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 * Robust futex support started by Ingo Molnar
12 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
15 * PI-futex support started by Ingo Molnar and Thomas Gleixner
16 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
19 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
20 * enough at me, Linus for the original (flawed) idea, Matthew
21 * Kirkwood for proof-of-concept implementation.
23 * "The futexes are also cursed."
24 * "But they come in a choice of three flavours!"
26 * This program is free software; you can redistribute it and/or modify
27 * it under the terms of the GNU General Public License as published by
28 * the Free Software Foundation; either version 2 of the License, or
29 * (at your option) any later version.
31 * This program is distributed in the hope that it will be useful,
32 * but WITHOUT ANY WARRANTY; without even the implied warranty of
33 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
34 * GNU General Public License for more details.
36 * You should have received a copy of the GNU General Public License
37 * along with this program; if not, write to the Free Software
38 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
40 #include <linux/slab.h>
41 #include <linux/poll.h>
43 #include <linux/file.h>
44 #include <linux/jhash.h>
45 #include <linux/init.h>
46 #include <linux/futex.h>
47 #include <linux/mount.h>
48 #include <linux/pagemap.h>
49 #include <linux/syscalls.h>
50 #include <linux/signal.h>
51 #include <asm/futex.h>
53 #include "rtmutex_common.h"
55 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
58 * Futexes are matched on equal values of this key.
59 * The key type depends on whether it's a shared or private mapping.
60 * Don't rearrange members without looking at hash_futex().
62 * offset is aligned to a multiple of sizeof(u32) (== 4) by definition.
63 * We set bit 0 to indicate if it's an inode-based key.
72 unsigned long address;
84 * Priority Inheritance state:
86 struct futex_pi_state {
88 * list of 'owned' pi_state instances - these have to be
89 * cleaned up in do_exit() if the task exits prematurely:
91 struct list_head list;
96 struct rt_mutex pi_mutex;
98 struct task_struct *owner;
105 * We use this hashed waitqueue instead of a normal wait_queue_t, so
106 * we can wake only the relevant ones (hashed queues may be shared).
108 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
109 * It is considered woken when list_empty(&q->list) || q->lock_ptr == 0.
110 * The order of wakup is always to make the first condition true, then
111 * wake up q->waiters, then make the second condition true.
114 struct list_head list;
115 wait_queue_head_t waiters;
117 /* Which hash list lock to use: */
118 spinlock_t *lock_ptr;
120 /* Key which the futex is hashed on: */
123 /* For fd, sigio sent using these: */
127 /* Optional priority inheritance state: */
128 struct futex_pi_state *pi_state;
129 struct task_struct *task;
133 * Split the global futex_lock into every hash list lock.
135 struct futex_hash_bucket {
137 struct list_head chain;
140 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
142 /* Futex-fs vfsmount entry: */
143 static struct vfsmount *futex_mnt;
146 * We hash on the keys returned from get_futex_key (see below).
148 static struct futex_hash_bucket *hash_futex(union futex_key *key)
150 u32 hash = jhash2((u32*)&key->both.word,
151 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
153 return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
157 * Return 1 if two futex_keys are equal, 0 otherwise.
159 static inline int match_futex(union futex_key *key1, union futex_key *key2)
161 return (key1->both.word == key2->both.word
162 && key1->both.ptr == key2->both.ptr
163 && key1->both.offset == key2->both.offset);
167 * Get parameters which are the keys for a futex.
169 * For shared mappings, it's (page->index, vma->vm_file->f_dentry->d_inode,
170 * offset_within_page). For private mappings, it's (uaddr, current->mm).
171 * We can usually work out the index without swapping in the page.
173 * Returns: 0, or negative error code.
174 * The key words are stored in *key on success.
176 * Should be called with ¤t->mm->mmap_sem but NOT any spinlocks.
178 static int get_futex_key(u32 __user *uaddr, union futex_key *key)
180 unsigned long address = (unsigned long)uaddr;
181 struct mm_struct *mm = current->mm;
182 struct vm_area_struct *vma;
187 * The futex address must be "naturally" aligned.
189 key->both.offset = address % PAGE_SIZE;
190 if (unlikely((key->both.offset % sizeof(u32)) != 0))
192 address -= key->both.offset;
195 * The futex is hashed differently depending on whether
196 * it's in a shared or private mapping. So check vma first.
198 vma = find_extend_vma(mm, address);
205 if (unlikely((vma->vm_flags & (VM_IO|VM_READ)) != VM_READ))
206 return (vma->vm_flags & VM_IO) ? -EPERM : -EACCES;
209 * Private mappings are handled in a simple way.
211 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
212 * it's a read-only handle, it's expected that futexes attach to
213 * the object not the particular process. Therefore we use
214 * VM_MAYSHARE here, not VM_SHARED which is restricted to shared
215 * mappings of _writable_ handles.
217 if (likely(!(vma->vm_flags & VM_MAYSHARE))) {
218 key->private.mm = mm;
219 key->private.address = address;
224 * Linear file mappings are also simple.
