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 #include <trace/hooks/fs.h>
43
44 /*
45 * LOCKING:
46 * There are three level of locking required by epoll :
47 *
48 * 1) epmutex (mutex)
49 * 2) ep->mtx (mutex)
50 * 3) ep->lock (rwlock)
51 *
52 * The acquire order is the one listed above, from 1 to 3.
53 * We need a rwlock (ep->lock) because we manipulate objects
54 * from inside the poll callback, that might be triggered from
55 * a wake_up() that in turn might be called from IRQ context.
56 * So we can't sleep inside the poll callback and hence we need
57 * a spinlock. During the event transfer loop (from kernel to
58 * user space) we could end up sleeping due a copy_to_user(), so
59 * we need a lock that will allow us to sleep. This lock is a
60 * mutex (ep->mtx). It is acquired during the event transfer loop,
61 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
62 * The epmutex is acquired when inserting an epoll fd onto another epoll
63 * fd. We do this so that we walk the epoll tree and ensure that this
64 * insertion does not create a cycle of epoll file descriptors, which
65 * could lead to deadlock. We need a global mutex to prevent two
66 * simultaneous inserts (A into B and B into A) from racing and
67 * constructing a cycle without either insert observing that it is
68 * going to.
69 * It is necessary to acquire multiple "ep->mtx"es at once in the
70 * case when one epoll fd is added to another. In this case, we
71 * always acquire the locks in the order of nesting (i.e. after
72 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
73 * before e2->mtx). Since we disallow cycles of epoll file
74 * descriptors, this ensures that the mutexes are well-ordered. In
75 * order to communicate this nesting to lockdep, when walking a tree
76 * of epoll file descriptors, we use the current recursion depth as
77 * the lockdep subkey.
78 * It is possible to drop the "ep->mtx" and to use the global
79 * mutex "epmutex" (together with "ep->lock") to have it working,
80 * but having "ep->mtx" will make the interface more scalable.
81 * Events that require holding "epmutex" are very rare, while for
82 * normal operations the epoll private "ep->mtx" will guarantee
83 * a better scalability.
84 */
85
86 /* Epoll private bits inside the event mask */
87 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET | EPOLLEXCLUSIVE)
88
89 #define EPOLLINOUT_BITS (EPOLLIN | EPOLLOUT)
90
91 #define EPOLLEXCLUSIVE_OK_BITS (EPOLLINOUT_BITS | EPOLLERR | EPOLLHUP | \
92 EPOLLWAKEUP | EPOLLET | EPOLLEXCLUSIVE)
93
94 /* Maximum number of nesting allowed inside epoll sets */
95 #define EP_MAX_NESTS 4
96
97 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
98
99 #define EP_UNACTIVE_PTR ((void *) -1L)
100
101 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
102
103 struct epoll_filefd {
104 struct file *file;
105 int fd;
106 } __packed;
107
108 /* Wait structure used by the poll hooks */
109 struct eppoll_entry {
110 /* List header used to link this structure to the "struct epitem" */
111 struct eppoll_entry *next;
112
113 /* The "base" pointer is set to the container "struct epitem" */
114 struct epitem *base;
115
116 /*
117 * Wait queue item that will be linked to the target file wait
118 * queue head.
119 */
120 wait_queue_entry_t wait;
121
122 /* The wait queue head that linked the "wait" wait queue item */
123 wait_queue_head_t *whead;
124 };
125
126 /*
127 * Each file descriptor added to the eventpoll interface will
128 * have an entry of this type linked to the "rbr" RB tree.
129 * Avoid increasing the size of this struct, there can be many thousands
130 * of these on a server and we do not want this to take another cache line.
131 */
132 struct epitem {
133 union {
134 /* RB tree node links this structure to the eventpoll RB tree */
135 struct rb_node rbn;
136 /* Used to free the struct epitem */
137 struct rcu_head rcu;
138 };
139
140 /* List header used to link this structure to the eventpoll ready list */
141 struct list_head rdllink;
142
143 /*
144 * Works together "struct eventpoll"->ovflist in keeping the
145 * single linked chain of items.
146 */
147 struct epitem *next;
148
149 /* The file descriptor information this item refers to */
150 struct epoll_filefd ffd;
151
152 /*
153 * Protected by file->f_lock, true for to-be-released epitem already
154 * removed from the "struct file" items list; together with
155 * eventpoll->refcount orchestrates "struct eventpoll" disposal
156 */
157 bool dying;
158
159 /* List containing poll wait queues */
160 struct eppoll_entry *pwqlist;
161
162 /* The "container" of this item */
163 struct eventpoll *ep;
164
165 /* List header used to link this item to the "struct file" items list */
166 struct hlist_node fllink;
167
168 /* wakeup_source used when EPOLLWAKEUP is set */
169 struct wakeup_source __rcu *ws;
170
171 /* The structure that describe the interested events and the source fd */
172 struct epoll_event event;
173 };
174
175 /*
176 * This structure is stored inside the "private_data" member of the file
177 * structure and represents the main data structure for the eventpoll
178 * interface.
179 */
180 struct eventpoll {
181 /*
182 * This mutex is used to ensure that files are not removed
183 * while epoll is using them. This is held during the event
184 * collection loop, the file cleanup path, the epoll file exit
185 * code and the ctl operations.
186 */
187 struct mutex mtx;
188
189 /* Wait queue used by sys_epoll_wait() */
190 wait_queue_head_t wq;
191
192 /* Wait queue used by file->poll() */
193 wait_queue_head_t poll_wait;
194
195 /* List of ready file descriptors */
196 struct list_head rdllist;
197
198 /* Lock which protects rdllist and ovflist */
199 rwlock_t lock;
200
201 /* RB tree root used to store monitored fd structs */
202 struct rb_root_cached rbr;
203
204 /*
205 * This is a single linked list that chains all the "struct epitem" that
206 * happened while transferring ready events to userspace w/out
207 * holding ->lock.
208 */
209 struct epitem *ovflist;
210
211 /* wakeup_source used when ep_scan_ready_list is running */
212 struct wakeup_source *ws;
213
214 /* The user that created the eventpoll descriptor */
215 struct user_struct *user;
216
217 struct file *file;
218
219 /* used to optimize loop detection check */
220 u64 gen;
221 struct hlist_head refs;
222
223 /*
224 * usage count, used together with epitem->dying to
225 * orchestrate the disposal of this struct
226 */
227 refcount_t refcount;
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 #ifdef CONFIG_DEBUG_LOCK_ALLOC
235 /* tracks wakeup nests for lockdep validation */
236 u8 nests;
237 #endif
238 };
239
240 /* Wrapper struct used by poll queueing */
241 struct ep_pqueue {
242 poll_table pt;
243 struct epitem *epi;
244 };
245
246 /*
247 * Configuration options available inside /proc/sys/fs/epoll/
248 */
249 /* Maximum number of epoll watched descriptors, per user */
250 static long max_user_watches __read_mostly;
251
252 /* Used for cycles detection */
253 static DEFINE_MUTEX(epmutex);
254
255 static u64 loop_check_gen = 0;
256
257 /* Used to check for epoll file descriptor inclusion loops */
258 static struct eventpoll *inserting_into;
259
260 /* Slab cache used to allocate "struct epitem" */
261 static struct kmem_cache *epi_cache __read_mostly;
262
263 /* Slab cache used to allocate "struct eppoll_entry" */
264 static struct kmem_cache *pwq_cache __read_mostly;
265
266 /*
267 * List of files with newly added links, where we may need to limit the number
268 * of emanating paths. Protected by the epmutex.
269 */
270 struct epitems_head {
271 struct hlist_head epitems;
272 struct epitems_head *next;
273 };
274 static struct epitems_head *tfile_check_list = EP_UNACTIVE_PTR;
275
276 static struct kmem_cache *ephead_cache __read_mostly;
277
free_ephead(struct epitems_head * head)278 static inline void free_ephead(struct epitems_head *head)
279 {
280 if (head)
281 kmem_cache_free(ephead_cache, head);
282 }
283
list_file(struct file * file)284 static void list_file(struct file *file)
285 {
286 struct epitems_head *head;
287
288 head = container_of(file->f_ep, struct epitems_head, epitems);
289 if (!head->next) {
290 head->next = tfile_check_list;
291 tfile_check_list = head;
292 }
293 }
294
unlist_file(struct epitems_head * head)295 static void unlist_file(struct epitems_head *head)
296 {
297 struct epitems_head *to_free = head;
298 struct hlist_node *p = rcu_dereference(hlist_first_rcu(&head->epitems));
299 if (p) {
300 struct epitem *epi= container_of(p, struct epitem, fllink);
301 spin_lock(&epi->ffd.file->f_lock);
302 if (!hlist_empty(&head->epitems))
303 to_free = NULL;
304 head->next = NULL;
305 spin_unlock(&epi->ffd.file->f_lock);
306 }
307 free_ephead(to_free);
308 }
309
310 #ifdef CONFIG_SYSCTL
311
312 #include <linux/sysctl.h>
313
314 static long long_zero;
315 static long long_max = LONG_MAX;
316
317 static struct ctl_table epoll_table[] = {
318 {
319 .procname = "max_user_watches",
320 .data = &max_user_watches,
321 .maxlen = sizeof(max_user_watches),
322 .mode = 0644,
323 .proc_handler = proc_doulongvec_minmax,
324 .extra1 = &long_zero,
325 .extra2 = &long_max,
326 },
327 { }
328 };
329
epoll_sysctls_init(void)330 static void __init epoll_sysctls_init(void)
331 {
332 register_sysctl("fs/epoll", epoll_table);
333 }
334 #else
335 #define epoll_sysctls_init() do { } while (0)
336 #endif /* CONFIG_SYSCTL */
337
338 static const struct file_operations eventpoll_fops;
339
is_file_epoll(struct file * f)340 static inline int is_file_epoll(struct file *f)
341 {
342 return f->f_op == &eventpoll_fops;
343 }
344
345 /* Setup the structure that is used as key for the RB tree */
ep_set_ffd(struct epoll_filefd * ffd,struct file * file,int fd)346 static inline void ep_set_ffd(struct epoll_filefd *ffd,
347 struct file *file, int fd)
348 {
349 ffd->file = file;
350 ffd->fd = fd;
351 }
352
353 /* Compare RB tree keys */
ep_cmp_ffd(struct epoll_filefd * p1,struct epoll_filefd * p2)354 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
355 struct epoll_filefd *p2)
356 {
357 return (p1->file > p2->file ? +1:
358 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
359 }
360
361 /* Tells us if the item is currently linked */
ep_is_linked(struct epitem * epi)362 static inline int ep_is_linked(struct epitem *epi)
363 {
364 return !list_empty(&epi->rdllink);
365 }
366
ep_pwq_from_wait(wait_queue_entry_t * p)367 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_entry_t *p)
368 {
369 return container_of(p, struct eppoll_entry, wait);
370 }
371
372 /* Get the "struct epitem" from a wait queue pointer */
ep_item_from_wait(wait_queue_entry_t * p)373 static inline struct epitem *ep_item_from_wait(wait_queue_entry_t *p)
374 {
375 return container_of(p, struct eppoll_entry, wait)->base;
376 }
377
378 /**
379 * ep_events_available - Checks if ready events might be available.
