1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * fs/eventpoll.c (Efficient event retrieval implementation)
4 * Copyright (C) 2001,...,2009 Davide Libenzi
5 *
6 * Davide Libenzi <davidel@xmailserver.org>
7 */
8
9 #include <linux/init.h>
10 #include <linux/kernel.h>
11 #include <linux/sched/signal.h>
12 #include <linux/fs.h>
13 #include <linux/file.h>
14 #include <linux/signal.h>
15 #include <linux/errno.h>
16 #include <linux/mm.h>
17 #include <linux/slab.h>
18 #include <linux/poll.h>
19 #include <linux/string.h>
20 #include <linux/list.h>
21 #include <linux/hash.h>
22 #include <linux/spinlock.h>
23 #include <linux/syscalls.h>
24 #include <linux/rbtree.h>
25 #include <linux/wait.h>
26 #include <linux/eventpoll.h>
27 #include <linux/mount.h>
28 #include <linux/bitops.h>
29 #include <linux/mutex.h>
30 #include <linux/anon_inodes.h>
31 #include <linux/device.h>
32 #include <linux/uaccess.h>
33 #include <asm/io.h>
34 #include <asm/mman.h>
35 #include <linux/atomic.h>
36 #include <linux/proc_fs.h>
37 #include <linux/seq_file.h>
38 #include <linux/compat.h>
39 #include <linux/rculist.h>
40 #include <net/busy_poll.h>
41
42 /*
43 * LOCKING:
44 * There are three level of locking required by epoll :
45 *
46 * 1) epmutex (mutex)
47 * 2) ep->mtx (mutex)
48 * 3) ep->lock (rwlock)
49 *
50 * The acquire order is the one listed above, from 1 to 3.
51 * We need a rwlock (ep->lock) because we manipulate objects
52 * from inside the poll callback, that might be triggered from
53 * a wake_up() that in turn might be called from IRQ context.
54 * So we can't sleep inside the poll callback and hence we need
55 * a spinlock. During the event transfer loop (from kernel to
56 * user space) we could end up sleeping due a copy_to_user(), so
57 * we need a lock that will allow us to sleep. This lock is a
58 * mutex (ep->mtx). It is acquired during the event transfer loop,
59 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
60 * Then we also need a global mutex to serialize eventpoll_release_file()
61 * and ep_free().
62 * This mutex is acquired by ep_free() during the epoll file
63 * cleanup path and it is also acquired by eventpoll_release_file()
64 * if a file has been pushed inside an epoll set and it is then
65 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
66 * It is also acquired when inserting an epoll fd onto another epoll
67 * fd. We do this so that we walk the epoll tree and ensure that this
68 * insertion does not create a cycle of epoll file descriptors, which
69 * could lead to deadlock. We need a global mutex to prevent two
70 * simultaneous inserts (A into B and B into A) from racing and
71 * constructing a cycle without either insert observing that it is
72 * going to.
73 * It is necessary to acquire multiple "ep->mtx"es at once in the
74 * case when one epoll fd is added to another. In this case, we
75 * always acquire the locks in the order of nesting (i.e. after
76 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
77 * before e2->mtx). Since we disallow cycles of epoll file
78 * descriptors, this ensures that the mutexes are well-ordered. In
79 * order to communicate this nesting to lockdep, when walking a tree
80 * of epoll file descriptors, we use the current recursion depth as
81 * the lockdep subkey.
82 * It is possible to drop the "ep->mtx" and to use the global
83 * mutex "epmutex" (together with "ep->lock") to have it working,
84 * but having "ep->mtx" will make the interface more scalable.
85 * Events that require holding "epmutex" are very rare, while for
86 * normal operations the epoll private "ep->mtx" will guarantee
87 * a better scalability.
88 */
89
90 /* Epoll private bits inside the event mask */
91 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET | EPOLLEXCLUSIVE)
92
93 #define EPOLLINOUT_BITS (EPOLLIN | EPOLLOUT)
94
95 #define EPOLLEXCLUSIVE_OK_BITS (EPOLLINOUT_BITS | EPOLLERR | EPOLLHUP | \
96 EPOLLWAKEUP | EPOLLET | EPOLLEXCLUSIVE)
97
98 /* Maximum number of nesting allowed inside epoll sets */
99 #define EP_MAX_NESTS 4
100
101 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
102
103 #define EP_UNACTIVE_PTR ((void *) -1L)
104
105 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
106
107 struct epoll_filefd {
108 struct file *file;
109 int fd;
110 } __packed;
111
112 /*
113 * Structure used to track possible nested calls, for too deep recursions
114 * and loop cycles.
115 */
116 struct nested_call_node {
117 struct list_head llink;
118 void *cookie;
119 void *ctx;
120 };
121
122 /*
123 * This structure is used as collector for nested calls, to check for
124 * maximum recursion dept and loop cycles.
125 */
126 struct nested_calls {
127 struct list_head tasks_call_list;
128 spinlock_t lock;
129 };
130
131 /*
132 * Each file descriptor added to the eventpoll interface will
133 * have an entry of this type linked to the "rbr" RB tree.
134 * Avoid increasing the size of this struct, there can be many thousands
135 * of these on a server and we do not want this to take another cache line.
136 */
137 struct epitem {
138 union {
139 /* RB tree node links this structure to the eventpoll RB tree */
140 struct rb_node rbn;
141 /* Used to free the struct epitem */
142 struct rcu_head rcu;
143 };
144
145 /* List header used to link this structure to the eventpoll ready list */
146 struct list_head rdllink;
147
148 /*
149 * Works together "struct eventpoll"->ovflist in keeping the
150 * single linked chain of items.
151 */
152 struct epitem *next;
153
154 /* The file descriptor information this item refers to */
155 struct epoll_filefd ffd;
156
157 /* Number of active wait queue attached to poll operations */
158 int nwait;
159
160 /* List containing poll wait queues */
161 struct list_head pwqlist;
162
163 /* The "container" of this item */
164 struct eventpoll *ep;
165
166 /* List header used to link this item to the "struct file" items list */
167 struct list_head fllink;
168
169 /* wakeup_source used when EPOLLWAKEUP is set */
170 struct wakeup_source __rcu *ws;
171
172 /* The structure that describe the interested events and the source fd */
173 struct epoll_event event;
174 };
175
176 /*
177 * This structure is stored inside the "private_data" member of the file
178 * structure and represents the main data structure for the eventpoll
179 * interface.
180 */
181 struct eventpoll {
182 /*
183 * This mutex is used to ensure that files are not removed
184 * while epoll is using them. This is held during the event
185 * collection loop, the file cleanup path, the epoll file exit
186 * code and the ctl operations.
187 */
188 struct mutex mtx;
189
190 /* Wait queue used by sys_epoll_wait() */
191 wait_queue_head_t wq;
192
193 /* Wait queue used by file->poll() */
194 wait_queue_head_t poll_wait;
195
196 /* List of ready file descriptors */
197 struct list_head rdllist;
198
199 /* Lock which protects rdllist and ovflist */
200 rwlock_t lock;
201
202 /* RB tree root used to store monitored fd structs */
203 struct rb_root_cached rbr;
204
205 /*
206 * This is a single linked list that chains all the "struct epitem" that
207 * happened while transferring ready events to userspace w/out
208 * holding ->lock.
209 */
210 struct epitem *ovflist;
211
212 /* wakeup_source used when ep_scan_ready_list is running */
213 struct wakeup_source *ws;
214
215 /* The user that created the eventpoll descriptor */
216 struct user_struct *user;
217
218 struct file *file;
219
220 /* used to optimize loop detection check */
221 u64 gen;
222
223 #ifdef CONFIG_NET_RX_BUSY_POLL
224 /* used to track busy poll napi_id */
225 unsigned int napi_id;
226 #endif
227
228 #ifdef CONFIG_DEBUG_LOCK_ALLOC
229 /* tracks wakeup nests for lockdep validation */
230 u8 nests;
231 #endif
232 };
233
234 /* Wait structure used by the poll hooks */
235 struct eppoll_entry {
236 /* List header used to link this structure to the "struct epitem" */
237 struct list_head llink;
238
239 /* The "base" pointer is set to the container "struct epitem" */
240 struct epitem *base;
241
242 /*
243 * Wait queue item that will be linked to the target file wait
244 * queue head.
245 */
246 wait_queue_entry_t wait;
247
248 /* The wait queue head that linked the "wait" wait queue item */
249 wait_queue_head_t *whead;
250 };
251
252 /* Wrapper struct used by poll queueing */
253 struct ep_pqueue {
254 poll_table pt;
255 struct epitem *epi;
256 };
257
258 /* Used by the ep_send_events() function as callback private data */
259 struct ep_send_events_data {
260 int maxevents;
261 struct epoll_event __user *events;
262 int res;
263 };
264
265 /*
266 * Configuration options available inside /proc/sys/fs/epoll/
267 */
268 /* Maximum number of epoll watched descriptors, per user */
269 static long max_user_watches __read_mostly;
270
271 /*
272 * This mutex is used to serialize ep_free() and eventpoll_release_file().
273 */
274 static DEFINE_MUTEX(epmutex);
275
276 static u64 loop_check_gen = 0;
277
278 /* Used to check for epoll file descriptor inclusion loops */
279 static struct nested_calls poll_loop_ncalls;
280
281 /* Slab cache used to allocate "struct epitem" */
282 static struct kmem_cache *epi_cache __read_mostly;
283
284 /* Slab cache used to allocate "struct eppoll_entry" */
285 static struct kmem_cache *pwq_cache __read_mostly;
286
287 /*
288 * List of files with newly added links, where we may need to limit the number
289 * of emanating paths. Protected by the epmutex.
