2 * fs/eventpoll.c ( Efficent event polling implementation )
3 * Copyright (C) 2001,...,2003 Davide Libenzi
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.
10 * Davide Libenzi <davidel@xmailserver.org>
14 #include <linux/module.h>
15 #include <linux/init.h>
16 #include <linux/kernel.h>
17 #include <linux/sched.h>
19 #include <linux/file.h>
20 #include <linux/signal.h>
21 #include <linux/errno.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/wait.h>
33 #include <linux/eventpoll.h>
34 #include <linux/mount.h>
35 #include <asm/bitops.h>
36 #include <asm/uaccess.h>
37 #include <asm/system.h>
40 #include <asm/atomic.h>
41 #include <asm/semaphore.h>
46 * There are three level of locking required by epoll :
48 * 1) epsem (semaphore)
49 * 2) ep->sem (rw_semaphore)
50 * 3) ep->lock (rw_lock)
52 * The acquire order is the one listed above, from 1 to 3.
53 * We need a spinlock (ep->lock) because we manipulate objects
54 * from inside the poll callback, that might be triggered from
55 * a wake_up() that in turn might be called from IRQ context.
56 * So we can't sleep inside the poll callback and hence we need
57 * a spinlock. During the event transfer loop (from kernel to
58 * user space) we could end up sleeping due a copy_to_user(), so
59 * we need a lock that will allow us to sleep. This lock is a
60 * read-write semaphore (ep->sem). It is acquired on read during
61 * the event transfer loop and in write during epoll_ctl(EPOLL_CTL_DEL)
62 * and during eventpoll_release_file(). Then we also need a global
63 * semaphore to serialize eventpoll_release_file() and ep_free().
64 * This semaphore is acquired by ep_free() during the epoll file
65 * cleanup path and it is also acquired by eventpoll_release_file()
66 * if a file has been pushed inside an epoll set and it is then
67 * close()d without a previous call toepoll_ctl(EPOLL_CTL_DEL).
68 * It is possible to drop the "ep->sem" and to use the global
69 * semaphore "epsem" (together with "ep->lock") to have it working,
70 * but having "ep->sem" will make the interface more scalable.
71 * Events that require holding "epsem" are very rare, while for
72 * normal operations the epoll private "ep->sem" will guarantee
73 * a greater scalability.
77 #define EVENTPOLLFS_MAGIC 0x03111965 /* My birthday should work for this :) */
82 #define DPRINTK(x) printk x
83 #define DNPRINTK(n, x) do { if ((n) <= DEBUG_EPOLL) printk x; } while (0)
84 #else /* #if DEBUG_EPOLL > 0 */
85 #define DPRINTK(x) (void) 0
86 #define DNPRINTK(n, x) (void) 0
87 #endif /* #if DEBUG_EPOLL > 0 */
92 #define EPI_SLAB_DEBUG (SLAB_DEBUG_FREE | SLAB_RED_ZONE /* | SLAB_POISON */)
93 #else /* #if DEBUG_EPI != 0 */
94 #define EPI_SLAB_DEBUG 0
95 #endif /* #if DEBUG_EPI != 0 */
97 /* Epoll private bits inside the event mask */
98 #define EP_PRIVATE_BITS (EPOLLONESHOT | EPOLLET)
100 /* Maximum number of poll wake up nests we are allowing */
101 #define EP_MAX_POLLWAKE_NESTS 4
103 /* Maximum size of the hash in bits ( 2^N ) */
104 #define EP_MAX_HASH_BITS 17
106 /* Minimum size of the hash in bits ( 2^N ) */
107 #define EP_MIN_HASH_BITS 9
109 /* Number of hash entries ( "struct list_head" ) inside a page */
110 #define EP_HENTRY_X_PAGE (PAGE_SIZE / sizeof(struct list_head))
112 /* Maximum size of the hash in pages */
113 #define EP_MAX_HPAGES ((1 << EP_MAX_HASH_BITS) / EP_HENTRY_X_PAGE + 1)
115 /* Number of pages allocated for an "hbits" sized hash table */
116 #define EP_HASH_PAGES(hbits) ((int) ((1 << (hbits)) / EP_HENTRY_X_PAGE + \
117 ((1 << (hbits)) % EP_HENTRY_X_PAGE ? 1: 0)))
119 /* Macro to allocate a "struct epitem" from the slab cache */
120 #define EPI_MEM_ALLOC() (struct epitem *) kmem_cache_alloc(epi_cache, SLAB_KERNEL)
122 /* Macro to free a "struct epitem" to the slab cache */
123 #define EPI_MEM_FREE(p) kmem_cache_free(epi_cache, p)
125 /* Macro to allocate a "struct eppoll_entry" from the slab cache */
126 #define PWQ_MEM_ALLOC() (struct eppoll_entry *) kmem_cache_alloc(pwq_cache, SLAB_KERNEL)
128 /* Macro to free a "struct eppoll_entry" to the slab cache */
129 #define PWQ_MEM_FREE(p) kmem_cache_free(pwq_cache, p)
131 /* Fast test to see if the file is an evenpoll file */
132 #define IS_FILE_EPOLL(f) ((f)->f_op == &eventpoll_fops)
135 * Remove the item from the list and perform its initialization.
136 * This is useful for us because we can test if the item is linked
137 * using "EP_IS_LINKED(p)".