226 key->shared.inode = vma->vm_file->f_dentry->d_inode;
227 key->both.offset++; /* Bit 0 of offset indicates inode-based key. */
228 if (likely(!(vma->vm_flags & VM_NONLINEAR))) {
229 key->shared.pgoff = (((address - vma->vm_start) >> PAGE_SHIFT)
235 * We could walk the page table to read the non-linear
236 * pte, and get the page index without fetching the page
237 * from swap. But that's a lot of code to duplicate here
238 * for a rare case, so we simply fetch the page.
240 err = get_user_pages(current, mm, address, 1, 0, 0, &page, NULL);
243 page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
251 * Take a reference to the resource addressed by a key.
252 * Can be called while holding spinlocks.
254 * NOTE: mmap_sem MUST be held between get_futex_key() and calling this
255 * function, if it is called at all. mmap_sem keeps key->shared.inode valid.
257 static inline void get_key_refs(union futex_key *key)
259 if (key->both.ptr != 0) {
260 if (key->both.offset & 1)
261 atomic_inc(&key->shared.inode->i_count);
263 atomic_inc(&key->private.mm->mm_count);
268 * Drop a reference to the resource addressed by a key.
269 * The hash bucket spinlock must not be held.
271 static void drop_key_refs(union futex_key *key)
273 if (key->both.ptr != 0) {
274 if (key->both.offset & 1)
275 iput(key->shared.inode);
277 mmdrop(key->private.mm);
281 static inline int get_futex_value_locked(u32 *dest, u32 __user *from)
286 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
289 return ret ? -EFAULT : 0;
293 * Fault handling. Called with current->mm->mmap_sem held.
295 static int futex_handle_fault(unsigned long address, int attempt)
297 struct vm_area_struct * vma;
298 struct mm_struct *mm = current->mm;
300 if (attempt >= 2 || !(vma = find_vma(mm, address)) ||
301 vma->vm_start > address || !(vma->vm_flags & VM_WRITE))
304 switch (handle_mm_fault(mm, vma, address, 1)) {
320 static int refill_pi_state_cache(void)
322 struct futex_pi_state *pi_state;
324 if (likely(current->pi_state_cache))
327 pi_state = kmalloc(sizeof(*pi_state), GFP_KERNEL);
332 memset(pi_state, 0, sizeof(*pi_state));
333 INIT_LIST_HEAD(&pi_state->list);
334 /* pi_mutex gets initialized later */
335 pi_state->owner = NULL;
336 atomic_set(&pi_state->refcount, 1);
338 current->pi_state_cache = pi_state;
343 static struct futex_pi_state * alloc_pi_state(void)
345 struct futex_pi_state *pi_state = current->pi_state_cache;
348 current->pi_state_cache = NULL;
353 static void free_pi_state(struct futex_pi_state *pi_state)
355 if (!atomic_dec_and_test(&pi_state->refcount))
359 * If pi_state->owner is NULL, the owner is most probably dying
360 * and has cleaned up the pi_state already
362 if (pi_state->owner) {
363 spin_lock_irq(&pi_state->owner->pi_lock);
364 list_del_init(&pi_state->list);
365 spin_unlock_irq(&pi_state->owner->pi_lock);
367 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
370 if (current->pi_state_cache)
374 * pi_state->list is already empty.
375 * clear pi_state->owner.
376 * refcount is at 0 - put it back to 1.
378 pi_state->owner = NULL;
379 atomic_set(&pi_state->refcount, 1);
380 current->pi_state_cache = pi_state;
385 * Look up the task based on what TID userspace gave us.
388 static struct task_struct * futex_find_get_task(pid_t pid)
390 struct task_struct *p;
392 read_lock(&tasklist_lock);
393 p = find_task_by_pid(pid);
396 if ((current->euid != p->euid) && (current->euid != p->uid)) {
400 if (p->state == EXIT_ZOMBIE || p->exit_state == EXIT_ZOMBIE) {
406 read_unlock(&tasklist_lock);
412 * This task is holding PI mutexes at exit time => bad.
413 * Kernel cleans up PI-state, but userspace is likely hosed.
414 * (Robust-futex cleanup is separate and might save the day for userspace.)
