait queue head at all. This will
585  * enable to have a hierarchy of epoll file descriptor of no more than
586  * EP_MAX_NESTS deep.
587  */
588 static void ep_poll_safewake(wait_queue_head_t *wq)
589 {
590 	int this_cpu = get_cpu();
591 
592 	ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
593 		       ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
594 
595 	put_cpu();
596 }
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 	/* If it is cleared by POLLFREE, it should be rcu-safe */
604 	whead = rcu_dereference(pwq->whead);
605 	if (whead)
606 		remove_wait_queue(whead, &pwq->wait);
607 	rcu_read_unlock();
608 }
609 
610 /*
611  * This function unregisters poll callbacks from the associated file
612  * descriptor.  Must be called with "mtx" held (or "epmutex" if called from
613  * ep_free).
614  */
615 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
616 {
617 	struct list_head *lsthead = &epi->pwqlist;
618 	struct eppoll_entry *pwq;
619 
620 	while (!list_empty(lsthead)) {
621 		pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
622 
623 		list_del(&pwq->llink);
624 		ep_remove_wait_queue(pwq);
625 		kmem_cache_free(pwq_cache, pwq);
626 	}
627 }
628 
629 /* call only when ep->mtx is held */
630 static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
631 {
632 	return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
633 }
634 
635 /* call only when ep->mtx is held */
636 static inline void ep_pm_stay_awake(struct epitem *epi)
637 {
638 	struct wakeup_source *ws = ep_wakeup_source(epi);
639 
640 	if (ws)
641 		__pm_stay_awake(ws);
642 }
643 
644 static inline bool ep_has_wakeup_source(struct epitem *epi)
645 {
646 	return rcu_access_pointer(epi->ws) ? true : false;
647 }
648 
649 /* call when ep->mtx cannot be held (ep_poll_callback) */
650 static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
651 {
652 	struct wakeup_source *ws;
653 
654 	rcu_read_lock();
655 	ws = rcu_dereference(epi->ws);
656 	if (ws)
657 		__pm_stay_awake(ws);
658 	rcu_read_unlock();
659 }
660 
661 /**
662  * ep_scan_ready_list - Scans the ready list in a way that makes possible for
663  *                      the scan code, to call f_op->poll(). Also allows for
664  *                      O(NumReady) performance.
665  *
666  * @ep: Pointer to the epoll private data structure.
667  * @sproc: Pointer to the scan callback.
668  * @priv: Private opaque data passed to the @sproc callback.
669  * @depth: The current depth of recursive f_op->poll calls.
670  * @ep_locked: caller already holds ep->mtx
671  *
672  * Returns: The same integer error code returned by the @sproc callback.
673  */
674 static int ep_scan_ready_list(struct eventpoll *ep,
675 			      int (*sproc)(struct eventpoll *,
676 					   struct list_head *, void *),
677 			      void *priv, int depth, bool ep_locked)
678 {
679 	int error, pwake = 0;
680 	unsigned long flags;
681 	struct epitem *epi, *nepi;
682 	LIST_HEAD(txlist);
683 
684 	/*
685 	 * We need to lock this because we could be hit by
686 	 * eventpoll_release_file() and epoll_ctl().
687 	 */
688 
689 	if (!ep_locked)
690 		mutex_lock_nested(&ep->mtx, depth);
691 
692 	/*
693 	 * Steal the ready list, and re-init the original one to the
694 	 * empty list. Also, set ep->ovflist to NULL so that events
695 	 * happening while looping w/out locks, are not lost. We cannot
696 	 * have the poll callback to queue directly on ep->rdllist,
697 	 * because we want the "sproc" callback to be able to do it
698 	 * in a lockless way.
699 	 */
700 	spin_lock_irqsave(&ep->lock, flags);
701 	list_splice_init(&ep->rdllist, &txlist);
702 	ep->ovflist = NULL;
703 	spin_unlock_irqrestore(&ep->lock, flags);
704 
705 	/*
706 	 * Now call the callback function.
707 	 */
708 	error = (*sproc)(ep, &txlist, priv);
709 
710 	spin_lock_irqsave(&ep->lock, flags);
711 	/*
712 	 * During the time we spent inside the "sproc" callback, some
713 	 * other events might have been queued by the poll callback.
714 	 * We re-insert them inside the main ready-list here.
715 	 */
716 	for (nepi = ep->ovflist; (epi = nepi) != NULL;
717 	     nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
718 		/*
719 		 * We need to check if the item is already in the list.
720 		 * During the "sproc" callback execution time, items are
721 		 * queued into ->ovflist but the "txlist" might already
722 		 * contain them, and the list_splice() below takes care of them.
723 		 */
724 		if (!ep_is_linked(&epi->rdllink)) {
725 			list_add_tail(&epi->rdllink, &ep->rdllist);
726 			ep_pm_stay_awake(epi);
727 		}
728 	}
729 	/*
730 	 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
731 	 * releasing the lock, events will be queued in the normal way inside
732 	 * ep->rdllist.
733 	 */
734 	ep->ovflist = EP_UNACTIVE_PTR;
735 
736 	/*
737 	 * Quickly re-inject items left on "txlist".
738 	 */
739 	list_splice(&txlist, &ep->rdllist);
740 	__pm_relax(ep->ws);
741 
742 	if (!list_empty(&ep->rdllist)) {
743 		/*
744 		 * Wake up (if active) both the eventpoll wait list and
745 		 * the ->poll() wait list (delayed after we release the lock).
746 		 */
747 		if (waitqueue_active(&ep->wq))
748 			wake_up_locked(&ep->wq);
749 		if (waitqueue_active(&ep->poll_wait))
750 			pwake++;
751 	}
752 	spin_unlock_irqrestore(&ep->lock, flags);
753 
754 	if (!ep_locked)
755 		mutex_unlock(&ep->mtx);
756 
757 	/* We have to call this outside the lock */
758 	if (pwake)
759 		ep_poll_safewake(&ep->poll_wait);
760 
761 	return error;
762 }
763 
764 static void epi_rcu_free(struct rcu_head *head)
765 {
766 	struct epitem *epi = container_of(head, struct epitem, rcu);
767 	kmem_cache_free(epi_cache, epi);
768 }
769 
770 /*
771  * Removes a "struct epitem" from the eventpoll RB tree and deallocates
772  * all the associated resources. Must be called with "mtx" held.
773  */
774 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
775 {
776 	unsigned long flags;
777 	struct file *file = epi->ffd.file;
778 
779 	/*
780 	 * Removes poll wait queue hooks. We _have_ to do this without holding
781 	 * the "ep->lock" otherwise a deadlock might occur. This because of the
782 	 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
783 	 * queue head lock when unregistering the wait queue. The wakeup callback
784 	 * will run by holding the wait queue head lock and will call our callback
785 	 * that will try to get "ep->lock".
786 	 */
787 	ep_unregister_pollwait(ep, epi);
788 
789 	/* Remove the current item from the list of epoll hooks */
790 	spin_lock(&file->f_lock);
791 	list_del_rcu(&epi->fllink);
792 	spin_unlock(&file->f_lock);
793 
794 	rb_erase(&epi->rbn, &ep->rbr);
795 
796 	spin_lock_irqsave(&ep->lock, flags);
797 	if (ep_is_linked(&epi->rdllink))
798 		list_del_init(&epi->rdllink);
799 	spin_unlock_irqrestore(&ep->lock, flags);
800 
801 	wakeup_source_unregister(ep_wakeup_source(epi));
802 	/*
803 	 * At this point it is safe to free the eventpoll item. Use the union
804 	 * field epi->rcu, since we are trying to minimize the size of
805 	 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
806 	 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
807 	 * use of the rbn field.
