2 * Fast Userspace Mutexes (which I call "Futexes!").
3 * (C) Rusty Russell, IBM 2002
5 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
8 * Removed page pinning, fix privately mapped COW pages and other cleanups
9 * (C) Copyright 2003, 2004 Jamie Lokier
11 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
12 * enough at me, Linus for the original (flawed) idea, Matthew
13 * Kirkwood for proof-of-concept implementation.
15 * "The futexes are also cursed."
16 * "But they come in a choice of three flavours!"
18 * This program is free software; you can redistribute it and/or modify
19 * it under the terms of the GNU General Public License as published by
20 * the Free Software Foundation; either version 2 of the License, or
21 * (at your option) any later version.
23 * This program is distributed in the hope that it will be useful,
24 * but WITHOUT ANY WARRANTY; without even the implied warranty of
25 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
26 * GNU General Public License for more details.
28 * You should have received a copy of the GNU General Public License
29 * along with this program; if not, write to the Free Software
30 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
32 #include <linux/slab.h>
33 #include <linux/poll.h>
35 #include <linux/file.h>
36 #include <linux/jhash.h>
37 #include <linux/init.h>
38 #include <linux/futex.h>
39 #include <linux/mount.h>
40 #include <linux/pagemap.h>
41 #include <linux/syscalls.h>
42 #include <linux/signal.h>
43 #include <asm/futex.h>
45 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
48 * Futexes are matched on equal values of this key.
49 * The key type depends on whether it's a shared or private mapping.
50 * Don't rearrange members without looking at hash_futex().
52 * offset is aligned to a multiple of sizeof(u32) (== 4) by definition.
53 * We set bit 0 to indicate if it's an inode-based key.
74 * We use this hashed waitqueue instead of a normal wait_queue_t, so
75 * we can wake only the relevant ones (hashed queues may be shared).
77 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
78 * It is considered woken when list_empty(&q->list) || q->lock_ptr == 0.
79 * The order of wakup is always to make the first condition true, then
80 * wake up q->waiters, then make the second condition true.
83 struct list_head list
;
84 wait_queue_head_t waiters
;
86 /* Which hash list lock to use. */
89 /* Key which the futex is hashed on. */
92 /* For fd, sigio sent using these. */
98 * Split the global futex_lock into every hash list lock.
100 struct futex_hash_bucket
{
102 struct list_head chain
;
105 static struct futex_hash_bucket futex_queues
[1<<FUTEX_HASHBITS
];
107 /* Futex-fs vfsmount entry: */
108 static struct vfsmount
*futex_mnt
;
111 * We hash on the keys returned from get_futex_key (see below).
113 static struct futex_hash_bucket
*hash_futex(union futex_key
*key
)
115 u32 hash
= jhash2((u32
*)&key
->both
.word
,
116 (sizeof(key
->both
.word
)+sizeof(key
->both
.ptr
))/4,
118 return &futex_queues
[hash
& ((1 << FUTEX_HASHBITS
)-1)];
122 * Return 1 if two futex_keys are equal, 0 otherwise.
124 static inline int match_futex(union futex_key
*key1
, union futex_key
*key2
)
126 return (key1
->both
.word
== key2
->both
.word
127 && key1
->both
.ptr
== key2
->both
.ptr
128 && key1
->both
.offset
== key2
->both
.offset
);
132 * Get parameters which are the keys for a futex.
134 * For shared mappings, it's (page->index, vma->vm_file->f_dentry->d_inode,
135 * offset_within_page). For private mappings, it's (uaddr, current->mm).
136 * We can usually work out the index without swapping in the page.
138 * Returns: 0, or negative error code.
139 * The key words are stored in *key on success.
141 * Should be called with ¤t->mm->mmap_sem but NOT any spinlocks.
143 static int get_futex_key(unsigned long uaddr
, union futex_key
*key
)
145 struct mm_struct
*mm
= current
->mm
;
146 struct vm_area_struct
*vma
;
151 * The futex address must be "naturally" aligned.
153 key
->both
.offset
= uaddr
% PAGE_SIZE
;
154 if (unlikely((key
->both
.offset
% sizeof(u32
)) != 0))
156 uaddr
-= key
->both
.offset
;
159 * The futex is hashed differently depending on whether
160 * it's in a shared or private mapping. So check vma first.
162 vma
= find_extend_vma(mm
, uaddr
);
169 if (unlikely((vma
->vm_flags
& (VM_IO
|VM_READ
)) != VM_READ
))
170 return (vma
->vm_flags
& VM_IO
) ? -EPERM
: -EACCES
;
173 * Private mappings are handled in a simple way.
