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 * This is used to optimize the event transfer to userspace. Since this
152 * is kept on stack, it should be pretty small.
154 #define EP_MAX_BUF_EVENTS 32
159 * Node that is linked into the "wake_task_list" member of the "struct poll_safewake".
160 * It is used to keep track on all tasks that are currently inside the wake_up() code
161 * to 1) short-circuit the one coming from the same task and same wait queue head
162 * ( loop ) 2) allow a maximum number of epoll descriptors inclusion nesting
163 * 3) let go the ones coming from other tasks.
165 struct wake_task_node {
166 struct list_head llink;
168 wait_queue_head_t *wq;
172 * This is used to implement the safe poll wake up avoiding to reenter
173 * the poll callback from inside wake_up().
175 struct poll_safewake {
176 struct list_head wake_task_list;
181 * This structure is stored inside the "private_data" member of the file
182 * structure and rapresent the main data sructure for the eventpoll
186 /* Protect the this structure access */
190 * This semaphore is used to ensure that files are not removed
191 * while epoll is using them. This is read-held during the event
192 * collection loop and it is write-held during the file cleanup
193 * path, the epoll file exit code and the ctl operations.
195 struct rw_semaphore sem;
197 /* Wait queue used by sys_epoll_wait() */
198 wait_queue_head_t wq;
200 /* Wait queue used by file->poll() */
201 wait_queue_head_t poll_wait;
203 /* List of ready file descriptors */
204 struct list_head rdllist;
206 /* Size of the hash */
207 unsigned int hashbits;
209 /* Pages for the "struct epitem" hash */
210 char *hpages[EP_MAX_HPAGES];
213 /* Wait structure used by the poll hooks */
214 struct eppoll_entry {
215 /* List header used to link this structure to the "struct epitem" */
216 struct list_head llink;
218 /* The "base" pointer is set to the container "struct epitem" */
222 * Wait queue item that will be linked to the target file wait
227 /* The wait queue head that linked the "wait" wait queue item */
228 wait_queue_head_t *whead;
232 * Each file descriptor added to the eventpoll interface will
233 * have an entry of this type linked to the hash.
236 /* List header used to link this structure to the eventpoll hash */
237 struct list_head llink;
239 /* List header used to link this structure to the eventpoll ready list */
240 struct list_head rdllink;
242 /* The file descriptor this item refers to */
245 /* Number of active wait queue attached to poll operations */
248 /* List containing poll wait queues */
249 struct list_head pwqlist;
251 /* The "container" of this item */
252 struct eventpoll *ep;
254 /* The file this item refers to */
257 /* The structure that describe the interested events and the source fd */
258 struct epoll_event event;
261 * Used to keep track of the usage count of the structure. This avoids
262 * that the structure will desappear from underneath our processing.
266 /* List header used to link this item to the "struct file" items list */
267 struct list_head fllink;
269 /* List header used to link the item to the transfer list */
270 struct list_head txlink;
273 * This is used during the collection/transfer of events to userspace
274 * to pin items empty events set.
276 unsigned int revents;
279 /* Wrapper struct used by poll queueing */
287 static void ep_poll_safewake_init(struct poll_safewake *psw);
288 static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq);
289 static unsigned int ep_get_hash_bits(unsigned int hintsize);
290 static int ep_getfd(int *efd, struct inode **einode, struct file **efile);
291 static int ep_alloc_pages(char **pages, int numpages);
292 static int ep_free_pages(char **pages, int numpages);
293 static int ep_file_init(struct file *file, unsigned int hashbits);
294 static unsigned int ep_hash_index(struct eventpoll *ep, struct file *file,
296 static struct list_head *ep_hash_entry(struct eventpoll *ep,
298 static int ep_init(struct eventpoll *ep, unsigned int hashbits);
299 static void ep_free(struct eventpoll *ep);
300 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd);
301 static void ep_use_epitem(struct epitem *epi);
302 static void ep_release_epitem(struct epitem *epi);
303 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
305 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
306 struct file *tfile, int fd);
307 static int ep_modify(struct eventpoll *ep, struct epitem *epi,
308 struct epoll_event *event);
309 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi);
310 static int ep_unlink(struct eventpoll *ep, struct epitem *epi);
311 static int ep_remove(struct eventpoll *ep, struct epitem *epi);
312 static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync);
313 static int ep_eventpoll_close(struct inode *inode, struct file *file);
314 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait);
315 static int ep_collect_ready_items(struct eventpoll *ep,
316 struct list_head *txlist, int maxevents);
317 static int ep_send_events(struct eventpoll *ep, struct list_head *txlist,
318 struct epoll_event __user *events);
319 static void ep_reinject_items(struct eventpoll *ep, struct list_head *txlist);
320 static int ep_events_transfer(struct eventpoll *ep,
321 struct epoll_event __user *events,
323 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
324 int maxevents, long timeout);
325 static int eventpollfs_delete_dentry(struct dentry *dentry);
326 static struct inode *ep_eventpoll_inode(void);
327 static struct super_block *eventpollfs_get_sb(struct file_system_type *fs_type,
328 int flags, const char *dev_name,
332 * This semaphore is used to serialize ep_free() and eventpoll_release_file().
