]>
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> | |
1da177e4 LT |
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> | |
144efe3e | 34 | #include <linux/mutex.h> |
da66f7cb | 35 | #include <linux/anon_inodes.h> |
1da177e4 LT |
36 | #include <asm/uaccess.h> |
37 | #include <asm/system.h> | |
38 | #include <asm/io.h> | |
39 | #include <asm/mman.h> | |
40 | #include <asm/atomic.h> | |
1da177e4 | 41 | |
1da177e4 LT |
42 | /* |
43 | * LOCKING: | |
44 | * There are three level of locking required by epoll : | |
45 | * | |
144efe3e | 46 | * 1) epmutex (mutex) |
d47de16c | 47 | * 2) ep->mtx (mutes) |
1da177e4 LT |
48 | * 3) ep->lock (rw_lock) |
49 | * | |
50 | * The acquire order is the one listed above, from 1 to 3. | |
51 | * We need a spinlock (ep->lock) because we manipulate objects | |
52 | * from inside the poll callback, that might be triggered from | |
53 | * a wake_up() that in turn might be called from IRQ context. | |
54 | * So we can't sleep inside the poll callback and hence we need | |
55 | * a spinlock. During the event transfer loop (from kernel to | |
56 | * user space) we could end up sleeping due a copy_to_user(), so | |
57 | * we need a lock that will allow us to sleep. This lock is a | |
d47de16c DL |
58 | * mutex (ep->mtx). It is acquired during the event transfer loop, |
59 | * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file(). | |
60 | * Then we also need a global mutex to serialize eventpoll_release_file() | |
61 | * and ep_free(). | |
62 | * This mutex is acquired by ep_free() during the epoll file | |
1da177e4 LT |
63 | * cleanup path and it is also acquired by eventpoll_release_file() |
64 | * if a file has been pushed inside an epoll set and it is then | |
65 | * close()d without a previous call toepoll_ctl(EPOLL_CTL_DEL). | |
d47de16c DL |
66 | * It is possible to drop the "ep->mtx" and to use the global |
67 | * mutex "epmutex" (together with "ep->lock") to have it working, | |
68 | * but having "ep->mtx" will make the interface more scalable. | |
144efe3e | 69 | * Events that require holding "epmutex" are very rare, while for |
d47de16c DL |
70 | * normal operations the epoll private "ep->mtx" will guarantee |
71 | * a better scalability. | |
1da177e4 LT |
72 | */ |
73 | ||
1da177e4 LT |
74 | #define DEBUG_EPOLL 0 |
75 | ||
76 | #if DEBUG_EPOLL > 0 | |
77 | #define DPRINTK(x) printk x | |
78 | #define DNPRINTK(n, x) do { if ((n) <= DEBUG_EPOLL) printk x; } while (0) | |
79 | #else /* #if DEBUG_EPOLL > 0 */ | |
80 | #define DPRINTK(x) (void) 0 | |
81 | #define DNPRINTK(n, x) (void) 0 | |
82 | #endif /* #if DEBUG_EPOLL > 0 */ | |
83 | ||
84 | #define DEBUG_EPI 0 | |
85 | ||
86 | #if DEBUG_EPI != 0 | |
87 | #define EPI_SLAB_DEBUG (SLAB_DEBUG_FREE | SLAB_RED_ZONE /* | SLAB_POISON */) | |
88 | #else /* #if DEBUG_EPI != 0 */ | |
89 | #define EPI_SLAB_DEBUG 0 | |
90 | #endif /* #if DEBUG_EPI != 0 */ | |
91 | ||
92 | /* Epoll private bits inside the event mask */ | |
93 | #define EP_PRIVATE_BITS (EPOLLONESHOT | EPOLLET) | |
94 | ||
95 | /* Maximum number of poll wake up nests we are allowing */ | |
96 | #define EP_MAX_POLLWAKE_NESTS 4 | |
97 | ||
e3306dd5 DL |
98 | /* Maximum msec timeout value storeable in a long int */ |
99 | #define EP_MAX_MSTIMEO min(1000ULL * MAX_SCHEDULE_TIMEOUT / HZ, (LONG_MAX - 999ULL) / HZ) | |
100 | ||
b611967d DL |
101 | #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event)) |
102 | ||
d47de16c DL |
103 | #define EP_UNACTIVE_PTR ((void *) -1L) |
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 | |
d47de16c | 114 | * (loop) 2) allow a maximum number of epoll descriptors inclusion nesting |
1da177e4 LT |
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 | ||
d47de16c DL |
132 | /* |
133 | * Each file descriptor added to the eventpoll interface will | |
134 | * have an entry of this type linked to the "rbr" RB tree. | |
135 | */ | |
136 | struct epitem { | |
137 | /* RB-Tree node used to link this structure to the eventpoll rb-tree */ | |
138 | struct rb_node rbn; | |
139 | ||
140 | /* List header used to link this structure to the eventpoll ready list */ | |
141 | struct list_head rdllink; | |
142 | ||
143 | /* The file descriptor information this item refers to */ | |
144 | struct epoll_filefd ffd; | |
145 | ||
146 | /* Number of active wait queue attached to poll operations */ | |
147 | int nwait; | |
148 | ||
149 | /* List containing poll wait queues */ | |
150 | struct list_head pwqlist; | |
151 | ||
152 | /* The "container" of this item */ | |
153 | struct eventpoll *ep; | |
154 | ||
155 | /* The structure that describe the interested events and the source fd */ | |
156 | struct epoll_event event; | |
157 | ||
158 | /* | |
159 | * Used to keep track of the usage count of the structure. This avoids | |
160 | * that the structure will desappear from underneath our processing. | |
161 | */ | |
162 | atomic_t usecnt; | |
163 | ||
164 | /* List header used to link this item to the "struct file" items list */ | |
165 | struct list_head fllink; | |
166 | ||
167 | /* | |
168 | * Works together "struct eventpoll"->ovflist in keeping the | |
169 | * single linked chain of items. | |
170 | */ | |
171 | struct epitem *next; | |
172 | }; | |
173 | ||
1da177e4 LT |
174 | /* |
175 | * This structure is stored inside the "private_data" member of the file | |
176 | * structure and rapresent the main data sructure for the eventpoll | |
177 | * interface. | |
178 | */ | |
179 | struct eventpoll { | |
180 | /* Protect the this structure access */ | |
181 | rwlock_t lock; | |
182 | ||
183 | /* | |
d47de16c DL |
184 | * This mutex is used to ensure that files are not removed |
185 | * while epoll is using them. This is held during the event | |
186 | * collection loop, the file cleanup path, the epoll file exit | |
187 | * code and the ctl operations. | |
1da177e4 | 188 | */ |
d47de16c | 189 | struct mutex mtx; |
1da177e4 LT |
190 | |
191 | /* Wait queue used by sys_epoll_wait() */ | |
192 | wait_queue_head_t wq; | |
193 | ||
194 | /* Wait queue used by file->poll() */ | |
195 | wait_queue_head_t poll_wait; | |
196 | ||
197 | /* List of ready file descriptors */ | |
198 | struct list_head rdllist; | |
199 | ||
200 | /* RB-Tree root used to store monitored fd structs */ | |
201 | struct rb_root rbr; | |
d47de16c DL |
202 | |
203 | /* | |
204 | * This is a single linked list that chains all the "struct epitem" that | |
205 | * happened while transfering ready events to userspace w/out | |
206 | * holding ->lock. | |
207 | */ | |
208 | struct epitem *ovflist; | |
1da177e4 LT |
209 | }; |
210 | ||
211 | /* Wait structure used by the poll hooks */ | |
212 | struct eppoll_entry { | |
213 | /* List header used to link this structure to the "struct epitem" */ | |
