blob: f5c88435c6be4042183713ac55d2026495df67bf [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * fs/eventpoll.c ( Efficent event polling implementation )
Davide Libenzi3419b232006-06-25 05:48:14 -07003 * Copyright (C) 2001,...,2006 Davide Libenzi
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 *
10 * Davide Libenzi <davidel@xmailserver.org>
11 *
12 */
13
14#include <linux/module.h>
15#include <linux/init.h>
16#include <linux/kernel.h>
17#include <linux/sched.h>
18#include <linux/fs.h>
19#include <linux/file.h>
20#include <linux/signal.h>
21#include <linux/errno.h>
22#include <linux/mm.h>
23#include <linux/slab.h>
24#include <linux/poll.h>
25#include <linux/smp_lock.h>
26#include <linux/string.h>
27#include <linux/list.h>
28#include <linux/hash.h>
29#include <linux/spinlock.h>
30#include <linux/syscalls.h>
31#include <linux/rwsem.h>
32#include <linux/rbtree.h>
33#include <linux/wait.h>
34#include <linux/eventpoll.h>
35#include <linux/mount.h>
36#include <linux/bitops.h>
Arjan van de Ven144efe32006-03-23 03:00:32 -080037#include <linux/mutex.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <asm/uaccess.h>
39#include <asm/system.h>
40#include <asm/io.h>
41#include <asm/mman.h>
42#include <asm/atomic.h>
43#include <asm/semaphore.h>
44
45
46/*
47 * LOCKING:
48 * There are three level of locking required by epoll :
49 *
Arjan van de Ven144efe32006-03-23 03:00:32 -080050 * 1) epmutex (mutex)
Linus Torvalds1da177e2005-04-16 15:20:36 -070051 * 2) ep->sem (rw_semaphore)
52 * 3) ep->lock (rw_lock)
53 *
54 * The acquire order is the one listed above, from 1 to 3.
55 * We need a spinlock (ep->lock) because we manipulate objects
56 * from inside the poll callback, that might be triggered from
57 * a wake_up() that in turn might be called from IRQ context.
58 * So we can't sleep inside the poll callback and hence we need
59 * a spinlock. During the event transfer loop (from kernel to
60 * user space) we could end up sleeping due a copy_to_user(), so
61 * we need a lock that will allow us to sleep. This lock is a
62 * read-write semaphore (ep->sem). It is acquired on read during
63 * the event transfer loop and in write during epoll_ctl(EPOLL_CTL_DEL)
64 * and during eventpoll_release_file(). Then we also need a global
65 * semaphore to serialize eventpoll_release_file() and ep_free().
66 * This semaphore is acquired by ep_free() during the epoll file
67 * cleanup path and it is also acquired by eventpoll_release_file()
68 * if a file has been pushed inside an epoll set and it is then
69 * close()d without a previous call toepoll_ctl(EPOLL_CTL_DEL).
70 * It is possible to drop the "ep->sem" and to use the global
Arjan van de Ven144efe32006-03-23 03:00:32 -080071 * semaphore "epmutex" (together with "ep->lock") to have it working,
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 * but having "ep->sem" will make the interface more scalable.
Arjan van de Ven144efe32006-03-23 03:00:32 -080073 * Events that require holding "epmutex" are very rare, while for
Linus Torvalds1da177e2005-04-16 15:20:36 -070074 * normal operations the epoll private "ep->sem" will guarantee
75 * a greater scalability.
76 */
77
78
79#define EVENTPOLLFS_MAGIC 0x03111965 /* My birthday should work for this :) */
80
81#define DEBUG_EPOLL 0
82
83#if DEBUG_EPOLL > 0
84#define DPRINTK(x) printk x
85#define DNPRINTK(n, x) do { if ((n) <= DEBUG_EPOLL) printk x; } while (0)
86#else /* #if DEBUG_EPOLL > 0 */
87#define DPRINTK(x) (void) 0
88#define DNPRINTK(n, x) (void) 0
89#endif /* #if DEBUG_EPOLL > 0 */
90
91#define DEBUG_EPI 0
92
93#if DEBUG_EPI != 0
94#define EPI_SLAB_DEBUG (SLAB_DEBUG_FREE | SLAB_RED_ZONE /* | SLAB_POISON */)
95#else /* #if DEBUG_EPI != 0 */
96#define EPI_SLAB_DEBUG 0
97#endif /* #if DEBUG_EPI != 0 */
98
99/* Epoll private bits inside the event mask */
100#define EP_PRIVATE_BITS (EPOLLONESHOT | EPOLLET)
101
102/* Maximum number of poll wake up nests we are allowing */
103#define EP_MAX_POLLWAKE_NESTS 4
104
Davide Libenzie3306dd2005-09-27 21:45:33 -0700105/* Maximum msec timeout value storeable in a long int */
106#define EP_MAX_MSTIMEO min(1000ULL * MAX_SCHEDULE_TIMEOUT / HZ, (LONG_MAX - 999ULL) / HZ)
107
Davide Libenzib6119672006-10-11 01:21:44 -0700108#define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
109
Davide Libenzie3306dd2005-09-27 21:45:33 -0700110
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111struct epoll_filefd {
112 struct file *file;
113 int fd;
114};
115
116/*
117 * Node that is linked into the "wake_task_list" member of the "struct poll_safewake".
118 * It is used to keep track on all tasks that are currently inside the wake_up() code
119 * to 1) short-circuit the one coming from the same task and same wait queue head
120 * ( loop ) 2) allow a maximum number of epoll descriptors inclusion nesting
121 * 3) let go the ones coming from other tasks.
122 */
123struct wake_task_node {
124 struct list_head llink;
Ingo Molnar36c8b582006-07-03 00:25:41 -0700125 struct task_struct *task;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126 wait_queue_head_t *wq;
127};
128
129/*
130 * This is used to implement the safe poll wake up avoiding to reenter
131 * the poll callback from inside wake_up().
132 */
133struct poll_safewake {
134 struct list_head wake_task_list;
135 spinlock_t lock;
136};
137
138/*
139 * This structure is stored inside the "private_data" member of the file
140 * structure and rapresent the main data sructure for the eventpoll
141 * interface.
142 */
143struct eventpoll {
144 /* Protect the this structure access */
145 rwlock_t lock;
146
147 /*
148 * This semaphore is used to ensure that files are not removed
149 * while epoll is using them. This is read-held during the event
150 * collection loop and it is write-held during the file cleanup
151 * path, the epoll file exit code and the ctl operations.
152 */
153 struct rw_semaphore sem;
154
155 /* Wait queue used by sys_epoll_wait() */
156 wait_queue_head_t wq;
157
158 /* Wait queue used by file->poll() */
159 wait_queue_head_t poll_wait;
160
161 /* List of ready file descriptors */
162 struct list_head rdllist;
163
164 /* RB-Tree root used to store monitored fd structs */
165 struct rb_root rbr;
166};
167
168/* Wait structure used by the poll hooks */
169struct eppoll_entry {
170 /* List header used to link this structure to the "struct epitem" */
171 struct list_head llink;
172
173 /* The "base" pointer is set to the container "struct epitem" */
174 void *base;
175
176 /*
177 * Wait queue item that will be linked to the target file wait
178 * queue head.