416 void exit_pi_state_list(struct task_struct *curr)
418 struct futex_hash_bucket *hb;
419 struct list_head *next, *head = &curr->pi_state_list;
420 struct futex_pi_state *pi_state;
424 * We are a ZOMBIE and nobody can enqueue itself on
425 * pi_state_list anymore, but we have to be careful
426 * versus waiters unqueueing themselfs
428 spin_lock_irq(&curr->pi_lock);
429 while (!list_empty(head)) {
432 pi_state = list_entry(next, struct futex_pi_state, list);
434 spin_unlock_irq(&curr->pi_lock);
436 hb = hash_futex(&key);
437 spin_lock(&hb->lock);
439 spin_lock_irq(&curr->pi_lock);
440 if (head->next != next) {
441 spin_unlock(&hb->lock);
445 list_del_init(&pi_state->list);
447 WARN_ON(pi_state->owner != curr);
449 pi_state->owner = NULL;
450 spin_unlock_irq(&curr->pi_lock);
452 rt_mutex_unlock(&pi_state->pi_mutex);
454 spin_unlock(&hb->lock);
456 spin_lock_irq(&curr->pi_lock);
458 spin_unlock_irq(&curr->pi_lock);
462 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb, struct futex_q *me)
464 struct futex_pi_state *pi_state = NULL;
465 struct futex_q *this, *next;
466 struct list_head *head;
467 struct task_struct *p;
472 list_for_each_entry_safe(this, next, head, list) {
473 if (match_futex (&this->key, &me->key)) {
475 * Another waiter already exists - bump up
476 * the refcount and return its pi_state:
478 pi_state = this->pi_state;
479 atomic_inc(&pi_state->refcount);
480 me->pi_state = pi_state;
487 * We are the first waiter - try to look up the real owner and
488 * attach the new pi_state to it:
490 pid = uval & FUTEX_TID_MASK;
491 p = futex_find_get_task(pid);
495 pi_state = alloc_pi_state();
498 * Initialize the pi_mutex in locked state and make 'p'
501 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
503 /* Store the key for possible exit cleanups: */
504 pi_state->key = me->key;
506 spin_lock_irq(&p->pi_lock);
507 list_add(&pi_state->list, &p->pi_state_list);
509 spin_unlock_irq(&p->pi_lock);
513 me->pi_state = pi_state;
519 * The hash bucket lock must be held when this is called.
520 * Afterwards, the futex_q must not be accessed.
522 static void wake_futex(struct futex_q *q)
524 list_del_init(&q->list);
526 send_sigio(&q->filp->f_owner, q->fd, POLL_IN);
528 * The lock in wake_up_all() is a crucial memory barrier after the
529 * list_del_init() and also before assigning to q->lock_ptr.
531 wake_up_all(&q->waiters);
533 * The waiting task can free the futex_q as soon as this is written,
534 * without taking any locks. This must come last.
536 * A memory barrier is required here to prevent the following store
537 * to lock_ptr from getting ahead of the wakeup. Clearing the lock
538 * at the end of wake_up_all() does not prevent this store from
545 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
547 struct task_struct *new_owner;
548 struct futex_pi_state *pi_state = this->pi_state;
554 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
557 * This happens when we have stolen the lock and the original
558 * pending owner did not enqueue itself back on the rt_mutex.
559 * Thats not a tragedy. We know that way, that a lock waiter
560 * is on the fly. We make the futex_q waiter the pending owner.
563 new_owner = this->task;
566 * We pass it to the next owner. (The WAITERS bit is always
567 * kept enabled while there is PI state around. We must also
568 * preserve the owner died bit.)
570 newval = (uval & FUTEX_OWNER_DIED) | FUTEX_WAITERS | new_owner->pid;
573 curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
576 if (curval == -EFAULT)
581 list_del_init(&pi_state->owner->pi_state_list);
582 list_add(&pi_state->list, &new_owner->pi_state_list);
583 pi_state->owner = new_owner;
584 rt_mutex_unlock(&pi_state->pi_mutex);
589 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
594 * There is no waiter, so we unlock the futex. The owner died
595 * bit has not to be preserved here. We are the owner:
598 oldval = futex_atomic_cmpxchg_inatomic(uaddr, uval, 0);
601 if (oldval == -EFAULT)
610 * Wake up all waiters hashed on the physical page that is mapped
611 * to this virtual address:
613 static int futex_wake(u32 __user *uaddr, int nr_wake)
615 struct futex_hash_bucket *hb;
616 struct futex_q *this, *next;
617 struct list_head *head;
621 down_read(¤t->mm->mmap_sem);
623 ret = get_futex_key(uaddr, &key);
624 if (unlikely(ret != 0))
627 hb = hash_futex(&key);
628 spin_lock(&hb->lock);
631 list_for_each_entry_safe(this, next, head, list) {
632 if (match_futex (&this->key, &key)) {
636 if (++ret >= nr_wake)
641 spin_unlock(&hb->lock);
643 up_read(¤t->mm->mmap_sem);
648 * Wake up all waiters hashed on the physical page that is mapped
649 * to this virtual address:
652 futex_wake_op(u32 __user *uaddr1, u32 __user *uaddr2,
653 int nr_wake, int nr_wake2, int op)
655 union futex_key key1, key2;
656 struct futex_hash_bucket *hb1, *hb2;
657 struct list_head *head;
658 struct futex_q *this, *next;
659 int ret, op_ret, attempt = 0;
662 down_read(¤t->mm->mmap_sem);
664 ret = get_futex_key(uaddr1, &key1);
665 if (unlikely(ret != 0))
667 ret = get_futex_key(uaddr2, &key2);
668 if (unlikely(ret != 0))
671 hb1 = hash_futex(&key1);
672 hb2 = hash_futex(&key2);
676 spin_lock(&hb1->lock);
677 spin_lock(&hb2->lock);
679 spin_lock(&hb1->lock);
681 op_ret = futex_atomic_op_inuser(op, uaddr2);
682 if (unlikely(op_ret < 0)) {
685 spin_unlock(&hb1->lock);
687 spin_unlock(&hb2->lock);
691 * we don't get EFAULT from MMU faults if we don't have an MMU,
692 * but we might get them from range checking
698 if (unlikely(op_ret != -EFAULT)) {
704 * futex_atomic_op_inuser needs to both read and write
705 * *(int __user *)uaddr2, but we can't modify it
706 * non-atomically. Therefore, if get_user below is not
707 * enough, we need to handle the fault ourselves, while
708 * still holding the mmap_sem.