214 | struct list_head llink; | |
215 | ||
216 | /* The "base" pointer is set to the container "struct epitem" */ | |
217 | void *base; | |
218 | ||
219 | /* | |
220 | * Wait queue item that will be linked to the target file wait | |
221 | * queue head. | |
222 | */ | |
223 | wait_queue_t wait; | |
224 | ||
225 | /* The wait queue head that linked the "wait" wait queue item */ | |
226 | wait_queue_head_t *whead; | |
227 | }; | |
228 | ||
1da177e4 LT |
229 | /* Wrapper struct used by poll queueing */ |
230 | struct ep_pqueue { | |
231 | poll_table pt; | |
232 | struct epitem *epi; | |
233 | }; | |
234 | ||
1da177e4 | 235 | /* |
d47de16c | 236 | * This mutex is used to serialize ep_free() and eventpoll_release_file(). |
1da177e4 | 237 | */ |
144efe3e | 238 | static struct mutex epmutex; |
1da177e4 LT |
239 | |
240 | /* Safe wake up implementation */ | |
241 | static struct poll_safewake psw; | |
242 | ||
243 | /* Slab cache used to allocate "struct epitem" */ | |
e18b890b | 244 | static struct kmem_cache *epi_cache __read_mostly; |
1da177e4 LT |
245 | |
246 | /* Slab cache used to allocate "struct eppoll_entry" */ | |
e18b890b | 247 | static struct kmem_cache *pwq_cache __read_mostly; |
1da177e4 | 248 | |
b030a4dd PE |
249 | |
250 | /* Setup the structure that is used as key for the rb-tree */ | |
251 | static inline void ep_set_ffd(struct epoll_filefd *ffd, | |
252 | struct file *file, int fd) | |
253 | { | |
254 | ffd->file = file; | |
255 | ffd->fd = fd; | |
256 | } | |
257 | ||
258 | /* Compare rb-tree keys */ | |
259 | static inline int ep_cmp_ffd(struct epoll_filefd *p1, | |
260 | struct epoll_filefd *p2) | |
261 | { | |
262 | return (p1->file > p2->file ? +1: | |
263 | (p1->file < p2->file ? -1 : p1->fd - p2->fd)); | |
264 | } | |
265 | ||
266 | /* Special initialization for the rb-tree node to detect linkage */ | |
267 | static inline void ep_rb_initnode(struct rb_node *n) | |
268 | { | |
c569882b | 269 | rb_set_parent(n, n); |
b030a4dd PE |
270 | } |
271 | ||
272 | /* Removes a node from the rb-tree and marks it for a fast is-linked check */ | |
273 | static inline void ep_rb_erase(struct rb_node *n, struct rb_root *r) | |
274 | { | |
275 | rb_erase(n, r); | |
c569882b | 276 | rb_set_parent(n, n); |
b030a4dd PE |
277 | } |
278 | ||
279 | /* Fast check to verify that the item is linked to the main rb-tree */ | |
280 | static inline int ep_rb_linked(struct rb_node *n) | |
281 | { | |
c569882b | 282 | return rb_parent(n) != n; |
b030a4dd PE |
283 | } |
284 | ||
b030a4dd PE |
285 | /* Tells us if the item is currently linked */ |
286 | static inline int ep_is_linked(struct list_head *p) | |
287 | { | |
288 | return !list_empty(p); | |
289 | } | |
290 | ||
291 | /* Get the "struct epitem" from a wait queue pointer */ | |
292 | static inline struct epitem * ep_item_from_wait(wait_queue_t *p) | |
293 | { | |
294 | return container_of(p, struct eppoll_entry, wait)->base; | |
295 | } | |
296 | ||
297 | /* Get the "struct epitem" from an epoll queue wrapper */ | |
298 | static inline struct epitem * ep_item_from_epqueue(poll_table *p) | |
299 | { | |
300 | return container_of(p, struct ep_pqueue, pt)->epi; | |
301 | } | |
302 | ||
303 | /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */ | |
6192bd53 | 304 | static inline int ep_op_has_event(int op) |
b030a4dd PE |
305 | { |
306 | return op != EPOLL_CTL_DEL; | |
307 | } | |
308 | ||
1da177e4 LT |
309 | /* Initialize the poll safe wake up structure */ |
310 | static void ep_poll_safewake_init(struct poll_safewake *psw) | |
311 | { | |
312 | ||
313 | INIT_LIST_HEAD(&psw->wake_task_list); | |
314 | spin_lock_init(&psw->lock); | |
315 | } | |
316 | ||
1da177e4 LT |
317 | /* |
318 | * Perform a safe wake up of the poll wait list. The problem is that | |
319 | * with the new callback'd wake up system, it is possible that the | |
320 | * poll callback is reentered from inside the call to wake_up() done | |
321 | * on the poll wait queue head. The rule is that we cannot reenter the | |
322 | * wake up code from the same task more than EP_MAX_POLLWAKE_NESTS times, | |
323 | * and we cannot reenter the same wait queue head at all. This will | |
324 | * enable to have a hierarchy of epoll file descriptor of no more than | |
325 | * EP_MAX_POLLWAKE_NESTS deep. We need the irq version of the spin lock | |
326 | * because this one gets called by the poll callback, that in turn is called | |
327 | * from inside a wake_up(), that might be called from irq context. | |
328 | */ | |
329 | static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq) | |
330 | { | |
331 | int wake_nests = 0; | |
332 | unsigned long flags; | |
36c8b586 | 333 | struct task_struct *this_task = current; |
1da177e4 LT |
334 | struct list_head *lsthead = &psw->wake_task_list, *lnk; |
335 | struct wake_task_node *tncur; | |
336 | struct wake_task_node tnode; | |
337 | ||
338 | spin_lock_irqsave(&psw->lock, flags); | |
339 | ||
340 | /* Try to see if the current task is already inside this wakeup call */ | |
341 | list_for_each(lnk, lsthead) { | |
342 | tncur = list_entry(lnk, struct wake_task_node, llink); | |
343 | ||
344 | if (tncur->wq == wq || | |
345 | (tncur->task == this_task && ++wake_nests > EP_MAX_POLLWAKE_NESTS)) { | |
346 | /* | |
347 | * Ops ... loop detected or maximum nest level reached. | |
348 | * We abort this wake by breaking the cycle itself. | |
349 | */ | |
350 | spin_unlock_irqrestore(&psw->lock, flags); | |
351 | return; | |
352 | } | |
353 | } | |
354 | ||
355 | /* Add the current task to the list */ | |
356 | tnode.task = this_task; | |
357 | tnode.wq = wq; | |
358 | list_add(&tnode.llink, lsthead); | |
359 | ||
360 | spin_unlock_irqrestore(&psw->lock, flags); | |
361 | ||
362 | /* Do really wake up now */ | |
363 | wake_up(wq); | |
364 | ||
365 | /* Remove the current task from the list */ | |
366 | spin_lock_irqsave(&psw->lock, flags); | |
367 | list_del(&tnode.llink); | |
368 | spin_unlock_irqrestore(&psw->lock, flags); | |
369 | } | |
370 | ||
1da177e4 | 371 | /* |
7699acd1 DL |
372 | * This function unregister poll callbacks from the associated file descriptor. |
373 | * Since this must be called without holding "ep->lock" the atomic exchange trick | |
374 | * will protect us from multiple unregister. | |
1da177e4 | 375 | */ |
7699acd1 | 376 | static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi) |
1da177e4 | 377 | { |
7699acd1 DL |
378 | int nwait; |
379 | struct list_head *lsthead = &epi->pwqlist; | |
380 | struct eppoll_entry *pwq; | |
1da177e4 | 381 | |
7699acd1 DL |
382 | /* This is called without locks, so we need the atomic exchange */ |
383 | nwait = xchg(&epi->nwait, 0); | |
1da177e4 | 384 | |