139 #define EP_LIST_DEL(p) do { list_del(p); INIT_LIST_HEAD(p); } while (0)
141 /* Tells us if the item is currently linked */
142 #define EP_IS_LINKED(p) (!list_empty(p))
144 /* Get the "struct epitem" from a wait queue pointer */
145 #define EP_ITEM_FROM_WAIT(p) ((struct epitem *) container_of(p, struct eppoll_entry, wait)->base)
147 /* Get the "struct epitem" from an epoll queue wrapper */
148 #define EP_ITEM_FROM_EPQUEUE(p) (container_of(p, struct ep_pqueue, pt)->epi)
151 * Node that is linked into the "wake_task_list" member of the "struct poll_safewake".
152 * It is used to keep track on all tasks that are currently inside the wake_up() code
153 * to 1) short-circuit the one coming from the same task and same wait queue head
154 * ( loop ) 2) allow a maximum number of epoll descriptors inclusion nesting
155 * 3) let go the ones coming from other tasks.
157 struct wake_task_node {
158 struct list_head llink;
160 wait_queue_head_t *wq;
164 * This is used to implement the safe poll wake up avoiding to reenter
165 * the poll callback from inside wake_up().
167 struct poll_safewake {
168 struct list_head wake_task_list;
173 * This structure is stored inside the "private_data" member of the file
174 * structure and rapresent the main data sructure for the eventpoll
178 /* Protect the this structure access */
182 * This semaphore is used to ensure that files are not removed
183 * while epoll is using them. This is read-held during the event
184 * collection loop and it is write-held during the file cleanup
185 * path, the epoll file exit code and the ctl operations.
187 struct rw_semaphore sem;
189 /* Wait queue used by sys_epoll_wait() */
190 wait_queue_head_t wq;
192 /* Wait queue used by file->poll() */
193 wait_queue_head_t poll_wait;
195 /* List of ready file descriptors */
196 struct list_head rdllist;
198 /* Size of the hash */
199 unsigned int hashbits;
201 /* Pages for the "struct epitem" hash */
202 char *hpages[EP_MAX_HPAGES];
205 /* Wait structure used by the poll hooks */
206 struct eppoll_entry {
207 /* List header used to link this structure to the "struct epitem" */
208 struct list_head llink;
210 /* The "base" pointer is set to the container "struct epitem" */
214 * Wait queue item that will be linked to the target file wait
219 /* The wait queue head that linked the "wait" wait queue item */
220 wait_queue_head_t *whead;
224 * Each file descriptor added to the eventpoll interface will
225 * have an entry of this type linked to the hash.
228 /* List header used to link this structure to the eventpoll hash */
229 struct list_head llink;
231 /* List header used to link this structure to the eventpoll ready list */
232 struct list_head rdllink;
234 /* The file descriptor this item refers to */
237 /* Number of active wait queue attached to poll operations */
240 /* List containing poll wait queues */
241 struct list_head pwqlist;
243 /* The "container" of this item */
244 struct eventpoll *ep;
246 /* The file this item refers to */
249 /* The structure that describe the interested events and the source fd */
250 struct epoll_event event;
253 * Used to keep track of the usage count of the structure. This avoids
254 * that the structure will desappear from underneath our processing.
258 /* List header used to link this item to the "struct file" items list */
259 struct list_head fllink;
261 /* List header used to link the item to the transfer list */
262 struct list_head txlink;
265 * This is used during the collection/transfer of events to userspace
266 * to pin items empty events set.
268 unsigned int revents;
271 /* Wrapper struct used by poll queueing */
279 static void ep_poll_safewake_init(struct poll_safewake *psw);
280 static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq);
281 static unsigned int ep_get_hash_bits(unsigned int hintsize);
282 static int ep_getfd(int *efd, struct inode **einode, struct file **efile);
283 static int ep_alloc_pages(char **pages, int numpages);
284 static int ep_free_pages(char **pages, int numpages);
285 static int ep_file_init(struct file *file, unsigned int hashbits);
286 static unsigned int ep_hash_index(struct eventpoll *ep, struct file *file,
288 static struct list_head *ep_hash_entry(struct eventpoll *ep,
290 static int ep_init(struct eventpoll *ep, unsigned int hashbits);
291 static void ep_free(struct eventpoll *ep);
292 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd);
293 static void ep_use_epitem(struct epitem *epi);
294 static void ep_release_epitem(struct epitem *epi);
295 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
297 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
298 struct file *tfile, int fd);
299 static int ep_modify(struct eventpoll *ep, struct epitem *epi,
300 struct epoll_event *event);
301 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi);
302 static int ep_unlink(struct eventpoll *ep, struct epitem *epi);
303 static int ep_remove(struct eventpoll *ep, struct epitem *epi);
304 static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key);
305 static int ep_eventpoll_close(struct inode *inode, struct file *file);
306 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait);
307 static int ep_collect_ready_items(struct eventpoll *ep,
308 struct list_head *txlist, int maxevents);
309 static int ep_send_events(struct eventpoll *ep, struct list_head *txlist,
310 struct epoll_event __user *events);
311 static void ep_reinject_items(struct eventpoll *ep, struct list_head *txlist);
312 static int ep_events_transfer(struct eventpoll *ep,
313 struct epoll_event __user *events,
315 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
316 int maxevents, long timeout);
317 static int eventpollfs_delete_dentry(struct dentry *dentry);
318 static struct inode *ep_eventpoll_inode(void);
319 static struct super_block *eventpollfs_get_sb(struct file_system_type *fs_type,
320 int flags, const char *dev_name,
324 * This semaphore is used to serialize ep_free() and eventpoll_release_file().
326 struct semaphore epsem;
328 /* Safe wake up implementation */
329 static struct poll_safewake psw;
331 /* Slab cache used to allocate "struct epitem" */
332 static kmem_cache_t *epi_cache;
334 /* Slab cache used to allocate "struct eppoll_entry" */
335 static kmem_cache_t *pwq_cache;
337 /* Virtual fs used to allocate inodes for eventpoll files */
338 static struct vfsmount *eventpoll_mnt;
340 /* File callbacks that implement the eventpoll file behaviour */
341 static struct file_operations eventpoll_fops = {
342 .release = ep_eventpoll_close,
343 .poll = ep_eventpoll_poll
347 * This is used to register the virtual file system from where
348 * eventpoll inodes are allocated.