334 struct semaphore epsem;
336 /* Safe wake up implementation */
337 static struct poll_safewake psw;
339 /* Slab cache used to allocate "struct epitem" */
340 static kmem_cache_t *epi_cache;
342 /* Slab cache used to allocate "struct eppoll_entry" */
343 static kmem_cache_t *pwq_cache;
345 /* Virtual fs used to allocate inodes for eventpoll files */
346 static struct vfsmount *eventpoll_mnt;
348 /* File callbacks that implement the eventpoll file behaviour */
349 static struct file_operations eventpoll_fops = {
350 .release = ep_eventpoll_close,
351 .poll = ep_eventpoll_poll
355 * This is used to register the virtual file system from where
356 * eventpoll inodes are allocated.
358 static struct file_system_type eventpoll_fs_type = {
359 .name = "eventpollfs",
360 .get_sb = eventpollfs_get_sb,
361 .kill_sb = kill_anon_super,
364 /* Very basic directory entry operations for the eventpoll virtual file system */
365 static struct dentry_operations eventpollfs_dentry_operations = {
366 .d_delete = eventpollfs_delete_dentry,
371 /* Initialize the poll safe wake up structure */
372 static void ep_poll_safewake_init(struct poll_safewake *psw)
375 INIT_LIST_HEAD(&psw->wake_task_list);
376 spin_lock_init(&psw->lock);
381 * Perform a safe wake up of the poll wait list. The problem is that
382 * with the new callback'd wake up system, it is possible that the
383 * poll callback is reentered from inside the call to wake_up() done
384 * on the poll wait queue head. The rule is that we cannot reenter the
385 * wake up code from the same task more than EP_MAX_POLLWAKE_NESTS times,
386 * and we cannot reenter the same wait queue head at all. This will
387 * enable to have a hierarchy of epoll file descriptor of no more than
388 * EP_MAX_POLLWAKE_NESTS deep. We need the irq version of the spin lock
389 * because this one gets called by the poll callback, that in turn is called
390 * from inside a wake_up(), that might be called from irq context.
392 static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq)
396 task_t *this_task = current;
397 struct list_head *lsthead = &psw->wake_task_list, *lnk;
398 struct wake_task_node *tncur;
399 struct wake_task_node tnode;
401 spin_lock_irqsave(&psw->lock, flags);
403 /* Try to see if the current task is already inside this wakeup call */
404 list_for_each(lnk, lsthead) {
405 tncur = list_entry(lnk, struct wake_task_node, llink);
407 if (tncur->wq == wq ||
408 (tncur->task == this_task && ++wake_nests > EP_MAX_POLLWAKE_NESTS)) {
410 * Ops ... loop detected or maximum nest level reached.
411 * We abort this wake by breaking the cycle itself.
413 spin_unlock_irqrestore(&psw->lock, flags);
418 /* Add the current task to the list */
419 tnode.task = this_task;
421 list_add(&tnode.llink, lsthead);
423 spin_unlock_irqrestore(&psw->lock, flags);
425 /* Do really wake up now */
428 /* Remove the current task from the list */
429 spin_lock_irqsave(&psw->lock, flags);
430 list_del(&tnode.llink);
431 spin_unlock_irqrestore(&psw->lock, flags);
436 * Calculate the size of the hash in bits. The returned size will be
437 * bounded between EP_MIN_HASH_BITS and EP_MAX_HASH_BITS.
439 static unsigned int ep_get_hash_bits(unsigned int hintsize)
443 for (i = 0, val = 1; val < hintsize && i < EP_MAX_HASH_BITS; i++, val <<= 1);
444 return i < EP_MIN_HASH_BITS ? EP_MIN_HASH_BITS: i;
448 /* Used to initialize the epoll bits inside the "struct file" */
449 void eventpoll_init_file(struct file *file)
452 INIT_LIST_HEAD(&file->f_ep_links);
453 spin_lock_init(&file->f_ep_lock);
458 * This is called from eventpoll_release() to unlink files from the eventpoll
459 * interface. We need to have this facility to cleanup correctly files that are
460 * closed without being removed from the eventpoll interface.
462 void eventpoll_release_file(struct file *file)
464 struct list_head *lsthead = &file->f_ep_links;
465 struct eventpoll *ep;
469 * We don't want to get "file->f_ep_lock" because it is not
470 * necessary. It is not necessary because we're in the "struct file"
471 * cleanup path, and this means that noone is using this file anymore.
472 * The only hit might come from ep_free() but by holding the semaphore
473 * will correctly serialize the operation. We do need to acquire
474 * "ep->sem" after "epsem" because ep_remove() requires it when called
475 * from anywhere but ep_free().