808 	 */
809 	call_rcu(&epi->rcu, epi_rcu_free);
810 
811 	atomic_long_dec(&ep->user->epoll_watches);
812 
813 	return 0;
814 }
815 
816 static void ep_free(struct eventpoll *ep)
817 {
818 	struct rb_node *rbp;
819 	struct epitem *epi;
820 
821 	/* We need to release all tasks waiting for these file */
822 	if (waitqueue_active(&ep->poll_wait))
823 		ep_poll_safewake(&ep->poll_wait);
824 
825 	/*
826 	 * We need to lock this because we could be hit by
827 	 * eventpoll_release_file() while we're freeing the "struct eventpoll".
828 	 * We do not need to hold "ep->mtx" here because the epoll file
829 	 * is on the way to be removed and no one has references to it
830 	 * anymore. The only hit might come from eventpoll_release_file() but
831 	 * holding "epmutex" is sufficient here.
832 	 */
833 	mutex_lock(&epmutex);
834 
835 	/*
836 	 * Walks through the whole tree by unregistering poll callbacks.
837 	 */
838 	for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
839 		epi = rb_entry(rbp, struct epitem, rbn);
840 
841 		ep_unregister_pollwait(ep, epi);
842 		cond_resched();
843 	}
844 
845 	/*
846 	 * Walks through the whole tree by freeing each "struct epitem". At this
847 	 * point we are sure no poll callbacks will be lingering around, and also by
848 	 * holding "epmutex" we can be sure that no file cleanup code will hit
849 	 * us during this operation. So we can avoid the lock on "ep->lock".
850 	 * We do not need to lock ep->mtx, either, we only do it to prevent
851 	 * a lockdep warning.
852 	 */
853 	mutex_lock(&ep->mtx);
854 	while ((rbp = rb_first(&ep->rbr)) != NULL) {
855 		epi = rb_entry(rbp, struct epitem, rbn);
856 		ep_remove(ep, epi);
857 		cond_resched();
858 	}
859 	mutex_unlock(&ep->mtx);
860 
861 	mutex_unlock(&epmutex);
862 	mutex_destroy(&ep->mtx);
863 	free_uid(ep->user);
864 	wakeup_source_unregister(ep->ws);
865 	kfree(ep);
866 }
867 
868 static int ep_eventpoll_release(struct inode *inode, struct file *file)
869 {
870 	struct eventpoll *ep = file->private_data;
871 
872 	if (ep)
873 		ep_free(ep);
874 
875 	return 0;
876 }
877 
878 static inline unsigned int ep_item_poll(struct epitem *epi, poll_table *pt)
879 {
880 	pt->_key = epi->event.events;
881 
882 	return epi->ffd.file->f_op->poll(epi->ffd.file, pt) & epi->event.events;
883 }
884 
885 static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
886 			       void *priv)
887 {
888 	struct epitem *epi, *tmp;
889 	poll_table pt;
890 
891 	init_poll_funcptr(&pt, NULL);
892 
893 	list_for_each_entry_safe(epi, tmp, head, rdllink) {
894 		if (ep_item_poll(epi, &pt))
895 			return POLLIN | POLLRDNORM;
896 		else {
897 			/*
898 			 * Item has been dropped into the ready list by the poll
899 			 * callback, but it's not actually ready, as far as
900 			 * caller requested events goes. We can remove it here.
901 			 */
902 			__pm_relax(ep_wakeup_source(epi));
903 			list_del_init(&epi->rdllink);
904 		}
905 	}
906 
907 	return 0;
908 }
909 
910 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
911 				 poll_table *pt);
912 
913 struct readyevents_arg {
914 	struct eventpoll *ep;
915 	bool locked;
916 };
917 
918 static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
919 {
920 	struct readyevents_arg *arg = priv;
921 
922 	return ep_scan_ready_list(arg->ep, ep_read_events_proc, NULL,
923 				  call_nests + 1, arg->locked);
924 }
925 
926 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
927 {
928 	int pollflags;
929 	struct eventpoll *ep = file->private_data;
930 	struct readyevents_arg arg;
931 
932 	/*
933 	 * During ep_insert() we already hold the ep->mtx for the tfile.
934 	 * Prevent re-aquisition.
935 	 */
936 	arg.locked = wait && (wait->_qproc == ep_ptable_queue_proc);
937 	arg.ep = ep;
938 
939 	/* Insert inside our poll wait queue */
940 	poll_wait(file, &ep->poll_wait, wait);
941 
942 	/*
943 	 * Proceed to find out if wanted events are really available inside
944 	 * the ready list. This need to be done under ep_call_nested()
945 	 * supervision, since the call to f_op->poll() done on listed files
946 	 * could re-enter here.
947 	 */
948 	pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
949 				   ep_poll_readyevents_proc, &arg, ep, current);
950 
951 	return pollflags != -1 ? pollflags : 0;
952 }
953 
954 #ifdef CONFIG_PROC_FS
955 static void ep_show_fdinfo(struct seq_file *m, struct file *f)
956 {
957 	struct eventpoll *ep = f->private_data;
958 	struct rb_node *rbp;
959 
960 	mutex_lock(&ep->mtx);
961 	for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
962 		struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
963 		struct inode *inode = file_inode(epi->ffd.file);
964 
965 		seq_printf(m, "tfd: %8d events: %8x data: %16llx "
966 			   " pos:%lli ino:%lx sdev:%x\n",
967 			   epi->ffd.fd, epi->event.events,
968 			   (long long)epi->event.data,
969 			   (long long)epi->ffd.file->f_pos,
970 			   inode->i_ino, inode->i_sb->s_dev);
971 		if (seq_has_overflowed(m))
972 			break;
973 	}
974 	mutex_unlock(&ep->mtx);
975 }
976 #endif
977 
978 /* File callbacks that implement the eventpoll file behaviour */
979 static const struct file_operations eventpoll_fops = {
980 #ifdef CONFIG_PROC_FS
981 	.show_fdinfo	= ep_show_fdinfo,
982 #endif
983 	.release	= ep_eventpoll_release,
984 	.poll		= ep_eventpoll_poll,
985 	.llseek		= noop_llseek,
986 };
987 
988 /*
989  * This is called from eventpoll_release() to unlink files from the eventpoll
990  * interface. We need to have this facility to cleanup correctly files that are
991  * closed without being removed from the eventpoll interface.
992  */
993 void eventpoll_release_file(struct file *file)
994 {
995 	struct eventpoll *ep;
996 	struct epitem *epi, *next;
997 
998 	/*
999 	 * We don't want to get "file->f_lock" because it is not
1000 	 * necessary. It is not necessary because we're in the "struct file"
1001 	 * cleanup path, and this means that no one is using this file anymore.
1002 	 * So, for example, epoll_ctl() cannot hit here since if we reach this
1003 	 * point, the file counter already went to zero and fget() would fail.
1004 	 * The only hit might come from ep_free() but by holding the mutex
1005 	 * will correctly serialize the operation. We do need to acquire
1006 	 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
1007 	 * from anywhere but ep_free().
1008 	 *
1009 	 * Besides, ep_remove() acquires the lock, so we can't hold it here.