7699acd1 DL |
385 | if (nwait) { |
386 | while (!list_empty(lsthead)) { | |
387 | pwq = list_first_entry(lsthead, struct eppoll_entry, llink); | |
1da177e4 | 388 | |
7699acd1 DL |
389 | list_del_init(&pwq->llink); |
390 | remove_wait_queue(pwq->whead, &pwq->wait); | |
391 | kmem_cache_free(pwq_cache, pwq); | |
392 | } | |
1da177e4 | 393 | } |
1da177e4 LT |
394 | } |
395 | ||
1da177e4 | 396 | /* |
7699acd1 DL |
397 | * Unlink the "struct epitem" from all places it might have been hooked up. |
398 | * This function must be called with write IRQ lock on "ep->lock". | |
1da177e4 | 399 | */ |
7699acd1 | 400 | static int ep_unlink(struct eventpoll *ep, struct epitem *epi) |
1da177e4 | 401 | { |
7699acd1 | 402 | int error; |
1da177e4 | 403 | |
7699acd1 DL |
404 | /* |
405 | * It can happen that this one is called for an item already unlinked. | |
406 | * The check protect us from doing a double unlink ( crash ). | |
407 | */ | |
408 | error = -ENOENT; | |
409 | if (!ep_rb_linked(&epi->rbn)) | |
410 | goto error_return; | |
1da177e4 | 411 | |
53d2be79 | 412 | /* |
7699acd1 DL |
413 | * Clear the event mask for the unlinked item. This will avoid item |
414 | * notifications to be sent after the unlink operation from inside | |
415 | * the kernel->userspace event transfer loop. | |
53d2be79 | 416 | */ |
7699acd1 | 417 | epi->event.events = 0; |
1da177e4 LT |
418 | |
419 | /* | |
7699acd1 DL |
420 | * At this point is safe to do the job, unlink the item from our rb-tree. |
421 | * This operation togheter with the above check closes the door to | |
422 | * double unlinks. | |
1da177e4 | 423 | */ |
7699acd1 | 424 | ep_rb_erase(&epi->rbn, &ep->rbr); |
1da177e4 | 425 | |
7699acd1 DL |
426 | /* |
427 | * If the item we are going to remove is inside the ready file descriptors | |
428 | * we want to remove it from this list to avoid stale events. | |
429 | */ | |
430 | if (ep_is_linked(&epi->rdllink)) | |
431 | list_del_init(&epi->rdllink); | |
1da177e4 | 432 | |
7699acd1 DL |
433 | error = 0; |
434 | error_return: | |
435 | ||
436 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_unlink(%p, %p) = %d\n", | |
437 | current, ep, epi->ffd.file, error)); | |
1da177e4 | 438 | |
1da177e4 LT |
439 | return error; |
440 | } | |
441 | ||
1da177e4 | 442 | /* |
7699acd1 DL |
443 | * Increment the usage count of the "struct epitem" making it sure |
444 | * that the user will have a valid pointer to reference. | |
1da177e4 | 445 | */ |
7699acd1 | 446 | static void ep_use_epitem(struct epitem *epi) |
1da177e4 | 447 | { |
7699acd1 DL |
448 | atomic_inc(&epi->usecnt); |
449 | } | |
1da177e4 | 450 | |
7699acd1 DL |
451 | /* |
452 | * Decrement ( release ) the usage count by signaling that the user | |
453 | * has finished using the structure. It might lead to freeing the | |
454 | * structure itself if the count goes to zero. | |
455 | */ | |
456 | static void ep_release_epitem(struct epitem *epi) | |
457 | { | |
458 | if (atomic_dec_and_test(&epi->usecnt)) | |
459 | kmem_cache_free(epi_cache, epi); | |
460 | } | |
1da177e4 | 461 | |
7699acd1 DL |
462 | /* |
463 | * Removes a "struct epitem" from the eventpoll RB tree and deallocates | |
464 | * all the associated resources. | |
465 | */ | |
466 | static int ep_remove(struct eventpoll *ep, struct epitem *epi) | |
467 | { | |
468 | int error; | |
469 | unsigned long flags; | |
470 | struct file *file = epi->ffd.file; | |
1da177e4 LT |
471 | |
472 | /* | |
7699acd1 DL |
473 | * Removes poll wait queue hooks. We _have_ to do this without holding |
474 | * the "ep->lock" otherwise a deadlock might occur. This because of the | |
475 | * sequence of the lock acquisition. Here we do "ep->lock" then the wait | |
476 | * queue head lock when unregistering the wait queue. The wakeup callback | |
477 | * will run by holding the wait queue head lock and will call our callback | |
478 | * that will try to get "ep->lock". | |
1da177e4 | 479 | */ |
7699acd1 | 480 | ep_unregister_pollwait(ep, epi); |
1da177e4 | 481 | |
7699acd1 DL |
482 | /* Remove the current item from the list of epoll hooks */ |
483 | spin_lock(&file->f_ep_lock); | |
484 | if (ep_is_linked(&epi->fllink)) | |
485 | list_del_init(&epi->fllink); | |
486 | spin_unlock(&file->f_ep_lock); | |
1da177e4 | 487 | |
7699acd1 DL |
488 | /* We need to acquire the write IRQ lock before calling ep_unlink() */ |
489 | write_lock_irqsave(&ep->lock, flags); | |
1da177e4 | 490 | |
7699acd1 DL |
491 | /* Really unlink the item from the RB tree */ |
492 | error = ep_unlink(ep, epi); | |
1da177e4 | 493 | |
7699acd1 | 494 | write_unlock_irqrestore(&ep->lock, flags); |
1da177e4 | 495 | |
7699acd1 DL |
496 | if (error) |
497 | goto error_return; | |
1da177e4 | 498 | |
7699acd1 DL |
499 | /* At this point it is safe to free the eventpoll item */ |
500 | ep_release_epitem(epi); | |
1da177e4 | 501 | |
7699acd1 DL |
502 | error = 0; |
503 | error_return: | |
504 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_remove(%p, %p) = %d\n", | |
505 | current, ep, file, error)); | |
1da177e4 LT |
506 | |
507 | return error; | |
508 | } | |
509 | ||
7699acd1 | 510 | static void ep_free(struct eventpoll *ep) |
1da177e4 | 511 | { |
7699acd1 DL |
512 | struct rb_node *rbp; |
513 | struct epitem *epi; | |
1da177e4 | 514 | |
7699acd1 DL |
515 | /* We need to release all tasks waiting for these file */ |
516 | if (waitqueue_active(&ep->poll_wait)) | |
517 | ep_poll_safewake(&psw, &ep->poll_wait); | |
1da177e4 | 518 | |
7699acd1 DL |
519 | /* |
520 | * We need to lock this because we could be hit by | |
521 | * eventpoll_release_file() while we're freeing the "struct eventpoll". | |
d47de16c | 522 | * We do not need to hold "ep->mtx" here because the epoll file |
7699acd1 DL |
523 | * is on the way to be removed and no one has references to it |
524 | * anymore. The only hit might come from eventpoll_release_file() but | |
525 | * holding "epmutex" is sufficent here. | |
526 | */ | |
527 | mutex_lock(&epmutex); | |
1da177e4 LT |
528 | |
529 | /* | |
7699acd1 | 530 | * Walks through the whole tree by unregistering poll callbacks. |
1da177e4 | 531 | */ |
7699acd1 DL |
532 | for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) { |
533 | epi = rb_entry(rbp, struct epitem, rbn); | |
534 | ||
535 | ep_unregister_pollwait(ep, epi); | |
536 | } | |
1da177e4 LT |
537 | |
538 | /* | |
7699acd1 DL |
539 | * Walks through the whole tree by freeing each "struct epitem". At this |
540 | * point we are sure no poll callbacks will be lingering around, and also by | |
d47de16c | 541 | * holding "epmutex" we can be sure that no file cleanup code will hit |
7699acd1 | 542 | * us during this operation. So we can avoid the lock on "ep->lock". |
1da177e4 | 543 | */ |
