Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * fs/eventpoll.c ( Efficent event polling implementation ) | |
3419b23a | 3 | * Copyright (C) 2001,...,2006 Davide Libenzi |
1da177e4 LT |
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 | ||
1da177e4 LT |
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> | |
1da177e4 LT |
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/rwsem.h> | |
30 | #include <linux/rbtree.h> | |
31 | #include <linux/wait.h> | |
32 | #include <linux/eventpoll.h> | |
33 | #include <linux/mount.h> | |
34 | #include <linux/bitops.h> | |
144efe3e | 35 | #include <linux/mutex.h> |
da66f7cb | 36 | #include <linux/anon_inodes.h> |
1da177e4 LT |
37 | #include <asm/uaccess.h> |
38 | #include <asm/system.h> | |
39 | #include <asm/io.h> | |
40 | #include <asm/mman.h> | |
41 | #include <asm/atomic.h> | |
42 | #include <asm/semaphore.h> | |
43 | ||
1da177e4 LT |
44 | /* |
45 | * LOCKING: | |
46 | * There are three level of locking required by epoll : | |
47 | * | |
144efe3e | 48 | * 1) epmutex (mutex) |
1da177e4 LT |
49 | * 2) ep->sem (rw_semaphore) |
50 | * 3) ep->lock (rw_lock) | |
51 | * | |
52 | * The acquire order is the one listed above, from 1 to 3. | |
53 | * We need a spinlock (ep->lock) because we manipulate objects | |
54 | * from inside the poll callback, that might be triggered from | |
55 | * a wake_up() that in turn might be called from IRQ context. | |
56 | * So we can't sleep inside the poll callback and hence we need | |
57 | * a spinlock. During the event transfer loop (from kernel to | |
58 | * user space) we could end up sleeping due a copy_to_user(), so | |
59 | * we need a lock that will allow us to sleep. This lock is a | |
60 | * read-write semaphore (ep->sem). It is acquired on read during | |
61 | * the event transfer loop and in write during epoll_ctl(EPOLL_CTL_DEL) | |
62 | * and during eventpoll_release_file(). Then we also need a global | |
63 | * semaphore to serialize eventpoll_release_file() and ep_free(). | |
64 | * This semaphore is acquired by ep_free() during the epoll file | |
65 | * cleanup path and it is also acquired by eventpoll_release_file() | |
66 | * if a file has been pushed inside an epoll set and it is then | |
67 | * close()d without a previous call toepoll_ctl(EPOLL_CTL_DEL). | |
68 | * It is possible to drop the "ep->sem" and to use the global | |
144efe3e | 69 | * semaphore "epmutex" (together with "ep->lock") to have it working, |
1da177e4 | 70 | * but having "ep->sem" will make the interface more scalable. |
144efe3e | 71 | * Events that require holding "epmutex" are very rare, while for |
1da177e4 LT |
72 | * normal operations the epoll private "ep->sem" will guarantee |
73 | * a greater scalability. | |
74 | */ | |
75 | ||
1da177e4 LT |
76 | #define DEBUG_EPOLL 0 |
77 | ||
78 | #if DEBUG_EPOLL > 0 | |
79 | #define DPRINTK(x) printk x | |
80 | #define DNPRINTK(n, x) do { if ((n) <= DEBUG_EPOLL) printk x; } while (0) | |
81 | #else /* #if DEBUG_EPOLL > 0 */ | |
82 | #define DPRINTK(x) (void) 0 | |
83 | #define DNPRINTK(n, x) (void) 0 | |
84 | #endif /* #if DEBUG_EPOLL > 0 */ | |
85 | ||
86 | #define DEBUG_EPI 0 | |
87 | ||
88 | #if DEBUG_EPI != 0 | |
89 | #define EPI_SLAB_DEBUG (SLAB_DEBUG_FREE | SLAB_RED_ZONE /* | SLAB_POISON */) | |
90 | #else /* #if DEBUG_EPI != 0 */ | |
91 | #define EPI_SLAB_DEBUG 0 | |
92 | #endif /* #if DEBUG_EPI != 0 */ | |
93 | ||
94 | /* Epoll private bits inside the event mask */ | |
95 | #define EP_PRIVATE_BITS (EPOLLONESHOT | EPOLLET) | |
96 | ||
97 | /* Maximum number of poll wake up nests we are allowing */ | |
98 | #define EP_MAX_POLLWAKE_NESTS 4 | |
99 | ||
e3306dd5 DL |
100 | /* Maximum msec timeout value storeable in a long int */ |
101 | #define EP_MAX_MSTIMEO min(1000ULL * MAX_SCHEDULE_TIMEOUT / HZ, (LONG_MAX - 999ULL) / HZ) | |
102 | ||
b611967d DL |
103 | #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event)) |
104 | ||
1da177e4 LT |
105 | struct epoll_filefd { |
106 | struct file *file; | |
107 | int fd; | |
108 | }; | |
109 | ||
110 | /* | |
111 | * Node that is linked into the "wake_task_list" member of the "struct poll_safewake". | |
112 | * It is used to keep track on all tasks that are currently inside the wake_up() code | |
113 | * to 1) short-circuit the one coming from the same task and same wait queue head | |
114 | * ( loop ) 2) allow a maximum number of epoll descriptors inclusion nesting | |
115 | * 3) let go the ones coming from other tasks. | |
116 | */ | |
117 | struct wake_task_node { | |
118 | struct list_head llink; | |
36c8b586 | 119 | struct task_struct *task; |
1da177e4 LT |
120 | wait_queue_head_t *wq; |
121 | }; | |
122 | ||
123 | /* | |
124 | * This is used to implement the safe poll wake up avoiding to reenter | |
125 | * the poll callback from inside wake_up(). | |
126 | */ | |
127 | struct poll_safewake { | |
128 | struct list_head wake_task_list; | |
129 | spinlock_t lock; | |
130 | }; | |
131 | ||
132 | /* | |
133 | * This structure is stored inside the "private_data" member of the file | |
134 | * structure and rapresent the main data sructure for the eventpoll | |
135 | * interface. | |
136 | */ | |
137 | struct eventpoll { | |
138 | /* Protect the this structure access */ | |
139 | rwlock_t lock; | |
140 | ||
141 | /* | |
142 | * This semaphore is used to ensure that files are not removed | |
143 | * while epoll is using them. This is read-held during the event | |
144 | * collection loop and it is write-held during the file cleanup | |
145 | * path, the epoll file exit code and the ctl operations. | |
146 | */ | |
147 | struct rw_semaphore sem; | |
148 | ||
149 | /* Wait queue used by sys_epoll_wait() */ | |
150 | wait_queue_head_t wq; | |
151 | ||
152 | /* Wait queue used by file->poll() */ | |
153 | wait_queue_head_t poll_wait; | |
154 | ||
155 | /* List of ready file descriptors */ | |
156 | struct list_head rdllist; | |
157 | ||
158 | /* RB-Tree root used to store monitored fd structs */ | |
159 | struct rb_root rbr; | |
160 | }; | |
161 | ||
162 | /* Wait structure used by the poll hooks */ | |
163 | struct eppoll_entry { | |
164 | /* List header used to link this structure to the "struct epitem" */ | |
165 | struct list_head llink; | |
166 | ||
167 | /* The "base" pointer is set to the container "struct epitem" */ | |
168 | void *base; | |
169 | ||
170 | /* | |
171 | * Wait queue item that will be linked to the target file wait | |
172 | * queue head. | |
173 | */ | |
174 | wait_queue_t wait; | |
175 | ||
176 | /* The wait queue head that linked the "wait" wait queue item */ | |
177 | wait_queue_head_t *whead; | |
178 | }; | |
179 | ||
180 | /* | |
181 | * Each file descriptor added to the eventpoll interface will | |
6192bd53 | 182 | * have an entry of this type linked to the "rbr" RB tree. |