179 */
180 wait_queue_t wait;
181
182 /* The wait queue head that linked the "wait" wait queue item */
183 wait_queue_head_t *whead;
184};
185
186/*
187 * Each file descriptor added to the eventpoll interface will
188 * have an entry of this type linked to the hash.
189 */
190struct epitem {
191 /* RB-Tree node used to link this structure to the eventpoll rb-tree */
192 struct rb_node rbn;
193
194 /* List header used to link this structure to the eventpoll ready list */
195 struct list_head rdllink;
196
197 /* The file descriptor information this item refers to */
198 struct epoll_filefd ffd;
199
200 /* Number of active wait queue attached to poll operations */
201 int nwait;
202
203 /* List containing poll wait queues */
204 struct list_head pwqlist;
205
206 /* The "container" of this item */
207 struct eventpoll *ep;
208
209 /* The structure that describe the interested events and the source fd */
210 struct epoll_event event;
211
212 /*
213 * Used to keep track of the usage count of the structure. This avoids
214 * that the structure will desappear from underneath our processing.
215 */
216 atomic_t usecnt;
217
218 /* List header used to link this item to the "struct file" items list */
219 struct list_head fllink;
220
221 /* List header used to link the item to the transfer list */
222 struct list_head txlink;
223
224 /*
225 * This is used during the collection/transfer of events to userspace
226 * to pin items empty events set.
227 */
228 unsigned int revents;
229};
230
231/* Wrapper struct used by poll queueing */
232struct ep_pqueue {
233 poll_table pt;
234 struct epitem *epi;
235};
236
237
238
239static void ep_poll_safewake_init(struct poll_safewake *psw);
240static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq);
Davide Libenzi53d2be72005-09-16 19:28:06 -0700241static int ep_getfd(int *efd, struct inode **einode, struct file **efile,
242 struct eventpoll *ep);
243static int ep_alloc(struct eventpoll **pep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244static void ep_free(struct eventpoll *ep);
245static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd);
246static void ep_use_epitem(struct epitem *epi);
247static void ep_release_epitem(struct epitem *epi);
248static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
249 poll_table *pt);
250static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi);
251static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
252 struct file *tfile, int fd);
253static int ep_modify(struct eventpoll *ep, struct epitem *epi,
254 struct epoll_event *event);
255static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi);
256static int ep_unlink(struct eventpoll *ep, struct epitem *epi);
257static int ep_remove(struct eventpoll *ep, struct epitem *epi);
258static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key);
259static int ep_eventpoll_close(struct inode *inode, struct file *file);
260static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait);
261static int ep_collect_ready_items(struct eventpoll *ep,
262 struct list_head *txlist, int maxevents);
263static int ep_send_events(struct eventpoll *ep, struct list_head *txlist,
264 struct epoll_event __user *events);
265static void ep_reinject_items(struct eventpoll *ep, struct list_head *txlist);
266static int ep_events_transfer(struct eventpoll *ep,
267 struct epoll_event __user *events,
268 int maxevents);
269static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
270 int maxevents, long timeout);
271static int eventpollfs_delete_dentry(struct dentry *dentry);
272static struct inode *ep_eventpoll_inode(void);
David Howells454e2392006-06-23 02:02:57 -0700273static int eventpollfs_get_sb(struct file_system_type *fs_type,
274 int flags, const char *dev_name,
275 void *data, struct vfsmount *mnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276
277/*
278 * This semaphore is used to serialize ep_free() and eventpoll_release_file().
279 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800280static struct mutex epmutex;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281
282/* Safe wake up implementation */
283static struct poll_safewake psw;
284
285/* Slab cache used to allocate "struct epitem" */
Eric Dumazetfa3536c2006-03-26 01:37:24 -0800286static kmem_cache_t *epi_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287
288/* Slab cache used to allocate "struct eppoll_entry" */
Eric Dumazetfa3536c2006-03-26 01:37:24 -0800289static kmem_cache_t *pwq_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290
291/* Virtual fs used to allocate inodes for eventpoll files */
Eric Dumazetfa3536c2006-03-26 01:37:24 -0800292static struct vfsmount *eventpoll_mnt __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293
294/* File callbacks that implement the eventpoll file behaviour */
Arjan van de Ven4b6f5d22006-03-28 01:56:42 -0800295static const struct file_operations eventpoll_fops = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296 .release = ep_eventpoll_close,
297 .poll = ep_eventpoll_poll
298};
299
300/*
301 * This is used to register the virtual file system from where
302 * eventpoll inodes are allocated.
303 */
304static struct file_system_type eventpoll_fs_type = {
305 .name = "eventpollfs",
306 .get_sb = eventpollfs_get_sb,
307 .kill_sb = kill_anon_super,
308};
309
310/* Very basic directory entry operations for the eventpoll virtual file system */
311static struct dentry_operations eventpollfs_dentry_operations = {
312 .d_delete = eventpollfs_delete_dentry,
313};
314
315
316
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700317/* Fast test to see if the file is an evenpoll file */
318static inline int is_file_epoll(struct file *f)
319{
320 return f->f_op == &eventpoll_fops;
321}
322
323/* Setup the structure that is used as key for the rb-tree */
324static inline void ep_set_ffd(struct epoll_filefd *ffd,
325 struct file *file, int fd)
326{
327 ffd->file = file;
328 ffd->fd = fd;
329}
330
331/* Compare rb-tree keys */
332static inline int ep_cmp_ffd(struct epoll_filefd *p1,
333 struct epoll_filefd *p2)
334{
335 return (p1->file > p2->file ? +1:
336 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
337}
338
339/* Special initialization for the rb-tree node to detect linkage */
340static inline void ep_rb_initnode(struct rb_node *n)
341{
David Woodhousec5698822006-04-21 13:17:24 +0100342 rb_set_parent(n, n);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700343}
344
345/* Removes a node from the rb-tree and marks it for a fast is-linked check */
346static inline void ep_rb_erase(struct rb_node *n, struct rb_root *r)
347{
348 rb_erase(n, r);
David Woodhousec5698822006-04-21 13:17:24 +0100349 rb_set_parent(n, n);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700350}
351
352/* Fast check to verify that the item is linked to the main rb-tree */
353static inline int ep_rb_linked(struct rb_node *n)
354{
David Woodhousec5698822006-04-21 13:17:24 +0100355 return rb_parent(n) != n;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700356}
357
358/*
359 * Remove the item from the list and perform its initialization.
360 * This is useful for us because we can test if the item is linked
361 * using "ep_is_linked(p)".
