]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * fs/eventpoll.c (Efficient event retrieval implementation) | |
3 | * Copyright (C) 2001,...,2009 Davide Libenzi | |
4 | * | |
5 | * This program is free software; you can redistribute it and/or modify | |
6 | * it under the terms of the GNU General Public License as published by | |
7 | * the Free Software Foundation; either version 2 of the License, or | |
8 | * (at your option) any later version. | |
9 | * | |
10 | * Davide Libenzi <davidel@xmailserver.org> | |
11 | * | |
12 | */ | |
13 | ||
14 | #include <linux/init.h> | |
15 | #include <linux/kernel.h> | |
16 | #include <linux/sched.h> | |
17 | #include <linux/fs.h> | |
18 | #include <linux/file.h> | |
19 | #include <linux/signal.h> | |
20 | #include <linux/errno.h> | |
21 | #include <linux/mm.h> | |
22 | #include <linux/slab.h> | |
23 | #include <linux/poll.h> | |
24 | #include <linux/string.h> | |
25 | #include <linux/list.h> | |
26 | #include <linux/hash.h> | |
27 | #include <linux/spinlock.h> | |
28 | #include <linux/syscalls.h> | |
29 | #include <linux/rbtree.h> | |
30 | #include <linux/wait.h> | |
31 | #include <linux/eventpoll.h> | |
32 | #include <linux/mount.h> | |
33 | #include <linux/bitops.h> | |
34 | #include <linux/mutex.h> | |
35 | #include <linux/anon_inodes.h> | |
36 | #include <linux/device.h> | |
37 | #include <asm/uaccess.h> | |
38 | #include <asm/io.h> | |
39 | #include <asm/mman.h> | |
40 | #include <linux/atomic.h> | |
41 | #include <linux/proc_fs.h> | |
42 | #include <linux/seq_file.h> | |
43 | #include <linux/compat.h> | |
44 | #include <linux/rculist.h> | |
45 | ||
46 | /* | |
47 | * LOCKING: | |
48 | * There are three level of locking required by epoll : | |
49 | * | |
50 | * 1) epmutex (mutex) | |
51 | * 2) ep->mtx (mutex) | |
52 | * 3) ep->lock (spinlock) | |
53 | * | |
54 | * The acquire order is the one listed above, from 1 to 3. | |
55 | * We need a spinlock (ep->lock) because we manipulate objects | |
56 | * from inside the poll callback, that might be triggered from | |
57 | * a wake_up() that in turn might be called from IRQ context. | |
58 | * So we can't sleep inside the poll callback and hence we need | |
59 | * a spinlock. During the event transfer loop (from kernel to | |
60 | * user space) we could end up sleeping due a copy_to_user(), so | |
61 | * we need a lock that will allow us to sleep. This lock is a | |
62 | * mutex (ep->mtx). It is acquired during the event transfer loop, | |
63 | * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file(). | |
64 | * Then we also need a global mutex to serialize eventpoll_release_file() | |
65 | * and ep_free(). | |
66 | * This mutex is acquired by ep_free() during the epoll file | |
67 | * cleanup path and it is also acquired by eventpoll_release_file() | |
68 | * if a file has been pushed inside an epoll set and it is then | |
69 | * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL). | |
70 | * It is also acquired when inserting an epoll fd onto another epoll | |
71 | * fd. We do this so that we walk the epoll tree and ensure that this | |
72 | * insertion does not create a cycle of epoll file descriptors, which | |
73 | * could lead to deadlock. We need a global mutex to prevent two | |
74 | * simultaneous inserts (A into B and B into A) from racing and | |
75 | * constructing a cycle without either insert observing that it is | |
76 | * going to. | |
77 | * It is necessary to acquire multiple "ep->mtx"es at once in the | |
78 | * case when one epoll fd is added to another. In this case, we | |
79 | * always acquire the locks in the order of nesting (i.e. after | |
80 | * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired | |
81 | * before e2->mtx). Since we disallow cycles of epoll file | |
82 | * descriptors, this ensures that the mutexes are well-ordered. In | |
83 | * order to communicate this nesting to lockdep, when walking a tree | |
84 | * of epoll file descriptors, we use the current recursion depth as | |
85 | * the lockdep subkey. | |
86 | * It is possible to drop the "ep->mtx" and to use the global | |
87 | * mutex "epmutex" (together with "ep->lock") to have it working, | |
88 | * but having "ep->mtx" will make the interface more scalable. | |
89 | * Events that require holding "epmutex" are very rare, while for | |
90 | * normal operations the epoll private "ep->mtx" will guarantee | |
91 | * a better scalability. | |
92 | */ | |
93 | ||
94 | /* Epoll private bits inside the event mask */ | |
95 | #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET) | |
96 | ||
97 | /* Maximum number of nesting allowed inside epoll sets */ | |
98 | #define EP_MAX_NESTS 4 | |
99 | ||
100 | #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event)) | |
101 | ||
102 | #define EP_UNACTIVE_PTR ((void *) -1L) | |
103 | ||
104 | #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry)) | |
105 | ||
106 | struct epoll_filefd { | |
107 | struct file *file; | |
108 | int fd; | |
109 | } __packed; | |
110 | ||
111 | /* | |
112 | * Structure used to track possible nested calls, for too deep recursions | |
113 | * and loop cycles. | |
114 | */ | |
115 | struct nested_call_node { | |
116 | struct list_head llink; | |
117 | void *cookie; | |
118 | void *ctx; | |
119 | }; | |
120 | ||
121 | /* | |
122 | * This structure is used as collector for nested calls, to check for | |
123 | * maximum recursion dept and loop cycles. | |
124 | */ | |
125 | struct nested_calls { | |
126 | struct list_head tasks_call_list; | |
127 | spinlock_t lock; | |
128 | }; | |
129 | ||
130 | /* | |
131 | * Each file descriptor added to the eventpoll interface will | |
132 | * have an entry of this type linked to the "rbr" RB tree. | |
133 | * Avoid increasing the size of this struct, there can be many thousands | |
134 | * of these on a server and we do not want this to take another cache line. | |
135 | */ | |
136 | struct epitem { | |
137 | union { | |
138 | /* RB tree node links this structure to the eventpoll RB tree */ | |
139 | struct rb_node rbn; | |
140 | /* Used to free the struct epitem */ | |
141 | struct rcu_head rcu; | |
142 | }; | |
143 | ||
144 | /* List header used to link this structure to the eventpoll ready list */ | |
145 | struct list_head rdllink; | |
146 | ||
147 | /* | |
148 | * Works together "struct eventpoll"->ovflist in keeping the | |
149 | * single linked chain of items. | |
150 | */ | |
151 | struct epitem *next; | |
152 | ||
153 | /* The file descriptor information this item refers to */ | |
154 | struct epoll_filefd ffd; | |
155 | ||
156 | /* Number of active wait queue attached to poll operations */ | |
157 | int nwait; | |
158 | ||
159 | /* List containing poll wait queues */ | |
160 | struct list_head 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 list_head 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 | /* Protect the access to this structure */ | |
182 | spinlock_t lock; | |
183 | ||
184 | /* | |
185 | * This mutex is used to ensure that files are not removed | |
186 | * while epoll is using them. This is held during the event | |
187 | * collection loop, the file cleanup path, the epoll file exit | |
188 | * code and the ctl operations. | |
189 | */ | |
190 | struct mutex mtx; | |
191 | ||
192 | /* Wait queue used by sys_epoll_wait() */ | |
193 | wait_queue_head_t wq; | |
194 | ||
195 | /* Wait queue used by file->poll() */ | |
196 | wait_queue_head_t poll_wait; | |
197 | ||
198 | /* List of ready file descriptors */ | |
199 | struct list_head rdllist; | |
200 | ||
201 | /* RB tree root used to store monitored fd structs */ | |
202 | struct rb_root 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 | int visited; | |
221 | struct list_head visited_list_link; | |
222 | }; | |
223 | ||
224 | /* Wait structure used by the poll hooks */ | |
225 | struct eppoll_entry { | |
226 | /* List header used to link this structure to the "struct epitem" */ | |
227 | struct list_head llink; | |
228 | ||
229 | /* The "base" pointer is set to the container "struct epitem" */ | |
230 | struct epitem *base; | |
231 | ||
232 | /* | |
233 | * Wait queue item that will be linked to the target file wait | |
234 | * queue head. | |
235 | */ | |
236 | wait_queue_t wait; | |
237 | ||
238 | /* The wait queue head that linked the "wait" wait queue item */ | |
239 | wait_queue_head_t *whead; | |
240 | }; | |
241 | ||
242 | /* Wrapper struct used by poll queueing */ | |
243 | struct ep_pqueue { | |
244 | poll_table pt; | |
245 | struct epitem *epi; | |
246 | }; | |
247 | ||
248 | /* Used by the ep_send_events() function as callback private data */ | |
249 | struct ep_send_events_data { | |
250 | int maxevents; | |
251 | struct epoll_event __user *events; | |
252 | }; | |
253 | ||
254 | /* | |
255 | * Configuration options available inside /proc/sys/fs/epoll/ | |
256 | */ | |
257 | /* Maximum number of epoll watched descriptors, per user */ | |
258 | static long max_user_watches __read_mostly; | |
259 | ||
260 | /* | |
261 | * This mutex is used to serialize ep_free() and eventpoll_release_file(). | |
262 | */ | |
263 | static DEFINE_MUTEX(epmutex); | |
264 | ||
265 | /* Used to check for epoll file descriptor inclusion loops */ | |
266 | static struct nested_calls poll_loop_ncalls; | |
267 | ||
268 | /* Used for safe wake up implementation */ | |
269 | static struct nested_calls poll_safewake_ncalls; | |
270 | ||
271 | /* Used to call file's f_op->poll() under the nested calls boundaries */ | |
272 | static struct nested_calls poll_readywalk_ncalls; | |