350 static struct file_system_type eventpoll_fs_type = {
351 .name = "eventpollfs",
352 .get_sb = eventpollfs_get_sb,
353 .kill_sb = kill_anon_super,
356 /* Very basic directory entry operations for the eventpoll virtual file system */
357 static struct dentry_operations eventpollfs_dentry_operations = {
358 .d_delete = eventpollfs_delete_dentry,
363 /* Initialize the poll safe wake up structure */
364 static void ep_poll_safewake_init(struct poll_safewake *psw)
367 INIT_LIST_HEAD(&psw->wake_task_list);
368 spin_lock_init(&psw->lock);
373 * Perform a safe wake up of the poll wait list. The problem is that
374 * with the new callback'd wake up system, it is possible that the
375 * poll callback is reentered from inside the call to wake_up() done
376 * on the poll wait queue head. The rule is that we cannot reenter the
377 * wake up code from the same task more than EP_MAX_POLLWAKE_NESTS times,
378 * and we cannot reenter the same wait queue head at all. This will
379 * enable to have a hierarchy of epoll file descriptor of no more than
380 * EP_MAX_POLLWAKE_NESTS deep. We need the irq version of the spin lock
381 * because this one gets called by the poll callback, that in turn is called
382 * from inside a wake_up(), that might be called from irq context.
384 static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq)
388 task_t *this_task = current;
389 struct list_head *lsthead = &psw->wake_task_list, *lnk;
390 struct wake_task_node *tncur;
391 struct wake_task_node tnode;
393 spin_lock_irqsave(&psw->lock, flags);
395 /* Try to see if the current task is already inside this wakeup call */
396 list_for_each(lnk, lsthead) {
397 tncur = list_entry(lnk, struct wake_task_node, llink);
399 if (tncur->wq == wq ||
400 (tncur->task == this_task && ++wake_nests > EP_MAX_POLLWAKE_NESTS)) {
402 * Ops ... loop detected or maximum nest level reached.
403 * We abort this wake by breaking the cycle itself.
405 spin_unlock_irqrestore(&psw->lock, flags);
410 /* Add the current task to the list */
411 tnode.task = this_task;
413 list_add(&tnode.llink, lsthead);
415 spin_unlock_irqrestore(&psw->lock, flags);
417 /* Do really wake up now */
420 /* Remove the current task from the list */
421 spin_lock_irqsave(&psw->lock, flags);
422 list_del(&tnode.llink);
423 spin_unlock_irqrestore(&psw->lock, flags);
428 * Calculate the size of the hash in bits. The returned size will be
429 * bounded between EP_MIN_HASH_BITS and EP_MAX_HASH_BITS.
431 static unsigned int ep_get_hash_bits(unsigned int hintsize)
435 for (i = 0, val = 1; val < hintsize && i < EP_MAX_HASH_BITS; i++, val <<= 1);
436 return i < EP_MIN_HASH_BITS ? EP_MIN_HASH_BITS: i;
440 /* Used to initialize the epoll bits inside the "struct file" */
441 void eventpoll_init_file(struct file *file)
444 INIT_LIST_HEAD(&file->f_ep_links);
445 spin_lock_init(&file->f_ep_lock);
450 * This is called from eventpoll_release() to unlink files from the eventpoll
451 * interface. We need to have this facility to cleanup correctly files that are
452 * closed without being removed from the eventpoll interface.
454 void eventpoll_release_file(struct file *file)
456 struct list_head *lsthead = &file->f_ep_links;
457 struct eventpoll *ep;
461 * We don't want to get "file->f_ep_lock" because it is not
462 * necessary. It is not necessary because we're in the "struct file"
463 * cleanup path, and this means that noone is using this file anymore.
464 * The only hit might come from ep_free() but by holding the semaphore
465 * will correctly serialize the operation. We do need to acquire
466 * "ep->sem" after "epsem" because ep_remove() requires it when called
467 * from anywhere but ep_free().
471 while (!list_empty(lsthead)) {
472 epi = list_entry(lsthead->next, struct epitem, fllink);
475 EP_LIST_DEL(&epi->fllink);
476 down_write(&ep->sem);
486 * It opens an eventpoll file descriptor by suggesting a storage of "size"
487 * file descriptors. The size parameter is just an hint about how to size
488 * data structures. It won't prevent the user to store more than "size"
489 * file descriptors inside the epoll interface. It is the kernel part of
490 * the userspace epoll_create(2).
492 asmlinkage long sys_epoll_create(int size)
495 unsigned int hashbits;
499 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d)\n",
502 /* Correctly size the hash */
503 hashbits = ep_get_hash_bits((unsigned int) size);
506 * Creates all the items needed to setup an eventpoll file. That is,
507 * a file structure, and inode and a free file descriptor.
509 error = ep_getfd(&fd, &inode, &file);
513 /* Setup the file internal data structure ( "struct eventpoll" ) */
514 error = ep_file_init(file, hashbits);
519 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
527 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
528 current, size, error));
534 * The following function implements the controller interface for
535 * the eventpoll file that enables the insertion/removal/change of
536 * file descriptors inside the interest set. It represents
537 * the kernel part of the user space epoll_ctl(2).