290 */
291 static LIST_HEAD(tfile_check_list);
292
293 #ifdef CONFIG_SYSCTL
294
295 #include <linux/sysctl.h>
296
297 static long long_zero;
298 static long long_max = LONG_MAX;
299
300 struct ctl_table epoll_table[] = {
301 {
302 .procname = "max_user_watches",
303 .data = &max_user_watches,
304 .maxlen = sizeof(max_user_watches),
305 .mode = 0644,
306 .proc_handler = proc_doulongvec_minmax,
307 .extra1 = &long_zero,
308 .extra2 = &long_max,
309 },
310 { }
311 };
312 #endif /* CONFIG_SYSCTL */
313
314 static const struct file_operations eventpoll_fops;
315
is_file_epoll(struct file * f)316 static inline int is_file_epoll(struct file *f)
317 {
318 return f->f_op == &eventpoll_fops;
319 }
320
321 /* Setup the structure that is used as key for the RB tree */
ep_set_ffd(struct epoll_filefd * ffd,struct file * file,int fd)322 static inline void ep_set_ffd(struct epoll_filefd *ffd,
323 struct file *file, int fd)
324 {
325 ffd->file = file;
326 ffd->fd = fd;
327 }
328
329 /* Compare RB tree keys */
ep_cmp_ffd(struct epoll_filefd * p1,struct epoll_filefd * p2)330 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
331 struct epoll_filefd *p2)
332 {
333 return (p1->file > p2->file ? +1:
334 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
335 }
336
337 /* Tells us if the item is currently linked */
ep_is_linked(struct epitem * epi)338 static inline int ep_is_linked(struct epitem *epi)
339 {
340 return !list_empty(&epi->rdllink);
341 }
342
ep_pwq_from_wait(wait_queue_entry_t * p)343 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_entry_t *p)
344 {
345 return container_of(p, struct eppoll_entry, wait);
346 }
347
348 /* Get the "struct epitem" from a wait queue pointer */
ep_item_from_wait(wait_queue_entry_t * p)349 static inline struct epitem *ep_item_from_wait(wait_queue_entry_t *p)
350 {
351 return container_of(p, struct eppoll_entry, wait)->base;
352 }
353
354 /* Get the "struct epitem" from an epoll queue wrapper */
ep_item_from_epqueue(poll_table * p)355 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
356 {
357 return container_of(p, struct ep_pqueue, pt)->epi;
358 }
359
360 /* Initialize the poll safe wake up structure */
ep_nested_calls_init(struct nested_calls * ncalls)361 static void ep_nested_calls_init(struct nested_calls *ncalls)
362 {
363 INIT_LIST_HEAD(&ncalls->tasks_call_list);
364 spin_lock_init(&ncalls->lock);
365 }
366
367 /**
368 * ep_events_available - Checks if ready events might be available.
369 *
370 * @ep: Pointer to the eventpoll context.
371 *
372 * Returns: Returns a value different than zero if ready events are available,
373 * or zero otherwise.
374 */
ep_events_available(struct eventpoll * ep)375 static inline int ep_events_available(struct eventpoll *ep)
376 {
377 return !list_empty_careful(&ep->rdllist) ||
378 READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR;
379 }
380
381 #ifdef CONFIG_NET_RX_BUSY_POLL
ep_busy_loop_end(void * p,unsigned long start_time)382 static bool ep_busy_loop_end(void *p, unsigned long start_time)
383 {
384 struct eventpoll *ep = p;
385
386 return ep_events_available(ep) || busy_loop_timeout(start_time);
387 }
388
389 /*
390 * Busy poll if globally on and supporting sockets found && no events,
391 * busy loop will return if need_resched or ep_events_available.
392 *
393 * we must do our busy polling with irqs enabled
394 */
ep_busy_loop(struct eventpoll * ep,int nonblock)395 static void ep_busy_loop(struct eventpoll *ep, int nonblock)
396 {
397 unsigned int napi_id = READ_ONCE(ep->napi_id);
398
399 if ((napi_id >= MIN_NAPI_ID) && net_busy_loop_on())
400 napi_busy_loop(napi_id, nonblock ? NULL : ep_busy_loop_end, ep);
401 }
402
ep_reset_busy_poll_napi_id(struct eventpoll * ep)403 static inline void ep_reset_busy_poll_napi_id(struct eventpoll *ep)
404 {
405 if (ep->napi_id)
406 ep->napi_id = 0;
407 }
408
409 /*
410 * Set epoll busy poll NAPI ID from sk.
411 */
ep_set_busy_poll_napi_id(struct epitem * epi)412 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
413 {
414 struct eventpoll *ep;
415 unsigned int napi_id;
416 struct socket *sock;
417 struct sock *sk;
418 int err;
419
420 if (!net_busy_loop_on())
421 return;
422
423 sock = sock_from_file(epi->ffd.file, &err);
424 if (!sock)
425 return;
426
427 sk = sock->sk;
428 if (!sk)
429 return;
430
431 napi_id = READ_ONCE(sk->sk_napi_id);
432 ep = epi->ep;
433
434 /* Non-NAPI IDs can be rejected
435 * or
436 * Nothing to do if we already have this ID
437 */
438 if (napi_id < MIN_NAPI_ID || napi_id == ep->napi_id)
439 return;
440
441 /* record NAPI ID for use in next busy poll */
442 ep->napi_id = napi_id;
443 }
444
445 #else
446
ep_busy_loop(struct eventpoll * ep,int nonblock)447 static inline void ep_busy_loop(struct eventpoll *ep, int nonblock)
448 {
449 }
450
ep_reset_busy_poll_napi_id(struct eventpoll * ep)451 static inline void ep_reset_busy_poll_napi_id(struct eventpoll *ep)
452 {
453 }
454
ep_set_busy_poll_napi_id(struct epitem * epi)455 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
456 {
457 }
458
459 #endif /* CONFIG_NET_RX_BUSY_POLL */
460
461 /**
462 * ep_call_nested - Perform a bound (possibly) nested call, by checking
463 * that the recursion limit is not exceeded, and that
464 * the same nested call (by the meaning of same cookie) is
465 * no re-entered.
466 *
467 * @ncalls: Pointer to the nested_calls structure to be used for this call.
468 * @nproc: Nested call core function pointer.
469 * @priv: Opaque data to be passed to the @nproc callback.
470 * @cookie: Cookie to be used to identify this nested call.
471 * @ctx: This instance context.
472 *
473 * Returns: Returns the code returned by the @nproc callback, or -1 if
474 * the maximum recursion limit has been exceeded.
475 */
ep_call_nested(struct nested_calls * ncalls,int (* nproc)(void *,void *,int),void * priv,void * cookie,void * ctx)476 static int ep_call_nested(struct nested_calls *ncalls,
477 int (*nproc)(void *, void *, int), void *priv,
478 void *cookie, void *ctx)
479 {
480 int error, call_nests = 0;
481 unsigned long flags;
482 struct list_head *lsthead = &ncalls->tasks_call_list;
483 struct nested_call_node *tncur;
484 struct nested_call_node tnode;
485
486 spin_lock_irqsave(&ncalls->lock, flags);
487
488 /*
489 * Try to see if the current task is already inside this wakeup call.
490 * We use a list here, since the population inside this set is always
491 * very much limited.
492 */
493 list_for_each_entry(tncur, lsthead, llink) {
494 if (tncur->ctx == ctx &&
495 (tncur->cookie == cookie || ++call_nests > EP_MAX_NESTS)) {
496 /*
497 * Ops ... loop detected or maximum nest level reached.
498 * We abort this wake by breaking the cycle itself.
499 */
500 error = -1;
501 goto out_unlock;
502 }
503 }
504
505 /* Add the current task and cookie to the list */
506 tnode.ctx = ctx;
507 tnode.cookie = cookie;
508 list_add(&tnode.llink, lsthead);
509
510 spin_unlock_irqrestore(&ncalls->lock, flags);
511
512 /* Call the nested function */
513 error = (*nproc)(priv, cookie, call_nests);
514
515 /* Remove the current task from the list */
516 spin_lock_irqsave(&ncalls->lock, flags);
517 list_del(&tnode.llink);
518 out_unlock:
519 spin_unlock_irqrestore(&ncalls->lock, flags);
520
521 return error;
522 }
523
524 /*
525 * As described in commit 0ccf831cb lockdep: annotate epoll
526 * the use of wait queues used by epoll is done in a very controlled
527 * manner. Wake ups can nest inside each other, but are never done
528 * with the same locking. For example:
529 *
530 * dfd = socket(...);
531 * efd1 = epoll_create();
532 * efd2 = epoll_create();
533 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
534 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
535 *
536 * When a packet arrives to the device underneath "dfd", the net code will
537 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
538 * callback wakeup entry on that queue, and the wake_up() performed by the
539 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
540 * (efd1) notices that it may have some event ready, so it needs to wake up
541 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
542 * that ends up in another wake_up(), after having checked about the
543 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
544 * avoid stack blasting.
545 *
546 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
547 * this special case of epoll.
548 */
549 #ifdef CONFIG_DEBUG_LOCK_ALLOC
550
ep_poll_safewake(struct eventpoll * ep,struct epitem * epi,unsigned pollflags)551 static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi,
552 unsigned pollflags)
553 {
554 struct eventpoll *ep_src;
555 unsigned long flags;
556 u8 nests = 0;
557
558 /*
559 * To set the subclass or nesting level for spin_lock_irqsave_nested()
560 * it might be natural to create a per-cpu nest count. However, since
561 * we can recurse on ep->poll_wait.lock, and a non-raw spinlock can
562 * schedule() in the -rt kernel, the per-cpu variable are no longer
563 * protected. Thus, we are introducing a per eventpoll nest field.
564 * If we are not being call from ep_poll_callback(), epi is NULL and
565 * we are at the first level of nesting, 0. Otherwise, we are being
566 * called from ep_poll_callback() and if a previous wakeup source is
567 * not an epoll file itself, we are at depth 1 since the wakeup source
568 * is depth 0. If the wakeup source is a previous epoll file in the
569 * wakeup chain then we use its nests value and record ours as
570 * nests + 1. The previous epoll file nests value is stable since its
571 * already holding its own poll_wait.lock.
572 */
573 if (epi) {
574 if ((is_file_epoll(epi->ffd.file))) {
575 ep_src = epi->ffd.file->private_data;
576 nests = ep_src->nests;
577 } else {
578 nests = 1;
579 }
580 }
581 spin_lock_irqsave_nested(&ep->poll_wait.lock, flags, nests);
582 ep->nests = nests + 1;
583 wake_up_locked_poll(&ep->poll_wait, EPOLLIN | pollflags);
584 ep->nests = 0;
585 spin_unlock_irqrestore(&ep->poll_wait.lock, flags);
586 }
587
588 #else
589
ep_poll_safewake(struct eventpoll * ep,struct epitem * epi,unsigned pollflags)590 static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi,
591 unsigned pollflags)
592 {
593 wake_up_poll(&ep->poll_wait, EPOLLIN | pollflags);
594 }
595
596 #endif
597
ep_remove_wait_queue(struct eppoll_entry * pwq)598 static void ep_remove_wait_queue(struct eppoll_entry *pwq)
599 {
600 wait_queue_head_t *whead;
601
602 rcu_read_lock();
603 /*
604 * If it is cleared by POLLFREE, it should be rcu-safe.
605 * If we read NULL we need a barrier paired with
606 * smp_store_release() in ep_poll_callback(), otherwise
607 * we rely on whead->lock.
608 */
609 whead = smp_load_acquire(&pwq->whead);
610 if (whead)
611 remove_wait_queue(whead, &pwq->wait);
612 rcu_read_unlock();
613 }
614
615 /*
616 * This function unregisters poll callbacks from the associated file
617 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
618 * ep_free).