7699acd1 DL |
544 | while ((rbp = rb_first(&ep->rbr)) != 0) { |
545 | epi = rb_entry(rbp, struct epitem, rbn); | |
546 | ep_remove(ep, epi); | |
547 | } | |
1da177e4 | 548 | |
7699acd1 | 549 | mutex_unlock(&epmutex); |
d47de16c DL |
550 | |
551 | mutex_destroy(&ep->mtx); | |
7699acd1 | 552 | } |
1da177e4 | 553 | |
7699acd1 DL |
554 | static int ep_eventpoll_release(struct inode *inode, struct file *file) |
555 | { | |
556 | struct eventpoll *ep = file->private_data; | |
1da177e4 | 557 | |
7699acd1 DL |
558 | if (ep) { |
559 | ep_free(ep); | |
560 | kfree(ep); | |
561 | } | |
562 | ||
563 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: close() ep=%p\n", current, ep)); | |
564 | return 0; | |
1da177e4 LT |
565 | } |
566 | ||
7699acd1 DL |
567 | static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait) |
568 | { | |
569 | unsigned int pollflags = 0; | |
570 | unsigned long flags; | |
571 | struct eventpoll *ep = file->private_data; | |
1da177e4 | 572 | |
7699acd1 DL |
573 | /* Insert inside our poll wait queue */ |
574 | poll_wait(file, &ep->poll_wait, wait); | |
575 | ||
576 | /* Check our condition */ | |
577 | read_lock_irqsave(&ep->lock, flags); | |
578 | if (!list_empty(&ep->rdllist)) | |
579 | pollflags = POLLIN | POLLRDNORM; | |
580 | read_unlock_irqrestore(&ep->lock, flags); | |
581 | ||
582 | return pollflags; | |
583 | } | |
584 | ||
585 | /* File callbacks that implement the eventpoll file behaviour */ | |
586 | static const struct file_operations eventpoll_fops = { | |
587 | .release = ep_eventpoll_release, | |
588 | .poll = ep_eventpoll_poll | |
589 | }; | |
590 | ||
591 | /* Fast test to see if the file is an evenpoll file */ | |
592 | static inline int is_file_epoll(struct file *f) | |
593 | { | |
594 | return f->f_op == &eventpoll_fops; | |
595 | } | |
b611967d DL |
596 | |
597 | /* | |
7699acd1 DL |
598 | * This is called from eventpoll_release() to unlink files from the eventpoll |
599 | * interface. We need to have this facility to cleanup correctly files that are | |
600 | * closed without being removed from the eventpoll interface. | |
b611967d | 601 | */ |
7699acd1 | 602 | void eventpoll_release_file(struct file *file) |
b611967d | 603 | { |
7699acd1 DL |
604 | struct list_head *lsthead = &file->f_ep_links; |
605 | struct eventpoll *ep; | |
606 | struct epitem *epi; | |
b611967d DL |
607 | |
608 | /* | |
7699acd1 DL |
609 | * We don't want to get "file->f_ep_lock" because it is not |
610 | * necessary. It is not necessary because we're in the "struct file" | |
611 | * cleanup path, and this means that noone is using this file anymore. | |
d47de16c | 612 | * The only hit might come from ep_free() but by holding the mutex |
7699acd1 | 613 | * will correctly serialize the operation. We do need to acquire |
d47de16c | 614 | * "ep->mtx" after "epmutex" because ep_remove() requires it when called |
7699acd1 | 615 | * from anywhere but ep_free(). |
b611967d | 616 | */ |
7699acd1 | 617 | mutex_lock(&epmutex); |
b611967d | 618 | |
7699acd1 DL |
619 | while (!list_empty(lsthead)) { |
620 | epi = list_first_entry(lsthead, struct epitem, fllink); | |
b611967d | 621 | |
7699acd1 DL |
622 | ep = epi->ep; |
623 | list_del_init(&epi->fllink); | |
d47de16c | 624 | mutex_lock(&ep->mtx); |
7699acd1 | 625 | ep_remove(ep, epi); |
d47de16c | 626 | mutex_unlock(&ep->mtx); |
b611967d DL |
627 | } |
628 | ||
7699acd1 | 629 | mutex_unlock(&epmutex); |
b611967d DL |
630 | } |
631 | ||
53d2be79 | 632 | static int ep_alloc(struct eventpoll **pep) |
1da177e4 | 633 | { |
53d2be79 | 634 | struct eventpoll *ep = kzalloc(sizeof(*ep), GFP_KERNEL); |
1da177e4 | 635 | |
53d2be79 | 636 | if (!ep) |
1da177e4 LT |
637 | return -ENOMEM; |
638 | ||
1da177e4 | 639 | rwlock_init(&ep->lock); |
d47de16c | 640 | mutex_init(&ep->mtx); |
1da177e4 LT |
641 | init_waitqueue_head(&ep->wq); |
642 | init_waitqueue_head(&ep->poll_wait); | |
643 | INIT_LIST_HEAD(&ep->rdllist); | |
644 | ep->rbr = RB_ROOT; | |
d47de16c | 645 | ep->ovflist = EP_UNACTIVE_PTR; |
1da177e4 | 646 | |
53d2be79 | 647 | *pep = ep; |
1da177e4 | 648 | |
53d2be79 | 649 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_alloc() ep=%p\n", |
1da177e4 LT |
650 | current, ep)); |
651 | return 0; | |
652 | } | |
653 | ||
1da177e4 | 654 | /* |
6192bd53 | 655 | * Search the file inside the eventpoll tree. It add usage count to |
1da177e4 LT |
656 | * the returned item, so the caller must call ep_release_epitem() |
657 | * after finished using the "struct epitem". | |
658 | */ | |
659 | static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd) | |
660 | { | |
661 | int kcmp; | |
662 | unsigned long flags; | |
663 | struct rb_node *rbp; | |
664 | struct epitem *epi, *epir = NULL; | |
665 | struct epoll_filefd ffd; | |
666 | ||
b030a4dd | 667 | ep_set_ffd(&ffd, file, fd); |
1da177e4 LT |
668 | read_lock_irqsave(&ep->lock, flags); |
669 | for (rbp = ep->rbr.rb_node; rbp; ) { | |
670 | epi = rb_entry(rbp, struct epitem, rbn); | |
b030a4dd | 671 | kcmp = ep_cmp_ffd(&ffd, &epi->ffd); |
1da177e4 LT |
672 | if (kcmp > 0) |
673 | rbp = rbp->rb_right; | |
674 | else if (kcmp < 0) | |
675 | rbp = rbp->rb_left; | |
676 | else { | |
677 | ep_use_epitem(epi); | |
678 | epir = epi; | |
679 | break; | |
680 | } | |
681 | } | |
682 | read_unlock_irqrestore(&ep->lock, flags); | |
683 | ||
684 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_find(%p) -> %p\n", | |
685 | current, file, epir)); | |
686 | ||
687 | return epir; | |
688 | } | |
689 | ||
1da177e4 | 690 | /* |
7699acd1 DL |
691 | * This is the callback that is passed to the wait queue wakeup |
692 | * machanism. It is called by the stored file descriptors when they | |
693 | * have events to report. | |
1da177e4 | 694 | */ |
7699acd1 | 695 | static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key) |
1da177e4 | 696 | { |
7699acd1 DL |
697 | int pwake = 0; |
698 | unsigned long flags; | |
699 | struct epitem *epi = ep_item_from_wait(wait); | |
700 | struct eventpoll *ep = epi->ep; | |
1da177e4 | 701 | |
7699acd1 DL |
702 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: poll_callback(%p) epi=%p ep=%p\n", |
703 | current, epi->ffd.file, epi, ep)); | |
1da177e4 | 704 | |
7699acd1 | 705 | write_lock_irqsave(&ep->lock, flags); |
1da177e4 | 706 | |
7699acd1 DL |
707 | /* |
708 | * If the event mask does not contain any poll(2) event, we consider the | |
709 | * descriptor to be disabled. This condition is likely the effect of the | |
710 | * EPOLLONESHOT bit that disables the descriptor when an event is received, | |
711 | * until the next EPOLL_CTL_MOD will be issued. | |
712 | */ | |
713 | if (!(epi->event.events & ~EP_PRIVATE_BITS)) | |
d47de16c DL |
714 | goto out_unlock; |
715 | ||
716 | /* | |