540 sys_epoll_ctl(int epfd, int op, int fd, struct epoll_event __user *event)
543 struct file *file, *tfile;
544 struct eventpoll *ep;
546 struct epoll_event epds;
548 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p)\n",
549 current, epfd, op, fd, event));
552 if (copy_from_user(&epds, event, sizeof(struct epoll_event)))
555 /* Get the "struct file *" for the eventpoll file */
561 /* Get the "struct file *" for the target file */
566 /* The target file descriptor must support poll */
568 if (!tfile->f_op || !tfile->f_op->poll)
572 * We have to check that the file structure underneath the file descriptor
573 * the user passed to us _is_ an eventpoll file. And also we do not permit
574 * adding an epoll file descriptor inside itself.
577 if (file == tfile || !IS_FILE_EPOLL(file))
581 * At this point it is safe to assume that the "private_data" contains
582 * our own data structure.
584 ep = file->private_data;
586 down_write(&ep->sem);
588 /* Try to lookup the file inside our hash table */
589 epi = ep_find(ep, tfile, fd);
595 epds.events |= POLLERR | POLLHUP;
597 error = ep_insert(ep, &epds, tfile, fd);
603 error = ep_remove(ep, epi);
609 epds.events |= POLLERR | POLLHUP;
610 error = ep_modify(ep, epi, &epds);
617 * The function ep_find() increments the usage count of the structure
618 * so, if this is not NULL, we need to release it.
621 ep_release_epitem(epi);
630 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p) = %d\n",
631 current, epfd, op, fd, event, error));
638 * Implement the event wait interface for the eventpoll file. It is the kernel
639 * part of the user space epoll_wait(2).
641 asmlinkage long sys_epoll_wait(int epfd, struct epoll_event __user *events,
642 int maxevents, int timeout)
646 struct eventpoll *ep;
648 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d)\n",
649 current, epfd, events, maxevents, timeout));
651 /* The maximum number of event must be greater than zero */
655 /* Verify that the area passed by the user is writeable */
656 if ((error = verify_area(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))))
659 /* Get the "struct file *" for the eventpoll file */
666 * We have to check that the file structure underneath the fd
667 * the user passed to us _is_ an eventpoll file.
670 if (!IS_FILE_EPOLL(file))
674 * At this point it is safe to assume that the "private_data" contains
675 * our own data structure.
677 ep = file->private_data;
679 /* Time to fish for events ... */
680 error = ep_poll(ep, events, maxevents, timeout);
685 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d) = %d\n",
686 current, epfd, events, maxevents, timeout, error));
693 * Creates the file descriptor to be used by the epoll interface.
695 static int ep_getfd(int *efd, struct inode **einode, struct file **efile)
699 struct dentry *dentry;
704 /* Get an ready to use file */
706 file = get_empty_filp();
710 /* Allocates an inode from the eventpoll file system */
711 inode = ep_eventpoll_inode();
712 error = PTR_ERR(inode);
716 /* Allocates a free descriptor to plug the file onto */
717 error = get_unused_fd();
723 * Link the inode to a directory entry by creating a unique name
724 * using the inode number.
727 sprintf(name, "[%lu]", inode->i_ino);
729 this.len = strlen(name);
730 this.hash = inode->i_ino;
731 dentry = d_alloc(eventpoll_mnt->mnt_sb->s_root, &this);
734 dentry->d_op = &eventpollfs_dentry_operations;
735 d_add(dentry, inode);
736 file->f_vfsmnt = mntget(eventpoll_mnt);
737 file->f_dentry = dget(dentry);
738 file->f_mapping = inode->i_mapping;
741 file->f_flags = O_RDONLY;
742 file->f_op = &eventpoll_fops;
743 file->f_mode = FMODE_READ;
745 file->private_data = NULL;
747 /* Install the new setup file into the allocated fd. */
748 fd_install(fd, file);
766 static int ep_alloc_pages(char **pages, int numpages)
770 for (i = 0; i < numpages; i++) {
771 pages[i] = (char *) __get_free_pages(GFP_KERNEL, 0);
773 for (--i; i >= 0; i--) {
774 ClearPageReserved(virt_to_page(pages[i]));
775 free_pages((unsigned long) pages[i], 0);
779 SetPageReserved(virt_to_page(pages[i]));
785 static int ep_free_pages(char **pages, int numpages)
789 for (i = 0; i < numpages; i++) {
790 ClearPageReserved(virt_to_page(pages[i]));
791 free_pages((unsigned long) pages[i], 0);
797 static int ep_file_init(struct file *file, unsigned int hashbits)
800 struct eventpoll *ep;
802 if (!(ep = kmalloc(sizeof(struct eventpoll), GFP_KERNEL)))
805 memset(ep, 0, sizeof(*ep));
807 error = ep_init(ep, hashbits);
813 file->private_data = ep;
815 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_file_init() ep=%p\n",
822 * Calculate the index of the hash relative to "file".
824 static unsigned int ep_hash_index(struct eventpoll *ep, struct file *file, int fd)
826 unsigned long ptr = (unsigned long) file ^ (fd << ep->hashbits);
828 return (unsigned int) hash_ptr((void *) ptr, ep->hashbits);
833 * Returns the hash entry ( struct list_head * ) of the passed index.