175 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
176 * it's a read-only handle, it's expected that futexes attach to
177 * the object not the particular process. Therefore we use
178 * VM_MAYSHARE here, not VM_SHARED which is restricted to shared
179 * mappings of _writable_ handles.
181 if (likely(!(vma
->vm_flags
& VM_MAYSHARE
))) {
182 key
->private.mm
= mm
;
183 key
->private.uaddr
= uaddr
;
188 * Linear file mappings are also simple.
190 key
->shared
.inode
= vma
->vm_file
->f_dentry
->d_inode
;
191 key
->both
.offset
++; /* Bit 0 of offset indicates inode-based key. */
192 if (likely(!(vma
->vm_flags
& VM_NONLINEAR
))) {
193 key
->shared
.pgoff
= (((uaddr
- vma
->vm_start
) >> PAGE_SHIFT
)
199 * We could walk the page table to read the non-linear
200 * pte, and get the page index without fetching the page
201 * from swap. But that's a lot of code to duplicate here
202 * for a rare case, so we simply fetch the page.
204 err
= get_user_pages(current
, mm
, uaddr
, 1, 0, 0, &page
, NULL
);
207 page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
215 * Take a reference to the resource addressed by a key.
216 * Can be called while holding spinlocks.
218 * NOTE: mmap_sem MUST be held between get_futex_key() and calling this
219 * function, if it is called at all. mmap_sem keeps key->shared.inode valid.
221 static inline void get_key_refs(union futex_key
*key
)
223 if (key
->both
.ptr
!= 0) {
224 if (key
->both
.offset
& 1)
225 atomic_inc(&key
->shared
.inode
->i_count
);
227 atomic_inc(&key
->private.mm
->mm_count
);
232 * Drop a reference to the resource addressed by a key.
233 * The hash bucket spinlock must not be held.
235 static void drop_key_refs(union futex_key
*key
)
237 if (key
->both
.ptr
!= 0) {
238 if (key
->both
.offset
& 1)
239 iput(key
->shared
.inode
);
241 mmdrop(key
->private.mm
);
245 static inline int get_futex_value_locked(int *dest
, int __user
*from
)
250 ret
= __copy_from_user_inatomic(dest
, from
, sizeof(int));
253 return ret
? -EFAULT
: 0;
257 * The hash bucket lock must be held when this is called.
258 * Afterwards, the futex_q must not be accessed.
260 static void wake_futex(struct futex_q
*q
)
262 list_del_init(&q
->list
);
264 send_sigio(&q
->filp
->f_owner
, q
->fd
, POLL_IN
);
266 * The lock in wake_up_all() is a crucial memory barrier after the
267 * list_del_init() and also before assigning to q->lock_ptr.
269 wake_up_all(&q
->waiters
);
271 * The waiting task can free the futex_q as soon as this is written,
272 * without taking any locks. This must come last.
278 * Wake up all waiters hashed on the physical page that is mapped
279 * to this virtual address:
281 static int futex_wake(unsigned long uaddr
, int nr_wake
)
284 struct futex_hash_bucket
*bh
;
285 struct list_head
*head
;
286 struct futex_q
*this, *next
;
289 down_read(¤t
->mm
->mmap_sem
);
291 ret
= get_futex_key(uaddr
, &key
);
292 if (unlikely(ret
!= 0))
295 bh
= hash_futex(&key
);
296 spin_lock(&bh
->lock
);
299 list_for_each_entry_safe(this, next
, head
, list
) {
300 if (match_futex (&this->key
, &key
)) {
302 if (++ret
>= nr_wake
)
307 spin_unlock(&bh
->lock
);
309 up_read(¤t
->mm
->mmap_sem
);
314 * Wake up all waiters hashed on the physical page that is mapped
315 * to this virtual address:
317 static int futex_wake_op(unsigned long uaddr1
, unsigned long uaddr2
, int nr_wake
, int nr_wake2
, int op
)
319 union futex_key key1
, key2
;
320 struct futex_hash_bucket
*bh1
, *bh2
;
321 struct list_head
*head
;
322 struct futex_q
*this, *next
;
323 int ret
, op_ret
, attempt
= 0;
326 down_read(¤t
->mm
->mmap_sem
);
328 ret
= get_futex_key(uaddr1
, &key1
);
329 if (unlikely(ret
!= 0))
331 ret
= get_futex_key(uaddr2
, &key2
);
332 if (unlikely(ret
!= 0))
335 bh1
= hash_futex(&key1