619 */
ep_unregister_pollwait(struct eventpoll * ep,struct epitem * epi)620 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
621 {
622 struct list_head *lsthead = &epi->pwqlist;
623 struct eppoll_entry *pwq;
624
625 while (!list_empty(lsthead)) {
626 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
627
628 list_del(&pwq->llink);
629 ep_remove_wait_queue(pwq);
630 kmem_cache_free(pwq_cache, pwq);
631 }
632 }
633
634 /* call only when ep->mtx is held */
ep_wakeup_source(struct epitem * epi)635 static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
636 {
637 return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
638 }
639
640 /* call only when ep->mtx is held */
ep_pm_stay_awake(struct epitem * epi)641 static inline void ep_pm_stay_awake(struct epitem *epi)
642 {
643 struct wakeup_source *ws = ep_wakeup_source(epi);
644
645 if (ws)
646 __pm_stay_awake(ws);
647 }
648
ep_has_wakeup_source(struct epitem * epi)649 static inline bool ep_has_wakeup_source(struct epitem *epi)
650 {
651 return rcu_access_pointer(epi->ws) ? true : false;
652 }
653
654 /* call when ep->mtx cannot be held (ep_poll_callback) */
ep_pm_stay_awake_rcu(struct epitem * epi)655 static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
656 {
657 struct wakeup_source *ws;
658
659 rcu_read_lock();
660 ws = rcu_dereference(epi->ws);
661 if (ws)
662 __pm_stay_awake(ws);
663 rcu_read_unlock();
664 }
665
666 /**
667 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
668 * the scan code, to call f_op->poll(). Also allows for
669 * O(NumReady) performance.
670 *
671 * @ep: Pointer to the epoll private data structure.
672 * @sproc: Pointer to the scan callback.
673 * @priv: Private opaque data passed to the @sproc callback.
674 * @depth: The current depth of recursive f_op->poll calls.
675 * @ep_locked: caller already holds ep->mtx
676 *
677 * Returns: The same integer error code returned by the @sproc callback.
678 */
ep_scan_ready_list(struct eventpoll * ep,__poll_t (* sproc)(struct eventpoll *,struct list_head *,void *),void * priv,int depth,bool ep_locked)679 static __poll_t ep_scan_ready_list(struct eventpoll *ep,
680 __poll_t (*sproc)(struct eventpoll *,
681 struct list_head *, void *),
682 void *priv, int depth, bool ep_locked)
683 {
684 __poll_t res;
685 struct epitem *epi, *nepi;
686 LIST_HEAD(txlist);
687
688 lockdep_assert_irqs_enabled();
689
690 /*
691 * We need to lock this because we could be hit by
692 * eventpoll_release_file() and epoll_ctl().
693 */
694
695 if (!ep_locked)
696 mutex_lock_nested(&ep->mtx, depth);
697
698 /*
699 * Steal the ready list, and re-init the original one to the
700 * empty list. Also, set ep->ovflist to NULL so that events
701 * happening while looping w/out locks, are not lost. We cannot
702 * have the poll callback to queue directly on ep->rdllist,
703 * because we want the "sproc" callback to be able to do it
704 * in a lockless way.
705 */
706 write_lock_irq(&ep->lock);
707 list_splice_init(&ep->rdllist, &txlist);
708 WRITE_ONCE(ep->ovflist, NULL);
709 write_unlock_irq(&ep->lock);
710
711 /*
712 * Now call the callback function.
713 */
714 res = (*sproc)(ep, &txlist, priv);
715
716 write_lock_irq(&ep->lock);
717 /*
718 * During the time we spent inside the "sproc" callback, some
719 * other events might have been queued by the poll callback.
720 * We re-insert them inside the main ready-list here.
721 */
722 for (nepi = READ_ONCE(ep->ovflist); (epi = nepi) != NULL;
723 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
724 /*
725 * We need to check if the item is already in the list.
726 * During the "sproc" callback execution time, items are
727 * queued into ->ovflist but the "txlist" might already
728 * contain them, and the list_splice() below takes care of them.
729 */
730 if (!ep_is_linked(epi)) {
731 /*
732 * ->ovflist is LIFO, so we have to reverse it in order
733 * to keep in FIFO.
734 */
735 list_add(&epi->rdllink, &ep->rdllist);
736 ep_pm_stay_awake(epi);
737 }
738 }
739 /*
740 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
741 * releasing the lock, events will be queued in the normal way inside
742 * ep->rdllist.
743 */
744 WRITE_ONCE(ep->ovflist, EP_UNACTIVE_PTR);
745
746 /*
747 * Quickly re-inject items left on "txlist".
748 */
749 list_splice(&txlist, &ep->rdllist);
750 __pm_relax(ep->ws);
751
752 if (!list_empty(&ep->rdllist)) {
753 if (waitqueue_active(&ep->wq))
754 wake_up(&ep->wq);
755 }
756
757 write_unlock_irq(&ep->lock);
758
759 if (!ep_locked)
760 mutex_unlock(&ep->mtx);
761
762 return res;
763 }
764
epi_rcu_free(struct rcu_head * head)765 static void epi_rcu_free(struct rcu_head *head)
766 {
767 struct epitem *epi = container_of(head, struct epitem, rcu);
768 kmem_cache_free(epi_cache, epi);
769 }
770
771 /*
772 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
773 * all the associated resources. Must be called with "mtx" held.
774 */
ep_remove(struct eventpoll * ep,struct epitem * epi)775 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
776 {
777 struct file *file = epi->ffd.file;
778
779 lockdep_assert_irqs_enabled();
780
781 /*
782 * Removes poll wait queue hooks.
783 */
784 ep_unregister_pollwait(ep, epi);
785
786 /* Remove the current item from the list of epoll hooks */
787 spin_lock(&file->f_lock);
788 list_del_rcu(&epi->fllink);
789 spin_unlock(&file->f_lock);
790
791 rb_erase_cached(&epi->rbn, &ep->rbr);
792
793 write_lock_irq(&ep->lock);
794 if (ep_is_linked(epi))
795 list_del_init(&epi->rdllink);
796 write_unlock_irq(&ep->lock);
797
798 wakeup_source_unregister(ep_wakeup_source(epi));
799 /*
800 * At this point it is safe to free the eventpoll item. Use the union
801 * field epi->rcu, since we are trying to minimize the size of
802 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
803 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
804 * use of the rbn field.
805 */
806 call_rcu(&epi->rcu, epi_rcu_free);
807
808 atomic_long_dec(&ep->user->epoll_watches);
809
810 return 0;
811 }
812
ep_free(struct eventpoll * ep)813 static void ep_free(struct eventpoll *ep)
814 {
815 struct rb_node *rbp;
816 struct epitem *epi;
817
818 /* We need to release all tasks waiting for these file */
819 if (waitqueue_active(&ep->poll_wait))
820 ep_poll_safewake(ep, NULL, 0);
821
822 /*
823 * We need to lock this because we could be hit by
824 * eventpoll_release_file() while we're freeing the "struct eventpoll".
825 * We do not need to hold "ep->mtx" here because the epoll file
826 * is on the way to be removed and no one has references to it
827 * anymore. The only hit might come from eventpoll_release_file() but
828 * holding "epmutex" is sufficient here.
829 */
830 mutex_lock(&epmutex);
831
832 /*
833 * Walks through the whole tree by unregistering poll callbacks.
834 */
835 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
836 epi = rb_entry(rbp, struct epitem, rbn);
837
838 ep_unregister_pollwait(ep, epi);
839 cond_resched();
840 }
841
842 /*
843 * Walks through the whole tree by freeing each "struct epitem". At this
844 * point we are sure no poll callbacks will be lingering around, and also by
845 * holding "epmutex" we can be sure that no file cleanup code will hit
846 * us during this operation. So we can avoid the lock on "ep->lock".
847 * We do not need to lock ep->mtx, either, we only do it to prevent
848 * a lockdep warning.
849 */
850 mutex_lock(&ep->mtx);
851 while ((rbp = rb_first_cached(&ep->rbr)) != NULL) {
852 epi = rb_entry(rbp, struct epitem, rbn);
853 ep_remove(ep, epi);
854 cond_resched();
855 }
856 mutex_unlock(&ep->mtx);
857
858 mutex_unlock(&epmutex);
859 mutex_destroy(&ep->mtx);
860 free_uid(ep->user);
861 wakeup_source_unregister(ep->ws);
862 kfree(ep);
863 }
864
ep_eventpoll_release(struct inode * inode,struct file * file)865 static int ep_eventpoll_release(struct inode *inode, struct file *file)
866 {
867 struct eventpoll *ep = file->private_data;
868
869 if (ep)
870 ep_free(ep);
871
872 return 0;
873 }
874
875 static __poll_t ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
876 void *priv);
877 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
878 poll_table *pt);
879
880 /*
881 * Differs from ep_eventpoll_poll() in that internal callers already have
882 * the ep->mtx so we need to start from depth=1, such that mutex_lock_nested()
883 * is correctly annotated.
884 */
ep_item_poll(const struct epitem * epi,poll_table * pt,int depth)885 static __poll_t ep_item_poll(const struct epitem *epi, poll_table *pt,
886 int depth)
887 {
888 struct eventpoll *ep;
889 bool locked;
890
891 pt->_key = epi->event.events;
892 if (!is_file_epoll(epi->ffd.file))
893 return vfs_poll(epi->ffd.file, pt) & epi->event.events;
894
895 ep = epi->ffd.file->private_data;
896 poll_wait(epi->ffd.file, &ep->poll_wait, pt);
897 locked = pt && (pt->_qproc == ep_ptable_queue_proc);
898
899 return ep_scan_ready_list(epi->ffd.file->private_data,
900 ep_read_events_proc, &depth, depth,
901 locked) & epi->event.events;
902 }
903
ep_read_events_proc(struct eventpoll * ep,struct list_head * head,void * priv)904 static __poll_t ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
905 void *priv)
906 {
907 struct epitem *epi, *tmp;
908 poll_table pt;
909 int depth = *(int *)priv;
910
911 init_poll_funcptr(&pt, NULL);
912 depth++;
913
914 list_for_each_entry_safe(epi, tmp, head, rdllink) {
915 if (ep_item_poll(epi, &pt, depth)) {
916 return EPOLLIN | EPOLLRDNORM;
917 } else {
918 /*
919 * Item has been dropped into the ready list by the poll
920 * callback, but it's not actually ready, as far as
921 * caller requested events goes. We can remove it here.
922 */
923 __pm_relax(ep_wakeup_source(epi));
924 list_del_init(&epi->rdllink);
925 }
926 }
927
928 return 0;
929 }
930
ep_eventpoll_poll(struct file * file,poll_table * wait)931 static __poll_t ep_eventpoll_poll(struct file *file, poll_table *wait)
932 {
933 struct eventpoll *ep = file->private_data;
934 int depth = 0;
935
936 /* Insert inside our poll wait queue */
937 poll_wait(file, &ep->poll_wait, wait);
938
939 /*
940 * Proceed to find out if wanted events are really available inside
941 * the ready list.