273 | ||
274 | /* Slab cache used to allocate "struct epitem" */ | |
275 | static struct kmem_cache *epi_cache __read_mostly; | |
276 | ||
277 | /* Slab cache used to allocate "struct eppoll_entry" */ | |
278 | static struct kmem_cache *pwq_cache __read_mostly; | |
279 | ||
280 | /* Visited nodes during ep_loop_check(), so we can unset them when we finish */ | |
281 | static LIST_HEAD(visited_list); | |
282 | ||
283 | /* | |
284 | * List of files with newly added links, where we may need to limit the number | |
285 | * of emanating paths. Protected by the epmutex. | |
286 | */ | |
287 | static LIST_HEAD(tfile_check_list); | |
288 | ||
289 | #ifdef CONFIG_SYSCTL | |
290 | ||
291 | #include <linux/sysctl.h> | |
292 | ||
293 | static long zero; | |
294 | static long long_max = LONG_MAX; | |
295 | ||
296 | ctl_table epoll_table[] = { | |
297 | { | |
298 | .procname = "max_user_watches", | |
299 | .data = &max_user_watches, | |
300 | .maxlen = sizeof(max_user_watches), | |
301 | .mode = 0644, | |
302 | .proc_handler = proc_doulongvec_minmax, | |
303 | .extra1 = &zero, | |
304 | .extra2 = &long_max, | |
305 | }, | |
306 | { } | |
307 | }; | |
308 | #endif /* CONFIG_SYSCTL */ | |
309 | ||
310 | static const struct file_operations eventpoll_fops; | |
311 | ||
312 | static inline int is_file_epoll(struct file *f) | |
313 | { | |
314 | return f->f_op == &eventpoll_fops; | |
315 | } | |
316 | ||
317 | /* Setup the structure that is used as key for the RB tree */ | |
318 | static inline void ep_set_ffd(struct epoll_filefd *ffd, | |
319 | struct file *file, int fd) | |
320 | { | |
321 | ffd->file = file; | |
322 | ffd->fd = fd; | |
323 | } | |
324 | ||
325 | /* Compare RB tree keys */ | |
326 | static inline int ep_cmp_ffd(struct epoll_filefd *p1, | |
327 | struct epoll_filefd *p2) | |
328 | { | |
329 | return (p1->file > p2->file ? +1: | |
330 | (p1->file < p2->file ? -1 : p1->fd - p2->fd)); | |
331 | } | |
332 | ||
333 | /* Tells us if the item is currently linked */ | |
334 | static inline int ep_is_linked(struct list_head *p) | |
335 | { | |
336 | return !list_empty(p); | |
337 | } | |
338 | ||
339 | static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_t *p) | |
340 | { | |
341 | return container_of(p, struct eppoll_entry, wait); | |
342 | } | |
343 | ||
344 | /* Get the "struct epitem" from a wait queue pointer */ | |
345 | static inline struct epitem *ep_item_from_wait(wait_queue_t *p) | |
346 | { | |
347 | return container_of(p, struct eppoll_entry, wait)->base; | |
348 | } | |
349 | ||
350 | /* Get the "struct epitem" from an epoll queue wrapper */ | |
351 | static inline struct epitem *ep_item_from_epqueue(poll_table *p) | |
352 | { | |
353 | return container_of(p, struct ep_pqueue, pt)->epi; | |
354 | } | |
355 | ||
356 | /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */ | |
357 | static inline int ep_op_has_event(int op) | |
358 | { | |
359 | return op != EPOLL_CTL_DEL; | |
360 | } | |
361 | ||
362 | /* Initialize the poll safe wake up structure */ | |
363 | static void ep_nested_calls_init(struct nested_calls *ncalls) | |
364 | { | |
365 | INIT_LIST_HEAD(&ncalls->tasks_call_list); | |
366 | spin_lock_init(&ncalls->lock); | |
367 | } | |
368 | ||
369 | /** | |
370 | * ep_events_available - Checks if ready events might be available. | |
371 | * | |
372 | * @ep: Pointer to the eventpoll context. | |
373 | * | |
374 | * Returns: Returns a value different than zero if ready events are available, | |
375 | * or zero otherwise. | |
376 | */ | |
377 | static inline int ep_events_available(struct eventpoll *ep) | |
378 | { | |
379 | return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR; | |
380 | } | |
381 | ||
382 | /** | |
383 | * ep_call_nested - Perform a bound (possibly) nested call, by checking | |
384 | * that the recursion limit is not exceeded, and that | |
385 | * the same nested call (by the meaning of same cookie) is | |
386 | * no re-entered. | |
387 | * | |
388 | * @ncalls: Pointer to the nested_calls structure to be used for this call. | |
389 | * @max_nests: Maximum number of allowed nesting calls. | |
390 | * @nproc: Nested call core function pointer. | |
391 | * @priv: Opaque data to be passed to the @nproc callback. | |
392 | * @cookie: Cookie to be used to identify this nested call. | |
393 | * @ctx: This instance context. | |
394 | * | |
395 | * Returns: Returns the code returned by the @nproc callback, or -1 if | |
396 | * the maximum recursion limit has been exceeded. | |
397 | */ | |
398 | static int ep_call_nested(struct nested_calls *ncalls, int max_nests, | |
399 | int (*nproc)(void *, void *, int), void *priv, | |
400 | void *cookie, void *ctx) | |
401 | { | |
402 | int error, call_nests = 0; | |
403 | unsigned long flags; | |
404 | struct list_head *lsthead = &ncalls->tasks_call_list; | |
405 | struct nested_call_node *tncur; | |
406 | struct nested_call_node tnode; | |
407 | ||
408 | spin_lock_irqsave(&ncalls->lock, flags); | |
409 | ||
410 | /* | |
411 | * Try to see if the current task is already inside this wakeup call. | |
412 | * We use a list here, since the population inside this set is always | |
413 | * very much limited. | |
414 | */ | |
415 | list_for_each_entry(tncur, lsthead, llink) { | |
416 | if (tncur->ctx == ctx && | |
417 | (tncur->cookie == cookie || ++call_nests > max_nests)) { | |
418 | /* | |
419 | * Ops ... loop detected or maximum nest level reached. | |
420 | * We abort this wake by breaking the cycle itself. | |
421 | */ | |
422 | error = -1; | |
423 | goto out_unlock; | |
424 | } | |
425 | } | |
426 | ||
427 | /* Add the current task and cookie to the list */ | |
428 | tnode.ctx = ctx; | |
429 | tnode.cookie = cookie; | |
430 | list_add(&tnode.llink, lsthead); | |
431 | ||
432 | spin_unlock_irqrestore(&ncalls->lock, flags); | |
433 | ||
434 | /* Call the nested function */ | |
435 | error = (*nproc)(priv, cookie, call_nests); | |
436 | ||
437 | /* Remove the current task from the list */ | |
438 | spin_lock_irqsave(&ncalls->lock, flags); | |
439 | list_del(&tnode.llink); | |
440 | out_unlock: | |
441 | spin_unlock_irqrestore(&ncalls->lock, flags); | |
442 | ||
443 | return error; | |
444 | } | |
445 | ||
446 | /* | |
447 | * As described in commit 0ccf831cb lockdep: annotate epoll | |
448 | * the use of wait queues used by epoll is done in a very controlled | |
449 | * manner. Wake ups can nest inside each other, but are never done | |
450 | * with the same locking. For example: | |
451 | * | |
452 | * dfd = socket(...); | |
453 | * efd1 = epoll_create(); | |
454 | * efd2 = epoll_create(); | |
455 | * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...); | |
456 | * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...); | |
457 | * | |
458 | * When a packet arrives to the device underneath "dfd", the net code will | |
459 | * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a | |
460 | * callback wakeup entry on that queue, and the wake_up() performed by the | |
461 | * "dfd" net code will end up in ep_poll_callback(). At this point epoll | |
462 | * (efd1) notices that it may have some event ready, so it needs to wake up | |
463 | * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake() | |
464 | * that ends up in another wake_up(), after having checked about the | |
465 | * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to | |
466 | * avoid stack blasting. | |
467 | * | |
468 | * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle | |
469 | * this special case of epoll. | |
470 | */ | |
471 | #ifdef CONFIG_DEBUG_LOCK_ALLOC | |
472 | static inline void ep_wake_up_nested(wait_queue_head_t *wqueue, | |
473 | unsigned long events, int subclass) | |
474 | { | |
475 | unsigned long flags; | |
476 | ||
477 | spin_lock_irqsave_nested(&wqueue->lock, flags, subclass); | |
478 | wake_up_locked_poll(wqueue, events); | |
479 | spin_unlock_irqrestore(&wqueue->lock, flags); | |
480 | } | |
481 | #else | |
482 | static inline void ep_wake_up_nested(wait_queue_head_t *wqueue, | |
483 | unsigned long events, int subclass) | |
484 | { | |
485 | wake_up_poll(wqueue, events); | |
486 | } | |
487 | #endif | |
488 | ||
489 | static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests) | |
490 | { | |
491 | ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN, | |
492 | 1 + call_nests); | |
493 | return 0; | |
494 | } | |
495 | ||
496 | /* | |
497 | * Perform a safe wake up of the poll wait list. The problem is that | |
498 | * with the new callback'd wake up system, it is possible that the | |
499 | * poll callback is reentered from inside the call to wake_up() done | |
500 | * on the poll wait queue head. The rule is that we cannot reenter the | |
501 | * wake up code from the same task more than EP_MAX_NESTS times, | |
502 | * and we cannot reenter the same wait queue head at all. This will | |
503 | * enable to have a hierarchy of epoll file descriptor of no more than | |
504 | * EP_MAX_NESTS deep. | |
505 | */ | |
506 | static void ep_poll_safewake(wait_queue_head_t *wq) | |
507 | { | |
508 | int this_cpu = get_cpu(); | |
509 | ||
510 | ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS, | |
511 | ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu); | |
512 | ||
513 | put_cpu(); | |
514 | } | |
515 | ||
516 | static void ep_remove_wait_queue(struct eppoll_entry *pwq) | |
517 | { | |
518 | wait_queue_head_t *whead; | |
519 | ||
520 | rcu_read_lock(); | |
521 | /* If it is cleared by POLLFREE, it should be rcu-safe */ | |