1010 	 */
1011 	mutex_lock(&epmutex);
1012 	list_for_each_entry_safe(epi, next, &file->f_ep_links, fllink) {
1013 		ep = epi->ep;
1014 		mutex_lock_nested(&ep->mtx, 0);
1015 		ep_remove(ep, epi);
1016 		mutex_unlock(&ep->mtx);
1017 	}
1018 	mutex_unlock(&epmutex);
1019 }
1020 
1021 static int ep_alloc(struct eventpoll **pep)
1022 {
1023 	int error;
1024 	struct user_struct *user;
1025 	struct eventpoll *ep;
1026 
1027 	user = get_current_user();
1028 	error = -ENOMEM;
1029 	ep = kzalloc(sizeof(*ep), GFP_KERNEL);
1030 	if (unlikely(!ep))
1031 		goto free_uid;
1032 
1033 	spin_lock_init(&ep->lock);
1034 	mutex_init(&ep->mtx);
1035 	init_waitqueue_head(&ep->wq);
1036 	init_waitqueue_head(&ep->poll_wait);
1037 	INIT_LIST_HEAD(&ep->rdllist);
1038 	ep->rbr = RB_ROOT;
1039 	ep->ovflist = EP_UNACTIVE_PTR;
1040 	ep->user = user;
1041 
1042 	*pep = ep;
1043 
1044 	return 0;
1045 
1046 free_uid:
1047 	free_uid(user);
1048 	return error;
1049 }
1050 
1051 /*
1052  * Search the file inside the eventpoll tree. The RB tree operations
1053  * are protected by the "mtx" mutex, and ep_find() must be called with
1054  * "mtx" held.
1055  */
1056 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
1057 {
1058 	int kcmp;
1059 	struct rb_node *rbp;
1060 	struct epitem *epi, *epir = NULL;
1061 	struct epoll_filefd ffd;
1062 
1063 	ep_set_ffd(&ffd, file, fd);
1064 	for (rbp = ep->rbr.rb_node; rbp; ) {
1065 		epi = rb_entry(rbp, struct epitem, rbn);
1066 		kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
1067 		if (kcmp > 0)
1068 			rbp = rbp->rb_right;
1069 		else if (kcmp < 0)
1070 			rbp = rbp->rb_left;
1071 		else {
1072 			epir = epi;
1073 			break;
1074 		}
1075 	}
1076 
1077 	return epir;
1078 }
1079 
1080 #ifdef CONFIG_CHECKPOINT_RESTORE
1081 static struct epitem *ep_find_tfd(struct eventpoll *ep, int tfd, unsigned long toff)
1082 {
1083 	struct rb_node *rbp;
1084 	struct epitem *epi;
1085 
1086 	for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1087 		epi = rb_entry(rbp, struct epitem, rbn);
1088 		if (epi->ffd.fd == tfd) {
1089 			if (toff == 0)
1090 				return epi;
1091 			else
1092 				toff--;
1093 		}
1094 		cond_resched();
1095 	}
1096 
1097 	return NULL;
1098 }
1099 
1100 struct file *get_epoll_tfile_raw_ptr(struct file *file, int tfd,
1101 				     unsigned long toff)
1102 {
1103 	struct file *file_raw;
1104 	struct eventpoll *ep;
1105 	struct epitem *epi;
1106 
1107 	if (!is_file_epoll(file))
1108 		return ERR_PTR(-EINVAL);
1109 
1110 	ep = file->private_data;
1111 
1112 	mutex_lock(&ep->mtx);
1113 	epi = ep_find_tfd(ep, tfd, toff);
1114 	if (epi)
1115 		file_raw = epi->ffd.file;
1116 	else
1117 		file_raw = ERR_PTR(-ENOENT);
1118 	mutex_unlock(&ep->mtx);
1119 
1120 	return file_raw;
1121 }
1122 #endif /* CONFIG_CHECKPOINT_RESTORE */
1123 
1124 /*
1125  * This is the callback that is passed to the wait queue wakeup
1126  * mechanism. It is called by the stored file descriptors when they
1127  * have events to report.
1128  */
1129 static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1130 {
1131 	int pwake = 0;
1132 	unsigned long flags;
1133 	struct epitem *epi = ep_item_from_wait(wait);
1134 	struct eventpoll *ep = epi->ep;
1135 	int ewake = 0;
1136 
1137 	if ((unsigned long)key & POLLFREE) {
1138 		ep_pwq_from_wait(wait)->whead = NULL;
1139 		/*
1140 		 * whead = NULL above can race with ep_remove_wait_queue()
1141 		 * which can do another remove_wait_queue() after us, so we
1142 		 * can't use __remove_wait_queue(). whead->lock is held by
1143 		 * the caller.
1144 		 */
1145 		list_del_init(&wait->entry);
1146 	}
1147 
1148 	spin_lock_irqsave(&ep->lock, flags);
1149 
1150 	ep_set_busy_poll_napi_id(epi);
1151 
1152 	/*
1153 	 * If the event mask does not contain any poll(2) event, we consider the
1154 	 * descriptor to be disabled. This condition is likely the effect of the
1155 	 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1156 	 * until the next EPOLL_CTL_MOD will be issued.
1157 	 */
1158 	if (!(epi->event.events & ~EP_PRIVATE_BITS))
1159 		goto out_unlock;
1160 
1161 	/*
1162 	 * Check the events coming with the callback. At this stage, not
1163 	 * every device reports the events in the "key" parameter of the
1164 	 * callback. We need to be able to handle both cases here, hence the
1165 	 * test for "key" != NULL before the event match test.
1166 	 */
1167 	if (key && !((unsigned long) key & epi->event.events))
1168 		goto out_unlock;
1169 
1170 	/*
1171 	 * If we are transferring events to userspace, we can hold no locks
1172 	 * (because we're accessing user memory, and because of linux f_op->poll()
1173 	 * semantics). All the events that happen during that period of time are
1174 	 * chained in ep->ovflist and requeued later on.
1175 	 */
1176 	if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
1177 		if (epi->next == EP_UNACTIVE_PTR) {
1178 			epi->next = ep->ovflist;
1179 			ep->ovflist = epi;
1180 			if (epi->ws) {
1181 				/*
1182 				 * Activate ep->ws since epi->ws may get
1183 				 * deactivated at any time.
1184 				 */
1185 				__pm_stay_awake(ep->ws);
1186 			}
1187 
1188 		}
1189 		goto out_unlock;
1190 	}
1191 
1192 	/* If this file is already in the ready list we exit soon */
1193 	if (!ep_is_linked(&epi->rdllink)) {
1194 		list_add_tail(&epi->rdllink, &ep->rdllist);
1195 		ep_pm_stay_awake_rcu(epi);
1196 	}
1197 
1198 	/*
1199 	 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1200 	 * wait list.
1201 	 */
1202 	if (waitqueue_active(&ep->wq)) {
1203 		if ((epi->event.events & EPOLLEXCLUSIVE) &&
1204 					!((unsigned long)key & POLLFREE)) {
1205 			switch ((unsigned long)key & EPOLLINOUT_BITS) {
1206 			case POLLIN:
1207 				if (epi->event.events & POLLIN)
1208 					ewake = 1;
1209 				break;
1210 			case POLLOUT:
1211 				if (epi->event.events & POLLOUT)
1212 					ewake = 1;
1213 				break;
1214 			case 0:
1215 				ewake = 1;
1216 				break;
1217 			}
1218 		}
1219 		wake_up_locked(&ep->wq);
1220 	}
1221 	if (waitqueue_active(&ep->poll_wait))
1222 		pwake++;
1223 
1224 out_unlock:
1225 	spin_unlock_irqrestore(&ep->lock, flags);
1226 
1227 	/* We have to call this outside the lock */
1228 	if (pwake)
1229 		ep_poll_safewake(&ep->poll_wait);
1230 
1231 	if (epi->event.events & EPOLLEXCLUSIVE)
1232 		return ewake;
1233 
1234 	return 1;
1235 }
1236 
1237 /*
1238  * This is the callback that is used to add our wait queue to the
1239  * target file wakeup lists.