1da177e4 LT |
183 | */ |
184 | struct epitem { | |
185 | /* RB-Tree node used to link this structure to the eventpoll rb-tree */ | |
186 | struct rb_node rbn; | |
187 | ||
188 | /* List header used to link this structure to the eventpoll ready list */ | |
189 | struct list_head rdllink; | |
190 | ||
191 | /* The file descriptor information this item refers to */ | |
192 | struct epoll_filefd ffd; | |
193 | ||
194 | /* Number of active wait queue attached to poll operations */ | |
195 | int nwait; | |
196 | ||
197 | /* List containing poll wait queues */ | |
198 | struct list_head pwqlist; | |
199 | ||
200 | /* The "container" of this item */ | |
201 | struct eventpoll *ep; | |
202 | ||
203 | /* The structure that describe the interested events and the source fd */ | |
204 | struct epoll_event event; | |
205 | ||
206 | /* | |
207 | * Used to keep track of the usage count of the structure. This avoids | |
208 | * that the structure will desappear from underneath our processing. | |
209 | */ | |
210 | atomic_t usecnt; | |
211 | ||
212 | /* List header used to link this item to the "struct file" items list */ | |
213 | struct list_head fllink; | |
1da177e4 LT |
214 | }; |
215 | ||
216 | /* Wrapper struct used by poll queueing */ | |
217 | struct ep_pqueue { | |
218 | poll_table pt; | |
219 | struct epitem *epi; | |
220 | }; | |
221 | ||
1da177e4 LT |
222 | /* |
223 | * This semaphore is used to serialize ep_free() and eventpoll_release_file(). | |
224 | */ | |
144efe3e | 225 | static struct mutex epmutex; |
1da177e4 LT |
226 | |
227 | /* Safe wake up implementation */ | |
228 | static struct poll_safewake psw; | |
229 | ||
230 | /* Slab cache used to allocate "struct epitem" */ | |
e18b890b | 231 | static struct kmem_cache *epi_cache __read_mostly; |
1da177e4 LT |
232 | |
233 | /* Slab cache used to allocate "struct eppoll_entry" */ | |
e18b890b | 234 | static struct kmem_cache *pwq_cache __read_mostly; |
1da177e4 | 235 | |
b030a4dd PE |
236 | |
237 | /* Setup the structure that is used as key for the rb-tree */ | |
238 | static inline void ep_set_ffd(struct epoll_filefd *ffd, | |
239 | struct file *file, int fd) | |
240 | { | |
241 | ffd->file = file; | |
242 | ffd->fd = fd; | |
243 | } | |
244 | ||
245 | /* Compare rb-tree keys */ | |
246 | static inline int ep_cmp_ffd(struct epoll_filefd *p1, | |
247 | struct epoll_filefd *p2) | |
248 | { | |
249 | return (p1->file > p2->file ? +1: | |
250 | (p1->file < p2->file ? -1 : p1->fd - p2->fd)); | |
251 | } | |
252 | ||
253 | /* Special initialization for the rb-tree node to detect linkage */ | |
254 | static inline void ep_rb_initnode(struct rb_node *n) | |
255 | { | |
c569882b | 256 | rb_set_parent(n, n); |
b030a4dd PE |
257 | } |
258 | ||
259 | /* Removes a node from the rb-tree and marks it for a fast is-linked check */ | |
260 | static inline void ep_rb_erase(struct rb_node *n, struct rb_root *r) | |
261 | { | |
262 | rb_erase(n, r); | |
c569882b | 263 | rb_set_parent(n, n); |
b030a4dd PE |
264 | } |
265 | ||
266 | /* Fast check to verify that the item is linked to the main rb-tree */ | |
267 | static inline int ep_rb_linked(struct rb_node *n) | |
268 | { | |
c569882b | 269 | return rb_parent(n) != n; |
b030a4dd PE |
270 | } |
271 | ||
b030a4dd PE |
272 | /* Tells us if the item is currently linked */ |
273 | static inline int ep_is_linked(struct list_head *p) | |
274 | { | |
275 | return !list_empty(p); | |
276 | } | |
277 | ||
278 | /* Get the "struct epitem" from a wait queue pointer */ | |
279 | static inline struct epitem * ep_item_from_wait(wait_queue_t *p) | |
280 | { | |
281 | return container_of(p, struct eppoll_entry, wait)->base; | |
282 | } | |
283 | ||
284 | /* Get the "struct epitem" from an epoll queue wrapper */ | |
285 | static inline struct epitem * ep_item_from_epqueue(poll_table *p) | |
286 | { | |
287 | return container_of(p, struct ep_pqueue, pt)->epi; | |
288 | } | |
289 | ||
290 | /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */ | |
6192bd53 | 291 | static inline int ep_op_has_event(int op) |
b030a4dd PE |
292 | { |
293 | return op != EPOLL_CTL_DEL; | |
294 | } | |
295 | ||
1da177e4 LT |
296 | /* Initialize the poll safe wake up structure */ |
297 | static void ep_poll_safewake_init(struct poll_safewake *psw) | |
298 | { | |
299 | ||
300 | INIT_LIST_HEAD(&psw->wake_task_list); | |
301 | spin_lock_init(&psw->lock); | |
302 | } | |
303 | ||
1da177e4 LT |
304 | /* |
305 | * Perform a safe wake up of the poll wait list. The problem is that | |
306 | * with the new callback'd wake up system, it is possible that the | |
307 | * poll callback is reentered from inside the call to wake_up() done | |
308 | * on the poll wait queue head. The rule is that we cannot reenter the | |
309 | * wake up code from the same task more than EP_MAX_POLLWAKE_NESTS times, | |
310 | * and we cannot reenter the same wait queue head at all. This will | |
311 | * enable to have a hierarchy of epoll file descriptor of no more than | |
312 | * EP_MAX_POLLWAKE_NESTS deep. We need the irq version of the spin lock | |
313 | * because this one gets called by the poll callback, that in turn is called | |
314 | * from inside a wake_up(), that might be called from irq context. | |
315 | */ | |
316 | static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq) | |
317 | { | |
318 | int wake_nests = 0; | |
319 | unsigned long flags; | |
36c8b586 | 320 | struct task_struct *this_task = current; |
1da177e4 LT |
321 | struct list_head *lsthead = &psw->wake_task_list, *lnk; |
322 | struct wake_task_node *tncur; | |
323 | struct wake_task_node tnode; | |
324 | ||
325 | spin_lock_irqsave(&psw->lock, flags); | |
326 | ||
327 | /* Try to see if the current task is already inside this wakeup call */ | |
328 | list_for_each(lnk, lsthead) { | |
329 | tncur = list_entry(lnk, struct wake_task_node, llink); | |
330 | ||
331 | if (tncur->wq == wq || | |
332 | (tncur->task == this_task && ++wake_nests > EP_MAX_POLLWAKE_NESTS)) { | |
333 | /* | |
334 | * Ops ... loop detected or maximum nest level reached. | |
335 | * We abort this wake by breaking the cycle itself. | |
336 | */ | |
337 | spin_unlock_irqrestore(&psw->lock, flags); | |
338 | return; | |
339 | } | |
340 | } | |
341 | ||
342 | /* Add the current task to the list */ | |
343 | tnode.task = this_task; | |
344 | tnode.wq = wq; | |
345 | list_add(&tnode.llink, lsthead); | |
346 | ||
347 | spin_unlock_irqrestore(&psw->lock, flags); | |
348 | ||
349 | /* Do really wake up now */ | |
350 | wake_up(wq); | |
351 | ||
352 | /* Remove the current task from the list */ | |
353 | spin_lock_irqsave(&psw->lock, flags); | |
354 | list_del(&tnode.llink); | |
355 | spin_unlock_irqrestore(&psw->lock, flags); | |
356 | } | |
357 | ||
1da177e4 | 358 | /* |
