]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blob - fs/eventpoll.c
rtc: at91sam9: Remove unnecessary offset variable checks
[mirror_ubuntu-jammy-kernel.git] / fs / eventpoll.c
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 /* Wait structure used by the poll hooks */
113 struct eppoll_entry {
114 /* List header used to link this structure to the "struct epitem" */
115 struct eppoll_entry *next;
116
117 /* The "base" pointer is set to the container "struct epitem" */
118 struct epitem *base;
119
120 /*
121 * Wait queue item that will be linked to the target file wait
122 * queue head.
123 */
124 wait_queue_entry_t wait;
125
126 /* The wait queue head that linked the "wait" wait queue item */
127 wait_queue_head_t *whead;
128 };
129
130 /*
131 * Each file descriptor added to the eventpoll interface will
132 * have an entry of this type linked to the "rbr" RB tree.
133 * Avoid increasing the size of this struct, there can be many thousands
134 * of these on a server and we do not want this to take another cache line.
135 */
136 struct epitem {
137 union {
138 /* RB tree node links this structure to the eventpoll RB tree */
139 struct rb_node rbn;
140 /* Used to free the struct epitem */
141 struct rcu_head rcu;
142 };
143
144 /* List header used to link this structure to the eventpoll ready list */
145 struct list_head rdllink;
146
147 /*
148 * Works together "struct eventpoll"->ovflist in keeping the
149 * single linked chain of items.
150 */
151 struct epitem *next;
152
153 /* The file descriptor information this item refers to */
154 struct epoll_filefd ffd;
155
156 /* List containing poll wait queues */
157 struct eppoll_entry *pwqlist;
158
159 /* The "container" of this item */
160 struct eventpoll *ep;
161
162 /* List header used to link this item to the "struct file" items list */
163 struct hlist_node fllink;
164
165 /* wakeup_source used when EPOLLWAKEUP is set */
166 struct wakeup_source __rcu *ws;
167
168 /* The structure that describe the interested events and the source fd */
169 struct epoll_event event;
170 };
171
172 /*
173 * This structure is stored inside the "private_data" member of the file
174 * structure and represents the main data structure for the eventpoll
175 * interface.
176 */
177 struct eventpoll {
178 /*
179 * This mutex is used to ensure that files are not removed
180 * while epoll is using them. This is held during the event
181 * collection loop, the file cleanup path, the epoll file exit
182 * code and the ctl operations.
183 */
184 struct mutex mtx;
185
186 /* Wait queue used by sys_epoll_wait() */
187 wait_queue_head_t wq;
188
189 /* Wait queue used by file->poll() */
190 wait_queue_head_t poll_wait;
191
192 /* List of ready file descriptors */
193 struct list_head rdllist;
194
195 /* Lock which protects rdllist and ovflist */
196 rwlock_t lock;
197
198 /* RB tree root used to store monitored fd structs */
199 struct rb_root_cached rbr;
200
201 /*
202 * This is a single linked list that chains all the "struct epitem" that
203 * happened while transferring ready events to userspace w/out
204 * holding ->lock.
205 */
206 struct epitem *ovflist;
207
208 /* wakeup_source used when ep_scan_ready_list is running */
209 struct wakeup_source *ws;
210
211 /* The user that created the eventpoll descriptor */
212 struct user_struct *user;
213
214 struct file *file;
215
216 /* used to optimize loop detection check */
217 u64 gen;
218 struct hlist_head refs;
219
220 #ifdef CONFIG_NET_RX_BUSY_POLL
221 /* used to track busy poll napi_id */
222 unsigned int napi_id;
223 #endif
224
225 #ifdef CONFIG_DEBUG_LOCK_ALLOC
226 /* tracks wakeup nests for lockdep validation */
227 u8 nests;
228 #endif
229 };
230
231 /* Wrapper struct used by poll queueing */
232 struct ep_pqueue {
233 poll_table pt;
234 struct epitem *epi;
235 };
236
237 /*
238 * Configuration options available inside /proc/sys/fs/epoll/
239 */
240 /* Maximum number of epoll watched descriptors, per user */
241 static long max_user_watches __read_mostly;
242
243 /*
244 * This mutex is used to serialize ep_free() and eventpoll_release_file().
245 */
246 static DEFINE_MUTEX(epmutex);
247
248 static u64 loop_check_gen = 0;
249
250 /* Used to check for epoll file descriptor inclusion loops */
251 static struct eventpoll *inserting_into;
252
253 /* Slab cache used to allocate "struct epitem" */
254 static struct kmem_cache *epi_cache __read_mostly;
255
256 /* Slab cache used to allocate "struct eppoll_entry" */
257 static struct kmem_cache *pwq_cache __read_mostly;
258
259 /*
260 * List of files with newly added links, where we may need to limit the number
261 * of emanating paths. Protected by the epmutex.
262 */
263 struct epitems_head {
264 struct hlist_head epitems;
265 struct epitems_head *next;
266 };
267 static struct epitems_head *tfile_check_list = EP_UNACTIVE_PTR;
268
269 static struct kmem_cache *ephead_cache __read_mostly;
270
271 static inline void free_ephead(struct epitems_head *head)
272 {
273 if (head)
274 kmem_cache_free(ephead_cache, head);
275 }
276
277 static void list_file(struct file *file)
278 {
279 struct epitems_head *head;
280
281 head = container_of(file->f_ep, struct epitems_head, epitems);
282 if (!head->next) {
283 head->next = tfile_check_list;
284 tfile_check_list = head;
285 }
286 }
287
288 static void unlist_file(struct epitems_head *head)
289 {
290 struct epitems_head *to_free = head;
291 struct hlist_node *p = rcu_dereference(hlist_first_rcu(&head->epitems));
292 if (p) {
293 struct epitem *epi= container_of(p, struct epitem, fllink);
294 spin_lock(&epi->ffd.file->f_lock);
295 if (!hlist_empty(&head->epitems))
296 to_free = NULL;
297 head->next = NULL;
298 spin_unlock(&epi->ffd.file->f_lock);
299 }
300 free_ephead(to_free);
301 }
302
303 #ifdef CONFIG_SYSCTL
304
305 #include <linux/sysctl.h>
306
307 static long long_zero;
308 static long long_max = LONG_MAX;
309
310 struct ctl_table epoll_table[] = {
311 {
312 .procname = "max_user_watches",
313 .data = &max_user_watches,
314 .maxlen = sizeof(max_user_watches),
315 .mode = 0644,
316 .proc_handler = proc_doulongvec_minmax,
317 .extra1 = &long_zero,
318 .extra2 = &long_max,
319 },
320 { }
321 };
322 #endif /* CONFIG_SYSCTL */
323
324 static const struct file_operations eventpoll_fops;
325
326 static inline int is_file_epoll(struct file *f)
327 {
328 return f->f_op == &eventpoll_fops;
329 }
330
331 /* Setup the structure that is used as key for the RB tree */
332 static inline void ep_set_ffd(struct epoll_filefd *ffd,
333 struct file *file, int fd)
334 {
335 ffd->file = file;
336 ffd->fd = fd;
337 }
338
339 /* Compare RB tree keys */
340 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
341 struct epoll_filefd *p2)
342 {
343 return (p1->file > p2->file ? +1:
344 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
345 }
346
347 /* Tells us if the item is currently linked */
348 static inline int ep_is_linked(struct epitem *epi)
349 {
350 return !list_empty(&epi->rdllink);
351 }
352
353 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_entry_t *p)
354 {
355 return container_of(p, struct eppoll_entry, wait);
356 }
357
358 /* Get the "struct epitem" from a wait queue pointer */
359 static inline struct epitem *ep_item_from_wait(wait_queue_entry_t *p)
360 {
361 return container_of(p, struct eppoll_entry, wait)->base;
362 }
363
364 /**
365 * ep_events_available - Checks if ready events might be available.
366 *
367 * @ep: Pointer to the eventpoll context.
368 *
369 * Return: a value different than %zero if ready events are available,
370 * or %zero otherwise.
371 */
372 static inline int ep_events_available(struct eventpoll *ep)
373 {
374 return !list_empty_careful(&ep->rdllist) ||
375 READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR;
376 }
377
378 #ifdef CONFIG_NET_RX_BUSY_POLL
379 static bool ep_busy_loop_end(void *p, unsigned long start_time)
380 {
381 struct eventpoll *ep = p;
382
383 return ep_events_available(ep) || busy_loop_timeout(start_time);
384 }
385
386 /*
387 * Busy poll if globally on and supporting sockets found && no events,
388 * busy loop will return if need_resched or ep_events_available.
389 *
390 * we must do our busy polling with irqs enabled
391 */
392 static bool ep_busy_loop(struct eventpoll *ep, int nonblock)
393 {
394 unsigned int napi_id = READ_ONCE(ep->napi_id);
395
396 if ((napi_id >= MIN_NAPI_ID) && net_busy_loop_on()) {
397 napi_busy_loop(napi_id, nonblock ? NULL : ep_busy_loop_end, ep, false,
398 BUSY_POLL_BUDGET);
399 if (ep_events_available(ep))
400 return true;
401 /*
402 * Busy poll timed out. Drop NAPI ID for now, we can add
403 * it back in when we have moved a socket with a valid NAPI
404 * ID onto the ready list.
405 */
406 ep->napi_id = 0;
407 return false;
408 }
409 return false;
410 }
411
412 /*
413 * Set epoll busy poll NAPI ID from sk.