522 | whead = rcu_dereference(pwq->whead); | |
523 | if (whead) | |
524 | remove_wait_queue(whead, &pwq->wait); | |
525 | rcu_read_unlock(); | |
526 | } | |
527 | ||
528 | /* | |
529 | * This function unregisters poll callbacks from the associated file | |
530 | * descriptor. Must be called with "mtx" held (or "epmutex" if called from | |
531 | * ep_free). | |
532 | */ | |
533 | static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi) | |
534 | { | |
535 | struct list_head *lsthead = &epi->pwqlist; | |
536 | struct eppoll_entry *pwq; | |
537 | ||
538 | while (!list_empty(lsthead)) { | |
539 | pwq = list_first_entry(lsthead, struct eppoll_entry, llink); | |
540 | ||
541 | list_del(&pwq->llink); | |
542 | ep_remove_wait_queue(pwq); | |
543 | kmem_cache_free(pwq_cache, pwq); | |
544 | } | |
545 | } | |
546 | ||
547 | /* call only when ep->mtx is held */ | |
548 | static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi) | |
549 | { | |
550 | return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx)); | |
551 | } | |
552 | ||
553 | /* call only when ep->mtx is held */ | |
554 | static inline void ep_pm_stay_awake(struct epitem *epi) | |
555 | { | |
556 | struct wakeup_source *ws = ep_wakeup_source(epi); | |
557 | ||
558 | if (ws) | |
559 | __pm_stay_awake(ws); | |
560 | } | |
561 | ||
562 | static inline bool ep_has_wakeup_source(struct epitem *epi) | |
563 | { | |
564 | return rcu_access_pointer(epi->ws) ? true : false; | |
565 | } | |
566 | ||
567 | /* call when ep->mtx cannot be held (ep_poll_callback) */ | |
568 | static inline void ep_pm_stay_awake_rcu(struct epitem *epi) | |
569 | { | |
570 | struct wakeup_source *ws; | |
571 | ||
572 | rcu_read_lock(); | |
573 | ws = rcu_dereference(epi->ws); | |
574 | if (ws) | |
575 | __pm_stay_awake(ws); | |
576 | rcu_read_unlock(); | |
577 | } | |
578 | ||
579 | /** | |
580 | * ep_scan_ready_list - Scans the ready list in a way that makes possible for | |
581 | * the scan code, to call f_op->poll(). Also allows for | |
582 | * O(NumReady) performance. | |
583 | * | |
584 | * @ep: Pointer to the epoll private data structure. | |
585 | * @sproc: Pointer to the scan callback. | |
586 | * @priv: Private opaque data passed to the @sproc callback. | |
587 | * @depth: The current depth of recursive f_op->poll calls. | |
588 | * | |
589 | * Returns: The same integer error code returned by the @sproc callback. | |
590 | */ | |
591 | static int ep_scan_ready_list(struct eventpoll *ep, | |
592 | int (*sproc)(struct eventpoll *, | |
593 | struct list_head *, void *), | |
594 | void *priv, | |
595 | int depth) | |
596 | { | |
597 | int error, pwake = 0; | |
598 | unsigned long flags; | |
599 | struct epitem *epi, *nepi; | |
600 | LIST_HEAD(txlist); | |
601 | ||
602 | /* | |
603 | * We need to lock this because we could be hit by | |
604 | * eventpoll_release_file() and epoll_ctl(). | |
605 | */ | |
606 | mutex_lock_nested(&ep->mtx, depth); | |
607 | ||
608 | /* | |
609 | * Steal the ready list, and re-init the original one to the | |
610 | * empty list. Also, set ep->ovflist to NULL so that events | |
611 | * happening while looping w/out locks, are not lost. We cannot | |
612 | * have the poll callback to queue directly on ep->rdllist, | |
613 | * because we want the "sproc" callback to be able to do it | |
614 | * in a lockless way. | |
615 | */ | |
616 | spin_lock_irqsave(&ep->lock, flags); | |
617 | list_splice_init(&ep->rdllist, &txlist); | |
618 | ep->ovflist = NULL; | |
619 | spin_unlock_irqrestore(&ep->lock, flags); | |
620 | ||
621 | /* | |
622 | * Now call the callback function. | |
623 | */ | |
624 | error = (*sproc)(ep, &txlist, priv); | |
625 | ||
626 | spin_lock_irqsave(&ep->lock, flags); | |
627 | /* | |
628 | * During the time we spent inside the "sproc" callback, some | |
629 | * other events might have been queued by the poll callback. | |
630 | * We re-insert them inside the main ready-list here. | |
631 | */ | |
632 | for (nepi = ep->ovflist; (epi = nepi) != NULL; | |
633 | nepi = epi->next, epi->next = EP_UNACTIVE_PTR) { | |
634 | /* | |
635 | * We need to check if the item is already in the list. | |
636 | * During the "sproc" callback execution time, items are | |
637 | * queued into ->ovflist but the "txlist" might already | |
638 | * contain them, and the list_splice() below takes care of them. | |
639 | */ | |
640 | if (!ep_is_linked(&epi->rdllink)) { | |
641 | list_add_tail(&epi->rdllink, &ep->rdllist); | |
642 | ep_pm_stay_awake(epi); | |
643 | } | |
644 | } | |
645 | /* | |
646 | * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after | |
647 | * releasing the lock, events will be queued in the normal way inside | |
648 | * ep->rdllist. | |
649 | */ | |
650 | ep->ovflist = EP_UNACTIVE_PTR; | |
651 | ||
652 | /* | |
653 | * Quickly re-inject items left on "txlist". | |
654 | */ | |
655 | list_splice(&txlist, &ep->rdllist); | |
656 | __pm_relax(ep->ws); | |
657 | ||
658 | if (!list_empty(&ep->rdllist)) { | |
659 | /* | |
660 | * Wake up (if active) both the eventpoll wait list and | |
661 | * the ->poll() wait list (delayed after we release the lock). | |
662 | */ | |
663 | if (waitqueue_active(&ep->wq)) | |
664 | wake_up_locked(&ep->wq); | |
665 | if (waitqueue_active(&ep->poll_wait)) | |
666 | pwake++; | |
667 | } | |
668 | spin_unlock_irqrestore(&ep->lock, flags); | |
669 | ||
670 | mutex_unlock(&ep->mtx); | |
671 | ||
672 | /* We have to call this outside the lock */ | |
673 | if (pwake) | |
674 | ep_poll_safewake(&ep->poll_wait); | |
675 | ||
676 | return error; | |
677 | } | |
678 | ||
679 | static void epi_rcu_free(struct rcu_head *head) | |
680 | { | |
681 | struct epitem *epi = container_of(head, struct epitem, rcu); | |
682 | kmem_cache_free(epi_cache, epi); | |
683 | } | |
684 | ||
685 | /* | |
686 | * Removes a "struct epitem" from the eventpoll RB tree and deallocates | |
687 | * all the associated resources. Must be called with "mtx" held. | |
688 | */ | |
689 | static int ep_remove(struct eventpoll *ep, struct epitem *epi) | |
690 | { | |
691 | unsigned long flags; | |
692 | struct file *file = epi->ffd.file; | |
693 | ||
694 | /* | |
695 | * Removes poll wait queue hooks. We _have_ to do this without holding | |
696 | * the "ep->lock" otherwise a deadlock might occur. This because of the | |
697 | * sequence of the lock acquisition. Here we do "ep->lock" then the wait | |
698 | * queue head lock when unregistering the wait queue. The wakeup callback | |
699 | * will run by holding the wait queue head lock and will call our callback | |
700 | * that will try to get "ep->lock". | |
701 | */ | |
702 | ep_unregister_pollwait(ep, epi); | |
703 | ||
704 | /* Remove the current item from the list of epoll hooks */ | |
705 | spin_lock(&file->f_lock); | |
706 | list_del_rcu(&epi->fllink); | |
707 | spin_unlock(&file->f_lock); | |
708 | ||
709 | rb_erase(&epi->rbn, &ep->rbr); | |
710 | ||
711 | spin_lock_irqsave(&ep->lock, flags); | |
712 | if (ep_is_linked(&epi->rdllink)) | |
713 | list_del_init(&epi->rdllink); | |
714 | spin_unlock_irqrestore(&ep->lock, flags); | |
715 | ||
716 | wakeup_source_unregister(ep_wakeup_source(epi)); | |
717 | /* | |
718 | * At this point it is safe to free the eventpoll item. Use the union | |
719 | * field epi->rcu, since we are trying to minimize the size of | |
720 | * 'struct epitem'. The 'rbn' field is no longer in use. Protected by | |
721 | * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make | |
722 | * use of the rbn field. | |
723 | */ | |
724 | call_rcu(&epi->rcu, epi_rcu_free); | |
725 | ||
726 | atomic_long_dec(&ep->user->epoll_watches); | |
727 | ||
728 | return 0; | |
729 | } | |
730 | ||
731 | static void ep_free(struct eventpoll *ep) | |
732 | { | |
733 | struct rb_node *rbp; | |
734 | struct epitem *epi; | |
735 | ||
736 | /* We need to release all tasks waiting for these file */ | |
737 | if (waitqueue_active(&ep->poll_wait)) | |
738 | ep_poll_safewake(&ep->poll_wait); | |
739 | ||
740 | /* | |
741 | * We need to lock this because we could be hit by | |
742 | * eventpoll_release_file() while we're freeing the "struct eventpoll". | |
743 | * We do not need to hold "ep->mtx" here because the epoll file | |
744 | * is on the way to be removed and no one has references to it | |
745 | * anymore. The only hit might come from eventpoll_release_file() but | |
746 | * holding "epmutex" is sufficient here. | |
747 | */ | |
748 | mutex_lock(&epmutex); | |
749 | ||
750 | /* | |
751 | * Walks through the whole tree by unregistering poll callbacks. | |
752 | */ | |
753 | for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) { | |
754 | epi = rb_entry(rbp, struct epitem, rbn); | |
755 | ||
756 | ep_unregister_pollwait(ep, epi); | |
757 | cond_resched(); | |
758 | } | |
759 | ||
760 | /* | |
761 | * Walks through the whole tree by freeing each "struct epitem". At this | |
762 | * point we are sure no poll callbacks will be lingering around, and also by | |
763 | * holding "epmutex" we can be sure that no file cleanup code will hit | |
764 | * us during this operation. So we can avoid the lock on "ep->lock". | |
765 | * We do not need to lock ep->mtx, either, we only do it to prevent | |
766 | * a lockdep warning. | |
767 | */ | |
768 | mutex_lock(&ep->mtx); | |
769 | while ((rbp = rb_first(&ep->rbr)) != NULL) { | |
770 | epi = rb_entry(rbp, struct epitem, rbn); | |
771 | ep_remove(ep, epi); | |
772 | cond_resched(); | |
773 | } | |
774 | mutex_unlock(&ep->mtx); | |
775 | ||
776 | mutex_unlock(&epmutex); | |