711 if (futex_handle_fault((unsigned long)uaddr2,
718 * If we would have faulted, release mmap_sem,
719 * fault it in and start all over again.
721 up_read(¤t->mm->mmap_sem);
723 ret = get_user(dummy, uaddr2);
732 list_for_each_entry_safe(this, next, head, list) {
733 if (match_futex (&this->key, &key1)) {
735 if (++ret >= nr_wake)
744 list_for_each_entry_safe(this, next, head, list) {
745 if (match_futex (&this->key, &key2)) {
747 if (++op_ret >= nr_wake2)
754 spin_unlock(&hb1->lock);
756 spin_unlock(&hb2->lock);
758 up_read(¤t->mm->mmap_sem);
763 * Requeue all waiters hashed on one physical page to another
766 static int futex_requeue(u32 __user *uaddr1, u32 __user *uaddr2,
767 int nr_wake, int nr_requeue, u32 *cmpval)
769 union futex_key key1, key2;
770 struct futex_hash_bucket *hb1, *hb2;
771 struct list_head *head1;
772 struct futex_q *this, *next;
773 int ret, drop_count = 0;
776 down_read(¤t->mm->mmap_sem);
778 ret = get_futex_key(uaddr1, &key1);
779 if (unlikely(ret != 0))
781 ret = get_futex_key(uaddr2, &key2);
782 if (unlikely(ret != 0))
785 hb1 = hash_futex(&key1);
786 hb2 = hash_futex(&key2);
789 spin_lock(&hb1->lock);
790 spin_lock(&hb2->lock);
792 spin_lock(&hb1->lock);
794 if (likely(cmpval != NULL)) {
797 ret = get_futex_value_locked(&curval, uaddr1);
800 spin_unlock(&hb1->lock);
802 spin_unlock(&hb2->lock);
805 * If we would have faulted, release mmap_sem, fault
806 * it in and start all over again.
808 up_read(¤t->mm->mmap_sem);
810 ret = get_user(curval, uaddr1);
817 if (curval != *cmpval) {
824 list_for_each_entry_safe(this, next, head1, list) {
825 if (!match_futex (&this->key, &key1))
827 if (++ret <= nr_wake) {
830 list_move_tail(&this->list, &hb2->chain);
831 this->lock_ptr = &hb2->lock;
836 if (ret - nr_wake >= nr_requeue)
838 /* Make sure to stop if key1 == key2: */
839 if (head1 == &hb2->chain && head1 != &next->list)
845 spin_unlock(&hb1->lock);
847 spin_unlock(&hb2->lock);
849 /* drop_key_refs() must be called outside the spinlocks. */
850 while (--drop_count >= 0)
851 drop_key_refs(&key1);
854 up_read(¤t->mm->mmap_sem);
858 /* The key must be already stored in q->key. */
859 static inline struct futex_hash_bucket *
860 queue_lock(struct futex_q *q, int fd, struct file *filp)
862 struct futex_hash_bucket *hb;
867 init_waitqueue_head(&q->waiters);
869 get_key_refs(&q->key);
870 hb = hash_futex(&q->key);
871 q->lock_ptr = &hb->lock;
873 spin_lock(&hb->lock);
877 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
879 list_add_tail(&q->list, &hb->chain);
881 spin_unlock(&hb->lock);
885 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
887 spin_unlock(&hb->lock);
888 drop_key_refs(&q->key);
892 * queue_me and unqueue_me must be called as a pair, each
893 * exactly once. They are called with the hashed spinlock held.
896 /* The key must be already stored in q->key. */
897 static void queue_me(struct futex_q *q, int fd, struct file *filp)
899 struct futex_hash_bucket *hb;
901 hb = queue_lock(q, fd, filp);
905 /* Return 1 if we were still queued (ie. 0 means we were woken) */
906 static int unqueue_me(struct futex_q *q)
908 spinlock_t *lock_ptr;
911 /* In the common case we don't take the spinlock, which is nice. */
913 lock_ptr = q->lock_ptr;
917 * q->lock_ptr can change between reading it and
918 * spin_lock(), causing us to take the wrong lock. This
919 * corrects the race condition.