7699acd1 DL |
359 | * This function unregister poll callbacks from the associated file descriptor. |
360 | * Since this must be called without holding "ep->lock" the atomic exchange trick | |
361 | * will protect us from multiple unregister. | |
1da177e4 | 362 | */ |
7699acd1 | 363 | static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi) |
1da177e4 | 364 | { |
7699acd1 DL |
365 | int nwait; |
366 | struct list_head *lsthead = &epi->pwqlist; | |
367 | struct eppoll_entry *pwq; | |
1da177e4 | 368 | |
7699acd1 DL |
369 | /* This is called without locks, so we need the atomic exchange */ |
370 | nwait = xchg(&epi->nwait, 0); | |
1da177e4 | 371 | |
7699acd1 DL |
372 | if (nwait) { |
373 | while (!list_empty(lsthead)) { | |
374 | pwq = list_first_entry(lsthead, struct eppoll_entry, llink); | |
1da177e4 | 375 | |
7699acd1 DL |
376 | list_del_init(&pwq->llink); |
377 | remove_wait_queue(pwq->whead, &pwq->wait); | |
378 | kmem_cache_free(pwq_cache, pwq); | |
379 | } | |
1da177e4 | 380 | } |
1da177e4 LT |
381 | } |
382 | ||
1da177e4 | 383 | /* |
7699acd1 DL |
384 | * Unlink the "struct epitem" from all places it might have been hooked up. |
385 | * This function must be called with write IRQ lock on "ep->lock". | |
1da177e4 | 386 | */ |
7699acd1 | 387 | static int ep_unlink(struct eventpoll *ep, struct epitem *epi) |
1da177e4 | 388 | { |
7699acd1 | 389 | int error; |
1da177e4 | 390 | |
7699acd1 DL |
391 | /* |
392 | * It can happen that this one is called for an item already unlinked. | |
393 | * The check protect us from doing a double unlink ( crash ). | |
394 | */ | |
395 | error = -ENOENT; | |
396 | if (!ep_rb_linked(&epi->rbn)) | |
397 | goto error_return; | |
1da177e4 | 398 | |
53d2be79 | 399 | /* |
7699acd1 DL |
400 | * Clear the event mask for the unlinked item. This will avoid item |
401 | * notifications to be sent after the unlink operation from inside | |
402 | * the kernel->userspace event transfer loop. | |
53d2be79 | 403 | */ |
7699acd1 | 404 | epi->event.events = 0; |
1da177e4 LT |
405 | |
406 | /* | |
7699acd1 DL |
407 | * At this point is safe to do the job, unlink the item from our rb-tree. |
408 | * This operation togheter with the above check closes the door to | |
409 | * double unlinks. | |
1da177e4 | 410 | */ |
7699acd1 | 411 | ep_rb_erase(&epi->rbn, &ep->rbr); |
1da177e4 | 412 | |
7699acd1 DL |
413 | /* |
414 | * If the item we are going to remove is inside the ready file descriptors | |
415 | * we want to remove it from this list to avoid stale events. | |
416 | */ | |
417 | if (ep_is_linked(&epi->rdllink)) | |
418 | list_del_init(&epi->rdllink); | |
1da177e4 | 419 | |
7699acd1 DL |
420 | error = 0; |
421 | error_return: | |
422 | ||
423 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_unlink(%p, %p) = %d\n", | |
424 | current, ep, epi->ffd.file, error)); | |
1da177e4 | 425 | |
1da177e4 LT |
426 | return error; |
427 | } | |
428 | ||
1da177e4 | 429 | /* |
7699acd1 DL |
430 | * Increment the usage count of the "struct epitem" making it sure |
431 | * that the user will have a valid pointer to reference. | |
1da177e4 | 432 | */ |
7699acd1 | 433 | static void ep_use_epitem(struct epitem *epi) |
1da177e4 | 434 | { |
7699acd1 DL |
435 | atomic_inc(&epi->usecnt); |
436 | } | |
1da177e4 | 437 | |
7699acd1 DL |
438 | /* |
439 | * Decrement ( release ) the usage count by signaling that the user | |
440 | * has finished using the structure. It might lead to freeing the | |
441 | * structure itself if the count goes to zero. | |
442 | */ | |
443 | static void ep_release_epitem(struct epitem *epi) | |
444 | { | |
445 | if (atomic_dec_and_test(&epi->usecnt)) | |
446 | kmem_cache_free(epi_cache, epi); | |
447 | } | |
1da177e4 | 448 | |
7699acd1 DL |
449 | /* |
450 | * Removes a "struct epitem" from the eventpoll RB tree and deallocates | |
451 | * all the associated resources. | |
452 | */ | |
453 | static int ep_remove(struct eventpoll *ep, struct epitem *epi) | |
454 | { | |
455 | int error; | |
456 | unsigned long flags; | |
457 | struct file *file = epi->ffd.file; | |
1da177e4 LT |
458 | |
459 | /* | |
7699acd1 DL |
460 | * Removes poll wait queue hooks. We _have_ to do this without holding |
461 | * the "ep->lock" otherwise a deadlock might occur. This because of the | |
462 | * sequence of the lock acquisition. Here we do "ep->lock" then the wait | |
463 | * queue head lock when unregistering the wait queue. The wakeup callback | |
464 | * will run by holding the wait queue head lock and will call our callback | |
465 | * that will try to get "ep->lock". | |
1da177e4 | 466 | */ |
7699acd1 | 467 | ep_unregister_pollwait(ep, epi); |
1da177e4 | 468 | |
7699acd1 DL |
469 | /* Remove the current item from the list of epoll hooks */ |
470 | spin_lock(&file->f_ep_lock); | |
471 | if (ep_is_linked(&epi->fllink)) | |
472 | list_del_init(&epi->fllink); | |
473 | spin_unlock(&file->f_ep_lock); | |
1da177e4 | 474 | |
7699acd1 DL |
475 | /* We need to acquire the write IRQ lock before calling ep_unlink() */ |
476 | write_lock_irqsave(&ep->lock, flags); | |
1da177e4 | 477 | |
7699acd1 DL |
478 | /* Really unlink the item from the RB tree */ |
479 | error = ep_unlink(ep, epi); | |
1da177e4 | 480 | |
7699acd1 | 481 | write_unlock_irqrestore(&ep->lock, flags); |
1da177e4 | 482 | |
7699acd1 DL |
483 | if (error) |
484 | goto error_return; | |
1da177e4 | 485 | |
7699acd1 DL |
486 | /* At this point it is safe to free the eventpoll item */ |
487 | ep_release_epitem(epi); | |
1da177e4 | 488 | |
7699acd1 DL |
489 | error = 0; |
490 | error_return: | |
491 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_remove(%p, %p) = %d\n", | |
492 | current, ep, file, error)); | |
1da177e4 LT |
493 | |
494 | return error; | |
495 | } | |
496 | ||
7699acd1 | 497 | static void ep_free(struct eventpoll *ep) |
1da177e4 | 498 | { |
7699acd1 DL |
499 | struct rb_node *rbp; |
500 | struct epitem *epi; | |
1da177e4 | 501 | |
7699acd1 DL |
502 | /* We need to release all tasks waiting for these file */ |
503 | if (waitqueue_active(&ep->poll_wait)) | |
504 | ep_poll_safewake(&psw, &ep->poll_wait); | |
1da177e4 | 505 | |
7699acd1 DL |
506 | /* |
507 | * We need to lock this because we could be hit by | |
508 | * eventpoll_release_file() while we're freeing the "struct eventpoll". | |
509 | * We do not need to hold "ep->sem" here because the epoll file | |
510 | * is on the way to be removed and no one has references to it | |
511 | * anymore. The only hit might come from eventpoll_release_file() but | |
512 | * holding "epmutex" is sufficent here. | |
513 | */ | |
514 | mutex_lock(&epmutex); | |
1da177e4 LT |
515 | |
516 | /* | |
7699acd1 | 517 | * Walks through the whole tree by unregistering poll callbacks. |
1da177e4 | 518 | */ |
7699acd1 DL |