362 */
363static inline void ep_list_del(struct list_head *p)
364{
365 list_del(p);
366 INIT_LIST_HEAD(p);
367}
368
369/* Tells us if the item is currently linked */
370static inline int ep_is_linked(struct list_head *p)
371{
372 return !list_empty(p);
373}
374
375/* Get the "struct epitem" from a wait queue pointer */
376static inline struct epitem * ep_item_from_wait(wait_queue_t *p)
377{
378 return container_of(p, struct eppoll_entry, wait)->base;
379}
380
381/* Get the "struct epitem" from an epoll queue wrapper */
382static inline struct epitem * ep_item_from_epqueue(poll_table *p)
383{
384 return container_of(p, struct ep_pqueue, pt)->epi;
385}
386
387/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
388static inline int ep_op_hash_event(int op)
389{
390 return op != EPOLL_CTL_DEL;
391}
392
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393/* Initialize the poll safe wake up structure */
394static void ep_poll_safewake_init(struct poll_safewake *psw)
395{
396
397 INIT_LIST_HEAD(&psw->wake_task_list);
398 spin_lock_init(&psw->lock);
399}
400
401
402/*
403 * Perform a safe wake up of the poll wait list. The problem is that
404 * with the new callback'd wake up system, it is possible that the
405 * poll callback is reentered from inside the call to wake_up() done
406 * on the poll wait queue head. The rule is that we cannot reenter the
407 * wake up code from the same task more than EP_MAX_POLLWAKE_NESTS times,
408 * and we cannot reenter the same wait queue head at all. This will
409 * enable to have a hierarchy of epoll file descriptor of no more than
410 * EP_MAX_POLLWAKE_NESTS deep. We need the irq version of the spin lock
411 * because this one gets called by the poll callback, that in turn is called
412 * from inside a wake_up(), that might be called from irq context.
413 */
414static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq)
415{
416 int wake_nests = 0;
417 unsigned long flags;
Ingo Molnar36c8b582006-07-03 00:25:41 -0700418 struct task_struct *this_task = current;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419 struct list_head *lsthead = &psw->wake_task_list, *lnk;
420 struct wake_task_node *tncur;
421 struct wake_task_node tnode;
422
423 spin_lock_irqsave(&psw->lock, flags);
424
425 /* Try to see if the current task is already inside this wakeup call */
426 list_for_each(lnk, lsthead) {
427 tncur = list_entry(lnk, struct wake_task_node, llink);
428
429 if (tncur->wq == wq ||
430 (tncur->task == this_task && ++wake_nests > EP_MAX_POLLWAKE_NESTS)) {
431 /*
432 * Ops ... loop detected or maximum nest level reached.
433 * We abort this wake by breaking the cycle itself.
434 */
435 spin_unlock_irqrestore(&psw->lock, flags);
436 return;
437 }
438 }
439
440 /* Add the current task to the list */
441 tnode.task = this_task;
442 tnode.wq = wq;
443 list_add(&tnode.llink, lsthead);
444
445 spin_unlock_irqrestore(&psw->lock, flags);
446
447 /* Do really wake up now */
448 wake_up(wq);
449
450 /* Remove the current task from the list */
451 spin_lock_irqsave(&psw->lock, flags);
452 list_del(&tnode.llink);
453 spin_unlock_irqrestore(&psw->lock, flags);
454}
455
456
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457/*
458 * This is called from eventpoll_release() to unlink files from the eventpoll
459 * interface. We need to have this facility to cleanup correctly files that are
460 * closed without being removed from the eventpoll interface.
461 */
462void eventpoll_release_file(struct file *file)
463{
464 struct list_head *lsthead = &file->f_ep_links;
465 struct eventpoll *ep;
466 struct epitem *epi;
467
468 /*
469 * We don't want to get "file->f_ep_lock" because it is not
470 * necessary. It is not necessary because we're in the "struct file"
471 * cleanup path, and this means that noone is using this file anymore.
472 * The only hit might come from ep_free() but by holding the semaphore
473 * will correctly serialize the operation. We do need to acquire
Arjan van de Ven144efe32006-03-23 03:00:32 -0800474 * "ep->sem" after "epmutex" because ep_remove() requires it when called
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475 * from anywhere but ep_free().
476 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800477 mutex_lock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478
479 while (!list_empty(lsthead)) {
480 epi = list_entry(lsthead->next, struct epitem, fllink);
481
482 ep = epi->ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700483 ep_list_del(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484 down_write(&ep->sem);
485 ep_remove(ep, epi);
486 up_write(&ep->sem);
487 }
488
Arjan van de Ven144efe32006-03-23 03:00:32 -0800489 mutex_unlock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490}
491
492
493/*
494 * It opens an eventpoll file descriptor by suggesting a storage of "size"
495 * file descriptors. The size parameter is just an hint about how to size
496 * data structures. It won't prevent the user to store more than "size"
497 * file descriptors inside the epoll interface. It is the kernel part of
498 * the userspace epoll_create(2).
499 */
500asmlinkage long sys_epoll_create(int size)
501{
Davide Libenzib6119672006-10-11 01:21:44 -0700502 int error, fd = -1;
Davide Libenzi53d2be72005-09-16 19:28:06 -0700503 struct eventpoll *ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504 struct inode *inode;
505 struct file *file;
506
507 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d)\n",
508 current, size));
509
Davide Libenzi53d2be72005-09-16 19:28:06 -0700510 /*
511 * Sanity check on the size parameter, and create the internal data
512 * structure ( "struct eventpoll" ).
513 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514 error = -EINVAL;
Davide Libenzi53d2be72005-09-16 19:28:06 -0700515 if (size <= 0 || (error = ep_alloc(&ep)) != 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 goto eexit_1;
517
518 /*
519 * Creates all the items needed to setup an eventpoll file. That is,
520 * a file structure, and inode and a free file descriptor.
521 */
Davide Libenzi53d2be72005-09-16 19:28:06 -0700522 error = ep_getfd(&fd, &inode, &file, ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523 if (error)
524 goto eexit_2;
525
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
527 current, size, fd));
528
529 return fd;
530
531eexit_2:
Davide Libenzi53d2be72005-09-16 19:28:06 -0700532 ep_free(ep);
533 kfree(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534eexit_1:
535 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
536 current, size, error));
537 return error;
538}
539
540
541/*
542 * The following function implements the controller interface for
543 * the eventpoll file that enables the insertion/removal/change of
544 * file descriptors inside the interest set. It represents
545 * the kernel part of the user space epoll_ctl(2).
546 */
547asmlinkage long
548sys_epoll_ctl(int epfd, int op, int fd, struct epoll_event __user *event)
549{
550 int error;
551 struct file *file, *tfile;
552 struct eventpoll *ep;
553 struct epitem *epi;
554 struct epoll_event epds;
555
556 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p)\n",
557 current, epfd, op, fd, event));
558
559 error = -EFAULT;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700560 if (ep_op_hash_event(op) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561 copy_from_user(&epds, event, sizeof(struct epoll_event)))
562 goto eexit_1;
563
564 /* Get the "struct file *" for the eventpoll file */
565 error = -EBADF;
566 file = fget(epfd);
567 if (!file)
568 goto eexit_1;
569
570 /* Get the "struct file *" for the target file */
571 tfile = fget(fd);
572 if (!tfile)
573 goto eexit_2;
574
575 /* The target file descriptor must support poll */
576 error = -EPERM;
577 if (!tfile->f_op || !tfile->f_op->poll)
578 goto eexit_3;
579
580 /*
581 * We have to check that the file structure underneath the file descriptor
582 * the user passed to us _is_ an eventpoll file. And also we do not permit
583 * adding an epoll file descriptor inside itself.
584 */
585 error = -EINVAL;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700586 if (file == tfile || !is_file_epoll(file))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587 goto eexit_3;
588
589 /*
590 * At this point it is safe to assume that the "private_data" contains
591 * our own data structure.