835 static struct list_head *ep_hash_entry(struct eventpoll *ep, unsigned int index)
838 return (struct list_head *) (ep->hpages[index / EP_HENTRY_X_PAGE] +
839 (index % EP_HENTRY_X_PAGE) * sizeof(struct list_head));
843 static int ep_init(struct eventpoll *ep, unsigned int hashbits)
846 unsigned int i, hsize;
848 rwlock_init(&ep->lock);
849 init_rwsem(&ep->sem);
850 init_waitqueue_head(&ep->wq);
851 init_waitqueue_head(&ep->poll_wait);
852 INIT_LIST_HEAD(&ep->rdllist);
854 /* Hash allocation and setup */
855 ep->hashbits = hashbits;
856 error = ep_alloc_pages(ep->hpages, EP_HASH_PAGES(ep->hashbits));
860 /* Initialize hash buckets */
861 for (i = 0, hsize = 1 << hashbits; i < hsize; i++)
862 INIT_LIST_HEAD(ep_hash_entry(ep, i));
870 static void ep_free(struct eventpoll *ep)
872 unsigned int i, hsize;
873 struct list_head *lsthead, *lnk;
876 /* We need to release all tasks waiting for these file */
877 if (waitqueue_active(&ep->poll_wait))
878 ep_poll_safewake(&psw, &ep->poll_wait);
881 * We need to lock this because we could be hit by
882 * eventpoll_release_file() while we're freeing the "struct eventpoll".
883 * We do not need to hold "ep->sem" here because the epoll file
884 * is on the way to be removed and no one has references to it
885 * anymore. The only hit might come from eventpoll_release_file() but
886 * holding "epsem" is sufficent here.
891 * Walks through the whole hash by unregistering poll callbacks.
893 for (i = 0, hsize = 1 << ep->hashbits; i < hsize; i++) {
894 lsthead = ep_hash_entry(ep, i);
896 list_for_each(lnk, lsthead) {
897 epi = list_entry(lnk, struct epitem, llink);
899 ep_unregister_pollwait(ep, epi);
904 * Walks through the whole hash by freeing each "struct epitem". At this
905 * point we are sure no poll callbacks will be lingering around, and also by
906 * write-holding "sem" we can be sure that no file cleanup code will hit
907 * us during this operation. So we can avoid the lock on "ep->lock".
909 for (i = 0, hsize = 1 << ep->hashbits; i < hsize; i++) {
910 lsthead = ep_hash_entry(ep, i);
912 while (!list_empty(lsthead)) {
913 epi = list_entry(lsthead->next, struct epitem, llink);
921 /* Free hash pages */
922 ep_free_pages(ep->hpages, EP_HASH_PAGES(ep->hashbits));
927 * Search the file inside the eventpoll hash. It add usage count to
928 * the returned item, so the caller must call ep_release_epitem()
929 * after finished using the "struct epitem".
931 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
934 struct list_head *lsthead, *lnk;
935 struct epitem *epi = NULL;
937 read_lock_irqsave(&ep->lock, flags);
939 lsthead = ep_hash_entry(ep, ep_hash_index(ep, file, fd));
940 list_for_each(lnk, lsthead) {
941 epi = list_entry(lnk, struct epitem, llink);
943 if (epi->file == file && epi->fd == fd) {
950 read_unlock_irqrestore(&ep->lock, flags);
952 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_find(%p) -> %p\n",
953 current, file, epi));
960 * Increment the usage count of the "struct epitem" making it sure
961 * that the user will have a valid pointer to reference.
963 static void ep_use_epitem(struct epitem *epi)
966 atomic_inc(&epi->usecnt);
971 * Decrement ( release ) the usage count by signaling that the user
972 * has finished using the structure. It might lead to freeing the
973 * structure itself if the count goes to zero.
975 static void ep_release_epitem(struct epitem *epi)
978 if (atomic_dec_and_test(&epi->usecnt))
984 * This is the callback that is used to add our wait queue to the
985 * target file wakeup lists.
987 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
990 struct epitem *epi = EP_ITEM_FROM_EPQUEUE(pt);
991 struct eppoll_entry *pwq;
993 if (epi->nwait >= 0 && (pwq = PWQ_MEM_ALLOC())) {
994 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
997 add_wait_queue(whead, &pwq->wait);
998 list_add_tail(&pwq->llink, &epi->pwqlist);
1001 /* We have to signal that an error occurred */
1007 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
1008 struct file *tfile, int fd)
1010 int error, revents, pwake = 0;
1011 unsigned long flags;
1013 struct ep_pqueue epq;
1016 if (!(epi = EPI_MEM_ALLOC()))
1019 /* Item initialization follow here ... */
1020 INIT_LIST_HEAD(&epi->llink);
1021 INIT_LIST_HEAD(&epi->rdllink);
1022 INIT_LIST_HEAD(&epi->fllink);
1023 INIT_LIST_HEAD(&epi->txlink);
1024 INIT_LIST_HEAD(&epi->pwqlist);
1028 epi->event = *event;
1029 atomic_set(&epi->usecnt, 1);
1032 /* Initialize the poll table using the queue callback */
1034 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1037 * Attach the item to the poll hooks and get current event bits.
1038 * We can safely use the file* here because its usage count has
1039 * been increased by the caller of this function.
1041 revents = tfile->f_op->poll(tfile, &epq.pt);
1044 * We have to check if something went wrong during the poll wait queue
1045 * install process. Namely an allocation for a wait queue failed due
1046 * high memory pressure.
1051 /* Add the current item to the list of active epoll hook for this file */
1052 spin_lock(&tfile->f_ep_lock);
1053 list_add_tail(&epi->fllink, &tfile->f_ep_links);
1054 spin_unlock(&tfile->f_ep_lock);
1056 /* We have to drop the new item inside our item list to keep track of it */
1057 write_lock_irqsave(&ep->lock, flags);
1059 /* Add the current item to the hash table */
1060 list_add(&epi->llink, ep_hash_entry(ep, ep_hash_index(ep, tfile, fd)));
1062 /* If the file is already "ready" we drop it inside the ready list */
1063 if ((revents & event->events) && !EP_IS_LINKED(&epi->rdllink)) {
1064 list_add_tail(&epi->rdllink, &ep->rdllist);
1066 /* Notify waiting tasks that events are available */
1067 if (waitqueue_active(&ep->wq))
1069 if (waitqueue_active(&ep->poll_wait))
1073 write_unlock_irqrestore(&ep->lock, flags);
1075 /* We have to call this outside the lock */
1077 ep_poll_safewake(&psw, &ep->poll_wait);
1079 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_insert(%p, %p, %d)\n",
1080 current, ep, tfile, fd));
1085 ep_unregister_pollwait(ep, epi);
1088 * We need to do this because an event could have been arrived on some
1089 * allocated wait queue.
