blob: 19ffb043abbccfb7fc564d114acabeaae8984a60 [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
108
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109struct epoll_filefd {
110 struct file *file;
111 int fd;
112};
113
114/*
115 * Node that is linked into the "wake_task_list" member of the "struct poll_safewake".
116 * It is used to keep track on all tasks that are currently inside the wake_up() code
117 * to 1) short-circuit the one coming from the same task and same wait queue head
118 * ( loop ) 2) allow a maximum number of epoll descriptors inclusion nesting
119 * 3) let go the ones coming from other tasks.
120 */
121struct wake_task_node {
122 struct list_head llink;
Ingo Molnar36c8b582006-07-03 00:25:41 -0700123 struct task_struct *task;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124 wait_queue_head_t *wq;
125};
126
127/*
128 * This is used to implement the safe poll wake up avoiding to reenter
129 * the poll callback from inside wake_up().
130 */
131struct poll_safewake {
132 struct list_head wake_task_list;
133 spinlock_t lock;
134};
135
136/*
137 * This structure is stored inside the "private_data" member of the file
138 * structure and rapresent the main data sructure for the eventpoll
139 * interface.
140 */
141struct eventpoll {
142 /* Protect the this structure access */
143 rwlock_t lock;
144
145 /*
146 * This semaphore is used to ensure that files are not removed
147 * while epoll is using them. This is read-held during the event
148 * collection loop and it is write-held during the file cleanup
149 * path, the epoll file exit code and the ctl operations.
150 */
151 struct rw_semaphore sem;
152
153 /* Wait queue used by sys_epoll_wait() */
154 wait_queue_head_t wq;
155
156 /* Wait queue used by file->poll() */
157 wait_queue_head_t poll_wait;
158
159 /* List of ready file descriptors */
160 struct list_head rdllist;
161
162 /* RB-Tree root used to store monitored fd structs */
163 struct rb_root rbr;
164};
165
166/* Wait structure used by the poll hooks */
167struct eppoll_entry {
168 /* List header used to link this structure to the "struct epitem" */
169 struct list_head llink;
170
171 /* The "base" pointer is set to the container "struct epitem" */
172 void *base;
173
174 /*
175 * Wait queue item that will be linked to the target file wait
176 * queue head.
177 */
178 wait_queue_t wait;
179
180 /* The wait queue head that linked the "wait" wait queue item */
181 wait_queue_head_t *whead;
182};
183
184/*
185 * Each file descriptor added to the eventpoll interface will
186 * have an entry of this type linked to the hash.
187 */
188struct epitem {
189 /* RB-Tree node used to link this structure to the eventpoll rb-tree */
190 struct rb_node rbn;
191
192 /* List header used to link this structure to the eventpoll ready list */
193 struct list_head rdllink;
194
195 /* The file descriptor information this item refers to */
196 struct epoll_filefd ffd;
197
198 /* Number of active wait queue attached to poll operations */
199 int nwait;
200
201 /* List containing poll wait queues */
202 struct list_head pwqlist;
203
204 /* The "container" of this item */
205 struct eventpoll *ep;
206
207 /* The structure that describe the interested events and the source fd */
208 struct epoll_event event;
209
210 /*
211 * Used to keep track of the usage count of the structure. This avoids
212 * that the structure will desappear from underneath our processing.
213 */
214 atomic_t usecnt;
215
216 /* List header used to link this item to the "struct file" items list */
217 struct list_head fllink;
218
219 /* List header used to link the item to the transfer list */
220 struct list_head txlink;
221
222 /*
223 * This is used during the collection/transfer of events to userspace
224 * to pin items empty events set.
225 */
226 unsigned int revents;
227};
228
229/* Wrapper struct used by poll queueing */
230struct ep_pqueue {
231 poll_table pt;
232 struct epitem *epi;
233};
234
235
236
237static void ep_poll_safewake_init(struct poll_safewake *psw);
238static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq);
Davide Libenzi53d2be72005-09-16 19:28:06 -0700239static int ep_getfd(int *efd, struct inode **einode, struct file **efile,
240 struct eventpoll *ep);
241static int ep_alloc(struct eventpoll **pep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242static void ep_free(struct eventpoll *ep);
243static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd);
244static void ep_use_epitem(struct epitem *epi);
245static void ep_release_epitem(struct epitem *epi);
246static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
247 poll_table *pt);
248static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi);
249static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
250 struct file *tfile, int fd);
251static int ep_modify(struct eventpoll *ep, struct epitem *epi,
252 struct epoll_event *event);
253static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi);
254static int ep_unlink(struct eventpoll *ep, struct epitem *epi);
255static int ep_remove(struct eventpoll *ep, struct epitem *epi);
256static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key);
257static int ep_eventpoll_close(struct inode *inode, struct file *file);
258static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait);
259static int ep_collect_ready_items(struct eventpoll *ep,
260 struct list_head *txlist, int maxevents);
261static int ep_send_events(struct eventpoll *ep, struct list_head *txlist,
262 struct epoll_event __user *events);
263static void ep_reinject_items(struct eventpoll *ep, struct list_head *txlist);
264static int ep_events_transfer(struct eventpoll *ep,
265 struct epoll_event __user *events,
266 int maxevents);
267static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
268 int maxevents, long timeout);
269static int eventpollfs_delete_dentry(struct dentry *dentry);
270static struct inode *ep_eventpoll_inode(void);
David Howells454e2392006-06-23 02:02:57 -0700271static int eventpollfs_get_sb(struct file_system_type *fs_type,
272 int flags, const char *dev_name,
273 void *data, struct vfsmount *mnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274
275/*
276 * This semaphore is used to serialize ep_free() and eventpoll_release_file().
277 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800278static struct mutex epmutex;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279
280/* Safe wake up implementation */
281static struct poll_safewake psw;
282
283/* Slab cache used to allocate "struct epitem" */
Eric Dumazetfa3536c2006-03-26 01:37:24 -0800284static kmem_cache_t *epi_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285
286/* Slab cache used to allocate "struct eppoll_entry" */
Eric Dumazetfa3536c2006-03-26 01:37:24 -0800287static kmem_cache_t *pwq_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288
289/* Virtual fs used to allocate inodes for eventpoll files */
Eric Dumazetfa3536c2006-03-26 01:37:24 -0800290static struct vfsmount *eventpoll_mnt __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291
292/* File callbacks that implement the eventpoll file behaviour */
Arjan van de Ven4b6f5d22006-03-28 01:56:42 -0800293static const struct file_operations eventpoll_fops = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294 .release = ep_eventpoll_close,
295 .poll = ep_eventpoll_poll
296};
297
298/*
299 * This is used to register the virtual file system from where
300 * eventpoll inodes are allocated.