414 */
415 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
416 {
417 struct eventpoll *ep;
418 unsigned int napi_id;
419 struct socket *sock;
420 struct sock *sk;
421
422 if (!net_busy_loop_on())
423 return;
424
425 sock = sock_from_file(epi->ffd.file);
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 bool ep_busy_loop(struct eventpoll *ep, int nonblock)
450 {
451 return false;
452 }
453
454 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
455 {
456 }
457
458 #endif /* CONFIG_NET_RX_BUSY_POLL */
459
460 /*
461 * As described in commit 0ccf831cb lockdep: annotate epoll
462 * the use of wait queues used by epoll is done in a very controlled
463 * manner. Wake ups can nest inside each other, but are never done
464 * with the same locking. For example:
465 *
466 * dfd = socket(...);
467 * efd1 = epoll_create();
468 * efd2 = epoll_create();
469 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
470 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
471 *
472 * When a packet arrives to the device underneath "dfd", the net code will
473 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
474 * callback wakeup entry on that queue, and the wake_up() performed by the
475 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
476 * (efd1) notices that it may have some event ready, so it needs to wake up
477 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
478 * that ends up in another wake_up(), after having checked about the
479 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
480 * avoid stack blasting.
481 *
482 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
483 * this special case of epoll.
484 */
485 #ifdef CONFIG_DEBUG_LOCK_ALLOC
486
487 static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi)
488 {
489 struct eventpoll *ep_src;
490 unsigned long flags;
491 u8 nests = 0;
492
493 /*
494 * To set the subclass or nesting level for spin_lock_irqsave_nested()
495 * it might be natural to create a per-cpu nest count. However, since
496 * we can recurse on ep->poll_wait.lock, and a non-raw spinlock can
497 * schedule() in the -rt kernel, the per-cpu variable are no longer
498 * protected. Thus, we are introducing a per eventpoll nest field.
499 * If we are not being call from ep_poll_callback(), epi is NULL and
500 * we are at the first level of nesting, 0. Otherwise, we are being
501 * called from ep_poll_callback() and if a previous wakeup source is
502 * not an epoll file itself, we are at depth 1 since the wakeup source
503 * is depth 0. If the wakeup source is a previous epoll file in the
504 * wakeup chain then we use its nests value and record ours as
505 * nests + 1. The previous epoll file nests value is stable since its
506 * already holding its own poll_wait.lock.
507 */
508 if (epi) {
509 if ((is_file_epoll(epi->ffd.file))) {
510 ep_src = epi->ffd.file->private_data;
511 nests = ep_src->nests;
512 } else {
513 nests = 1;
514 }
515 }
516 spin_lock_irqsave_nested(&ep->poll_wait.lock, flags, nests);
517 ep->nests = nests + 1;
518 wake_up_locked_poll(&ep->poll_wait, EPOLLIN);
519 ep->nests = 0;
520 spin_unlock_irqrestore(&ep->poll_wait.lock, flags);
521 }
522
523 #else
524
525 static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi)
526 {
527 wake_up_poll(&ep->poll_wait, EPOLLIN);
528 }
529
530 #endif
531
532 static void ep_remove_wait_queue(struct eppoll_entry *pwq)
533 {
534 wait_queue_head_t *whead;
535
536 rcu_read_lock();
537 /*
538 * If it is cleared by POLLFREE, it should be rcu-safe.
539 * If we read NULL we need a barrier paired with
540 * smp_store_release() in ep_poll_callback(), otherwise
541 * we rely on whead->lock.
542 */
543 whead = smp_load_acquire(&pwq->whead);
544 if (whead)
545 remove_wait_queue(whead, &pwq->wait);
546 rcu_read_unlock();
547 }
548
549 /*
550 * This function unregisters poll callbacks from the associated file
551 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
552 * ep_free).
553 */
554 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
555 {
556 struct eppoll_entry **p = &epi->pwqlist;
557 struct eppoll_entry *pwq;
558
559 while ((pwq = *p) != NULL) {
560 *p = pwq->next;
561 ep_remove_wait_queue(pwq);
562 kmem_cache_free(pwq_cache, pwq);
563 }
564 }
565
566 /* call only when ep->mtx is held */
567 static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
568 {
569 return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
570 }
571
572 /* call only when ep->mtx is held */
573 static inline void ep_pm_stay_awake(struct epitem *epi)
574 {
575 struct wakeup_source *ws = ep_wakeup_source(epi);
576
577 if (ws)
578 __pm_stay_awake(ws);
579 }
580
581 static inline bool ep_has_wakeup_source(struct epitem *epi)
582 {
583 return rcu_access_pointer(epi->ws) ? true : false;
584 }
585
586 /* call when ep->mtx cannot be held (ep_poll_callback) */
587 static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
588 {
589 struct wakeup_source *ws;
590
591 rcu_read_lock();
592 ws = rcu_dereference(epi->ws);
593 if (ws)
594 __pm_stay_awake(ws);
595 rcu_read_unlock();
596 }
597
598
599 /*
600 * ep->mutex needs to be held because we could be hit by
601 * eventpoll_release_file() and epoll_ctl().
602 */
603 static void ep_start_scan(struct eventpoll *ep, struct list_head *txlist)
604 {
605 /*
606 * Steal the ready list, and re-init the original one to the
607 * empty list. Also, set ep->ovflist to NULL so that events
608 * happening while looping w/out locks, are not lost. We cannot
609 * have the poll callback to queue directly on ep->rdllist,
610 * because we want the "sproc" callback to be able to do it
611 * in a lockless way.
612 */
613 lockdep_assert_irqs_enabled();
614 write_lock_irq(&ep->lock);
615 list_splice_init(&ep->rdllist, txlist);
616 WRITE_ONCE(ep->ovflist, NULL);
617 write_unlock_irq(&ep->lock);
618 }
619
620 static void ep_done_scan(struct eventpoll *ep,
621 struct list_head *txlist)
622 {
623 struct epitem *epi, *nepi;
624
625 write_lock_irq(&ep->lock);
626 /*
627 * During the time we spent inside the "sproc" callback, some
628 * other events might have been queued by the poll callback.
629 * We re-insert them inside the main ready-list here.
630 */
631 for (nepi = READ_ONCE(ep->ovflist); (epi = nepi) != NULL;
632 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
633 /*
634 * We need to check if the item is already in the list.
635 * During the "sproc" callback execution time, items are
636 * queued into ->ovflist but the "txlist" might already
637 * contain them, and the list_splice() below takes care of them.
638 */
639 if (!ep_is_linked(epi)) {
640 /*
641 * ->ovflist is LIFO, so we have to reverse it in order
642 * to keep in FIFO.
643 */
644 list_add(&epi->rdllink, &ep->rdllist);
645 ep_pm_stay_awake(epi);
646 }
647 }
648 /*
649 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
650 * releasing the lock, events will be queued in the normal way inside
651 * ep->rdllist.
652 */
653 WRITE_ONCE(ep->ovflist, EP_UNACTIVE_PTR);
654
655 /*
656 * Quickly re-inject items left on "txlist".
657 */
658 list_splice(txlist, &ep->rdllist);
659 __pm_relax(ep->ws);
660
661 if (!list_empty(&ep->rdllist)) {
662 if (waitqueue_active(&ep->wq))
663 wake_up(&ep->wq);
664 }
665
666 write_unlock_irq(&ep->lock);
667 }
668
669 static void epi_rcu_free(struct rcu_head *head)
670 {
671 struct epitem *epi = container_of(head, struct epitem, rcu);
672 kmem_cache_free(epi_cache, epi);
673 }
674
675 /*
676 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
677 * all the associated resources. Must be called with "mtx" held.
678 */
679 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
680 {
681 struct file *file = epi->ffd.file;
682 struct epitems_head *to_free;
683 struct hlist_head *head;
684
685 lockdep_assert_irqs_enabled();
686
687 /*
688 * Removes poll wait queue hooks.
689 */
690 ep_unregister_pollwait(ep, epi);
691
692 /* Remove the current item from the list of epoll hooks */
693 spin_lock(&file->f_lock);
694 to_free = NULL;
695 head = file->f_ep;
696 if (head->first == &epi->fllink && !epi->fllink.next) {
697 file->f_ep = NULL;
698 if (!is_file_epoll(file)) {
699 struct epitems_head *v;
700 v = container_of(head, struct epitems_head, epitems);
701 if (!smp_load_acquire(&v->next))
702 to_free = v;
703 }
704 }
705 hlist_del_rcu(&epi->fllink);
706 spin_unlock(&file->f_lock);
707 free_ephead(to_free);
708
709 rb_erase_cached(&epi->rbn, &ep->rbr);
710
711 write_lock_irq(&ep->lock);
712 if (ep_is_linked(epi))
713 list_del_init(&epi->rdllink);
714 write_unlock_irq(&ep->lock);
715
716 wakeup_source_unregister(ep_wakeup_source(epi));
717 /*
718 * At this point it is safe to free the eventpoll item. Use the union
719 * field epi->rcu, since we are trying to minimize the size of
720 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
721 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
722 * use of the rbn field.