717 | * If we are trasfering events to userspace, we can hold no locks | |
718 | * (because we're accessing user memory, and because of linux f_op->poll() | |
719 | * semantics). All the events that happens during that period of time are | |
720 | * chained in ep->ovflist and requeued later on. | |
721 | */ | |
722 | if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) { | |
723 | if (epi->next == EP_UNACTIVE_PTR) { | |
724 | epi->next = ep->ovflist; | |
725 | ep->ovflist = epi; | |
726 | } | |
727 | goto out_unlock; | |
728 | } | |
1da177e4 | 729 | |
7699acd1 DL |
730 | /* If this file is already in the ready list we exit soon */ |
731 | if (ep_is_linked(&epi->rdllink)) | |
732 | goto is_linked; | |
733 | ||
734 | list_add_tail(&epi->rdllink, &ep->rdllist); | |
735 | ||
736 | is_linked: | |
737 | /* | |
738 | * Wake up ( if active ) both the eventpoll wait list and the ->poll() | |
739 | * wait list. | |
740 | */ | |
741 | if (waitqueue_active(&ep->wq)) | |
742 | __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE | | |
743 | TASK_INTERRUPTIBLE); | |
744 | if (waitqueue_active(&ep->poll_wait)) | |
745 | pwake++; | |
746 | ||
d47de16c | 747 | out_unlock: |
7699acd1 | 748 | write_unlock_irqrestore(&ep->lock, flags); |
1da177e4 | 749 | |
7699acd1 DL |
750 | /* We have to call this outside the lock */ |
751 | if (pwake) | |
752 | ep_poll_safewake(&psw, &ep->poll_wait); | |
753 | ||
754 | return 1; | |
755 | } | |
1da177e4 LT |
756 | |
757 | /* | |
758 | * This is the callback that is used to add our wait queue to the | |
759 | * target file wakeup lists. | |
760 | */ | |
761 | static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead, | |
762 | poll_table *pt) | |
763 | { | |
b030a4dd | 764 | struct epitem *epi = ep_item_from_epqueue(pt); |
1da177e4 LT |
765 | struct eppoll_entry *pwq; |
766 | ||
e94b1766 | 767 | if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) { |
1da177e4 LT |
768 | init_waitqueue_func_entry(&pwq->wait, ep_poll_callback); |
769 | pwq->whead = whead; | |
770 | pwq->base = epi; | |
771 | add_wait_queue(whead, &pwq->wait); | |
772 | list_add_tail(&pwq->llink, &epi->pwqlist); | |
773 | epi->nwait++; | |
774 | } else { | |
775 | /* We have to signal that an error occurred */ | |
776 | epi->nwait = -1; | |
777 | } | |
778 | } | |
779 | ||
1da177e4 LT |
780 | static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi) |
781 | { | |
782 | int kcmp; | |
783 | struct rb_node **p = &ep->rbr.rb_node, *parent = NULL; | |
784 | struct epitem *epic; | |
785 | ||
786 | while (*p) { | |
787 | parent = *p; | |
788 | epic = rb_entry(parent, struct epitem, rbn); | |
b030a4dd | 789 | kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd); |
1da177e4 LT |
790 | if (kcmp > 0) |
791 | p = &parent->rb_right; | |
792 | else | |
793 | p = &parent->rb_left; | |
794 | } | |
795 | rb_link_node(&epi->rbn, parent, p); | |
796 | rb_insert_color(&epi->rbn, &ep->rbr); | |
797 | } | |
798 | ||
1da177e4 LT |
799 | static int ep_insert(struct eventpoll *ep, struct epoll_event *event, |
800 | struct file *tfile, int fd) | |
801 | { | |
802 | int error, revents, pwake = 0; | |
803 | unsigned long flags; | |
804 | struct epitem *epi; | |
805 | struct ep_pqueue epq; | |
806 | ||
807 | error = -ENOMEM; | |
e94b1766 | 808 | if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL))) |
7699acd1 | 809 | goto error_return; |
1da177e4 LT |
810 | |
811 | /* Item initialization follow here ... */ | |
b030a4dd | 812 | ep_rb_initnode(&epi->rbn); |
1da177e4 LT |
813 | INIT_LIST_HEAD(&epi->rdllink); |
814 | INIT_LIST_HEAD(&epi->fllink); | |
1da177e4 LT |
815 | INIT_LIST_HEAD(&epi->pwqlist); |
816 | epi->ep = ep; | |
b030a4dd | 817 | ep_set_ffd(&epi->ffd, tfile, fd); |
1da177e4 LT |
818 | epi->event = *event; |
819 | atomic_set(&epi->usecnt, 1); | |
820 | epi->nwait = 0; | |
d47de16c | 821 | epi->next = EP_UNACTIVE_PTR; |
1da177e4 LT |
822 | |
823 | /* Initialize the poll table using the queue callback */ | |
824 | epq.epi = epi; | |
825 | init_poll_funcptr(&epq.pt, ep_ptable_queue_proc); | |
826 | ||
827 | /* | |
828 | * Attach the item to the poll hooks and get current event bits. | |
829 | * We can safely use the file* here because its usage count has | |
830 | * been increased by the caller of this function. | |
831 | */ | |
832 | revents = tfile->f_op->poll(tfile, &epq.pt); | |
833 | ||
834 | /* | |
835 | * We have to check if something went wrong during the poll wait queue | |
836 | * install process. Namely an allocation for a wait queue failed due | |
837 | * high memory pressure. | |
838 | */ | |
839 | if (epi->nwait < 0) | |
7699acd1 | 840 | goto error_unregister; |
1da177e4 LT |
841 | |
842 | /* Add the current item to the list of active epoll hook for this file */ | |
843 | spin_lock(&tfile->f_ep_lock); | |
844 | list_add_tail(&epi->fllink, &tfile->f_ep_links); | |
845 | spin_unlock(&tfile->f_ep_lock); | |
846 | ||
847 | /* We have to drop the new item inside our item list to keep track of it */ | |
848 | write_lock_irqsave(&ep->lock, flags); | |
849 | ||
850 | /* Add the current item to the rb-tree */ | |
851 | ep_rbtree_insert(ep, epi); | |
852 | ||
853 | /* If the file is already "ready" we drop it inside the ready list */ | |
b030a4dd | 854 | if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) { |
1da177e4 LT |
855 | list_add_tail(&epi->rdllink, &ep->rdllist); |
856 | ||
857 | /* Notify waiting tasks that events are available */ | |
858 | if (waitqueue_active(&ep->wq)) | |
3419b23a | 859 | __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE); |
1da177e4 LT |
860 | if (waitqueue_active(&ep->poll_wait)) |
861 | pwake++; | |
862 | } | |
863 | ||
864 | write_unlock_irqrestore(&ep->lock, flags); | |
865 | ||
866 | /* We have to call this outside the lock */ | |
867 | if (pwake) | |
868 | ep_poll_safewake(&psw, &ep->poll_wait); | |
869 | ||
870 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_insert(%p, %p, %d)\n", | |
871 | current, ep, tfile, fd)); | |
872 | ||
873 | return 0; | |
874 | ||
7699acd1 | 875 | error_unregister: |
1da177e4 LT |
876 | ep_unregister_pollwait(ep, epi); |
877 | ||
878 | /* | |
879 | * We need to do this because an event could have been arrived on some | |
880 | * allocated wait queue. | |
881 | */ | |
882 | write_lock_irqsave(&ep->lock, flags); | |
b030a4dd | 883 | if (ep_is_linked(&epi->rdllink)) |
6192bd53 | 884 | list_del_init(&epi->rdllink); |
1da177e4 LT |
885 | write_unlock_irqrestore(&ep->lock, flags); |
886 | ||
b030a4dd | 887 | kmem_cache_free(epi_cache, epi); |
7699acd1 | 888 | error_return: |
1da177e4 LT |
889 | return error; |
890 | } | |
891 | ||
1da177e4 LT |
892 | /* |
893 | * Modify the interest event mask by dropping an event if the new mask | |