301 */
302static struct file_system_type eventpoll_fs_type = {
303 .name = "eventpollfs",
304 .get_sb = eventpollfs_get_sb,
305 .kill_sb = kill_anon_super,
306};
307
308/* Very basic directory entry operations for the eventpoll virtual file system */
309static struct dentry_operations eventpollfs_dentry_operations = {
310 .d_delete = eventpollfs_delete_dentry,
311};
312
313
314
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700315/* Fast test to see if the file is an evenpoll file */
316static inline int is_file_epoll(struct file *f)
317{
318 return f->f_op == &eventpoll_fops;
319}
320
321/* Setup the structure that is used as key for the rb-tree */
322static inline void ep_set_ffd(struct epoll_filefd *ffd,
323 struct file *file, int fd)
324{
325 ffd->file = file;
326 ffd->fd = fd;
327}
328
329/* Compare rb-tree keys */
330static inline int ep_cmp_ffd(struct epoll_filefd *p1,
331 struct epoll_filefd *p2)
332{
333 return (p1->file > p2->file ? +1:
334 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
335}
336
337/* Special initialization for the rb-tree node to detect linkage */
338static inline void ep_rb_initnode(struct rb_node *n)
339{
David Woodhousec5698822006-04-21 13:17:24 +0100340 rb_set_parent(n, n);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700341}
342
343/* Removes a node from the rb-tree and marks it for a fast is-linked check */
344static inline void ep_rb_erase(struct rb_node *n, struct rb_root *r)
345{
346 rb_erase(n, r);
David Woodhousec5698822006-04-21 13:17:24 +0100347 rb_set_parent(n, n);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700348}
349
350/* Fast check to verify that the item is linked to the main rb-tree */
351static inline int ep_rb_linked(struct rb_node *n)
352{
David Woodhousec5698822006-04-21 13:17:24 +0100353 return rb_parent(n) != n;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700354}
355
356/*
357 * Remove the item from the list and perform its initialization.
358 * This is useful for us because we can test if the item is linked
359 * using "ep_is_linked(p)".
360 */
361static inline void ep_list_del(struct list_head *p)
362{
363 list_del(p);
364 INIT_LIST_HEAD(p);
365}
366
367/* Tells us if the item is currently linked */
368static inline int ep_is_linked(struct list_head *p)
369{
370 return !list_empty(p);
371}
372
373/* Get the "struct epitem" from a wait queue pointer */
374static inline struct epitem * ep_item_from_wait(wait_queue_t *p)
375{
376 return container_of(p, struct eppoll_entry, wait)->base;
377}
378
379/* Get the "struct epitem" from an epoll queue wrapper */
380static inline struct epitem * ep_item_from_epqueue(poll_table *p)
381{
382 return container_of(p, struct ep_pqueue, pt)->epi;
383}
384
385/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
386static inline int ep_op_hash_event(int op)
387{
388 return op != EPOLL_CTL_DEL;
389}
390
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391/* Initialize the poll safe wake up structure */
392static void ep_poll_safewake_init(struct poll_safewake *psw)
393{
394
395 INIT_LIST_HEAD(&psw->wake_task_list);
396 spin_lock_init(&psw->lock);
397}
398
399
400/*
401 * Perform a safe wake up of the poll wait list. The problem is that
402 * with the new callback'd wake up system, it is possible that the
403 * poll callback is reentered from inside the call to wake_up() done
404 * on the poll wait queue head. The rule is that we cannot reenter the
405 * wake up code from the same task more than EP_MAX_POLLWAKE_NESTS times,
406 * and we cannot reenter the same wait queue head at all. This will
407 * enable to have a hierarchy of epoll file descriptor of no more than
408 * EP_MAX_POLLWAKE_NESTS deep. We need the irq version of the spin lock
409 * because this one gets called by the poll callback, that in turn is called
410 * from inside a wake_up(), that might be called from irq context.
411 */
412static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq)
413{
414 int wake_nests = 0;
415 unsigned long flags;
Ingo Molnar36c8b582006-07-03 00:25:41 -0700416 struct task_struct *this_task = current;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417 struct list_head *lsthead = &psw->wake_task_list, *lnk;
418 struct wake_task_node *tncur;
419 struct wake_task_node tnode;
420
421 spin_lock_irqsave(&psw->lock, flags);
422
423 /* Try to see if the current task is already inside this wakeup call */
424 list_for_each(lnk, lsthead) {
425 tncur = list_entry(lnk, struct wake_task_node, llink);
426
427 if (tncur->wq == wq ||
428 (tncur->task == this_task && ++wake_nests > EP_MAX_POLLWAKE_NESTS)) {
429 /*
430 * Ops ... loop detected or maximum nest level reached.
431 * We abort this wake by breaking the cycle itself.
432 */
433 spin_unlock_irqrestore(&psw->lock, flags);
434 return;
435 }
436 }
437
438 /* Add the current task to the list */
439 tnode.task = this_task;
440 tnode.wq = wq;
441 list_add(&tnode.llink, lsthead);
442
443 spin_unlock_irqrestore(&psw->lock, flags);
444
445 /* Do really wake up now */
446 wake_up(wq);
447
448 /* Remove the current task from the list */
449 spin_lock_irqsave(&psw->lock, flags);
450 list_del(&tnode.llink);
451 spin_unlock_irqrestore(&psw->lock, flags);
452}
453
454
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455/*
456 * This is called from eventpoll_release() to unlink files from the eventpoll
457 * interface. We need to have this facility to cleanup correctly files that are
458 * closed without being removed from the eventpoll interface.
459 */
460void eventpoll_release_file(struct file *file)
461{
462 struct list_head *lsthead = &file->f_ep_links;
463 struct eventpoll *ep;
464 struct epitem *epi;
465
466 /*
467 * We don't want to get "file->f_ep_lock" because it is not
468 * necessary. It is not necessary because we're in the "struct file"
469 * cleanup path, and this means that noone is using this file anymore.