380 *
381 * @ep: Pointer to the eventpoll context.
382 *
383 * Return: a value different than %zero if ready events are available,
384 * or %zero otherwise.
385 */
ep_events_available(struct eventpoll * ep)386 static inline int ep_events_available(struct eventpoll *ep)
387 {
388 return !list_empty_careful(&ep->rdllist) ||
389 READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR;
390 }
391
392 #ifdef CONFIG_NET_RX_BUSY_POLL
ep_busy_loop_end(void * p,unsigned long start_time)393 static bool ep_busy_loop_end(void *p, unsigned long start_time)
394 {
395 struct eventpoll *ep = p;
396
397 return ep_events_available(ep) || busy_loop_timeout(start_time);
398 }
399
400 /*
401 * Busy poll if globally on and supporting sockets found && no events,
402 * busy loop will return if need_resched or ep_events_available.
403 *
404 * we must do our busy polling with irqs enabled
405 */
ep_busy_loop(struct eventpoll * ep,int nonblock)406 static bool ep_busy_loop(struct eventpoll *ep, int nonblock)
407 {
408 unsigned int napi_id = READ_ONCE(ep->napi_id);
409
410 if ((napi_id >= MIN_NAPI_ID) && net_busy_loop_on()) {
411 napi_busy_loop(napi_id, nonblock ? NULL : ep_busy_loop_end, ep, false,
412 BUSY_POLL_BUDGET);
413 if (ep_events_available(ep))
414 return true;
415 /*
416 * Busy poll timed out. Drop NAPI ID for now, we can add
417 * it back in when we have moved a socket with a valid NAPI
418 * ID onto the ready list.
419 */
420 ep->napi_id = 0;
421 return false;
422 }
423 return false;
424 }
425
426 /*
427 * Set epoll busy poll NAPI ID from sk.
428 */
ep_set_busy_poll_napi_id(struct epitem * epi)429 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
430 {
431 struct eventpoll *ep;
432 unsigned int napi_id;
433 struct socket *sock;
434 struct sock *sk;
435
436 if (!net_busy_loop_on())
437 return;
438
439 sock = sock_from_file(epi->ffd.file);
440 if (!sock)
441 return;
442
443 sk = sock->sk;
444 if (!sk)
445 return;
446
447 napi_id = READ_ONCE(sk->sk_napi_id);
448 ep = epi->ep;
449
450 /* Non-NAPI IDs can be rejected
451 * or
452 * Nothing to do if we already have this ID
453 */
454 if (napi_id < MIN_NAPI_ID || napi_id == ep->napi_id)
455 return;
456
457 /* record NAPI ID for use in next busy poll */
458 ep->napi_id = napi_id;
459 }
460
461 #else
462
ep_busy_loop(struct eventpoll * ep,int nonblock)463 static inline bool ep_busy_loop(struct eventpoll *ep, int nonblock)
464 {
465 return false;
466 }
467
ep_set_busy_poll_napi_id(struct epitem * epi)468 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
469 {
470 }
471
472 #endif /* CONFIG_NET_RX_BUSY_POLL */
473
474 /*
475 * As described in commit 0ccf831cb lockdep: annotate epoll
476 * the use of wait queues used by epoll is done in a very controlled
477 * manner. Wake ups can nest inside each other, but are never done
478 * with the same locking. For example:
479 *
480 * dfd = socket(...);
481 * efd1 = epoll_create();
482 * efd2 = epoll_create();
483 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
484 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
485 *
486 * When a packet arrives to the device underneath "dfd", the net code will
487 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
488 * callback wakeup entry on that queue, and the wake_up() performed by the
489 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
490 * (efd1) notices that it may have some event ready, so it needs to wake up
491 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
492 * that ends up in another wake_up(), after having checked about the
493 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
494 * avoid stack blasting.
495 *
496 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
497 * this special case of epoll.
498 */
499 #ifdef CONFIG_DEBUG_LOCK_ALLOC
500
ep_poll_safewake(struct eventpoll * ep,struct epitem * epi,unsigned pollflags)501 static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi,
502 unsigned pollflags)
503 {
504 struct eventpoll *ep_src;
505 unsigned long flags;
506 u8 nests = 0;
507
508 /*
509 * To set the subclass or nesting level for spin_lock_irqsave_nested()
510 * it might be natural to create a per-cpu nest count. However, since
511 * we can recurse on ep->poll_wait.lock, and a non-raw spinlock can
512 * schedule() in the -rt kernel, the per-cpu variable are no longer
513 * protected. Thus, we are introducing a per eventpoll nest field.
514 * If we are not being call from ep_poll_callback(), epi is NULL and
515 * we are at the first level of nesting, 0. Otherwise, we are being
516 * called from ep_poll_callback() and if a previous wakeup source is
517 * not an epoll file itself, we are at depth 1 since the wakeup source
518 * is depth 0. If the wakeup source is a previous epoll file in the
519 * wakeup chain then we use its nests value and record ours as
520 * nests + 1. The previous epoll file nests value is stable since its
521 * already holding its own poll_wait.lock.
522 */
523 if (epi) {
524 if ((is_file_epoll(epi->ffd.file))) {
525 ep_src = epi->ffd.file->private_data;
526 nests = ep_src->nests;
527 } else {
528 nests = 1;
529 }
530 }
531 spin_lock_irqsave_nested(&ep->poll_wait.lock, flags, nests);
532 ep->nests = nests + 1;
533 wake_up_locked_poll(&ep->poll_wait, EPOLLIN | pollflags);
534 ep->nests = 0;
535 spin_unlock_irqrestore(&ep->poll_wait.lock, flags);
536 }
537
538 #else
539
ep_poll_safewake(struct eventpoll * ep,struct epitem * epi,unsigned pollflags)540 static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi,
541 unsigned pollflags)
542 {
543 wake_up_poll(&ep->poll_wait, EPOLLIN | pollflags);
544 }
545
546 #endif
547
ep_remove_wait_queue(struct eppoll_entry * pwq)548 static void ep_remove_wait_queue(struct eppoll_entry *pwq)
549 {
550 wait_queue_head_t *whead;
551
552 rcu_read_lock();
553 /*
554 * If it is cleared by POLLFREE, it should be rcu-safe.
555 * If we read NULL we need a barrier paired with
556 * smp_store_release() in ep_poll_callback(), otherwise
557 * we rely on whead->lock.
558 */
559 whead = smp_load_acquire(&pwq->whead);
560 if (whead)
561 remove_wait_queue(whead, &pwq->wait);
562 rcu_read_unlock();
563 }
564
565 /*
566 * This function unregisters poll callbacks from the associated file
567 * descriptor. Must be called with "mtx" held.
568 */
ep_unregister_pollwait(struct eventpoll * ep,struct epitem * epi)569 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
570 {
571 struct eppoll_entry **p = &epi->pwqlist;
572 struct eppoll_entry *pwq;
573
574 while ((pwq = *p) != NULL) {
575 *p = pwq->next;
576 ep_remove_wait_queue(pwq);
577 kmem_cache_free(pwq_cache, pwq);
578 }
579 }
580
581 /* call only when ep->mtx is held */
ep_wakeup_source(struct epitem * epi)582 static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
583 {
584 return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
585 }
586
587 /* call only when ep->mtx is held */
ep_pm_stay_awake(struct epitem * epi)588 static inline void ep_pm_stay_awake(struct epitem *epi)
589 {
590 struct wakeup_source *ws = ep_wakeup_source(epi);
591
592 if (ws)
593 __pm_stay_awake(ws);
594 }
595
ep_has_wakeup_source(struct epitem * epi)596 static inline bool ep_has_wakeup_source(struct epitem *epi)
597 {
598 return rcu_access_pointer(epi->ws) ? true : false;
599 }
600
601 /* call when ep->mtx cannot be held (ep_poll_callback) */
ep_pm_stay_awake_rcu(struct epitem * epi)602 static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
603 {
604 struct wakeup_source *ws;
605
606 rcu_read_lock();
607 ws = rcu_dereference(epi->ws);
608 if (ws)
609 __pm_stay_awake(ws);
610 rcu_read_unlock();
611 }
612
613
614 /*
615 * ep->mutex needs to be held because we could be hit by
616 * eventpoll_release_file() and epoll_ctl().
617 */
ep_start_scan(struct eventpoll * ep,struct list_head * txlist)618 static void ep_start_scan(struct eventpoll *ep, struct list_head *txlist)
619 {
620 /*
621 * Steal the ready list, and re-init the original one to the
622 * empty list. Also, set ep->ovflist to NULL so that events
623 * happening while looping w/out locks, are not lost. We cannot
624 * have the poll callback to queue directly on ep->rdllist,
625 * because we want the "sproc" callback to be able to do it
626 * in a lockless way.
627 */
628 lockdep_assert_irqs_enabled();
629 write_lock_irq(&ep->lock);
630 list_splice_init(&ep->rdllist, txlist);
631 WRITE_ONCE(ep->ovflist, NULL);
632 write_unlock_irq(&ep->lock);
633 }
634
ep_done_scan(struct eventpoll * ep,struct list_head * txlist)635 static void ep_done_scan(struct eventpoll *ep,
636 struct list_head *txlist)
637 {
638 struct epitem *epi, *nepi;
639
640 write_lock_irq(&ep->lock);
641 /*
642 * During the time we spent inside the "sproc" callback, some
643 * other events might have been queued by the poll callback.
644 * We re-insert them inside the main ready-list here.
645 */
646 for (nepi = READ_ONCE(ep->ovflist); (epi = nepi) != NULL;
647 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
648 /*
649 * We need to check if the item is already in the list.
650 * During the "sproc" callback execution time, items are
651 * queued into ->ovflist but the "txlist" might already
652 * contain them, and the list_splice() below takes care of them.
653 */
654 if (!ep_is_linked(epi)) {
655 /*
656 * ->ovflist is LIFO, so we have to reverse it in order
657 * to keep in FIFO.
658 */
659 list_add(&epi->rdllink, &ep->rdllist);
660 ep_pm_stay_awake(epi);
661 }
662 }
663 /*
664 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
665 * releasing the lock, events will be queued in the normal way inside
666 * ep->rdllist.
667 */
668 WRITE_ONCE(ep->ovflist, EP_UNACTIVE_PTR);
669
670 /*
671 * Quickly re-inject items left on "txlist".