519 | for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) { |
520 | epi = rb_entry(rbp, struct epitem, rbn); | |
521 | ||
522 | ep_unregister_pollwait(ep, epi); | |
523 | } | |
1da177e4 LT |
524 | |
525 | /* | |
7699acd1 DL |
526 | * Walks through the whole tree by freeing each "struct epitem". At this |
527 | * point we are sure no poll callbacks will be lingering around, and also by | |
528 | * write-holding "sem" we can be sure that no file cleanup code will hit | |
529 | * us during this operation. So we can avoid the lock on "ep->lock". | |
1da177e4 | 530 | */ |
7699acd1 DL |
531 | while ((rbp = rb_first(&ep->rbr)) != 0) { |
532 | epi = rb_entry(rbp, struct epitem, rbn); | |
533 | ep_remove(ep, epi); | |
534 | } | |
1da177e4 | 535 | |
7699acd1 DL |
536 | mutex_unlock(&epmutex); |
537 | } | |
1da177e4 | 538 | |
7699acd1 DL |
539 | static int ep_eventpoll_release(struct inode *inode, struct file *file) |
540 | { | |
541 | struct eventpoll *ep = file->private_data; | |
1da177e4 | 542 | |
7699acd1 DL |
543 | if (ep) { |
544 | ep_free(ep); | |
545 | kfree(ep); | |
546 | } | |
547 | ||
548 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: close() ep=%p\n", current, ep)); | |
549 | return 0; | |
1da177e4 LT |
550 | } |
551 | ||
7699acd1 DL |
552 | static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait) |
553 | { | |
554 | unsigned int pollflags = 0; | |
555 | unsigned long flags; | |
556 | struct eventpoll *ep = file->private_data; | |
1da177e4 | 557 | |
7699acd1 DL |
558 | /* Insert inside our poll wait queue */ |
559 | poll_wait(file, &ep->poll_wait, wait); | |
560 | ||
561 | /* Check our condition */ | |
562 | read_lock_irqsave(&ep->lock, flags); | |
563 | if (!list_empty(&ep->rdllist)) | |
564 | pollflags = POLLIN | POLLRDNORM; | |
565 | read_unlock_irqrestore(&ep->lock, flags); | |
566 | ||
567 | return pollflags; | |
568 | } | |
569 | ||
570 | /* File callbacks that implement the eventpoll file behaviour */ | |
571 | static const struct file_operations eventpoll_fops = { | |
572 | .release = ep_eventpoll_release, | |
573 | .poll = ep_eventpoll_poll | |
574 | }; | |
575 | ||
576 | /* Fast test to see if the file is an evenpoll file */ | |
577 | static inline int is_file_epoll(struct file *f) | |
578 | { | |
579 | return f->f_op == &eventpoll_fops; | |
580 | } | |
b611967d DL |
581 | |
582 | /* | |
7699acd1 DL |
583 | * This is called from eventpoll_release() to unlink files from the eventpoll |
584 | * interface. We need to have this facility to cleanup correctly files that are | |
585 | * closed without being removed from the eventpoll interface. | |
b611967d | 586 | */ |
7699acd1 | 587 | void eventpoll_release_file(struct file *file) |
b611967d | 588 | { |
7699acd1 DL |
589 | struct list_head *lsthead = &file->f_ep_links; |
590 | struct eventpoll *ep; | |
591 | struct epitem *epi; | |
b611967d DL |
592 | |
593 | /* | |
7699acd1 DL |
594 | * We don't want to get "file->f_ep_lock" because it is not |
595 | * necessary. It is not necessary because we're in the "struct file" | |
596 | * cleanup path, and this means that noone is using this file anymore. | |
597 | * The only hit might come from ep_free() but by holding the semaphore | |
598 | * will correctly serialize the operation. We do need to acquire | |
599 | * "ep->sem" after "epmutex" because ep_remove() requires it when called | |
600 | * from anywhere but ep_free(). | |
b611967d | 601 | */ |
7699acd1 | 602 | mutex_lock(&epmutex); |
b611967d | 603 | |
7699acd1 DL |
604 | while (!list_empty(lsthead)) { |
605 | epi = list_first_entry(lsthead, struct epitem, fllink); | |
b611967d | 606 | |
7699acd1 DL |
607 | ep = epi->ep; |
608 | list_del_init(&epi->fllink); | |
609 | down_write(&ep->sem); | |
610 | ep_remove(ep, epi); | |
611 | up_write(&ep->sem); | |
b611967d DL |
612 | } |
613 | ||
7699acd1 | 614 | mutex_unlock(&epmutex); |
b611967d DL |
615 | } |
616 | ||
53d2be79 | 617 | static int ep_alloc(struct eventpoll **pep) |
1da177e4 | 618 | { |
53d2be79 | 619 | struct eventpoll *ep = kzalloc(sizeof(*ep), GFP_KERNEL); |
1da177e4 | 620 | |
53d2be79 | 621 | if (!ep) |
1da177e4 LT |
622 | return -ENOMEM; |
623 | ||
1da177e4 LT |
624 | rwlock_init(&ep->lock); |
625 | init_rwsem(&ep->sem); | |
626 | init_waitqueue_head(&ep->wq); | |
627 | init_waitqueue_head(&ep->poll_wait); | |
628 | INIT_LIST_HEAD(&ep->rdllist); | |
629 | ep->rbr = RB_ROOT; | |
630 | ||
53d2be79 | 631 | *pep = ep; |
1da177e4 | 632 | |
53d2be79 | 633 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_alloc() ep=%p\n", |
1da177e4 LT |
634 | current, ep)); |
635 | return 0; | |
636 | } | |
637 | ||
1da177e4 | 638 | /* |
6192bd53 | 639 | * Search the file inside the eventpoll tree. It add usage count to |
1da177e4 LT |
640 | * the returned item, so the caller must call ep_release_epitem() |
641 | * after finished using the "struct epitem". | |
642 | */ | |
643 | static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd) | |
644 | { | |
645 | int kcmp; | |
646 | unsigned long flags; | |
647 | struct rb_node *rbp; | |
648 | struct epitem *epi, *epir = NULL; | |
649 | struct epoll_filefd ffd; | |
650 | ||
b030a4dd | 651 | ep_set_ffd(&ffd, file, fd); |
1da177e4 LT |
652 | read_lock_irqsave(&ep->lock, flags); |
653 | for (rbp = ep->rbr.rb_node; rbp; ) { | |
654 | epi = rb_entry(rbp, struct epitem, rbn); | |
b030a4dd | 655 | kcmp = ep_cmp_ffd(&ffd, &epi->ffd); |
1da177e4 LT |
656 | if (kcmp > 0) |
657 | rbp = rbp->rb_right; | |
658 | else if (kcmp < 0) | |
659 | rbp = rbp->rb_left; | |
660 | else { | |
661 | ep_use_epitem(epi); | |
662 | epir = epi; | |
663 | break; | |
664 | } | |
665 | } | |
666 | read_unlock_irqrestore(&ep->lock, flags); | |
667 | ||
668 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_find(%p) -> %p\n", | |
669 | current, file, epir)); | |
670 | ||
671 | return epir; | |
672 | } | |
673 | ||
1da177e4 | 674 | /* |
7699acd1 DL |
675 | * This is the callback that is passed to the wait queue wakeup |
676 | * machanism. It is called by the stored file descriptors when they | |
677 | * have events to report. | |
1da177e4 | 678 | */ |
7699acd1 | 679 | static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key) |
1da177e4 | 680 | { |
7699acd1 DL |
681 | int pwake = 0; |
682 | unsigned long flags; | |
683 | struct epitem *epi = ep_item_from_wait(wait); | |
684 | struct eventpoll *ep = epi->ep; | |
1da177e4 | 685 | |
7699acd1 DL |
686 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: poll_callback(%p) epi=%p ep=%p\n", |
687 | current, epi->ffd.file, epi, ep)); | |
1da177e4 | 688 | |
7699acd1 | 689 | write_lock_irqsave(&ep->lock, flags); |
1da177e4 | 690 | |
7699acd1 DL |
691 | /* |
692 | * If the event mask does not contain any poll(2) event, we consider the | |