1240  */
1241 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1242 				 poll_table *pt)
1243 {
1244 	struct epitem *epi = ep_item_from_epqueue(pt);
1245 	struct eppoll_entry *pwq;
1246 
1247 	if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
1248 		init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1249 		pwq->whead = whead;
1250 		pwq->base = epi;
1251 		if (epi->event.events & EPOLLEXCLUSIVE)
1252 			add_wait_queue_exclusive(whead, &pwq->wait);
1253 		else
1254 			add_wait_queue(whead, &pwq->wait);
1255 		list_add_tail(&pwq->llink, &epi->pwqlist);
1256 		epi->nwait++;
1257 	} else {
1258 		/* We have to signal that an error occurred */
1259 		epi->nwait = -1;
1260 	}
1261 }
1262 
1263 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1264 {
1265 	int kcmp;
1266 	struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
1267 	struct epitem *epic;
1268 
1269 	while (*p) {
1270 		parent = *p;
1271 		epic = rb_entry(parent, struct epitem, rbn);
1272 		kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1273 		if (kcmp > 0)
1274 			p = &parent->rb_right;
1275 		else
1276 			p = &parent->rb_left;
1277 	}
1278 	rb_link_node(&epi->rbn, parent, p);
1279 	rb_insert_color(&epi->rbn, &ep->rbr);
1280 }
1281 
1282 
1283 
1284 #define PATH_ARR_SIZE 5
1285 /*
1286  * These are the number paths of length 1 to 5, that we are allowing to emanate
1287  * from a single file of interest. For example, we allow 1000 paths of length
1288  * 1, to emanate from each file of interest. This essentially represents the
1289  * potential wakeup paths, which need to be limited in order to avoid massive
1290  * uncontrolled wakeup storms. The common use case should be a single ep which
1291  * is connected to n file sources. In this case each file source has 1 path
1292  * of length 1. Thus, the numbers below should be more than sufficient. These
1293  * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1294  * and delete can't add additional paths. Protected by the epmutex.
1295  */
1296 static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1297 static int path_count[PATH_ARR_SIZE];
1298 
1299 static int path_count_inc(int nests)
1300 {
1301 	/* Allow an arbitrary number of depth 1 paths */
1302 	if (nests == 0)
1303 		return 0;
1304 
1305 	if (++path_count[nests] > path_limits[nests])
1306 		return -1;
1307 	return 0;
1308 }
1309 
1310 static void path_count_init(void)
1311 {
1312 	int i;
1313 
1314 	for (i = 0; i < PATH_ARR_SIZE; i++)
1315 		path_count[i] = 0;
1316 }
1317 
1318 static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1319 {
1320 	int error = 0;
1321 	struct file *file = priv;
1322 	struct file *child_file;
1323 	struct epitem *epi;
1324 
1325 	/* CTL_DEL can remove links here, but that can't increase our count */
1326 	rcu_read_lock();
1327 	list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) {
1328 		child_file = epi->ep->file;
1329 		if (is_file_epoll(child_file)) {
1330 			if (list_empty(&child_file->f_ep_links)) {
1331 				if (path_count_inc(call_nests)) {
1332 					error = -1;
1333 					break;
1334 				}
1335 			} else {
1336 				error = ep_call_nested(&poll_loop_ncalls,
1337 							EP_MAX_NESTS,
1338 							reverse_path_check_proc,
1339 							child_file, child_file,
1340 							current);
1341 			}
1342 			if (error != 0)
1343 				break;
1344 		} else {
1345 			printk(KERN_ERR "reverse_path_check_proc: "
1346 				"file is not an ep!\n");
1347 		}
1348 	}
1349 	rcu_read_unlock();
1350 	return error;
1351 }
1352 
1353 /**
1354  * reverse_path_check - The tfile_check_list is list of file *, which have
1355  *                      links that are proposed to be newly added. We need to
1356  *                      make sure that those added links don't add too many
1357  *                      paths such that we will spend all our time waking up
1358  *                      eventpoll objects.
1359  *
1360  * Returns: Returns zero if the proposed links don't create too many paths,
1361  *	    -1 otherwise.
1362  */
1363 static int reverse_path_check(void)
1364 {
1365 	int error = 0;
1366 	struct file *current_file;
1367 
1368 	/* let's call this for all tfiles */
1369 	list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
1370 		path_count_init();
1371 		error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1372 					reverse_path_check_proc, current_file,
1373 					current_file, current);
1374 		if (error)
1375 			break;
1376 	}
1377 	return error;
1378 }
1379 
1380 static int ep_create_wakeup_source(struct epitem *epi)
1381 {
1382 	const char *name;
1383 	struct wakeup_source *ws;
1384 
1385 	if (!epi->ep->ws) {
1386 		epi->ep->ws = wakeup_source_register("eventpoll");
1387 		if (!epi->ep->ws)
1388 			return -ENOMEM;
1389 	}
1390 
1391 	name = epi->ffd.file->f_path.dentry->d_name.name;
1392 	ws = wakeup_source_register(name);
1393 
1394 	if (!ws)
1395 		return -ENOMEM;
1396 	rcu_assign_pointer(epi->ws, ws);
1397 
1398 	return 0;
1399 }
1400 
1401 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1402 static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1403 {
1404 	struct wakeup_source *ws = ep_wakeup_source(epi);
1405 
1406 	RCU_INIT_POINTER(epi->ws, NULL);
1407 
1408 	/*
1409 	 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1410 	 * used internally by wakeup_source_remove, too (called by
1411 	 * wakeup_source_unregister), so we cannot use call_rcu
1412 	 */
1413 	synchronize_rcu();
1414 	wakeup_source_unregister(ws);
1415 }
1416 
1417 /*
1418  * Must be called with "mtx" held.
1419  */
1420 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
1421 		     struct file *tfile, int fd, int full_check)
1422 {
1423 	int error, revents, pwake = 0;
1424 	unsigned long flags;
1425 	long user_watches;
1426 	struct epitem *epi;
1427 	struct ep_pqueue epq;
1428 
1429 	user_watches = atomic_long_read(&ep->user->epoll_watches);
1430 	if (unlikely(user_watches >= max_user_watches))
1431 		return -ENOSPC;
1432 	if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1433 		return -ENOMEM;
1434 
1435 	/* Item initialization follow here ... */
1436 	INIT_LIST_HEAD(&epi->rdllink);
1437 	INIT_LIST_HEAD(&epi->fllink);
1438 	INIT_LIST_HEAD(&epi->pwqlist);
1439 	epi->ep = ep;
1440 	ep_set_ffd(&epi->ffd, tfile, fd);
1441 	epi->event = *event;
1442 	epi->nwait = 0;
1443 	epi->next = EP_UNACTIVE_PTR;
1444 	if (epi->event.events & EPOLLWAKEUP) {
1445 		error = ep_create_wakeup_source(epi);
1446 		if (error)
1447 			goto error_create_wakeup_source;
1448 	} else {
1449 		RCU_INIT_POINTER(epi->ws, NULL);
1450 	}
1451 
1452 	/* Initialize the poll table using the queue callback */
1453 	epq.epi = epi;
1454 	init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1455 
1456 	/*
1457 	 * Attach the item to the poll hooks and get current event bits.
1458 	 * We can safely use the file* here because its usage count has
1459 	 * been increased by the caller of this function. Note that after
1460 	 * this operation completes, the poll callback can start hitting
1461 	 * the new item.
1462 	 */
1463 	revents = ep_item_poll(epi, &epq.pt);
1464 
1465 	/*
1466 	 * We have to check if something went wrong during the poll wait queue
1467 	 * install process. Namely an allocation for a wait queue failed due
1468 	 * high memory pressure.
1469 	 */
1470 	error = -ENOMEM;
1471 	if (epi->nwait < 0)
1472 		goto error_unregister;
1473 
1474 	/* Add the current item to the list of active epoll hook for this file */
1475 	spin_lock(&tfile->f_lock);
1476 	list_add_tail_rcu(&epi->fllink, &tfile->f_ep_links);
1477 	spin_unlock(&tfile->f_lock);
1478 
1479 	/*
1480 	 * Add the current item to the RB tree. All RB tree operations are
1481 	 * protected by "mtx", and ep_insert() is called with "mtx" held.