470 * The only hit might come from ep_free() but by holding the semaphore
471 * will correctly serialize the operation. We do need to acquire
Arjan van de Ven144efe32006-03-23 03:00:32 -0800472 * "ep->sem" after "epmutex" because ep_remove() requires it when called
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473 * from anywhere but ep_free().
474 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800475 mutex_lock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476
477 while (!list_empty(lsthead)) {
478 epi = list_entry(lsthead->next, struct epitem, fllink);
479
480 ep = epi->ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700481 ep_list_del(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482 down_write(&ep->sem);
483 ep_remove(ep, epi);
484 up_write(&ep->sem);
485 }
486
Arjan van de Ven144efe32006-03-23 03:00:32 -0800487 mutex_unlock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488}
489
490
491/*
492 * It opens an eventpoll file descriptor by suggesting a storage of "size"
493 * file descriptors. The size parameter is just an hint about how to size
494 * data structures. It won't prevent the user to store more than "size"
495 * file descriptors inside the epoll interface. It is the kernel part of
496 * the userspace epoll_create(2).
497 */
498asmlinkage long sys_epoll_create(int size)
499{
500 int error, fd;
Davide Libenzi53d2be72005-09-16 19:28:06 -0700501 struct eventpoll *ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 struct inode *inode;
503 struct file *file;
504
505 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d)\n",
506 current, size));
507
Davide Libenzi53d2be72005-09-16 19:28:06 -0700508 /*
509 * Sanity check on the size parameter, and create the internal data
510 * structure ( "struct eventpoll" ).
511 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 error = -EINVAL;
Davide Libenzi53d2be72005-09-16 19:28:06 -0700513 if (size <= 0 || (error = ep_alloc(&ep)) != 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514 goto eexit_1;
515
516 /*
517 * Creates all the items needed to setup an eventpoll file. That is,
518 * a file structure, and inode and a free file descriptor.
519 */
Davide Libenzi53d2be72005-09-16 19:28:06 -0700520 error = ep_getfd(&fd, &inode, &file, ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 if (error)
522 goto eexit_2;
523
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
525 current, size, fd));
526
527 return fd;
528
529eexit_2:
Davide Libenzi53d2be72005-09-16 19:28:06 -0700530 ep_free(ep);
531 kfree(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532eexit_1:
533 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
534 current, size, error));
535 return error;
536}
537
538
539/*
540 * The following function implements the controller interface for
541 * the eventpoll file that enables the insertion/removal/change of
542 * file descriptors inside the interest set. It represents
543 * the kernel part of the user space epoll_ctl(2).
544 */
545asmlinkage long
546sys_epoll_ctl(int epfd, int op, int fd, struct epoll_event __user *event)
547{
548 int error;
549 struct file *file, *tfile;
550 struct eventpoll *ep;
551 struct epitem *epi;
552 struct epoll_event epds;
553
554 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p)\n",
555 current, epfd, op, fd, event));
556
557 error = -EFAULT;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700558 if (ep_op_hash_event(op) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559 copy_from_user(&epds, event, sizeof(struct epoll_event)))
560 goto eexit_1;
561
562 /* Get the "struct file *" for the eventpoll file */
563 error = -EBADF;
564 file = fget(epfd);
565 if (!file)
566 goto eexit_1;
567
568 /* Get the "struct file *" for the target file */
569 tfile = fget(fd);
570 if (!tfile)
571 goto eexit_2;
572
573 /* The target file descriptor must support poll */
574 error = -EPERM;
575 if (!tfile->f_op || !tfile->f_op->poll)
576 goto eexit_3;
577
578 /*
579 * We have to check that the file structure underneath the file descriptor
580 * the user passed to us _is_ an eventpoll file. And also we do not permit
581 * adding an epoll file descriptor inside itself.
582 */
583 error = -EINVAL;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700584 if (file == tfile || !is_file_epoll(file))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 goto eexit_3;
586
587 /*
588 * At this point it is safe to assume that the "private_data" contains
589 * our own data structure.
590 */
591 ep = file->private_data;
592
593 down_write(&ep->sem);
594
595 /* Try to lookup the file inside our hash table */
596 epi = ep_find(ep, tfile, fd);
597
598 error = -EINVAL;
599 switch (op) {
600 case EPOLL_CTL_ADD:
601 if (!epi) {
Davide Libenzi23951402006-04-10 22:54:12 -0700602 epds.events |= POLLERR | POLLHUP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603
604 error = ep_insert(ep, &epds, tfile, fd);
605 } else
606 error = -EEXIST;
607 break;
608 case EPOLL_CTL_DEL:
609 if (epi)
610 error = ep_remove(ep, epi);
611 else
612 error = -ENOENT;
613 break;
614 case EPOLL_CTL_MOD:
615 if (epi) {
Davide Libenzi23951402006-04-10 22:54:12 -0700616 epds.events |= POLLERR | POLLHUP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617 error = ep_modify(ep, epi, &epds);
618 } else
619 error = -ENOENT;
620 break;
621 }
622
623 /*
624 * The function ep_find() increments the usage count of the structure
625 * so, if this is not NULL, we need to release it.
626 */
627 if (epi)
628 ep_release_epitem(epi);
629
630 up_write(&ep->sem);
631
632eexit_3:
633 fput(tfile);
634eexit_2:
635 fput(file);
636eexit_1:
637 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p) = %d\n",
638 current, epfd, op, fd, event, error));
639
640 return error;
641}
642
643#define MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
644
645/*
646 * Implement the event wait interface for the eventpoll file. It is the kernel
647 * part of the user space epoll_wait(2).
648 */
649asmlinkage long sys_epoll_wait(int epfd, struct epoll_event __user *events,
650 int maxevents, int timeout)
651{
652 int error;
653 struct file *file;
654 struct eventpoll *ep;
655
656 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d)\n",
657 current, epfd, events, maxevents, timeout));
658
659 /* The maximum number of event must be greater than zero */
660 if (maxevents <= 0 || maxevents > MAX_EVENTS)
661 return -EINVAL;
662
663 /* Verify that the area passed by the user is writeable */
664 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
665 error = -EFAULT;
666 goto eexit_1;
667 }
668
669 /* Get the "struct file *" for the eventpoll file */
670 error = -EBADF;
671 file = fget(epfd);
672 if (!file)
673 goto eexit_1;
674
675 /*
676 * We have to check that the file structure underneath the fd
677 * the user passed to us _is_ an eventpoll file.
678 */
679 error = -EINVAL;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700680 if (!is_file_epoll(file))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681 goto eexit_2;
682
683 /*
684 * At this point it is safe to assume that the "private_data" contains
685 * our own data structure.