693 | * descriptor to be disabled. This condition is likely the effect of the | |
694 | * EPOLLONESHOT bit that disables the descriptor when an event is received, | |
695 | * until the next EPOLL_CTL_MOD will be issued. | |
696 | */ | |
697 | if (!(epi->event.events & ~EP_PRIVATE_BITS)) | |
698 | goto is_disabled; | |
1da177e4 | 699 | |
7699acd1 DL |
700 | /* If this file is already in the ready list we exit soon */ |
701 | if (ep_is_linked(&epi->rdllink)) | |
702 | goto is_linked; | |
703 | ||
704 | list_add_tail(&epi->rdllink, &ep->rdllist); | |
705 | ||
706 | is_linked: | |
707 | /* | |
708 | * Wake up ( if active ) both the eventpoll wait list and the ->poll() | |
709 | * wait list. | |
710 | */ | |
711 | if (waitqueue_active(&ep->wq)) | |
712 | __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE | | |
713 | TASK_INTERRUPTIBLE); | |
714 | if (waitqueue_active(&ep->poll_wait)) | |
715 | pwake++; | |
716 | ||
717 | is_disabled: | |
718 | write_unlock_irqrestore(&ep->lock, flags); | |
1da177e4 | 719 | |
7699acd1 DL |
720 | /* We have to call this outside the lock */ |
721 | if (pwake) | |
722 | ep_poll_safewake(&psw, &ep->poll_wait); | |
723 | ||
724 | return 1; | |
725 | } | |
1da177e4 LT |
726 | |
727 | /* | |
728 | * This is the callback that is used to add our wait queue to the | |
729 | * target file wakeup lists. | |
730 | */ | |
731 | static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead, | |
732 | poll_table *pt) | |
733 | { | |
b030a4dd | 734 | struct epitem *epi = ep_item_from_epqueue(pt); |
1da177e4 LT |
735 | struct eppoll_entry *pwq; |
736 | ||
e94b1766 | 737 | if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) { |
1da177e4 LT |
738 | init_waitqueue_func_entry(&pwq->wait, ep_poll_callback); |
739 | pwq->whead = whead; | |
740 | pwq->base = epi; | |
741 | add_wait_queue(whead, &pwq->wait); | |
742 | list_add_tail(&pwq->llink, &epi->pwqlist); | |
743 | epi->nwait++; | |
744 | } else { | |
745 | /* We have to signal that an error occurred */ | |
746 | epi->nwait = -1; | |
747 | } | |
748 | } | |
749 | ||
1da177e4 LT |
750 | static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi) |
751 | { | |
752 | int kcmp; | |
753 | struct rb_node **p = &ep->rbr.rb_node, *parent = NULL; | |
754 | struct epitem *epic; | |
755 | ||
756 | while (*p) { | |
757 | parent = *p; | |
758 | epic = rb_entry(parent, struct epitem, rbn); | |
b030a4dd | 759 | kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd); |
1da177e4 LT |
760 | if (kcmp > 0) |
761 | p = &parent->rb_right; | |
762 | else | |
763 | p = &parent->rb_left; | |
764 | } | |
765 | rb_link_node(&epi->rbn, parent, p); | |
766 | rb_insert_color(&epi->rbn, &ep->rbr); | |
767 | } | |
768 | ||
1da177e4 LT |
769 | static int ep_insert(struct eventpoll *ep, struct epoll_event *event, |
770 | struct file *tfile, int fd) | |
771 | { | |
772 | int error, revents, pwake = 0; | |
773 | unsigned long flags; | |
774 | struct epitem *epi; | |
775 | struct ep_pqueue epq; | |
776 | ||
777 | error = -ENOMEM; | |
e94b1766 | 778 | if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL))) |
7699acd1 | 779 | goto error_return; |
1da177e4 LT |
780 | |
781 | /* Item initialization follow here ... */ | |
b030a4dd | 782 | ep_rb_initnode(&epi->rbn); |
1da177e4 LT |
783 | INIT_LIST_HEAD(&epi->rdllink); |
784 | INIT_LIST_HEAD(&epi->fllink); | |
1da177e4 LT |
785 | INIT_LIST_HEAD(&epi->pwqlist); |
786 | epi->ep = ep; | |
b030a4dd | 787 | ep_set_ffd(&epi->ffd, tfile, fd); |
1da177e4 LT |
788 | epi->event = *event; |
789 | atomic_set(&epi->usecnt, 1); | |
790 | epi->nwait = 0; | |
791 | ||
792 | /* Initialize the poll table using the queue callback */ | |
793 | epq.epi = epi; | |
794 | init_poll_funcptr(&epq.pt, ep_ptable_queue_proc); | |
795 | ||
796 | /* | |
797 | * Attach the item to the poll hooks and get current event bits. | |
798 | * We can safely use the file* here because its usage count has | |
799 | * been increased by the caller of this function. | |
800 | */ | |
801 | revents = tfile->f_op->poll(tfile, &epq.pt); | |
802 | ||
803 | /* | |
804 | * We have to check if something went wrong during the poll wait queue | |
805 | * install process. Namely an allocation for a wait queue failed due | |
806 | * high memory pressure. | |
807 | */ | |
808 | if (epi->nwait < 0) | |
7699acd1 | 809 | goto error_unregister; |
1da177e4 LT |
810 | |
811 | /* Add the current item to the list of active epoll hook for this file */ | |
812 | spin_lock(&tfile->f_ep_lock); | |
813 | list_add_tail(&epi->fllink, &tfile->f_ep_links); | |
814 | spin_unlock(&tfile->f_ep_lock); | |
815 | ||
816 | /* We have to drop the new item inside our item list to keep track of it */ | |
817 | write_lock_irqsave(&ep->lock, flags); | |
818 | ||
819 | /* Add the current item to the rb-tree */ | |
820 | ep_rbtree_insert(ep, epi); | |
821 | ||
822 | /* If the file is already "ready" we drop it inside the ready list */ | |
b030a4dd | 823 | if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) { |
1da177e4 LT |
824 | list_add_tail(&epi->rdllink, &ep->rdllist); |
825 | ||
826 | /* Notify waiting tasks that events are available */ | |
827 | if (waitqueue_active(&ep->wq)) | |
3419b23a | 828 | __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE); |
1da177e4 LT |
829 | if (waitqueue_active(&ep->poll_wait)) |
830 | pwake++; | |
831 | } | |
832 | ||
833 | write_unlock_irqrestore(&ep->lock, flags); | |
834 | ||
835 | /* We have to call this outside the lock */ | |
836 | if (pwake) | |
837 | ep_poll_safewake(&psw, &ep->poll_wait); | |
838 | ||
839 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_insert(%p, %p, %d)\n", | |
840 | current, ep, tfile, fd)); | |
841 | ||
842 | return 0; | |
843 | ||
7699acd1 | 844 | error_unregister: |
1da177e4 LT |
845 | ep_unregister_pollwait(ep, epi); |
846 | ||
847 | /* | |
848 | * We need to do this because an event could have been arrived on some | |
849 | * allocated wait queue. | |
850 | */ | |
851 | write_lock_irqsave(&ep->lock, flags); | |
b030a4dd | 852 | if (ep_is_linked(&epi->rdllink)) |
6192bd53 | 853 | list_del_init(&epi->rdllink); |
1da177e4 LT |
854 | write_unlock_irqrestore(&ep->lock, flags); |
855 | ||
b030a4dd | 856 | kmem_cache_free(epi_cache, epi); |
7699acd1 | 857 | error_return: |
1da177e4 LT |
858 | return error; |
859 | } | |
860 | ||
1da177e4 LT |
861 | /* |
862 | * Modify the interest event mask by dropping an event if the new mask | |
863 | * has a match in the current file status. | |
864 | */ | |
865 | static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event) | |
866 | { | |
867 | int pwake = 0; | |
868 | unsigned int revents; | |
869 | unsigned long flags; | |
870 | ||
871 | /* | |
872 | * Set the new event interest mask before calling f_op->poll(), otherwise | |