894 | * has a match in the current file status. | |
895 | */ | |
896 | static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event) | |
897 | { | |
898 | int pwake = 0; | |
899 | unsigned int revents; | |
900 | unsigned long flags; | |
901 | ||
902 | /* | |
903 | * Set the new event interest mask before calling f_op->poll(), otherwise | |
904 | * a potential race might occur. In fact if we do this operation inside | |
905 | * the lock, an event might happen between the f_op->poll() call and the | |
906 | * new event set registering. | |
907 | */ | |
908 | epi->event.events = event->events; | |
909 | ||
910 | /* | |
911 | * Get current event bits. We can safely use the file* here because | |
912 | * its usage count has been increased by the caller of this function. | |
913 | */ | |
914 | revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL); | |
915 | ||
916 | write_lock_irqsave(&ep->lock, flags); | |
917 | ||
918 | /* Copy the data member from inside the lock */ | |
919 | epi->event.data = event->data; | |
920 | ||
921 | /* | |
6192bd53 | 922 | * If the item is not linked to the RB tree it means that it's on its |
1da177e4 LT |
923 | * way toward the removal. Do nothing in this case. |
924 | */ | |
b030a4dd | 925 | if (ep_rb_linked(&epi->rbn)) { |
1da177e4 LT |
926 | /* |
927 | * If the item is "hot" and it is not registered inside the ready | |
928 | * list, push it inside. If the item is not "hot" and it is currently | |
929 | * registered inside the ready list, unlink it. | |
930 | */ | |
931 | if (revents & event->events) { | |
b030a4dd | 932 | if (!ep_is_linked(&epi->rdllink)) { |
1da177e4 LT |
933 | list_add_tail(&epi->rdllink, &ep->rdllist); |
934 | ||
935 | /* Notify waiting tasks that events are available */ | |
936 | if (waitqueue_active(&ep->wq)) | |
3419b23a DL |
937 | __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE | |
938 | TASK_INTERRUPTIBLE); | |
7699acd1 DL |
939 | if (waitqueue_active(&ep->poll_wait)) |
940 | pwake++; | |
941 | } | |
942 | } | |
943 | } | |
1da177e4 | 944 | |
7699acd1 | 945 | write_unlock_irqrestore(&ep->lock, flags); |
1da177e4 | 946 | |
7699acd1 DL |
947 | /* We have to call this outside the lock */ |
948 | if (pwake) | |
949 | ep_poll_safewake(&psw, &ep->poll_wait); | |
1da177e4 | 950 | |
7699acd1 | 951 | return 0; |
1da177e4 LT |
952 | } |
953 | ||
d47de16c DL |
954 | static int ep_send_events(struct eventpoll *ep, struct epoll_event __user *events, |
955 | int maxevents) | |
1da177e4 | 956 | { |
6192bd53 | 957 | int eventcnt, error = -EFAULT, pwake = 0; |
1da177e4 | 958 | unsigned int revents; |
6192bd53 | 959 | unsigned long flags; |
d47de16c DL |
960 | struct epitem *epi, *nepi; |
961 | struct list_head txlist; | |
962 | ||
963 | INIT_LIST_HEAD(&txlist); | |
6192bd53 | 964 | |
d47de16c DL |
965 | /* |
966 | * We need to lock this because we could be hit by | |
967 | * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL). | |
968 | */ | |
969 | mutex_lock(&ep->mtx); | |
970 | ||
971 | /* | |
972 | * Steal the ready list, and re-init the original one to the | |
973 | * empty list. Also, set ep->ovflist to NULL so that events | |
974 | * happening while looping w/out locks, are not lost. We cannot | |
975 | * have the poll callback to queue directly on ep->rdllist, | |
976 | * because we are doing it in the loop below, in a lockless way. | |
977 | */ | |
978 | write_lock_irqsave(&ep->lock, flags); | |
979 | list_splice(&ep->rdllist, &txlist); | |
980 | INIT_LIST_HEAD(&ep->rdllist); | |
981 | ep->ovflist = NULL; | |
982 | write_unlock_irqrestore(&ep->lock, flags); | |
1da177e4 LT |
983 | |
984 | /* | |
985 | * We can loop without lock because this is a task private list. | |
6192bd53 | 986 | * We just splice'd out the ep->rdllist in ep_collect_ready_items(). |
d47de16c | 987 | * Items cannot vanish during the loop because we are holding "mtx". |
1da177e4 | 988 | */ |
d47de16c DL |
989 | for (eventcnt = 0; !list_empty(&txlist) && eventcnt < maxevents;) { |
990 | epi = list_first_entry(&txlist, struct epitem, rdllink); | |
991 | ||
992 | list_del_init(&epi->rdllink); | |
1da177e4 LT |
993 | |
994 | /* | |
995 | * Get the ready file event set. We can safely use the file | |
d47de16c DL |
996 | * because we are holding the "mtx" and this will guarantee |
997 | * that both the file and the item will not vanish. | |
1da177e4 LT |
998 | */ |
999 | revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL); | |
6192bd53 | 1000 | revents &= epi->event.events; |
1da177e4 LT |
1001 | |
1002 | /* | |
6192bd53 DL |
1003 | * Is the event mask intersect the caller-requested one, |
1004 | * deliver the event to userspace. Again, we are holding | |
d47de16c DL |
1005 | * "mtx", so no operations coming from userspace can change |
1006 | * the item. | |
1da177e4 | 1007 | */ |
6192bd53 DL |
1008 | if (revents) { |
1009 | if (__put_user(revents, | |
1da177e4 LT |
1010 | &events[eventcnt].events) || |
1011 | __put_user(epi->event.data, | |
1012 | &events[eventcnt].data)) | |
6192bd53 | 1013 | goto errxit; |
1da177e4 LT |
1014 | if (epi->event.events & EPOLLONESHOT) |
1015 | epi->event.events &= EP_PRIVATE_BITS; | |
1016 | eventcnt++; | |
1017 | } | |
1da177e4 | 1018 | /* |
d47de16c DL |
1019 | * At this point, noone can insert into ep->rdllist besides |
1020 | * us. The epoll_ctl() callers are locked out by us holding | |
1021 | * "mtx" and the poll callback will queue them in ep->ovflist. | |
1da177e4 | 1022 | */ |
6192bd53 | 1023 | if (!(epi->event.events & EPOLLET) && |
d47de16c DL |
1024 | (revents & epi->event.events)) |
1025 | list_add_tail(&epi->rdllink, &ep->rdllist); | |
1da177e4 | 1026 | } |
6192bd53 | 1027 | error = 0; |
1da177e4 | 1028 | |
d47de16c | 1029 | errxit: |
6192bd53 | 1030 | |
d47de16c DL |
1031 | write_lock_irqsave(&ep->lock, flags); |
1032 | /* | |
1033 | * During the time we spent in the loop above, some other events | |
1034 | * might have been queued by the poll callback. We re-insert them | |
1035 | * here (in case they are not already queued, or they're one-shot). | |
1036 | */ | |
1037 | for (nepi = ep->ovflist; (epi = nepi) != NULL; | |
1038 | nepi = epi->next, epi->next = EP_UNACTIVE_PTR) { | |
1039 | if (!ep_is_linked(&epi->rdllink) && | |
1040 | (epi->event.events & ~EP_PRIVATE_BITS)) | |
1041 | list_add_tail(&epi->rdllink, &ep->rdllist); | |
1042 | } | |
6192bd53 | 1043 | /* |
d47de16c DL |
1044 | * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after |
1045 | * releasing the lock, events will be queued in the normal way inside | |
1046 | * ep->rdllist. | |
6192bd53 | 1047 | */ |
d47de16c | 1048 | ep->ovflist = EP_UNACTIVE_PTR; |
6192bd53 | 1049 | |
d47de16c DL |
1050 | /* |