686 */
687 ep = file->private_data;
688
689 /* Time to fish for events ... */
690 error = ep_poll(ep, events, maxevents, timeout);
691
692eexit_2:
693 fput(file);
694eexit_1:
695 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d) = %d\n",
696 current, epfd, events, maxevents, timeout, error));
697
698 return error;
699}
700
701
702/*
703 * Creates the file descriptor to be used by the epoll interface.
704 */
Davide Libenzi53d2be72005-09-16 19:28:06 -0700705static int ep_getfd(int *efd, struct inode **einode, struct file **efile,
706 struct eventpoll *ep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707{
708 struct qstr this;
709 char name[32];
710 struct dentry *dentry;
711 struct inode *inode;
712 struct file *file;
713 int error, fd;
714
715 /* Get an ready to use file */
716 error = -ENFILE;
717 file = get_empty_filp();
718 if (!file)
719 goto eexit_1;
720
721 /* Allocates an inode from the eventpoll file system */
722 inode = ep_eventpoll_inode();
723 error = PTR_ERR(inode);
724 if (IS_ERR(inode))
725 goto eexit_2;
726
727 /* Allocates a free descriptor to plug the file onto */
728 error = get_unused_fd();
729 if (error < 0)
730 goto eexit_3;
731 fd = error;
732
733 /*
734 * Link the inode to a directory entry by creating a unique name
735 * using the inode number.
736 */
737 error = -ENOMEM;
738 sprintf(name, "[%lu]", inode->i_ino);
739 this.name = name;
740 this.len = strlen(name);
741 this.hash = inode->i_ino;
742 dentry = d_alloc(eventpoll_mnt->mnt_sb->s_root, &this);
743 if (!dentry)
744 goto eexit_4;
745 dentry->d_op = &eventpollfs_dentry_operations;
746 d_add(dentry, inode);
747 file->f_vfsmnt = mntget(eventpoll_mnt);
748 file->f_dentry = dentry;
749 file->f_mapping = inode->i_mapping;
750
751 file->f_pos = 0;
752 file->f_flags = O_RDONLY;
753 file->f_op = &eventpoll_fops;
754 file->f_mode = FMODE_READ;
755 file->f_version = 0;
Davide Libenzi53d2be72005-09-16 19:28:06 -0700756 file->private_data = ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757
758 /* Install the new setup file into the allocated fd. */
759 fd_install(fd, file);
760
761 *efd = fd;
762 *einode = inode;
763 *efile = file;
764 return 0;
765
766eexit_4:
767 put_unused_fd(fd);
768eexit_3:
769 iput(inode);
770eexit_2:
771 put_filp(file);
772eexit_1:
773 return error;
774}
775
776
Davide Libenzi53d2be72005-09-16 19:28:06 -0700777static int ep_alloc(struct eventpoll **pep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778{
Davide Libenzi53d2be72005-09-16 19:28:06 -0700779 struct eventpoll *ep = kzalloc(sizeof(*ep), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780
Davide Libenzi53d2be72005-09-16 19:28:06 -0700781 if (!ep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782 return -ENOMEM;
783
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 rwlock_init(&ep->lock);
785 init_rwsem(&ep->sem);
786 init_waitqueue_head(&ep->wq);
787 init_waitqueue_head(&ep->poll_wait);
788 INIT_LIST_HEAD(&ep->rdllist);
789 ep->rbr = RB_ROOT;
790
Davide Libenzi53d2be72005-09-16 19:28:06 -0700791 *pep = ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792
Davide Libenzi53d2be72005-09-16 19:28:06 -0700793 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_alloc() ep=%p\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 current, ep));
795 return 0;
796}
797
798
799static void ep_free(struct eventpoll *ep)
800{
801 struct rb_node *rbp;
802 struct epitem *epi;
803
804 /* We need to release all tasks waiting for these file */
805 if (waitqueue_active(&ep->poll_wait))
806 ep_poll_safewake(&psw, &ep->poll_wait);
807
808 /*
809 * We need to lock this because we could be hit by
810 * eventpoll_release_file() while we're freeing the "struct eventpoll".
811 * We do not need to hold "ep->sem" here because the epoll file
812 * is on the way to be removed and no one has references to it
813 * anymore. The only hit might come from eventpoll_release_file() but
Arjan van de Ven144efe32006-03-23 03:00:32 -0800814 * holding "epmutex" is sufficent here.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800816 mutex_lock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817
818 /*
819 * Walks through the whole tree by unregistering poll callbacks.
820 */
821 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
822 epi = rb_entry(rbp, struct epitem, rbn);
823
824 ep_unregister_pollwait(ep, epi);
825 }
826
827 /*
828 * Walks through the whole hash by freeing each "struct epitem". At this
829 * point we are sure no poll callbacks will be lingering around, and also by
830 * write-holding "sem" we can be sure that no file cleanup code will hit
831 * us during this operation. So we can avoid the lock on "ep->lock".
832 */
833 while ((rbp = rb_first(&ep->rbr)) != 0) {
834 epi = rb_entry(rbp, struct epitem, rbn);
835 ep_remove(ep, epi);
836 }
837
Arjan van de Ven144efe32006-03-23 03:00:32 -0800838 mutex_unlock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839}
840
841
842/*
843 * Search the file inside the eventpoll hash. It add usage count to
844 * the returned item, so the caller must call ep_release_epitem()
845 * after finished using the "struct epitem".
846 */
847static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
848{
849 int kcmp;
850 unsigned long flags;
851 struct rb_node *rbp;
852 struct epitem *epi, *epir = NULL;
853 struct epoll_filefd ffd;
854
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700855 ep_set_ffd(&ffd, file, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856 read_lock_irqsave(&ep->lock, flags);
857 for (rbp = ep->rbr.rb_node; rbp; ) {
858 epi = rb_entry(rbp, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700859 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700860 if (kcmp > 0)
861 rbp = rbp->rb_right;
862 else if (kcmp < 0)
863 rbp = rbp->rb_left;
864 else {
865 ep_use_epitem(epi);
866 epir = epi;
867 break;
868 }
869 }
870 read_unlock_irqrestore(&ep->lock, flags);
871
872 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_find(%p) -> %p\n",
873 current, file, epir));
874
875 return epir;
876}
877
878
879/*
880 * Increment the usage count of the "struct epitem" making it sure
881 * that the user will have a valid pointer to reference.