592 */
593 ep = file->private_data;
594
595 down_write(&ep->sem);
596
597 /* Try to lookup the file inside our hash table */
598 epi = ep_find(ep, tfile, fd);
599
600 error = -EINVAL;
601 switch (op) {
602 case EPOLL_CTL_ADD:
603 if (!epi) {
Davide Libenzi23951402006-04-10 22:54:12 -0700604 epds.events |= POLLERR | POLLHUP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605
606 error = ep_insert(ep, &epds, tfile, fd);
607 } else
608 error = -EEXIST;
609 break;
610 case EPOLL_CTL_DEL:
611 if (epi)
612 error = ep_remove(ep, epi);
613 else
614 error = -ENOENT;
615 break;
616 case EPOLL_CTL_MOD:
617 if (epi) {
Davide Libenzi23951402006-04-10 22:54:12 -0700618 epds.events |= POLLERR | POLLHUP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619 error = ep_modify(ep, epi, &epds);
620 } else
621 error = -ENOENT;
622 break;
623 }
624
625 /*
626 * The function ep_find() increments the usage count of the structure
627 * so, if this is not NULL, we need to release it.
628 */
629 if (epi)
630 ep_release_epitem(epi);
631
632 up_write(&ep->sem);
633
634eexit_3:
635 fput(tfile);
636eexit_2:
637 fput(file);
638eexit_1:
639 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p) = %d\n",
640 current, epfd, op, fd, event, error));
641
642 return error;
643}
644
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645
646/*
647 * Implement the event wait interface for the eventpoll file. It is the kernel
648 * part of the user space epoll_wait(2).
649 */
650asmlinkage long sys_epoll_wait(int epfd, struct epoll_event __user *events,
651 int maxevents, int timeout)
652{
653 int error;
654 struct file *file;
655 struct eventpoll *ep;
656
657 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d)\n",
658 current, epfd, events, maxevents, timeout));
659
660 /* The maximum number of event must be greater than zero */
Davide Libenzib6119672006-10-11 01:21:44 -0700661 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662 return -EINVAL;
663
664 /* Verify that the area passed by the user is writeable */
665 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
666 error = -EFAULT;
667 goto eexit_1;
668 }
669
670 /* Get the "struct file *" for the eventpoll file */
671 error = -EBADF;
672 file = fget(epfd);
673 if (!file)
674 goto eexit_1;
675
676 /*
677 * We have to check that the file structure underneath the fd
678 * the user passed to us _is_ an eventpoll file.
679 */
680 error = -EINVAL;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700681 if (!is_file_epoll(file))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682 goto eexit_2;
683
684 /*
685 * At this point it is safe to assume that the "private_data" contains
686 * our own data structure.
687 */
688 ep = file->private_data;
689
690 /* Time to fish for events ... */
691 error = ep_poll(ep, events, maxevents, timeout);
692
693eexit_2:
694 fput(file);
695eexit_1:
696 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d) = %d\n",
697 current, epfd, events, maxevents, timeout, error));
698
699 return error;
700}
701
702
Davide Libenzib6119672006-10-11 01:21:44 -0700703#ifdef TIF_RESTORE_SIGMASK
704
705/*
706 * Implement the event wait interface for the eventpoll file. It is the kernel
707 * part of the user space epoll_pwait(2).
708 */
709asmlinkage long sys_epoll_pwait(int epfd, struct epoll_event __user *events,
710 int maxevents, int timeout, const sigset_t __user *sigmask,
711 size_t sigsetsize)
712{
713 int error;
714 sigset_t ksigmask, sigsaved;
715
716 /*
717 * If the caller wants a certain signal mask to be set during the wait,
718 * we apply it here.
719 */
720 if (sigmask) {
721 if (sigsetsize != sizeof(sigset_t))
722 return -EINVAL;
723 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
724 return -EFAULT;
725 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
726 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
727 }
728
729 error = sys_epoll_wait(epfd, events, maxevents, timeout);
730
731 /*
732 * If we changed the signal mask, we need to restore the original one.
733 * In case we've got a signal while waiting, we do not restore the
734 * signal mask yet, and we allow do_signal() to deliver the signal on
735 * the way back to userspace, before the signal mask is restored.
736 */
737 if (sigmask) {
738 if (error == -EINTR) {
739 memcpy(&current->saved_sigmask, &sigsaved,
740 sizeof(sigsaved));
741 set_thread_flag(TIF_RESTORE_SIGMASK);
742 } else
743 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
744 }
745
746 return error;
747}
748
749#endif /* #ifdef TIF_RESTORE_SIGMASK */
750
751
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752/*
753 * Creates the file descriptor to be used by the epoll interface.
754 */
Davide Libenzi53d2be72005-09-16 19:28:06 -0700755static int ep_getfd(int *efd, struct inode **einode, struct file **efile,
756 struct eventpoll *ep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757{
758 struct qstr this;
759 char name[32];
760 struct dentry *dentry;
761 struct inode *inode;
762 struct file *file;
763 int error, fd;
764
765 /* Get an ready to use file */
766 error = -ENFILE;
767 file = get_empty_filp();
768 if (!file)
769 goto eexit_1;
770
771 /* Allocates an inode from the eventpoll file system */
772 inode = ep_eventpoll_inode();
Jeff Garzikc3b65712006-10-03 01:13:52 -0700773 if (IS_ERR(inode)) {
774 error = PTR_ERR(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775 goto eexit_2;
Jeff Garzikc3b65712006-10-03 01:13:52 -0700776 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777
778 /* Allocates a free descriptor to plug the file onto */
779 error = get_unused_fd();
780 if (error < 0)
781 goto eexit_3;
782 fd = error;
783
784 /*
785 * Link the inode to a directory entry by creating a unique name
786 * using the inode number.
787 */
788 error = -ENOMEM;
789 sprintf(name, "[%lu]", inode->i_ino);
790 this.name = name;
791 this.len = strlen(name);
792 this.hash = inode->i_ino;
793 dentry = d_alloc(eventpoll_mnt->mnt_sb->s_root, &this);
794 if (!dentry)
795 goto eexit_4;
796 dentry->d_op = &eventpollfs_dentry_operations;
797 d_add(dentry, inode);
798 file->f_vfsmnt = mntget(eventpoll_mnt);
799 file->f_dentry = dentry;
800 file->f_mapping = inode->i_mapping;
801
802 file->f_pos = 0;
803 file->f_flags = O_RDONLY;
804 file->f_op = &eventpoll_fops;
805 file->f_mode = FMODE_READ;
806 file->f_version = 0;
Davide Libenzi53d2be72005-09-16 19:28:06 -0700807 file->private_data = ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808
809 /* Install the new setup file into the allocated fd. */
810 fd_install(fd, file);
811
812 *efd = fd;
813 *einode = inode;
814 *efile = file;
815 return 0;
816
817eexit_4:
818 put_unused_fd(fd);
819eexit_3:
820 iput(inode);
821eexit_2:
822 put_filp(file);
823eexit_1:
824 return error;
825}
826
827
Davide Libenzi53d2be72005-09-16 19:28:06 -0700828static int ep_alloc(struct eventpoll **pep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829{
Davide Libenzi53d2be72005-09-16 19:28:06 -0700830 struct eventpoll *ep = kzalloc(sizeof(*ep), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831
Davide Libenzi53d2be72005-09-16 19:28:06 -0700832 if (!ep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833 return -ENOMEM;
834
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835 rwlock_init(&ep->lock);
836 init_rwsem(&ep->sem);
837 init_waitqueue_head(&ep->wq);
838 init_waitqueue_head(&ep->poll_wait);
839 INIT_LIST_HEAD(&ep->rdllist);
840 ep->rbr = RB_ROOT;
841
Davide Libenzi53d2be72005-09-16 19:28:06 -0700842 *pep = ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843
Davide Libenzi53d2be72005-09-16 19:28:06 -0700844 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_alloc() ep=%p\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845 current, ep));
846 return 0;
847}
848
849
850static void ep_free(struct eventpoll *ep)
851{
852 struct rb_node *rbp;
853 struct epitem *epi;
854
855 /* We need to release all tasks waiting for these file */
856 if (waitqueue_active(&ep->poll_wait))
857 ep_poll_safewake(&psw, &ep->poll_wait);
858
859 /*
860 * We need to lock this because we could be hit by
861 * eventpoll_release_file() while we're freeing the "struct eventpoll".