);
336 bh2
= hash_futex(&key2
);
340 spin_lock(&bh1
->lock
);
341 spin_lock(&bh2
->lock
);
343 spin_lock(&bh1
->lock
);
345 op_ret
= futex_atomic_op_inuser(op
, (int __user
*)uaddr2
);
346 if (unlikely(op_ret
< 0)) {
349 spin_unlock(&bh1
->lock
);
351 spin_unlock(&bh2
->lock
);
353 if (unlikely(op_ret
!= -EFAULT
)) {
358 /* futex_atomic_op_inuser needs to both read and write
359 * *(int __user *)uaddr2, but we can't modify it
360 * non-atomically. Therefore, if get_user below is not
361 * enough, we need to handle the fault ourselves, while
362 * still holding the mmap_sem. */
364 struct vm_area_struct
* vma
;
365 struct mm_struct
*mm
= current
->mm
;
369 !(vma
= find_vma(mm
, uaddr2
)) ||
370 vma
->vm_start
> uaddr2
||
371 !(vma
->vm_flags
& VM_WRITE
))
374 switch (handle_mm_fault(mm
, vma
, uaddr2
, 1)) {
387 /* If we would have faulted, release mmap_sem,
388 * fault it in and start all over again. */
389 up_read(¤t
->mm
->mmap_sem
);
391 ret
= get_user(dummy
, (int __user
*)uaddr2
);
400 list_for_each_entry_safe(this, next
, head
, list
) {
401 if (match_futex (&this->key
, &key1
)) {
403 if (++ret
>= nr_wake
)
412 list_for_each_entry_safe(this, next
, head
, list
) {
413 if (match_futex (&this->key
, &key2
)) {
415 if (++op_ret
>= nr_wake2
)
422 spin_unlock(&bh1
->lock
);
424 spin_unlock(&bh2
->lock
);
426 up_read(¤t
->mm
->mmap_sem
);
431 * Requeue all waiters hashed on one physical page to another
434 static int futex_requeue(unsigned long uaddr1
, unsigned long uaddr2
,
435 int nr_wake
, int nr_requeue
, int *valp
)
437 union futex_key key1
, key2
;
438 struct futex_hash_bucket
*bh1
, *bh2
;
439 struct list_head
*head1
;
440 struct futex_q
*this, *next
;
441 int ret
, drop_count
= 0;
444 down_read(¤t
->mm
->mmap_sem
);
446 ret
= get_futex_key(uaddr1
, &key1
);
447 if (unlikely(ret
!= 0))
449 ret
= get_futex_key(uaddr2
, &key2
);
450 if (unlikely(ret
!= 0))
453 bh1
= hash_futex(&key1
);
454 bh2
= hash_futex(&key2
);
457 spin_lock(&bh1
->lock
);
458 spin_lock(&bh2
->lock
);
460 spin_lock(&bh1
->lock
);
462 if (likely(valp
!= NULL
)) {
465 ret
= get_futex_value_locked(&curval
, (int __user
*)uaddr1
);
468 spin_unlock(&bh1
->lock
);
470 spin_unlock(&bh2
->lock
);
472 /* If we would have faulted, release mmap_sem, fault
473 * it in and start all over again.
475 up_read(¤t
->mm
->mmap_sem
);
477 ret
= get_user(curval
, (int __user
*)uaddr1
);
484 if (curval
!= *valp
) {
491 list_for_each_entry_safe(this, next
, head1
, list
) {
492 if (!match_futex (&this->key
, &key1
))
494 if (++ret
<= nr_wake
) {
497 list_move_tail(&this->list
, &bh2
->chain
);
498 this->lock_ptr
= &bh2
->lock
;
503 if (ret
- nr_wake
>= nr_requeue
)
505 /* Make sure to stop if key1 == key2 */
506 if (head1
== &bh2
->chain
&& head1
!= &next
->list
)
512 spin_unlock(&bh1
->lock
);
514 spin_unlock(&bh2
->lock
);
516 /* drop_key_refs() must be called outside the spinlocks. */
517 while (--drop_count
>= 0)
518 drop_key_refs(&key1
);
521 up_read(¤t
->mm
->mmap_sem
);
525 /* The key must be already stored in q->key. */
526 static inline struct futex_hash_bucket
*
527 queue_lock(struct futex_q
*q
, int fd
, struct file
*filp
)
529 struct futex_hash_bucket
*bh
;
534 init_waitqueue_head(&q
->waiters
);
536 get_key_refs(&q
->key
);
537 bh
= hash_futex(&q
->key
);
538 q
->lock_ptr
= &bh
->lock
;
540 spin_lock(&bh
->lock
);
544 static inline void __queue_me(struct futex_q
*q
, struct futex_hash_bucket
*bh
)
546 list_add_tail(&q
->list
, &bh
->chain
);
547 spin_unlock(&bh
->lock
);
551 queue_unlock(struct futex_q
*q
, struct futex_hash_bucket
*bh
)
553 spin_unlock(&bh
->lock
);
554 drop_key_refs(&q
->key
);
558 * queue_me and unqueue_me must be called as a pair, each
559 * exactly once. They are called with the hashed spinlock held.