1091 write_lock_irqsave(&ep->lock, flags);
1092 if (EP_IS_LINKED(&epi->rdllink))
1093 EP_LIST_DEL(&epi->rdllink);
1094 write_unlock_irqrestore(&ep->lock, flags);
1103 * Modify the interest event mask by dropping an event if the new mask
1104 * has a match in the current file status.
1106 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1109 unsigned int revents;
1110 unsigned long flags;
1113 * Set the new event interest mask before calling f_op->poll(), otherwise
1114 * a potential race might occur. In fact if we do this operation inside
1115 * the lock, an event might happen between the f_op->poll() call and the
1116 * new event set registering.
1118 epi->event.events = event->events;
1121 * Get current event bits. We can safely use the file* here because
1122 * its usage count has been increased by the caller of this function.
1124 revents = epi->file->f_op->poll(epi->file, NULL);
1126 write_lock_irqsave(&ep->lock, flags);
1128 /* Copy the data member from inside the lock */
1129 epi->event.data = event->data;
1132 * If the item is not linked to the hash it means that it's on its
1133 * way toward the removal. Do nothing in this case.
1135 if (EP_IS_LINKED(&epi->llink)) {
1137 * If the item is "hot" and it is not registered inside the ready
1138 * list, push it inside. If the item is not "hot" and it is currently
1139 * registered inside the ready list, unlink it.
1141 if (revents & event->events) {
1142 if (!EP_IS_LINKED(&epi->rdllink)) {
1143 list_add_tail(&epi->rdllink, &ep->rdllist);
1145 /* Notify waiting tasks that events are available */
1146 if (waitqueue_active(&ep->wq))
1148 if (waitqueue_active(&ep->poll_wait))
1154 write_unlock_irqrestore(&ep->lock, flags);
1156 /* We have to call this outside the lock */
1158 ep_poll_safewake(&psw, &ep->poll_wait);
1165 * This function unregister poll callbacks from the associated file descriptor.
1166 * Since this must be called without holding "ep->lock" the atomic exchange trick
1167 * will protect us from multiple unregister.
1169 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
1172 struct list_head *lsthead = &epi->pwqlist;
1173 struct eppoll_entry *pwq;
1175 /* This is called without locks, so we need the atomic exchange */
1176 nwait = xchg(&epi->nwait, 0);
1179 while (!list_empty(lsthead)) {
1180 pwq = list_entry(lsthead->next, struct eppoll_entry, llink);
1182 EP_LIST_DEL(&pwq->llink);
1183 remove_wait_queue(pwq->whead, &pwq->wait);
1191 * Unlink the "struct epitem" from all places it might have been hooked up.
1192 * This function must be called with write IRQ lock on "ep->lock".
1194 static int ep_unlink(struct eventpoll *ep, struct epitem *epi)
1199 * It can happen that this one is called for an item already unlinked.
1200 * The check protect us from doing a double unlink ( crash ).
1203 if (!EP_IS_LINKED(&epi->llink))
1207 * Clear the event mask for the unlinked item. This will avoid item
1208 * notifications to be sent after the unlink operation from inside
1209 * the kernel->userspace event transfer loop.
1211 epi->event.events = 0;
1214 * At this point is safe to do the job, unlink the item from our list.
1215 * This operation togheter with the above check closes the door to
1218 EP_LIST_DEL(&epi->llink);
1221 * If the item we are going to remove is inside the ready file descriptors
1222 * we want to remove it from this list to avoid stale events.
1224 if (EP_IS_LINKED(&epi->rdllink))
1225 EP_LIST_DEL(&epi->rdllink);
1230 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_unlink(%p, %p) = %d\n",
1231 current, ep, epi->file, error));
1238 * Removes a "struct epitem" from the eventpoll hash and deallocates
1239 * all the associated resources.
1241 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
1244 unsigned long flags;
1245 struct file *file = epi->file;
1248 * Removes poll wait queue hooks. We _have_ to do this without holding
1249 * the "ep->lock" otherwise a deadlock might occur. This because of the
1250 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
1251 * queue head lock when unregistering the wait queue. The wakeup callback
1252 * will run by holding the wait queue head lock and will call our callback
1253 * that will try to get "ep->lock".
1255 ep_unregister_pollwait(ep, epi);
1257 /* Remove the current item from the list of epoll hooks */
1258 spin_lock(&file->f_ep_lock);
1259 if (EP_IS_LINKED(&epi->fllink))
1260 EP_LIST_DEL(&epi->fllink);
1261 spin_unlock(&file->f_ep_lock);
1263 /* We need to acquire the write IRQ lock before calling ep_unlink() */
1264 write_lock_irqsave(&ep->lock, flags);
1266 /* Really unlink the item from the hash */
1267 error = ep_unlink(ep, epi);
1269 write_unlock_irqrestore(&ep->lock, flags);
1274 /* At this point it is safe to free the eventpoll item */
1275 ep_release_epitem(epi);
1279 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_remove(%p, %p) = %d\n",
1280 current, ep, file, error));
1287 * This is the callback that is passed to the wait queue wakeup
1288 * machanism. It is called by the stored file descriptors when they
1289 * have events to report.