862 * We do not need to hold "ep->sem" here because the epoll file
863 * is on the way to be removed and no one has references to it
864 * anymore. The only hit might come from eventpoll_release_file() but
Arjan van de Ven144efe32006-03-23 03:00:32 -0800865 * holding "epmutex" is sufficent here.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800867 mutex_lock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700868
869 /*
870 * Walks through the whole tree by unregistering poll callbacks.
871 */
872 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
873 epi = rb_entry(rbp, struct epitem, rbn);
874
875 ep_unregister_pollwait(ep, epi);
876 }
877
878 /*
879 * Walks through the whole hash by freeing each "struct epitem". At this
880 * point we are sure no poll callbacks will be lingering around, and also by
881 * write-holding "sem" we can be sure that no file cleanup code will hit
882 * us during this operation. So we can avoid the lock on "ep->lock".
883 */
884 while ((rbp = rb_first(&ep->rbr)) != 0) {
885 epi = rb_entry(rbp, struct epitem, rbn);
886 ep_remove(ep, epi);
887 }
888
Arjan van de Ven144efe32006-03-23 03:00:32 -0800889 mutex_unlock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890}
891
892
893/*
894 * Search the file inside the eventpoll hash. It add usage count to
895 * the returned item, so the caller must call ep_release_epitem()
896 * after finished using the "struct epitem".
897 */
898static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
899{
900 int kcmp;
901 unsigned long flags;
902 struct rb_node *rbp;
903 struct epitem *epi, *epir = NULL;
904 struct epoll_filefd ffd;
905
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700906 ep_set_ffd(&ffd, file, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700907 read_lock_irqsave(&ep->lock, flags);
908 for (rbp = ep->rbr.rb_node; rbp; ) {
909 epi = rb_entry(rbp, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700910 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 if (kcmp > 0)
912 rbp = rbp->rb_right;
913 else if (kcmp < 0)
914 rbp = rbp->rb_left;
915 else {
916 ep_use_epitem(epi);
917 epir = epi;
918 break;
919 }
920 }
921 read_unlock_irqrestore(&ep->lock, flags);
922
923 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_find(%p) -> %p\n",
924 current, file, epir));
925
926 return epir;
927}
928
929
930/*
931 * Increment the usage count of the "struct epitem" making it sure
932 * that the user will have a valid pointer to reference.
933 */
934static void ep_use_epitem(struct epitem *epi)
935{
936
937 atomic_inc(&epi->usecnt);
938}
939
940
941/*
942 * Decrement ( release ) the usage count by signaling that the user
943 * has finished using the structure. It might lead to freeing the
944 * structure itself if the count goes to zero.
945 */
946static void ep_release_epitem(struct epitem *epi)
947{
948
949 if (atomic_dec_and_test(&epi->usecnt))
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700950 kmem_cache_free(epi_cache, epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951}
952
953
954/*
955 * This is the callback that is used to add our wait queue to the
956 * target file wakeup lists.
957 */
958static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
959 poll_table *pt)
960{
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700961 struct epitem *epi = ep_item_from_epqueue(pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962 struct eppoll_entry *pwq;
963
Christoph Lametere94b1762006-12-06 20:33:17 -0800964 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
966 pwq->whead = whead;
967 pwq->base = epi;
968 add_wait_queue(whead, &pwq->wait);
969 list_add_tail(&pwq->llink, &epi->pwqlist);
970 epi->nwait++;
971 } else {
972 /* We have to signal that an error occurred */
973 epi->nwait = -1;
974 }
975}
976
977
978static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
979{
980 int kcmp;
981 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
982 struct epitem *epic;
983
984 while (*p) {
985 parent = *p;
986 epic = rb_entry(parent, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700987 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700988 if (kcmp > 0)
989 p = &parent->rb_right;
990 else
991 p = &parent->rb_left;
992 }
993 rb_link_node(&epi->rbn, parent, p);
994 rb_insert_color(&epi->rbn, &ep->rbr);
995}
996
997
998static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
999 struct file *tfile, int fd)
1000{
1001 int error, revents, pwake = 0;
1002 unsigned long flags;
1003 struct epitem *epi;
1004 struct ep_pqueue epq;
1005
1006 error = -ENOMEM;
Christoph Lametere94b1762006-12-06 20:33:17 -08001007 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001008 goto eexit_1;
1009
1010 /* Item initialization follow here ... */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001011 ep_rb_initnode(&epi->rbn);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012 INIT_LIST_HEAD(&epi->rdllink);
1013 INIT_LIST_HEAD(&epi->fllink);
1014 INIT_LIST_HEAD(&epi->txlink);
1015 INIT_LIST_HEAD(&epi->pwqlist);
1016 epi->ep = ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001017 ep_set_ffd(&epi->ffd, tfile, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001018 epi->event = *event;
1019 atomic_set(&epi->usecnt, 1);
1020 epi->nwait = 0;
1021
1022 /* Initialize the poll table using the queue callback */
1023 epq.epi = epi;
1024 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1025
1026 /*
1027 * Attach the item to the poll hooks and get current event bits.
1028 * We can safely use the file* here because its usage count has
1029 * been increased by the caller of this function.
1030 */
1031 revents = tfile->f_op->poll(tfile, &epq.pt);
1032
1033 /*
1034 * We have to check if something went wrong during the poll wait queue
1035 * install process. Namely an allocation for a wait queue failed due
1036 * high memory pressure.
1037 */
1038 if (epi->nwait < 0)
1039 goto eexit_2;
1040
1041 /* Add the current item to the list of active epoll hook for this file */
1042 spin_lock(&tfile->f_ep_lock);
1043 list_add_tail(&epi->fllink, &tfile->f_ep_links);
1044 spin_unlock(&tfile->f_ep_lock);
1045
1046 /* We have to drop the new item inside our item list to keep track of it */
1047 write_lock_irqsave(&ep->lock, flags);
1048
1049 /* Add the current item to the rb-tree */
1050 ep_rbtree_insert(ep, epi);
1051
1052 /* If the file is already "ready" we drop it inside the ready list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001053 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001054 list_add_tail(&epi->rdllink, &ep->rdllist);
1055
1056 /* Notify waiting tasks that events are available */
1057 if (waitqueue_active(&ep->wq))
Davide Libenzi3419b232006-06-25 05:48:14 -07001058 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059 if (waitqueue_active(&ep->poll_wait))
1060 pwake++;
1061 }
1062
1063 write_unlock_irqrestore(&ep->lock, flags);
1064
1065 /* We have to call this outside the lock */
1066 if (pwake)
1067 ep_poll_safewake(&psw, &ep->poll_wait);
1068
1069 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_insert(%p, %p, %d)\n",
1070 current, ep, tfile, fd));
1071
1072 return 0;
1073
1074eexit_2:
1075 ep_unregister_pollwait(ep, epi);
1076
1077 /*
1078 * We need to do this because an event could have been arrived on some
1079 * allocated wait queue.