562 /* The key must be already stored in q->key. */
563 static void queue_me(struct futex_q
*q
, int fd
, struct file
*filp
)
565 struct futex_hash_bucket
*bh
;
566 bh
= queue_lock(q
, fd
, filp
);
570 /* Return 1 if we were still queued (ie. 0 means we were woken) */
571 static int unqueue_me(struct futex_q
*q
)
574 spinlock_t
*lock_ptr
;
576 /* In the common case we don't take the spinlock, which is nice. */
578 lock_ptr
= q
->lock_ptr
;
582 * q->lock_ptr can change between reading it and
583 * spin_lock(), causing us to take the wrong lock. This
584 * corrects the race condition.
586 * Reasoning goes like this: if we have the wrong lock,
587 * q->lock_ptr must have changed (maybe several times)
588 * between reading it and the spin_lock(). It can
589 * change again after the spin_lock() but only if it was
590 * already changed before the spin_lock(). It cannot,
591 * however, change back to the original value. Therefore
592 * we can detect whether we acquired the correct lock.
594 if (unlikely(lock_ptr
!= q
->lock_ptr
)) {
595 spin_unlock(lock_ptr
);
598 WARN_ON(list_empty(&q
->list
));
600 spin_unlock(lock_ptr
);
604 drop_key_refs(&q
->key
);
608 static int futex_wait(unsigned long uaddr
, int val
, unsigned long time
)
610 DECLARE_WAITQUEUE(wait
, current
);
613 struct futex_hash_bucket
*bh
;
616 down_read(¤t
->mm
->mmap_sem
);
618 ret
= get_futex_key(uaddr
, &q
.key
);
619 if (unlikely(ret
!= 0))
620 goto out_release_sem
;
622 bh
= queue_lock(&q
, -1, NULL
);
625 * Access the page AFTER the futex is queued.
626 * Order is important:
628 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
629 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
631 * The basic logical guarantee of a futex is that it blocks ONLY
632 * if cond(var) is known to be true at the time of blocking, for
633 * any cond. If we queued after testing *uaddr, that would open
634 * a race condition where we could block indefinitely with
635 * cond(var) false, which would violate the guarantee.
637 * A consequence is that futex_wait() can return zero and absorb
638 * a wakeup when *uaddr != val on entry to the syscall. This is
641 * We hold the mmap semaphore, so the mapping cannot have changed
642 * since we looked it up in get_futex_key.
645 ret
= get_futex_value_locked(&curval
, (int __user
*)uaddr
);
648 queue_unlock(&q
, bh
);
650 /* If we would have faulted, release mmap_sem, fault it in and
651 * start all over again.
653 up_read(¤t
->mm
->mmap_sem
);
655 ret
= get_user(curval
, (int __user
*)uaddr
);
663 queue_unlock(&q
, bh
);
664 goto out_release_sem
;
667 /* Only actually queue if *uaddr contained val. */
671 * Now the futex is queued and we have checked the data, we
672 * don't want to hold mmap_sem while we sleep.
674 up_read(¤t
->mm
->mmap_sem
);
677 * There might have been scheduling since the queue_me(), as we
678 * cannot hold a spinlock across the get_user() in case it
679 * faults, and we cannot just set TASK_INTERRUPTIBLE state when
680 * queueing ourselves into the futex hash. This code thus has to
681 * rely on the futex_wake() code removing us from hash when it
685 /* add_wait_queue is the barrier after __set_current_state. */
686 __set_current_state(TASK_INTERRUPTIBLE
);
687 add_wait_queue(&q
.waiters
, &wait
);
689 * !list_empty() is safe here without any lock.
690 * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
692 if (likely(!list_empty(&q
.list
)))
693 time
= schedule_timeout(time
);
694 __set_current_state(TASK_RUNNING
);
697 * NOTE: we don't remove ourselves from the waitqueue because
698 * we are the only user of it.