479 while (!list_empty(lsthead)) {
480 epi = list_entry(lsthead->next, struct epitem, fllink);
483 EP_LIST_DEL(&epi->fllink);
484 down_write(&ep->sem);
494 * It opens an eventpoll file descriptor by suggesting a storage of "size"
495 * file descriptors. The size parameter is just an hint about how to size
496 * data structures. It won't prevent the user to store more than "size"
497 * file descriptors inside the epoll interface. It is the kernel part of
498 * the userspace epoll_create(2).
500 asmlinkage long sys_epoll_create(int size)
503 unsigned int hashbits;
507 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d)\n",
510 /* Correctly size the hash */
511 hashbits = ep_get_hash_bits((unsigned int) size);
514 * Creates all the items needed to setup an eventpoll file. That is,
515 * a file structure, and inode and a free file descriptor.
517 error = ep_getfd(&fd, &inode, &file);
521 /* Setup the file internal data structure ( "struct eventpoll" ) */
522 error = ep_file_init(file, hashbits);
527 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
535 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
536 current, size, error));
542 * The following function implements the controller interface for
543 * the eventpoll file that enables the insertion/removal/change of
544 * file descriptors inside the interest set. It represents
545 * the kernel part of the user space epoll_ctl(2).
548 sys_epoll_ctl(int epfd, int op, int fd, struct epoll_event __user *event)
551 struct file *file, *tfile;
552 struct eventpoll *ep;
554 struct epoll_event epds;
556 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p)\n",
557 current, epfd, op, fd, event));
560 if (copy_from_user(&epds, event, sizeof(struct epoll_event)))
563 /* Get the "struct file *" for the eventpoll file */
569 /* Get the "struct file *" for the target file */
574 /* The target file descriptor must support poll */
576 if (!tfile->f_op || !tfile->f_op->poll)
580 * We have to check that the file structure underneath the file descriptor
581 * the user passed to us _is_ an eventpoll file. And also we do not permit
582 * adding an epoll file descriptor inside itself.
585 if (file == tfile || !IS_FILE_EPOLL(file))
589 * At this point it is safe to assume that the "private_data" contains
590 * our own data structure.
592 ep = file->private_data;
594 down_write(&ep->sem);
596 /* Try to lookup the file inside our hash table */
597 epi = ep_find(ep, tfile, fd);
603 epds.events |= POLLERR | POLLHUP;
605 error = ep_insert(ep, &epds, tfile, fd);
611 error = ep_remove(ep, epi);
617 epds.events |= POLLERR | POLLHUP;
618 error = ep_modify(ep, epi, &epds);
625 * The function ep_find() increments the usage count of the structure
626 * so, if this is not NULL, we need to release it.
629 ep_release_epitem(epi);
638 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p) = %d\n",
639 current, epfd, op, fd, event, error));
646 * Implement the event wait interface for the eventpoll file. It is the kernel
647 * part of the user space epoll_wait(2).
649 asmlinkage long sys_epoll_wait(int epfd, struct epoll_event __user *events,
650 int maxevents, int timeout)
654 struct eventpoll *ep;
656 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d)\n",
657 current, epfd, events, maxevents, timeout));
659 /* The maximum number of event must be greater than zero */
663 /* Verify that the area passed by the user is writeable */
664 if ((error = verify_area(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))))
667 /* Get the "struct file *" for the eventpoll file */
674 * We have to check that the file structure underneath the fd
675 * the user passed to us _is_ an eventpoll file.
678 if (!IS_FILE_EPOLL(file))
682 * At this point it is safe to assume that the "private_data" contains
683 * our own data structure.
685 ep = file->private_data;
687 /* Time to fish for events ... */
688 error = ep_poll(ep, events, maxevents, timeout);
693 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d) = %d\n",
694 current, epfd, events, maxevents, timeout, error));
701 * Creates the file descriptor to be used by the epoll interface.
703 static int ep_getfd(int *efd, struct inode **einode, struct file **efile)
707 struct dentry *dentry;
712 /* Get an ready to use file */
714 file = get_empty_filp();
718 /* Allocates an inode from the eventpoll file system */
719 inode = ep_eventpoll_inode();
720 error = PTR_ERR(inode);
724 /* Allocates a free descriptor to plug the file onto */
725 error = get_unused_fd();
731 * Link the inode to a directory entry by creating a unique name
732 * using the inode number.
735 sprintf(name, "[%lu]", inode->i_ino);
737 this.len = strlen(name);
738 this.hash = inode->i_ino;
739 dentry = d_alloc(eventpoll_mnt->mnt_sb->s_root, &this);
742 dentry->d_op = &eventpollfs_dentry_operations;
743 d_add(dentry, inode);
744 file->f_vfsmnt = mntget(eventpoll_mnt);
745 file->f_dentry = dget(dentry);
746 file->f_mapping = inode->i_mapping;
749 file->f_flags = O_RDONLY;
750 file->f_op = &eventpoll_fops;
751 file->f_mode = FMODE_READ;
753 file->private_data = NULL;
755 /* Install the new setup file into the allocated fd. */
756 fd_install(fd, file);
774 static int ep_alloc_pages(char **pages, int numpages)
778 for (i = 0; i < numpages; i++) {
779 pages[i] = (char *) __get_free_pages(GFP_KERNEL, 0);
781 for (--i; i >= 0; i--) {
782 ClearPageReserved(virt_to_page(pages[i]));
783 free_pages((unsigned long) pages[i], 0);
787 SetPageReserved(virt_to_page(pages[i]));
793 static int ep_free_pages(char **pages, int numpages)
797 for (i = 0; i < numpages; i++) {
798 ClearPageReserved(virt_to_page(pages[i]));
799 free_pages((unsigned long) pages[i], 0);
805 static int ep_file_init(struct file *file, unsigned int hashbits)
808 struct eventpoll *ep;
810 if (!(ep = kmalloc(sizeof(struct eventpoll), GFP_KERNEL)))
813 memset(ep, 0, sizeof(*ep));
815 error = ep_init(ep, hashbits);
821 file->private_data = ep;
823 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_file_init() ep=%p\n",
830 * Calculate the index of the hash relative to "file".