882 */
883static void ep_use_epitem(struct epitem *epi)
884{
885
886 atomic_inc(&epi->usecnt);
887}
888
889
890/*
891 * Decrement ( release ) the usage count by signaling that the user
892 * has finished using the structure. It might lead to freeing the
893 * structure itself if the count goes to zero.
894 */
895static void ep_release_epitem(struct epitem *epi)
896{
897
898 if (atomic_dec_and_test(&epi->usecnt))
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700899 kmem_cache_free(epi_cache, epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900}
901
902
903/*
904 * This is the callback that is used to add our wait queue to the
905 * target file wakeup lists.
906 */
907static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
908 poll_table *pt)
909{
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700910 struct epitem *epi = ep_item_from_epqueue(pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 struct eppoll_entry *pwq;
912
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700913 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, SLAB_KERNEL))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
915 pwq->whead = whead;
916 pwq->base = epi;
917 add_wait_queue(whead, &pwq->wait);
918 list_add_tail(&pwq->llink, &epi->pwqlist);
919 epi->nwait++;
920 } else {
921 /* We have to signal that an error occurred */
922 epi->nwait = -1;
923 }
924}
925
926
927static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
928{
929 int kcmp;
930 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
931 struct epitem *epic;
932
933 while (*p) {
934 parent = *p;
935 epic = rb_entry(parent, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700936 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937 if (kcmp > 0)
938 p = &parent->rb_right;
939 else
940 p = &parent->rb_left;
941 }
942 rb_link_node(&epi->rbn, parent, p);
943 rb_insert_color(&epi->rbn, &ep->rbr);
944}
945
946
947static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
948 struct file *tfile, int fd)
949{
950 int error, revents, pwake = 0;
951 unsigned long flags;
952 struct epitem *epi;
953 struct ep_pqueue epq;
954
955 error = -ENOMEM;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700956 if (!(epi = kmem_cache_alloc(epi_cache, SLAB_KERNEL)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957 goto eexit_1;
958
959 /* Item initialization follow here ... */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700960 ep_rb_initnode(&epi->rbn);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961 INIT_LIST_HEAD(&epi->rdllink);
962 INIT_LIST_HEAD(&epi->fllink);
963 INIT_LIST_HEAD(&epi->txlink);
964 INIT_LIST_HEAD(&epi->pwqlist);
965 epi->ep = ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700966 ep_set_ffd(&epi->ffd, tfile, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967 epi->event = *event;
968 atomic_set(&epi->usecnt, 1);
969 epi->nwait = 0;
970
971 /* Initialize the poll table using the queue callback */
972 epq.epi = epi;
973 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
974
975 /*
976 * Attach the item to the poll hooks and get current event bits.
977 * We can safely use the file* here because its usage count has
978 * been increased by the caller of this function.
979 */
980 revents = tfile->f_op->poll(tfile, &epq.pt);
981
982 /*
983 * We have to check if something went wrong during the poll wait queue
984 * install process. Namely an allocation for a wait queue failed due
985 * high memory pressure.
986 */
987 if (epi->nwait < 0)
988 goto eexit_2;
989
990 /* Add the current item to the list of active epoll hook for this file */
991 spin_lock(&tfile->f_ep_lock);
992 list_add_tail(&epi->fllink, &tfile->f_ep_links);
993 spin_unlock(&tfile->f_ep_lock);
994
995 /* We have to drop the new item inside our item list to keep track of it */
996 write_lock_irqsave(&ep->lock, flags);
997
998 /* Add the current item to the rb-tree */
999 ep_rbtree_insert(ep, epi);
1000
1001 /* If the file is already "ready" we drop it inside the ready list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001002 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001003 list_add_tail(&epi->rdllink, &ep->rdllist);
1004
1005 /* Notify waiting tasks that events are available */
1006 if (waitqueue_active(&ep->wq))
Davide Libenzi3419b232006-06-25 05:48:14 -07001007 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001008 if (waitqueue_active(&ep->poll_wait))
1009 pwake++;
1010 }
1011
1012 write_unlock_irqrestore(&ep->lock, flags);
1013
1014 /* We have to call this outside the lock */
1015 if (pwake)
1016 ep_poll_safewake(&psw, &ep->poll_wait);
1017
1018 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_insert(%p, %p, %d)\n",
1019 current, ep, tfile, fd));
1020
1021 return 0;
1022
1023eexit_2:
1024 ep_unregister_pollwait(ep, epi);
1025
1026 /*
1027 * We need to do this because an event could have been arrived on some
1028 * allocated wait queue.
1029 */
1030 write_lock_irqsave(&ep->lock, flags);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001031 if (ep_is_linked(&epi->rdllink))
1032 ep_list_del(&epi->rdllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001033 write_unlock_irqrestore(&ep->lock, flags);
1034
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001035 kmem_cache_free(epi_cache, epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036eexit_1:
1037 return error;
1038}
1039
1040
1041/*
1042 * Modify the interest event mask by dropping an event if the new mask
1043 * has a match in the current file status.
1044 */
1045static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1046{
1047 int pwake = 0;
1048 unsigned int revents;
1049 unsigned long flags;
1050
1051 /*
1052 * Set the new event interest mask before calling f_op->poll(), otherwise
1053 * a potential race might occur. In fact if we do this operation inside
1054 * the lock, an event might happen between the f_op->poll() call and the
1055 * new event set registering.
1056 */
1057 epi->event.events = event->events;
1058
1059 /*
1060 * Get current event bits. We can safely use the file* here because
1061 * its usage count has been increased by the caller of this function.
1062 */
1063 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
1064
1065 write_lock_irqsave(&ep->lock, flags);
1066
1067 /* Copy the data member from inside the lock */
1068 epi->event.data = event->data;
1069
1070 /*
1071 * If the item is not linked to the hash it means that it's on its
1072 * way toward the removal. Do nothing in this case.
1073 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001074 if (ep_rb_linked(&epi->rbn)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075 /*
1076 * If the item is "hot" and it is not registered inside the ready
1077 * list, push it inside. If the item is not "hot" and it is currently
1078 * registered inside the ready list, unlink it.