1482 	 */
1483 	ep_rbtree_insert(ep, epi);
1484 
1485 	/* now check if we've created too many backpaths */
1486 	error = -EINVAL;
1487 	if (full_check && reverse_path_check())
1488 		goto error_remove_epi;
1489 
1490 	/* We have to drop the new item inside our item list to keep track of it */
1491 	spin_lock_irqsave(&ep->lock, flags);
1492 
1493 	/* record NAPI ID of new item if present */
1494 	ep_set_busy_poll_napi_id(epi);
1495 
1496 	/* If the file is already "ready" we drop it inside the ready list */
1497 	if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
1498 		list_add_tail(&epi->rdllink, &ep->rdllist);
1499 		ep_pm_stay_awake(epi);
1500 
1501 		/* Notify waiting tasks that events are available */
1502 		if (waitqueue_active(&ep->wq))
1503 			wake_up_locked(&ep->wq);
1504 		if (waitqueue_active(&ep->poll_wait))
1505 			pwake++;
1506 	}
1507 
1508 	spin_unlock_irqrestore(&ep->lock, flags);
1509 
1510 	atomic_long_inc(&ep->user->epoll_watches);
1511 
1512 	/* We have to call this outside the lock */
1513 	if (pwake)
1514 		ep_poll_safewake(&ep->poll_wait);
1515 
1516 	return 0;
1517 
1518 error_remove_epi:
1519 	spin_lock(&tfile->f_lock);
1520 	list_del_rcu(&epi->fllink);
1521 	spin_unlock(&tfile->f_lock);
1522 
1523 	rb_erase(&epi->rbn, &ep->rbr);
1524 
1525 error_unregister:
1526 	ep_unregister_pollwait(ep, epi);
1527 
1528 	/*
1529 	 * We need to do this because an event could have been arrived on some
1530 	 * allocated wait queue. Note that we don't care about the ep->ovflist
1531 	 * list, since that is used/cleaned only inside a section bound by "mtx".
1532 	 * And ep_insert() is called with "mtx" held.
1533 	 */
1534 	spin_lock_irqsave(&ep->lock, flags);
1535 	if (ep_is_linked(&epi->rdllink))
1536 		list_del_init(&epi->rdllink);
1537 	spin_unlock_irqrestore(&ep->lock, flags);
1538 
1539 	wakeup_source_unregister(ep_wakeup_source(epi));
1540 
1541 error_create_wakeup_source:
1542 	kmem_cache_free(epi_cache, epi);
1543 
1544 	return error;
1545 }
1546 
1547 /*
1548  * Modify the interest event mask by dropping an event if the new mask
1549  * has a match in the current file status. Must be called with "mtx" held.
1550  */
1551 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1552 {
1553 	int pwake = 0;
1554 	unsigned int revents;
1555 	poll_table pt;
1556 
1557 	init_poll_funcptr(&pt, NULL);
1558 
1559 	/*
1560 	 * Set the new event interest mask before calling f_op->poll();
1561 	 * otherwise we might miss an event that happens between the
1562 	 * f_op->poll() call and the new event set registering.
1563 	 */
1564 	epi->event.events = event->events; /* need barrier below */
1565 	epi->event.data = event->data; /* protected by mtx */
1566 	if (epi->event.events & EPOLLWAKEUP) {
1567 		if (!ep_has_wakeup_source(epi))
1568 			ep_create_wakeup_source(epi);
1569 	} else if (ep_has_wakeup_source(epi)) {
1570 		ep_destroy_wakeup_source(epi);
1571 	}
1572 
1573 	/*
1574 	 * The following barrier has two effects:
1575 	 *
1576 	 * 1) Flush epi changes above to other CPUs.  This ensures
1577 	 *    we do not miss events from ep_poll_callback if an
1578 	 *    event occurs immediately after we call f_op->poll().
1579 	 *    We need this because we did not take ep->lock while
1580 	 *    changing epi above (but ep_poll_callback does take
1581 	 *    ep->lock).
1582 	 *
1583 	 * 2) We also need to ensure we do not miss _past_ events
1584 	 *    when calling f_op->poll().  This barrier also
1585 	 *    pairs with the barrier in wq_has_sleeper (see
1586 	 *    comments for wq_has_sleeper).
1587 	 *
1588 	 * This barrier will now guarantee ep_poll_callback or f_op->poll
1589 	 * (or both) will notice the readiness of an item.
1590 	 */
1591 	smp_mb();
1592 
1593 	/*
1594 	 * Get current event bits. We can safely use the file* here because
1595 	 * its usage count has been increased by the caller of this function.
1596 	 */
1597 	revents = ep_item_poll(epi, &pt);
1598 
1599 	/*
1600 	 * If the item is "hot" and it is not registered inside the ready
1601 	 * list, push it inside.
1602 	 */
1603 	if (revents & event->events) {
1604 		spin_lock_irq(&ep->lock);
1605 		if (!ep_is_linked(&epi->rdllink)) {
1606 			list_add_tail(&epi->rdllink, &ep->rdllist);
1607 			ep_pm_stay_awake(epi);
1608 
1609 			/* Notify waiting tasks that events are available */
1610 			if (waitqueue_active(&ep->wq))
1611 				wake_up_locked(&ep->wq);
1612 			if (waitqueue_active(&ep->poll_wait))
1613 				pwake++;
1614 		}
1615 		spin_unlock_irq(&ep->lock);
1616 	}
1617 
1618 	/* We have to call this outside the lock */
1619 	if (pwake)
1620 		ep_poll_safewake(&ep->poll_wait);
1621 
1622 	return 0;
1623 }
1624 
1625 static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1626 			       void *priv)
1627 {
1628 	struct ep_send_events_data *esed = priv;
1629 	int eventcnt;
1630 	unsigned int revents;
1631 	struct epitem *epi;
1632 	struct epoll_event __user *uevent;
1633 	struct wakeup_source *ws;
1634 	poll_table pt;
1635 
1636 	init_poll_funcptr(&pt, NULL);
1637 
1638 	/*
1639 	 * We can loop without lock because we are passed a task private list.
1640 	 * Items cannot vanish during the loop because ep_scan_ready_list() is
1641 	 * holding "mtx" during this call.
1642 	 */
1643 	for (eventcnt = 0, uevent = esed->events;
1644 	     !list_empty(head) && eventcnt < esed->maxevents;) {
1645 		epi = list_first_entry(head, struct epitem, rdllink);
1646 
1647 		/*
1648 		 * Activate ep->ws before deactivating epi->ws to prevent
1649 		 * triggering auto-suspend here (in case we reactive epi->ws
1650 		 * below).
1651 		 *
1652 		 * This could be rearranged to delay the deactivation of epi->ws
1653 		 * instead, but then epi->ws would temporarily be out of sync
1654 		 * with ep_is_linked().
1655 		 */
1656 		ws = ep_wakeup_source(epi);
1657 		if (ws) {
1658 			if (ws->active)
1659 				__pm_stay_awake(ep->ws);
1660 			__pm_relax(ws);
1661 		}
1662 
1663 		list_del_init(&epi->rdllink);
1664 
1665 		revents = ep_item_poll(epi, &pt);
1666 
1667 		/*
1668 		 * If the event mask intersect the caller-requested one,
1669 		 * deliver the event to userspace. Again, ep_scan_ready_list()
1670 		 * is holding "mtx", so no operations coming from userspace
1671 		 * can change the item.