701 /* If we were woken (and unqueued), we succeeded, whatever. */
706 /* We expect signal_pending(current), but another thread may
707 * have handled it for us already. */
711 up_read(¤t
->mm
->mmap_sem
);
715 static int futex_close(struct inode
*inode
, struct file
*filp
)
717 struct futex_q
*q
= filp
->private_data
;
724 /* This is one-shot: once it's gone off you need a new fd */
725 static unsigned int futex_poll(struct file
*filp
,
726 struct poll_table_struct
*wait
)
728 struct futex_q
*q
= filp
->private_data
;
731 poll_wait(filp
, &q
->waiters
, wait
);
734 * list_empty() is safe here without any lock.
735 * q->lock_ptr != 0 is not safe, because of ordering against wakeup.
737 if (list_empty(&q
->list
))
738 ret
= POLLIN
| POLLRDNORM
;
743 static struct file_operations futex_fops
= {
744 .release
= futex_close
,
749 * Signal allows caller to avoid the race which would occur if they
750 * set the sigio stuff up afterwards.
752 static int futex_fd(unsigned long uaddr
, int signal
)
759 if (!valid_signal(signal
))
762 ret
= get_unused_fd();
765 filp
= get_empty_filp();
771 filp
->f_op
= &futex_fops
;
772 filp
->f_vfsmnt
= mntget(futex_mnt
);
773 filp
->f_dentry
= dget(futex_mnt
->mnt_root
);
774 filp
->f_mapping
= filp
->f_dentry
->d_inode
->i_mapping
;
777 err
= f_setown(filp
, current
->pid
, 1);
781 filp
->f_owner
.signum
= signal
;
784 q
= kmalloc(sizeof(*q
), GFP_KERNEL
);
790 down_read(¤t
->mm
->mmap_sem
);
791 err
= get_futex_key(uaddr
, &q
->key
);
793 if (unlikely(err
!= 0)) {
794 up_read(¤t
->mm
->mmap_sem
);
800 * queue_me() must be called before releasing mmap_sem, because
801 * key->shared.inode needs to be referenced while holding it.
803 filp
->private_data
= q
;
805 queue_me(q
, ret
, filp
);
806 up_read(¤t
->mm
->mmap_sem
);
808 /* Now we map fd to filp, so userspace can access it */
809 fd_install(ret
, filp
);
819 long do_futex(unsigned long uaddr
, int op
, int val
, unsigned long timeout
,
820 unsigned long uaddr2
, int val2
, int val3
)
826 ret
= futex_wait(uaddr
, val
, timeout
);
829 ret
= futex_wake(uaddr
, val
);
832 /* non-zero val means F_SETOWN(getpid()) & F_SETSIG(val) */
833 ret
= futex_fd(uaddr
, val
);
836 ret
= futex_requeue(uaddr
, uaddr2
, val
, val2
, NULL
);
838 case FUTEX_CMP_REQUEUE
:
839 ret
= futex_requeue(uaddr
, uaddr2
, val
, val2
, &val3
);
842 ret
= futex_wake_op(uaddr
, uaddr2
, val
, val2
, val3
);
851 asmlinkage
long sys_futex(u32 __user
*uaddr
, int op
, int val
,
852 struct timespec __user
*utime
, u32 __user
*uaddr2
,
856 unsigned long timeout
= MAX_SCHEDULE_TIMEOUT
;
859 if ((op
== FUTEX_WAIT
) && utime
) {
860 if (copy_from_user(&t
, utime
, sizeof(t
)) != 0)
862 timeout
= timespec_to_jiffies(&t
) + 1;
865 * requeue parameter in 'utime' if op == FUTEX_REQUEUE.
867 if (op
>= FUTEX_REQUEUE
)
868 val2
= (int) (unsigned long) utime
;
870 return do_futex((unsigned long)uaddr
, op
, val
, timeout
,
871 (unsigned long)uaddr2
, val2
, val3
);
874 static struct super_block
*
875 futexfs_get_sb(struct file_system_type
*fs_type
,
876 int flags
, const char *dev_name
, void *data
)
878 return get_sb_pseudo(fs_type
, "futex", NULL
, 0xBAD1DEA);
881 static struct file_system_type futex_fs_type
= {
883 .get_sb
= futexfs_get_sb
,
884 .kill_sb
= kill_anon_super
,
887 static int __init
init(void)
891 register_filesystem(&futex_fs_type
);
892 futex_mnt
= kern_mount(&futex_fs_type
);
894 for (i
= 0; i
< ARRAY_SIZE(futex_queues
); i
++) {
895 INIT_LIST_HEAD(&futex_queues
[i
].chain
);
896 spin_lock_init(&futex_queues
[i
].lock
);