723 */
724 call_rcu(&epi->rcu, epi_rcu_free);
725
726 atomic_long_dec(&ep->user->epoll_watches);
727
728 return 0;
729 }
730
731 static void ep_free(struct eventpoll *ep)
732 {
733 struct rb_node *rbp;
734 struct epitem *epi;
735
736 /* We need to release all tasks waiting for these file */
737 if (waitqueue_active(&ep->poll_wait))
738 ep_poll_safewake(ep, NULL);
739
740 /*
741 * We need to lock this because we could be hit by
742 * eventpoll_release_file() while we're freeing the "struct eventpoll".
743 * We do not need to hold "ep->mtx" here because the epoll file
744 * is on the way to be removed and no one has references to it
745 * anymore. The only hit might come from eventpoll_release_file() but
746 * holding "epmutex" is sufficient here.
747 */
748 mutex_lock(&epmutex);
749
750 /*
751 * Walks through the whole tree by unregistering poll callbacks.
752 */
753 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
754 epi = rb_entry(rbp, struct epitem, rbn);
755
756 ep_unregister_pollwait(ep, epi);
757 cond_resched();
758 }
759
760 /*
761 * Walks through the whole tree by freeing each "struct epitem". At this
762 * point we are sure no poll callbacks will be lingering around, and also by
763 * holding "epmutex" we can be sure that no file cleanup code will hit
764 * us during this operation. So we can avoid the lock on "ep->lock".
765 * We do not need to lock ep->mtx, either, we only do it to prevent
766 * a lockdep warning.
767 */
768 mutex_lock(&ep->mtx);
769 while ((rbp = rb_first_cached(&ep->rbr)) != NULL) {
770 epi = rb_entry(rbp, struct epitem, rbn);
771 ep_remove(ep, epi);
772 cond_resched();
773 }
774 mutex_unlock(&ep->mtx);
775
776 mutex_unlock(&epmutex);
777 mutex_destroy(&ep->mtx);
778 free_uid(ep->user);
779 wakeup_source_unregister(ep->ws);
780 kfree(ep);
781 }
782
783 static int ep_eventpoll_release(struct inode *inode, struct file *file)
784 {
785 struct eventpoll *ep = file->private_data;
786
787 if (ep)
788 ep_free(ep);
789
790 return 0;
791 }
792
793 static __poll_t ep_item_poll(const struct epitem *epi, poll_table *pt, int depth);
794
795 static __poll_t __ep_eventpoll_poll(struct file *file, poll_table *wait, int depth)
796 {
797 struct eventpoll *ep = file->private_data;
798 LIST_HEAD(txlist);
799 struct epitem *epi, *tmp;
800 poll_table pt;
801 __poll_t res = 0;
802
803 init_poll_funcptr(&pt, NULL);
804
805 /* Insert inside our poll wait queue */
806 poll_wait(file, &ep->poll_wait, wait);
807
808 /*
809 * Proceed to find out if wanted events are really available inside
810 * the ready list.
811 */
812 mutex_lock_nested(&ep->mtx, depth);
813 ep_start_scan(ep, &txlist);
814 list_for_each_entry_safe(epi, tmp, &txlist, rdllink) {
815 if (ep_item_poll(epi, &pt, depth + 1)) {
816 res = EPOLLIN | EPOLLRDNORM;
817 break;
818 } else {
819 /*
820 * Item has been dropped into the ready list by the poll
821 * callback, but it's not actually ready, as far as
822 * caller requested events goes. We can remove it here.
823 */
824 __pm_relax(ep_wakeup_source(epi));
825 list_del_init(&epi->rdllink);
826 }
827 }
828 ep_done_scan(ep, &txlist);
829 mutex_unlock(&ep->mtx);
830 return res;
831 }
832
833 /*
834 * Differs from ep_eventpoll_poll() in that internal callers already have
835 * the ep->mtx so we need to start from depth=1, such that mutex_lock_nested()
836 * is correctly annotated.
837 */
838 static __poll_t ep_item_poll(const struct epitem *epi, poll_table *pt,
839 int depth)
840 {
841 struct file *file = epi->ffd.file;
842 __poll_t res;
843
844 pt->_key = epi->event.events;
845 if (!is_file_epoll(file))
846 res = vfs_poll(file, pt);
847 else
848 res = __ep_eventpoll_poll(file, pt, depth);
849 return res & epi->event.events;
850 }
851
852 static __poll_t ep_eventpoll_poll(struct file *file, poll_table *wait)
853 {
854 return __ep_eventpoll_poll(file, wait, 0);
855 }
856
857 #ifdef CONFIG_PROC_FS
858 static void ep_show_fdinfo(struct seq_file *m, struct file *f)
859 {
860 struct eventpoll *ep = f->private_data;
861 struct rb_node *rbp;
862
863 mutex_lock(&ep->mtx);
864 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
865 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
866 struct inode *inode = file_inode(epi->ffd.file);
867
868 seq_printf(m, "tfd: %8d events: %8x data: %16llx "
869 " pos:%lli ino:%lx sdev:%x\n",
870 epi->ffd.fd, epi->event.events,
871 (long long)epi->event.data,
872 (long long)epi->ffd.file->f_pos,
873 inode->i_ino, inode->i_sb->s_dev);
874 if (seq_has_overflowed(m))
875 break;
876 }
877 mutex_unlock(&ep->mtx);
878 }
879 #endif
880
881 /* File callbacks that implement the eventpoll file behaviour */
882 static const struct file_operations eventpoll_fops = {
883 #ifdef CONFIG_PROC_FS
884 .show_fdinfo = ep_show_fdinfo,
885 #endif
886 .release = ep_eventpoll_release,
887 .poll = ep_eventpoll_poll,
888 .llseek = noop_llseek,
889 };
890
891 /*
892 * This is called from eventpoll_release() to unlink files from the eventpoll
893 * interface. We need to have this facility to cleanup correctly files that are
894 * closed without being removed from the eventpoll interface.
895 */
896 void eventpoll_release_file(struct file *file)
897 {
898 struct eventpoll *ep;
899 struct epitem *epi;
900 struct hlist_node *next;
901
902 /*
903 * We don't want to get "file->f_lock" because it is not
904 * necessary. It is not necessary because we're in the "struct file"
905 * cleanup path, and this means that no one is using this file anymore.
906 * So, for example, epoll_ctl() cannot hit here since if we reach this
907 * point, the file counter already went to zero and fget() would fail.
908 * The only hit might come from ep_free() but by holding the mutex
909 * will correctly serialize the operation. We do need to acquire
910 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
911 * from anywhere but ep_free().
912 *
913 * Besides, ep_remove() acquires the lock, so we can't hold it here.
914 */
915 mutex_lock(&epmutex);
916 if (unlikely(!file->f_ep)) {
917 mutex_unlock(&epmutex);
918 return;
919 }
920 hlist_for_each_entry_safe(epi, next, file->f_ep, fllink) {
921 ep = epi->ep;
922 mutex_lock_nested(&ep->mtx, 0);
923 ep_remove(ep, epi);
924 mutex_unlock(&ep->mtx);
925 }
926 mutex_unlock(&epmutex);
927 }
928
929 static int ep_alloc(struct eventpoll **pep)
930 {
931 int error;
932 struct user_struct *user;
933 struct eventpoll *ep;
934
935 user = get_current_user();
936 error = -ENOMEM;
937 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
938 if (unlikely(!ep))
939 goto free_uid;
940
941 mutex_init(&ep->mtx);
942 rwlock_init(&ep->lock);
943 init_waitqueue_head(&ep->wq);
944 init_waitqueue_head(&ep->poll_wait);
945 INIT_LIST_HEAD(&ep->rdllist);
946 ep->rbr = RB_ROOT_CACHED;
947 ep->ovflist = EP_UNACTIVE_PTR;
948 ep->user = user;
949
950 *pep = ep;
951
952 return 0;
953
954 free_uid:
955 free_uid(user);
956 return error;
957 }
958
959 /*
960 * Search the file inside the eventpoll tree. The RB tree operations
961 * are protected by the "mtx" mutex, and ep_find() must be called with
962 * "mtx" held.
963 */
964 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
965 {
966 int kcmp;
967 struct rb_node *rbp;
968 struct epitem *epi, *epir = NULL;
969 struct epoll_filefd ffd;
970
971 ep_set_ffd(&ffd, file, fd);
972 for (rbp = ep->rbr.rb_root.rb_node; rbp; ) {
973 epi = rb_entry(rbp, struct epitem, rbn);
974 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
975 if (kcmp > 0)
976 rbp = rbp->rb_right;
977 else if (kcmp < 0)
978 rbp = rbp->rb_left;
979 else {
980 epir = epi;
981 break;
982 }
983 }
984
985 return epir;
986 }
987
988 #ifdef CONFIG_KCMP
989 static struct epitem *ep_find_tfd(struct eventpoll *ep, int tfd, unsigned long toff)
990 {
991 struct rb_node *rbp;
992 struct epitem *epi;
993
994 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
995 epi = rb_entry(rbp, struct epitem, rbn);
996 if (epi->ffd.fd == tfd) {
997 if (toff == 0)
998 return epi;
999 else
1000 toff--;
1001 }
1002 cond_resched();
1003 }
1004
1005 return NULL;
1006 }
1007
1008 struct file *get_epoll_tfile_raw_ptr(struct file *file, int tfd,
1009 unsigned long toff)
1010 {
1011 struct file *file_raw;
1012 struct eventpoll *ep;
1013 struct epitem *epi;
1014
1015 if (!is_file_epoll(file))
1016 return ERR_PTR(-EINVAL);
1017
1018 ep = file->private_data;
1019
1020 mutex_lock(&ep->mtx);
1021 epi = ep_find_tfd(ep, tfd, toff);
1022 if (epi)
1023 file_raw = epi->ffd.file;
1024 else
1025 file_raw = ERR_PTR(-ENOENT);
1026 mutex_unlock(&ep->mtx);
1027
1028 return file_raw;
1029 }
1030 #endif /* CONFIG_KCMP */
1031
1032 /*
1033 * Adds a new entry to the tail of the list in a lockless way, i.e.