873 | * a potential race might occur. In fact if we do this operation inside | |
874 | * the lock, an event might happen between the f_op->poll() call and the | |
875 | * new event set registering. | |
876 | */ | |
877 | epi->event.events = event->events; | |
878 | ||
879 | /* | |
880 | * Get current event bits. We can safely use the file* here because | |
881 | * its usage count has been increased by the caller of this function. | |
882 | */ | |
883 | revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL); | |
884 | ||
885 | write_lock_irqsave(&ep->lock, flags); | |
886 | ||
887 | /* Copy the data member from inside the lock */ | |
888 | epi->event.data = event->data; | |
889 | ||
890 | /* | |
6192bd53 | 891 | * If the item is not linked to the RB tree it means that it's on its |
1da177e4 LT |
892 | * way toward the removal. Do nothing in this case. |
893 | */ | |
b030a4dd | 894 | if (ep_rb_linked(&epi->rbn)) { |
1da177e4 LT |
895 | /* |
896 | * If the item is "hot" and it is not registered inside the ready | |
897 | * list, push it inside. If the item is not "hot" and it is currently | |
898 | * registered inside the ready list, unlink it. | |
899 | */ | |
900 | if (revents & event->events) { | |
b030a4dd | 901 | if (!ep_is_linked(&epi->rdllink)) { |
1da177e4 LT |
902 | list_add_tail(&epi->rdllink, &ep->rdllist); |
903 | ||
904 | /* Notify waiting tasks that events are available */ | |
905 | if (waitqueue_active(&ep->wq)) | |
3419b23a DL |
906 | __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE | |
907 | TASK_INTERRUPTIBLE); | |
7699acd1 DL |
908 | if (waitqueue_active(&ep->poll_wait)) |
909 | pwake++; | |
910 | } | |
911 | } | |
912 | } | |
1da177e4 | 913 | |
7699acd1 | 914 | write_unlock_irqrestore(&ep->lock, flags); |
1da177e4 | 915 | |
7699acd1 DL |
916 | /* We have to call this outside the lock */ |
917 | if (pwake) | |
918 | ep_poll_safewake(&psw, &ep->poll_wait); | |
1da177e4 | 919 | |
7699acd1 | 920 | return 0; |
1da177e4 LT |
921 | } |
922 | ||
1da177e4 LT |
923 | /* |
924 | * This function is called without holding the "ep->lock" since the call to | |
925 | * __copy_to_user() might sleep, and also f_op->poll() might reenable the IRQ | |
926 | * because of the way poll() is traditionally implemented in Linux. | |
927 | */ | |
928 | static int ep_send_events(struct eventpoll *ep, struct list_head *txlist, | |
6192bd53 | 929 | struct epoll_event __user *events, int maxevents) |
1da177e4 | 930 | { |
6192bd53 | 931 | int eventcnt, error = -EFAULT, pwake = 0; |
1da177e4 | 932 | unsigned int revents; |
6192bd53 | 933 | unsigned long flags; |
1da177e4 | 934 | struct epitem *epi; |
6192bd53 DL |
935 | struct list_head injlist; |
936 | ||
937 | INIT_LIST_HEAD(&injlist); | |
1da177e4 LT |
938 | |
939 | /* | |
940 | * We can loop without lock because this is a task private list. | |
6192bd53 DL |
941 | * We just splice'd out the ep->rdllist in ep_collect_ready_items(). |
942 | * Items cannot vanish during the loop because we are holding "sem" in | |
943 | * read. | |
1da177e4 | 944 | */ |
6192bd53 | 945 | for (eventcnt = 0; !list_empty(txlist) && eventcnt < maxevents;) { |
b5e61818 | 946 | epi = list_first_entry(txlist, struct epitem, rdllink); |
6192bd53 | 947 | prefetch(epi->rdllink.next); |
1da177e4 LT |
948 | |
949 | /* | |
950 | * Get the ready file event set. We can safely use the file | |
951 | * because we are holding the "sem" in read and this will | |
952 | * guarantee that both the file and the item will not vanish. | |
953 | */ | |
954 | revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL); | |
6192bd53 | 955 | revents &= epi->event.events; |
1da177e4 LT |
956 | |
957 | /* | |
6192bd53 DL |
958 | * Is the event mask intersect the caller-requested one, |
959 | * deliver the event to userspace. Again, we are holding | |
960 | * "sem" in read, so no operations coming from userspace | |
961 | * can change the item. | |
1da177e4 | 962 | */ |
6192bd53 DL |
963 | if (revents) { |
964 | if (__put_user(revents, | |
1da177e4 LT |
965 | &events[eventcnt].events) || |
966 | __put_user(epi->event.data, | |
967 | &events[eventcnt].data)) | |
6192bd53 | 968 | goto errxit; |
1da177e4 LT |
969 | if (epi->event.events & EPOLLONESHOT) |
970 | epi->event.events &= EP_PRIVATE_BITS; | |
971 | eventcnt++; | |
972 | } | |
1da177e4 LT |
973 | |
974 | /* | |
6192bd53 DL |
975 | * This is tricky. We are holding the "sem" in read, and this |
976 | * means that the operations that can change the "linked" status | |
977 | * of the epoll item (epi->rbn and epi->rdllink), cannot touch | |
978 | * them. Also, since we are "linked" from a epi->rdllink POV | |
979 | * (the item is linked to our transmission list we just | |
980 | * spliced), the ep_poll_callback() cannot touch us either, | |
981 | * because of the check present in there. Another parallel | |
982 | * epoll_wait() will not get the same result set, since we | |
983 | * spliced the ready list before. Note that list_del() still | |
984 | * shows the item as linked to the test in ep_poll_callback(). | |
1da177e4 | 985 | */ |
6192bd53 DL |
986 | list_del(&epi->rdllink); |
987 | if (!(epi->event.events & EPOLLET) && | |
988 | (revents & epi->event.events)) | |
989 | list_add_tail(&epi->rdllink, &injlist); | |
990 | else { | |
991 | /* | |
992 | * Be sure the item is totally detached before re-init | |
993 | * the list_head. After INIT_LIST_HEAD() is committed, | |
994 | * the ep_poll_callback() can requeue the item again, | |
995 | * but we don't care since we are already past it. | |
996 | */ | |
997 | smp_mb(); | |
998 | INIT_LIST_HEAD(&epi->rdllink); | |
1da177e4 LT |
999 | } |
1000 | } | |
6192bd53 | 1001 | error = 0; |
1da177e4 | 1002 | |
6192bd53 DL |
1003 | errxit: |
1004 | ||
1005 | /* | |
1006 | * If the re-injection list or the txlist are not empty, re-splice | |
1007 | * them to the ready list and do proper wakeups. | |
1008 | */ | |
1009 | if (!list_empty(&injlist) || !list_empty(txlist)) { | |
1010 | write_lock_irqsave(&ep->lock, flags); | |
1011 | ||
1012 | list_splice(txlist, &ep->rdllist); | |
1013 | list_splice(&injlist, &ep->rdllist); | |
1da177e4 LT |
1014 | /* |
1015 | * Wake up ( if active ) both the eventpoll wait list and the ->poll() | |
1016 | * wait list. | |
1017 | */ | |
1018 | if (waitqueue_active(&ep->wq)) | |
3419b23a DL |
1019 | __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE | |
1020 | TASK_INTERRUPTIBLE); | |
1da177e4 LT |
1021 | if (waitqueue_active(&ep->poll_wait)) |
1022 | pwake++; | |
1da177e4 | 1023 | |
6192bd53 DL |
1024 | write_unlock_irqrestore(&ep->lock, flags); |
1025 | } | |
1da177e4 LT |
1026 | |
1027 | /* We have to call this outside the lock */ | |
1028 | if (pwake) | |
1029 | ep_poll_safewake(&psw, &ep->poll_wait); | |
6192bd53 DL |
1030 | |