672 */
673 list_splice(txlist, &ep->rdllist);
674 __pm_relax(ep->ws);
675
676 if (!list_empty(&ep->rdllist)) {
677 if (waitqueue_active(&ep->wq))
678 wake_up(&ep->wq);
679 }
680
681 write_unlock_irq(&ep->lock);
682 }
683
epi_rcu_free(struct rcu_head * head)684 static void epi_rcu_free(struct rcu_head *head)
685 {
686 struct epitem *epi = container_of(head, struct epitem, rcu);
687 kmem_cache_free(epi_cache, epi);
688 }
689
ep_get(struct eventpoll * ep)690 static void ep_get(struct eventpoll *ep)
691 {
692 refcount_inc(&ep->refcount);
693 }
694
695 /*
696 * Returns true if the event poll can be disposed
697 */
ep_refcount_dec_and_test(struct eventpoll * ep)698 static bool ep_refcount_dec_and_test(struct eventpoll *ep)
699 {
700 if (!refcount_dec_and_test(&ep->refcount))
701 return false;
702
703 WARN_ON_ONCE(!RB_EMPTY_ROOT(&ep->rbr.rb_root));
704 return true;
705 }
706
ep_free(struct eventpoll * ep)707 static void ep_free(struct eventpoll *ep)
708 {
709 mutex_destroy(&ep->mtx);
710 free_uid(ep->user);
711 wakeup_source_unregister(ep->ws);
712 kfree(ep);
713 }
714
715 /*
716 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
717 * all the associated resources. Must be called with "mtx" held.
718 * If the dying flag is set, do the removal only if force is true.
719 * This prevents ep_clear_and_put() from dropping all the ep references
720 * while running concurrently with eventpoll_release_file().
721 * Returns true if the eventpoll can be disposed.
722 */
__ep_remove(struct eventpoll * ep,struct epitem * epi,bool force)723 static bool __ep_remove(struct eventpoll *ep, struct epitem *epi, bool force)
724 {
725 struct file *file = epi->ffd.file;
726 struct epitems_head *to_free;
727 struct hlist_head *head;
728
729 lockdep_assert_irqs_enabled();
730
731 /*
732 * Removes poll wait queue hooks.
733 */
734 ep_unregister_pollwait(ep, epi);
735
736 /* Remove the current item from the list of epoll hooks */
737 spin_lock(&file->f_lock);
738 if (epi->dying && !force) {
739 spin_unlock(&file->f_lock);
740 return false;
741 }
742
743 to_free = NULL;
744 head = file->f_ep;
745 if (head->first == &epi->fllink && !epi->fllink.next) {
746 file->f_ep = NULL;
747 if (!is_file_epoll(file)) {
748 struct epitems_head *v;
749 v = container_of(head, struct epitems_head, epitems);
750 if (!smp_load_acquire(&v->next))
751 to_free = v;
752 }
753 }
754 hlist_del_rcu(&epi->fllink);
755 spin_unlock(&file->f_lock);
756 free_ephead(to_free);
757
758 rb_erase_cached(&epi->rbn, &ep->rbr);
759
760 write_lock_irq(&ep->lock);
761 if (ep_is_linked(epi))
762 list_del_init(&epi->rdllink);
763 write_unlock_irq(&ep->lock);
764
765 wakeup_source_unregister(ep_wakeup_source(epi));
766 /*
767 * At this point it is safe to free the eventpoll item. Use the union
768 * field epi->rcu, since we are trying to minimize the size of
769 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
770 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
771 * use of the rbn field.
772 */
773 call_rcu(&epi->rcu, epi_rcu_free);
774
775 percpu_counter_dec(&ep->user->epoll_watches);
776 return ep_refcount_dec_and_test(ep);
777 }
778
779 /*
780 * ep_remove variant for callers owing an additional reference to the ep
781 */
ep_remove_safe(struct eventpoll * ep,struct epitem * epi)782 static void ep_remove_safe(struct eventpoll *ep, struct epitem *epi)
783 {
784 WARN_ON_ONCE(__ep_remove(ep, epi, false));
785 }
786
ep_clear_and_put(struct eventpoll * ep)787 static void ep_clear_and_put(struct eventpoll *ep)
788 {
789 struct rb_node *rbp, *next;
790 struct epitem *epi;
791 bool dispose;
792
793 /* We need to release all tasks waiting for these file */
794 if (waitqueue_active(&ep->poll_wait))
795 ep_poll_safewake(ep, NULL, 0);
796
797 mutex_lock(&ep->mtx);
798
799 /*
800 * Walks through the whole tree by unregistering poll callbacks.
801 */
802 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
803 epi = rb_entry(rbp, struct epitem, rbn);
804
805 ep_unregister_pollwait(ep, epi);
806 cond_resched();
807 }
808
809 /*
810 * Walks through the whole tree and try to free each "struct epitem".
811 * Note that ep_remove_safe() will not remove the epitem in case of a
812 * racing eventpoll_release_file(); the latter will do the removal.
813 * At this point we are sure no poll callbacks will be lingering around.
814 * Since we still own a reference to the eventpoll struct, the loop can't
815 * dispose it.
816 */
817 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = next) {
818 next = rb_next(rbp);
819 epi = rb_entry(rbp, struct epitem, rbn);
820 ep_remove_safe(ep, epi);
821 cond_resched();
822 }
823
824 dispose = ep_refcount_dec_and_test(ep);
825 mutex_unlock(&ep->mtx);
826
827 if (dispose)
828 ep_free(ep);
829 }
830
ep_eventpoll_release(struct inode * inode,struct file * file)831 static int ep_eventpoll_release(struct inode *inode, struct file *file)
832 {
833 struct eventpoll *ep = file->private_data;
834
835 if (ep)
836 ep_clear_and_put(ep);
837
838 return 0;
839 }
840
841 static __poll_t ep_item_poll(const struct epitem *epi, poll_table *pt, int depth);
842
__ep_eventpoll_poll(struct file * file,poll_table * wait,int depth)843 static __poll_t __ep_eventpoll_poll(struct file *file, poll_table *wait, int depth)
844 {
845 struct eventpoll *ep = file->private_data;
846 LIST_HEAD(txlist);
847 struct epitem *epi, *tmp;
848 poll_table pt;
849 __poll_t res = 0;
850
851 init_poll_funcptr(&pt, NULL);
852
853 /* Insert inside our poll wait queue */
854 poll_wait(file, &ep->poll_wait, wait);
855
856 /*
857 * Proceed to find out if wanted events are really available inside
858 * the ready list.
859 */
860 mutex_lock_nested(&ep->mtx, depth);
861 ep_start_scan(ep, &txlist);
862 list_for_each_entry_safe(epi, tmp, &txlist, rdllink) {
863 if (ep_item_poll(epi, &pt, depth + 1)) {
864 res = EPOLLIN | EPOLLRDNORM;
865 break;
866 } else {
867 /*
868 * Item has been dropped into the ready list by the poll
869 * callback, but it's not actually ready, as far as
870 * caller requested events goes. We can remove it here.
871 */
872 __pm_relax(ep_wakeup_source(epi));
873 list_del_init(&epi->rdllink);
874 }
875 }
876 ep_done_scan(ep, &txlist);
877 mutex_unlock(&ep->mtx);
878 return res;
879 }
880
881 /*
882 * Differs from ep_eventpoll_poll() in that internal callers already have
883 * the ep->mtx so we need to start from depth=1, such that mutex_lock_nested()
884 * is correctly annotated.
885 */
ep_item_poll(const struct epitem * epi,poll_table * pt,int depth)886 static __poll_t ep_item_poll(const struct epitem *epi, poll_table *pt,
887 int depth)
888 {
889 struct file *file = epi->ffd.file;
890 __poll_t res;
891
892 pt->_key = epi->event.events;
893 if (!is_file_epoll(file))
894 res = vfs_poll(file, pt);
895 else
896 res = __ep_eventpoll_poll(file, pt, depth);
897 return res & epi->event.events;
898 }
899
ep_eventpoll_poll(struct file * file,poll_table * wait)900 static __poll_t ep_eventpoll_poll(struct file *file, poll_table *wait)
901 {
902 return __ep_eventpoll_poll(file, wait, 0);
903 }
904
905 #ifdef CONFIG_PROC_FS
ep_show_fdinfo(struct seq_file * m,struct file * f)906 static void ep_show_fdinfo(struct seq_file *m, struct file *f)
907 {
908 struct eventpoll *ep = f->private_data;
909 struct rb_node *rbp;
910
911 mutex_lock(&ep->mtx);
912 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
913 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
914 struct inode *inode = file_inode(epi->ffd.file);
915
916 seq_printf(m, "tfd: %8d events: %8x data: %16llx "
917 " pos:%lli ino:%lx sdev:%x\n",
918 epi->ffd.fd, epi->event.events,
919 (long long)epi->event.data,
920 (long long)epi->ffd.file->f_pos,
921 inode->i_ino, inode->i_sb->s_dev);
922 if (seq_has_overflowed(m))
923 break;
924 }
925 mutex_unlock(&ep->mtx);
926 }
927 #endif
928
929 /* File callbacks that implement the eventpoll file behaviour */
930 static const struct file_operations eventpoll_fops = {
931 #ifdef CONFIG_PROC_FS
932 .show_fdinfo = ep_show_fdinfo,
933 #endif
934 .release = ep_eventpoll_release,
935 .poll = ep_eventpoll_poll,
936 .llseek = noop_llseek,
937 };
938
939 /*
940 * This is called from eventpoll_release() to unlink files from the eventpoll
941 * interface. We need to have this facility to cleanup correctly files that are
942 * closed without being removed from the eventpoll interface.
943 */
eventpoll_release_file(struct file * file)944 void eventpoll_release_file(struct file *file)
945 {
946 struct eventpoll *ep;
947 struct epitem *epi;
948 bool dispose;
949
950 /*
951 * Use the 'dying' flag to prevent a concurrent ep_clear_and_put() from
952 * touching the epitems list before eventpoll_release_file() can access
953 * the ep->mtx.
954 */
955 again:
956 spin_lock(&file->f_lock);
957 if (file->f_ep && file->f_ep->first) {
958 epi = hlist_entry(file->f_ep->first, struct epitem, fllink);
959 epi->dying = true;
960 spin_unlock(&file->f_lock);
961
962 /*
963 * ep access is safe as we still own a reference to the ep
964 * struct
965 */
966 ep = epi->ep;
967 mutex_lock(&ep->mtx);
968 dispose = __ep_remove(ep, epi, true);
969 mutex_unlock(&ep->mtx);
970
971 if (dispose)
972 ep_free(ep);
973 goto again;
974 }
975 spin_unlock(&file->f_lock);
976 }
977
ep_alloc(struct eventpoll ** pep)978 static int ep_alloc(struct eventpoll **pep)
979 {
980 int error;
981 struct user_struct *user;
982 struct eventpoll *ep;
983
984 user = get_current_user();
985 error = -ENOMEM;
986 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
987 if (unlikely(!ep))
988 goto free_uid;
989
990 mutex_init(&ep->mtx);
991 rwlock_init(&ep->lock);
992 init_waitqueue_head(&ep->wq);
993 init_waitqueue_head(&ep->poll_wait);
994 INIT_LIST_HEAD(&ep->rdllist);
995 ep->rbr = RB_ROOT_CACHED;
996 ep->ovflist = EP_UNACTIVE_PTR;
997 ep->user = user;
998 refcount_set(&ep->refcount, 1);
999
1000 *pep = ep;
1001
1002 return 0;
1003
1004 free_uid:
1005 free_uid(user);
1006 return error;
1007 }
1008
1009 /*
1010 * Search the file inside the eventpoll tree. The RB tree operations
1011 * are protected by the "mtx" mutex, and ep_find() must be called with
1012 * "mtx" held.