832 static unsigned int ep_hash_index(struct eventpoll *ep, struct file *file, int fd)
834 unsigned long ptr = (unsigned long) file ^ (fd << ep->hashbits);
836 return (unsigned int) hash_ptr((void *) ptr, ep->hashbits);
841 * Returns the hash entry ( struct list_head * ) of the passed index.
843 static struct list_head *ep_hash_entry(struct eventpoll *ep, unsigned int index)
846 return (struct list_head *) (ep->hpages[index / EP_HENTRY_X_PAGE] +
847 (index % EP_HENTRY_X_PAGE) * sizeof(struct list_head));
851 static int ep_init(struct eventpoll *ep, unsigned int hashbits)
854 unsigned int i, hsize;
856 rwlock_init(&ep->lock);
857 init_rwsem(&ep->sem);
858 init_waitqueue_head(&ep->wq);
859 init_waitqueue_head(&ep->poll_wait);
860 INIT_LIST_HEAD(&ep->rdllist);
862 /* Hash allocation and setup */
863 ep->hashbits = hashbits;
864 error = ep_alloc_pages(ep->hpages, EP_HASH_PAGES(ep->hashbits));
868 /* Initialize hash buckets */
869 for (i = 0, hsize = 1 << hashbits; i < hsize; i++)
870 INIT_LIST_HEAD(ep_hash_entry(ep, i));
878 static void ep_free(struct eventpoll *ep)
880 unsigned int i, hsize;
881 struct list_head *lsthead, *lnk;
884 /* We need to release all tasks waiting for these file */
885 if (waitqueue_active(&ep->poll_wait))
886 ep_poll_safewake(&psw, &ep->poll_wait);
889 * We need to lock this because we could be hit by
890 * eventpoll_release_file() while we're freeing the "struct eventpoll".
891 * We do not need to hold "ep->sem" here because the epoll file
892 * is on the way to be removed and no one has references to it
893 * anymore. The only hit might come from eventpoll_release_file() but
894 * holding "epsem" is sufficent here.
899 * Walks through the whole hash by unregistering poll callbacks.
901 for (i = 0, hsize = 1 << ep->hashbits; i < hsize; i++) {
902 lsthead = ep_hash_entry(ep, i);
904 list_for_each(lnk, lsthead) {
905 epi = list_entry(lnk, struct epitem, llink);
907 ep_unregister_pollwait(ep, epi);
912 * Walks through the whole hash by freeing each "struct epitem". At this
913 * point we are sure no poll callbacks will be lingering around, and also by
914 * write-holding "sem" we can be sure that no file cleanup code will hit
915 * us during this operation. So we can avoid the lock on "ep->lock".
917 for (i = 0, hsize = 1 << ep->hashbits; i < hsize; i++) {
918 lsthead = ep_hash_entry(ep, i);
920 while (!list_empty(lsthead)) {
921 epi = list_entry(lsthead->next, struct epitem, llink);
929 /* Free hash pages */
930 ep_free_pages(ep->hpages, EP_HASH_PAGES(ep->hashbits));
935 * Search the file inside the eventpoll hash. It add usage count to
936 * the returned item, so the caller must call ep_release_epitem()
937 * after finished using the "struct epitem".
939 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
942 struct list_head *lsthead, *lnk;
943 struct epitem *epi = NULL;
945 read_lock_irqsave(&ep->lock, flags);
947 lsthead = ep_hash_entry(ep, ep_hash_index(ep, file, fd));
948 list_for_each(lnk, lsthead) {
949 epi = list_entry(lnk, struct epitem, llink);
951 if (epi->file == file && epi->fd == fd) {
958 read_unlock_irqrestore(&ep->lock, flags);
960 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_find(%p) -> %p\n",
961 current, file, epi));
968 * Increment the usage count of the "struct epitem" making it sure
969 * that the user will have a valid pointer to reference.
971 static void ep_use_epitem(struct epitem *epi)
974 atomic_inc(&epi->usecnt);
979 * Decrement ( release ) the usage count by signaling that the user
980 * has finished using the structure. It might lead to freeing the
981 * structure itself if the count goes to zero.