921 * Reasoning goes like this: if we have the wrong lock,
922 * q->lock_ptr must have changed (maybe several times)
923 * between reading it and the spin_lock(). It can
924 * change again after the spin_lock() but only if it was
925 * already changed before the spin_lock(). It cannot,
926 * however, change back to the original value. Therefore
927 * we can detect whether we acquired the correct lock.
929 if (unlikely(lock_ptr != q->lock_ptr)) {
930 spin_unlock(lock_ptr);
933 WARN_ON(list_empty(&q->list));
938 spin_unlock(lock_ptr);
942 drop_key_refs(&q->key);
947 * PI futexes can not be requeued and must remove themself from the
948 * hash bucket. The hash bucket lock is held on entry and dropped here.
950 static void unqueue_me_pi(struct futex_q *q, struct futex_hash_bucket *hb)
952 WARN_ON(list_empty(&q->list));
955 BUG_ON(!q->pi_state);
956 free_pi_state(q->pi_state);
959 spin_unlock(&hb->lock);
961 drop_key_refs(&q->key);
964 static int futex_wait(u32 __user *uaddr, u32 val, unsigned long time)
966 struct task_struct *curr = current;
967 DECLARE_WAITQUEUE(wait, curr);
968 struct futex_hash_bucket *hb;
975 down_read(&curr->mm->mmap_sem);
977 ret = get_futex_key(uaddr, &q.key);
978 if (unlikely(ret != 0))
979 goto out_release_sem;
981 hb = queue_lock(&q, -1, NULL);
984 * Access the page AFTER the futex is queued.
985 * Order is important:
987 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
988 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
990 * The basic logical guarantee of a futex is that it blocks ONLY
991 * if cond(var) is known to be true at the time of blocking, for
992 * any cond. If we queued after testing *uaddr, that would open
993 * a race condition where we could block indefinitely with
994 * cond(var) false, which would violate the guarantee.
996 * A consequence is that futex_wait() can return zero and absorb
997 * a wakeup when *uaddr != val on entry to the syscall. This is
1000 * We hold the mmap semaphore, so the mapping cannot have changed
1001 * since we looked it up in get_futex_key.
1003 ret = get_futex_value_locked(&uval, uaddr);
1005 if (unlikely(ret)) {
1006 queue_unlock(&q, hb);
1009 * If we would have faulted, release mmap_sem, fault it in and
1010 * start all over again.
1012 up_read(&curr->mm->mmap_sem);
1014 ret = get_user(uval, uaddr);
1022 goto out_unlock_release_sem;
1024 /* Only actually queue if *uaddr contained val. */
1028 * Now the futex is queued and we have checked the data, we
1029 * don't want to hold mmap_sem while we sleep.
1031 up_read(&curr->mm->mmap_sem);
1034 * There might have been scheduling since the queue_me(), as we
1035 * cannot hold a spinlock across the get_user() in case it
1036 * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1037 * queueing ourselves into the futex hash. This code thus has to
1038 * rely on the futex_wake() code removing us from hash when it
1042 /* add_wait_queue is the barrier after __set_current_state. */
1043 __set_current_state(TASK_INTERRUPTIBLE);
1044 add_wait_queue(&q.waiters, &wait);
1046 * !list_empty() is safe here without any lock.
1047 * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1049 if (likely(!list_empty(&q.list)))
1050 time = schedule_timeout(time);
1051 __set_current_state(TASK_RUNNING);
1054 * NOTE: we don't remove ourselves from the waitqueue because
1055 * we are the only user of it.
1058 /* If we were woken (and unqueued), we succeeded, whatever. */
1059 if (!unqueue_me(&q))
1064 * We expect signal_pending(current), but another thread may
1065 * have handled it for us already.
1069 out_unlock_release_sem:
1070 queue_unlock(&q, hb);
1073 up_read(&curr->mm->mmap_sem);
1078 * Userspace tried a 0 -> TID atomic transition of the futex value
1079 * and failed. The kernel side here does the whole locking operation:
1080 * if there are waiters then it will block, it does PI, etc. (Due to
1081 * races the kernel might see a 0 value of the futex too.)
1083 static int do_futex_lock_pi(u32 __user *uaddr, int detect, int trylock,
1084 struct hrtimer_sleeper *to)
1086 struct task_struct *curr = current;
1087 struct futex_hash_bucket *hb;
1088 u32 uval, newval, curval;
1090 int ret, attempt = 0;
1092 if (refill_pi_state_cache())
1097 down_read(&curr->mm->mmap_sem);
1099 ret = get_futex_key(uaddr, &q.key);
1100 if (unlikely(ret != 0))
1101 goto out_release_sem;
1103 hb = queue_lock(&q, -1, NULL);
1107 * To avoid races, we attempt to take the lock here again
1108 * (by doing a 0 -> TID atomic cmpxchg), while holding all
1109 * the locks. It will most likely not succeed.