1013 */
ep_find(struct eventpoll * ep,struct file * file,int fd)1014 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
1015 {
1016 int kcmp;
1017 struct rb_node *rbp;
1018 struct epitem *epi, *epir = NULL;
1019 struct epoll_filefd ffd;
1020
1021 ep_set_ffd(&ffd, file, fd);
1022 for (rbp = ep->rbr.rb_root.rb_node; rbp; ) {
1023 epi = rb_entry(rbp, struct epitem, rbn);
1024 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
1025 if (kcmp > 0)
1026 rbp = rbp->rb_right;
1027 else if (kcmp < 0)
1028 rbp = rbp->rb_left;
1029 else {
1030 epir = epi;
1031 break;
1032 }
1033 }
1034
1035 return epir;
1036 }
1037
1038 #ifdef CONFIG_KCMP
ep_find_tfd(struct eventpoll * ep,int tfd,unsigned long toff)1039 static struct epitem *ep_find_tfd(struct eventpoll *ep, int tfd, unsigned long toff)
1040 {
1041 struct rb_node *rbp;
1042 struct epitem *epi;
1043
1044 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1045 epi = rb_entry(rbp, struct epitem, rbn);
1046 if (epi->ffd.fd == tfd) {
1047 if (toff == 0)
1048 return epi;
1049 else
1050 toff--;
1051 }
1052 cond_resched();
1053 }
1054
1055 return NULL;
1056 }
1057
get_epoll_tfile_raw_ptr(struct file * file,int tfd,unsigned long toff)1058 struct file *get_epoll_tfile_raw_ptr(struct file *file, int tfd,
1059 unsigned long toff)
1060 {
1061 struct file *file_raw;
1062 struct eventpoll *ep;
1063 struct epitem *epi;
1064
1065 if (!is_file_epoll(file))
1066 return ERR_PTR(-EINVAL);
1067
1068 ep = file->private_data;
1069
1070 mutex_lock(&ep->mtx);
1071 epi = ep_find_tfd(ep, tfd, toff);
1072 if (epi)
1073 file_raw = epi->ffd.file;
1074 else
1075 file_raw = ERR_PTR(-ENOENT);
1076 mutex_unlock(&ep->mtx);
1077
1078 return file_raw;
1079 }
1080 #endif /* CONFIG_KCMP */
1081
1082 /*
1083 * Adds a new entry to the tail of the list in a lockless way, i.e.
1084 * multiple CPUs are allowed to call this function concurrently.
1085 *
1086 * Beware: it is necessary to prevent any other modifications of the
1087 * existing list until all changes are completed, in other words
1088 * concurrent list_add_tail_lockless() calls should be protected
1089 * with a read lock, where write lock acts as a barrier which
1090 * makes sure all list_add_tail_lockless() calls are fully
1091 * completed.
1092 *
1093 * Also an element can be locklessly added to the list only in one
1094 * direction i.e. either to the tail or to the head, otherwise
1095 * concurrent access will corrupt the list.
1096 *
1097 * Return: %false if element has been already added to the list, %true
1098 * otherwise.
1099 */
list_add_tail_lockless(struct list_head * new,struct list_head * head)1100 static inline bool list_add_tail_lockless(struct list_head *new,
1101 struct list_head *head)
1102 {
1103 struct list_head *prev;
1104
1105 /*
1106 * This is simple 'new->next = head' operation, but cmpxchg()
1107 * is used in order to detect that same element has been just
1108 * added to the list from another CPU: the winner observes
1109 * new->next == new.
1110 */
1111 if (!try_cmpxchg(&new->next, &new, head))
1112 return false;
1113
1114 /*
1115 * Initially ->next of a new element must be updated with the head
1116 * (we are inserting to the tail) and only then pointers are atomically
1117 * exchanged. XCHG guarantees memory ordering, thus ->next should be
1118 * updated before pointers are actually swapped and pointers are
1119 * swapped before prev->next is updated.
1120 */
1121
1122 prev = xchg(&head->prev, new);
1123
1124 /*
1125 * It is safe to modify prev->next and new->prev, because a new element
1126 * is added only to the tail and new->next is updated before XCHG.
1127 */
1128
1129 prev->next = new;
1130 new->prev = prev;
1131
1132 return true;
1133 }
1134
1135 /*
1136 * Chains a new epi entry to the tail of the ep->ovflist in a lockless way,
1137 * i.e. multiple CPUs are allowed to call this function concurrently.
1138 *
1139 * Return: %false if epi element has been already chained, %true otherwise.
1140 */
chain_epi_lockless(struct epitem * epi)1141 static inline bool chain_epi_lockless(struct epitem *epi)
1142 {
1143 struct eventpoll *ep = epi->ep;
1144
1145 /* Fast preliminary check */
1146 if (epi->next != EP_UNACTIVE_PTR)
1147 return false;
1148
1149 /* Check that the same epi has not been just chained from another CPU */
1150 if (cmpxchg(&epi->next, EP_UNACTIVE_PTR, NULL) != EP_UNACTIVE_PTR)
1151 return false;
1152
1153 /* Atomically exchange tail */
1154 epi->next = xchg(&ep->ovflist, epi);
1155
1156 return true;
1157 }
1158
1159 /*
1160 * This is the callback that is passed to the wait queue wakeup
1161 * mechanism. It is called by the stored file descriptors when they
1162 * have events to report.
1163 *
1164 * This callback takes a read lock in order not to contend with concurrent
1165 * events from another file descriptor, thus all modifications to ->rdllist
1166 * or ->ovflist are lockless. Read lock is paired with the write lock from
1167 * ep_scan_ready_list(), which stops all list modifications and guarantees
1168 * that lists state is seen correctly.
1169 *
1170 * Another thing worth to mention is that ep_poll_callback() can be called
1171 * concurrently for the same @epi from different CPUs if poll table was inited
1172 * with several wait queues entries. Plural wakeup from different CPUs of a
1173 * single wait queue is serialized by wq.lock, but the case when multiple wait
1174 * queues are used should be detected accordingly. This is detected using
1175 * cmpxchg() operation.
1176 */
ep_poll_callback(wait_queue_entry_t * wait,unsigned mode,int sync,void * key)1177 static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1178 {
1179 int pwake = 0;
1180 struct epitem *epi = ep_item_from_wait(wait);
1181 struct eventpoll *ep = epi->ep;
1182 __poll_t pollflags = key_to_poll(key);
1183 unsigned long flags;
1184 int ewake = 0;
1185
1186 read_lock_irqsave(&ep->lock, flags);
1187
1188 ep_set_busy_poll_napi_id(epi);
1189
1190 /*
1191 * If the event mask does not contain any poll(2) event, we consider the
1192 * descriptor to be disabled. This condition is likely the effect of the
1193 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1194 * until the next EPOLL_CTL_MOD will be issued.
1195 */
1196 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1197 goto out_unlock;
1198
1199 /*
1200 * Check the events coming with the callback. At this stage, not
1201 * every device reports the events in the "key" parameter of the
1202 * callback. We need to be able to handle both cases here, hence the
1203 * test for "key" != NULL before the event match test.
1204 */
1205 if (pollflags && !(pollflags & epi->event.events))
1206 goto out_unlock;
1207
1208 /*
1209 * If we are transferring events to userspace, we can hold no locks
1210 * (because we're accessing user memory, and because of linux f_op->poll()
1211 * semantics). All the events that happen during that period of time are
1212 * chained in ep->ovflist and requeued later on.
1213 */
1214 if (READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR) {
1215 if (chain_epi_lockless(epi))
1216 ep_pm_stay_awake_rcu(epi);
1217 } else if (!ep_is_linked(epi)) {
1218 /* In the usual case, add event to ready list. */
1219 if (list_add_tail_lockless(&epi->rdllink, &ep->rdllist))
1220 ep_pm_stay_awake_rcu(epi);
1221 }
1222
1223 /*
1224 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1225 * wait list.
1226 */
1227 if (waitqueue_active(&ep->wq)) {
1228 if ((epi->event.events & EPOLLEXCLUSIVE) &&
1229 !(pollflags & POLLFREE)) {
1230 switch (pollflags & EPOLLINOUT_BITS) {
1231 case EPOLLIN:
1232 if (epi->event.events & EPOLLIN)
1233 ewake = 1;
1234 break;
1235 case EPOLLOUT:
1236 if (epi->event.events & EPOLLOUT)
1237 ewake = 1;
1238 break;
1239 case 0:
1240 ewake = 1;
1241 break;
1242 }
1243 }
1244 wake_up(&ep->wq);
1245 }
1246 if (waitqueue_active(&ep->poll_wait))
1247 pwake++;
1248
1249 out_unlock:
1250 read_unlock_irqrestore(&ep->lock, flags);
1251
1252 /* We have to call this outside the lock */
1253 if (pwake)
1254 ep_poll_safewake(ep, epi, pollflags & EPOLL_URING_WAKE);
1255
1256 if (!(epi->event.events & EPOLLEXCLUSIVE))
1257 ewake = 1;
1258
1259 if (pollflags & POLLFREE) {
1260 /*
1261 * If we race with ep_remove_wait_queue() it can miss
1262 * ->whead = NULL and do another remove_wait_queue() after
1263 * us, so we can't use __remove_wait_queue().
1264 */
1265 list_del_init(&wait->entry);
1266 /*
1267 * ->whead != NULL protects us from the race with
1268 * ep_clear_and_put() or ep_remove(), ep_remove_wait_queue()
1269 * takes whead->lock held by the caller. Once we nullify it,
1270 * nothing protects ep/epi or even wait.
1271 */
1272 smp_store_release(&ep_pwq_from_wait(wait)->whead, NULL);
1273 }
1274
1275 return ewake;
1276 }
1277
1278 /*
1279 * This is the callback that is used to add our wait queue to the
1280 * target file wakeup lists.