777 | mutex_destroy(&ep->mtx); | |
778 | free_uid(ep->user); | |
779 | wakeup_source_unregister(ep->ws); | |
780 | kfree(ep); | |
781 | } | |
782 | ||
783 | static int ep_eventpoll_release(struct inode *inode, struct file *file) | |
784 | { | |
785 | struct eventpoll *ep = file->private_data; | |
786 | ||
787 | if (ep) | |
788 | ep_free(ep); | |
789 | ||
790 | return 0; | |
791 | } | |
792 | ||
793 | static inline unsigned int ep_item_poll(struct epitem *epi, poll_table *pt) | |
794 | { | |
795 | pt->_key = epi->event.events; | |
796 | ||
797 | return epi->ffd.file->f_op->poll(epi->ffd.file, pt) & epi->event.events; | |
798 | } | |
799 | ||
800 | static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head, | |
801 | void *priv) | |
802 | { | |
803 | struct epitem *epi, *tmp; | |
804 | poll_table pt; | |
805 | ||
806 | init_poll_funcptr(&pt, NULL); | |
807 | ||
808 | list_for_each_entry_safe(epi, tmp, head, rdllink) { | |
809 | if (ep_item_poll(epi, &pt)) | |
810 | return POLLIN | POLLRDNORM; | |
811 | else { | |
812 | /* | |
813 | * Item has been dropped into the ready list by the poll | |
814 | * callback, but it's not actually ready, as far as | |
815 | * caller requested events goes. We can remove it here. | |
816 | */ | |
817 | __pm_relax(ep_wakeup_source(epi)); | |
818 | list_del_init(&epi->rdllink); | |
819 | } | |
820 | } | |
821 | ||
822 | return 0; | |
823 | } | |
824 | ||
825 | static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests) | |
826 | { | |
827 | return ep_scan_ready_list(priv, ep_read_events_proc, NULL, call_nests + 1); | |
828 | } | |
829 | ||
830 | static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait) | |
831 | { | |
832 | int pollflags; | |
833 | struct eventpoll *ep = file->private_data; | |
834 | ||
835 | /* Insert inside our poll wait queue */ | |
836 | poll_wait(file, &ep->poll_wait, wait); | |
837 | ||
838 | /* | |
839 | * Proceed to find out if wanted events are really available inside | |
840 | * the ready list. This need to be done under ep_call_nested() | |
841 | * supervision, since the call to f_op->poll() done on listed files | |
842 | * could re-enter here. | |
843 | */ | |
844 | pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS, | |
845 | ep_poll_readyevents_proc, ep, ep, current); | |
846 | ||
847 | return pollflags != -1 ? pollflags : 0; | |
848 | } | |
849 | ||
850 | #ifdef CONFIG_PROC_FS | |
851 | static int ep_show_fdinfo(struct seq_file *m, struct file *f) | |
852 | { | |
853 | struct eventpoll *ep = f->private_data; | |
854 | struct rb_node *rbp; | |
855 | int ret = 0; | |
856 | ||
857 | mutex_lock(&ep->mtx); | |
858 | for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) { | |
859 | struct epitem *epi = rb_entry(rbp, struct epitem, rbn); | |
860 | ||
861 | ret = seq_printf(m, "tfd: %8d events: %8x data: %16llx\n", | |
862 | epi->ffd.fd, epi->event.events, | |
863 | (long long)epi->event.data); | |
864 | if (ret) | |
865 | break; | |
866 | } | |
867 | mutex_unlock(&ep->mtx); | |
868 | ||
869 | return ret; | |
870 | } | |
871 | #endif | |
872 | ||
873 | /* File callbacks that implement the eventpoll file behaviour */ | |
874 | static const struct file_operations eventpoll_fops = { | |
875 | #ifdef CONFIG_PROC_FS | |
876 | .show_fdinfo = ep_show_fdinfo, | |
877 | #endif | |
878 | .release = ep_eventpoll_release, | |
879 | .poll = ep_eventpoll_poll, | |
880 | .llseek = noop_llseek, | |
881 | }; | |
882 | ||
883 | /* | |
884 | * This is called from eventpoll_release() to unlink files from the eventpoll | |
885 | * interface. We need to have this facility to cleanup correctly files that are | |
886 | * closed without being removed from the eventpoll interface. | |
887 | */ | |
888 | void eventpoll_release_file(struct file *file) | |
889 | { | |
890 | struct eventpoll *ep; | |
891 | struct epitem *epi; | |
892 | ||
893 | /* | |
894 | * We don't want to get "file->f_lock" because it is not | |
895 | * necessary. It is not necessary because we're in the "struct file" | |
896 | * cleanup path, and this means that no one is using this file anymore. | |
897 | * So, for example, epoll_ctl() cannot hit here since if we reach this | |
898 | * point, the file counter already went to zero and fget() would fail. | |
899 | * The only hit might come from ep_free() but by holding the mutex | |
900 | * will correctly serialize the operation. We do need to acquire | |
901 | * "ep->mtx" after "epmutex" because ep_remove() requires it when called | |
902 | * from anywhere but ep_free(). | |
903 | * | |
904 | * Besides, ep_remove() acquires the lock, so we can't hold it here. | |
905 | */ | |
906 | mutex_lock(&epmutex); | |
907 | list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) { | |
908 | ep = epi->ep; | |
909 | mutex_lock_nested(&ep->mtx, 0); | |
910 | ep_remove(ep, epi); | |
911 | mutex_unlock(&ep->mtx); | |
912 | } | |
913 | mutex_unlock(&epmutex); | |
914 | } | |
915 | ||
916 | static int ep_alloc(struct eventpoll **pep) | |
917 | { | |
918 | int error; | |
919 | struct user_struct *user; | |
920 | struct eventpoll *ep; | |
921 | ||
922 | user = get_current_user(); | |
923 | error = -ENOMEM; | |
924 | ep = kzalloc(sizeof(*ep), GFP_KERNEL); | |
925 | if (unlikely(!ep)) | |
926 | goto free_uid; | |
927 | ||
928 | spin_lock_init(&ep->lock); | |
929 | mutex_init(&ep->mtx); | |
930 | init_waitqueue_head(&ep->wq); | |
931 | init_waitqueue_head(&ep->poll_wait); | |
932 | INIT_LIST_HEAD(&ep->rdllist); | |
933 | ep->rbr = RB_ROOT; | |
934 | ep->ovflist = EP_UNACTIVE_PTR; | |
935 | ep->user = user; | |
936 | ||
937 | *pep = ep; | |
938 | ||
939 | return 0; | |
940 | ||
941 | free_uid: | |
942 | free_uid(user); | |
943 | return error; | |
944 | } | |
945 | ||
946 | /* | |
947 | * Search the file inside the eventpoll tree. The RB tree operations | |
948 | * are protected by the "mtx" mutex, and ep_find() must be called with | |
949 | * "mtx" held. | |
950 | */ | |
951 | static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd) | |
952 | { | |
953 | int kcmp; | |
954 | struct rb_node *rbp; | |
955 | struct epitem *epi, *epir = NULL; | |
956 | struct epoll_filefd ffd; | |
957 | ||
958 | ep_set_ffd(&ffd, file, fd); | |
959 | for (rbp = ep->rbr.rb_node; rbp; ) { | |
960 | epi = rb_entry(rbp, struct epitem, rbn); | |
961 | kcmp = ep_cmp_ffd(&ffd, &epi->ffd); | |
962 | if (kcmp > 0) | |
963 | rbp = rbp->rb_right; | |
964 | else if (kcmp < 0) | |
965 | rbp = rbp->rb_left; | |
966 | else { | |
967 | epir = epi; | |
968 | break; | |
969 | } | |
970 | } | |
971 | ||
972 | return epir; | |
973 | } | |
974 | ||
975 | /* | |
976 | * This is the callback that is passed to the wait queue wakeup | |
977 | * mechanism. It is called by the stored file descriptors when they | |
978 | * have events to report. | |
979 | */ | |
980 | static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key) | |
981 | { | |
982 | int pwake = 0; | |
983 | unsigned long flags; | |
984 | struct epitem *epi = ep_item_from_wait(wait); | |
985 | struct eventpoll *ep = epi->ep; | |
986 | ||
987 | if ((unsigned long)key & POLLFREE) { | |
988 | ep_pwq_from_wait(wait)->whead = NULL; | |
989 | /* | |
990 | * whead = NULL above can race with ep_remove_wait_queue() | |
991 | * which can do another remove_wait_queue() after us, so we | |
992 | * can't use __remove_wait_queue(). whead->lock is held by | |
993 | * the caller. | |
994 | */ | |
995 | list_del_init(&wait->task_list); | |
996 | } | |
997 | ||
998 | spin_lock_irqsave(&ep->lock, flags); | |
999 | ||
1000 | /* | |
1001 | * If the event mask does not contain any poll(2) event, we consider the | |
1002 | * descriptor to be disabled. This condition is likely the effect of the | |
1003 | * EPOLLONESHOT bit that disables the descriptor when an event is received, | |
1004 | * until the next EPOLL_CTL_MOD will be issued. | |
1005 | */ | |
1006 | if (!(epi->event.events & ~EP_PRIVATE_BITS)) | |
1007 | goto out_unlock; | |
1008 | ||
1009 | /* | |
1010 | * Check the events coming with the callback. At this stage, not | |
1011 | * every device reports the events in the "key" parameter of the | |
1012 | * callback. We need to be able to handle both cases here, hence the | |
1013 | * test for "key" != NULL before the event match test. | |
1014 | */ | |
1015 | if (key && !((unsigned long) key & epi->event.events)) | |
1016 | goto out_unlock; | |
1017 | ||
1018 | /* | |
1019 | * If we are transferring events to userspace, we can hold no locks | |
1020 | * (because we're accessing user memory, and because of linux f_op->poll() | |
1021 | * semantics). All the events that happen during that period of time are | |
1022 | * chained in ep->ovflist and requeued later on. | |
1023 | */ | |
1024 | if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) { | |
1025 | if (epi->next == EP_UNACTIVE_PTR) { | |
1026 | epi->next = ep->ovflist; | |
1027 | ep->ovflist = epi; | |
1028 | if (epi->ws) { | |
1029 | /* | |
1030 | * Activate ep->ws since epi->ws may get | |
1031 | * deactivated at any time. | |
1032 | */ | |
1033 | __pm_stay_awake(ep->ws); | |
1034 | } | |
1035 | ||
1036 | } | |
1037 | goto out_unlock; | |
1038 | } | |
1039 | ||
1040 | /* If this file is already in the ready list we exit soon */ | |
1041 | if (!ep_is_linked(&epi->rdllink)) { | |
1042 | list_add_tail(&epi->rdllink, &ep->rdllist); | |
1043 | ep_pm_stay_awake_rcu(epi); | |
1044 | } | |
1045 | ||
1046 | /* | |
1047 | * Wake up ( if active ) both the eventpoll wait list and the ->poll() | |
1048 | * wait list. | |
1049 | */ | |