1291 static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
1294 unsigned long flags;
1295 struct epitem *epi = EP_ITEM_FROM_WAIT(wait);
1296 struct eventpoll *ep = epi->ep;
1298 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: poll_callback(%p) epi=%p ep=%p\n",
1299 current, epi->file, epi, ep));
1301 write_lock_irqsave(&ep->lock, flags);
1304 * If the event mask does not contain any poll(2) event, we consider the
1305 * descriptor to be disabled. This condition is likely the effect of the
1306 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1307 * until the next EPOLL_CTL_MOD will be issued.
1309 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1312 /* If this file is already in the ready list we exit soon */
1313 if (EP_IS_LINKED(&epi->rdllink))
1316 list_add_tail(&epi->rdllink, &ep->rdllist);
1320 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1323 if (waitqueue_active(&ep->wq))
1325 if (waitqueue_active(&ep->poll_wait))
1329 write_unlock_irqrestore(&ep->lock, flags);
1331 /* We have to call this outside the lock */
1333 ep_poll_safewake(&psw, &ep->poll_wait);
1339 static int ep_eventpoll_close(struct inode *inode, struct file *file)
1341 struct eventpoll *ep = file->private_data;
1348 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: close() ep=%p\n", current, ep));
1353 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
1355 unsigned int pollflags = 0;
1356 unsigned long flags;
1357 struct eventpoll *ep = file->private_data;
1359 /* Insert inside our poll wait queue */
1360 poll_wait(file, &ep->poll_wait, wait);
1362 /* Check our condition */
1363 read_lock_irqsave(&ep->lock, flags);
1364 if (!list_empty(&ep->rdllist))
1365 pollflags = POLLIN | POLLRDNORM;
1366 read_unlock_irqrestore(&ep->lock, flags);
1373 * Since we have to release the lock during the __copy_to_user() operation and
1374 * during the f_op->poll() call, we try to collect the maximum number of items
1375 * by reducing the irqlock/irqunlock switching rate.
1377 static int ep_collect_ready_items(struct eventpoll *ep, struct list_head *txlist, int maxevents)
1380 unsigned long flags;
1381 struct list_head *lsthead = &ep->rdllist, *lnk;
1384 write_lock_irqsave(&ep->lock, flags);
1386 for (nepi = 0, lnk = lsthead->next; lnk != lsthead && nepi < maxevents;) {
1387 epi = list_entry(lnk, struct epitem, rdllink);
1391 /* If this file is already in the ready list we exit soon */
1392 if (!EP_IS_LINKED(&epi->txlink)) {
1394 * This is initialized in this way so that the default
1395 * behaviour of the reinjecting code will be to push back
1396 * the item inside the ready list.
1398 epi->revents = epi->event.events;
1400 /* Link the ready item into the transfer list */
1401 list_add(&epi->txlink, txlist);
1405 * Unlink the item from the ready list.
1407 EP_LIST_DEL(&epi->rdllink);
1411 write_unlock_irqrestore(&ep->lock, flags);
1418 * This function is called without holding the "ep->lock" since the call to
1419 * __copy_to_user() might sleep, and also f_op->poll() might reenable the IRQ
1420 * because of the way poll() is traditionally implemented in Linux.
1422 static int ep_send_events(struct eventpoll *ep, struct list_head *txlist,
1423 struct epoll_event __user *events)
1426 unsigned int revents;
1427 struct list_head *lnk;
1431 * We can loop without lock because this is a task private list.
1432 * The test done during the collection loop will guarantee us that
1433 * another task will not try to collect this file. Also, items
1434 * cannot vanish during the loop because we are holding "sem".
1436 list_for_each(lnk, txlist) {
1437 epi = list_entry(lnk, struct epitem, txlink);
1440 * Get the ready file event set. We can safely use the file
1441 * because we are holding the "sem" in read and this will
1442 * guarantee that both the file and the item will not vanish.
1444 revents = epi->file->f_op->poll(epi->file, NULL);
1447 * Set the return event set for the current file descriptor.
1448 * Note that only the task task was successfully able to link
1449 * the item to its "txlist" will write this field.
1451 epi->revents = revents & epi->event.events;
1454 if (__put_user(epi->revents,
1455 &events[eventcnt].events) ||
1456 __put_user(epi->event.data,
1457 &events[eventcnt].data))
1459 if (epi->event.events & EPOLLONESHOT)
1460 epi->event.events &= EP_PRIVATE_BITS;
1469 * Walk through the transfer list we collected with ep_collect_ready_items()
1470 * and, if 1) the item is still "alive" 2) its event set is not empty 3) it's
1471 * not already linked, links it to the ready list. Same as above, we are holding
1472 * "sem" so items cannot vanish underneath our nose.
1474 static void ep_reinject_items(struct eventpoll *ep, struct list_head *txlist)
1476 int ricnt = 0, pwake = 0;
1477 unsigned long flags;
1480 write_lock_irqsave(&ep->lock, flags);
1482 while (!list_empty(txlist)) {
1483 epi = list_entry(txlist->next, struct epitem, txlink);
1485 /* Unlink the current item from the transfer list */
1486 EP_LIST_DEL(&epi->txlink);
1489 * If the item is no more linked to the interest set, we don't
1490 * have to push it inside the ready list because the following
1491 * ep_release_epitem() is going to drop it. Also, if the current
1492 * item is set to have an Edge Triggered behaviour, we don't have
1493 * to push it back either.