942 */
943 return ep_scan_ready_list(ep, ep_read_events_proc,
944 &depth, depth, false);
945 }
946
947 #ifdef CONFIG_PROC_FS
ep_show_fdinfo(struct seq_file * m,struct file * f)948 static void ep_show_fdinfo(struct seq_file *m, struct file *f)
949 {
950 struct eventpoll *ep = f->private_data;
951 struct rb_node *rbp;
952
953 mutex_lock(&ep->mtx);
954 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
955 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
956 struct inode *inode = file_inode(epi->ffd.file);
957
958 seq_printf(m, "tfd: %8d events: %8x data: %16llx "
959 " pos:%lli ino:%lx sdev:%x\n",
960 epi->ffd.fd, epi->event.events,
961 (long long)epi->event.data,
962 (long long)epi->ffd.file->f_pos,
963 inode->i_ino, inode->i_sb->s_dev);
964 if (seq_has_overflowed(m))
965 break;
966 }
967 mutex_unlock(&ep->mtx);
968 }
969 #endif
970
971 /* File callbacks that implement the eventpoll file behaviour */
972 static const struct file_operations eventpoll_fops = {
973 #ifdef CONFIG_PROC_FS
974 .show_fdinfo = ep_show_fdinfo,
975 #endif
976 .release = ep_eventpoll_release,
977 .poll = ep_eventpoll_poll,
978 .llseek = noop_llseek,
979 };
980
981 /*
982 * This is called from eventpoll_release() to unlink files from the eventpoll
983 * interface. We need to have this facility to cleanup correctly files that are
984 * closed without being removed from the eventpoll interface.
985 */
eventpoll_release_file(struct file * file)986 void eventpoll_release_file(struct file *file)
987 {
988 struct eventpoll *ep;
989 struct epitem *epi, *next;
990
991 /*
992 * We don't want to get "file->f_lock" because it is not
993 * necessary. It is not necessary because we're in the "struct file"
994 * cleanup path, and this means that no one is using this file anymore.
995 * So, for example, epoll_ctl() cannot hit here since if we reach this
996 * point, the file counter already went to zero and fget() would fail.
997 * The only hit might come from ep_free() but by holding the mutex
998 * will correctly serialize the operation. We do need to acquire
999 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
1000 * from anywhere but ep_free().
1001 *
1002 * Besides, ep_remove() acquires the lock, so we can't hold it here.
1003 */
1004 mutex_lock(&epmutex);
1005 list_for_each_entry_safe(epi, next, &file->f_ep_links, fllink) {
1006 ep = epi->ep;
1007 mutex_lock_nested(&ep->mtx, 0);
1008 ep_remove(ep, epi);
1009 mutex_unlock(&ep->mtx);
1010 }
1011 mutex_unlock(&epmutex);
1012 }
1013
ep_alloc(struct eventpoll ** pep)1014 static int ep_alloc(struct eventpoll **pep)
1015 {
1016 int error;
1017 struct user_struct *user;
1018 struct eventpoll *ep;
1019
1020 user = get_current_user();
1021 error = -ENOMEM;
1022 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
1023 if (unlikely(!ep))
1024 goto free_uid;
1025
1026 mutex_init(&ep->mtx);
1027 rwlock_init(&ep->lock);
1028 init_waitqueue_head(&ep->wq);
1029 init_waitqueue_head(&ep->poll_wait);
1030 INIT_LIST_HEAD(&ep->rdllist);
1031 ep->rbr = RB_ROOT_CACHED;
1032 ep->ovflist = EP_UNACTIVE_PTR;
1033 ep->user = user;
1034
1035 *pep = ep;
1036
1037 return 0;
1038
1039 free_uid:
1040 free_uid(user);
1041 return error;
1042 }
1043
1044 /*
1045 * Search the file inside the eventpoll tree. The RB tree operations
1046 * are protected by the "mtx" mutex, and ep_find() must be called with
1047 * "mtx" held.
1048 */
ep_find(struct eventpoll * ep,struct file * file,int fd)1049 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
1050 {
1051 int kcmp;
1052 struct rb_node *rbp;
1053 struct epitem *epi, *epir = NULL;
1054 struct epoll_filefd ffd;
1055
1056 ep_set_ffd(&ffd, file, fd);
1057 for (rbp = ep->rbr.rb_root.rb_node; rbp; ) {
1058 epi = rb_entry(rbp, struct epitem, rbn);
1059 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
1060 if (kcmp > 0)
1061 rbp = rbp->rb_right;
1062 else if (kcmp < 0)
1063 rbp = rbp->rb_left;
1064 else {
1065 epir = epi;
1066 break;
1067 }
1068 }
1069
1070 return epir;
1071 }
1072
1073 #ifdef CONFIG_KCMP
ep_find_tfd(struct eventpoll * ep,int tfd,unsigned long toff)1074 static struct epitem *ep_find_tfd(struct eventpoll *ep, int tfd, unsigned long toff)
1075 {
1076 struct rb_node *rbp;
1077 struct epitem *epi;
1078
1079 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1080 epi = rb_entry(rbp, struct epitem, rbn);
1081 if (epi->ffd.fd == tfd) {
1082 if (toff == 0)
1083 return epi;
1084 else
1085 toff--;
1086 }
1087 cond_resched();
1088 }
1089
1090 return NULL;
1091 }
1092
get_epoll_tfile_raw_ptr(struct file * file,int tfd,unsigned long toff)1093 struct file *get_epoll_tfile_raw_ptr(struct file *file, int tfd,
1094 unsigned long toff)
1095 {
1096 struct file *file_raw;
1097 struct eventpoll *ep;
1098 struct epitem *epi;
1099
1100 if (!is_file_epoll(file))
1101 return ERR_PTR(-EINVAL);
1102
1103 ep = file->private_data;
1104
1105 mutex_lock(&ep->mtx);
1106 epi = ep_find_tfd(ep, tfd, toff);
1107 if (epi)
1108 file_raw = epi->ffd.file;
1109 else
1110 file_raw = ERR_PTR(-ENOENT);
1111 mutex_unlock(&ep->mtx);
1112
1113 return file_raw;
1114 }
1115 #endif /* CONFIG_KCMP */
1116
1117 /**
1118 * Adds a new entry to the tail of the list in a lockless way, i.e.
1119 * multiple CPUs are allowed to call this function concurrently.
1120 *
1121 * Beware: it is necessary to prevent any other modifications of the
1122 * existing list until all changes are completed, in other words
1123 * concurrent list_add_tail_lockless() calls should be protected
1124 * with a read lock, where write lock acts as a barrier which
1125 * makes sure all list_add_tail_lockless() calls are fully
1126 * completed.
1127 *
1128 * Also an element can be locklessly added to the list only in one
1129 * direction i.e. either to the tail either to the head, otherwise
1130 * concurrent access will corrupt the list.
1131 *
1132 * Returns %false if element has been already added to the list, %true
1133 * otherwise.
1134 */
list_add_tail_lockless(struct list_head * new,struct list_head * head)1135 static inline bool list_add_tail_lockless(struct list_head *new,
1136 struct list_head *head)
1137 {
1138 struct list_head *prev;
1139
1140 /*
1141 * This is simple 'new->next = head' operation, but cmpxchg()
1142 * is used in order to detect that same element has been just
1143 * added to the list from another CPU: the winner observes
1144 * new->next == new.
1145 */
1146 if (cmpxchg(&new->next, new, head) != new)
1147 return false;
1148
1149 /*
1150 * Initially ->next of a new element must be updated with the head
1151 * (we are inserting to the tail) and only then pointers are atomically
1152 * exchanged. XCHG guarantees memory ordering, thus ->next should be
1153 * updated before pointers are actually swapped and pointers are
1154 * swapped before prev->next is updated.
1155 */
1156
1157 prev = xchg(&head->prev, new);
1158
1159 /*
1160 * It is safe to modify prev->next and new->prev, because a new element
1161 * is added only to the tail and new->next is updated before XCHG.
1162 */
1163
1164 prev->next = new;
1165 new->prev = prev;
1166
1167 return true;
1168 }
1169
1170 /**
1171 * Chains a new epi entry to the tail of the ep->ovflist in a lockless way,
1172 * i.e. multiple CPUs are allowed to call this function concurrently.
1173 *
1174 * Returns %false if epi element has been already chained, %true otherwise.
1175 */
chain_epi_lockless(struct epitem * epi)1176 static inline bool chain_epi_lockless(struct epitem *epi)
1177 {
1178 struct eventpoll *ep = epi->ep;
1179
1180 /* Fast preliminary check */
1181 if (epi->next != EP_UNACTIVE_PTR)
1182 return false;
1183
1184 /* Check that the same epi has not been just chained from another CPU */
1185 if (cmpxchg(&epi->next, EP_UNACTIVE_PTR, NULL) != EP_UNACTIVE_PTR)
1186 return false;
1187
1188 /* Atomically exchange tail */
1189 epi->next = xchg(&ep->ovflist, epi);
1190
1191 return true;
1192 }
1193
1194 /*
1195 * This is the callback that is passed to the wait queue wakeup
1196 * mechanism. It is called by the stored file descriptors when they
1197 * have events to report.
1198 *
1199 * This callback takes a read lock in order not to content with concurrent
1200 * events from another file descriptors, thus all modifications to ->rdllist
1201 * or ->ovflist are lockless. Read lock is paired with the write lock from
1202 * ep_scan_ready_list(), which stops all list modifications and guarantees
1203 * that lists state is seen correctly.
1204 *
1205 * Another thing worth to mention is that ep_poll_callback() can be called
1206 * concurrently for the same @epi from different CPUs if poll table was inited
1207 * with several wait queues entries. Plural wakeup from different CPUs of a
1208 * single wait queue is serialized by wq.lock, but the case when multiple wait
1209 * queues are used should be detected accordingly. This is detected using
1210 * cmpxchg() operation.
1211 */
ep_poll_callback(wait_queue_entry_t * wait,unsigned mode,int sync,void * key)1212 static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1213 {
1214 int pwake = 0;
1215 struct epitem *epi = ep_item_from_wait(wait);
1216 struct eventpoll *ep = epi->ep;
1217 __poll_t pollflags = key_to_poll(key);
1218 unsigned long flags;
1219 int ewake = 0;
1220
1221 read_lock_irqsave(&ep->lock, flags);
1222
1223 ep_set_busy_poll_napi_id(epi);
1224
1225 /*
1226 * If the event mask does not contain any poll(2) event, we consider the
1227 * descriptor to be disabled. This condition is likely the effect of the
1228 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1229 * until the next EPOLL_CTL_MOD will be issued.