1111 newval = current->pid;
1113 inc_preempt_count();
1114 curval = futex_atomic_cmpxchg_inatomic(uaddr, 0, newval);
1115 dec_preempt_count();
1117 if (unlikely(curval == -EFAULT))
1120 /* We own the lock already */
1121 if (unlikely((curval & FUTEX_TID_MASK) == current->pid)) {
1123 force_sig(SIGKILL, current);
1125 goto out_unlock_release_sem;
1129 * Surprise - we got the lock. Just return
1132 if (unlikely(!curval))
1133 goto out_unlock_release_sem;
1136 newval = uval | FUTEX_WAITERS;
1138 inc_preempt_count();
1139 curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
1140 dec_preempt_count();
1142 if (unlikely(curval == -EFAULT))
1144 if (unlikely(curval != uval))
1148 * We dont have the lock. Look up the PI state (or create it if
1149 * we are the first waiter):
1151 ret = lookup_pi_state(uval, hb, &q);
1153 if (unlikely(ret)) {
1155 * There were no waiters and the owner task lookup
1156 * failed. When the OWNER_DIED bit is set, then we
1157 * know that this is a robust futex and we actually
1158 * take the lock. This is safe as we are protected by
1159 * the hash bucket lock. We also set the waiters bit
1160 * unconditionally here, to simplify glibc handling of
1161 * multiple tasks racing to acquire the lock and
1162 * cleanup the problems which were left by the dead
1165 if (curval & FUTEX_OWNER_DIED) {
1167 newval = current->pid |
1168 FUTEX_OWNER_DIED | FUTEX_WAITERS;
1170 inc_preempt_count();
1171 curval = futex_atomic_cmpxchg_inatomic(uaddr,
1173 dec_preempt_count();
1175 if (unlikely(curval == -EFAULT))
1177 if (unlikely(curval != uval))
1181 goto out_unlock_release_sem;
1185 * Only actually queue now that the atomic ops are done:
1190 * Now the futex is queued and we have checked the data, we
1191 * don't want to hold mmap_sem while we sleep.
1193 up_read(&curr->mm->mmap_sem);
1195 WARN_ON(!q.pi_state);
1197 * Block on the PI mutex:
1200 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1202 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1203 /* Fixup the trylock return value: */
1204 ret = ret ? 0 : -EWOULDBLOCK;
1207 down_read(&curr->mm->mmap_sem);
1208 hb = queue_lock(&q, -1, NULL);
1211 * Got the lock. We might not be the anticipated owner if we
1212 * did a lock-steal - fix up the PI-state in that case.
1214 if (!ret && q.pi_state->owner != curr) {
1215 u32 newtid = current->pid | FUTEX_WAITERS;
1218 if (q.pi_state->owner != NULL) {
1219 spin_lock_irq(&q.pi_state->owner->pi_lock);
1220 list_del_init(&q.pi_state->list);
1221 spin_unlock_irq(&q.pi_state->owner->pi_lock);
1223 newtid |= FUTEX_OWNER_DIED;
1225 q.pi_state->owner = current;
1227 spin_lock_irq(¤t->pi_lock);
1228 list_add(&q.pi_state->list, ¤t->pi_state_list);
1229 spin_unlock_irq(¤t->pi_lock);
1231 /* Unqueue and drop the lock */
1232 unqueue_me_pi(&q, hb);
1233 up_read(&curr->mm->mmap_sem);
1235 * We own it, so we have to replace the pending owner
1236 * TID. This must be atomic as we have preserve the
1237 * owner died bit here.
1239 ret = get_user(uval, uaddr);
1241 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1242 curval = futex_atomic_cmpxchg_inatomic(uaddr,
1244 if (curval == -EFAULT)
1252 * Catch the rare case, where the lock was released
1253 * when we were on the way back before we locked
1256 if (ret && q.pi_state->owner == curr) {
1257 if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1260 /* Unqueue and drop the lock */
1261 unqueue_me_pi(&q, hb);
1262 up_read(&curr->mm->mmap_sem);
1265 if (!detect && ret == -EDEADLK && 0)
1266 force_sig(SIGKILL, current);
1270 out_unlock_release_sem:
1271 queue_unlock(&q, hb);
1274 up_read(&curr->mm->mmap_sem);
1279 * We have to r/w *(int __user *)uaddr, but we can't modify it
1280 * non-atomically. Therefore, if get_user below is not
1281 * enough, we need to handle the fault ourselves, while
1282 * still holding the mmap_sem.
1285 if (futex_handle_fault((unsigned long)uaddr, attempt))
1286 goto out_unlock_release_sem;
1291 queue_unlock(&q, hb);
1292 up_read(&curr->mm->mmap_sem);
1294 ret = get_user(uval, uaddr);
1295 if (!ret && (uval != -EFAULT))
1304 static long futex_lock_pi_restart(struct restart_block *restart)
1306 struct hrtimer_sleeper timeout, *to = NULL;
1309 restart->fn = do_no_restart_syscall;
1311 if (restart->arg2 || restart->arg3) {
1313 hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_ABS);
1314 hrtimer_init_sleeper(to, current);
1315 to->timer.expires.tv64 = ((u64)restart->arg1 << 32) |
1316 (u64) restart->arg0;
1319 pr_debug("lock_pi restart: %p, %d (%d)\n",
1320 (u32 __user *)restart->arg0, current->pid);
1322 ret = do_futex_lock_pi((u32 __user *)restart->arg0, restart->arg1,
1328 restart->fn = futex_lock_pi_restart;
1330 /* The other values are filled in */
1331 return -ERESTART_RESTARTBLOCK;
1335 * Called from the syscall entry below.