1050 | if (waitqueue_active(&ep->wq)) | |
1051 | wake_up_locked(&ep->wq); | |
1052 | if (waitqueue_active(&ep->poll_wait)) | |
1053 | pwake++; | |
1054 | ||
1055 | out_unlock: | |
1056 | spin_unlock_irqrestore(&ep->lock, flags); | |
1057 | ||
1058 | /* We have to call this outside the lock */ | |
1059 | if (pwake) | |
1060 | ep_poll_safewake(&ep->poll_wait); | |
1061 | ||
1062 | return 1; | |
1063 | } | |
1064 | ||
1065 | /* | |
1066 | * This is the callback that is used to add our wait queue to the | |
1067 | * target file wakeup lists. | |
1068 | */ | |
1069 | static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead, | |
1070 | poll_table *pt) | |
1071 | { | |
1072 | struct epitem *epi = ep_item_from_epqueue(pt); | |
1073 | struct eppoll_entry *pwq; | |
1074 | ||
1075 | if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) { | |
1076 | init_waitqueue_func_entry(&pwq->wait, ep_poll_callback); | |
1077 | pwq->whead = whead; | |
1078 | pwq->base = epi; | |
1079 | add_wait_queue(whead, &pwq->wait); | |
1080 | list_add_tail(&pwq->llink, &epi->pwqlist); | |
1081 | epi->nwait++; | |
1082 | } else { | |
1083 | /* We have to signal that an error occurred */ | |
1084 | epi->nwait = -1; | |
1085 | } | |
1086 | } | |
1087 | ||
1088 | static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi) | |
1089 | { | |
1090 | int kcmp; | |
1091 | struct rb_node **p = &ep->rbr.rb_node, *parent = NULL; | |
1092 | struct epitem *epic; | |
1093 | ||
1094 | while (*p) { | |
1095 | parent = *p; | |
1096 | epic = rb_entry(parent, struct epitem, rbn); | |
1097 | kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd); | |
1098 | if (kcmp > 0) | |
1099 | p = &parent->rb_right; | |
1100 | else | |
1101 | p = &parent->rb_left; | |
1102 | } | |
1103 | rb_link_node(&epi->rbn, parent, p); | |
1104 | rb_insert_color(&epi->rbn, &ep->rbr); | |
1105 | } | |
1106 | ||
1107 | ||
1108 | ||
1109 | #define PATH_ARR_SIZE 5 | |
1110 | /* | |
1111 | * These are the number paths of length 1 to 5, that we are allowing to emanate | |
1112 | * from a single file of interest. For example, we allow 1000 paths of length | |
1113 | * 1, to emanate from each file of interest. This essentially represents the | |
1114 | * potential wakeup paths, which need to be limited in order to avoid massive | |
1115 | * uncontrolled wakeup storms. The common use case should be a single ep which | |
1116 | * is connected to n file sources. In this case each file source has 1 path | |
1117 | * of length 1. Thus, the numbers below should be more than sufficient. These | |
1118 | * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify | |
1119 | * and delete can't add additional paths. Protected by the epmutex. | |
1120 | */ | |
1121 | static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 }; | |
1122 | static int path_count[PATH_ARR_SIZE]; | |
1123 | ||
1124 | static int path_count_inc(int nests) | |
1125 | { | |
1126 | /* Allow an arbitrary number of depth 1 paths */ | |
1127 | if (nests == 0) | |
1128 | return 0; | |
1129 | ||
1130 | if (++path_count[nests] > path_limits[nests]) | |
1131 | return -1; | |
1132 | return 0; | |
1133 | } | |
1134 | ||
1135 | static void path_count_init(void) | |
1136 | { | |
1137 | int i; | |
1138 | ||
1139 | for (i = 0; i < PATH_ARR_SIZE; i++) | |
1140 | path_count[i] = 0; | |
1141 | } | |
1142 | ||
1143 | static int reverse_path_check_proc(void *priv, void *cookie, int call_nests) | |
1144 | { | |
1145 | int error = 0; | |
1146 | struct file *file = priv; | |
1147 | struct file *child_file; | |
1148 | struct epitem *epi; | |
1149 | ||
1150 | /* CTL_DEL can remove links here, but that can't increase our count */ | |
1151 | rcu_read_lock(); | |
1152 | list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) { | |
1153 | child_file = epi->ep->file; | |
1154 | if (is_file_epoll(child_file)) { | |
1155 | if (list_empty(&child_file->f_ep_links)) { | |
1156 | if (path_count_inc(call_nests)) { | |
1157 | error = -1; | |
1158 | break; | |
1159 | } | |
1160 | } else { | |
1161 | error = ep_call_nested(&poll_loop_ncalls, | |
1162 | EP_MAX_NESTS, | |
1163 | reverse_path_check_proc, | |
1164 | child_file, child_file, | |
1165 | current); | |
1166 | } | |
1167 | if (error != 0) | |
1168 | break; | |
1169 | } else { | |
1170 | printk(KERN_ERR "reverse_path_check_proc: " | |
1171 | "file is not an ep!\n"); | |
1172 | } | |
1173 | } | |
1174 | rcu_read_unlock(); | |
1175 | return error; | |
1176 | } | |
1177 | ||
1178 | /** | |
1179 | * reverse_path_check - The tfile_check_list is list of file *, which have | |
1180 | * links that are proposed to be newly added. We need to | |
1181 | * make sure that those added links don't add too many | |
1182 | * paths such that we will spend all our time waking up | |
1183 | * eventpoll objects. | |
1184 | * | |
1185 | * Returns: Returns zero if the proposed links don't create too many paths, | |
1186 | * -1 otherwise. | |
1187 | */ | |
1188 | static int reverse_path_check(void) | |
1189 | { | |
1190 | int error = 0; | |
1191 | struct file *current_file; | |
1192 | ||
1193 | /* let's call this for all tfiles */ | |
1194 | list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) { | |
1195 | path_count_init(); | |
1196 | error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS, | |
1197 | reverse_path_check_proc, current_file, | |
1198 | current_file, current); | |
1199 | if (error) | |
1200 | break; | |
1201 | } | |
1202 | return error; | |
1203 | } | |
1204 | ||
1205 | static int ep_create_wakeup_source(struct epitem *epi) | |
1206 | { | |
1207 | const char *name; | |
1208 | struct wakeup_source *ws; | |
1209 | ||
1210 | if (!epi->ep->ws) { | |
1211 | epi->ep->ws = wakeup_source_register("eventpoll"); | |
1212 | if (!epi->ep->ws) | |
1213 | return -ENOMEM; | |
1214 | } | |
1215 | ||
1216 | name = epi->ffd.file->f_path.dentry->d_name.name; | |
1217 | ws = wakeup_source_register(name); | |
1218 | ||
1219 | if (!ws) | |
1220 | return -ENOMEM; | |
1221 | rcu_assign_pointer(epi->ws, ws); | |
1222 | ||
1223 | return 0; | |
1224 | } | |
1225 | ||
1226 | /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */ | |
1227 | static noinline void ep_destroy_wakeup_source(struct epitem *epi) | |
1228 | { | |
1229 | struct wakeup_source *ws = ep_wakeup_source(epi); | |
1230 | ||
1231 | RCU_INIT_POINTER(epi->ws, NULL); | |
1232 | ||
1233 | /* | |
1234 | * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is | |
1235 | * used internally by wakeup_source_remove, too (called by | |
1236 | * wakeup_source_unregister), so we cannot use call_rcu | |
1237 | */ | |
1238 | synchronize_rcu(); | |
1239 | wakeup_source_unregister(ws); | |
1240 | } | |
1241 | ||
1242 | /* | |
1243 | * Must be called with "mtx" held. | |
1244 | */ | |
1245 | static int ep_insert(struct eventpoll *ep, struct epoll_event *event, | |
1246 | struct file *tfile, int fd) | |
1247 | { | |
1248 | int error, revents, pwake = 0; | |
1249 | unsigned long flags; | |
1250 | long user_watches; | |
1251 | struct epitem *epi; | |
1252 | struct ep_pqueue epq; | |
1253 | ||
1254 | user_watches = atomic_long_read(&ep->user->epoll_watches); | |
1255 | if (unlikely(user_watches >= max_user_watches)) | |
1256 | return -ENOSPC; | |
1257 | if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL))) | |
1258 | return -ENOMEM; | |
1259 | ||
1260 | /* Item initialization follow here ... */ | |
1261 | INIT_LIST_HEAD(&epi->rdllink); | |
1262 | INIT_LIST_HEAD(&epi->fllink); | |
1263 | INIT_LIST_HEAD(&epi->pwqlist); | |
1264 | epi->ep = ep; | |
1265 | ep_set_ffd(&epi->ffd, tfile, fd); | |
1266 | epi->event = *event; | |
1267 | epi->nwait = 0; | |
1268 | epi->next = EP_UNACTIVE_PTR; | |
1269 | if (epi->event.events & EPOLLWAKEUP) { | |
1270 | error = ep_create_wakeup_source(epi); | |
1271 | if (error) | |
1272 | goto error_create_wakeup_source; | |
1273 | } else { | |
1274 | RCU_INIT_POINTER(epi->ws, NULL); | |
1275 | } | |
1276 | ||
1277 | /* Initialize the poll table using the queue callback */ | |
1278 | epq.epi = epi; | |
1279 | init_poll_funcptr(&epq.pt, ep_ptable_queue_proc); | |
1280 | ||
1281 | /* | |
1282 | * Attach the item to the poll hooks and get current event bits. | |
1283 | * We can safely use the file* here because its usage count has | |
1284 | * been increased by the caller of this function. Note that after | |
1285 | * this operation completes, the poll callback can start hitting | |
1286 | * the new item. | |
1287 | */ | |
1288 | revents = ep_item_poll(epi, &epq.pt); | |
1289 | ||
1290 | /* | |
1291 | * We have to check if something went wrong during the poll wait queue | |
1292 | * install process. Namely an allocation for a wait queue failed due | |
1293 | * high memory pressure. | |
1294 | */ | |
1295 | error = -ENOMEM; | |
1296 | if (epi->nwait < 0) | |
1297 | goto error_unregister; | |
1298 | ||
1299 | /* Add the current item to the list of active epoll hook for this file */ | |
1300 | spin_lock(&tfile->f_lock); | |
1301 | list_add_tail_rcu(&epi->fllink, &tfile->f_ep_links); | |
1302 | spin_unlock(&tfile->f_lock); | |
1303 | ||
1304 | /* | |
1305 | * Add the current item to the RB tree. All RB tree operations are | |
1306 | * protected by "mtx", and ep_insert() is called with "mtx" held. | |
1307 | */ | |
1308 | ep_rbtree_insert(ep, epi); | |
1309 | ||
1310 | /* now check if we've created too many backpaths */ | |
1311 | error = -EINVAL; | |
1312 | if (reverse_path_check()) | |
1313 | goto error_remove_epi; | |
1314 | ||
1315 | /* We have to drop the new item inside our item list to keep track of it */ | |