1495 if (EP_IS_LINKED(&epi->llink) && !(epi->event.events & EPOLLET) &&
1496 (epi->revents & epi->event.events) && !EP_IS_LINKED(&epi->rdllink)) {
1497 list_add_tail(&epi->rdllink, &ep->rdllist);
1504 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1507 if (waitqueue_active(&ep->wq))
1509 if (waitqueue_active(&ep->poll_wait))
1513 write_unlock_irqrestore(&ep->lock, flags);
1515 /* We have to call this outside the lock */
1517 ep_poll_safewake(&psw, &ep->poll_wait);
1522 * Perform the transfer of events to user space.
1524 static int ep_events_transfer(struct eventpoll *ep,
1525 struct epoll_event __user *events, int maxevents)
1528 struct list_head txlist;
1530 INIT_LIST_HEAD(&txlist);
1533 * We need to lock this because we could be hit by
1534 * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL).
1536 down_read(&ep->sem);
1538 /* Collect/extract ready items */
1539 if (ep_collect_ready_items(ep, &txlist, maxevents) > 0) {
1540 /* Build result set in userspace */
1541 eventcnt = ep_send_events(ep, &txlist, events);
1543 /* Reinject ready items into the ready list */
1544 ep_reinject_items(ep, &txlist);
1553 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1554 int maxevents, long timeout)
1557 unsigned long flags;
1562 * Calculate the timeout by checking for the "infinite" value ( -1 )
1563 * and the overflow condition. The passed timeout is in milliseconds,
1564 * that why (t * HZ) / 1000.
1566 jtimeout = timeout == -1 || timeout > (MAX_SCHEDULE_TIMEOUT - 1000) / HZ ?
1567 MAX_SCHEDULE_TIMEOUT: (timeout * HZ + 999) / 1000;
1570 write_lock_irqsave(&ep->lock, flags);
1573 if (list_empty(&ep->rdllist)) {
1575 * We don't have any available event to return to the caller.
1576 * We need to sleep here, and we will be wake up by
1577 * ep_poll_callback() when events will become available.
1579 init_waitqueue_entry(&wait, current);
1580 add_wait_queue(&ep->wq, &wait);
1584 * We don't want to sleep if the ep_poll_callback() sends us
1585 * a wakeup in between. That's why we set the task state
1586 * to TASK_INTERRUPTIBLE before doing the checks.
1588 set_current_state(TASK_INTERRUPTIBLE);
1589 if (!list_empty(&ep->rdllist) || !jtimeout)
1591 if (signal_pending(current)) {
1596 write_unlock_irqrestore(&ep->lock, flags);
1597 jtimeout = schedule_timeout(jtimeout);
1598 write_lock_irqsave(&ep->lock, flags);
1600 remove_wait_queue(&ep->wq, &wait);
1602 set_current_state(TASK_RUNNING);
1605 /* Is it worth to try to dig for events ? */
1606 eavail = !list_empty(&ep->rdllist);
1608 write_unlock_irqrestore(&ep->lock, flags);
1611 * Try to transfer events to user space. In case we get 0 events and
1612 * there's still timeout left over, we go trying again in search of
1615 if (!res && eavail &&
1616 !(res = ep_events_transfer(ep, events, maxevents)) && jtimeout)
1623 static int eventpollfs_delete_dentry(struct dentry *dentry)
1630 static struct inode *ep_eventpoll_inode(void)
1632 int error = -ENOMEM;
1633 struct inode *inode = new_inode(eventpoll_mnt->mnt_sb);
1638 inode->i_fop = &eventpoll_fops;
1641 * Mark the inode dirty from the very beginning,
1642 * that way it will never be moved to the dirty
1643 * list because mark_inode_dirty() will think
1644 * that it already _is_ on the dirty list.
1646 inode->i_state = I_DIRTY;
1647 inode->i_mode = S_IRUSR | S_IWUSR;
1648 inode->i_uid = current->fsuid;
1649 inode->i_gid = current->fsgid;
1650 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1651 inode->i_blksize = PAGE_SIZE;
1655 return ERR_PTR(error);
1659 static struct super_block *
1660 eventpollfs_get_sb(struct file_system_type *fs_type, int flags,
1661 const char *dev_name, void *data)
1663 return get_sb_pseudo(fs_type, "eventpoll:", NULL, EVENTPOLLFS_MAGIC);
1667 static int __init eventpoll_init(void)
1673 /* Initialize the structure used to perform safe poll wait head wake ups */
1674 ep_poll_safewake_init(&psw);
1676 /* Allocates slab cache used to allocate "struct epitem" items */
1677 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1678 0, SLAB_HWCACHE_ALIGN|EPI_SLAB_DEBUG|SLAB_PANIC,
1681 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1682 pwq_cache = kmem_cache_create("eventpoll_pwq",
1683 sizeof(struct eppoll_entry), 0,
1684 EPI_SLAB_DEBUG|SLAB_PANIC, NULL, NULL);
1687 * Register the virtual file system that will be the source of inodes
1688 * for the eventpoll files
1690 error = register_filesystem(&eventpoll_fs_type);
1694 /* Mount the above commented virtual file system */
1695 eventpoll_mnt = kern_mount(&eventpoll_fs_type);
1696 error = PTR_ERR(eventpoll_mnt);
1697 if (IS_ERR(eventpoll_mnt))
1700 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: successfully initialized.\n",
1705 panic("eventpoll_init() failed\n");
1709 static void __exit eventpoll_exit(void)
1711 /* Undo all operations done inside eventpoll_init() */
1712 unregister_filesystem(&eventpoll_fs_type);
1713 mntput(eventpoll_mnt);
1714 kmem_cache_destroy(pwq_cache);
1715 kmem_cache_destroy(epi_cache);
1718 module_init(eventpoll_init);
1719 module_exit(eventpoll_exit);
1721 MODULE_LICENSE("GPL");