1337 static int futex_lock_pi(u32 __user *uaddr, int detect, unsigned long sec,
1338 long nsec, int trylock)
1340 struct hrtimer_sleeper timeout, *to = NULL;
1341 struct restart_block *restart;
1344 if (sec != MAX_SCHEDULE_TIMEOUT) {
1346 hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_ABS);
1347 hrtimer_init_sleeper(to, current);
1348 to->timer.expires = ktime_set(sec, nsec);
1351 ret = do_futex_lock_pi(uaddr, detect, trylock, to);
1356 pr_debug("lock_pi interrupted: %p, %d (%d)\n", uaddr, current->pid);
1358 restart = ¤t_thread_info()->restart_block;
1359 restart->fn = futex_lock_pi_restart;
1360 restart->arg0 = (unsigned long) uaddr;
1361 restart->arg1 = detect;
1363 restart->arg2 = to->timer.expires.tv64 & 0xFFFFFFFF;
1364 restart->arg3 = to->timer.expires.tv64 >> 32;
1366 restart->arg2 = restart->arg3 = 0;
1368 return -ERESTART_RESTARTBLOCK;
1372 * Userspace attempted a TID -> 0 atomic transition, and failed.
1373 * This is the in-kernel slowpath: we look up the PI state (if any),
1374 * and do the rt-mutex unlock.
1376 static int futex_unlock_pi(u32 __user *uaddr)
1378 struct futex_hash_bucket *hb;
1379 struct futex_q *this, *next;
1381 struct list_head *head;
1382 union futex_key key;
1383 int ret, attempt = 0;
1386 if (get_user(uval, uaddr))
1389 * We release only a lock we actually own:
1391 if ((uval & FUTEX_TID_MASK) != current->pid)
1394 * First take all the futex related locks:
1396 down_read(¤t->mm->mmap_sem);
1398 ret = get_futex_key(uaddr, &key);
1399 if (unlikely(ret != 0))
1402 hb = hash_futex(&key);
1403 spin_lock(&hb->lock);
1407 * To avoid races, try to do the TID -> 0 atomic transition
1408 * again. If it succeeds then we can return without waking
1411 inc_preempt_count();
1412 uval = futex_atomic_cmpxchg_inatomic(uaddr, current->pid, 0);
1413 dec_preempt_count();
1415 if (unlikely(uval == -EFAULT))
1418 * Rare case: we managed to release the lock atomically,
1419 * no need to wake anyone else up:
1421 if (unlikely(uval == current->pid))
1425 * Ok, other tasks may need to be woken up - check waiters
1426 * and do the wakeup if necessary:
1430 list_for_each_entry_safe(this, next, head, list) {
1431 if (!match_futex (&this->key, &key))
1433 ret = wake_futex_pi(uaddr, uval, this);
1435 * The atomic access to the futex value
1436 * generated a pagefault, so retry the
1437 * user-access and the wakeup:
1444 * No waiters - kernel unlocks the futex:
1446 ret = unlock_futex_pi(uaddr, uval);
1451 spin_unlock(&hb->lock);
1453 up_read(¤t->mm->mmap_sem);
1459 * We have to r/w *(int __user *)uaddr, but we can't modify it
1460 * non-atomically. Therefore, if get_user below is not
1461 * enough, we need to handle the fault ourselves, while
1462 * still holding the mmap_sem.
1465 if (futex_handle_fault((unsigned long)uaddr, attempt))
1471 spin_unlock(&hb->lock);
1472 up_read(¤t->mm->mmap_sem);
1474 ret = get_user(uval, uaddr);
1475 if (!ret && (uval != -EFAULT))
1481 static int futex_close(struct inode *inode, struct file *filp)
1483 struct futex_q *q = filp->private_data;
1491 /* This is one-shot: once it's gone off you need a new fd */
1492 static unsigned int futex_poll(struct file *filp,
1493 struct poll_table_struct *wait)
1495 struct futex_q *q = filp->private_data;
1498 poll_wait(filp, &q->waiters, wait);
1501 * list_empty() is safe here without any lock.
1502 * q->lock_ptr != 0 is not safe, because of ordering against wakeup.
1504 if (list_empty(&q->list))
1505 ret = POLLIN | POLLRDNORM;
1510 static struct file_operations futex_fops = {
1511 .release = futex_close,
1516 * Signal allows caller to avoid the race which would occur if they
1517 * set the sigio stuff up afterwards.