1316 | spin_lock_irqsave(&ep->lock, flags); | |
1317 | ||
1318 | /* If the file is already "ready" we drop it inside the ready list */ | |
1319 | if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) { | |
1320 | list_add_tail(&epi->rdllink, &ep->rdllist); | |
1321 | ep_pm_stay_awake(epi); | |
1322 | ||
1323 | /* Notify waiting tasks that events are available */ | |
1324 | if (waitqueue_active(&ep->wq)) | |
1325 | wake_up_locked(&ep->wq); | |
1326 | if (waitqueue_active(&ep->poll_wait)) | |
1327 | pwake++; | |
1328 | } | |
1329 | ||
1330 | spin_unlock_irqrestore(&ep->lock, flags); | |
1331 | ||
1332 | atomic_long_inc(&ep->user->epoll_watches); | |
1333 | ||
1334 | /* We have to call this outside the lock */ | |
1335 | if (pwake) | |
1336 | ep_poll_safewake(&ep->poll_wait); | |
1337 | ||
1338 | return 0; | |
1339 | ||
1340 | error_remove_epi: | |
1341 | spin_lock(&tfile->f_lock); | |
1342 | list_del_rcu(&epi->fllink); | |
1343 | spin_unlock(&tfile->f_lock); | |
1344 | ||
1345 | rb_erase(&epi->rbn, &ep->rbr); | |
1346 | ||
1347 | error_unregister: | |
1348 | ep_unregister_pollwait(ep, epi); | |
1349 | ||
1350 | /* | |
1351 | * We need to do this because an event could have been arrived on some | |
1352 | * allocated wait queue. Note that we don't care about the ep->ovflist | |
1353 | * list, since that is used/cleaned only inside a section bound by "mtx". | |
1354 | * And ep_insert() is called with "mtx" held. | |
1355 | */ | |
1356 | spin_lock_irqsave(&ep->lock, flags); | |
1357 | if (ep_is_linked(&epi->rdllink)) | |
1358 | list_del_init(&epi->rdllink); | |
1359 | spin_unlock_irqrestore(&ep->lock, flags); | |
1360 | ||
1361 | wakeup_source_unregister(ep_wakeup_source(epi)); | |
1362 | ||
1363 | error_create_wakeup_source: | |
1364 | kmem_cache_free(epi_cache, epi); | |
1365 | ||
1366 | return error; | |
1367 | } | |
1368 | ||
1369 | /* | |
1370 | * Modify the interest event mask by dropping an event if the new mask | |
1371 | * has a match in the current file status. Must be called with "mtx" held. | |
1372 | */ | |
1373 | static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event) | |
1374 | { | |
1375 | int pwake = 0; | |
1376 | unsigned int revents; | |
1377 | poll_table pt; | |
1378 | ||
1379 | init_poll_funcptr(&pt, NULL); | |
1380 | ||
1381 | /* | |
1382 | * Set the new event interest mask before calling f_op->poll(); | |
1383 | * otherwise we might miss an event that happens between the | |
1384 | * f_op->poll() call and the new event set registering. | |
1385 | */ | |
1386 | epi->event.events = event->events; /* need barrier below */ | |
1387 | epi->event.data = event->data; /* protected by mtx */ | |
1388 | if (epi->event.events & EPOLLWAKEUP) { | |
1389 | if (!ep_has_wakeup_source(epi)) | |
1390 | ep_create_wakeup_source(epi); | |
1391 | } else if (ep_has_wakeup_source(epi)) { | |
1392 | ep_destroy_wakeup_source(epi); | |
1393 | } | |
1394 | ||
1395 | /* | |
1396 | * The following barrier has two effects: | |
1397 | * | |
1398 | * 1) Flush epi changes above to other CPUs. This ensures | |
1399 | * we do not miss events from ep_poll_callback if an | |
1400 | * event occurs immediately after we call f_op->poll(). | |
1401 | * We need this because we did not take ep->lock while | |
1402 | * changing epi above (but ep_poll_callback does take | |
1403 | * ep->lock). | |
1404 | * | |
1405 | * 2) We also need to ensure we do not miss _past_ events | |
1406 | * when calling f_op->poll(). This barrier also | |
1407 | * pairs with the barrier in wq_has_sleeper (see | |
1408 | * comments for wq_has_sleeper). | |
1409 | * | |
1410 | * This barrier will now guarantee ep_poll_callback or f_op->poll | |
1411 | * (or both) will notice the readiness of an item. | |
1412 | */ | |
1413 | smp_mb(); | |
1414 | ||
1415 | /* | |
1416 | * Get current event bits. We can safely use the file* here because | |
1417 | * its usage count has been increased by the caller of this function. | |
1418 | */ | |
1419 | revents = ep_item_poll(epi, &pt); | |
1420 | ||
1421 | /* | |
1422 | * If the item is "hot" and it is not registered inside the ready | |
1423 | * list, push it inside. | |
1424 | */ | |
1425 | if (revents & event->events) { | |
1426 | spin_lock_irq(&ep->lock); | |
1427 | if (!ep_is_linked(&epi->rdllink)) { | |
1428 | list_add_tail(&epi->rdllink, &ep->rdllist); | |
1429 | ep_pm_stay_awake(epi); | |
1430 | ||
1431 | /* Notify waiting tasks that events are available */ | |
1432 | if (waitqueue_active(&ep->wq)) | |
1433 | wake_up_locked(&ep->wq); | |
1434 | if (waitqueue_active(&ep->poll_wait)) | |
1435 | pwake++; | |
1436 | } | |
1437 | spin_unlock_irq(&ep->lock); | |
1438 | } | |
1439 | ||
1440 | /* We have to call this outside the lock */ | |
1441 | if (pwake) | |
1442 | ep_poll_safewake(&ep->poll_wait); | |
1443 | ||
1444 | return 0; | |
1445 | } | |
1446 | ||
1447 | static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head, | |
1448 | void *priv) | |
1449 | { | |
1450 | struct ep_send_events_data *esed = priv; | |
1451 | int eventcnt; | |
1452 | unsigned int revents; | |
1453 | struct epitem *epi; | |
1454 | struct epoll_event __user *uevent; | |
1455 | struct wakeup_source *ws; | |
1456 | poll_table pt; | |
1457 | ||
1458 | init_poll_funcptr(&pt, NULL); | |
1459 | ||
1460 | /* | |
1461 | * We can loop without lock because we are passed a task private list. | |
1462 | * Items cannot vanish during the loop because ep_scan_ready_list() is | |
1463 | * holding "mtx" during this call. | |
1464 | */ | |
1465 | for (eventcnt = 0, uevent = esed->events; | |
1466 | !list_empty(head) && eventcnt < esed->maxevents;) { | |
1467 | epi = list_first_entry(head, struct epitem, rdllink); | |
1468 | ||
1469 | /* | |
1470 | * Activate ep->ws before deactivating epi->ws to prevent | |
1471 | * triggering auto-suspend here (in case we reactive epi->ws | |
1472 | * below). | |
1473 | * | |
1474 | * This could be rearranged to delay the deactivation of epi->ws | |
1475 | * instead, but then epi->ws would temporarily be out of sync | |
1476 | * with ep_is_linked(). | |
1477 | */ | |
1478 | ws = ep_wakeup_source(epi); | |
1479 | if (ws) { | |
1480 | if (ws->active) | |
1481 | __pm_stay_awake(ep->ws); | |
1482 | __pm_relax(ws); | |
1483 | } | |
1484 | ||
1485 | list_del_init(&epi->rdllink); | |
1486 | ||
1487 | revents = ep_item_poll(epi, &pt); | |
1488 | ||
1489 | /* | |
1490 | * If the event mask intersect the caller-requested one, | |
1491 | * deliver the event to userspace. Again, ep_scan_ready_list() | |
1492 | * is holding "mtx", so no operations coming from userspace | |
1493 | * can change the item. | |
1494 | */ | |
1495 | if (revents) { | |
1496 | if (__put_user(revents, &uevent->events) || | |
1497 | __put_user(epi->event.data, &uevent->data)) { | |
1498 | list_add(&epi->rdllink, head); | |
1499 | ep_pm_stay_awake(epi); | |
1500 | return eventcnt ? eventcnt : -EFAULT; | |
1501 | } | |
1502 | eventcnt++; | |
1503 | uevent++; | |
1504 | if (epi->event.events & EPOLLONESHOT) | |
1505 | epi->event.events &= EP_PRIVATE_BITS; | |
1506 | else if (!(epi->event.events & EPOLLET)) { | |
1507 | /* | |
1508 | * If this file has been added with Level | |
1509 | * Trigger mode, we need to insert back inside | |
1510 | * the ready list, so that the next call to | |
1511 | * epoll_wait() will check again the events | |
1512 | * availability. At this point, no one can insert | |
1513 | * into ep->rdllist besides us. The epoll_ctl() | |
1514 | * callers are locked out by | |
1515 | * ep_scan_ready_list() holding "mtx" and the | |
1516 | * poll callback will queue them in ep->ovflist. | |
1517 | */ | |
1518 | list_add_tail(&epi->rdllink, &ep->rdllist); | |
1519 | ep_pm_stay_awake(epi); | |
1520 | } | |
1521 | } | |
1522 | } | |
1523 | ||
1524 | return eventcnt; | |
1525 | } | |
1526 | ||
1527 | static int ep_send_events(struct eventpoll *ep, | |
1528 | struct epoll_event __user *events, int maxevents) | |
1529 | { | |
1530 | struct ep_send_events_data esed; | |
1531 | ||
1532 | esed.maxevents = maxevents; | |
1533 | esed.events = events; | |
1534 | ||
1535 | return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0); | |
1536 | } | |
1537 | ||
1538 | static inline struct timespec ep_set_mstimeout(long ms) | |
1539 | { | |
1540 | struct timespec now, ts = { | |
1541 | .tv_sec = ms / MSEC_PER_SEC, | |
1542 | .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC), | |
1543 | }; | |
1544 | ||
1545 | ktime_get_ts(&now); | |
1546 | return timespec_add_safe(now, ts); | |
1547 | } | |
1548 | ||
1549 | /** | |
1550 | * ep_poll - Retrieves ready events, and delivers them to the caller supplied | |
1551 | * event buffer. | |
1552 | * | |
1553 | * @ep: Pointer to the eventpoll context. | |
1554 | * @events: Pointer to the userspace buffer where the ready events should be | |
1555 | * stored. | |
1556 | * @maxevents: Size (in terms of number of events) of the caller event buffer. | |
1557 | * @timeout: Maximum timeout for the ready events fetch operation, in | |
1558 | * milliseconds. If the @timeout is zero, the function will not block, | |
1559 | * while if the @timeout is less than zero, the function will block | |
1560 | * until at least one event has been retrieved (or an error | |
1561 | * occurred). | |
1562 | * | |
1563 | * Returns: Returns the number of ready events which have been fetched, or an | |
1564 | * error code, in case of error. | |
1565 | */ | |
1566 | static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events, | |