1051 | * In case of error in the event-send loop, we might still have items | |
1052 | * inside the "txlist". We need to splice them back inside ep->rdllist. | |
1053 | */ | |
1054 | list_splice(&txlist, &ep->rdllist); | |
1055 | ||
1056 | if (!list_empty(&ep->rdllist)) { | |
1da177e4 | 1057 | /* |
d47de16c | 1058 | * Wake up (if active) both the eventpoll wait list and the ->poll() |
1da177e4 LT |
1059 | * wait list. |
1060 | */ | |
1061 | if (waitqueue_active(&ep->wq)) | |
3419b23a DL |
1062 | __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE | |
1063 | TASK_INTERRUPTIBLE); | |
1da177e4 LT |
1064 | if (waitqueue_active(&ep->poll_wait)) |
1065 | pwake++; | |
6192bd53 | 1066 | } |
d47de16c DL |
1067 | write_unlock_irqrestore(&ep->lock, flags); |
1068 | ||
1069 | mutex_unlock(&ep->mtx); | |
1da177e4 LT |
1070 | |
1071 | /* We have to call this outside the lock */ | |
1072 | if (pwake) | |
1073 | ep_poll_safewake(&psw, &ep->poll_wait); | |
6192bd53 DL |
1074 | |
1075 | return eventcnt == 0 ? error: eventcnt; | |
1da177e4 LT |
1076 | } |
1077 | ||
1da177e4 LT |
1078 | static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events, |
1079 | int maxevents, long timeout) | |
1080 | { | |
1081 | int res, eavail; | |
1082 | unsigned long flags; | |
1083 | long jtimeout; | |
1084 | wait_queue_t wait; | |
1085 | ||
1086 | /* | |
1087 | * Calculate the timeout by checking for the "infinite" value ( -1 ) | |
1088 | * and the overflow condition. The passed timeout is in milliseconds, | |
1089 | * that why (t * HZ) / 1000. | |
1090 | */ | |
e3306dd5 DL |
1091 | jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ? |
1092 | MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000; | |
1da177e4 LT |
1093 | |
1094 | retry: | |
1095 | write_lock_irqsave(&ep->lock, flags); | |
1096 | ||
1097 | res = 0; | |
1098 | if (list_empty(&ep->rdllist)) { | |
1099 | /* | |
1100 | * We don't have any available event to return to the caller. | |
1101 | * We need to sleep here, and we will be wake up by | |
1102 | * ep_poll_callback() when events will become available. | |
1103 | */ | |
1104 | init_waitqueue_entry(&wait, current); | |
d47de16c | 1105 | wait.flags |= WQ_FLAG_EXCLUSIVE; |
3419b23a | 1106 | __add_wait_queue(&ep->wq, &wait); |
1da177e4 LT |
1107 | |
1108 | for (;;) { | |
1109 | /* | |
1110 | * We don't want to sleep if the ep_poll_callback() sends us | |
1111 | * a wakeup in between. That's why we set the task state | |
1112 | * to TASK_INTERRUPTIBLE before doing the checks. | |
1113 | */ | |
1114 | set_current_state(TASK_INTERRUPTIBLE); | |
1115 | if (!list_empty(&ep->rdllist) || !jtimeout) | |
1116 | break; | |
1117 | if (signal_pending(current)) { | |
1118 | res = -EINTR; | |
1119 | break; | |
1120 | } | |
1121 | ||
1122 | write_unlock_irqrestore(&ep->lock, flags); | |
1123 | jtimeout = schedule_timeout(jtimeout); | |
1124 | write_lock_irqsave(&ep->lock, flags); | |
1125 | } | |
3419b23a | 1126 | __remove_wait_queue(&ep->wq, &wait); |
1da177e4 LT |
1127 | |
1128 | set_current_state(TASK_RUNNING); | |
1129 | } | |
1130 | ||
1131 | /* Is it worth to try to dig for events ? */ | |
1132 | eavail = !list_empty(&ep->rdllist); | |
1133 | ||
1134 | write_unlock_irqrestore(&ep->lock, flags); | |
1135 | ||
1136 | /* | |
1137 | * Try to transfer events to user space. In case we get 0 events and | |
1138 | * there's still timeout left over, we go trying again in search of | |
1139 | * more luck. | |
1140 | */ | |
1141 | if (!res && eavail && | |
d47de16c | 1142 | !(res = ep_send_events(ep, events, maxevents)) && jtimeout) |
1da177e4 LT |
1143 | goto retry; |
1144 | ||
1145 | return res; | |
1146 | } | |
1147 | ||
7699acd1 DL |
1148 | /* |
1149 | * It opens an eventpoll file descriptor by suggesting a storage of "size" | |
1150 | * file descriptors. The size parameter is just an hint about how to size | |
1151 | * data structures. It won't prevent the user to store more than "size" | |
1152 | * file descriptors inside the epoll interface. It is the kernel part of | |
1153 | * the userspace epoll_create(2). | |
1154 | */ | |
1155 | asmlinkage long sys_epoll_create(int size) | |
1156 | { | |
1157 | int error, fd = -1; | |
1158 | struct eventpoll *ep; | |
1159 | struct inode *inode; | |
1160 | struct file *file; | |
1161 | ||
1162 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d)\n", | |
1163 | current, size)); | |
1164 | ||
1165 | /* | |
1166 | * Sanity check on the size parameter, and create the internal data | |
1167 | * structure ( "struct eventpoll" ). | |
1168 | */ | |
1169 | error = -EINVAL; | |
1170 | if (size <= 0 || (error = ep_alloc(&ep)) != 0) | |
1171 | goto error_return; | |
1172 | ||
1173 | /* | |
1174 | * Creates all the items needed to setup an eventpoll file. That is, | |
1175 | * a file structure, and inode and a free file descriptor. | |
1176 | */ | |
1177 | error = anon_inode_getfd(&fd, &inode, &file, "[eventpoll]", | |
1178 | &eventpoll_fops, ep); | |
1179 | if (error) | |
1180 | goto error_free; | |
1181 | ||
1182 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n", | |
1183 | current, size, fd)); | |
1184 | ||
1185 | return fd; | |
1186 | ||
1187 | error_free: | |
1188 | ep_free(ep); | |
1189 | kfree(ep); | |
1190 | error_return: | |
1191 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n", | |
1192 | current, size, error)); | |
1193 | return error; | |
1194 | } | |
1195 | ||
1196 | /* | |
1197 | * The following function implements the controller interface for | |
1198 | * the eventpoll file that enables the insertion/removal/change of | |
1199 | * file descriptors inside the interest set. It represents | |
1200 | * the kernel part of the user space epoll_ctl(2). | |
1201 | */ | |
1202 | asmlinkage long sys_epoll_ctl(int epfd, int op, int fd, | |
1203 | struct epoll_event __user *event) | |
1204 | { | |
1205 | int error; | |
1206 | struct file *file, *tfile; | |
1207 | struct eventpoll *ep; | |
1208 | struct epitem *epi; | |
1209 | struct epoll_event epds; | |
1210 | ||
1211 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p)\n", | |
1212 | current, epfd, op, fd, event)); | |
1213 | ||
1214 | error = -EFAULT; | |
1215 | if (ep_op_has_event(op) && | |
1216 | copy_from_user(&epds, event, sizeof(struct epoll_event))) | |
1217 | goto error_return; | |
1218 | ||
1219 | /* Get the "struct file *" for the eventpoll file */ | |
1220 | error = -EBADF; | |
1221 | file = fget(epfd); | |
1222 | if (!file) | |
1223 | goto error_return; | |
1224 | ||
1225 | /* Get the "struct file *" for the target file */ | |
1226 | tfile = fget(fd); | |
1227 | if (!tfile) | |
1228 | goto error_fput; | |
1229 | ||
1230 | /* The target file descriptor must support poll */ | |
1231 | error = -EPERM; | |