1031 | return eventcnt == 0 ? error: eventcnt; | |
1da177e4 LT |
1032 | } |
1033 | ||
1da177e4 LT |
1034 | /* |
1035 | * Perform the transfer of events to user space. | |
1036 | */ | |
1037 | static int ep_events_transfer(struct eventpoll *ep, | |
1038 | struct epoll_event __user *events, int maxevents) | |
1039 | { | |
6192bd53 DL |
1040 | int eventcnt; |
1041 | unsigned long flags; | |
1da177e4 LT |
1042 | struct list_head txlist; |
1043 | ||
1044 | INIT_LIST_HEAD(&txlist); | |
1045 | ||
1046 | /* | |
1047 | * We need to lock this because we could be hit by | |
1048 | * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL). | |
1049 | */ | |
1050 | down_read(&ep->sem); | |
1051 | ||
6192bd53 DL |
1052 | /* |
1053 | * Steal the ready list, and re-init the original one to the | |
1054 | * empty list. | |
1055 | */ | |
1056 | write_lock_irqsave(&ep->lock, flags); | |
1057 | list_splice(&ep->rdllist, &txlist); | |
1058 | INIT_LIST_HEAD(&ep->rdllist); | |
1059 | write_unlock_irqrestore(&ep->lock, flags); | |
1da177e4 | 1060 | |
6192bd53 DL |
1061 | /* Build result set in userspace */ |
1062 | eventcnt = ep_send_events(ep, &txlist, events, maxevents); | |
1da177e4 LT |
1063 | |
1064 | up_read(&ep->sem); | |
1065 | ||
1066 | return eventcnt; | |
1067 | } | |
1068 | ||
1da177e4 LT |
1069 | static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events, |
1070 | int maxevents, long timeout) | |
1071 | { | |
1072 | int res, eavail; | |
1073 | unsigned long flags; | |
1074 | long jtimeout; | |
1075 | wait_queue_t wait; | |
1076 | ||
1077 | /* | |
1078 | * Calculate the timeout by checking for the "infinite" value ( -1 ) | |
1079 | * and the overflow condition. The passed timeout is in milliseconds, | |
1080 | * that why (t * HZ) / 1000. | |
1081 | */ | |
e3306dd5 DL |
1082 | jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ? |
1083 | MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000; | |
1da177e4 LT |
1084 | |
1085 | retry: | |
1086 | write_lock_irqsave(&ep->lock, flags); | |
1087 | ||
1088 | res = 0; | |
1089 | if (list_empty(&ep->rdllist)) { | |
1090 | /* | |
1091 | * We don't have any available event to return to the caller. | |
1092 | * We need to sleep here, and we will be wake up by | |
1093 | * ep_poll_callback() when events will become available. | |
1094 | */ | |
1095 | init_waitqueue_entry(&wait, current); | |
3419b23a | 1096 | __add_wait_queue(&ep->wq, &wait); |
1da177e4 LT |
1097 | |
1098 | for (;;) { | |
1099 | /* | |
1100 | * We don't want to sleep if the ep_poll_callback() sends us | |
1101 | * a wakeup in between. That's why we set the task state | |
1102 | * to TASK_INTERRUPTIBLE before doing the checks. | |
1103 | */ | |
1104 | set_current_state(TASK_INTERRUPTIBLE); | |
1105 | if (!list_empty(&ep->rdllist) || !jtimeout) | |
1106 | break; | |
1107 | if (signal_pending(current)) { | |
1108 | res = -EINTR; | |
1109 | break; | |
1110 | } | |
1111 | ||
1112 | write_unlock_irqrestore(&ep->lock, flags); | |
1113 | jtimeout = schedule_timeout(jtimeout); | |
1114 | write_lock_irqsave(&ep->lock, flags); | |
1115 | } | |
3419b23a | 1116 | __remove_wait_queue(&ep->wq, &wait); |
1da177e4 LT |
1117 | |
1118 | set_current_state(TASK_RUNNING); | |
1119 | } | |
1120 | ||
1121 | /* Is it worth to try to dig for events ? */ | |
1122 | eavail = !list_empty(&ep->rdllist); | |
1123 | ||
1124 | write_unlock_irqrestore(&ep->lock, flags); | |
1125 | ||
1126 | /* | |
1127 | * Try to transfer events to user space. In case we get 0 events and | |
1128 | * there's still timeout left over, we go trying again in search of | |
1129 | * more luck. | |
1130 | */ | |
1131 | if (!res && eavail && | |
1132 | !(res = ep_events_transfer(ep, events, maxevents)) && jtimeout) | |
1133 | goto retry; | |
1134 | ||
1135 | return res; | |
1136 | } | |
1137 | ||
7699acd1 DL |
1138 | /* |
1139 | * It opens an eventpoll file descriptor by suggesting a storage of "size" | |
1140 | * file descriptors. The size parameter is just an hint about how to size | |
1141 | * data structures. It won't prevent the user to store more than "size" | |
1142 | * file descriptors inside the epoll interface. It is the kernel part of | |
1143 | * the userspace epoll_create(2). | |
1144 | */ | |
1145 | asmlinkage long sys_epoll_create(int size) | |
1146 | { | |
1147 | int error, fd = -1; | |
1148 | struct eventpoll *ep; | |
1149 | struct inode *inode; | |
1150 | struct file *file; | |
1151 | ||
1152 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d)\n", | |
1153 | current, size)); | |
1154 | ||
1155 | /* | |
1156 | * Sanity check on the size parameter, and create the internal data | |
1157 | * structure ( "struct eventpoll" ). | |
1158 | */ | |
1159 | error = -EINVAL; | |
1160 | if (size <= 0 || (error = ep_alloc(&ep)) != 0) | |
1161 | goto error_return; | |
1162 | ||
1163 | /* | |
1164 | * Creates all the items needed to setup an eventpoll file. That is, | |
1165 | * a file structure, and inode and a free file descriptor. | |
1166 | */ | |
1167 | error = anon_inode_getfd(&fd, &inode, &file, "[eventpoll]", | |
1168 | &eventpoll_fops, ep); | |
1169 | if (error) | |
1170 | goto error_free; | |
1171 | ||
1172 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n", | |
1173 | current, size, fd)); | |
1174 | ||
1175 | return fd; | |
1176 | ||
1177 | error_free: | |
1178 | ep_free(ep); | |
1179 | kfree(ep); | |
1180 | error_return: | |
1181 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n", | |
1182 | current, size, error)); | |
1183 | return error; | |
1184 | } | |
1185 | ||
1186 | /* | |
1187 | * The following function implements the controller interface for | |
1188 | * the eventpoll file that enables the insertion/removal/change of | |
1189 | * file descriptors inside the interest set. It represents | |
1190 | * the kernel part of the user space epoll_ctl(2). | |
1191 | */ | |
1192 | asmlinkage long sys_epoll_ctl(int epfd, int op, int fd, | |
1193 | struct epoll_event __user *event) | |
1194 | { | |
1195 | int error; | |
1196 | struct file *file, *tfile; | |
1197 | struct eventpoll *ep; | |
1198 | struct epitem *epi; | |
1199 | struct epoll_event epds; | |
1200 | ||
1201 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p)\n", | |
1202 | current, epfd, op, fd, event)); | |
1203 | ||
1204 | error = -EFAULT; | |
1205 | if (ep_op_has_event(op) && | |
1206 | copy_from_user(&epds, event, sizeof(struct epoll_event))) | |
1207 | goto error_return; | |
1208 | ||
1209 | /* Get the "struct file *" for the eventpoll file */ | |
1210 | error = -EBADF; | |
1211 | file = fget(epfd); | |
1212 | if (!file) | |
1213 | goto error_return; | |
1214 | ||
1215 | /* Get the "struct file *" for the target file */ | |
1216 | tfile = fget(fd); | |
1217 | if (!tfile) | |
1218 | goto error_fput; | |