1281 */
ep_ptable_queue_proc(struct file * file,wait_queue_head_t * whead,poll_table * pt)1282 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1283 poll_table *pt)
1284 {
1285 struct ep_pqueue *epq = container_of(pt, struct ep_pqueue, pt);
1286 struct epitem *epi = epq->epi;
1287 struct eppoll_entry *pwq;
1288
1289 if (unlikely(!epi)) // an earlier allocation has failed
1290 return;
1291
1292 pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL);
1293 if (unlikely(!pwq)) {
1294 epq->epi = NULL;
1295 return;
1296 }
1297
1298 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1299 pwq->whead = whead;
1300 pwq->base = epi;
1301 if (epi->event.events & EPOLLEXCLUSIVE)
1302 add_wait_queue_exclusive(whead, &pwq->wait);
1303 else
1304 add_wait_queue(whead, &pwq->wait);
1305 pwq->next = epi->pwqlist;
1306 epi->pwqlist = pwq;
1307 }
1308
ep_rbtree_insert(struct eventpoll * ep,struct epitem * epi)1309 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1310 {
1311 int kcmp;
1312 struct rb_node **p = &ep->rbr.rb_root.rb_node, *parent = NULL;
1313 struct epitem *epic;
1314 bool leftmost = true;
1315
1316 while (*p) {
1317 parent = *p;
1318 epic = rb_entry(parent, struct epitem, rbn);
1319 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1320 if (kcmp > 0) {
1321 p = &parent->rb_right;
1322 leftmost = false;
1323 } else
1324 p = &parent->rb_left;
1325 }
1326 rb_link_node(&epi->rbn, parent, p);
1327 rb_insert_color_cached(&epi->rbn, &ep->rbr, leftmost);
1328 }
1329
1330
1331
1332 #define PATH_ARR_SIZE 5
1333 /*
1334 * These are the number paths of length 1 to 5, that we are allowing to emanate
1335 * from a single file of interest. For example, we allow 1000 paths of length
1336 * 1, to emanate from each file of interest. This essentially represents the
1337 * potential wakeup paths, which need to be limited in order to avoid massive
1338 * uncontrolled wakeup storms. The common use case should be a single ep which
1339 * is connected to n file sources. In this case each file source has 1 path
1340 * of length 1. Thus, the numbers below should be more than sufficient. These
1341 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1342 * and delete can't add additional paths. Protected by the epmutex.
1343 */
1344 static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1345 static int path_count[PATH_ARR_SIZE];
1346
path_count_inc(int nests)1347 static int path_count_inc(int nests)
1348 {
1349 /* Allow an arbitrary number of depth 1 paths */
1350 if (nests == 0)
1351 return 0;
1352
1353 if (++path_count[nests] > path_limits[nests])
1354 return -1;
1355 return 0;
1356 }
1357
path_count_init(void)1358 static void path_count_init(void)
1359 {
1360 int i;
1361
1362 for (i = 0; i < PATH_ARR_SIZE; i++)
1363 path_count[i] = 0;
1364 }
1365
reverse_path_check_proc(struct hlist_head * refs,int depth)1366 static int reverse_path_check_proc(struct hlist_head *refs, int depth)
1367 {
1368 int error = 0;
1369 struct epitem *epi;
1370
1371 if (depth > EP_MAX_NESTS) /* too deep nesting */
1372 return -1;
1373
1374 /* CTL_DEL can remove links here, but that can't increase our count */
1375 hlist_for_each_entry_rcu(epi, refs, fllink) {
1376 struct hlist_head *refs = &epi->ep->refs;
1377 if (hlist_empty(refs))
1378 error = path_count_inc(depth);
1379 else
1380 error = reverse_path_check_proc(refs, depth + 1);
1381 if (error != 0)
1382 break;
1383 }
1384 return error;
1385 }
1386
1387 /**
1388 * reverse_path_check - The tfile_check_list is list of epitem_head, which have
1389 * links that are proposed to be newly added. We need to
1390 * make sure that those added links don't add too many
1391 * paths such that we will spend all our time waking up
1392 * eventpoll objects.
1393 *
1394 * Return: %zero if the proposed links don't create too many paths,
1395 * %-1 otherwise.
1396 */
reverse_path_check(void)1397 static int reverse_path_check(void)
1398 {
1399 struct epitems_head *p;
1400
1401 for (p = tfile_check_list; p != EP_UNACTIVE_PTR; p = p->next) {
1402 int error;
1403 path_count_init();
1404 rcu_read_lock();
1405 error = reverse_path_check_proc(&p->epitems, 0);
1406 rcu_read_unlock();
1407 if (error)
1408 return error;
1409 }
1410 return 0;
1411 }
1412
ep_create_wakeup_source(struct epitem * epi)1413 static int ep_create_wakeup_source(struct epitem *epi)
1414 {
1415 struct name_snapshot n;
1416 struct wakeup_source *ws;
1417 char ws_name[64];
1418
1419 strlcpy(ws_name, "eventpoll", sizeof(ws_name));
1420 trace_android_vh_ep_create_wakeup_source(ws_name, sizeof(ws_name));
1421 if (!epi->ep->ws) {
1422 epi->ep->ws = wakeup_source_register(NULL, ws_name);
1423 if (!epi->ep->ws)
1424 return -ENOMEM;
1425 }
1426
1427 take_dentry_name_snapshot(&n, epi->ffd.file->f_path.dentry);
1428 strlcpy(ws_name, n.name.name, sizeof(ws_name));
1429 trace_android_vh_ep_create_wakeup_source(ws_name, sizeof(ws_name));
1430 ws = wakeup_source_register(NULL, ws_name);
1431 release_dentry_name_snapshot(&n);
1432
1433 if (!ws)
1434 return -ENOMEM;
1435 rcu_assign_pointer(epi->ws, ws);
1436
1437 return 0;
1438 }
1439
1440 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
ep_destroy_wakeup_source(struct epitem * epi)1441 static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1442 {
1443 struct wakeup_source *ws = ep_wakeup_source(epi);
1444
1445 RCU_INIT_POINTER(epi->ws, NULL);
1446
1447 /*
1448 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1449 * used internally by wakeup_source_remove, too (called by
1450 * wakeup_source_unregister), so we cannot use call_rcu
1451 */
1452 synchronize_rcu();
1453 wakeup_source_unregister(ws);
1454 }
1455
attach_epitem(struct file * file,struct epitem * epi)1456 static int attach_epitem(struct file *file, struct epitem *epi)
1457 {
1458 struct epitems_head *to_free = NULL;
1459 struct hlist_head *head = NULL;
1460 struct eventpoll *ep = NULL;
1461
1462 if (is_file_epoll(file))
1463 ep = file->private_data;
1464
1465 if (ep) {
1466 head = &ep->refs;
1467 } else if (!READ_ONCE(file->f_ep)) {
1468 allocate:
1469 to_free = kmem_cache_zalloc(ephead_cache, GFP_KERNEL);
1470 if (!to_free)
1471 return -ENOMEM;
1472 head = &to_free->epitems;
1473 }
1474 spin_lock(&file->f_lock);
1475 if (!file->f_ep) {
1476 if (unlikely(!head)) {
1477 spin_unlock(&file->f_lock);
1478 goto allocate;
1479 }
1480 file->f_ep = head;
1481 to_free = NULL;
1482 }
1483 hlist_add_head_rcu(&epi->fllink, file->f_ep);
1484 spin_unlock(&file->f_lock);
1485 free_ephead(to_free);
1486 return 0;
1487 }
1488
1489 /*
1490 * Must be called with "mtx" held.
1491 */
ep_insert(struct eventpoll * ep,const struct epoll_event * event,struct file * tfile,int fd,int full_check)1492 static int ep_insert(struct eventpoll *ep, const struct epoll_event *event,
1493 struct file *tfile, int fd, int full_check)
1494 {
1495 int error, pwake = 0;
1496 __poll_t revents;
1497 struct epitem *epi;
1498 struct ep_pqueue epq;
1499 struct eventpoll *tep = NULL;
1500
1501 if (is_file_epoll(tfile))
1502 tep = tfile->private_data;
1503
1504 lockdep_assert_irqs_enabled();
1505
1506 if (unlikely(percpu_counter_compare(&ep->user->epoll_watches,
1507 max_user_watches) >= 0))
1508 return -ENOSPC;
1509 percpu_counter_inc(&ep->user->epoll_watches);
1510
1511 if (!(epi = kmem_cache_zalloc(epi_cache, GFP_KERNEL))) {
1512 percpu_counter_dec(&ep->user->epoll_watches);
1513 return -ENOMEM;
1514 }
1515
1516 /* Item initialization follow here ... */
1517 INIT_LIST_HEAD(&epi->rdllink);
1518 epi->ep = ep;
1519 ep_set_ffd(&epi->ffd, tfile, fd);
1520 epi->event = *event;
1521 epi->next = EP_UNACTIVE_PTR;
1522
1523 if (tep)
1524 mutex_lock_nested(&tep->mtx, 1);
1525 /* Add the current item to the list of active epoll hook for this file */
1526 if (unlikely(attach_epitem(tfile, epi) < 0)) {
1527 if (tep)
1528 mutex_unlock(&tep->mtx);
1529 kmem_cache_free(epi_cache, epi);
1530 percpu_counter_dec(&ep->user->epoll_watches);
1531 return -ENOMEM;
1532 }
1533
1534 if (full_check && !tep)
1535 list_file(tfile);
1536
1537 /*
1538 * Add the current item to the RB tree. All RB tree operations are
1539 * protected by "mtx", and ep_insert() is called with "mtx" held.
1540 */
1541 ep_rbtree_insert(ep, epi);
1542 if (tep)
1543 mutex_unlock(&tep->mtx);
1544
1545 /*
1546 * ep_remove_safe() calls in the later error paths can't lead to
1547 * ep_free() as the ep file itself still holds an ep reference.
1548 */
1549 ep_get(ep);
1550
1551 /* now check if we've created too many backpaths */
1552 if (unlikely(full_check && reverse_path_check())) {
1553 ep_remove_safe(ep, epi);
1554 return -EINVAL;
1555 }
1556
1557 if (epi->event.events & EPOLLWAKEUP) {
1558 error = ep_create_wakeup_source(epi);
1559 if (error) {
1560 ep_remove_safe(ep, epi);
1561 return error;
1562 }
1563 }
1564
1565 /* Initialize the poll table using the queue callback */
1566 epq.epi = epi;
1567 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1568
1569 /*
1570 * Attach the item to the poll hooks and get current event bits.
1571 * We can safely use the file* here because its usage count has
1572 * been increased by the caller of this function. Note that after
1573 * this operation completes, the poll callback can start hitting
1574 * the new item.
1575 */
1576 revents = ep_item_poll(epi, &epq.pt, 1);
1577
1578 /*
1579 * We have to check if something went wrong during the poll wait queue
1580 * install process. Namely an allocation for a wait queue failed due
1581 * high memory pressure.