1230 */
1231 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1232 goto out_unlock;
1233
1234 /*
1235 * Check the events coming with the callback. At this stage, not
1236 * every device reports the events in the "key" parameter of the
1237 * callback. We need to be able to handle both cases here, hence the
1238 * test for "key" != NULL before the event match test.
1239 */
1240 if (pollflags && !(pollflags & epi->event.events))
1241 goto out_unlock;
1242
1243 /*
1244 * If we are transferring events to userspace, we can hold no locks
1245 * (because we're accessing user memory, and because of linux f_op->poll()
1246 * semantics). All the events that happen during that period of time are
1247 * chained in ep->ovflist and requeued later on.
1248 */
1249 if (READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR) {
1250 if (chain_epi_lockless(epi))
1251 ep_pm_stay_awake_rcu(epi);
1252 } else if (!ep_is_linked(epi)) {
1253 /* In the usual case, add event to ready list. */
1254 if (list_add_tail_lockless(&epi->rdllink, &ep->rdllist))
1255 ep_pm_stay_awake_rcu(epi);
1256 }
1257
1258 /*
1259 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1260 * wait list.
1261 */
1262 if (waitqueue_active(&ep->wq)) {
1263 if ((epi->event.events & EPOLLEXCLUSIVE) &&
1264 !(pollflags & POLLFREE)) {
1265 switch (pollflags & EPOLLINOUT_BITS) {
1266 case EPOLLIN:
1267 if (epi->event.events & EPOLLIN)
1268 ewake = 1;
1269 break;
1270 case EPOLLOUT:
1271 if (epi->event.events & EPOLLOUT)
1272 ewake = 1;
1273 break;
1274 case 0:
1275 ewake = 1;
1276 break;
1277 }
1278 }
1279 wake_up(&ep->wq);
1280 }
1281 if (waitqueue_active(&ep->poll_wait))
1282 pwake++;
1283
1284 out_unlock:
1285 read_unlock_irqrestore(&ep->lock, flags);
1286
1287 /* We have to call this outside the lock */
1288 if (pwake)
1289 ep_poll_safewake(ep, epi, pollflags & EPOLL_URING_WAKE);
1290
1291 if (!(epi->event.events & EPOLLEXCLUSIVE))
1292 ewake = 1;
1293
1294 if (pollflags & POLLFREE) {
1295 /*
1296 * If we race with ep_remove_wait_queue() it can miss
1297 * ->whead = NULL and do another remove_wait_queue() after
1298 * us, so we can't use __remove_wait_queue().
1299 */
1300 list_del_init(&wait->entry);
1301 /*
1302 * ->whead != NULL protects us from the race with ep_free()
1303 * or ep_remove(), ep_remove_wait_queue() takes whead->lock
1304 * held by the caller. Once we nullify it, nothing protects
1305 * ep/epi or even wait.
1306 */
1307 smp_store_release(&ep_pwq_from_wait(wait)->whead, NULL);
1308 }
1309
1310 return ewake;
1311 }
1312
1313 /*
1314 * This is the callback that is used to add our wait queue to the
1315 * target file wakeup lists.
1316 */
ep_ptable_queue_proc(struct file * file,wait_queue_head_t * whead,poll_table * pt)1317 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1318 poll_table *pt)
1319 {
1320 struct epitem *epi = ep_item_from_epqueue(pt);
1321 struct eppoll_entry *pwq;
1322
1323 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
1324 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1325 pwq->whead = whead;
1326 pwq->base = epi;
1327 if (epi->event.events & EPOLLEXCLUSIVE)
1328 add_wait_queue_exclusive(whead, &pwq->wait);
1329 else
1330 add_wait_queue(whead, &pwq->wait);
1331 list_add_tail(&pwq->llink, &epi->pwqlist);
1332 epi->nwait++;
1333 } else {
1334 /* We have to signal that an error occurred */
1335 epi->nwait = -1;
1336 }
1337 }
1338
ep_rbtree_insert(struct eventpoll * ep,struct epitem * epi)1339 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1340 {
1341 int kcmp;
1342 struct rb_node **p = &ep->rbr.rb_root.rb_node, *parent = NULL;
1343 struct epitem *epic;
1344 bool leftmost = true;
1345
1346 while (*p) {
1347 parent = *p;
1348 epic = rb_entry(parent, struct epitem, rbn);
1349 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1350 if (kcmp > 0) {
1351 p = &parent->rb_right;
1352 leftmost = false;
1353 } else
1354 p = &parent->rb_left;
1355 }
1356 rb_link_node(&epi->rbn, parent, p);
1357 rb_insert_color_cached(&epi->rbn, &ep->rbr, leftmost);
1358 }
1359
1360
1361
1362 #define PATH_ARR_SIZE 5
1363 /*
1364 * These are the number paths of length 1 to 5, that we are allowing to emanate
1365 * from a single file of interest. For example, we allow 1000 paths of length
1366 * 1, to emanate from each file of interest. This essentially represents the
1367 * potential wakeup paths, which need to be limited in order to avoid massive
1368 * uncontrolled wakeup storms. The common use case should be a single ep which
1369 * is connected to n file sources. In this case each file source has 1 path
1370 * of length 1. Thus, the numbers below should be more than sufficient. These
1371 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1372 * and delete can't add additional paths. Protected by the epmutex.
1373 */
1374 static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1375 static int path_count[PATH_ARR_SIZE];
1376
path_count_inc(int nests)1377 static int path_count_inc(int nests)
1378 {
1379 /* Allow an arbitrary number of depth 1 paths */
1380 if (nests == 0)
1381 return 0;
1382
1383 if (++path_count[nests] > path_limits[nests])
1384 return -1;
1385 return 0;
1386 }
1387
path_count_init(void)1388 static void path_count_init(void)
1389 {
1390 int i;
1391
1392 for (i = 0; i < PATH_ARR_SIZE; i++)
1393 path_count[i] = 0;
1394 }
1395
reverse_path_check_proc(void * priv,void * cookie,int call_nests)1396 static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1397 {
1398 int error = 0;
1399 struct file *file = priv;
1400 struct file *child_file;
1401 struct epitem *epi;
1402
1403 /* CTL_DEL can remove links here, but that can't increase our count */
1404 rcu_read_lock();
1405 list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) {
1406 child_file = epi->ep->file;
1407 if (is_file_epoll(child_file)) {
1408 if (list_empty(&child_file->f_ep_links)) {
1409 if (path_count_inc(call_nests)) {
1410 error = -1;
1411 break;
1412 }
1413 } else {
1414 error = ep_call_nested(&poll_loop_ncalls,
1415 reverse_path_check_proc,
1416 child_file, child_file,
1417 current);
1418 }
1419 if (error != 0)
1420 break;
1421 } else {
1422 printk(KERN_ERR "reverse_path_check_proc: "
1423 "file is not an ep!\n");
1424 }
1425 }
1426 rcu_read_unlock();
1427 return error;
1428 }
1429
1430 /**
1431 * reverse_path_check - The tfile_check_list is list of file *, which have
1432 * links that are proposed to be newly added. We need to
1433 * make sure that those added links don't add too many
1434 * paths such that we will spend all our time waking up
1435 * eventpoll objects.
1436 *
1437 * Returns: Returns zero if the proposed links don't create too many paths,
1438 * -1 otherwise.
1439 */
reverse_path_check(void)1440 static int reverse_path_check(void)
1441 {
1442 int error = 0;
1443 struct file *current_file;
1444
1445 /* let's call this for all tfiles */
1446 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
1447 path_count_init();
1448 error = ep_call_nested(&poll_loop_ncalls,
1449 reverse_path_check_proc, current_file,
1450 current_file, current);
1451 if (error)
1452 break;
1453 }
1454 return error;
1455 }
1456
ep_create_wakeup_source(struct epitem * epi)1457 static int ep_create_wakeup_source(struct epitem *epi)
1458 {
1459 struct name_snapshot n;
1460 struct wakeup_source *ws;
1461
1462 if (!epi->ep->ws) {
1463 epi->ep->ws = wakeup_source_register(NULL, "eventpoll");
1464 if (!epi->ep->ws)
1465 return -ENOMEM;
1466 }
1467
1468 take_dentry_name_snapshot(&n, epi->ffd.file->f_path.dentry);
1469 ws = wakeup_source_register(NULL, n.name.name);
1470 release_dentry_name_snapshot(&n);
1471
1472 if (!ws)
1473 return -ENOMEM;
1474 rcu_assign_pointer(epi->ws, ws);
1475
1476 return 0;
1477 }
1478
1479 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
ep_destroy_wakeup_source(struct epitem * epi)1480 static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1481 {
1482 struct wakeup_source *ws = ep_wakeup_source(epi);
1483
1484 RCU_INIT_POINTER(epi->ws, NULL);
1485
1486 /*
1487 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1488 * used internally by wakeup_source_remove, too (called by
1489 * wakeup_source_unregister), so we cannot use call_rcu
1490 */
1491 synchronize_rcu();
1492 wakeup_source_unregister(ws);
1493 }
1494
1495 /*
1496 * Must be called with "mtx" held.
1497 */
ep_insert(struct eventpoll * ep,const struct epoll_event * event,struct file * tfile,int fd,int full_check)1498 static int ep_insert(struct eventpoll *ep, const struct epoll_event *event,
1499 struct file *tfile, int fd, int full_check)
1500 {
1501 int error, pwake = 0;
1502 __poll_t revents;
1503 long user_watches;
1504 struct epitem *epi;
1505 struct ep_pqueue epq;
1506
1507 lockdep_assert_irqs_enabled();
1508
1509 user_watches = atomic_long_read(&ep->user->epoll_watches);
1510 if (unlikely(user_watches >= max_user_watches))
1511 return -ENOSPC;
1512 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1513 return -ENOMEM;
1514
1515 /* Item initialization follow here ... */
1516 INIT_LIST_HEAD(&epi->rdllink);
1517 INIT_LIST_HEAD(&epi->fllink);
1518 INIT_LIST_HEAD(&epi->pwqlist);
1519 epi->ep = ep;
1520 ep_set_ffd(&epi->ffd, tfile, fd);
1521 epi->event = *event;
1522 epi->nwait = 0;
1523 epi->next = EP_UNACTIVE_PTR;
1524 if (epi->event.events & EPOLLWAKEUP) {
1525 error = ep_create_wakeup_source(epi);
1526 if (error)
1527 goto error_create_wakeup_source;
1528 } else {
1529 RCU_INIT_POINTER(epi->ws, NULL);
1530 }
1531
1532 /* Add the current item to the list of active epoll hook for this file */
1533 spin_lock(&tfile->f_lock);
1534 list_add_tail_rcu(&epi->fllink, &tfile->f_ep_links);
1535 spin_unlock(&tfile->f_lock);
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
1543 /* now check if we've created too many backpaths */
1544 error = -EINVAL;
1545 if (full_check && reverse_path_check())
1546 goto error_remove_epi;
1547
1548 /* Initialize the poll table using the queue callback */
1549 epq.epi = epi;
1550 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1551
1552 /*
1553 * Attach the item to the poll hooks and get current event bits.