1219 | ||
1220 | /* The target file descriptor must support poll */ | |
1221 | error = -EPERM; | |
1222 | if (!tfile->f_op || !tfile->f_op->poll) | |
1223 | goto error_tgt_fput; | |
1224 | ||
1225 | /* | |
1226 | * We have to check that the file structure underneath the file descriptor | |
1227 | * the user passed to us _is_ an eventpoll file. And also we do not permit | |
1228 | * adding an epoll file descriptor inside itself. | |
1229 | */ | |
1230 | error = -EINVAL; | |
1231 | if (file == tfile || !is_file_epoll(file)) | |
1232 | goto error_tgt_fput; | |
1233 | ||
1234 | /* | |
1235 | * At this point it is safe to assume that the "private_data" contains | |
1236 | * our own data structure. | |
1237 | */ | |
1238 | ep = file->private_data; | |
1239 | ||
1240 | down_write(&ep->sem); | |
1241 | ||
1242 | /* Try to lookup the file inside our RB tree */ | |
1243 | epi = ep_find(ep, tfile, fd); | |
1244 | ||
1245 | error = -EINVAL; | |
1246 | switch (op) { | |
1247 | case EPOLL_CTL_ADD: | |
1248 | if (!epi) { | |
1249 | epds.events |= POLLERR | POLLHUP; | |
1250 | ||
1251 | error = ep_insert(ep, &epds, tfile, fd); | |
1252 | } else | |
1253 | error = -EEXIST; | |
1254 | break; | |
1255 | case EPOLL_CTL_DEL: | |
1256 | if (epi) | |
1257 | error = ep_remove(ep, epi); | |
1258 | else | |
1259 | error = -ENOENT; | |
1260 | break; | |
1261 | case EPOLL_CTL_MOD: | |
1262 | if (epi) { | |
1263 | epds.events |= POLLERR | POLLHUP; | |
1264 | error = ep_modify(ep, epi, &epds); | |
1265 | } else | |
1266 | error = -ENOENT; | |
1267 | break; | |
1268 | } | |
1269 | /* | |
1270 | * The function ep_find() increments the usage count of the structure | |
1271 | * so, if this is not NULL, we need to release it. | |
1272 | */ | |
1273 | if (epi) | |
1274 | ep_release_epitem(epi); | |
1275 | up_write(&ep->sem); | |
1276 | ||
1277 | error_tgt_fput: | |
1278 | fput(tfile); | |
1279 | error_fput: | |
1280 | fput(file); | |
1281 | error_return: | |
1282 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p) = %d\n", | |
1283 | current, epfd, op, fd, event, error)); | |
1284 | ||
1285 | return error; | |
1286 | } | |
1287 | ||
1288 | /* | |
1289 | * Implement the event wait interface for the eventpoll file. It is the kernel | |
1290 | * part of the user space epoll_wait(2). | |
1291 | */ | |
1292 | asmlinkage long sys_epoll_wait(int epfd, struct epoll_event __user *events, | |
1293 | int maxevents, int timeout) | |
1294 | { | |
1295 | int error; | |
1296 | struct file *file; | |
1297 | struct eventpoll *ep; | |
1298 | ||
1299 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d)\n", | |
1300 | current, epfd, events, maxevents, timeout)); | |
1301 | ||
1302 | /* The maximum number of event must be greater than zero */ | |
1303 | if (maxevents <= 0 || maxevents > EP_MAX_EVENTS) | |
1304 | return -EINVAL; | |
1305 | ||
1306 | /* Verify that the area passed by the user is writeable */ | |
1307 | if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) { | |
1308 | error = -EFAULT; | |
1309 | goto error_return; | |
1310 | } | |
1311 | ||
1312 | /* Get the "struct file *" for the eventpoll file */ | |
1313 | error = -EBADF; | |
1314 | file = fget(epfd); | |
1315 | if (!file) | |
1316 | goto error_return; | |
1317 | ||
1318 | /* | |
1319 | * We have to check that the file structure underneath the fd | |
1320 | * the user passed to us _is_ an eventpoll file. | |
1321 | */ | |
1322 | error = -EINVAL; | |
1323 | if (!is_file_epoll(file)) | |
1324 | goto error_fput; | |
1325 | ||
1326 | /* | |
1327 | * At this point it is safe to assume that the "private_data" contains | |
1328 | * our own data structure. | |
1329 | */ | |
1330 | ep = file->private_data; | |
1331 | ||
1332 | /* Time to fish for events ... */ | |
1333 | error = ep_poll(ep, events, maxevents, timeout); | |
1334 | ||
1335 | error_fput: | |
1336 | fput(file); | |
1337 | error_return: | |
1338 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d) = %d\n", | |
1339 | current, epfd, events, maxevents, timeout, error)); | |
1340 | ||
1341 | return error; | |
1342 | } | |
1343 | ||
1344 | #ifdef TIF_RESTORE_SIGMASK | |
1345 | ||
1346 | /* | |
1347 | * Implement the event wait interface for the eventpoll file. It is the kernel | |
1348 | * part of the user space epoll_pwait(2). | |
1349 | */ | |
1350 | asmlinkage long sys_epoll_pwait(int epfd, struct epoll_event __user *events, | |
1351 | int maxevents, int timeout, const sigset_t __user *sigmask, | |
1352 | size_t sigsetsize) | |
1353 | { | |
1354 | int error; | |
1355 | sigset_t ksigmask, sigsaved; | |
1356 | ||
1357 | /* | |
1358 | * If the caller wants a certain signal mask to be set during the wait, | |
1359 | * we apply it here. | |
1360 | */ | |
1361 | if (sigmask) { | |
1362 | if (sigsetsize != sizeof(sigset_t)) | |
1363 | return -EINVAL; | |
1364 | if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask))) | |
1365 | return -EFAULT; | |
1366 | sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP)); | |
1367 | sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved); | |
1368 | } | |
1369 | ||
1370 | error = sys_epoll_wait(epfd, events, maxevents, timeout); | |
1371 | ||
1372 | /* | |
1373 | * If we changed the signal mask, we need to restore the original one. | |
1374 | * In case we've got a signal while waiting, we do not restore the | |
1375 | * signal mask yet, and we allow do_signal() to deliver the signal on | |
1376 | * the way back to userspace, before the signal mask is restored. | |
1377 | */ | |
1378 | if (sigmask) { | |
1379 | if (error == -EINTR) { | |
1380 | memcpy(¤t->saved_sigmask, &sigsaved, | |
1381 | sizeof(sigsaved)); | |
1382 | set_thread_flag(TIF_RESTORE_SIGMASK); | |
1383 | } else | |
1384 | sigprocmask(SIG_SETMASK, &sigsaved, NULL); | |
1385 | } | |
1386 | ||
1387 | return error; | |
1388 | } | |
1389 | ||
1390 | #endif /* #ifdef TIF_RESTORE_SIGMASK */ | |
1391 | ||
1da177e4 LT |
1392 | static int __init eventpoll_init(void) |
1393 | { | |
144efe3e | 1394 | mutex_init(&epmutex); |
1da177e4 LT |
1395 | |
1396 | /* Initialize the structure used to perform safe poll wait head wake ups */ | |
1397 | ep_poll_safewake_init(&psw); | |
1398 | ||
1399 | /* Allocates slab cache used to allocate "struct epitem" items */ | |
1400 | epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem), | |
1401 | 0, SLAB_HWCACHE_ALIGN|EPI_SLAB_DEBUG|SLAB_PANIC, | |
1402 | NULL, NULL); | |
1403 | ||
1404 | /* Allocates slab cache used to allocate "struct eppoll_entry" */ | |
1405 | pwq_cache = kmem_cache_create("eventpoll_pwq", | |
1406 | sizeof(struct eppoll_entry), 0, | |
1407 | EPI_SLAB_DEBUG|SLAB_PANIC, NULL, NULL); | |
1408 | ||
1da177e4 | 1409 | return 0; |
1da177e4 | 1410 | } |
cea69241 | 1411 | fs_initcall(eventpoll_init); |
1da177e4 | 1412 |