1672 		 */
1673 		if (revents) {
1674 			if (__put_user(revents, &uevent->events) ||
1675 			    __put_user(epi->event.data, &uevent->data)) {
1676 				list_add(&epi->rdllink, head);
1677 				ep_pm_stay_awake(epi);
1678 				return eventcnt ? eventcnt : -EFAULT;
1679 			}
1680 			eventcnt++;
1681 			uevent++;
1682 			if (epi->event.events & EPOLLONESHOT)
1683 				epi->event.events &= EP_PRIVATE_BITS;
1684 			else if (!(epi->event.events & EPOLLET)) {
1685 				/*
1686 				 * If this file has been added with Level
1687 				 * Trigger mode, we need to insert back inside
1688 				 * the ready list, so that the next call to
1689 				 * epoll_wait() will check again the events
1690 				 * availability. At this point, no one can insert
1691 				 * into ep->rdllist besides us. The epoll_ctl()
1692 				 * callers are locked out by
1693 				 * ep_scan_ready_list() holding "mtx" and the
1694 				 * poll callback will queue them in ep->ovflist.
1695 				 */
1696 				list_add_tail(&epi->rdllink, &ep->rdllist);
1697 				ep_pm_stay_awake(epi);
1698 			}
1699 		}
1700 	}
1701 
1702 	return eventcnt;
1703 }
1704 
1705 static int ep_send_events(struct eventpoll *ep,
1706 			  struct epoll_event __user *events, int maxevents)
1707 {
1708 	struct ep_send_events_data esed;
1709 
1710 	esed.maxevents = maxevents;
1711 	esed.events = events;
1712 
1713 	return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false);
1714 }
1715 
1716 static inline struct timespec64 ep_set_mstimeout(long ms)
1717 {
1718 	struct timespec64 now, ts = {
1719 		.tv_sec = ms / MSEC_PER_SEC,
1720 		.tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1721 	};
1722 
1723 	ktime_get_ts64(&now);
1724 	return timespec64_add_safe(now, ts);
1725 }
1726 
1727 /**
1728  * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1729  *           event buffer.
1730  *
1731  * @ep: Pointer to the eventpoll context.
1732  * @events: Pointer to the userspace buffer where the ready events should be
1733  *          stored.
1734  * @maxevents: Size (in terms of number of events) of the caller event buffer.
1735  * @timeout: Maximum timeout for the ready events fetch operation, in
1736  *           milliseconds. If the @timeout is zero, the function will not block,
1737  *           while if the @timeout is less than zero, the function will block
1738  *           until at least one event has been retrieved (or an error
1739  *           occurred).
1740  *
1741  * Returns: Returns the number of ready events which have been fetched, or an
1742  *          error code, in case of error.
1743  */
1744 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1745 		   int maxevents, long timeout)
1746 {
1747 	int res = 0, eavail, timed_out = 0;
1748 	unsigned long flags;
1749 	u64 slack = 0;
1750 	wait_queue_entry_t wait;
1751 	ktime_t expires, *to = NULL;
1752 
1753 	if (timeout > 0) {
1754 		struct timespec64 end_time = ep_set_mstimeout(timeout);
1755 
1756 		slack = select_estimate_accuracy(&end_time);
1757 		to = &expires;
1758 		*to = timespec64_to_ktime(end_time);
1759 	} else if (timeout == 0) {
1760 		/*
1761 		 * Avoid the unnecessary trip to the wait queue loop, if the
1762 		 * caller specified a non blocking operation.
1763 		 */
1764 		timed_out = 1;
1765 		spin_lock_irqsave(&ep->lock, flags);
1766 		goto check_events;
1767 	}
1768 
1769 fetch_events:
1770 
1771 	if (!ep_events_available(ep))
1772 		ep_busy_loop(ep, timed_out);
1773 
1774 	spin_lock_irqsave(&ep->lock, flags);
1775 
1776 	if (!ep_events_available(ep)) {
1777 		/*
1778 		 * Busy poll timed out.  Drop NAPI ID for now, we can add
1779 		 * it back in when we have moved a socket with a valid NAPI
1780 		 * ID onto the ready list.
1781 		 */
1782 		ep_reset_busy_poll_napi_id(ep);
1783 
1784 		/*
1785 		 * We don't have any available event to return to the caller.
1786 		 * We need to sleep here, and we will be wake up by
1787 		 * ep_poll_callback() when events will become available.
1788 		 */
1789 		init_waitqueue_entry(&wait, current);
1790 		__add_wait_queue_exclusive(&ep->wq, &wait);
1791 
1792 		for (;;) {
1793 			/*
1794 			 * We don't want to sleep if the ep_poll_callback() sends us
1795 			 * a wakeup in between. That's why we set the task state
1796 			 * to TASK_INTERRUPTIBLE before doing the checks.
1797 			 */
1798 			set_current_state(TASK_INTERRUPTIBLE);
1799 			/*
1800 			 * Always short-circuit for fatal signals to allow
1801 			 * threads to make a timely exit without the chance of
1802 			 * finding more events available and fetching
1803 			 * repeatedly.
1804 			 */
1805 			if (fatal_signal_pending(current)) {
1806 				res = -EINTR;
1807 				break;
1808 			}
1809 			if (ep_events_available(ep) || timed_out)
1810 				break;
1811 			if (signal_pending(current)) {
1812 				res = -EINTR;
1813 				break;
1814 			}
1815 
1816 			spin_unlock_irqrestore(&ep->lock, flags);
1817 			if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
1818 				timed_out = 1;
1819 
1820 			spin_lock_irqsave(&ep->lock, flags);
1821 		}
1822 
1823 		__remove_wait_queue(&ep->wq, &wait);
1824 		__set_current_state(TASK_RUNNING);
1825 	}
1826 check_events:
1827 	/* Is it worth to try to dig for events ? */
1828 	eavail = ep_events_available(ep);
1829 
1830 	spin_unlock_irqrestore(&ep->lock, flags);
1831 
1832 	/*
1833 	 * Try to transfer events to user space. In case we get 0 events and
1834 	 * there's still timeout left over, we go trying again in search of
1835 	 * more luck.
1836 	 */
1837 	if (!res && eavail &&
1838 	    !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1839 		goto fetch_events;
1840 
1841 	return res;
1842 }
1843 
1844 /**
1845  * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1846  *                      API, to verify that adding an epoll file inside another
1847  *                      epoll structure, does not violate the constraints, in
1848  *                      terms of closed loops, or too deep chains (which can
1849  *                      result in excessive stack usage).
1850  *
1851  * @priv: Pointer to the epoll file to be currently checked.
1852  * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1853  *          data structure pointer.
1854  * @call_nests: Current dept of the @ep_call_nested() call stack.
1855  *
1856  * Returns: Returns zero if adding the epoll @file inside current epoll
1857  *          structure @ep does not violate the constraints, or -1 otherwise.
1858  */
1859 static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1860 {
1861 	int error = 0;
1862 	struct file *file = priv;
1863 	struct eventpoll *ep = file->private_data;
1864 	struct eventpoll *ep_tovisit;
1865 	struct rb_node *rbp;
1866 	struct epitem *epi;
1867 
1868 	mutex_lock_nested(&ep->mtx, call_nests + 1);
1869 	ep->visited = 1;
1870 	list_add(&ep->visited_list_link, &visited_list);
1871 	for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1872 		epi = rb_entry(rbp, struct epitem, rbn);
1873 		if (unlikely(is_file_epoll(epi->ffd.file))) {
1874 			ep_tovisit = epi->ffd.file->private_data;
1875 			if (ep_tovisit->visited)
1876 				continue;
1877 			error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1878 					ep_loop_check_proc, epi->ffd.file,
1879 					ep_tovisit, current);
1880 			if (error != 0)
1881 				break;
1882 		} else {
1883 			/*
1884 			 * If we've reached a file that is not associated with
1885 			 * an ep, then we need to check if the newly added
1886 			 * links are going to add too many wakeup paths. We do
1887 			 * this by adding it to the tfile_check_list, if it's
1888 			 * not already there, and calling reverse_path_check()
1889 			 * during ep_insert().