1554 * We can safely use the file* here because its usage count has
1555 * been increased by the caller of this function. Note that after
1556 * this operation completes, the poll callback can start hitting
1557 * the new item.
1558 */
1559 revents = ep_item_poll(epi, &epq.pt, 1);
1560
1561 /*
1562 * We have to check if something went wrong during the poll wait queue
1563 * install process. Namely an allocation for a wait queue failed due
1564 * high memory pressure.
1565 */
1566 error = -ENOMEM;
1567 if (epi->nwait < 0)
1568 goto error_unregister;
1569
1570 /* We have to drop the new item inside our item list to keep track of it */
1571 write_lock_irq(&ep->lock);
1572
1573 /* record NAPI ID of new item if present */
1574 ep_set_busy_poll_napi_id(epi);
1575
1576 /* If the file is already "ready" we drop it inside the ready list */
1577 if (revents && !ep_is_linked(epi)) {
1578 list_add_tail(&epi->rdllink, &ep->rdllist);
1579 ep_pm_stay_awake(epi);
1580
1581 /* Notify waiting tasks that events are available */
1582 if (waitqueue_active(&ep->wq))
1583 wake_up(&ep->wq);
1584 if (waitqueue_active(&ep->poll_wait))
1585 pwake++;
1586 }
1587
1588 write_unlock_irq(&ep->lock);
1589
1590 atomic_long_inc(&ep->user->epoll_watches);
1591
1592 /* We have to call this outside the lock */
1593 if (pwake)
1594 ep_poll_safewake(ep, NULL, 0);
1595
1596 return 0;
1597
1598 error_unregister:
1599 ep_unregister_pollwait(ep, epi);
1600 error_remove_epi:
1601 spin_lock(&tfile->f_lock);
1602 list_del_rcu(&epi->fllink);
1603 spin_unlock(&tfile->f_lock);
1604
1605 rb_erase_cached(&epi->rbn, &ep->rbr);
1606
1607 /*
1608 * We need to do this because an event could have been arrived on some
1609 * allocated wait queue. Note that we don't care about the ep->ovflist
1610 * list, since that is used/cleaned only inside a section bound by "mtx".
1611 * And ep_insert() is called with "mtx" held.
1612 */
1613 write_lock_irq(&ep->lock);
1614 if (ep_is_linked(epi))
1615 list_del_init(&epi->rdllink);
1616 write_unlock_irq(&ep->lock);
1617
1618 wakeup_source_unregister(ep_wakeup_source(epi));
1619
1620 error_create_wakeup_source:
1621 kmem_cache_free(epi_cache, epi);
1622
1623 return error;
1624 }
1625
1626 /*
1627 * Modify the interest event mask by dropping an event if the new mask
1628 * has a match in the current file status. Must be called with "mtx" held.
1629 */
ep_modify(struct eventpoll * ep,struct epitem * epi,const struct epoll_event * event)1630 static int ep_modify(struct eventpoll *ep, struct epitem *epi,
1631 const struct epoll_event *event)
1632 {
1633 int pwake = 0;
1634 poll_table pt;
1635
1636 lockdep_assert_irqs_enabled();
1637
1638 init_poll_funcptr(&pt, NULL);
1639
1640 /*
1641 * Set the new event interest mask before calling f_op->poll();
1642 * otherwise we might miss an event that happens between the
1643 * f_op->poll() call and the new event set registering.
1644 */
1645 epi->event.events = event->events; /* need barrier below */
1646 epi->event.data = event->data; /* protected by mtx */
1647 if (epi->event.events & EPOLLWAKEUP) {
1648 if (!ep_has_wakeup_source(epi))
1649 ep_create_wakeup_source(epi);
1650 } else if (ep_has_wakeup_source(epi)) {
1651 ep_destroy_wakeup_source(epi);
1652 }
1653
1654 /*
1655 * The following barrier has two effects:
1656 *
1657 * 1) Flush epi changes above to other CPUs. This ensures
1658 * we do not miss events from ep_poll_callback if an
1659 * event occurs immediately after we call f_op->poll().
1660 * We need this because we did not take ep->lock while
1661 * changing epi above (but ep_poll_callback does take
1662 * ep->lock).
1663 *
1664 * 2) We also need to ensure we do not miss _past_ events
1665 * when calling f_op->poll(). This barrier also
1666 * pairs with the barrier in wq_has_sleeper (see
1667 * comments for wq_has_sleeper).
1668 *
1669 * This barrier will now guarantee ep_poll_callback or f_op->poll
1670 * (or both) will notice the readiness of an item.
1671 */
1672 smp_mb();
1673
1674 /*
1675 * Get current event bits. We can safely use the file* here because
1676 * its usage count has been increased by the caller of this function.
1677 * If the item is "hot" and it is not registered inside the ready
1678 * list, push it inside.
1679 */
1680 if (ep_item_poll(epi, &pt, 1)) {
1681 write_lock_irq(&ep->lock);
1682 if (!ep_is_linked(epi)) {
1683 list_add_tail(&epi->rdllink, &ep->rdllist);
1684 ep_pm_stay_awake(epi);
1685
1686 /* Notify waiting tasks that events are available */
1687 if (waitqueue_active(&ep->wq))
1688 wake_up(&ep->wq);
1689 if (waitqueue_active(&ep->poll_wait))
1690 pwake++;
1691 }
1692 write_unlock_irq(&ep->lock);
1693 }
1694
1695 /* We have to call this outside the lock */
1696 if (pwake)
1697 ep_poll_safewake(ep, NULL, 0);
1698
1699 return 0;
1700 }
1701
ep_send_events_proc(struct eventpoll * ep,struct list_head * head,void * priv)1702 static __poll_t ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1703 void *priv)
1704 {
1705 struct ep_send_events_data *esed = priv;
1706 __poll_t revents;
1707 struct epitem *epi, *tmp;
1708 struct epoll_event __user *uevent = esed->events;
1709 struct wakeup_source *ws;
1710 poll_table pt;
1711
1712 init_poll_funcptr(&pt, NULL);
1713 esed->res = 0;
1714
1715 /*
1716 * We can loop without lock because we are passed a task private list.
1717 * Items cannot vanish during the loop because ep_scan_ready_list() is
1718 * holding "mtx" during this call.
1719 */
1720 lockdep_assert_held(&ep->mtx);
1721
1722 list_for_each_entry_safe(epi, tmp, head, rdllink) {
1723 if (esed->res >= esed->maxevents)
1724 break;
1725
1726 /*
1727 * Activate ep->ws before deactivating epi->ws to prevent
1728 * triggering auto-suspend here (in case we reactive epi->ws
1729 * below).
1730 *
1731 * This could be rearranged to delay the deactivation of epi->ws
1732 * instead, but then epi->ws would temporarily be out of sync
1733 * with ep_is_linked().
1734 */
1735 ws = ep_wakeup_source(epi);
1736 if (ws) {
1737 if (ws->active)
1738 __pm_stay_awake(ep->ws);
1739 __pm_relax(ws);
1740 }
1741
1742 list_del_init(&epi->rdllink);
1743
1744 /*
1745 * If the event mask intersect the caller-requested one,
1746 * deliver the event to userspace. Again, ep_scan_ready_list()
1747 * is holding ep->mtx, so no operations coming from userspace
1748 * can change the item.
1749 */
1750 revents = ep_item_poll(epi, &pt, 1);
1751 if (!revents)
1752 continue;
1753
1754 if (__put_user(revents, &uevent->events) ||
1755 __put_user(epi->event.data, &uevent->data)) {
1756 list_add(&epi->rdllink, head);
1757 ep_pm_stay_awake(epi);
1758 if (!esed->res)
1759 esed->res = -EFAULT;
1760 return 0;
1761 }
1762 esed->res++;
1763 uevent++;
1764 if (epi->event.events & EPOLLONESHOT)
1765 epi->event.events &= EP_PRIVATE_BITS;
1766 else if (!(epi->event.events & EPOLLET)) {
1767 /*
1768 * If this file has been added with Level
1769 * Trigger mode, we need to insert back inside
1770 * the ready list, so that the next call to
1771 * epoll_wait() will check again the events
1772 * availability. At this point, no one can insert
1773 * into ep->rdllist besides us. The epoll_ctl()
1774 * callers are locked out by
1775 * ep_scan_ready_list() holding "mtx" and the
1776 * poll callback will queue them in ep->ovflist.
1777 */
1778 list_add_tail(&epi->rdllink, &ep->rdllist);
1779 ep_pm_stay_awake(epi);
1780 }
1781 }
1782
1783 return 0;
1784 }
1785
ep_send_events(struct eventpoll * ep,struct epoll_event __user * events,int maxevents)1786 static int ep_send_events(struct eventpoll *ep,
1787 struct epoll_event __user *events, int maxevents)
1788 {
1789 struct ep_send_events_data esed;
1790
1791 esed.maxevents = maxevents;
1792 esed.events = events;
1793
1794 ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false);
1795 return esed.res;
1796 }
1797
ep_set_mstimeout(long ms)1798 static inline struct timespec64 ep_set_mstimeout(long ms)
1799 {
1800 struct timespec64 now, ts = {
1801 .tv_sec = ms / MSEC_PER_SEC,
1802 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1803 };
1804
1805 ktime_get_ts64(&now);
1806 return timespec64_add_safe(now, ts);
1807 }
1808
1809 /*
1810 * autoremove_wake_function, but remove even on failure to wake up, because we
1811 * know that default_wake_function/ttwu will only fail if the thread is already
1812 * woken, and in that case the ep_poll loop will remove the entry anyways, not
1813 * try to reuse it.
1814 */
ep_autoremove_wake_function(struct wait_queue_entry * wq_entry,unsigned int mode,int sync,void * key)1815 static int ep_autoremove_wake_function(struct wait_queue_entry *wq_entry,
1816 unsigned int mode, int sync, void *key)
1817 {
1818 int ret = default_wake_function(wq_entry, mode, sync, key);
1819
1820 list_del_init(&wq_entry->entry);
1821 return ret;
1822 }
1823
1824 /**
1825 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1826 * event buffer.
1827 *
1828 * @ep: Pointer to the eventpoll context.
1829 * @events: Pointer to the userspace buffer where the ready events should be
1830 * stored.
1831 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1832 * @timeout: Maximum timeout for the ready events fetch operation, in
1833 * milliseconds. If the @timeout is zero, the function will not block,
1834 * while if the @timeout is less than zero, the function will block
1835 * until at least one event has been retrieved (or an error
1836 * occurred).
1837 *
1838 * Returns: Returns the number of ready events which have been fetched, or an
1839 * error code, in case of error.