1034 * multiple CPUs are allowed to call this function concurrently.
1035 *
1036 * Beware: it is necessary to prevent any other modifications of the
1037 * existing list until all changes are completed, in other words
1038 * concurrent list_add_tail_lockless() calls should be protected
1039 * with a read lock, where write lock acts as a barrier which
1040 * makes sure all list_add_tail_lockless() calls are fully
1041 * completed.
1042 *
1043 * Also an element can be locklessly added to the list only in one
1044 * direction i.e. either to the tail or to the head, otherwise
1045 * concurrent access will corrupt the list.
1046 *
1047 * Return: %false if element has been already added to the list, %true
1048 * otherwise.
1049 */
1050 static inline bool list_add_tail_lockless(struct list_head *new,
1051 struct list_head *head)
1052 {
1053 struct list_head *prev;
1054
1055 /*
1056 * This is simple 'new->next = head' operation, but cmpxchg()
1057 * is used in order to detect that same element has been just
1058 * added to the list from another CPU: the winner observes
1059 * new->next == new.
1060 */
1061 if (cmpxchg(&new->next, new, head) != new)
1062 return false;
1063
1064 /*
1065 * Initially ->next of a new element must be updated with the head
1066 * (we are inserting to the tail) and only then pointers are atomically
1067 * exchanged. XCHG guarantees memory ordering, thus ->next should be
1068 * updated before pointers are actually swapped and pointers are
1069 * swapped before prev->next is updated.
1070 */
1071
1072 prev = xchg(&head->prev, new);
1073
1074 /*
1075 * It is safe to modify prev->next and new->prev, because a new element
1076 * is added only to the tail and new->next is updated before XCHG.
1077 */
1078
1079 prev->next = new;
1080 new->prev = prev;
1081
1082 return true;
1083 }
1084
1085 /*
1086 * Chains a new epi entry to the tail of the ep->ovflist in a lockless way,
1087 * i.e. multiple CPUs are allowed to call this function concurrently.
1088 *
1089 * Return: %false if epi element has been already chained, %true otherwise.
1090 */
1091 static inline bool chain_epi_lockless(struct epitem *epi)
1092 {
1093 struct eventpoll *ep = epi->ep;
1094
1095 /* Fast preliminary check */
1096 if (epi->next != EP_UNACTIVE_PTR)
1097 return false;
1098
1099 /* Check that the same epi has not been just chained from another CPU */
1100 if (cmpxchg(&epi->next, EP_UNACTIVE_PTR, NULL) != EP_UNACTIVE_PTR)
1101 return false;
1102
1103 /* Atomically exchange tail */
1104 epi->next = xchg(&ep->ovflist, epi);
1105
1106 return true;
1107 }
1108
1109 /*
1110 * This is the callback that is passed to the wait queue wakeup
1111 * mechanism. It is called by the stored file descriptors when they
1112 * have events to report.
1113 *
1114 * This callback takes a read lock in order not to contend with concurrent
1115 * events from another file descriptor, thus all modifications to ->rdllist
1116 * or ->ovflist are lockless. Read lock is paired with the write lock from
1117 * ep_scan_ready_list(), which stops all list modifications and guarantees
1118 * that lists state is seen correctly.
1119 *
1120 * Another thing worth to mention is that ep_poll_callback() can be called
1121 * concurrently for the same @epi from different CPUs if poll table was inited
1122 * with several wait queues entries. Plural wakeup from different CPUs of a
1123 * single wait queue is serialized by wq.lock, but the case when multiple wait
1124 * queues are used should be detected accordingly. This is detected using
1125 * cmpxchg() operation.
1126 */
1127 static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1128 {
1129 int pwake = 0;
1130 struct epitem *epi = ep_item_from_wait(wait);
1131 struct eventpoll *ep = epi->ep;
1132 __poll_t pollflags = key_to_poll(key);
1133 unsigned long flags;
1134 int ewake = 0;
1135
1136 read_lock_irqsave(&ep->lock, flags);
1137
1138 ep_set_busy_poll_napi_id(epi);
1139
1140 /*
1141 * If the event mask does not contain any poll(2) event, we consider the
1142 * descriptor to be disabled. This condition is likely the effect of the
1143 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1144 * until the next EPOLL_CTL_MOD will be issued.
1145 */
1146 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1147 goto out_unlock;
1148
1149 /*
1150 * Check the events coming with the callback. At this stage, not
1151 * every device reports the events in the "key" parameter of the
1152 * callback. We need to be able to handle both cases here, hence the
1153 * test for "key" != NULL before the event match test.
1154 */
1155 if (pollflags && !(pollflags & epi->event.events))
1156 goto out_unlock;
1157
1158 /*
1159 * If we are transferring events to userspace, we can hold no locks
1160 * (because we're accessing user memory, and because of linux f_op->poll()
1161 * semantics). All the events that happen during that period of time are
1162 * chained in ep->ovflist and requeued later on.
1163 */
1164 if (READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR) {
1165 if (chain_epi_lockless(epi))
1166 ep_pm_stay_awake_rcu(epi);
1167 } else if (!ep_is_linked(epi)) {
1168 /* In the usual case, add event to ready list. */
1169 if (list_add_tail_lockless(&epi->rdllink, &ep->rdllist))
1170 ep_pm_stay_awake_rcu(epi);
1171 }
1172
1173 /*
1174 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1175 * wait list.
1176 */
1177 if (waitqueue_active(&ep->wq)) {
1178 if ((epi->event.events & EPOLLEXCLUSIVE) &&
1179 !(pollflags & POLLFREE)) {
1180 switch (pollflags & EPOLLINOUT_BITS) {
1181 case EPOLLIN:
1182 if (epi->event.events & EPOLLIN)
1183 ewake = 1;
1184 break;
1185 case EPOLLOUT:
1186 if (epi->event.events & EPOLLOUT)
1187 ewake = 1;
1188 break;
1189 case 0:
1190 ewake = 1;
1191 break;
1192 }
1193 }
1194 wake_up(&ep->wq);
1195 }
1196 if (waitqueue_active(&ep->poll_wait))
1197 pwake++;
1198
1199 out_unlock:
1200 read_unlock_irqrestore(&ep->lock, flags);
1201
1202 /* We have to call this outside the lock */
1203 if (pwake)
1204 ep_poll_safewake(ep, epi);
1205
1206 if (!(epi->event.events & EPOLLEXCLUSIVE))
1207 ewake = 1;
1208
1209 if (pollflags & POLLFREE) {
1210 /*
1211 * If we race with ep_remove_wait_queue() it can miss
1212 * ->whead = NULL and do another remove_wait_queue() after
1213 * us, so we can't use __remove_wait_queue().
1214 */
1215 list_del_init(&wait->entry);
1216 /*
1217 * ->whead != NULL protects us from the race with ep_free()
1218 * or ep_remove(), ep_remove_wait_queue() takes whead->lock
1219 * held by the caller. Once we nullify it, nothing protects
1220 * ep/epi or even wait.
1221 */
1222 smp_store_release(&ep_pwq_from_wait(wait)->whead, NULL);
1223 }
1224
1225 return ewake;
1226 }
1227
1228 /*
1229 * This is the callback that is used to add our wait queue to the
1230 * target file wakeup lists.
1231 */
1232 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1233 poll_table *pt)
1234 {
1235 struct ep_pqueue *epq = container_of(pt, struct ep_pqueue, pt);
1236 struct epitem *epi = epq->epi;
1237 struct eppoll_entry *pwq;
1238
1239 if (unlikely(!epi)) // an earlier allocation has failed
1240 return;
1241
1242 pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL);
1243 if (unlikely(!pwq)) {
1244 epq->epi = NULL;
1245 return;
1246 }
1247
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 pwq->next = epi->pwqlist;
1256 epi->pwqlist = pwq;
1257 }
1258
1259 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1260 {
1261 int kcmp;
1262 struct rb_node **p = &ep->rbr.rb_root.rb_node, *parent = NULL;
1263 struct epitem *epic;
1264 bool leftmost = true;
1265
1266 while (*p) {
1267 parent = *p;
1268 epic = rb_entry(parent, struct epitem, rbn);
1269 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1270 if (kcmp > 0) {
1271 p = &parent->rb_right;
1272 leftmost = false;
1273 } else
1274 p = &parent->rb_left;
1275 }
1276 rb_link_node(&epi->rbn, parent, p);
1277 rb_insert_color_cached(&epi->rbn, &ep->rbr, leftmost);
1278 }
1279
1280
1281
1282 #define PATH_ARR_SIZE 5
1283 /*
1284 * These are the number paths of length 1 to 5, that we are allowing to emanate
1285 * from a single file of interest. For example, we allow 1000 paths of length
1286 * 1, to emanate from each file of interest. This essentially represents the
1287 * potential wakeup paths, which need to be limited in order to avoid massive
1288 * uncontrolled wakeup storms. The common use case should be a single ep which
1289 * is connected to n file sources. In this case each file source has 1 path
1290 * of length 1. Thus, the numbers below should be more than sufficient. These
1291 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1292 * and delete can't add additional paths. Protected by the epmutex.