1582 */
1583 if (unlikely(!epq.epi)) {
1584 ep_remove_safe(ep, epi);
1585 return -ENOMEM;
1586 }
1587
1588 /* We have to drop the new item inside our item list to keep track of it */
1589 write_lock_irq(&ep->lock);
1590
1591 /* record NAPI ID of new item if present */
1592 ep_set_busy_poll_napi_id(epi);
1593
1594 /* If the file is already "ready" we drop it inside the ready list */
1595 if (revents && !ep_is_linked(epi)) {
1596 list_add_tail(&epi->rdllink, &ep->rdllist);
1597 ep_pm_stay_awake(epi);
1598
1599 /* Notify waiting tasks that events are available */
1600 if (waitqueue_active(&ep->wq))
1601 wake_up(&ep->wq);
1602 if (waitqueue_active(&ep->poll_wait))
1603 pwake++;
1604 }
1605
1606 write_unlock_irq(&ep->lock);
1607
1608 /* We have to call this outside the lock */
1609 if (pwake)
1610 ep_poll_safewake(ep, NULL, 0);
1611
1612 return 0;
1613 }
1614
1615 /*
1616 * Modify the interest event mask by dropping an event if the new mask
1617 * has a match in the current file status. Must be called with "mtx" held.
1618 */
ep_modify(struct eventpoll * ep,struct epitem * epi,const struct epoll_event * event)1619 static int ep_modify(struct eventpoll *ep, struct epitem *epi,
1620 const struct epoll_event *event)
1621 {
1622 int pwake = 0;
1623 poll_table pt;
1624
1625 lockdep_assert_irqs_enabled();
1626
1627 init_poll_funcptr(&pt, NULL);
1628
1629 /*
1630 * Set the new event interest mask before calling f_op->poll();
1631 * otherwise we might miss an event that happens between the
1632 * f_op->poll() call and the new event set registering.
1633 */
1634 epi->event.events = event->events; /* need barrier below */
1635 epi->event.data = event->data; /* protected by mtx */
1636 if (epi->event.events & EPOLLWAKEUP) {
1637 if (!ep_has_wakeup_source(epi))
1638 ep_create_wakeup_source(epi);
1639 } else if (ep_has_wakeup_source(epi)) {
1640 ep_destroy_wakeup_source(epi);
1641 }
1642
1643 /*
1644 * The following barrier has two effects:
1645 *
1646 * 1) Flush epi changes above to other CPUs. This ensures
1647 * we do not miss events from ep_poll_callback if an
1648 * event occurs immediately after we call f_op->poll().
1649 * We need this because we did not take ep->lock while
1650 * changing epi above (but ep_poll_callback does take
1651 * ep->lock).
1652 *
1653 * 2) We also need to ensure we do not miss _past_ events
1654 * when calling f_op->poll(). This barrier also
1655 * pairs with the barrier in wq_has_sleeper (see
1656 * comments for wq_has_sleeper).
1657 *
1658 * This barrier will now guarantee ep_poll_callback or f_op->poll
1659 * (or both) will notice the readiness of an item.
1660 */
1661 smp_mb();
1662
1663 /*
1664 * Get current event bits. We can safely use the file* here because
1665 * its usage count has been increased by the caller of this function.
1666 * If the item is "hot" and it is not registered inside the ready
1667 * list, push it inside.
1668 */
1669 if (ep_item_poll(epi, &pt, 1)) {
1670 write_lock_irq(&ep->lock);
1671 if (!ep_is_linked(epi)) {
1672 list_add_tail(&epi->rdllink, &ep->rdllist);
1673 ep_pm_stay_awake(epi);
1674
1675 /* Notify waiting tasks that events are available */
1676 if (waitqueue_active(&ep->wq))
1677 wake_up(&ep->wq);
1678 if (waitqueue_active(&ep->poll_wait))
1679 pwake++;
1680 }
1681 write_unlock_irq(&ep->lock);
1682 }
1683
1684 /* We have to call this outside the lock */
1685 if (pwake)
1686 ep_poll_safewake(ep, NULL, 0);
1687
1688 return 0;
1689 }
1690
ep_send_events(struct eventpoll * ep,struct epoll_event __user * events,int maxevents)1691 static int ep_send_events(struct eventpoll *ep,
1692 struct epoll_event __user *events, int maxevents)
1693 {
1694 struct epitem *epi, *tmp;
1695 LIST_HEAD(txlist);
1696 poll_table pt;
1697 int res = 0;
1698
1699 /*
1700 * Always short-circuit for fatal signals to allow threads to make a
1701 * timely exit without the chance of finding more events available and
1702 * fetching repeatedly.
1703 */
1704 if (fatal_signal_pending(current))
1705 return -EINTR;
1706
1707 init_poll_funcptr(&pt, NULL);
1708
1709 mutex_lock(&ep->mtx);
1710 ep_start_scan(ep, &txlist);
1711
1712 /*
1713 * We can loop without lock because we are passed a task private list.
1714 * Items cannot vanish during the loop we are holding ep->mtx.
1715 */
1716 list_for_each_entry_safe(epi, tmp, &txlist, rdllink) {
1717 struct wakeup_source *ws;
1718 __poll_t revents;
1719
1720 if (res >= maxevents)
1721 break;
1722
1723 /*
1724 * Activate ep->ws before deactivating epi->ws to prevent
1725 * triggering auto-suspend here (in case we reactive epi->ws
1726 * below).
1727 *
1728 * This could be rearranged to delay the deactivation of epi->ws
1729 * instead, but then epi->ws would temporarily be out of sync
1730 * with ep_is_linked().
1731 */
1732 ws = ep_wakeup_source(epi);
1733 if (ws) {
1734 if (ws->active)
1735 __pm_stay_awake(ep->ws);
1736 __pm_relax(ws);
1737 }
1738
1739 list_del_init(&epi->rdllink);
1740
1741 /*
1742 * If the event mask intersect the caller-requested one,
1743 * deliver the event to userspace. Again, we are holding ep->mtx,
1744 * so no operations coming from userspace can change the item.
1745 */
1746 revents = ep_item_poll(epi, &pt, 1);
1747 if (!revents)
1748 continue;
1749
1750 events = epoll_put_uevent(revents, epi->event.data, events);
1751 if (!events) {
1752 list_add(&epi->rdllink, &txlist);
1753 ep_pm_stay_awake(epi);
1754 if (!res)
1755 res = -EFAULT;
1756 break;
1757 }
1758 res++;
1759 if (epi->event.events & EPOLLONESHOT)
1760 epi->event.events &= EP_PRIVATE_BITS;
1761 else if (!(epi->event.events & EPOLLET)) {
1762 /*
1763 * If this file has been added with Level
1764 * Trigger mode, we need to insert back inside
1765 * the ready list, so that the next call to
1766 * epoll_wait() will check again the events
1767 * availability. At this point, no one can insert
1768 * into ep->rdllist besides us. The epoll_ctl()
1769 * callers are locked out by
1770 * ep_scan_ready_list() holding "mtx" and the
1771 * poll callback will queue them in ep->ovflist.
1772 */
1773 list_add_tail(&epi->rdllink, &ep->rdllist);
1774 ep_pm_stay_awake(epi);
1775 }
1776 }
1777 ep_done_scan(ep, &txlist);
1778 mutex_unlock(&ep->mtx);
1779
1780 return res;
1781 }
1782
ep_timeout_to_timespec(struct timespec64 * to,long ms)1783 static struct timespec64 *ep_timeout_to_timespec(struct timespec64 *to, long ms)
1784 {
1785 struct timespec64 now;
1786
1787 if (ms < 0)
1788 return NULL;
1789
1790 if (!ms) {
1791 to->tv_sec = 0;
1792 to->tv_nsec = 0;
1793 return to;
1794 }
1795
1796 to->tv_sec = ms / MSEC_PER_SEC;
1797 to->tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC);
1798
1799 ktime_get_ts64(&now);
1800 *to = timespec64_add_safe(now, *to);
1801 return to;
1802 }
1803
1804 /*
1805 * autoremove_wake_function, but remove even on failure to wake up, because we
1806 * know that default_wake_function/ttwu will only fail if the thread is already
1807 * woken, and in that case the ep_poll loop will remove the entry anyways, not
1808 * try to reuse it.
1809 */
ep_autoremove_wake_function(struct wait_queue_entry * wq_entry,unsigned int mode,int sync,void * key)1810 static int ep_autoremove_wake_function(struct wait_queue_entry *wq_entry,
1811 unsigned int mode, int sync, void *key)
1812 {
1813 int ret = default_wake_function(wq_entry, mode, sync, key);
1814
1815 /*
1816 * Pairs with list_empty_careful in ep_poll, and ensures future loop
1817 * iterations see the cause of this wakeup.
1818 */
1819 list_del_init_careful(&wq_entry->entry);
1820 return ret;
1821 }
1822
1823 /**
1824 * ep_poll - Retrieves ready events, and delivers them to the caller-supplied
1825 * event buffer.
1826 *
1827 * @ep: Pointer to the eventpoll context.
1828 * @events: Pointer to the userspace buffer where the ready events should be
1829 * stored.
1830 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1831 * @timeout: Maximum timeout for the ready events fetch operation, in
1832 * timespec. If the timeout is zero, the function will not block,
1833 * while if the @timeout ptr is NULL, the function will block
1834 * until at least one event has been retrieved (or an error
1835 * occurred).
1836 *
1837 * Return: the number of ready events which have been fetched, or an
1838 * error code, in case of error.
1839 */
ep_poll(struct eventpoll * ep,struct epoll_event __user * events,int maxevents,struct timespec64 * timeout)1840 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1841 int maxevents, struct timespec64 *timeout)
1842 {
1843 int res, eavail, timed_out = 0;
1844 u64 slack = 0;
1845 wait_queue_entry_t wait;
1846 ktime_t expires, *to = NULL;
1847
1848 lockdep_assert_irqs_enabled();
1849
1850 if (timeout && (timeout->tv_sec | timeout->tv_nsec)) {
1851 slack = select_estimate_accuracy(timeout);
1852 to = &expires;
1853 *to = timespec64_to_ktime(*timeout);
1854 } else if (timeout) {
1855 /*
1856 * Avoid the unnecessary trip to the wait queue loop, if the
1857 * caller specified a non blocking operation.
1858 */
1859 timed_out = 1;
1860 }
1861
1862 /*
1863 * This call is racy: We may or may not see events that are being added
1864 * to the ready list under the lock (e.g., in IRQ callbacks). For cases
1865 * with a non-zero timeout, this thread will check the ready list under
1866 * lock and will add to the wait queue. For cases with a zero
1867 * timeout, the user by definition should not care and will have to
1868 * recheck again.
1869 */
1870 eavail = ep_events_available(ep);
1871
1872 while (1) {
1873 if (eavail) {
1874 /*
1875 * Try to transfer events to user space. In case we get
1876 * 0 events and there's still timeout left over, we go
1877 * trying again in search of more luck.
1878 */
1879 res = ep_send_events(ep, events, maxevents);
1880 if (res)
1881 return res;
1882 }
1883
1884 if (timed_out)
1885 return 0;
1886
1887 eavail = ep_busy_loop(ep, timed_out);
1888 if (eavail)
1889 continue;
1890
1891 if (signal_pending(current))
1892 return -EINTR;
1893
1894 /*
1895 * Internally init_wait() uses autoremove_wake_function(),
1896 * thus wait entry is removed from the wait queue on each
1897 * wakeup. Why it is important? In case of several waiters
1898 * each new wakeup will hit the next waiter, giving it the
1899 * chance to harvest new event. Otherwise wakeup can be
1900 * lost. This is also good performance-wise, because on
1901 * normal wakeup path no need to call __remove_wait_queue()
1902 * explicitly, thus ep->lock is not taken, which halts the
1903 * event delivery.