983 static void ep_release_epitem(struct epitem *epi)
986 if (atomic_dec_and_test(&epi->usecnt))
992 * This is the callback that is used to add our wait queue to the
993 * target file wakeup lists.
995 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
998 struct epitem *epi = EP_ITEM_FROM_EPQUEUE(pt);
999 struct eppoll_entry *pwq;
1001 if (epi->nwait >= 0 && (pwq = PWQ_MEM_ALLOC())) {
1002 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1005 add_wait_queue(whead, &pwq->wait);
1006 list_add_tail(&pwq->llink, &epi->pwqlist);
1009 /* We have to signal that an error occurred */
1015 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
1016 struct file *tfile, int fd)
1018 int error, revents, pwake = 0;
1019 unsigned long flags;
1021 struct ep_pqueue epq;
1024 if (!(epi = EPI_MEM_ALLOC()))
1027 /* Item initialization follow here ... */
1028 INIT_LIST_HEAD(&epi->llink);
1029 INIT_LIST_HEAD(&epi->rdllink);
1030 INIT_LIST_HEAD(&epi->fllink);
1031 INIT_LIST_HEAD(&epi->txlink);
1032 INIT_LIST_HEAD(&epi->pwqlist);
1036 epi->event = *event;
1037 atomic_set(&epi->usecnt, 1);
1040 /* Initialize the poll table using the queue callback */
1042 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1045 * Attach the item to the poll hooks and get current event bits.
1046 * We can safely use the file* here because its usage count has
1047 * been increased by the caller of this function.
1049 revents = tfile->f_op->poll(tfile, &epq.pt);
1052 * We have to check if something went wrong during the poll wait queue
1053 * install process. Namely an allocation for a wait queue failed due
1054 * high memory pressure.
1059 /* Add the current item to the list of active epoll hook for this file */
1060 spin_lock(&tfile->f_ep_lock);
1061 list_add_tail(&epi->fllink, &tfile->f_ep_links);
1062 spin_unlock(&tfile->f_ep_lock);
1064 /* We have to drop the new item inside our item list to keep track of it */
1065 write_lock_irqsave(&ep->lock, flags);
1067 /* Add the current item to the hash table */
1068 list_add(&epi->llink, ep_hash_entry(ep, ep_hash_index(ep, tfile, fd)));
1070 /* If the file is already "ready" we drop it inside the ready list */
1071 if ((revents & event->events) && !EP_IS_LINKED(&epi->rdllink)) {
1072 list_add_tail(&epi->rdllink, &ep->rdllist);
1074 /* Notify waiting tasks that events are available */
1075 if (waitqueue_active(&ep->wq))
1077 if (waitqueue_active(&ep->poll_wait))
1081 write_unlock_irqrestore(&ep->lock, flags);
1083 /* We have to call this outside the lock */
1085 ep_poll_safewake(&psw, &ep->poll_wait);
1087 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_insert(%p, %p, %d)\n",
1088 current, ep, tfile, fd));
1093 ep_unregister_pollwait(ep, epi);
1096 * We need to do this because an event could have been arrived on some
1097 * allocated wait queue.
1099 write_lock_irqsave(&ep->lock, flags);
1100 if (EP_IS_LINKED(&epi->rdllink))
1101 EP_LIST_DEL(&epi->rdllink);
1102 write_unlock_irqrestore(&ep->lock, flags);
1111 * Modify the interest event mask by dropping an event if the new mask
1112 * has a match in the current file status.
1114 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1117 unsigned int revents;
1118 unsigned long flags;
1121 * Set the new event interest mask before calling f_op->poll(), otherwise
1122 * a potential race might occur. In fact if we do this operation inside
1123 * the lock, an event might happen between the f_op->poll() call and the
1124 * new event set registering.
1126 epi->event.events = event->events;
1129 * Get current event bits. We can safely use the file* here because
1130 * its usage count has been increased by the caller of this function.
1132 revents = epi->file->f_op->poll(epi->file, NULL);
1134 write_lock_irqsave(&ep->lock, flags);
1136 /* Copy the data member from inside the lock */
1137 epi->event.data = event->data;
1140 * If the item is not linked to the hash it means that it's on its
1141 * way toward the removal. Do nothing in this case.
1143 if (EP_IS_LINKED(&epi->llink)) {
1145 * If the item is "hot" and it is not registered inside the ready
1146 * list, push it inside. If the item is not "hot" and it is currently
1147 * registered inside the ready list, unlink it.
1149 if (revents & event->events) {
1150 if (!EP_IS_LINKED(&epi->rdllink)) {
1151 list_add_tail(&epi->rdllink, &ep->rdllist);
1153 /* Notify waiting tasks that events are available */
1154 if (waitqueue_active(&ep->wq))
1156 if (waitqueue_active(&ep->poll_wait))
1162 write_unlock_irqrestore(&ep->lock, flags);
1164 /* We have to call this outside the lock */
1166 ep_poll_safewake(&psw, &ep->poll_wait);
1173 * This function unregister poll callbacks from the associated file descriptor.