1293 */
1294 static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1295 static int path_count[PATH_ARR_SIZE];
1296
1297 static int path_count_inc(int nests)
1298 {
1299 /* Allow an arbitrary number of depth 1 paths */
1300 if (nests == 0)
1301 return 0;
1302
1303 if (++path_count[nests] > path_limits[nests])
1304 return -1;
1305 return 0;
1306 }
1307
1308 static void path_count_init(void)
1309 {
1310 int i;
1311
1312 for (i = 0; i < PATH_ARR_SIZE; i++)
1313 path_count[i] = 0;
1314 }
1315
1316 static int reverse_path_check_proc(struct hlist_head *refs, int depth)
1317 {
1318 int error = 0;
1319 struct epitem *epi;
1320
1321 if (depth > EP_MAX_NESTS) /* too deep nesting */
1322 return -1;
1323
1324 /* CTL_DEL can remove links here, but that can't increase our count */
1325 hlist_for_each_entry_rcu(epi, refs, fllink) {
1326 struct hlist_head *refs = &epi->ep->refs;
1327 if (hlist_empty(refs))
1328 error = path_count_inc(depth);
1329 else
1330 error = reverse_path_check_proc(refs, depth + 1);
1331 if (error != 0)
1332 break;
1333 }
1334 return error;
1335 }
1336
1337 /**
1338 * reverse_path_check - The tfile_check_list is list of epitem_head, which have
1339 * links that are proposed to be newly added. We need to
1340 * make sure that those added links don't add too many
1341 * paths such that we will spend all our time waking up
1342 * eventpoll objects.
1343 *
1344 * Return: %zero if the proposed links don't create too many paths,
1345 * %-1 otherwise.
1346 */
1347 static int reverse_path_check(void)
1348 {
1349 struct epitems_head *p;
1350
1351 for (p = tfile_check_list; p != EP_UNACTIVE_PTR; p = p->next) {
1352 int error;
1353 path_count_init();
1354 rcu_read_lock();
1355 error = reverse_path_check_proc(&p->epitems, 0);
1356 rcu_read_unlock();
1357 if (error)
1358 return error;
1359 }
1360 return 0;
1361 }
1362
1363 static int ep_create_wakeup_source(struct epitem *epi)
1364 {
1365 struct name_snapshot n;
1366 struct wakeup_source *ws;
1367
1368 if (!epi->ep->ws) {
1369 epi->ep->ws = wakeup_source_register(NULL, "eventpoll");
1370 if (!epi->ep->ws)
1371 return -ENOMEM;
1372 }
1373
1374 take_dentry_name_snapshot(&n, epi->ffd.file->f_path.dentry);
1375 ws = wakeup_source_register(NULL, n.name.name);
1376 release_dentry_name_snapshot(&n);
1377
1378 if (!ws)
1379 return -ENOMEM;
1380 rcu_assign_pointer(epi->ws, ws);
1381
1382 return 0;
1383 }
1384
1385 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1386 static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1387 {
1388 struct wakeup_source *ws = ep_wakeup_source(epi);
1389
1390 RCU_INIT_POINTER(epi->ws, NULL);
1391
1392 /*
1393 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1394 * used internally by wakeup_source_remove, too (called by
1395 * wakeup_source_unregister), so we cannot use call_rcu
1396 */
1397 synchronize_rcu();
1398 wakeup_source_unregister(ws);
1399 }
1400
1401 static int attach_epitem(struct file *file, struct epitem *epi)
1402 {
1403 struct epitems_head *to_free = NULL;
1404 struct hlist_head *head = NULL;
1405 struct eventpoll *ep = NULL;
1406
1407 if (is_file_epoll(file))
1408 ep = file->private_data;
1409
1410 if (ep) {
1411 head = &ep->refs;
1412 } else if (!READ_ONCE(file->f_ep)) {
1413 allocate:
1414 to_free = kmem_cache_zalloc(ephead_cache, GFP_KERNEL);
1415 if (!to_free)
1416 return -ENOMEM;
1417 head = &to_free->epitems;
1418 }
1419 spin_lock(&file->f_lock);
1420 if (!file->f_ep) {
1421 if (unlikely(!head)) {
1422 spin_unlock(&file->f_lock);
1423 goto allocate;
1424 }
1425 file->f_ep = head;
1426 to_free = NULL;
1427 }
1428 hlist_add_head_rcu(&epi->fllink, file->f_ep);
1429 spin_unlock(&file->f_lock);
1430 free_ephead(to_free);
1431 return 0;
1432 }
1433
1434 /*
1435 * Must be called with "mtx" held.
1436 */
1437 static int ep_insert(struct eventpoll *ep, const struct epoll_event *event,
1438 struct file *tfile, int fd, int full_check)
1439 {
1440 int error, pwake = 0;
1441 __poll_t revents;
1442 long user_watches;
1443 struct epitem *epi;
1444 struct ep_pqueue epq;
1445 struct eventpoll *tep = NULL;
1446
1447 if (is_file_epoll(tfile))
1448 tep = tfile->private_data;
1449
1450 lockdep_assert_irqs_enabled();
1451
1452 user_watches = atomic_long_read(&ep->user->epoll_watches);
1453 if (unlikely(user_watches >= max_user_watches))
1454 return -ENOSPC;
1455 if (!(epi = kmem_cache_zalloc(epi_cache, GFP_KERNEL)))
1456 return -ENOMEM;
1457
1458 /* Item initialization follow here ... */
1459 INIT_LIST_HEAD(&epi->rdllink);
1460 epi->ep = ep;
1461 ep_set_ffd(&epi->ffd, tfile, fd);
1462 epi->event = *event;
1463 epi->next = EP_UNACTIVE_PTR;
1464
1465 if (tep)
1466 mutex_lock_nested(&tep->mtx, 1);
1467 /* Add the current item to the list of active epoll hook for this file */
1468 if (unlikely(attach_epitem(tfile, epi) < 0)) {
1469 kmem_cache_free(epi_cache, epi);
1470 if (tep)
1471 mutex_unlock(&tep->mtx);
1472 return -ENOMEM;
1473 }
1474
1475 if (full_check && !tep)
1476 list_file(tfile);
1477
1478 atomic_long_inc(&ep->user->epoll_watches);
1479
1480 /*
1481 * Add the current item to the RB tree. All RB tree operations are
1482 * protected by "mtx", and ep_insert() is called with "mtx" held.
1483 */
1484 ep_rbtree_insert(ep, epi);
1485 if (tep)
1486 mutex_unlock(&tep->mtx);
1487
1488 /* now check if we've created too many backpaths */
1489 if (unlikely(full_check && reverse_path_check())) {
1490 ep_remove(ep, epi);
1491 return -EINVAL;
1492 }
1493
1494 if (epi->event.events & EPOLLWAKEUP) {
1495 error = ep_create_wakeup_source(epi);
1496 if (error) {
1497 ep_remove(ep, epi);
1498 return error;
1499 }
1500 }
1501
1502 /* Initialize the poll table using the queue callback */
1503 epq.epi = epi;
1504 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1505
1506 /*
1507 * Attach the item to the poll hooks and get current event bits.
1508 * We can safely use the file* here because its usage count has
1509 * been increased by the caller of this function. Note that after
1510 * this operation completes, the poll callback can start hitting
1511 * the new item.
1512 */
1513 revents = ep_item_poll(epi, &epq.pt, 1);
1514
1515 /*
1516 * We have to check if something went wrong during the poll wait queue
1517 * install process. Namely an allocation for a wait queue failed due
1518 * high memory pressure.
1519 */
1520 if (unlikely(!epq.epi)) {
1521 ep_remove(ep, epi);
1522 return -ENOMEM;
1523 }
1524
1525 /* We have to drop the new item inside our item list to keep track of it */
1526 write_lock_irq(&ep->lock);
1527
1528 /* record NAPI ID of new item if present */
1529 ep_set_busy_poll_napi_id(epi);
1530
1531 /* If the file is already "ready" we drop it inside the ready list */
1532 if (revents && !ep_is_linked(epi)) {
1533 list_add_tail(&epi->rdllink, &ep->rdllist);
1534 ep_pm_stay_awake(epi);
1535
1536 /* Notify waiting tasks that events are available */
1537 if (waitqueue_active(&ep->wq))
1538 wake_up(&ep->wq);
1539 if (waitqueue_active(&ep->poll_wait))
1540 pwake++;
1541 }
1542
1543 write_unlock_irq(&ep->lock);
1544
1545 /* We have to call this outside the lock */
1546 if (pwake)
1547 ep_poll_safewake(ep, NULL);
1548
1549 return 0;
1550 }
1551
1552 /*
1553 * Modify the interest event mask by dropping an event if the new mask
1554 * has a match in the current file status. Must be called with "mtx" held.
1555 */
1556 static int ep_modify(struct eventpoll *ep, struct epitem *epi,
1557 const struct epoll_event *event)
1558 {
1559 int pwake = 0;
1560 poll_table pt;
1561
1562 lockdep_assert_irqs_enabled();
1563
1564 init_poll_funcptr(&pt, NULL);
1565
1566 /*
1567 * Set the new event interest mask before calling f_op->poll();
1568 * otherwise we might miss an event that happens between the
1569 * f_op->poll() call and the new event set registering.