1567 | int maxevents, long timeout) | |
1568 | { | |
1569 | int res = 0, eavail, timed_out = 0; | |
1570 | unsigned long flags; | |
1571 | long slack = 0; | |
1572 | wait_queue_t wait; | |
1573 | ktime_t expires, *to = NULL; | |
1574 | ||
1575 | if (timeout > 0) { | |
1576 | struct timespec end_time = ep_set_mstimeout(timeout); | |
1577 | ||
1578 | slack = select_estimate_accuracy(&end_time); | |
1579 | to = &expires; | |
1580 | *to = timespec_to_ktime(end_time); | |
1581 | } else if (timeout == 0) { | |
1582 | /* | |
1583 | * Avoid the unnecessary trip to the wait queue loop, if the | |
1584 | * caller specified a non blocking operation. | |
1585 | */ | |
1586 | timed_out = 1; | |
1587 | spin_lock_irqsave(&ep->lock, flags); | |
1588 | goto check_events; | |
1589 | } | |
1590 | ||
1591 | fetch_events: | |
1592 | spin_lock_irqsave(&ep->lock, flags); | |
1593 | ||
1594 | if (!ep_events_available(ep)) { | |
1595 | /* | |
1596 | * We don't have any available event to return to the caller. | |
1597 | * We need to sleep here, and we will be wake up by | |
1598 | * ep_poll_callback() when events will become available. | |
1599 | */ | |
1600 | init_waitqueue_entry(&wait, current); | |
1601 | __add_wait_queue_exclusive(&ep->wq, &wait); | |
1602 | ||
1603 | for (;;) { | |
1604 | /* | |
1605 | * We don't want to sleep if the ep_poll_callback() sends us | |
1606 | * a wakeup in between. That's why we set the task state | |
1607 | * to TASK_INTERRUPTIBLE before doing the checks. | |
1608 | */ | |
1609 | set_current_state(TASK_INTERRUPTIBLE); | |
1610 | if (ep_events_available(ep) || timed_out) | |
1611 | break; | |
1612 | if (signal_pending(current)) { | |
1613 | res = -EINTR; | |
1614 | break; | |
1615 | } | |
1616 | ||
1617 | spin_unlock_irqrestore(&ep->lock, flags); | |
1618 | if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS)) | |
1619 | timed_out = 1; | |
1620 | ||
1621 | spin_lock_irqsave(&ep->lock, flags); | |
1622 | } | |
1623 | __remove_wait_queue(&ep->wq, &wait); | |
1624 | ||
1625 | set_current_state(TASK_RUNNING); | |
1626 | } | |
1627 | check_events: | |
1628 | /* Is it worth to try to dig for events ? */ | |
1629 | eavail = ep_events_available(ep); | |
1630 | ||
1631 | spin_unlock_irqrestore(&ep->lock, flags); | |
1632 | ||
1633 | /* | |
1634 | * Try to transfer events to user space. In case we get 0 events and | |
1635 | * there's still timeout left over, we go trying again in search of | |
1636 | * more luck. | |
1637 | */ | |
1638 | if (!res && eavail && | |
1639 | !(res = ep_send_events(ep, events, maxevents)) && !timed_out) | |
1640 | goto fetch_events; | |
1641 | ||
1642 | return res; | |
1643 | } | |
1644 | ||
1645 | /** | |
1646 | * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested() | |
1647 | * API, to verify that adding an epoll file inside another | |
1648 | * epoll structure, does not violate the constraints, in | |
1649 | * terms of closed loops, or too deep chains (which can | |
1650 | * result in excessive stack usage). | |
1651 | * | |
1652 | * @priv: Pointer to the epoll file to be currently checked. | |
1653 | * @cookie: Original cookie for this call. This is the top-of-the-chain epoll | |
1654 | * data structure pointer. | |
1655 | * @call_nests: Current dept of the @ep_call_nested() call stack. | |
1656 | * | |
1657 | * Returns: Returns zero if adding the epoll @file inside current epoll | |
1658 | * structure @ep does not violate the constraints, or -1 otherwise. | |
1659 | */ | |
1660 | static int ep_loop_check_proc(void *priv, void *cookie, int call_nests) | |
1661 | { | |
1662 | int error = 0; | |
1663 | struct file *file = priv; | |
1664 | struct eventpoll *ep = file->private_data; | |
1665 | struct eventpoll *ep_tovisit; | |
1666 | struct rb_node *rbp; | |
1667 | struct epitem *epi; | |
1668 | ||
1669 | mutex_lock_nested(&ep->mtx, call_nests + 1); | |
1670 | ep->visited = 1; | |
1671 | list_add(&ep->visited_list_link, &visited_list); | |
1672 | for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) { | |
1673 | epi = rb_entry(rbp, struct epitem, rbn); | |
1674 | if (unlikely(is_file_epoll(epi->ffd.file))) { | |
1675 | ep_tovisit = epi->ffd.file->private_data; | |
1676 | if (ep_tovisit->visited) | |
1677 | continue; | |
1678 | error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS, | |
1679 | ep_loop_check_proc, epi->ffd.file, | |
1680 | ep_tovisit, current); | |
1681 | if (error != 0) | |
1682 | break; | |
1683 | } else { | |
1684 | /* | |
1685 | * If we've reached a file that is not associated with | |
1686 | * an ep, then we need to check if the newly added | |
1687 | * links are going to add too many wakeup paths. We do | |
1688 | * this by adding it to the tfile_check_list, if it's | |
1689 | * not already there, and calling reverse_path_check() | |
1690 | * during ep_insert(). | |
1691 | */ | |
1692 | if (list_empty(&epi->ffd.file->f_tfile_llink)) | |
1693 | list_add(&epi->ffd.file->f_tfile_llink, | |
1694 | &tfile_check_list); | |
1695 | } | |
1696 | } | |
1697 | mutex_unlock(&ep->mtx); | |
1698 | ||
1699 | return error; | |
1700 | } | |
1701 | ||
1702 | /** | |
1703 | * ep_loop_check - Performs a check to verify that adding an epoll file (@file) | |
1704 | * another epoll file (represented by @ep) does not create | |
1705 | * closed loops or too deep chains. | |
1706 | * | |
1707 | * @ep: Pointer to the epoll private data structure. | |
1708 | * @file: Pointer to the epoll file to be checked. | |
1709 | * | |
1710 | * Returns: Returns zero if adding the epoll @file inside current epoll | |
1711 | * structure @ep does not violate the constraints, or -1 otherwise. | |
1712 | */ | |
1713 | static int ep_loop_check(struct eventpoll *ep, struct file *file) | |
1714 | { | |
1715 | int ret; | |
1716 | struct eventpoll *ep_cur, *ep_next; | |
1717 | ||
1718 | ret = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS, | |
1719 | ep_loop_check_proc, file, ep, current); | |
1720 | /* clear visited list */ | |
1721 | list_for_each_entry_safe(ep_cur, ep_next, &visited_list, | |
1722 | visited_list_link) { | |
1723 | ep_cur->visited = 0; | |
1724 | list_del(&ep_cur->visited_list_link); | |
1725 | } | |
1726 | return ret; | |
1727 | } | |
1728 | ||
1729 | static void clear_tfile_check_list(void) | |
1730 | { | |
1731 | struct file *file; | |
1732 | ||
1733 | /* first clear the tfile_check_list */ | |
1734 | while (!list_empty(&tfile_check_list)) { | |
1735 | file = list_first_entry(&tfile_check_list, struct file, | |
1736 | f_tfile_llink); | |
1737 | list_del_init(&file->f_tfile_llink); | |
1738 | } | |
1739 | INIT_LIST_HEAD(&tfile_check_list); | |
1740 | } | |
1741 | ||
1742 | /* | |
1743 | * Open an eventpoll file descriptor. | |
1744 | */ | |
1745 | SYSCALL_DEFINE1(epoll_create1, int, flags) | |
1746 | { | |
1747 | int error, fd; | |
1748 | struct eventpoll *ep = NULL; | |
1749 | struct file *file; | |
1750 | ||
1751 | /* Check the EPOLL_* constant for consistency. */ | |
1752 | BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC); | |
1753 | ||
1754 | if (flags & ~EPOLL_CLOEXEC) | |
1755 | return -EINVAL; | |
1756 | /* | |
1757 | * Create the internal data structure ("struct eventpoll"). | |
1758 | */ | |
1759 | error = ep_alloc(&ep); | |
1760 | if (error < 0) | |
1761 | return error; | |
1762 | /* | |
1763 | * Creates all the items needed to setup an eventpoll file. That is, | |
1764 | * a file structure and a free file descriptor. | |
1765 | */ | |
1766 | fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC)); | |
1767 | if (fd < 0) { | |
1768 | error = fd; | |
1769 | goto out_free_ep; | |
1770 | } | |
1771 | file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep, | |
1772 | O_RDWR | (flags & O_CLOEXEC)); | |
1773 | if (IS_ERR(file)) { | |
1774 | error = PTR_ERR(file); | |
1775 | goto out_free_fd; | |
1776 | } | |
1777 | ep->file = file; | |
1778 | fd_install(fd, file); | |
1779 | return fd; | |
1780 | ||
1781 | out_free_fd: | |
1782 | put_unused_fd(fd); | |
1783 | out_free_ep: | |
1784 | ep_free(ep); | |
1785 | return error; | |
1786 | } | |
1787 | ||
1788 | SYSCALL_DEFINE1(epoll_create, int, size) | |
1789 | { | |
1790 | if (size <= 0) | |
1791 | return -EINVAL; | |
1792 | ||
1793 | return sys_epoll_create1(0); | |
1794 | } | |
1795 | ||
1796 | /* | |
1797 | * The following function implements the controller interface for | |
1798 | * the eventpoll file that enables the insertion/removal/change of | |
1799 | * file descriptors inside the interest set. | |
1800 | */ | |
1801 | SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd, | |
1802 | struct epoll_event __user *, event) | |
1803 | { | |
1804 | int error; | |
1805 | int did_lock_epmutex = 0; | |
1806 | struct fd f, tf; | |
1807 | struct eventpoll *ep; | |
1808 | struct epitem *epi; | |
1809 | struct epoll_event epds; | |
1810 | ||
1811 | error = -EFAULT; | |
1812 | if (ep_op_has_event(op) && | |
1813 | copy_from_user(&epds, event, sizeof(struct epoll_event))) | |
1814 | goto error_return; | |
1815 | ||
1816 | error = -EBADF; | |
1817 | f = fdget(epfd); | |
1818 | if (!f.file) | |
1819 | goto error_return; | |
1820 | ||
1821 | /* Get the "struct file *" for the target file */ | |
1822 | tf = fdget(fd); | |
1823 | if (!tf.file) | |
1824 | goto error_fput; | |
1825 | ||
1826 | /* The target file descriptor must support poll */ | |
1827 | error = -EPERM; | |
1828 | if (!tf.file->f_op || !tf.file->f_op->poll) | |
1829 | goto error_tgt_fput; | |
1830 | ||
1831 | /* Check if EPOLLWAKEUP is allowed */ | |