1890 			 */
1891 			if (list_empty(&epi->ffd.file->f_tfile_llink))
1892 				list_add(&epi->ffd.file->f_tfile_llink,
1893 					 &tfile_check_list);
1894 		}
1895 	}
1896 	mutex_unlock(&ep->mtx);
1897 
1898 	return error;
1899 }
1900 
1901 /**
1902  * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1903  *                 another epoll file (represented by @ep) does not create
1904  *                 closed loops or too deep chains.
1905  *
1906  * @ep: Pointer to the epoll private data structure.
1907  * @file: Pointer to the epoll file to be checked.
1908  *
1909  * Returns: Returns zero if adding the epoll @file inside current epoll
1910  *          structure @ep does not violate the constraints, or -1 otherwise.
1911  */
1912 static int ep_loop_check(struct eventpoll *ep, struct file *file)
1913 {
1914 	int ret;
1915 	struct eventpoll *ep_cur, *ep_next;
1916 
1917 	ret = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1918 			      ep_loop_check_proc, file, ep, current);
1919 	/* clear visited list */
1920 	list_for_each_entry_safe(ep_cur, ep_next, &visited_list,
1921 							visited_list_link) {
1922 		ep_cur->visited = 0;
1923 		list_del(&ep_cur->visited_list_link);
1924 	}
1925 	return ret;
1926 }
1927 
1928 static void clear_tfile_check_list(void)
1929 {
1930 	struct file *file;
1931 
1932 	/* first clear the tfile_check_list */
1933 	while (!list_empty(&tfile_check_list)) {
1934 		file = list_first_entry(&tfile_check_list, struct file,
1935 					f_tfile_llink);
1936 		list_del_init(&file->f_tfile_llink);
1937 	}
1938 	INIT_LIST_HEAD(&tfile_check_list);
1939 }
1940 
1941 /*
1942  * Open an eventpoll file descriptor.
1943  */
1944 SYSCALL_DEFINE1(epoll_create1, int, flags)
1945 {
1946 	int error, fd;
1947 	struct eventpoll *ep = NULL;
1948 	struct file *file;
1949 
1950 	/* Check the EPOLL_* constant for consistency.  */
1951 	BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1952 
1953 	if (flags & ~EPOLL_CLOEXEC)
1954 		return -EINVAL;
1955 	/*
1956 	 * Create the internal data structure ("struct eventpoll").
1957 	 */
1958 	error = ep_alloc(&ep);
1959 	if (error < 0)
1960 		return error;
1961 	/*
1962 	 * Creates all the items needed to setup an eventpoll file. That is,
1963 	 * a file structure and a free file descriptor.
1964 	 */
1965 	fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1966 	if (fd < 0) {
1967 		error = fd;
1968 		goto out_free_ep;
1969 	}
1970 	file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
1971 				 O_RDWR | (flags & O_CLOEXEC));
1972 	if (IS_ERR(file)) {
1973 		error = PTR_ERR(file);
1974 		goto out_free_fd;
1975 	}
1976 	ep->file = file;
1977 	fd_install(fd, file);
1978 	return fd;
1979 
1980 out_free_fd:
1981 	put_unused_fd(fd);
1982 out_free_ep:
1983 	ep_free(ep);
1984 	return error;
1985 }
1986 
1987 SYSCALL_DEFINE1(epoll_create, int, size)
1988 {
1989 	if (size <= 0)
1990 		return -EINVAL;
1991 
1992 	return sys_epoll_create1(0);
1993 }
1994 
1995 /*
1996  * The following function implements the controller interface for
1997  * the eventpoll file that enables the insertion/removal/change of
1998  * file descriptors inside the interest set.
1999  */
2000 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
2001 		struct epoll_event __user *, event)
2002 {
2003 	int error;
2004 	int full_check = 0;
2005 	struct fd f, tf;
2006 	struct eventpoll *ep;
2007 	struct epitem *epi;
2008 	struct epoll_event epds;
2009 	struct eventpoll *tep = NULL;
2010 
2011 	error = -EFAULT;
2012 	if (ep_op_has_event(op) &&
2013 	    copy_from_user(&epds, event, sizeof(struct epoll_event)))
2014 		goto error_return;
2015 
2016 	error = -EBADF;
2017 	f = fdget(epfd);
2018 	if (!f.file)
2019 		goto error_return;
2020 
2021 	/* Get the "struct file *" for the target file */
2022 	tf = fdget(fd);
2023 	if (!tf.file)
2024 		goto error_fput;
2025 
2026 	/* The target file descriptor must support poll */
2027 	error = -EPERM;
2028 	if (!tf.file->f_op->poll)
2029 		goto error_tgt_fput;
2030 
2031 	/* Check if EPOLLWAKEUP is allowed */
2032 	if (ep_op_has_event(op))
2033 		ep_take_care_of_epollwakeup(&epds);
2034 
2035 	/*
2036 	 * We have to check that the file structure underneath the file descriptor
2037 	 * the user passed to us _is_ an eventpoll file. And also we do not permit
2038 	 * adding an epoll file descriptor inside itself.
2039 	 */
2040 	error = -EINVAL;
2041 	if (f.file == tf.file || !is_file_epoll(f.file))
2042 		goto error_tgt_fput;
2043 
2044 	/*
2045 	 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
2046 	 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
2047 	 * Also, we do not currently supported nested exclusive wakeups.
2048 	 */
2049 	if (ep_op_has_event(op) && (epds.events & EPOLLEXCLUSIVE)) {
2050 		if (op == EPOLL_CTL_MOD)
2051 			goto error_tgt_fput;
2052 		if (op == EPOLL_CTL_ADD && (is_file_epoll(tf.file) ||
2053 				(epds.events & ~EPOLLEXCLUSIVE_OK_BITS)))
2054 			goto error_tgt_fput;
2055 	}
2056 
2057 	/*
2058 	 * At this point it is safe to assume that the "private_data" contains
2059 	 * our own data structure.
2060 	 */
2061 	ep = f.file->private_data;
2062 
2063 	/*
2064 	 * When we insert an epoll file descriptor, inside another epoll file
2065 	 * descriptor, there is the change of creating closed loops, which are
2066 	 * better be handled here, than in more critical paths. While we are
2067 	 * checking for loops we also determine the list of files reachable
2068 	 * and hang them on the tfile_check_list, so we can check that we
2069 	 * haven't created too many possible wakeup paths.
2070 	 *
2071 	 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
2072 	 * the epoll file descriptor is attaching directly to a wakeup source,
2073 	 * unless the epoll file descriptor is nested. The purpose of taking the
2074 	 * 'epmutex' on add is to prevent complex toplogies such as loops and
2075 	 * deep wakeup paths from forming in parallel through multiple
2076 	 * EPOLL_CTL_ADD operations.
2077 	 */
2078 	mutex_lock_nested(&ep->mtx, 0);
2079 	if (op == EPOLL_CTL_ADD) {
2080 		if (!list_empty(&f.file->f_ep_links) ||
2081 						is_file_epoll(tf.file)) {
2082 			full_check = 1;
2083 			mutex_unlock(&ep->mtx);
2084 			mutex_lock(&epmutex);
2085 			if (is_file_epoll(tf.file)) {
2086 				error = -ELOOP;
2087 				if (ep_loop_check(ep, tf.file) != 0) {
2088 					clear_tfile_check_list();
2089 					goto error_tgt_fput;
2090 				}
2091 			} else
2092 				list_add(&tf.file->f_tfile_llink,
2093 							&tfile_check_list);
2094 			mutex_lock_nested(&ep->mtx, 0);
2095 			if (is_file_epoll(tf.file)) {
2096 				tep = tf.file->private_data;
2097 				mutex_lock_nested(&tep->mtx, 1);
2098 			}
2099 		}
2100 	}
2101 
2102 	/*
2103 	 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
2104 	 * above, we can be sure to be able to use the item looked up by
2105 	 * ep_find() till we release the mutex.