1080 */
1081 write_lock_irqsave(&ep->lock, flags);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001082 if (ep_is_linked(&epi->rdllink))
1083 ep_list_del(&epi->rdllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084 write_unlock_irqrestore(&ep->lock, flags);
1085
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001086 kmem_cache_free(epi_cache, epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087eexit_1:
1088 return error;
1089}
1090
1091
1092/*
1093 * Modify the interest event mask by dropping an event if the new mask
1094 * has a match in the current file status.
1095 */
1096static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1097{
1098 int pwake = 0;
1099 unsigned int revents;
1100 unsigned long flags;
1101
1102 /*
1103 * Set the new event interest mask before calling f_op->poll(), otherwise
1104 * a potential race might occur. In fact if we do this operation inside
1105 * the lock, an event might happen between the f_op->poll() call and the
1106 * new event set registering.
1107 */
1108 epi->event.events = event->events;
1109
1110 /*
1111 * Get current event bits. We can safely use the file* here because
1112 * its usage count has been increased by the caller of this function.
1113 */
1114 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
1115
1116 write_lock_irqsave(&ep->lock, flags);
1117
1118 /* Copy the data member from inside the lock */
1119 epi->event.data = event->data;
1120
1121 /*
1122 * If the item is not linked to the hash it means that it's on its
1123 * way toward the removal. Do nothing in this case.
1124 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001125 if (ep_rb_linked(&epi->rbn)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 /*
1127 * If the item is "hot" and it is not registered inside the ready
1128 * list, push it inside. If the item is not "hot" and it is currently
1129 * registered inside the ready list, unlink it.
1130 */
1131 if (revents & event->events) {
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001132 if (!ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 list_add_tail(&epi->rdllink, &ep->rdllist);
1134
1135 /* Notify waiting tasks that events are available */
1136 if (waitqueue_active(&ep->wq))
Davide Libenzi3419b232006-06-25 05:48:14 -07001137 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
1138 TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 if (waitqueue_active(&ep->poll_wait))
1140 pwake++;
1141 }
1142 }
1143 }
1144
1145 write_unlock_irqrestore(&ep->lock, flags);
1146
1147 /* We have to call this outside the lock */
1148 if (pwake)
1149 ep_poll_safewake(&psw, &ep->poll_wait);
1150
1151 return 0;
1152}
1153
1154
1155/*
1156 * This function unregister poll callbacks from the associated file descriptor.
1157 * Since this must be called without holding "ep->lock" the atomic exchange trick
1158 * will protect us from multiple unregister.
1159 */
1160static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
1161{
1162 int nwait;
1163 struct list_head *lsthead = &epi->pwqlist;
1164 struct eppoll_entry *pwq;
1165
1166 /* This is called without locks, so we need the atomic exchange */
1167 nwait = xchg(&epi->nwait, 0);
1168
1169 if (nwait) {
1170 while (!list_empty(lsthead)) {
1171 pwq = list_entry(lsthead->next, struct eppoll_entry, llink);
1172
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001173 ep_list_del(&pwq->llink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174 remove_wait_queue(pwq->whead, &pwq->wait);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001175 kmem_cache_free(pwq_cache, pwq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001176 }
1177 }
1178}
1179
1180
1181/*
1182 * Unlink the "struct epitem" from all places it might have been hooked up.
1183 * This function must be called with write IRQ lock on "ep->lock".
1184 */
1185static int ep_unlink(struct eventpoll *ep, struct epitem *epi)
1186{
1187 int error;
1188
1189 /*
1190 * It can happen that this one is called for an item already unlinked.
1191 * The check protect us from doing a double unlink ( crash ).
1192 */
1193 error = -ENOENT;
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001194 if (!ep_rb_linked(&epi->rbn))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195 goto eexit_1;
1196
1197 /*
1198 * Clear the event mask for the unlinked item. This will avoid item
1199 * notifications to be sent after the unlink operation from inside
1200 * the kernel->userspace event transfer loop.
1201 */
1202 epi->event.events = 0;
1203
1204 /*
1205 * At this point is safe to do the job, unlink the item from our rb-tree.
1206 * This operation togheter with the above check closes the door to
1207 * double unlinks.
1208 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001209 ep_rb_erase(&epi->rbn, &ep->rbr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001210
1211 /*
1212 * If the item we are going to remove is inside the ready file descriptors
1213 * we want to remove it from this list to avoid stale events.
1214 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001215 if (ep_is_linked(&epi->rdllink))
1216 ep_list_del(&epi->rdllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217
1218 error = 0;
1219eexit_1:
1220
1221 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_unlink(%p, %p) = %d\n",
Masoud Asgharifard Sharbiani45f17e02006-08-27 01:23:48 -07001222 current, ep, epi->ffd.file, error));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223
1224 return error;
1225}
1226
1227
1228/*
1229 * Removes a "struct epitem" from the eventpoll hash and deallocates
1230 * all the associated resources.
1231 */
1232static int ep_remove(struct eventpoll *ep, struct epitem *epi)
1233{
1234 int error;
1235 unsigned long flags;
1236 struct file *file = epi->ffd.file;
1237
1238 /*
1239 * Removes poll wait queue hooks. We _have_ to do this without holding
1240 * the "ep->lock" otherwise a deadlock might occur. This because of the
1241 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
1242 * queue head lock when unregistering the wait queue. The wakeup callback
1243 * will run by holding the wait queue head lock and will call our callback
1244 * that will try to get "ep->lock".
1245 */
1246 ep_unregister_pollwait(ep, epi);
1247
1248 /* Remove the current item from the list of epoll hooks */
1249 spin_lock(&file->f_ep_lock);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001250 if (ep_is_linked(&epi->fllink))
1251 ep_list_del(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252 spin_unlock(&file->f_ep_lock);
1253
1254 /* We need to acquire the write IRQ lock before calling ep_unlink() */
1255 write_lock_irqsave(&ep->lock, flags);
1256
1257 /* Really unlink the item from the hash */
1258 error = ep_unlink(ep, epi);
1259
1260 write_unlock_irqrestore(&ep->lock, flags);
1261
1262 if (error)
1263 goto eexit_1;
1264
1265 /* At this point it is safe to free the eventpoll item */
1266 ep_release_epitem(epi);
1267
1268 error = 0;
1269eexit_1:
1270 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_remove(%p, %p) = %d\n",
1271 current, ep, file, error));
1272
1273 return error;
1274}
1275
1276
1277/*
1278 * This is the callback that is passed to the wait queue wakeup
1279 * machanism. It is called by the stored file descriptors when they
1280 * have events to report.