1232 | if (!tfile->f_op || !tfile->f_op->poll) | |
1233 | goto error_tgt_fput; | |
1234 | ||
1235 | /* | |
1236 | * We have to check that the file structure underneath the file descriptor | |
1237 | * the user passed to us _is_ an eventpoll file. And also we do not permit | |
1238 | * adding an epoll file descriptor inside itself. | |
1239 | */ | |
1240 | error = -EINVAL; | |
1241 | if (file == tfile || !is_file_epoll(file)) | |
1242 | goto error_tgt_fput; | |
1243 | ||
1244 | /* | |
1245 | * At this point it is safe to assume that the "private_data" contains | |
1246 | * our own data structure. | |
1247 | */ | |
1248 | ep = file->private_data; | |
1249 | ||
d47de16c | 1250 | mutex_lock(&ep->mtx); |
7699acd1 DL |
1251 | |
1252 | /* Try to lookup the file inside our RB tree */ | |
1253 | epi = ep_find(ep, tfile, fd); | |
1254 | ||
1255 | error = -EINVAL; | |
1256 | switch (op) { | |
1257 | case EPOLL_CTL_ADD: | |
1258 | if (!epi) { | |
1259 | epds.events |= POLLERR | POLLHUP; | |
1260 | ||
1261 | error = ep_insert(ep, &epds, tfile, fd); | |
1262 | } else | |
1263 | error = -EEXIST; | |
1264 | break; | |
1265 | case EPOLL_CTL_DEL: | |
1266 | if (epi) | |
1267 | error = ep_remove(ep, epi); | |
1268 | else | |
1269 | error = -ENOENT; | |
1270 | break; | |
1271 | case EPOLL_CTL_MOD: | |
1272 | if (epi) { | |
1273 | epds.events |= POLLERR | POLLHUP; | |
1274 | error = ep_modify(ep, epi, &epds); | |
1275 | } else | |
1276 | error = -ENOENT; | |
1277 | break; | |
1278 | } | |
1279 | /* | |
1280 | * The function ep_find() increments the usage count of the structure | |
1281 | * so, if this is not NULL, we need to release it. | |
1282 | */ | |
1283 | if (epi) | |
1284 | ep_release_epitem(epi); | |
d47de16c | 1285 | mutex_unlock(&ep->mtx); |
7699acd1 DL |
1286 | |
1287 | error_tgt_fput: | |
1288 | fput(tfile); | |
1289 | error_fput: | |
1290 | fput(file); | |
1291 | error_return: | |
1292 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p) = %d\n", | |
1293 | current, epfd, op, fd, event, error)); | |
1294 | ||
1295 | return error; | |
1296 | } | |
1297 | ||
1298 | /* | |
1299 | * Implement the event wait interface for the eventpoll file. It is the kernel | |
1300 | * part of the user space epoll_wait(2). | |
1301 | */ | |
1302 | asmlinkage long sys_epoll_wait(int epfd, struct epoll_event __user *events, | |
1303 | int maxevents, int timeout) | |
1304 | { | |
1305 | int error; | |
1306 | struct file *file; | |
1307 | struct eventpoll *ep; | |
1308 | ||
1309 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d)\n", | |
1310 | current, epfd, events, maxevents, timeout)); | |
1311 | ||
1312 | /* The maximum number of event must be greater than zero */ | |
1313 | if (maxevents <= 0 || maxevents > EP_MAX_EVENTS) | |
1314 | return -EINVAL; | |
1315 | ||
1316 | /* Verify that the area passed by the user is writeable */ | |
1317 | if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) { | |
1318 | error = -EFAULT; | |
1319 | goto error_return; | |
1320 | } | |
1321 | ||
1322 | /* Get the "struct file *" for the eventpoll file */ | |
1323 | error = -EBADF; | |
1324 | file = fget(epfd); | |
1325 | if (!file) | |
1326 | goto error_return; | |
1327 | ||
1328 | /* | |
1329 | * We have to check that the file structure underneath the fd | |
1330 | * the user passed to us _is_ an eventpoll file. | |
1331 | */ | |
1332 | error = -EINVAL; | |
1333 | if (!is_file_epoll(file)) | |
1334 | goto error_fput; | |
1335 | ||
1336 | /* | |
1337 | * At this point it is safe to assume that the "private_data" contains | |
1338 | * our own data structure. | |
1339 | */ | |
1340 | ep = file->private_data; | |
1341 | ||
1342 | /* Time to fish for events ... */ | |
1343 | error = ep_poll(ep, events, maxevents, timeout); | |
1344 | ||
1345 | error_fput: | |
1346 | fput(file); | |
1347 | error_return: | |
1348 | DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d) = %d\n", | |
1349 | current, epfd, events, maxevents, timeout, error)); | |
1350 | ||
1351 | return error; | |
1352 | } | |
1353 | ||
1354 | #ifdef TIF_RESTORE_SIGMASK | |
1355 | ||
1356 | /* | |
1357 | * Implement the event wait interface for the eventpoll file. It is the kernel | |
1358 | * part of the user space epoll_pwait(2). | |
1359 | */ | |
1360 | asmlinkage long sys_epoll_pwait(int epfd, struct epoll_event __user *events, | |
1361 | int maxevents, int timeout, const sigset_t __user *sigmask, | |
1362 | size_t sigsetsize) | |
1363 | { | |
1364 | int error; | |
1365 | sigset_t ksigmask, sigsaved; | |
1366 | ||
1367 | /* | |
1368 | * If the caller wants a certain signal mask to be set during the wait, | |
1369 | * we apply it here. | |
1370 | */ | |
1371 | if (sigmask) { | |
1372 | if (sigsetsize != sizeof(sigset_t)) | |
1373 | return -EINVAL; | |
1374 | if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask))) | |
1375 | return -EFAULT; | |
1376 | sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP)); | |
1377 | sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved); | |
1378 | } | |
1379 | ||
1380 | error = sys_epoll_wait(epfd, events, maxevents, timeout); | |
1381 | ||
1382 | /* | |
1383 | * If we changed the signal mask, we need to restore the original one. | |
1384 | * In case we've got a signal while waiting, we do not restore the | |
1385 | * signal mask yet, and we allow do_signal() to deliver the signal on | |
1386 | * the way back to userspace, before the signal mask is restored. | |
1387 | */ | |
1388 | if (sigmask) { | |
1389 | if (error == -EINTR) { | |
1390 | memcpy(¤t->saved_sigmask, &sigsaved, | |
1391 | sizeof(sigsaved)); | |
1392 | set_thread_flag(TIF_RESTORE_SIGMASK); | |
1393 | } else | |
1394 | sigprocmask(SIG_SETMASK, &sigsaved, NULL); | |
1395 | } | |
1396 | ||
1397 | return error; | |
1398 | } | |
1399 | ||
1400 | #endif /* #ifdef TIF_RESTORE_SIGMASK */ | |
1401 | ||
1da177e4 LT |
1402 | static int __init eventpoll_init(void) |
1403 | { | |
144efe3e | 1404 | mutex_init(&epmutex); |
1da177e4 LT |
1405 | |
1406 | /* Initialize the structure used to perform safe poll wait head wake ups */ | |
1407 | ep_poll_safewake_init(&psw); | |
1408 | ||
1409 | /* Allocates slab cache used to allocate "struct epitem" items */ | |
1410 | epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem), | |
1411 | 0, SLAB_HWCACHE_ALIGN|EPI_SLAB_DEBUG|SLAB_PANIC, | |
1412 | NULL, NULL); | |
1413 | ||
1414 | /* Allocates slab cache used to allocate "struct eppoll_entry" */ | |
1415 | pwq_cache = kmem_cache_create("eventpoll_pwq", | |
1416 | sizeof(struct eppoll_entry), 0, | |
1417 | EPI_SLAB_DEBUG|SLAB_PANIC, NULL, NULL); | |
1418 | ||
1da177e4 | 1419 | return 0; |
1da177e4 | 1420 | } |
cea69241 | 1421 | fs_initcall(eventpoll_init); |
1da177e4 | 1422 |