1079 */
1080 if (revents & event->events) {
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001081 if (!ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001082 list_add_tail(&epi->rdllink, &ep->rdllist);
1083
1084 /* Notify waiting tasks that events are available */
1085 if (waitqueue_active(&ep->wq))
Davide Libenzi3419b232006-06-25 05:48:14 -07001086 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
1087 TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088 if (waitqueue_active(&ep->poll_wait))
1089 pwake++;
1090 }
1091 }
1092 }
1093
1094 write_unlock_irqrestore(&ep->lock, flags);
1095
1096 /* We have to call this outside the lock */
1097 if (pwake)
1098 ep_poll_safewake(&psw, &ep->poll_wait);
1099
1100 return 0;
1101}
1102
1103
1104/*
1105 * This function unregister poll callbacks from the associated file descriptor.
1106 * Since this must be called without holding "ep->lock" the atomic exchange trick
1107 * will protect us from multiple unregister.
1108 */
1109static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
1110{
1111 int nwait;
1112 struct list_head *lsthead = &epi->pwqlist;
1113 struct eppoll_entry *pwq;
1114
1115 /* This is called without locks, so we need the atomic exchange */
1116 nwait = xchg(&epi->nwait, 0);
1117
1118 if (nwait) {
1119 while (!list_empty(lsthead)) {
1120 pwq = list_entry(lsthead->next, struct eppoll_entry, llink);
1121
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001122 ep_list_del(&pwq->llink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123 remove_wait_queue(pwq->whead, &pwq->wait);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001124 kmem_cache_free(pwq_cache, pwq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 }
1126 }
1127}
1128
1129
1130/*
1131 * Unlink the "struct epitem" from all places it might have been hooked up.
1132 * This function must be called with write IRQ lock on "ep->lock".
1133 */
1134static int ep_unlink(struct eventpoll *ep, struct epitem *epi)
1135{
1136 int error;
1137
1138 /*
1139 * It can happen that this one is called for an item already unlinked.
1140 * The check protect us from doing a double unlink ( crash ).
1141 */
1142 error = -ENOENT;
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001143 if (!ep_rb_linked(&epi->rbn))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001144 goto eexit_1;
1145
1146 /*
1147 * Clear the event mask for the unlinked item. This will avoid item
1148 * notifications to be sent after the unlink operation from inside
1149 * the kernel->userspace event transfer loop.
1150 */
1151 epi->event.events = 0;
1152
1153 /*
1154 * At this point is safe to do the job, unlink the item from our rb-tree.
1155 * This operation togheter with the above check closes the door to
1156 * double unlinks.
1157 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001158 ep_rb_erase(&epi->rbn, &ep->rbr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159
1160 /*
1161 * If the item we are going to remove is inside the ready file descriptors
1162 * we want to remove it from this list to avoid stale events.
1163 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001164 if (ep_is_linked(&epi->rdllink))
1165 ep_list_del(&epi->rdllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166
1167 error = 0;
1168eexit_1:
1169
1170 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_unlink(%p, %p) = %d\n",
1171 current, ep, epi->file, error));
1172
1173 return error;
1174}
1175
1176
1177/*
1178 * Removes a "struct epitem" from the eventpoll hash and deallocates
1179 * all the associated resources.
1180 */
1181static int ep_remove(struct eventpoll *ep, struct epitem *epi)
1182{
1183 int error;
1184 unsigned long flags;
1185 struct file *file = epi->ffd.file;
1186
1187 /*
1188 * Removes poll wait queue hooks. We _have_ to do this without holding
1189 * the "ep->lock" otherwise a deadlock might occur. This because of the
1190 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
1191 * queue head lock when unregistering the wait queue. The wakeup callback
1192 * will run by holding the wait queue head lock and will call our callback
1193 * that will try to get "ep->lock".
1194 */
1195 ep_unregister_pollwait(ep, epi);
1196
1197 /* Remove the current item from the list of epoll hooks */
1198 spin_lock(&file->f_ep_lock);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001199 if (ep_is_linked(&epi->fllink))
1200 ep_list_del(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001201 spin_unlock(&file->f_ep_lock);
1202
1203 /* We need to acquire the write IRQ lock before calling ep_unlink() */
1204 write_lock_irqsave(&ep->lock, flags);
1205
1206 /* Really unlink the item from the hash */
1207 error = ep_unlink(ep, epi);
1208
1209 write_unlock_irqrestore(&ep->lock, flags);
1210
1211 if (error)
1212 goto eexit_1;
1213
1214 /* At this point it is safe to free the eventpoll item */
1215 ep_release_epitem(epi);
1216
1217 error = 0;
1218eexit_1:
1219 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_remove(%p, %p) = %d\n",
1220 current, ep, file, error));
1221
1222 return error;
1223}
1224
1225
1226/*
1227 * This is the callback that is passed to the wait queue wakeup
1228 * machanism. It is called by the stored file descriptors when they
1229 * have events to report.
1230 */
1231static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
1232{
1233 int pwake = 0;
1234 unsigned long flags;
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001235 struct epitem *epi = ep_item_from_wait(wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001236 struct eventpoll *ep = epi->ep;
1237
1238 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: poll_callback(%p) epi=%p ep=%p\n",
1239 current, epi->file, epi, ep));
1240
1241 write_lock_irqsave(&ep->lock, flags);
1242
1243 /*
1244 * If the event mask does not contain any poll(2) event, we consider the
1245 * descriptor to be disabled. This condition is likely the effect of the
1246 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1247 * until the next EPOLL_CTL_MOD will be issued.
1248 */
1249 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1250 goto is_disabled;
1251
1252 /* If this file is already in the ready list we exit soon */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001253 if (ep_is_linked(&epi->rdllink))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254 goto is_linked;
1255
1256 list_add_tail(&epi->rdllink, &ep->rdllist);
1257
1258is_linked:
1259 /*
1260 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1261 * wait list.