1281 */
1282static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
1283{
1284 int pwake = 0;
1285 unsigned long flags;
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001286 struct epitem *epi = ep_item_from_wait(wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 struct eventpoll *ep = epi->ep;
1288
1289 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: poll_callback(%p) epi=%p ep=%p\n",
Masoud Asgharifard Sharbiani45f17e02006-08-27 01:23:48 -07001290 current, epi->ffd.file, epi, ep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291
1292 write_lock_irqsave(&ep->lock, flags);
1293
1294 /*
1295 * If the event mask does not contain any poll(2) event, we consider the
1296 * descriptor to be disabled. This condition is likely the effect of the
1297 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1298 * until the next EPOLL_CTL_MOD will be issued.
1299 */
1300 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1301 goto is_disabled;
1302
1303 /* If this file is already in the ready list we exit soon */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001304 if (ep_is_linked(&epi->rdllink))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 goto is_linked;
1306
1307 list_add_tail(&epi->rdllink, &ep->rdllist);
1308
1309is_linked:
1310 /*
1311 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1312 * wait list.
1313 */
1314 if (waitqueue_active(&ep->wq))
Davide Libenzi3419b232006-06-25 05:48:14 -07001315 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
1316 TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317 if (waitqueue_active(&ep->poll_wait))
1318 pwake++;
1319
1320is_disabled:
1321 write_unlock_irqrestore(&ep->lock, flags);
1322
1323 /* We have to call this outside the lock */
1324 if (pwake)
1325 ep_poll_safewake(&psw, &ep->poll_wait);
1326
1327 return 1;
1328}
1329
1330
1331static int ep_eventpoll_close(struct inode *inode, struct file *file)
1332{
1333 struct eventpoll *ep = file->private_data;
1334
1335 if (ep) {
1336 ep_free(ep);
1337 kfree(ep);
1338 }
1339
1340 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: close() ep=%p\n", current, ep));
1341 return 0;
1342}
1343
1344
1345static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
1346{
1347 unsigned int pollflags = 0;
1348 unsigned long flags;
1349 struct eventpoll *ep = file->private_data;
1350
1351 /* Insert inside our poll wait queue */
1352 poll_wait(file, &ep->poll_wait, wait);
1353
1354 /* Check our condition */
1355 read_lock_irqsave(&ep->lock, flags);
1356 if (!list_empty(&ep->rdllist))
1357 pollflags = POLLIN | POLLRDNORM;
1358 read_unlock_irqrestore(&ep->lock, flags);
1359
1360 return pollflags;
1361}
1362
1363
1364/*
1365 * Since we have to release the lock during the __copy_to_user() operation and
1366 * during the f_op->poll() call, we try to collect the maximum number of items
1367 * by reducing the irqlock/irqunlock switching rate.
1368 */
1369static int ep_collect_ready_items(struct eventpoll *ep, struct list_head *txlist, int maxevents)
1370{
1371 int nepi;
1372 unsigned long flags;
1373 struct list_head *lsthead = &ep->rdllist, *lnk;
1374 struct epitem *epi;
1375
1376 write_lock_irqsave(&ep->lock, flags);
1377
1378 for (nepi = 0, lnk = lsthead->next; lnk != lsthead && nepi < maxevents;) {
1379 epi = list_entry(lnk, struct epitem, rdllink);
1380
1381 lnk = lnk->next;
1382
1383 /* If this file is already in the ready list we exit soon */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001384 if (!ep_is_linked(&epi->txlink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385 /*
1386 * This is initialized in this way so that the default
1387 * behaviour of the reinjecting code will be to push back
1388 * the item inside the ready list.
1389 */
1390 epi->revents = epi->event.events;
1391
1392 /* Link the ready item into the transfer list */
1393 list_add(&epi->txlink, txlist);
1394 nepi++;
1395
1396 /*
1397 * Unlink the item from the ready list.
1398 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001399 ep_list_del(&epi->rdllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001400 }
1401 }
1402
1403 write_unlock_irqrestore(&ep->lock, flags);
1404
1405 return nepi;
1406}
1407
1408
1409/*
1410 * This function is called without holding the "ep->lock" since the call to
1411 * __copy_to_user() might sleep, and also f_op->poll() might reenable the IRQ
1412 * because of the way poll() is traditionally implemented in Linux.
1413 */
1414static int ep_send_events(struct eventpoll *ep, struct list_head *txlist,
1415 struct epoll_event __user *events)
1416{
1417 int eventcnt = 0;
1418 unsigned int revents;
1419 struct list_head *lnk;
1420 struct epitem *epi;
1421
1422 /*
1423 * We can loop without lock because this is a task private list.
1424 * The test done during the collection loop will guarantee us that
1425 * another task will not try to collect this file. Also, items
1426 * cannot vanish during the loop because we are holding "sem".
1427 */
1428 list_for_each(lnk, txlist) {
1429 epi = list_entry(lnk, struct epitem, txlink);
1430
1431 /*
1432 * Get the ready file event set. We can safely use the file
1433 * because we are holding the "sem" in read and this will
1434 * guarantee that both the file and the item will not vanish.
1435 */
1436 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
1437
1438 /*
1439 * Set the return event set for the current file descriptor.
1440 * Note that only the task task was successfully able to link
1441 * the item to its "txlist" will write this field.
1442 */
1443 epi->revents = revents & epi->event.events;
1444
1445 if (epi->revents) {
1446 if (__put_user(epi->revents,
1447 &events[eventcnt].events) ||
1448 __put_user(epi->event.data,
1449 &events[eventcnt].data))
1450 return -EFAULT;
1451 if (epi->event.events & EPOLLONESHOT)
1452 epi->event.events &= EP_PRIVATE_BITS;
1453 eventcnt++;
1454 }
1455 }
1456 return eventcnt;
1457}
1458
1459
1460/*
1461 * Walk through the transfer list we collected with ep_collect_ready_items()
1462 * and, if 1) the item is still "alive" 2) its event set is not empty 3) it's
1463 * not already linked, links it to the ready list. Same as above, we are holding
1464 * "sem" so items cannot vanish underneath our nose.
1465 */
1466static void ep_reinject_items(struct eventpoll *ep, struct list_head *txlist)
1467{
1468 int ricnt = 0, pwake = 0;
1469 unsigned long flags;
1470 struct epitem *epi;
1471
1472 write_lock_irqsave(&ep->lock, flags);
1473
1474 while (!list_empty(txlist)) {
1475 epi = list_entry(txlist->next, struct epitem, txlink);
1476
1477 /* Unlink the current item from the transfer list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001478 ep_list_del(&epi->txlink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479
1480 /*
1481 * If the item is no more linked to the interest set, we don't
1482 * have to push it inside the ready list because the following
1483 * ep_release_epitem() is going to drop it. Also, if the current
1484 * item is set to have an Edge Triggered behaviour, we don't have
1485 * to push it back either.
1486 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001487 if (ep_rb_linked(&epi->rbn) && !(epi->event.events & EPOLLET) &&
1488 (epi->revents & epi->event.events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489 list_add_tail(&epi->rdllink, &ep->rdllist);
1490 ricnt++;
1491 }
1492 }
1493
1494 if (ricnt) {
1495 /*
1496 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1497 * wait list.