1840 */
ep_poll(struct eventpoll * ep,struct epoll_event __user * events,int maxevents,long timeout)1841 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1842 int maxevents, long timeout)
1843 {
1844 int res = 0, eavail, timed_out = 0;
1845 u64 slack = 0;
1846 wait_queue_entry_t wait;
1847 ktime_t expires, *to = NULL;
1848
1849 lockdep_assert_irqs_enabled();
1850
1851 if (timeout > 0) {
1852 struct timespec64 end_time = ep_set_mstimeout(timeout);
1853
1854 slack = select_estimate_accuracy(&end_time);
1855 to = &expires;
1856 *to = timespec64_to_ktime(end_time);
1857 } else if (timeout == 0) {
1858 /*
1859 * Avoid the unnecessary trip to the wait queue loop, if the
1860 * caller specified a non blocking operation. We still need
1861 * lock because we could race and not see an epi being added
1862 * to the ready list while in irq callback. Thus incorrectly
1863 * returning 0 back to userspace.
1864 */
1865 timed_out = 1;
1866
1867 write_lock_irq(&ep->lock);
1868 eavail = ep_events_available(ep);
1869 write_unlock_irq(&ep->lock);
1870
1871 goto send_events;
1872 }
1873
1874 fetch_events:
1875
1876 if (!ep_events_available(ep))
1877 ep_busy_loop(ep, timed_out);
1878
1879 eavail = ep_events_available(ep);
1880 if (eavail)
1881 goto send_events;
1882
1883 /*
1884 * Busy poll timed out. Drop NAPI ID for now, we can add
1885 * it back in when we have moved a socket with a valid NAPI
1886 * ID onto the ready list.
1887 */
1888 ep_reset_busy_poll_napi_id(ep);
1889
1890 do {
1891 /*
1892 * Internally init_wait() uses autoremove_wake_function(),
1893 * thus wait entry is removed from the wait queue on each
1894 * wakeup. Why it is important? In case of several waiters
1895 * each new wakeup will hit the next waiter, giving it the
1896 * chance to harvest new event. Otherwise wakeup can be
1897 * lost. This is also good performance-wise, because on
1898 * normal wakeup path no need to call __remove_wait_queue()
1899 * explicitly, thus ep->lock is not taken, which halts the
1900 * event delivery.
1901 *
1902 * In fact, we now use an even more aggressive function that
1903 * unconditionally removes, because we don't reuse the wait
1904 * entry between loop iterations. This lets us also avoid the
1905 * performance issue if a process is killed, causing all of its
1906 * threads to wake up without being removed normally.
1907 */
1908 init_wait(&wait);
1909 wait.func = ep_autoremove_wake_function;
1910
1911 write_lock_irq(&ep->lock);
1912 /*
1913 * Barrierless variant, waitqueue_active() is called under
1914 * the same lock on wakeup ep_poll_callback() side, so it
1915 * is safe to avoid an explicit barrier.
1916 */
1917 __set_current_state(TASK_INTERRUPTIBLE);
1918
1919 /*
1920 * Do the final check under the lock. ep_scan_ready_list()
1921 * plays with two lists (->rdllist and ->ovflist) and there
1922 * is always a race when both lists are empty for short
1923 * period of time although events are pending, so lock is
1924 * important.
1925 */
1926 eavail = ep_events_available(ep);
1927 if (!eavail) {
1928 if (signal_pending(current))
1929 res = -EINTR;
1930 else
1931 __add_wait_queue_exclusive(&ep->wq, &wait);
1932 }
1933 write_unlock_irq(&ep->lock);
1934
1935 if (!eavail && !res)
1936 timed_out = !schedule_hrtimeout_range(to, slack,
1937 HRTIMER_MODE_ABS);
1938
1939 /*
1940 * We were woken up, thus go and try to harvest some events.
1941 * If timed out and still on the wait queue, recheck eavail
1942 * carefully under lock, below.
1943 */
1944 eavail = 1;
1945 } while (0);
1946
1947 __set_current_state(TASK_RUNNING);
1948
1949 if (!list_empty_careful(&wait.entry)) {
1950 write_lock_irq(&ep->lock);
1951 /*
1952 * If the thread timed out and is not on the wait queue, it
1953 * means that the thread was woken up after its timeout expired
1954 * before it could reacquire the lock. Thus, when wait.entry is
1955 * empty, it needs to harvest events.
1956 */
1957 if (timed_out)
1958 eavail = list_empty(&wait.entry);
1959 __remove_wait_queue(&ep->wq, &wait);
1960 write_unlock_irq(&ep->lock);
1961 }
1962
1963 send_events:
1964 if (fatal_signal_pending(current)) {
1965 /*
1966 * Always short-circuit for fatal signals to allow
1967 * threads to make a timely exit without the chance of
1968 * finding more events available and fetching
1969 * repeatedly.
1970 */
1971 res = -EINTR;
1972 }
1973 /*
1974 * Try to transfer events to user space. In case we get 0 events and
1975 * there's still timeout left over, we go trying again in search of
1976 * more luck.
1977 */
1978 if (!res && eavail &&
1979 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1980 goto fetch_events;
1981
1982 return res;
1983 }
1984
1985 /**
1986 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1987 * API, to verify that adding an epoll file inside another
1988 * epoll structure, does not violate the constraints, in
1989 * terms of closed loops, or too deep chains (which can
1990 * result in excessive stack usage).
1991 *
1992 * @priv: Pointer to the epoll file to be currently checked.
1993 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1994 * data structure pointer.
1995 * @call_nests: Current dept of the @ep_call_nested() call stack.
1996 *
1997 * Returns: Returns zero if adding the epoll @file inside current epoll
1998 * structure @ep does not violate the constraints, or -1 otherwise.
1999 */
ep_loop_check_proc(void * priv,void * cookie,int call_nests)2000 static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
2001 {
2002 int error = 0;
2003 struct file *file = priv;
2004 struct eventpoll *ep = file->private_data;
2005 struct eventpoll *ep_tovisit;
2006 struct rb_node *rbp;
2007 struct epitem *epi;
2008
2009 mutex_lock_nested(&ep->mtx, call_nests + 1);
2010 ep->gen = loop_check_gen;
2011 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
2012 epi = rb_entry(rbp, struct epitem, rbn);
2013 if (unlikely(is_file_epoll(epi->ffd.file))) {
2014 ep_tovisit = epi->ffd.file->private_data;
2015 if (ep_tovisit->gen == loop_check_gen)
2016 continue;
2017 error = ep_call_nested(&poll_loop_ncalls,
2018 ep_loop_check_proc, epi->ffd.file,
2019 ep_tovisit, current);
2020 if (error != 0)
2021 break;
2022 } else {
2023 /*
2024 * If we've reached a file that is not associated with
2025 * an ep, then we need to check if the newly added
2026 * links are going to add too many wakeup paths. We do
2027 * this by adding it to the tfile_check_list, if it's
2028 * not already there, and calling reverse_path_check()
2029 * during ep_insert().
2030 */
2031 if (list_empty(&epi->ffd.file->f_tfile_llink)) {
2032 if (get_file_rcu(epi->ffd.file))
2033 list_add(&epi->ffd.file->f_tfile_llink,
2034 &tfile_check_list);
2035 }
2036 }
2037 }
2038 mutex_unlock(&ep->mtx);
2039
2040 return error;
2041 }
2042
2043 /**
2044 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
2045 * another epoll file (represented by @ep) does not create
2046 * closed loops or too deep chains.
2047 *
2048 * @ep: Pointer to the epoll private data structure.
2049 * @file: Pointer to the epoll file to be checked.
2050 *
2051 * Returns: Returns zero if adding the epoll @file inside current epoll
2052 * structure @ep does not violate the constraints, or -1 otherwise.
2053 */
ep_loop_check(struct eventpoll * ep,struct file * file)2054 static int ep_loop_check(struct eventpoll *ep, struct file *file)
2055 {
2056 return ep_call_nested(&poll_loop_ncalls,
2057 ep_loop_check_proc, file, ep, current);
2058 }
2059
clear_tfile_check_list(void)2060 static void clear_tfile_check_list(void)
2061 {
2062 struct file *file;
2063
2064 /* first clear the tfile_check_list */
2065 while (!list_empty(&tfile_check_list)) {
2066 file = list_first_entry(&tfile_check_list, struct file,
2067 f_tfile_llink);
2068 list_del_init(&file->f_tfile_llink);
2069 fput(file);
2070 }
2071 INIT_LIST_HEAD(&tfile_check_list);
2072 }
2073
2074 /*
2075 * Open an eventpoll file descriptor.
2076 */
do_epoll_create(int flags)2077 static int do_epoll_create(int flags)
2078 {
2079 int error, fd;
2080 struct eventpoll *ep = NULL;
2081 struct file *file;
2082
2083 /* Check the EPOLL_* constant for consistency. */
2084 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
2085
2086 if (flags & ~EPOLL_CLOEXEC)
2087 return -EINVAL;
2088 /*
2089 * Create the internal data structure ("struct eventpoll").
2090 */
2091 error = ep_alloc(&ep);
2092 if (error < 0)
2093 return error;
2094 /*
2095 * Creates all the items needed to setup an eventpoll file. That is,
2096 * a file structure and a free file descriptor.
2097 */
2098 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
2099 if (fd < 0) {
2100 error = fd;
2101 goto out_free_ep;
2102 }
2103 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
2104 O_RDWR | (flags & O_CLOEXEC));
2105 if (IS_ERR(file)) {
2106 error = PTR_ERR(file);
2107 goto out_free_fd;
2108 }
2109 ep->file = file;
2110 fd_install(fd, file);
2111 return fd;
2112
2113 out_free_fd:
2114 put_unused_fd(fd);
2115 out_free_ep:
2116 ep_free(ep);
2117 return error;
2118 }
2119
SYSCALL_DEFINE1(epoll_create1,int,flags)2120 SYSCALL_DEFINE1(epoll_create1, int, flags)
2121 {
2122 return do_epoll_create(flags);
2123 }
2124
SYSCALL_DEFINE1(epoll_create,int,size)2125 SYSCALL_DEFINE1(epoll_create, int, size)
2126 {
2127 if (size <= 0)
2128 return -EINVAL;
2129
2130 return do_epoll_create(0);
2131 }
2132
epoll_mutex_lock(struct mutex * mutex,int depth,bool nonblock)2133 static inline int epoll_mutex_lock(struct mutex *mutex, int depth,
2134 bool nonblock)
2135 {
2136 if (!nonblock) {
2137 mutex_lock_nested(mutex, depth);
2138 return 0;
2139 }
2140 if (mutex_trylock(mutex))
2141 return 0;
2142 return -EAGAIN;
2143 }
2144
do_epoll_ctl(int epfd,int op,int fd,struct epoll_event * epds,bool nonblock)2145 int do_epoll_ctl(int epfd, int op, int fd, struct epoll_event *epds,
2146 bool nonblock)
2147 {
2148 int error;
2149 int full_check = 0;
2150 struct fd f, tf;
2151 struct eventpoll *ep;
2152 struct epitem *epi;
2153 struct eventpoll *tep = NULL;
2154
2155 error = -EBADF;
2156 f = fdget(epfd);
2157 if (!f.file)
2158 goto error_return;
2159
2160 /* Get the "struct file *" for the target file */
2161 tf = fdget(fd);
2162 if (!tf.file)
2163 goto error_fput;
2164
2165 /* The target file descriptor must support poll */
2166 error = -EPERM;
2167 if (!file_can_poll(tf.file))
2168 goto error_tgt_fput;
2169
2170 /* Check if EPOLLWAKEUP is allowed */
2171 if (ep_op_has_event(op))
2172 ep_take_care_of_epollwakeup(epds);
2173
2174 /*
2175 * We have to check that the file structure underneath the file descriptor
2176 * the user passed to us _is_ an eventpoll file. And also we do not permit
2177 * adding an epoll file descriptor inside itself.