2106 	 */
2107 	epi = ep_find(ep, tf.file, fd);
2108 
2109 	error = -EINVAL;
2110 	switch (op) {
2111 	case EPOLL_CTL_ADD:
2112 		if (!epi) {
2113 			epds.events |= POLLERR | POLLHUP;
2114 			error = ep_insert(ep, &epds, tf.file, fd, full_check);
2115 		} else
2116 			error = -EEXIST;
2117 		if (full_check)
2118 			clear_tfile_check_list();
2119 		break;
2120 	case EPOLL_CTL_DEL:
2121 		if (epi)
2122 			error = ep_remove(ep, epi);
2123 		else
2124 			error = -ENOENT;
2125 		break;
2126 	case EPOLL_CTL_MOD:
2127 		if (epi) {
2128 			if (!(epi->event.events & EPOLLEXCLUSIVE)) {
2129 				epds.events |= POLLERR | POLLHUP;
2130 				error = ep_modify(ep, epi, &epds);
2131 			}
2132 		} else
2133 			error = -ENOENT;
2134 		break;
2135 	}
2136 	if (tep != NULL)
2137 		mutex_unlock(&tep->mtx);
2138 	mutex_unlock(&ep->mtx);
2139 
2140 error_tgt_fput:
2141 	if (full_check)
2142 		mutex_unlock(&epmutex);
2143 
2144 	fdput(tf);
2145 error_fput:
2146 	fdput(f);
2147 error_return:
2148 
2149 	return error;
2150 }
2151 
2152 /*
2153  * Implement the event wait interface for the eventpoll file. It is the kernel
2154  * part of the user space epoll_wait(2).
2155  */
2156 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
2157 		int, maxevents, int, timeout)
2158 {
2159 	int error;
2160 	struct fd f;
2161 	struct eventpoll *ep;
2162 
2163 	/* The maximum number of event must be greater than zero */
2164 	if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
2165 		return -EINVAL;
2166 
2167 	/* Verify that the area passed by the user is writeable */
2168 	if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event)))
2169 		return -EFAULT;
2170 
2171 	/* Get the "struct file *" for the eventpoll file */
2172 	f = fdget(epfd);
2173 	if (!f.file)
2174 		return -EBADF;
2175 
2176 	/*
2177 	 * We have to check that the file structure underneath the fd
2178 	 * the user passed to us _is_ an eventpoll file.
2179 	 */
2180 	error = -EINVAL;
2181 	if (!is_file_epoll(f.file))
2182 		goto error_fput;
2183 
2184 	/*
2185 	 * At this point it is safe to assume that the "private_data" contains
2186 	 * our own data structure.
2187 	 */
2188 	ep = f.file->private_data;
2189 
2190 	/* Time to fish for events ... */
2191 	error = ep_poll(ep, events, maxevents, timeout);
2192 
2193 error_fput:
2194 	fdput(f);
2195 	return error;
2196 }
2197 
2198 /*
2199  * Implement the event wait interface for the eventpoll file. It is the kernel
2200  * part of the user space epoll_pwait(2).
2201  */
2202 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2203 		int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2204 		size_t, sigsetsize)
2205 {
2206 	int error;
2207 	sigset_t ksigmask, sigsaved;
2208 
2209 	/*
2210 	 * If the caller wants a certain signal mask to be set during the wait,
2211 	 * we apply it here.
2212 	 */
2213 	if (sigmask) {
2214 		if (sigsetsize != sizeof(sigset_t))
2215 			return -EINVAL;
2216 		if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
2217 			return -EFAULT;
2218 		sigsaved = current->blocked;
2219 		set_current_blocked(&ksigmask);
2220 	}
2221 
2222 	error = sys_epoll_wait(epfd, events, maxevents, timeout);
2223 
2224 	/*
2225 	 * If we changed the signal mask, we need to restore the original one.
2226 	 * In case we've got a signal while waiting, we do not restore the
2227 	 * signal mask yet, and we allow do_signal() to deliver the signal on
2228 	 * the way back to userspace, before the signal mask is restored.
2229 	 */
2230 	if (sigmask) {
2231 		if (error == -EINTR) {
2232 			memcpy(&current->saved_sigmask, &sigsaved,
2233 			       sizeof(sigsaved));
2234 			set_restore_sigmask();
2235 		} else
2236 			set_current_blocked(&sigsaved);
2237 	}
2238 
2239 	return error;
2240 }
2241 
2242 #ifdef CONFIG_COMPAT
2243 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2244 			struct epoll_event __user *, events,
2245 			int, maxevents, int, timeout,
2246 			const compat_sigset_t __user *, sigmask,
2247 			compat_size_t, sigsetsize)
2248 {
2249 	long err;
2250 	compat_sigset_t csigmask;
2251 	sigset_t ksigmask, sigsaved;
2252 
2253 	/*
2254 	 * If the caller wants a certain signal mask to be set during the wait,
2255 	 * we apply it here.
2256 	 */
2257 	if (sigmask) {
2258 		if (sigsetsize != sizeof(compat_sigset_t))
2259 			return -EINVAL;
2260 		if (copy_from_user(&csigmask, sigmask, sizeof(csigmask)))
2261 			return -EFAULT;
2262 		sigset_from_compat(&ksigmask, &csigmask);
2263 		sigsaved = current->blocked;
2264 		set_current_blocked(&ksigmask);
2265 	}
2266 
2267 	err = sys_epoll_wait(epfd, events, maxevents, timeout);
2268 
2269 	/*
2270 	 * If we changed the signal mask, we need to restore the original one.
2271 	 * In case we've got a signal while waiting, we do not restore the
2272 	 * signal mask yet, and we allow do_signal() to deliver the signal on
2273 	 * the way back to userspace, before the signal mask is restored.
2274 	 */
2275 	if (sigmask) {
2276 		if (err == -EINTR) {
2277 			memcpy(&current->saved_sigmask, &sigsaved,
2278 			       sizeof(sigsaved));
2279 			set_restore_sigmask();
2280 		} else
2281 			set_current_blocked(&sigsaved);
2282 	}
2283 
2284 	return err;
2285 }
2286 #endif
2287 
2288 static int __init eventpoll_init(void)
2289 {
2290 	struct sysinfo si;
2291 
2292 	si_meminfo(&si);
2293 	/*
2294 	 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2295 	 */
2296 	max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2297 		EP_ITEM_COST;
2298 	BUG_ON(max_user_watches < 0);
2299 
2300 	/*
2301 	 * Initialize the structure used to perform epoll file descriptor
2302 	 * inclusion loops checks.
2303 	 */
2304 	ep_nested_calls_init(&poll_loop_ncalls);
2305 
2306 	/* Initialize the structure used to perform safe poll wait head wake ups */
2307 	ep_nested_calls_init(&poll_safewake_ncalls);
2308 
2309 	/* Initialize the structure used to perform file's f_op->poll() calls */
2310 	ep_nested_calls_init(&poll_readywalk_ncalls);
2311 
2312 	/*
2313 	 * We can have many thousands of epitems, so prevent this from
2314 	 * using an extra cache line on 64-bit (and smaller) CPUs
2315 	 */
2316 	BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2317 
2318 	/* Allocates slab cache used to allocate "struct epitem" items */
2319 	epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2320 			0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2321 
2322 	/* Allocates slab cache used to allocate "struct eppoll_entry" */
2323 	pwq_cache = kmem_cache_create("eventpoll_pwq",
2324 			sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
2325 
2326 	return 0;
2327 }
2328 fs_initcall(eventpoll_init);
2329