1904 *
1905 * In fact, we now use an even more aggressive function that
1906 * unconditionally removes, because we don't reuse the wait
1907 * entry between loop iterations. This lets us also avoid the
1908 * performance issue if a process is killed, causing all of its
1909 * threads to wake up without being removed normally.
1910 */
1911 init_wait(&wait);
1912 wait.func = ep_autoremove_wake_function;
1913
1914 write_lock_irq(&ep->lock);
1915 /*
1916 * Barrierless variant, waitqueue_active() is called under
1917 * the same lock on wakeup ep_poll_callback() side, so it
1918 * is safe to avoid an explicit barrier.
1919 */
1920 __set_current_state(TASK_INTERRUPTIBLE);
1921
1922 /*
1923 * Do the final check under the lock. ep_scan_ready_list()
1924 * plays with two lists (->rdllist and ->ovflist) and there
1925 * is always a race when both lists are empty for short
1926 * period of time although events are pending, so lock is
1927 * important.
1928 */
1929 eavail = ep_events_available(ep);
1930 if (!eavail)
1931 __add_wait_queue_exclusive(&ep->wq, &wait);
1932
1933 write_unlock_irq(&ep->lock);
1934
1935 if (!eavail)
1936 timed_out = !schedule_hrtimeout_range(to, slack,
1937 HRTIMER_MODE_ABS);
1938 __set_current_state(TASK_RUNNING);
1939
1940 /*
1941 * We were woken up, thus go and try to harvest some events.
1942 * If timed out and still on the wait queue, recheck eavail
1943 * carefully under lock, below.
1944 */
1945 eavail = 1;
1946
1947 if (!list_empty_careful(&wait.entry)) {
1948 write_lock_irq(&ep->lock);
1949 /*
1950 * If the thread timed out and is not on the wait queue,
1951 * it means that the thread was woken up after its
1952 * timeout expired before it could reacquire the lock.
1953 * Thus, when wait.entry is empty, it needs to harvest
1954 * events.
1955 */
1956 if (timed_out)
1957 eavail = list_empty(&wait.entry);
1958 __remove_wait_queue(&ep->wq, &wait);
1959 write_unlock_irq(&ep->lock);
1960 }
1961 }
1962 }
1963
1964 /**
1965 * ep_loop_check_proc - verify that adding an epoll file inside another
1966 * epoll structure does not violate the constraints, in
1967 * terms of closed loops, or too deep chains (which can
1968 * result in excessive stack usage).
1969 *
1970 * @ep: the &struct eventpoll to be currently checked.
1971 * @depth: Current depth of the path being checked.
1972 *
1973 * Return: %zero if adding the epoll @file inside current epoll
1974 * structure @ep does not violate the constraints, or %-1 otherwise.
1975 */
ep_loop_check_proc(struct eventpoll * ep,int depth)1976 static int ep_loop_check_proc(struct eventpoll *ep, int depth)
1977 {
1978 int error = 0;
1979 struct rb_node *rbp;
1980 struct epitem *epi;
1981
1982 mutex_lock_nested(&ep->mtx, depth + 1);
1983 ep->gen = loop_check_gen;
1984 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1985 epi = rb_entry(rbp, struct epitem, rbn);
1986 if (unlikely(is_file_epoll(epi->ffd.file))) {
1987 struct eventpoll *ep_tovisit;
1988 ep_tovisit = epi->ffd.file->private_data;
1989 if (ep_tovisit->gen == loop_check_gen)
1990 continue;
1991 if (ep_tovisit == inserting_into || depth > EP_MAX_NESTS)
1992 error = -1;
1993 else
1994 error = ep_loop_check_proc(ep_tovisit, depth + 1);
1995 if (error != 0)
1996 break;
1997 } else {
1998 /*
1999 * If we've reached a file that is not associated with
2000 * an ep, then we need to check if the newly added
2001 * links are going to add too many wakeup paths. We do
2002 * this by adding it to the tfile_check_list, if it's
2003 * not already there, and calling reverse_path_check()
2004 * during ep_insert().
2005 */
2006 list_file(epi->ffd.file);
2007 }
2008 }
2009 mutex_unlock(&ep->mtx);
2010
2011 return error;
2012 }
2013
2014 /**
2015 * ep_loop_check - Performs a check to verify that adding an epoll file (@to)
2016 * into another epoll file (represented by @ep) does not create
2017 * closed loops or too deep chains.
2018 *
2019 * @ep: Pointer to the epoll we are inserting into.
2020 * @to: Pointer to the epoll to be inserted.
2021 *
2022 * Return: %zero if adding the epoll @to inside the epoll @from
2023 * does not violate the constraints, or %-1 otherwise.
2024 */
ep_loop_check(struct eventpoll * ep,struct eventpoll * to)2025 static int ep_loop_check(struct eventpoll *ep, struct eventpoll *to)
2026 {
2027 inserting_into = ep;
2028 return ep_loop_check_proc(to, 0);
2029 }
2030
clear_tfile_check_list(void)2031 static void clear_tfile_check_list(void)
2032 {
2033 rcu_read_lock();
2034 while (tfile_check_list != EP_UNACTIVE_PTR) {
2035 struct epitems_head *head = tfile_check_list;
2036 tfile_check_list = head->next;
2037 unlist_file(head);
2038 }
2039 rcu_read_unlock();
2040 }
2041
2042 /*
2043 * Open an eventpoll file descriptor.
2044 */
do_epoll_create(int flags)2045 static int do_epoll_create(int flags)
2046 {
2047 int error, fd;
2048 struct eventpoll *ep = NULL;
2049 struct file *file;
2050
2051 /* Check the EPOLL_* constant for consistency. */
2052 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
2053
2054 if (flags & ~EPOLL_CLOEXEC)
2055 return -EINVAL;
2056 /*
2057 * Create the internal data structure ("struct eventpoll").
2058 */
2059 error = ep_alloc(&ep);
2060 if (error < 0)
2061 return error;
2062 /*
2063 * Creates all the items needed to setup an eventpoll file. That is,
2064 * a file structure and a free file descriptor.
2065 */
2066 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
2067 if (fd < 0) {
2068 error = fd;
2069 goto out_free_ep;
2070 }
2071 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
2072 O_RDWR | (flags & O_CLOEXEC));
2073 if (IS_ERR(file)) {
2074 error = PTR_ERR(file);
2075 goto out_free_fd;
2076 }
2077 ep->file = file;
2078 fd_install(fd, file);
2079 return fd;
2080
2081 out_free_fd:
2082 put_unused_fd(fd);
2083 out_free_ep:
2084 ep_clear_and_put(ep);
2085 return error;
2086 }
2087
SYSCALL_DEFINE1(epoll_create1,int,flags)2088 SYSCALL_DEFINE1(epoll_create1, int, flags)
2089 {
2090 return do_epoll_create(flags);
2091 }
2092
SYSCALL_DEFINE1(epoll_create,int,size)2093 SYSCALL_DEFINE1(epoll_create, int, size)
2094 {
2095 if (size <= 0)
2096 return -EINVAL;
2097
2098 return do_epoll_create(0);
2099 }
2100
epoll_mutex_lock(struct mutex * mutex,int depth,bool nonblock)2101 static inline int epoll_mutex_lock(struct mutex *mutex, int depth,
2102 bool nonblock)
2103 {
2104 if (!nonblock) {
2105 mutex_lock_nested(mutex, depth);
2106 return 0;
2107 }
2108 if (mutex_trylock(mutex))
2109 return 0;
2110 return -EAGAIN;
2111 }
2112
do_epoll_ctl(int epfd,int op,int fd,struct epoll_event * epds,bool nonblock)2113 int do_epoll_ctl(int epfd, int op, int fd, struct epoll_event *epds,
2114 bool nonblock)
2115 {
2116 int error;
2117 int full_check = 0;
2118 struct fd f, tf;
2119 struct eventpoll *ep;
2120 struct epitem *epi;
2121 struct eventpoll *tep = NULL;
2122
2123 error = -EBADF;
2124 f = fdget(epfd);
2125 if (!f.file)
2126 goto error_return;
2127
2128 /* Get the "struct file *" for the target file */
2129 tf = fdget(fd);
2130 if (!tf.file)
2131 goto error_fput;
2132
2133 /* The target file descriptor must support poll */
2134 error = -EPERM;
2135 if (!file_can_poll(tf.file))
2136 goto error_tgt_fput;
2137
2138 /* Check if EPOLLWAKEUP is allowed */
2139 if (ep_op_has_event(op))
2140 ep_take_care_of_epollwakeup(epds);
2141
2142 /*
2143 * We have to check that the file structure underneath the file descriptor
2144 * the user passed to us _is_ an eventpoll file. And also we do not permit
2145 * adding an epoll file descriptor inside itself.
2146 */
2147 error = -EINVAL;
2148 if (f.file == tf.file || !is_file_epoll(f.file))
2149 goto error_tgt_fput;
2150
2151 /*
2152 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
2153 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
2154 * Also, we do not currently supported nested exclusive wakeups.
2155 */
2156 if (ep_op_has_event(op) && (epds->events & EPOLLEXCLUSIVE)) {
2157 if (op == EPOLL_CTL_MOD)
2158 goto error_tgt_fput;
2159 if (op == EPOLL_CTL_ADD && (is_file_epoll(tf.file) ||
2160 (epds->events & ~EPOLLEXCLUSIVE_OK_BITS)))
2161 goto error_tgt_fput;
2162 }
2163
2164 /*
2165 * At this point it is safe to assume that the "private_data" contains
2166 * our own data structure.
2167 */
2168 ep = f.file->private_data;
2169
2170 /*
2171 * When we insert an epoll file descriptor inside another epoll file
2172 * descriptor, there is the chance of creating closed loops, which are
2173 * better be handled here, than in more critical paths. While we are
2174 * checking for loops we also determine the list of files reachable
2175 * and hang them on the tfile_check_list, so we can check that we
2176 * haven't created too many possible wakeup paths.
2177 *
2178 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
2179 * the epoll file descriptor is attaching directly to a wakeup source,
2180 * unless the epoll file descriptor is nested. The purpose of taking the
2181 * 'epmutex' on add is to prevent complex toplogies such as loops and
2182 * deep wakeup paths from forming in parallel through multiple
2183 * EPOLL_CTL_ADD operations.