1174 * Since this must be called without holding "ep->lock" the atomic exchange trick
1175 * will protect us from multiple unregister.
1177 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
1180 struct list_head *lsthead = &epi->pwqlist;
1181 struct eppoll_entry *pwq;
1183 /* This is called without locks, so we need the atomic exchange */
1184 nwait = xchg(&epi->nwait, 0);
1187 while (!list_empty(lsthead)) {
1188 pwq = list_entry(lsthead->next, struct eppoll_entry, llink);
1190 EP_LIST_DEL(&pwq->llink);
1191 remove_wait_queue(pwq->whead, &pwq->wait);
1199 * Unlink the "struct epitem" from all places it might have been hooked up.
1200 * This function must be called with write IRQ lock on "ep->lock".
1202 static int ep_unlink(struct eventpoll *ep, struct epitem *epi)
1207 * It can happen that this one is called for an item already unlinked.
1208 * The check protect us from doing a double unlink ( crash ).
1211 if (!EP_IS_LINKED(&epi->llink))
1215 * Clear the event mask for the unlinked item. This will avoid item
1216 * notifications to be sent after the unlink operation from inside
1217 * the kernel->userspace event transfer loop.
1219 epi->event.events = 0;
1222 * At this point is safe to do the job, unlink the item from our list.
1223 * This operation togheter with the above check closes the door to
1226 EP_LIST_DEL(&epi->llink);
1229 * If the item we are going to remove is inside the ready file descriptors
1230 * we want to remove it from this list to avoid stale events.
1232 if (EP_IS_LINKED(&epi->rdllink))
1233 EP_LIST_DEL(&epi->rdllink);
1238 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_unlink(%p, %p) = %d\n",
1239 current, ep, epi->file, error));
1246 * Removes a "struct epitem" from the eventpoll hash and deallocates
1247 * all the associated resources.
1249 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
1252 unsigned long flags;
1253 struct file *file = epi->file;
1256 * Removes poll wait queue hooks. We _have_ to do this without holding
1257 * the "ep->lock" otherwise a deadlock might occur. This because of the
1258 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
1259 * queue head lock when unregistering the wait queue. The wakeup callback
1260 * will run by holding the wait queue head lock and will call our callback
1261 * that will try to get "ep->lock".
1263 ep_unregister_pollwait(ep, epi);
1265 /* Remove the current item from the list of epoll hooks */
1266 spin_lock(&file->f_ep_lock);
1267 if (EP_IS_LINKED(&epi->fllink))
1268 EP_LIST_DEL(&epi->fllink);
1269 spin_unlock(&file->f_ep_lock);
1271 /* We need to acquire the write IRQ lock before calling ep_unlink() */
1272 write_lock_irqsave(&ep->lock, flags);
1274 /* Really unlink the item from the hash */
1275 error = ep_unlink(ep, epi);
1277 write_unlock_irqrestore(&ep->lock, flags);
1282 /* At this point it is safe to free the eventpoll item */
1283 ep_release_epitem(epi);
1287 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_remove(%p, %p) = %d\n",
1288 current, ep, file, error));
1295 * This is the callback that is passed to the wait queue wakeup
1296 * machanism. It is called by the stored file descriptors when they
1297 * have events to report.
1299 static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync)
1302 unsigned long flags;
1303 struct epitem *epi = EP_ITEM_FROM_WAIT(wait);
1304 struct eventpoll *ep = epi->ep;
1306 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: poll_callback(%p) epi=%p ep=%p\n",
1307 current, epi->file, epi, ep));
1309 write_lock_irqsave(&ep->lock, flags);
1312 * If the event mask does not contain any poll(2) event, we consider the
1313 * descriptor to be disabled. This condition is likely the effect of the
1314 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1315 * until the next EPOLL_CTL_MOD will be issued.
1317 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1320 /* If this file is already in the ready list we exit soon */
1321 if (EP_IS_LINKED(&epi->rdllink))
1324 list_add_tail(&epi->rdllink, &ep->rdllist);
1328 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1331 if (waitqueue_active(&ep->wq))
1333 if (waitqueue_active(&ep->poll_wait))
1337 write_unlock_irqrestore(&ep->lock, flags);
1339 /* We have to call this outside the lock */
1341 ep_poll_safewake(&psw, &ep->poll_wait);
1347 static int ep_eventpoll_close(struct inode *inode, struct file *file)
1349 struct eventpoll *ep = file->private_data;
1356 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: close() ep=%p\n", current, ep));
1361 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
1363 unsigned int pollflags = 0;
1364 unsigned long flags;
1365 struct eventpoll *ep = file->private_data;
1367 /* Insert inside our poll wait queue */
1368 poll_wait(file, &ep->poll_wait, wait);
1370 /* Check our condition */
1371 read_lock_irqsave(&ep->lock, flags);
1372 if (!list_empty(&ep->rdllist))
1373 pollflags = POLLIN | POLLRDNORM;
1374 read_unlock_irqrestore(&ep->lock, flags);
1381 * Since we have to release the lock during the __copy_to_user() operation and
1382 * during the f_op->poll() call, we try to collect the maximum number of items
1383 * by reducing the irqlock/irqunlock switching rate.