2178 */
2179 error = -EINVAL;
2180 if (f.file == tf.file || !is_file_epoll(f.file))
2181 goto error_tgt_fput;
2182
2183 /*
2184 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
2185 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
2186 * Also, we do not currently supported nested exclusive wakeups.
2187 */
2188 if (ep_op_has_event(op) && (epds->events & EPOLLEXCLUSIVE)) {
2189 if (op == EPOLL_CTL_MOD)
2190 goto error_tgt_fput;
2191 if (op == EPOLL_CTL_ADD && (is_file_epoll(tf.file) ||
2192 (epds->events & ~EPOLLEXCLUSIVE_OK_BITS)))
2193 goto error_tgt_fput;
2194 }
2195
2196 /*
2197 * At this point it is safe to assume that the "private_data" contains
2198 * our own data structure.
2199 */
2200 ep = f.file->private_data;
2201
2202 /*
2203 * When we insert an epoll file descriptor, inside another epoll file
2204 * descriptor, there is the change of creating closed loops, which are
2205 * better be handled here, than in more critical paths. While we are
2206 * checking for loops we also determine the list of files reachable
2207 * and hang them on the tfile_check_list, so we can check that we
2208 * haven't created too many possible wakeup paths.
2209 *
2210 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
2211 * the epoll file descriptor is attaching directly to a wakeup source,
2212 * unless the epoll file descriptor is nested. The purpose of taking the
2213 * 'epmutex' on add is to prevent complex toplogies such as loops and
2214 * deep wakeup paths from forming in parallel through multiple
2215 * EPOLL_CTL_ADD operations.
2216 */
2217 error = epoll_mutex_lock(&ep->mtx, 0, nonblock);
2218 if (error)
2219 goto error_tgt_fput;
2220 if (op == EPOLL_CTL_ADD) {
2221 if (!list_empty(&f.file->f_ep_links) ||
2222 ep->gen == loop_check_gen ||
2223 is_file_epoll(tf.file)) {
2224 mutex_unlock(&ep->mtx);
2225 error = epoll_mutex_lock(&epmutex, 0, nonblock);
2226 if (error)
2227 goto error_tgt_fput;
2228 loop_check_gen++;
2229 full_check = 1;
2230 if (is_file_epoll(tf.file)) {
2231 error = -ELOOP;
2232 if (ep_loop_check(ep, tf.file) != 0)
2233 goto error_tgt_fput;
2234 } else {
2235 get_file(tf.file);
2236 list_add(&tf.file->f_tfile_llink,
2237 &tfile_check_list);
2238 }
2239 error = epoll_mutex_lock(&ep->mtx, 0, nonblock);
2240 if (error)
2241 goto error_tgt_fput;
2242 if (is_file_epoll(tf.file)) {
2243 tep = tf.file->private_data;
2244 error = epoll_mutex_lock(&tep->mtx, 1, nonblock);
2245 if (error) {
2246 mutex_unlock(&ep->mtx);
2247 goto error_tgt_fput;
2248 }
2249 }
2250 }
2251 }
2252
2253 /*
2254 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
2255 * above, we can be sure to be able to use the item looked up by
2256 * ep_find() till we release the mutex.
2257 */
2258 epi = ep_find(ep, tf.file, fd);
2259
2260 error = -EINVAL;
2261 switch (op) {
2262 case EPOLL_CTL_ADD:
2263 if (!epi) {
2264 epds->events |= EPOLLERR | EPOLLHUP;
2265 error = ep_insert(ep, epds, tf.file, fd, full_check);
2266 } else
2267 error = -EEXIST;
2268 break;
2269 case EPOLL_CTL_DEL:
2270 if (epi)
2271 error = ep_remove(ep, epi);
2272 else
2273 error = -ENOENT;
2274 break;
2275 case EPOLL_CTL_MOD:
2276 if (epi) {
2277 if (!(epi->event.events & EPOLLEXCLUSIVE)) {
2278 epds->events |= EPOLLERR | EPOLLHUP;
2279 error = ep_modify(ep, epi, epds);
2280 }
2281 } else
2282 error = -ENOENT;
2283 break;
2284 }
2285 if (tep != NULL)
2286 mutex_unlock(&tep->mtx);
2287 mutex_unlock(&ep->mtx);
2288
2289 error_tgt_fput:
2290 if (full_check) {
2291 clear_tfile_check_list();
2292 loop_check_gen++;
2293 mutex_unlock(&epmutex);
2294 }
2295
2296 fdput(tf);
2297 error_fput:
2298 fdput(f);
2299 error_return:
2300
2301 return error;
2302 }
2303
2304 /*
2305 * The following function implements the controller interface for
2306 * the eventpoll file that enables the insertion/removal/change of
2307 * file descriptors inside the interest set.
2308 */
SYSCALL_DEFINE4(epoll_ctl,int,epfd,int,op,int,fd,struct epoll_event __user *,event)2309 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
2310 struct epoll_event __user *, event)
2311 {
2312 struct epoll_event epds;
2313
2314 if (ep_op_has_event(op) &&
2315 copy_from_user(&epds, event, sizeof(struct epoll_event)))
2316 return -EFAULT;
2317
2318 return do_epoll_ctl(epfd, op, fd, &epds, false);
2319 }
2320
2321 /*
2322 * Implement the event wait interface for the eventpoll file. It is the kernel
2323 * part of the user space epoll_wait(2).
2324 */
do_epoll_wait(int epfd,struct epoll_event __user * events,int maxevents,int timeout)2325 static int do_epoll_wait(int epfd, struct epoll_event __user *events,
2326 int maxevents, int timeout)
2327 {
2328 int error;
2329 struct fd f;
2330 struct eventpoll *ep;
2331
2332 /* The maximum number of event must be greater than zero */
2333 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
2334 return -EINVAL;
2335
2336 /* Verify that the area passed by the user is writeable */
2337 if (!access_ok(events, maxevents * sizeof(struct epoll_event)))
2338 return -EFAULT;
2339
2340 /* Get the "struct file *" for the eventpoll file */
2341 f = fdget(epfd);
2342 if (!f.file)
2343 return -EBADF;
2344
2345 /*
2346 * We have to check that the file structure underneath the fd
2347 * the user passed to us _is_ an eventpoll file.
2348 */
2349 error = -EINVAL;
2350 if (!is_file_epoll(f.file))
2351 goto error_fput;
2352
2353 /*
2354 * At this point it is safe to assume that the "private_data" contains
2355 * our own data structure.
2356 */
2357 ep = f.file->private_data;
2358
2359 /* Time to fish for events ... */
2360 error = ep_poll(ep, events, maxevents, timeout);
2361
2362 error_fput:
2363 fdput(f);
2364 return error;
2365 }
2366
SYSCALL_DEFINE4(epoll_wait,int,epfd,struct epoll_event __user *,events,int,maxevents,int,timeout)2367 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
2368 int, maxevents, int, timeout)
2369 {
2370 return do_epoll_wait(epfd, events, maxevents, timeout);
2371 }
2372
2373 /*
2374 * Implement the event wait interface for the eventpoll file. It is the kernel
2375 * part of the user space epoll_pwait(2).
2376 */
SYSCALL_DEFINE6(epoll_pwait,int,epfd,struct epoll_event __user *,events,int,maxevents,int,timeout,const sigset_t __user *,sigmask,size_t,sigsetsize)2377 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2378 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2379 size_t, sigsetsize)
2380 {
2381 int error;
2382
2383 /*
2384 * If the caller wants a certain signal mask to be set during the wait,
2385 * we apply it here.
2386 */
2387 error = set_user_sigmask(sigmask, sigsetsize);
2388 if (error)
2389 return error;
2390
2391 error = do_epoll_wait(epfd, events, maxevents, timeout);
2392 restore_saved_sigmask_unless(error == -EINTR);
2393
2394 return error;
2395 }
2396
2397 #ifdef CONFIG_COMPAT
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)2398 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2399 struct epoll_event __user *, events,
2400 int, maxevents, int, timeout,
2401 const compat_sigset_t __user *, sigmask,
2402 compat_size_t, sigsetsize)
2403 {
2404 long err;
2405
2406 /*
2407 * If the caller wants a certain signal mask to be set during the wait,
2408 * we apply it here.
2409 */
2410 err = set_compat_user_sigmask(sigmask, sigsetsize);
2411 if (err)
2412 return err;
2413
2414 err = do_epoll_wait(epfd, events, maxevents, timeout);
2415 restore_saved_sigmask_unless(err == -EINTR);
2416
2417 return err;
2418 }
2419 #endif
2420
eventpoll_init(void)2421 static int __init eventpoll_init(void)
2422 {
2423 struct sysinfo si;
2424
2425 si_meminfo(&si);
2426 /*
2427 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2428 */
2429 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2430 EP_ITEM_COST;
2431 BUG_ON(max_user_watches < 0);
2432
2433 /*
2434 * Initialize the structure used to perform epoll file descriptor
2435 * inclusion loops checks.
2436 */
2437 ep_nested_calls_init(&poll_loop_ncalls);
2438
2439 /*
2440 * We can have many thousands of epitems, so prevent this from
2441 * using an extra cache line on 64-bit (and smaller) CPUs
2442 */
2443 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2444
2445 /* Allocates slab cache used to allocate "struct epitem" items */
2446 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2447 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC|SLAB_ACCOUNT, NULL);
2448
2449 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2450 pwq_cache = kmem_cache_create("eventpoll_pwq",
2451 sizeof(struct eppoll_entry), 0, SLAB_PANIC|SLAB_ACCOUNT, NULL);
2452
2453 return 0;
2454 }
2455 fs_initcall(eventpoll_init);
2456