2184 */
2185 error = epoll_mutex_lock(&ep->mtx, 0, nonblock);
2186 if (error)
2187 goto error_tgt_fput;
2188 if (op == EPOLL_CTL_ADD) {
2189 if (READ_ONCE(f.file->f_ep) || ep->gen == loop_check_gen ||
2190 is_file_epoll(tf.file)) {
2191 mutex_unlock(&ep->mtx);
2192 error = epoll_mutex_lock(&epmutex, 0, nonblock);
2193 if (error)
2194 goto error_tgt_fput;
2195 loop_check_gen++;
2196 full_check = 1;
2197 if (is_file_epoll(tf.file)) {
2198 tep = tf.file->private_data;
2199 error = -ELOOP;
2200 if (ep_loop_check(ep, tep) != 0)
2201 goto error_tgt_fput;
2202 }
2203 error = epoll_mutex_lock(&ep->mtx, 0, nonblock);
2204 if (error)
2205 goto error_tgt_fput;
2206 }
2207 }
2208
2209 /*
2210 * Try to lookup the file inside our RB tree. Since we grabbed "mtx"
2211 * above, we can be sure to be able to use the item looked up by
2212 * ep_find() till we release the mutex.
2213 */
2214 epi = ep_find(ep, tf.file, fd);
2215
2216 error = -EINVAL;
2217 switch (op) {
2218 case EPOLL_CTL_ADD:
2219 if (!epi) {
2220 epds->events |= EPOLLERR | EPOLLHUP;
2221 error = ep_insert(ep, epds, tf.file, fd, full_check);
2222 } else
2223 error = -EEXIST;
2224 break;
2225 case EPOLL_CTL_DEL:
2226 if (epi) {
2227 /*
2228 * The eventpoll itself is still alive: the refcount
2229 * can't go to zero here.
2230 */
2231 ep_remove_safe(ep, epi);
2232 error = 0;
2233 } else {
2234 error = -ENOENT;
2235 }
2236 break;
2237 case EPOLL_CTL_MOD:
2238 if (epi) {
2239 if (!(epi->event.events & EPOLLEXCLUSIVE)) {
2240 epds->events |= EPOLLERR | EPOLLHUP;
2241 error = ep_modify(ep, epi, epds);
2242 }
2243 } else
2244 error = -ENOENT;
2245 break;
2246 }
2247 mutex_unlock(&ep->mtx);
2248
2249 error_tgt_fput:
2250 if (full_check) {
2251 clear_tfile_check_list();
2252 loop_check_gen++;
2253 mutex_unlock(&epmutex);
2254 }
2255
2256 fdput(tf);
2257 error_fput:
2258 fdput(f);
2259 error_return:
2260
2261 return error;
2262 }
2263
2264 /*
2265 * The following function implements the controller interface for
2266 * the eventpoll file that enables the insertion/removal/change of
2267 * file descriptors inside the interest set.
2268 */
SYSCALL_DEFINE4(epoll_ctl,int,epfd,int,op,int,fd,struct epoll_event __user *,event)2269 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
2270 struct epoll_event __user *, event)
2271 {
2272 struct epoll_event epds;
2273
2274 if (ep_op_has_event(op) &&
2275 copy_from_user(&epds, event, sizeof(struct epoll_event)))
2276 return -EFAULT;
2277
2278 return do_epoll_ctl(epfd, op, fd, &epds, false);
2279 }
2280
2281 /*
2282 * Implement the event wait interface for the eventpoll file. It is the kernel
2283 * part of the user space epoll_wait(2).
2284 */
do_epoll_wait(int epfd,struct epoll_event __user * events,int maxevents,struct timespec64 * to)2285 static int do_epoll_wait(int epfd, struct epoll_event __user *events,
2286 int maxevents, struct timespec64 *to)
2287 {
2288 int error;
2289 struct fd f;
2290 struct eventpoll *ep;
2291
2292 /* The maximum number of event must be greater than zero */
2293 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
2294 return -EINVAL;
2295
2296 /* Verify that the area passed by the user is writeable */
2297 if (!access_ok(events, maxevents * sizeof(struct epoll_event)))
2298 return -EFAULT;
2299
2300 /* Get the "struct file *" for the eventpoll file */
2301 f = fdget(epfd);
2302 if (!f.file)
2303 return -EBADF;
2304
2305 /*
2306 * We have to check that the file structure underneath the fd
2307 * the user passed to us _is_ an eventpoll file.
2308 */
2309 error = -EINVAL;
2310 if (!is_file_epoll(f.file))
2311 goto error_fput;
2312
2313 /*
2314 * At this point it is safe to assume that the "private_data" contains
2315 * our own data structure.
2316 */
2317 ep = f.file->private_data;
2318
2319 /* Time to fish for events ... */
2320 error = ep_poll(ep, events, maxevents, to);
2321
2322 error_fput:
2323 fdput(f);
2324 return error;
2325 }
2326
SYSCALL_DEFINE4(epoll_wait,int,epfd,struct epoll_event __user *,events,int,maxevents,int,timeout)2327 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
2328 int, maxevents, int, timeout)
2329 {
2330 struct timespec64 to;
2331
2332 return do_epoll_wait(epfd, events, maxevents,
2333 ep_timeout_to_timespec(&to, timeout));
2334 }
2335
2336 /*
2337 * Implement the event wait interface for the eventpoll file. It is the kernel
2338 * part of the user space epoll_pwait(2).
2339 */
do_epoll_pwait(int epfd,struct epoll_event __user * events,int maxevents,struct timespec64 * to,const sigset_t __user * sigmask,size_t sigsetsize)2340 static int do_epoll_pwait(int epfd, struct epoll_event __user *events,
2341 int maxevents, struct timespec64 *to,
2342 const sigset_t __user *sigmask, size_t sigsetsize)
2343 {
2344 int error;
2345
2346 /*
2347 * If the caller wants a certain signal mask to be set during the wait,
2348 * we apply it here.
2349 */
2350 error = set_user_sigmask(sigmask, sigsetsize);
2351 if (error)
2352 return error;
2353
2354 error = do_epoll_wait(epfd, events, maxevents, to);
2355
2356 restore_saved_sigmask_unless(error == -EINTR);
2357
2358 return error;
2359 }
2360
SYSCALL_DEFINE6(epoll_pwait,int,epfd,struct epoll_event __user *,events,int,maxevents,int,timeout,const sigset_t __user *,sigmask,size_t,sigsetsize)2361 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2362 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2363 size_t, sigsetsize)
2364 {
2365 struct timespec64 to;
2366
2367 return do_epoll_pwait(epfd, events, maxevents,
2368 ep_timeout_to_timespec(&to, timeout),
2369 sigmask, sigsetsize);
2370 }
2371
SYSCALL_DEFINE6(epoll_pwait2,int,epfd,struct epoll_event __user *,events,int,maxevents,const struct __kernel_timespec __user *,timeout,const sigset_t __user *,sigmask,size_t,sigsetsize)2372 SYSCALL_DEFINE6(epoll_pwait2, int, epfd, struct epoll_event __user *, events,
2373 int, maxevents, const struct __kernel_timespec __user *, timeout,
2374 const sigset_t __user *, sigmask, size_t, sigsetsize)
2375 {
2376 struct timespec64 ts, *to = NULL;
2377
2378 if (timeout) {
2379 if (get_timespec64(&ts, timeout))
2380 return -EFAULT;
2381 to = &ts;
2382 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
2383 return -EINVAL;
2384 }
2385
2386 return do_epoll_pwait(epfd, events, maxevents, to,
2387 sigmask, sigsetsize);
2388 }
2389
2390 #ifdef CONFIG_COMPAT
do_compat_epoll_pwait(int epfd,struct epoll_event __user * events,int maxevents,struct timespec64 * timeout,const compat_sigset_t __user * sigmask,compat_size_t sigsetsize)2391 static int do_compat_epoll_pwait(int epfd, struct epoll_event __user *events,
2392 int maxevents, struct timespec64 *timeout,
2393 const compat_sigset_t __user *sigmask,
2394 compat_size_t sigsetsize)
2395 {
2396 long err;
2397
2398 /*
2399 * If the caller wants a certain signal mask to be set during the wait,
2400 * we apply it here.
2401 */
2402 err = set_compat_user_sigmask(sigmask, sigsetsize);
2403 if (err)
2404 return err;
2405
2406 err = do_epoll_wait(epfd, events, maxevents, timeout);
2407
2408 restore_saved_sigmask_unless(err == -EINTR);
2409
2410 return err;
2411 }
2412
COMPAT_SYSCALL_DEFINE6(epoll_pwait,int,epfd,struct epoll_event __user *,events,int,maxevents,int,timeout,const compat_sigset_t __user *,sigmask,compat_size_t,sigsetsize)2413 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2414 struct epoll_event __user *, events,
2415 int, maxevents, int, timeout,
2416 const compat_sigset_t __user *, sigmask,
2417 compat_size_t, sigsetsize)
2418 {
2419 struct timespec64 to;
2420
2421 return do_compat_epoll_pwait(epfd, events, maxevents,
2422 ep_timeout_to_timespec(&to, timeout),
2423 sigmask, sigsetsize);
2424 }
2425
COMPAT_SYSCALL_DEFINE6(epoll_pwait2,int,epfd,struct epoll_event __user *,events,int,maxevents,const struct __kernel_timespec __user *,timeout,const compat_sigset_t __user *,sigmask,compat_size_t,sigsetsize)2426 COMPAT_SYSCALL_DEFINE6(epoll_pwait2, int, epfd,
2427 struct epoll_event __user *, events,
2428 int, maxevents,
2429 const struct __kernel_timespec __user *, timeout,
2430 const compat_sigset_t __user *, sigmask,
2431 compat_size_t, sigsetsize)
2432 {
2433 struct timespec64 ts, *to = NULL;
2434
2435 if (timeout) {
2436 if (get_timespec64(&ts, timeout))
2437 return -EFAULT;
2438 to = &ts;
2439 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
2440 return -EINVAL;
2441 }
2442
2443 return do_compat_epoll_pwait(epfd, events, maxevents, to,
2444 sigmask, sigsetsize);
2445 }
2446
2447 #endif
2448
eventpoll_init(void)2449 static int __init eventpoll_init(void)
2450 {
2451 struct sysinfo si;
2452
2453 si_meminfo(&si);
2454 /*
2455 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2456 */
2457 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2458 EP_ITEM_COST;
2459 BUG_ON(max_user_watches < 0);
2460
2461 /*
2462 * We can have many thousands of epitems, so prevent this from
2463 * using an extra cache line on 64-bit (and smaller) CPUs
2464 */
2465 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2466
2467 /* Allocates slab cache used to allocate "struct epitem" items */
2468 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2469 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC|SLAB_ACCOUNT, NULL);
2470
2471 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2472 pwq_cache = kmem_cache_create("eventpoll_pwq",
2473 sizeof(struct eppoll_entry), 0, SLAB_PANIC|SLAB_ACCOUNT, NULL);
2474 epoll_sysctls_init();
2475
2476 ephead_cache = kmem_cache_create("ep_head",
2477 sizeof(struct epitems_head), 0, SLAB_PANIC|SLAB_ACCOUNT, NULL);
2478
2479 return 0;
2480 }
2481 fs_initcall(eventpoll_init);
2482