1262 */
1263 if (waitqueue_active(&ep->wq))
Davide Libenzi3419b232006-06-25 05:48:14 -07001264 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
1265 TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266 if (waitqueue_active(&ep->poll_wait))
1267 pwake++;
1268
1269is_disabled:
1270 write_unlock_irqrestore(&ep->lock, flags);
1271
1272 /* We have to call this outside the lock */
1273 if (pwake)
1274 ep_poll_safewake(&psw, &ep->poll_wait);
1275
1276 return 1;
1277}
1278
1279
1280static int ep_eventpoll_close(struct inode *inode, struct file *file)
1281{
1282 struct eventpoll *ep = file->private_data;
1283
1284 if (ep) {
1285 ep_free(ep);
1286 kfree(ep);
1287 }
1288
1289 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: close() ep=%p\n", current, ep));
1290 return 0;
1291}
1292
1293
1294static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
1295{
1296 unsigned int pollflags = 0;
1297 unsigned long flags;
1298 struct eventpoll *ep = file->private_data;
1299
1300 /* Insert inside our poll wait queue */
1301 poll_wait(file, &ep->poll_wait, wait);
1302
1303 /* Check our condition */
1304 read_lock_irqsave(&ep->lock, flags);
1305 if (!list_empty(&ep->rdllist))
1306 pollflags = POLLIN | POLLRDNORM;
1307 read_unlock_irqrestore(&ep->lock, flags);
1308
1309 return pollflags;
1310}
1311
1312
1313/*
1314 * Since we have to release the lock during the __copy_to_user() operation and
1315 * during the f_op->poll() call, we try to collect the maximum number of items
1316 * by reducing the irqlock/irqunlock switching rate.
1317 */
1318static int ep_collect_ready_items(struct eventpoll *ep, struct list_head *txlist, int maxevents)
1319{
1320 int nepi;
1321 unsigned long flags;
1322 struct list_head *lsthead = &ep->rdllist, *lnk;
1323 struct epitem *epi;
1324
1325 write_lock_irqsave(&ep->lock, flags);
1326
1327 for (nepi = 0, lnk = lsthead->next; lnk != lsthead && nepi < maxevents;) {
1328 epi = list_entry(lnk, struct epitem, rdllink);
1329
1330 lnk = lnk->next;
1331
1332 /* If this file is already in the ready list we exit soon */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001333 if (!ep_is_linked(&epi->txlink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334 /*
1335 * This is initialized in this way so that the default
1336 * behaviour of the reinjecting code will be to push back
1337 * the item inside the ready list.
1338 */
1339 epi->revents = epi->event.events;
1340
1341 /* Link the ready item into the transfer list */
1342 list_add(&epi->txlink, txlist);
1343 nepi++;
1344
1345 /*
1346 * Unlink the item from the ready list.
1347 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001348 ep_list_del(&epi->rdllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349 }
1350 }
1351
1352 write_unlock_irqrestore(&ep->lock, flags);
1353
1354 return nepi;
1355}
1356
1357
1358/*
1359 * This function is called without holding the "ep->lock" since the call to
1360 * __copy_to_user() might sleep, and also f_op->poll() might reenable the IRQ
1361 * because of the way poll() is traditionally implemented in Linux.
1362 */
1363static int ep_send_events(struct eventpoll *ep, struct list_head *txlist,
1364 struct epoll_event __user *events)
1365{
1366 int eventcnt = 0;
1367 unsigned int revents;
1368 struct list_head *lnk;
1369 struct epitem *epi;
1370
1371 /*
1372 * We can loop without lock because this is a task private list.
1373 * The test done during the collection loop will guarantee us that
1374 * another task will not try to collect this file. Also, items
1375 * cannot vanish during the loop because we are holding "sem".
1376 */
1377 list_for_each(lnk, txlist) {
1378 epi = list_entry(lnk, struct epitem, txlink);
1379
1380 /*
1381 * Get the ready file event set. We can safely use the file
1382 * because we are holding the "sem" in read and this will
1383 * guarantee that both the file and the item will not vanish.
1384 */
1385 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
1386
1387 /*
1388 * Set the return event set for the current file descriptor.
1389 * Note that only the task task was successfully able to link
1390 * the item to its "txlist" will write this field.
1391 */
1392 epi->revents = revents & epi->event.events;
1393
1394 if (epi->revents) {
1395 if (__put_user(epi->revents,
1396 &events[eventcnt].events) ||
1397 __put_user(epi->event.data,
1398 &events[eventcnt].data))
1399 return -EFAULT;
1400 if (epi->event.events & EPOLLONESHOT)
1401 epi->event.events &= EP_PRIVATE_BITS;
1402 eventcnt++;
1403 }
1404 }
1405 return eventcnt;
1406}
1407
1408
1409/*
1410 * Walk through the transfer list we collected with ep_collect_ready_items()
1411 * and, if 1) the item is still "alive" 2) its event set is not empty 3) it's
1412 * not already linked, links it to the ready list. Same as above, we are holding
1413 * "sem" so items cannot vanish underneath our nose.
1414 */
1415static void ep_reinject_items(struct eventpoll *ep, struct list_head *txlist)
1416{
1417 int ricnt = 0, pwake = 0;
1418 unsigned long flags;
1419 struct epitem *epi;
1420
1421 write_lock_irqsave(&ep->lock, flags);
1422
1423 while (!list_empty(txlist)) {
1424 epi = list_entry(txlist->next, struct epitem, txlink);
1425
1426 /* Unlink the current item from the transfer list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001427 ep_list_del(&epi->txlink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428
1429 /*
1430 * If the item is no more linked to the interest set, we don't
1431 * have to push it inside the ready list because the following
1432 * ep_release_epitem() is going to drop it. Also, if the current
1433 * item is set to have an Edge Triggered behaviour, we don't have
1434 * to push it back either.
1435 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001436 if (ep_rb_linked(&epi->rbn) && !(epi->event.events & EPOLLET) &&
1437 (epi->revents & epi->event.events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438 list_add_tail(&epi->rdllink, &ep->rdllist);
1439 ricnt++;
1440 }
1441 }
1442
1443 if (ricnt) {
1444 /*
1445 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1446 * wait list.
1447 */
1448 if (waitqueue_active(&ep->wq))
Davide Libenzi3419b232006-06-25 05:48:14 -07001449 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
1450 TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001451 if (waitqueue_active(&ep->poll_wait))
1452 pwake++;
1453 }
1454
1455 write_unlock_irqrestore(&ep->lock, flags);
1456
1457 /* We have to call this outside the lock */
1458 if (pwake)
1459 ep_poll_safewake(&psw, &ep->poll_wait);
1460}
1461
1462
1463/*
1464 * Perform the transfer of events to user space.
1465 */
1466static int ep_events_transfer(struct eventpoll *ep,
1467 struct epoll_event __user *events, int maxevents)
1468{
1469 int eventcnt = 0;
1470 struct list_head txlist;
1471
1472 INIT_LIST_HEAD(&txlist);
1473
1474 /*
1475 * We need to lock this because we could be hit by
1476 * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL).