1832 | if ((epds.events & EPOLLWAKEUP) && !capable(CAP_BLOCK_SUSPEND)) | |
1833 | epds.events &= ~EPOLLWAKEUP; | |
1834 | ||
1835 | /* | |
1836 | * We have to check that the file structure underneath the file descriptor | |
1837 | * the user passed to us _is_ an eventpoll file. And also we do not permit | |
1838 | * adding an epoll file descriptor inside itself. | |
1839 | */ | |
1840 | error = -EINVAL; | |
1841 | if (f.file == tf.file || !is_file_epoll(f.file)) | |
1842 | goto error_tgt_fput; | |
1843 | ||
1844 | /* | |
1845 | * At this point it is safe to assume that the "private_data" contains | |
1846 | * our own data structure. | |
1847 | */ | |
1848 | ep = f.file->private_data; | |
1849 | ||
1850 | /* | |
1851 | * When we insert an epoll file descriptor, inside another epoll file | |
1852 | * descriptor, there is the change of creating closed loops, which are | |
1853 | * better be handled here, than in more critical paths. While we are | |
1854 | * checking for loops we also determine the list of files reachable | |
1855 | * and hang them on the tfile_check_list, so we can check that we | |
1856 | * haven't created too many possible wakeup paths. | |
1857 | * | |
1858 | * We need to hold the epmutex across ep_insert to prevent | |
1859 | * multple adds from creating loops in parallel. | |
1860 | */ | |
1861 | if (op == EPOLL_CTL_ADD) { | |
1862 | mutex_lock(&epmutex); | |
1863 | did_lock_epmutex = 1; | |
1864 | if (is_file_epoll(tf.file)) { | |
1865 | error = -ELOOP; | |
1866 | if (ep_loop_check(ep, tf.file) != 0) { | |
1867 | clear_tfile_check_list(); | |
1868 | goto error_tgt_fput; | |
1869 | } | |
1870 | } else | |
1871 | list_add(&tf.file->f_tfile_llink, &tfile_check_list); | |
1872 | } | |
1873 | ||
1874 | mutex_lock_nested(&ep->mtx, 0); | |
1875 | ||
1876 | /* | |
1877 | * Try to lookup the file inside our RB tree, Since we grabbed "mtx" | |
1878 | * above, we can be sure to be able to use the item looked up by | |
1879 | * ep_find() till we release the mutex. | |
1880 | */ | |
1881 | epi = ep_find(ep, tf.file, fd); | |
1882 | ||
1883 | error = -EINVAL; | |
1884 | switch (op) { | |
1885 | case EPOLL_CTL_ADD: | |
1886 | if (!epi) { | |
1887 | epds.events |= POLLERR | POLLHUP; | |
1888 | error = ep_insert(ep, &epds, tf.file, fd); | |
1889 | } else | |
1890 | error = -EEXIST; | |
1891 | clear_tfile_check_list(); | |
1892 | break; | |
1893 | case EPOLL_CTL_DEL: | |
1894 | if (epi) | |
1895 | error = ep_remove(ep, epi); | |
1896 | else | |
1897 | error = -ENOENT; | |
1898 | break; | |
1899 | case EPOLL_CTL_MOD: | |
1900 | if (epi) { | |
1901 | epds.events |= POLLERR | POLLHUP; | |
1902 | error = ep_modify(ep, epi, &epds); | |
1903 | } else | |
1904 | error = -ENOENT; | |
1905 | break; | |
1906 | } | |
1907 | mutex_unlock(&ep->mtx); | |
1908 | ||
1909 | error_tgt_fput: | |
1910 | if (did_lock_epmutex) | |
1911 | mutex_unlock(&epmutex); | |
1912 | ||
1913 | fdput(tf); | |
1914 | error_fput: | |
1915 | fdput(f); | |
1916 | error_return: | |
1917 | ||
1918 | return error; | |
1919 | } | |
1920 | ||
1921 | /* | |
1922 | * Implement the event wait interface for the eventpoll file. It is the kernel | |
1923 | * part of the user space epoll_wait(2). | |
1924 | */ | |
1925 | SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events, | |
1926 | int, maxevents, int, timeout) | |
1927 | { | |
1928 | int error; | |
1929 | struct fd f; | |
1930 | struct eventpoll *ep; | |
1931 | ||
1932 | /* The maximum number of event must be greater than zero */ | |
1933 | if (maxevents <= 0 || maxevents > EP_MAX_EVENTS) | |
1934 | return -EINVAL; | |
1935 | ||
1936 | /* Verify that the area passed by the user is writeable */ | |
1937 | if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) | |
1938 | return -EFAULT; | |
1939 | ||
1940 | /* Get the "struct file *" for the eventpoll file */ | |
1941 | f = fdget(epfd); | |
1942 | if (!f.file) | |
1943 | return -EBADF; | |
1944 | ||
1945 | /* | |
1946 | * We have to check that the file structure underneath the fd | |
1947 | * the user passed to us _is_ an eventpoll file. | |
1948 | */ | |
1949 | error = -EINVAL; | |
1950 | if (!is_file_epoll(f.file)) | |
1951 | goto error_fput; | |
1952 | ||
1953 | /* | |
1954 | * At this point it is safe to assume that the "private_data" contains | |
1955 | * our own data structure. | |
1956 | */ | |
1957 | ep = f.file->private_data; | |
1958 | ||
1959 | /* Time to fish for events ... */ | |
1960 | error = ep_poll(ep, events, maxevents, timeout); | |
1961 | ||
1962 | error_fput: | |
1963 | fdput(f); | |
1964 | return error; | |
1965 | } | |
1966 | ||
1967 | /* | |
1968 | * Implement the event wait interface for the eventpoll file. It is the kernel | |
1969 | * part of the user space epoll_pwait(2). | |
1970 | */ | |
1971 | SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events, | |
1972 | int, maxevents, int, timeout, const sigset_t __user *, sigmask, | |
1973 | size_t, sigsetsize) | |
1974 | { | |
1975 | int error; | |
1976 | sigset_t ksigmask, sigsaved; | |
1977 | ||
1978 | /* | |
1979 | * If the caller wants a certain signal mask to be set during the wait, | |
1980 | * we apply it here. | |
1981 | */ | |
1982 | if (sigmask) { | |
1983 | if (sigsetsize != sizeof(sigset_t)) | |
1984 | return -EINVAL; | |
1985 | if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask))) | |
1986 | return -EFAULT; | |
1987 | sigsaved = current->blocked; | |
1988 | set_current_blocked(&ksigmask); | |
1989 | } | |
1990 | ||
1991 | error = sys_epoll_wait(epfd, events, maxevents, timeout); | |
1992 | ||
1993 | /* | |
1994 | * If we changed the signal mask, we need to restore the original one. | |
1995 | * In case we've got a signal while waiting, we do not restore the | |
1996 | * signal mask yet, and we allow do_signal() to deliver the signal on | |
1997 | * the way back to userspace, before the signal mask is restored. | |
1998 | */ | |
1999 | if (sigmask) { | |
2000 | if (error == -EINTR) { | |
2001 | memcpy(¤t->saved_sigmask, &sigsaved, | |
2002 | sizeof(sigsaved)); | |
2003 | set_restore_sigmask(); | |
2004 | } else | |
2005 | set_current_blocked(&sigsaved); | |
2006 | } | |
2007 | ||
2008 | return error; | |
2009 | } | |
2010 | ||
2011 | #ifdef CONFIG_COMPAT | |
2012 | COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd, | |
2013 | struct epoll_event __user *, events, | |
2014 | int, maxevents, int, timeout, | |
2015 | const compat_sigset_t __user *, sigmask, | |
2016 | compat_size_t, sigsetsize) | |
2017 | { | |
2018 | long err; | |
2019 | compat_sigset_t csigmask; | |
2020 | sigset_t ksigmask, sigsaved; | |
2021 | ||
2022 | /* | |
2023 | * If the caller wants a certain signal mask to be set during the wait, | |
2024 | * we apply it here. | |
2025 | */ | |
2026 | if (sigmask) { | |
2027 | if (sigsetsize != sizeof(compat_sigset_t)) | |
2028 | return -EINVAL; | |
2029 | if (copy_from_user(&csigmask, sigmask, sizeof(csigmask))) | |
2030 | return -EFAULT; | |
2031 | sigset_from_compat(&ksigmask, &csigmask); | |
2032 | sigsaved = current->blocked; | |
2033 | set_current_blocked(&ksigmask); | |
2034 | } | |
2035 | ||
2036 | err = sys_epoll_wait(epfd, events, maxevents, timeout); | |
2037 | ||
2038 | /* | |
2039 | * If we changed the signal mask, we need to restore the original one. | |
2040 | * In case we've got a signal while waiting, we do not restore the | |
2041 | * signal mask yet, and we allow do_signal() to deliver the signal on | |
2042 | * the way back to userspace, before the signal mask is restored. | |
2043 | */ | |
2044 | if (sigmask) { | |
2045 | if (err == -EINTR) { | |
2046 | memcpy(¤t->saved_sigmask, &sigsaved, | |
2047 | sizeof(sigsaved)); | |
2048 | set_restore_sigmask(); | |
2049 | } else | |
2050 | set_current_blocked(&sigsaved); | |
2051 | } | |
2052 | ||
2053 | return err; | |
2054 | } | |
2055 | #endif | |
2056 | ||
2057 | static int __init eventpoll_init(void) | |
2058 | { | |
2059 | struct sysinfo si; | |
2060 | ||
2061 | si_meminfo(&si); | |
2062 | /* | |
2063 | * Allows top 4% of lomem to be allocated for epoll watches (per user). | |
2064 | */ | |
2065 | max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) / | |
2066 | EP_ITEM_COST; | |
2067 | BUG_ON(max_user_watches < 0); | |
2068 | ||
2069 | /* | |
2070 | * Initialize the structure used to perform epoll file descriptor | |
2071 | * inclusion loops checks. | |
2072 | */ | |
2073 | ep_nested_calls_init(&poll_loop_ncalls); | |
2074 | ||
2075 | /* Initialize the structure used to perform safe poll wait head wake ups */ | |
2076 | ep_nested_calls_init(&poll_safewake_ncalls); | |
2077 | ||
2078 | /* Initialize the structure used to perform file's f_op->poll() calls */ | |
2079 | ep_nested_calls_init(&poll_readywalk_ncalls); | |
2080 | ||
2081 | /* | |
2082 | * We can have many thousands of epitems, so prevent this from | |
2083 | * using an extra cache line on 64-bit (and smaller) CPUs | |
2084 | */ | |
2085 | BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128); | |
2086 | ||
2087 | /* Allocates slab cache used to allocate "struct epitem" items */ | |
2088 | epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem), | |
2089 | 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL); | |
2090 | ||
2091 | /* Allocates slab cache used to allocate "struct eppoll_entry" */ | |
2092 | pwq_cache = kmem_cache_create("eventpoll_pwq", | |
2093 | sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL); | |
2094 | ||
2095 | return 0; | |
2096 | } | |
2097 | fs_initcall(eventpoll_init); |