1570 */
1571 epi->event.events = event->events; /* need barrier below */
1572 epi->event.data = event->data; /* protected by mtx */
1573 if (epi->event.events & EPOLLWAKEUP) {
1574 if (!ep_has_wakeup_source(epi))
1575 ep_create_wakeup_source(epi);
1576 } else if (ep_has_wakeup_source(epi)) {
1577 ep_destroy_wakeup_source(epi);
1578 }
1579
1580 /*
1581 * The following barrier has two effects:
1582 *
1583 * 1) Flush epi changes above to other CPUs. This ensures
1584 * we do not miss events from ep_poll_callback if an
1585 * event occurs immediately after we call f_op->poll().
1586 * We need this because we did not take ep->lock while
1587 * changing epi above (but ep_poll_callback does take
1588 * ep->lock).
1589 *
1590 * 2) We also need to ensure we do not miss _past_ events
1591 * when calling f_op->poll(). This barrier also
1592 * pairs with the barrier in wq_has_sleeper (see
1593 * comments for wq_has_sleeper).
1594 *
1595 * This barrier will now guarantee ep_poll_callback or f_op->poll
1596 * (or both) will notice the readiness of an item.
1597 */
1598 smp_mb();
1599
1600 /*
1601 * Get current event bits. We can safely use the file* here because
1602 * its usage count has been increased by the caller of this function.
1603 * If the item is "hot" and it is not registered inside the ready
1604 * list, push it inside.
1605 */
1606 if (ep_item_poll(epi, &pt, 1)) {
1607 write_lock_irq(&ep->lock);
1608 if (!ep_is_linked(epi)) {
1609 list_add_tail(&epi->rdllink, &ep->rdllist);
1610 ep_pm_stay_awake(epi);
1611
1612 /* Notify waiting tasks that events are available */
1613 if (waitqueue_active(&ep->wq))
1614 wake_up(&ep->wq);
1615 if (waitqueue_active(&ep->poll_wait))
1616 pwake++;
1617 }
1618 write_unlock_irq(&ep->lock);
1619 }
1620
1621 /* We have to call this outside the lock */
1622 if (pwake)
1623 ep_poll_safewake(ep, NULL);
1624
1625 return 0;
1626 }
1627
1628 static int ep_send_events(struct eventpoll *ep,
1629 struct epoll_event __user *events, int maxevents)
1630 {
1631 struct epitem *epi, *tmp;
1632 LIST_HEAD(txlist);
1633 poll_table pt;
1634 int res = 0;
1635
1636 /*
1637 * Always short-circuit for fatal signals to allow threads to make a
1638 * timely exit without the chance of finding more events available and
1639 * fetching repeatedly.
1640 */
1641 if (fatal_signal_pending(current))
1642 return -EINTR;
1643
1644 init_poll_funcptr(&pt, NULL);
1645
1646 mutex_lock(&ep->mtx);
1647 ep_start_scan(ep, &txlist);
1648
1649 /*
1650 * We can loop without lock because we are passed a task private list.
1651 * Items cannot vanish during the loop we are holding ep->mtx.
1652 */
1653 list_for_each_entry_safe(epi, tmp, &txlist, rdllink) {
1654 struct wakeup_source *ws;
1655 __poll_t revents;
1656
1657 if (res >= maxevents)
1658 break;
1659
1660 /*
1661 * Activate ep->ws before deactivating epi->ws to prevent
1662 * triggering auto-suspend here (in case we reactive epi->ws
1663 * below).
1664 *
1665 * This could be rearranged to delay the deactivation of epi->ws
1666 * instead, but then epi->ws would temporarily be out of sync
1667 * with ep_is_linked().
1668 */
1669 ws = ep_wakeup_source(epi);
1670 if (ws) {
1671 if (ws->active)
1672 __pm_stay_awake(ep->ws);
1673 __pm_relax(ws);
1674 }
1675
1676 list_del_init(&epi->rdllink);
1677
1678 /*
1679 * If the event mask intersect the caller-requested one,
1680 * deliver the event to userspace. Again, we are holding ep->mtx,
1681 * so no operations coming from userspace can change the item.
1682 */
1683 revents = ep_item_poll(epi, &pt, 1);
1684 if (!revents)
1685 continue;
1686
1687 if (__put_user(revents, &events->events) ||
1688 __put_user(epi->event.data, &events->data)) {
1689 list_add(&epi->rdllink, &txlist);
1690 ep_pm_stay_awake(epi);
1691 if (!res)
1692 res = -EFAULT;
1693 break;
1694 }
1695 res++;
1696 events++;
1697 if (epi->event.events & EPOLLONESHOT)
1698 epi->event.events &= EP_PRIVATE_BITS;
1699 else if (!(epi->event.events & EPOLLET)) {
1700 /*
1701 * If this file has been added with Level
1702 * Trigger mode, we need to insert back inside
1703 * the ready list, so that the next call to
1704 * epoll_wait() will check again the events
1705 * availability. At this point, no one can insert
1706 * into ep->rdllist besides us. The epoll_ctl()
1707 * callers are locked out by
1708 * ep_scan_ready_list() holding "mtx" and the
1709 * poll callback will queue them in ep->ovflist.
1710 */
1711 list_add_tail(&epi->rdllink, &ep->rdllist);
1712 ep_pm_stay_awake(epi);
1713 }
1714 }
1715 ep_done_scan(ep, &txlist);
1716 mutex_unlock(&ep->mtx);
1717
1718 return res;
1719 }
1720
1721 static struct timespec64 *ep_timeout_to_timespec(struct timespec64 *to, long ms)
1722 {
1723 struct timespec64 now;
1724
1725 if (ms < 0)
1726 return NULL;
1727
1728 if (!ms) {
1729 to->tv_sec = 0;
1730 to->tv_nsec = 0;
1731 return to;
1732 }
1733
1734 to->tv_sec = ms / MSEC_PER_SEC;
1735 to->tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC);
1736
1737 ktime_get_ts64(&now);
1738 *to = timespec64_add_safe(now, *to);
1739 return to;
1740 }
1741
1742 /**
1743 * ep_poll - Retrieves ready events, and delivers them to the caller-supplied
1744 * event buffer.
1745 *
1746 * @ep: Pointer to the eventpoll context.
1747 * @events: Pointer to the userspace buffer where the ready events should be
1748 * stored.
1749 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1750 * @timeout: Maximum timeout for the ready events fetch operation, in
1751 * timespec. If the timeout is zero, the function will not block,
1752 * while if the @timeout ptr is NULL, the function will block
1753 * until at least one event has been retrieved (or an error
1754 * occurred).
1755 *
1756 * Return: the number of ready events which have been fetched, or an
1757 * error code, in case of error.
1758 */
1759 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1760 int maxevents, struct timespec64 *timeout)
1761 {
1762 int res, eavail, timed_out = 0;
1763 u64 slack = 0;
1764 wait_queue_entry_t wait;
1765 ktime_t expires, *to = NULL;
1766
1767 lockdep_assert_irqs_enabled();
1768
1769 if (timeout && (timeout->tv_sec | timeout->tv_nsec)) {
1770 slack = select_estimate_accuracy(timeout);
1771 to = &expires;
1772 *to = timespec64_to_ktime(*timeout);
1773 } else if (timeout) {
1774 /*
1775 * Avoid the unnecessary trip to the wait queue loop, if the
1776 * caller specified a non blocking operation.
1777 */
1778 timed_out = 1;
1779 }
1780
1781 /*
1782 * This call is racy: We may or may not see events that are being added
1783 * to the ready list under the lock (e.g., in IRQ callbacks). For cases
1784 * with a non-zero timeout, this thread will check the ready list under
1785 * lock and will add to the wait queue. For cases with a zero
1786 * timeout, the user by definition should not care and will have to
1787 * recheck again.
1788 */
1789 eavail = ep_events_available(ep);
1790
1791 while (1) {
1792 if (eavail) {
1793 /*
1794 * Try to transfer events to user space. In case we get
1795 * 0 events and there's still timeout left over, we go
1796 * trying again in search of more luck.
1797 */
1798 res = ep_send_events(ep, events, maxevents);
1799 if (res)
1800 return res;
1801 }
1802
1803 if (timed_out)
1804 return 0;
1805
1806 eavail = ep_busy_loop(ep, timed_out);
1807 if (eavail)
1808 continue;
1809
1810 if (signal_pending(current))
1811 return -EINTR;
1812
1813 /*
1814 * Internally init_wait() uses autoremove_wake_function(),
1815 * thus wait entry is removed from the wait queue on each
1816 * wakeup. Why it is important? In case of several waiters
1817 * each new wakeup will hit the next waiter, giving it the
1818 * chance to harvest new event. Otherwise wakeup can be
1819 * lost. This is also good performance-wise, because on
1820 * normal wakeup path no need to call __remove_wait_queue()
1821 * explicitly, thus ep->lock is not taken, which halts the
1822 * event delivery.
1823 */
1824 init_wait(&wait);
1825
1826 write_lock_irq(&ep->lock);
1827 /*
1828 * Barrierless variant, waitqueue_active() is called under
1829 * the same lock on wakeup ep_poll_callback() side, so it
1830 * is safe to avoid an explicit barrier.