1477 */
1478 down_read(&ep->sem);
1479
1480 /* Collect/extract ready items */
1481 if (ep_collect_ready_items(ep, &txlist, maxevents) > 0) {
1482 /* Build result set in userspace */
1483 eventcnt = ep_send_events(ep, &txlist, events);
1484
1485 /* Reinject ready items into the ready list */
1486 ep_reinject_items(ep, &txlist);
1487 }
1488
1489 up_read(&ep->sem);
1490
1491 return eventcnt;
1492}
1493
1494
1495static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1496 int maxevents, long timeout)
1497{
1498 int res, eavail;
1499 unsigned long flags;
1500 long jtimeout;
1501 wait_queue_t wait;
1502
1503 /*
1504 * Calculate the timeout by checking for the "infinite" value ( -1 )
1505 * and the overflow condition. The passed timeout is in milliseconds,
1506 * that why (t * HZ) / 1000.
1507 */
Davide Libenzie3306dd2005-09-27 21:45:33 -07001508 jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
1509 MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510
1511retry:
1512 write_lock_irqsave(&ep->lock, flags);
1513
1514 res = 0;
1515 if (list_empty(&ep->rdllist)) {
1516 /*
1517 * We don't have any available event to return to the caller.
1518 * We need to sleep here, and we will be wake up by
1519 * ep_poll_callback() when events will become available.
1520 */
1521 init_waitqueue_entry(&wait, current);
Davide Libenzi3419b232006-06-25 05:48:14 -07001522 __add_wait_queue(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523
1524 for (;;) {
1525 /*
1526 * We don't want to sleep if the ep_poll_callback() sends us
1527 * a wakeup in between. That's why we set the task state
1528 * to TASK_INTERRUPTIBLE before doing the checks.
1529 */
1530 set_current_state(TASK_INTERRUPTIBLE);
1531 if (!list_empty(&ep->rdllist) || !jtimeout)
1532 break;
1533 if (signal_pending(current)) {
1534 res = -EINTR;
1535 break;
1536 }
1537
1538 write_unlock_irqrestore(&ep->lock, flags);
1539 jtimeout = schedule_timeout(jtimeout);
1540 write_lock_irqsave(&ep->lock, flags);
1541 }
Davide Libenzi3419b232006-06-25 05:48:14 -07001542 __remove_wait_queue(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001543
1544 set_current_state(TASK_RUNNING);
1545 }
1546
1547 /* Is it worth to try to dig for events ? */
1548 eavail = !list_empty(&ep->rdllist);
1549
1550 write_unlock_irqrestore(&ep->lock, flags);
1551
1552 /*
1553 * Try to transfer events to user space. In case we get 0 events and
1554 * there's still timeout left over, we go trying again in search of
1555 * more luck.
1556 */
1557 if (!res && eavail &&
1558 !(res = ep_events_transfer(ep, events, maxevents)) && jtimeout)
1559 goto retry;
1560
1561 return res;
1562}
1563
1564
1565static int eventpollfs_delete_dentry(struct dentry *dentry)
1566{
1567
1568 return 1;
1569}
1570
1571
1572static struct inode *ep_eventpoll_inode(void)
1573{
1574 int error = -ENOMEM;
1575 struct inode *inode = new_inode(eventpoll_mnt->mnt_sb);
1576
1577 if (!inode)
1578 goto eexit_1;
1579
1580 inode->i_fop = &eventpoll_fops;
1581
1582 /*
1583 * Mark the inode dirty from the very beginning,
1584 * that way it will never be moved to the dirty
1585 * list because mark_inode_dirty() will think
1586 * that it already _is_ on the dirty list.
1587 */
1588 inode->i_state = I_DIRTY;
1589 inode->i_mode = S_IRUSR | S_IWUSR;
1590 inode->i_uid = current->fsuid;
1591 inode->i_gid = current->fsgid;
1592 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1593 inode->i_blksize = PAGE_SIZE;
1594 return inode;
1595
1596eexit_1:
1597 return ERR_PTR(error);
1598}
1599
1600
David Howells454e2392006-06-23 02:02:57 -07001601static int
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602eventpollfs_get_sb(struct file_system_type *fs_type, int flags,
David Howells454e2392006-06-23 02:02:57 -07001603 const char *dev_name, void *data, struct vfsmount *mnt)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604{
David Howells454e2392006-06-23 02:02:57 -07001605 return get_sb_pseudo(fs_type, "eventpoll:", NULL, EVENTPOLLFS_MAGIC,
1606 mnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001607}
1608
1609
1610static int __init eventpoll_init(void)
1611{
1612 int error;
1613
Arjan van de Ven144efe32006-03-23 03:00:32 -08001614 mutex_init(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615
1616 /* Initialize the structure used to perform safe poll wait head wake ups */
1617 ep_poll_safewake_init(&psw);
1618
1619 /* Allocates slab cache used to allocate "struct epitem" items */
1620 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1621 0, SLAB_HWCACHE_ALIGN|EPI_SLAB_DEBUG|SLAB_PANIC,
1622 NULL, NULL);
1623
1624 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1625 pwq_cache = kmem_cache_create("eventpoll_pwq",
1626 sizeof(struct eppoll_entry), 0,
1627 EPI_SLAB_DEBUG|SLAB_PANIC, NULL, NULL);
1628
1629 /*
1630 * Register the virtual file system that will be the source of inodes
1631 * for the eventpoll files
1632 */
1633 error = register_filesystem(&eventpoll_fs_type);
1634 if (error)
1635 goto epanic;
1636
1637 /* Mount the above commented virtual file system */
1638 eventpoll_mnt = kern_mount(&eventpoll_fs_type);
1639 error = PTR_ERR(eventpoll_mnt);
1640 if (IS_ERR(eventpoll_mnt))
1641 goto epanic;
1642
1643 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: successfully initialized.\n",
1644 current));
1645 return 0;
1646
1647epanic:
1648 panic("eventpoll_init() failed\n");
1649}
1650
1651
1652static void __exit eventpoll_exit(void)
1653{
1654 /* Undo all operations done inside eventpoll_init() */
1655 unregister_filesystem(&eventpoll_fs_type);
1656 mntput(eventpoll_mnt);
1657 kmem_cache_destroy(pwq_cache);
1658 kmem_cache_destroy(epi_cache);
1659}
1660
1661module_init(eventpoll_init);
1662module_exit(eventpoll_exit);
1663
1664MODULE_LICENSE("GPL");