1385 static int ep_collect_ready_items(struct eventpoll *ep, struct list_head *txlist, int maxevents)
1388 unsigned long flags;
1389 struct list_head *lsthead = &ep->rdllist, *lnk;
1392 write_lock_irqsave(&ep->lock, flags);
1394 for (nepi = 0, lnk = lsthead->next; lnk != lsthead && nepi < maxevents;) {
1395 epi = list_entry(lnk, struct epitem, rdllink);
1399 /* If this file is already in the ready list we exit soon */
1400 if (!EP_IS_LINKED(&epi->txlink)) {
1402 * This is initialized in this way so that the default
1403 * behaviour of the reinjecting code will be to push back
1404 * the item inside the ready list.
1406 epi->revents = epi->event.events;
1408 /* Link the ready item into the transfer list */
1409 list_add(&epi->txlink, txlist);
1413 * Unlink the item from the ready list.
1415 EP_LIST_DEL(&epi->rdllink);
1419 write_unlock_irqrestore(&ep->lock, flags);
1426 * This function is called without holding the "ep->lock" since the call to
1427 * __copy_to_user() might sleep, and also f_op->poll() might reenable the IRQ
1428 * because of the way poll() is traditionally implemented in Linux.
1430 static int ep_send_events(struct eventpoll *ep, struct list_head *txlist,
1431 struct epoll_event __user *events)
1433 int eventcnt = 0, eventbuf = 0;
1434 unsigned int revents;
1435 struct list_head *lnk;
1437 struct epoll_event event[EP_MAX_BUF_EVENTS];
1440 * We can loop without lock because this is a task private list.
1441 * The test done during the collection loop will guarantee us that
1442 * another task will not try to collect this file. Also, items
1443 * cannot vanish during the loop because we are holding "sem".
1445 list_for_each(lnk, txlist) {
1446 epi = list_entry(lnk, struct epitem, txlink);
1449 * Get the ready file event set. We can safely use the file
1450 * because we are holding the "sem" in read and this will
1451 * guarantee that both the file and the item will not vanish.
1453 revents = epi->file->f_op->poll(epi->file, NULL);
1456 * Set the return event set for the current file descriptor.
1457 * Note that only the task task was successfully able to link
1458 * the item to its "txlist" will write this field.
1460 epi->revents = revents & epi->event.events;
1463 event[eventbuf] = epi->event;
1464 event[eventbuf].events &= revents;
1466 if (eventbuf == EP_MAX_BUF_EVENTS) {
1467 if (__copy_to_user(&events[eventcnt], event,
1468 eventbuf * sizeof(struct epoll_event)))
1470 eventcnt += eventbuf;
1473 if (epi->event.events & EPOLLONESHOT)
1474 epi->event.events &= EP_PRIVATE_BITS;
1479 if (__copy_to_user(&events[eventcnt], event,
1480 eventbuf * sizeof(struct epoll_event)))
1482 eventcnt += eventbuf;
1490 * Walk through the transfer list we collected with ep_collect_ready_items()
1491 * and, if 1) the item is still "alive" 2) its event set is not empty 3) it's
1492 * not already linked, links it to the ready list. Same as above, we are holding
1493 * "sem" so items cannot vanish underneath our nose.
1495 static void ep_reinject_items(struct eventpoll *ep, struct list_head *txlist)
1497 int ricnt = 0, pwake = 0;
1498 unsigned long flags;
1501 write_lock_irqsave(&ep->lock, flags);
1503 while (!list_empty(txlist)) {
1504 epi = list_entry(txlist->next, struct epitem, txlink);
1506 /* Unlink the current item from the transfer list */
1507 EP_LIST_DEL(&epi->txlink);
1510 * If the item is no more linked to the interest set, we don't
1511 * have to push it inside the ready list because the following
1512 * ep_release_epitem() is going to drop it. Also, if the current
1513 * item is set to have an Edge Triggered behaviour, we don't have
1514 * to push it back either.
1516 if (EP_IS_LINKED(&epi->llink) && !(epi->event.events & EPOLLET) &&
1517 (epi->revents & epi->event.events) && !EP_IS_LINKED(&epi->rdllink)) {
1518 list_add_tail(&epi->rdllink, &ep->rdllist);
1525 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1528 if (waitqueue_active(&ep->wq))
1530 if (waitqueue_active(&ep->poll_wait))
1534 write_unlock_irqrestore(&ep->lock, flags);
1536 /* We have to call this outside the lock */
1538 ep_poll_safewake(&psw, &ep->poll_wait);
1543 * Perform the transfer of events to user space.
1545 static int ep_events_transfer(struct eventpoll *ep,
1546 struct epoll_event __user *events, int maxevents)
1549 struct list_head txlist;
1551 INIT_LIST_HEAD(&txlist);
1554 * We need to lock this because we could be hit by
1555 * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL).