1831 */
1832 __set_current_state(TASK_INTERRUPTIBLE);
1833
1834 /*
1835 * Do the final check under the lock. ep_scan_ready_list()
1836 * plays with two lists (->rdllist and ->ovflist) and there
1837 * is always a race when both lists are empty for short
1838 * period of time although events are pending, so lock is
1839 * important.
1840 */
1841 eavail = ep_events_available(ep);
1842 if (!eavail)
1843 __add_wait_queue_exclusive(&ep->wq, &wait);
1844
1845 write_unlock_irq(&ep->lock);
1846
1847 if (!eavail)
1848 timed_out = !schedule_hrtimeout_range(to, slack,
1849 HRTIMER_MODE_ABS);
1850 __set_current_state(TASK_RUNNING);
1851
1852 /*
1853 * We were woken up, thus go and try to harvest some events.
1854 * If timed out and still on the wait queue, recheck eavail
1855 * carefully under lock, below.
1856 */
1857 eavail = 1;
1858
1859 if (!list_empty_careful(&wait.entry)) {
1860 write_lock_irq(&ep->lock);
1861 /*
1862 * If the thread timed out and is not on the wait queue,
1863 * it means that the thread was woken up after its
1864 * timeout expired before it could reacquire the lock.
1865 * Thus, when wait.entry is empty, it needs to harvest
1866 * events.
1867 */
1868 if (timed_out)
1869 eavail = list_empty(&wait.entry);
1870 __remove_wait_queue(&ep->wq, &wait);
1871 write_unlock_irq(&ep->lock);
1872 }
1873 }
1874 }
1875
1876 /**
1877 * ep_loop_check_proc - verify that adding an epoll file inside another
1878 * epoll structure does not violate the constraints, in
1879 * terms of closed loops, or too deep chains (which can
1880 * result in excessive stack usage).
1881 *
1882 * @ep: the &struct eventpoll to be currently checked.
1883 * @depth: Current depth of the path being checked.
1884 *
1885 * Return: %zero if adding the epoll @file inside current epoll
1886 * structure @ep does not violate the constraints, or %-1 otherwise.
1887 */
1888 static int ep_loop_check_proc(struct eventpoll *ep, int depth)
1889 {
1890 int error = 0;
1891 struct rb_node *rbp;
1892 struct epitem *epi;
1893
1894 mutex_lock_nested(&ep->mtx, depth + 1);
1895 ep->gen = loop_check_gen;
1896 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1897 epi = rb_entry(rbp, struct epitem, rbn);
1898 if (unlikely(is_file_epoll(epi->ffd.file))) {
1899 struct eventpoll *ep_tovisit;
1900 ep_tovisit = epi->ffd.file->private_data;
1901 if (ep_tovisit->gen == loop_check_gen)
1902 continue;
1903 if (ep_tovisit == inserting_into || depth > EP_MAX_NESTS)
1904 error = -1;
1905 else
1906 error = ep_loop_check_proc(ep_tovisit, depth + 1);
1907 if (error != 0)
1908 break;
1909 } else {
1910 /*
1911 * If we've reached a file that is not associated with
1912 * an ep, then we need to check if the newly added
1913 * links are going to add too many wakeup paths. We do
1914 * this by adding it to the tfile_check_list, if it's
1915 * not already there, and calling reverse_path_check()
1916 * during ep_insert().
1917 */
1918 list_file(epi->ffd.file);
1919 }
1920 }
1921 mutex_unlock(&ep->mtx);
1922
1923 return error;
1924 }
1925
1926 /**
1927 * ep_loop_check - Performs a check to verify that adding an epoll file (@to)
1928 * into another epoll file (represented by @ep) does not create
1929 * closed loops or too deep chains.
1930 *
1931 * @ep: Pointer to the epoll we are inserting into.
1932 * @to: Pointer to the epoll to be inserted.
1933 *
1934 * Return: %zero if adding the epoll @to inside the epoll @from
1935 * does not violate the constraints, or %-1 otherwise.
1936 */
1937 static int ep_loop_check(struct eventpoll *ep, struct eventpoll *to)
1938 {
1939 inserting_into = ep;
1940 return ep_loop_check_proc(to, 0);
1941 }
1942
1943 static void clear_tfile_check_list(void)
1944 {
1945 rcu_read_lock();
1946 while (tfile_check_list != EP_UNACTIVE_PTR) {
1947 struct epitems_head *head = tfile_check_list;
1948 tfile_check_list = head->next;
1949 unlist_file(head);
1950 }
1951 rcu_read_unlock();
1952 }
1953
1954 /*
1955 * Open an eventpoll file descriptor.
1956 */
1957 static int do_epoll_create(int flags)
1958 {
1959 int error, fd;
1960 struct eventpoll *ep = NULL;
1961 struct file *file;
1962
1963 /* Check the EPOLL_* constant for consistency. */
1964 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1965
1966 if (flags & ~EPOLL_CLOEXEC)
1967 return -EINVAL;
1968 /*
1969 * Create the internal data structure ("struct eventpoll").
1970 */
1971 error = ep_alloc(&ep);
1972 if (error < 0)
1973 return error;
1974 /*
1975 * Creates all the items needed to setup an eventpoll file. That is,
1976 * a file structure and a free file descriptor.
1977 */
1978 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1979 if (fd < 0) {
1980 error = fd;
1981 goto out_free_ep;
1982 }
1983 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
1984 O_RDWR | (flags & O_CLOEXEC));
1985 if (IS_ERR(file)) {
1986 error = PTR_ERR(file);
1987 goto out_free_fd;
1988 }
1989 ep->file = file;
1990 fd_install(fd, file);
1991 return fd;
1992
1993 out_free_fd:
1994 put_unused_fd(fd);
1995 out_free_ep:
1996 ep_free(ep);
1997 return error;
1998 }
1999
2000 SYSCALL_DEFINE1(epoll_create1, int, flags)
2001 {
2002 return do_epoll_create(flags);
2003 }
2004
2005 SYSCALL_DEFINE1(epoll_create, int, size)
2006 {
2007 if (size <= 0)
2008 return -EINVAL;
2009
2010 return do_epoll_create(0);
2011 }
2012
2013 static inline int epoll_mutex_lock(struct mutex *mutex, int depth,
2014 bool nonblock)
2015 {
2016 if (!nonblock) {
2017 mutex_lock_nested(mutex, depth);
2018 return 0;
2019 }
2020 if (mutex_trylock(mutex))
2021 return 0;
2022 return -EAGAIN;
2023 }
2024
2025 int do_epoll_ctl(int epfd, int op, int fd, struct epoll_event *epds,
2026 bool nonblock)
2027 {
2028 int error;
2029 int full_check = 0;
2030 struct fd f, tf;
2031 struct eventpoll *ep;
2032 struct epitem *epi;
2033 struct eventpoll *tep = NULL;
2034
2035 error = -EBADF;
2036 f = fdget(epfd);
2037 if (!f.file)
2038 goto error_return;
2039
2040 /* Get the "struct file *" for the target file */
2041 tf = fdget(fd);
2042 if (!tf.file)
2043 goto error_fput;
2044
2045 /* The target file descriptor must support poll */
2046 error = -EPERM;
2047 if (!file_can_poll(tf.file))
2048 goto error_tgt_fput;
2049
2050 /* Check if EPOLLWAKEUP is allowed */
2051 if (ep_op_has_event(op))
2052 ep_take_care_of_epollwakeup(epds);
2053
2054 /*
2055 * We have to check that the file structure underneath the file descriptor
2056 * the user passed to us _is_ an eventpoll file. And also we do not permit
2057 * adding an epoll file descriptor inside itself.
2058 */
2059 error = -EINVAL;
2060 if (f.file == tf.file || !is_file_epoll(f.file))
2061 goto error_tgt_fput;
2062
2063 /*
2064 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
2065 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
2066 * Also, we do not currently supported nested exclusive wakeups.
2067 */
2068 if (ep_op_has_event(op) && (epds->events & EPOLLEXCLUSIVE)) {
2069 if (op == EPOLL_CTL_MOD)
2070 goto error_tgt_fput;
2071 if (op == EPOLL_CTL_ADD && (is_file_epoll(tf.file) ||
2072 (epds->events & ~EPOLLEXCLUSIVE_OK_BITS)))
2073 goto error_tgt_fput;
2074 }
2075
2076 /*
2077 * At this point it is safe to assume that the "private_data" contains
2078 * our own data structure.
2079 */
2080 ep = f.file->private_data;
2081
2082 /*
2083 * When we insert an epoll file descriptor inside another epoll file
2084 * descriptor, there is the chance of creating closed loops, which are
2085 * better be handled here, than in more critical paths. While we are
2086 * checking for loops we also determine the list of files reachable
2087 * and hang them on the tfile_check_list, so we can check that we
2088 * haven't created too many possible wakeup paths.
2089 *
2090 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
2091 * the epoll file descriptor is attaching directly to a wakeup source,
2092 * unless the epoll file descriptor is nested. The purpose of taking the
2093 * 'epmutex' on add is to prevent complex toplogies such as loops and
2094 * deep wakeup paths from forming in parallel through multiple
2095 * EPOLL_CTL_ADD operations.