1498 */
1499 if (waitqueue_active(&ep->wq))
Davide Libenzi3419b232006-06-25 05:48:14 -07001500 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
1501 TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502 if (waitqueue_active(&ep->poll_wait))
1503 pwake++;
1504 }
1505
1506 write_unlock_irqrestore(&ep->lock, flags);
1507
1508 /* We have to call this outside the lock */
1509 if (pwake)
1510 ep_poll_safewake(&psw, &ep->poll_wait);
1511}
1512
1513
1514/*
1515 * Perform the transfer of events to user space.
1516 */
1517static int ep_events_transfer(struct eventpoll *ep,
1518 struct epoll_event __user *events, int maxevents)
1519{
1520 int eventcnt = 0;
1521 struct list_head txlist;
1522
1523 INIT_LIST_HEAD(&txlist);
1524
1525 /*
1526 * We need to lock this because we could be hit by
1527 * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL).
1528 */
1529 down_read(&ep->sem);
1530
1531 /* Collect/extract ready items */
1532 if (ep_collect_ready_items(ep, &txlist, maxevents) > 0) {
1533 /* Build result set in userspace */
1534 eventcnt = ep_send_events(ep, &txlist, events);
1535
1536 /* Reinject ready items into the ready list */
1537 ep_reinject_items(ep, &txlist);
1538 }
1539
1540 up_read(&ep->sem);
1541
1542 return eventcnt;
1543}
1544
1545
1546static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1547 int maxevents, long timeout)
1548{
1549 int res, eavail;
1550 unsigned long flags;
1551 long jtimeout;
1552 wait_queue_t wait;
1553
1554 /*
1555 * Calculate the timeout by checking for the "infinite" value ( -1 )
1556 * and the overflow condition. The passed timeout is in milliseconds,
1557 * that why (t * HZ) / 1000.
1558 */
Davide Libenzie3306dd2005-09-27 21:45:33 -07001559 jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
1560 MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561
1562retry:
1563 write_lock_irqsave(&ep->lock, flags);
1564
1565 res = 0;
1566 if (list_empty(&ep->rdllist)) {
1567 /*
1568 * We don't have any available event to return to the caller.
1569 * We need to sleep here, and we will be wake up by
1570 * ep_poll_callback() when events will become available.
1571 */
1572 init_waitqueue_entry(&wait, current);
Davide Libenzi3419b232006-06-25 05:48:14 -07001573 __add_wait_queue(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001574
1575 for (;;) {
1576 /*
1577 * We don't want to sleep if the ep_poll_callback() sends us
1578 * a wakeup in between. That's why we set the task state
1579 * to TASK_INTERRUPTIBLE before doing the checks.
1580 */
1581 set_current_state(TASK_INTERRUPTIBLE);
1582 if (!list_empty(&ep->rdllist) || !jtimeout)
1583 break;
1584 if (signal_pending(current)) {
1585 res = -EINTR;
1586 break;
1587 }
1588
1589 write_unlock_irqrestore(&ep->lock, flags);
1590 jtimeout = schedule_timeout(jtimeout);
1591 write_lock_irqsave(&ep->lock, flags);
1592 }
Davide Libenzi3419b232006-06-25 05:48:14 -07001593 __remove_wait_queue(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594
1595 set_current_state(TASK_RUNNING);
1596 }
1597
1598 /* Is it worth to try to dig for events ? */
1599 eavail = !list_empty(&ep->rdllist);
1600
1601 write_unlock_irqrestore(&ep->lock, flags);
1602
1603 /*
1604 * Try to transfer events to user space. In case we get 0 events and
1605 * there's still timeout left over, we go trying again in search of
1606 * more luck.
1607 */
1608 if (!res && eavail &&
1609 !(res = ep_events_transfer(ep, events, maxevents)) && jtimeout)
1610 goto retry;
1611
1612 return res;
1613}
1614
1615
1616static int eventpollfs_delete_dentry(struct dentry *dentry)
1617{
1618
1619 return 1;
1620}
1621
1622
1623static struct inode *ep_eventpoll_inode(void)
1624{
1625 int error = -ENOMEM;
1626 struct inode *inode = new_inode(eventpoll_mnt->mnt_sb);
1627
1628 if (!inode)
1629 goto eexit_1;
1630
1631 inode->i_fop = &eventpoll_fops;
1632
1633 /*
1634 * Mark the inode dirty from the very beginning,
1635 * that way it will never be moved to the dirty
1636 * list because mark_inode_dirty() will think
1637 * that it already _is_ on the dirty list.
1638 */
1639 inode->i_state = I_DIRTY;
1640 inode->i_mode = S_IRUSR | S_IWUSR;
1641 inode->i_uid = current->fsuid;
1642 inode->i_gid = current->fsgid;
1643 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644 return inode;
1645
1646eexit_1:
1647 return ERR_PTR(error);
1648}
1649
1650
David Howells454e2392006-06-23 02:02:57 -07001651static int
Linus Torvalds1da177e2005-04-16 15:20:36 -07001652eventpollfs_get_sb(struct file_system_type *fs_type, int flags,
David Howells454e2392006-06-23 02:02:57 -07001653 const char *dev_name, void *data, struct vfsmount *mnt)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654{
David Howells454e2392006-06-23 02:02:57 -07001655 return get_sb_pseudo(fs_type, "eventpoll:", NULL, EVENTPOLLFS_MAGIC,
1656 mnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657}
1658
1659
1660static int __init eventpoll_init(void)
1661{
1662 int error;
1663
Arjan van de Ven144efe32006-03-23 03:00:32 -08001664 mutex_init(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001665
1666 /* Initialize the structure used to perform safe poll wait head wake ups */
1667 ep_poll_safewake_init(&psw);
1668
1669 /* Allocates slab cache used to allocate "struct epitem" items */
1670 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1671 0, SLAB_HWCACHE_ALIGN|EPI_SLAB_DEBUG|SLAB_PANIC,
1672 NULL, NULL);
1673
1674 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1675 pwq_cache = kmem_cache_create("eventpoll_pwq",
1676 sizeof(struct eppoll_entry), 0,
1677 EPI_SLAB_DEBUG|SLAB_PANIC, NULL, NULL);
1678
1679 /*
1680 * Register the virtual file system that will be the source of inodes
1681 * for the eventpoll files
1682 */
1683 error = register_filesystem(&eventpoll_fs_type);
1684 if (error)
1685 goto epanic;
1686
1687 /* Mount the above commented virtual file system */
1688 eventpoll_mnt = kern_mount(&eventpoll_fs_type);
1689 error = PTR_ERR(eventpoll_mnt);
1690 if (IS_ERR(eventpoll_mnt))
1691 goto epanic;
1692
1693 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: successfully initialized.\n",
1694 current));
1695 return 0;
1696
1697epanic:
1698 panic("eventpoll_init() failed\n");
1699}
1700
1701
1702static void __exit eventpoll_exit(void)
1703{
1704 /* Undo all operations done inside eventpoll_init() */
1705 unregister_filesystem(&eventpoll_fs_type);
1706 mntput(eventpoll_mnt);
1707 kmem_cache_destroy(pwq_cache);
1708 kmem_cache_destroy(epi_cache);
1709}
1710
1711module_init(eventpoll_init);
1712module_exit(eventpoll_exit);
1713
1714MODULE_LICENSE("GPL");