1557 down_read(&ep->sem);
1559 /* Collect/extract ready items */
1560 if (ep_collect_ready_items(ep, &txlist, maxevents) > 0) {
1561 /* Build result set in userspace */
1562 eventcnt = ep_send_events(ep, &txlist, events);
1564 /* Reinject ready items into the ready list */
1565 ep_reinject_items(ep, &txlist);
1574 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1575 int maxevents, long timeout)
1578 unsigned long flags;
1583 * Calculate the timeout by checking for the "infinite" value ( -1 )
1584 * and the overflow condition. The passed timeout is in milliseconds,
1585 * that why (t * HZ) / 1000.
1587 jtimeout = timeout == -1 || timeout > (MAX_SCHEDULE_TIMEOUT - 1000) / HZ ?
1588 MAX_SCHEDULE_TIMEOUT: (timeout * HZ + 999) / 1000;
1591 write_lock_irqsave(&ep->lock, flags);
1594 if (list_empty(&ep->rdllist)) {
1596 * We don't have any available event to return to the caller.
1597 * We need to sleep here, and we will be wake up by
1598 * ep_poll_callback() when events will become available.
1600 init_waitqueue_entry(&wait, current);
1601 add_wait_queue(&ep->wq, &wait);
1605 * We don't want to sleep if the ep_poll_callback() sends us
1606 * a wakeup in between. That's why we set the task state
1607 * to TASK_INTERRUPTIBLE before doing the checks.
1609 set_current_state(TASK_INTERRUPTIBLE);
1610 if (!list_empty(&ep->rdllist) || !jtimeout)
1612 if (signal_pending(current)) {
1617 write_unlock_irqrestore(&ep->lock, flags);
1618 jtimeout = schedule_timeout(jtimeout);
1619 write_lock_irqsave(&ep->lock, flags);
1621 remove_wait_queue(&ep->wq, &wait);
1623 set_current_state(TASK_RUNNING);
1626 /* Is it worth to try to dig for events ? */
1627 eavail = !list_empty(&ep->rdllist);
1629 write_unlock_irqrestore(&ep->lock, flags);
1632 * Try to transfer events to user space. In case we get 0 events and
1633 * there's still timeout left over, we go trying again in search of
1636 if (!res && eavail &&
1637 !(res = ep_events_transfer(ep, events, maxevents)) && jtimeout)
1644 static int eventpollfs_delete_dentry(struct dentry *dentry)
1651 static struct inode *ep_eventpoll_inode(void)
1653 int error = -ENOMEM;
1654 struct inode *inode = new_inode(eventpoll_mnt->mnt_sb);
1659 inode->i_fop = &eventpoll_fops;
1662 * Mark the inode dirty from the very beginning,
1663 * that way it will never be moved to the dirty
1664 * list because mark_inode_dirty() will think
1665 * that it already _is_ on the dirty list.
1667 inode->i_state = I_DIRTY;
1668 inode->i_mode = S_IRUSR | S_IWUSR;
1669 inode->i_uid = current->fsuid;
1670 inode->i_gid = current->fsgid;
1671 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1672 inode->i_blksize = PAGE_SIZE;
1676 return ERR_PTR(error);
1680 static struct super_block *
1681 eventpollfs_get_sb(struct file_system_type *fs_type, int flags,
1682 const char *dev_name, void *data)
1684 return get_sb_pseudo(fs_type, "eventpoll:", NULL, EVENTPOLLFS_MAGIC);
1688 static int __init eventpoll_init(void)
1694 /* Initialize the structure used to perform safe poll wait head wake ups */
1695 ep_poll_safewake_init(&psw);
1697 /* Allocates slab cache used to allocate "struct epitem" items */
1699 epi_cache = kmem_cache_create("eventpoll_epi",
1700 sizeof(struct epitem),
1702 SLAB_HWCACHE_ALIGN | EPI_SLAB_DEBUG, NULL, NULL);
1706 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1708 pwq_cache = kmem_cache_create("eventpoll_pwq",
1709 sizeof(struct eppoll_entry),
1711 EPI_SLAB_DEBUG, NULL, NULL);
1716 * Register the virtual file system that will be the source of inodes
1717 * for the eventpoll files
1719 error = register_filesystem(&eventpoll_fs_type);
1723 /* Mount the above commented virtual file system */
1724 eventpoll_mnt = kern_mount(&eventpoll_fs_type);
1725 error = PTR_ERR(eventpoll_mnt);
1726 if (IS_ERR(eventpoll_mnt))
1729 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: successfully initialized.\n", current));
1734 unregister_filesystem(&eventpoll_fs_type);
1736 kmem_cache_destroy(pwq_cache);
1738 kmem_cache_destroy(epi_cache);
1745 static void __exit eventpoll_exit(void)
1747 /* Undo all operations done inside eventpoll_init() */
1748 unregister_filesystem(&eventpoll_fs_type);
1749 mntput(eventpoll_mnt);
1750 kmem_cache_destroy(pwq_cache);
1751 kmem_cache_destroy(epi_cache);
1754 module_init(eventpoll_init);
1755 module_exit(eventpoll_exit);
1757 MODULE_LICENSE("GPL");