2096 */
2097 error = epoll_mutex_lock(&ep->mtx, 0, nonblock);
2098 if (error)
2099 goto error_tgt_fput;
2100 if (op == EPOLL_CTL_ADD) {
2101 if (READ_ONCE(f.file->f_ep) || ep->gen == loop_check_gen ||
2102 is_file_epoll(tf.file)) {
2103 mutex_unlock(&ep->mtx);
2104 error = epoll_mutex_lock(&epmutex, 0, nonblock);
2105 if (error)
2106 goto error_tgt_fput;
2107 loop_check_gen++;
2108 full_check = 1;
2109 if (is_file_epoll(tf.file)) {
2110 tep = tf.file->private_data;
2111 error = -ELOOP;
2112 if (ep_loop_check(ep, tep) != 0)
2113 goto error_tgt_fput;
2114 }
2115 error = epoll_mutex_lock(&ep->mtx, 0, nonblock);
2116 if (error)
2117 goto error_tgt_fput;
2118 }
2119 }
2120
2121 /*
2122 * Try to lookup the file inside our RB tree. Since we grabbed "mtx"
2123 * above, we can be sure to be able to use the item looked up by
2124 * ep_find() till we release the mutex.
2125 */
2126 epi = ep_find(ep, tf.file, fd);
2127
2128 error = -EINVAL;
2129 switch (op) {
2130 case EPOLL_CTL_ADD:
2131 if (!epi) {
2132 epds->events |= EPOLLERR | EPOLLHUP;
2133 error = ep_insert(ep, epds, tf.file, fd, full_check);
2134 } else
2135 error = -EEXIST;
2136 break;
2137 case EPOLL_CTL_DEL:
2138 if (epi)
2139 error = ep_remove(ep, epi);
2140 else
2141 error = -ENOENT;
2142 break;
2143 case EPOLL_CTL_MOD:
2144 if (epi) {
2145 if (!(epi->event.events & EPOLLEXCLUSIVE)) {
2146 epds->events |= EPOLLERR | EPOLLHUP;
2147 error = ep_modify(ep, epi, epds);
2148 }
2149 } else
2150 error = -ENOENT;
2151 break;
2152 }
2153 mutex_unlock(&ep->mtx);
2154
2155 error_tgt_fput:
2156 if (full_check) {
2157 clear_tfile_check_list();
2158 loop_check_gen++;
2159 mutex_unlock(&epmutex);
2160 }
2161
2162 fdput(tf);
2163 error_fput:
2164 fdput(f);
2165 error_return:
2166
2167 return error;
2168 }
2169
2170 /*
2171 * The following function implements the controller interface for
2172 * the eventpoll file that enables the insertion/removal/change of
2173 * file descriptors inside the interest set.
2174 */
2175 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
2176 struct epoll_event __user *, event)
2177 {
2178 struct epoll_event epds;
2179
2180 if (ep_op_has_event(op) &&
2181 copy_from_user(&epds, event, sizeof(struct epoll_event)))
2182 return -EFAULT;
2183
2184 return do_epoll_ctl(epfd, op, fd, &epds, false);
2185 }
2186
2187 /*
2188 * Implement the event wait interface for the eventpoll file. It is the kernel
2189 * part of the user space epoll_wait(2).
2190 */
2191 static int do_epoll_wait(int epfd, struct epoll_event __user *events,
2192 int maxevents, struct timespec64 *to)
2193 {
2194 int error;
2195 struct fd f;
2196 struct eventpoll *ep;
2197
2198 /* The maximum number of event must be greater than zero */
2199 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
2200 return -EINVAL;
2201
2202 /* Verify that the area passed by the user is writeable */
2203 if (!access_ok(events, maxevents * sizeof(struct epoll_event)))
2204 return -EFAULT;
2205
2206 /* Get the "struct file *" for the eventpoll file */
2207 f = fdget(epfd);
2208 if (!f.file)
2209 return -EBADF;
2210
2211 /*
2212 * We have to check that the file structure underneath the fd
2213 * the user passed to us _is_ an eventpoll file.
2214 */
2215 error = -EINVAL;
2216 if (!is_file_epoll(f.file))
2217 goto error_fput;
2218
2219 /*
2220 * At this point it is safe to assume that the "private_data" contains
2221 * our own data structure.
2222 */
2223 ep = f.file->private_data;
2224
2225 /* Time to fish for events ... */
2226 error = ep_poll(ep, events, maxevents, to);
2227
2228 error_fput:
2229 fdput(f);
2230 return error;
2231 }
2232
2233 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
2234 int, maxevents, int, timeout)
2235 {
2236 struct timespec64 to;
2237
2238 return do_epoll_wait(epfd, events, maxevents,
2239 ep_timeout_to_timespec(&to, timeout));
2240 }
2241
2242 /*
2243 * Implement the event wait interface for the eventpoll file. It is the kernel
2244 * part of the user space epoll_pwait(2).
2245 */
2246 static int do_epoll_pwait(int epfd, struct epoll_event __user *events,
2247 int maxevents, struct timespec64 *to,
2248 const sigset_t __user *sigmask, size_t sigsetsize)
2249 {
2250 int error;
2251
2252 /*
2253 * If the caller wants a certain signal mask to be set during the wait,
2254 * we apply it here.
2255 */
2256 error = set_user_sigmask(sigmask, sigsetsize);
2257 if (error)
2258 return error;
2259
2260 error = do_epoll_wait(epfd, events, maxevents, to);
2261
2262 restore_saved_sigmask_unless(error == -EINTR);
2263
2264 return error;
2265 }
2266
2267 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2268 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2269 size_t, sigsetsize)
2270 {
2271 struct timespec64 to;
2272
2273 return do_epoll_pwait(epfd, events, maxevents,
2274 ep_timeout_to_timespec(&to, timeout),
2275 sigmask, sigsetsize);
2276 }
2277
2278 SYSCALL_DEFINE6(epoll_pwait2, int, epfd, struct epoll_event __user *, events,
2279 int, maxevents, const struct __kernel_timespec __user *, timeout,
2280 const sigset_t __user *, sigmask, size_t, sigsetsize)
2281 {
2282 struct timespec64 ts, *to = NULL;
2283
2284 if (timeout) {
2285 if (get_timespec64(&ts, timeout))
2286 return -EFAULT;
2287 to = &ts;
2288 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
2289 return -EINVAL;
2290 }
2291
2292 return do_epoll_pwait(epfd, events, maxevents, to,
2293 sigmask, sigsetsize);
2294 }
2295
2296 #ifdef CONFIG_COMPAT
2297 static int do_compat_epoll_pwait(int epfd, struct epoll_event __user *events,
2298 int maxevents, struct timespec64 *timeout,
2299 const compat_sigset_t __user *sigmask,
2300 compat_size_t sigsetsize)
2301 {
2302 long err;
2303
2304 /*
2305 * If the caller wants a certain signal mask to be set during the wait,
2306 * we apply it here.
2307 */
2308 err = set_compat_user_sigmask(sigmask, sigsetsize);
2309 if (err)
2310 return err;
2311
2312 err = do_epoll_wait(epfd, events, maxevents, timeout);
2313
2314 restore_saved_sigmask_unless(err == -EINTR);
2315
2316 return err;
2317 }
2318
2319 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2320 struct epoll_event __user *, events,
2321 int, maxevents, int, timeout,
2322 const compat_sigset_t __user *, sigmask,
2323 compat_size_t, sigsetsize)
2324 {
2325 struct timespec64 to;
2326
2327 return do_compat_epoll_pwait(epfd, events, maxevents,
2328 ep_timeout_to_timespec(&to, timeout),
2329 sigmask, sigsetsize);
2330 }
2331
2332 COMPAT_SYSCALL_DEFINE6(epoll_pwait2, int, epfd,
2333 struct epoll_event __user *, events,
2334 int, maxevents,
2335 const struct __kernel_timespec __user *, timeout,
2336 const compat_sigset_t __user *, sigmask,
2337 compat_size_t, sigsetsize)
2338 {
2339 struct timespec64 ts, *to = NULL;
2340
2341 if (timeout) {
2342 if (get_timespec64(&ts, timeout))
2343 return -EFAULT;
2344 to = &ts;
2345 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
2346 return -EINVAL;
2347 }
2348
2349 return do_compat_epoll_pwait(epfd, events, maxevents, to,
2350 sigmask, sigsetsize);
2351 }
2352
2353 #endif
2354
2355 static int __init eventpoll_init(void)
2356 {
2357 struct sysinfo si;
2358
2359 si_meminfo(&si);
2360 /*
2361 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2362 */
2363 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2364 EP_ITEM_COST;
2365 BUG_ON(max_user_watches < 0);
2366
2367 /*
2368 * We can have many thousands of epitems, so prevent this from
2369 * using an extra cache line on 64-bit (and smaller) CPUs
2370 */
2371 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2372
2373 /* Allocates slab cache used to allocate "struct epitem" items */
2374 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2375 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC|SLAB_ACCOUNT, NULL);
2376
2377 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2378 pwq_cache = kmem_cache_create("eventpoll_pwq",
2379 sizeof(struct eppoll_entry), 0, SLAB_PANIC|SLAB_ACCOUNT, NULL);
2380
2381 ephead_cache = kmem_cache_create("ep_head",
2382 sizeof(struct epitems_head), 0, SLAB_PANIC|SLAB_ACCOUNT, NULL);
2383
2384 return 0;
2385 }
2386 fs_initcall(eventpoll_init);