1519 static int futex_fd(u32 __user *uaddr, int signal)
1526 if (!valid_signal(signal))
1529 ret = get_unused_fd();
1532 filp = get_empty_filp();
1538 filp->f_op = &futex_fops;
1539 filp->f_vfsmnt = mntget(futex_mnt);
1540 filp->f_dentry = dget(futex_mnt->mnt_root);
1541 filp->f_mapping = filp->f_dentry->d_inode->i_mapping;
1544 err = f_setown(filp, current->pid, 1);
1548 filp->f_owner.signum = signal;
1551 q = kmalloc(sizeof(*q), GFP_KERNEL);
1558 down_read(¤t->mm->mmap_sem);
1559 err = get_futex_key(uaddr, &q->key);
1561 if (unlikely(err != 0)) {
1562 up_read(¤t->mm->mmap_sem);
1568 * queue_me() must be called before releasing mmap_sem, because
1569 * key->shared.inode needs to be referenced while holding it.
1571 filp->private_data = q;
1573 queue_me(q, ret, filp);
1574 up_read(¤t->mm->mmap_sem);
1576 /* Now we map fd to filp, so userspace can access it */
1577 fd_install(ret, filp);
1588 * Support for robust futexes: the kernel cleans up held futexes at
1591 * Implementation: user-space maintains a per-thread list of locks it
1592 * is holding. Upon do_exit(), the kernel carefully walks this list,
1593 * and marks all locks that are owned by this thread with the
1594 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1595 * always manipulated with the lock held, so the list is private and
1596 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1597 * field, to allow the kernel to clean up if the thread dies after
1598 * acquiring the lock, but just before it could have added itself to
1599 * the list. There can only be one such pending lock.
1603 * sys_set_robust_list - set the robust-futex list head of a task
1604 * @head: pointer to the list-head
1605 * @len: length of the list-head, as userspace expects
1608 sys_set_robust_list(struct robust_list_head __user *head,
1612 * The kernel knows only one size for now:
1614 if (unlikely(len != sizeof(*head)))
1617 current->robust_list = head;
1623 * sys_get_robust_list - get the robust-futex list head of a task
1624 * @pid: pid of the process [zero for current task]
1625 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1626 * @len_ptr: pointer to a length field, the kernel fills in the header size
1629 sys_get_robust_list(int pid, struct robust_list_head __user **head_ptr,
1630 size_t __user *len_ptr)
1632 struct robust_list_head *head;
1636 head = current->robust_list;
1638 struct task_struct *p;
1641 read_lock(&tasklist_lock);
1642 p = find_task_by_pid(pid);
1646 if ((current->euid != p->euid) && (current->euid != p->uid) &&
1647 !capable(CAP_SYS_PTRACE))
1649 head = p->robust_list;
1650 read_unlock(&tasklist_lock);
1653 if (put_user(sizeof(*head), len_ptr))
1655 return put_user(head, head_ptr);
1658 read_unlock(&tasklist_lock);
1664 * Process a futex-list entry, check whether it's owned by the
1665 * dying task, and do notification if so:
1667 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr)
1672 if (get_user(uval, uaddr))
1675 if ((uval & FUTEX_TID_MASK) == curr->pid) {
1677 * Ok, this dying thread is truly holding a futex
1678 * of interest. Set the OWNER_DIED bit atomically
1679 * via cmpxchg, and if the value had FUTEX_WAITERS
1680 * set, wake up a waiter (if any). (We have to do a
1681 * futex_wake() even if OWNER_DIED is already set -
1682 * to handle the rare but possible case of recursive
1683 * thread-death.) The rest of the cleanup is done in
1686 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval,
1687 uval | FUTEX_OWNER_DIED);
1688 if (nval == -EFAULT)
1694 if (uval & FUTEX_WAITERS)
1695 futex_wake(uaddr, 1);
1701 * Walk curr->robust_list (very carefully, it's a userspace list!)
1702 * and mark any locks found there dead, and notify any waiters.
1704 * We silently return on any sign of list-walking problem.
1706 void exit_robust_list(struct task_struct *curr)
1708 struct robust_list_head __user *head = curr->robust_list;
1709 struct robust_list __user *entry, *pending;
1710 unsigned int limit = ROBUST_LIST_LIMIT;
1711 unsigned long futex_offset;
1714 * Fetch the list head (which was registered earlier, via
1715 * sys_set_robust_list()):
1717 if (get_user(entry, &head->list.next))
1720 * Fetch the relative futex offset:
1722 if (get_user(futex_offset, &head->futex_offset))
1725 * Fetch any possibly pending lock-add first, and handle it
1728 if (get_user(pending, &head->list_op_pending))
1731 handle_futex_death((void *)pending + futex_offset, curr);
1733 while (entry != &head->list) {
1735 * A pending lock might already be on the list, so
1736 * don't process it twice:
1738 if (entry != pending)
1739 if (handle_futex_death((void *)entry + futex_offset,
1743 * Fetch the next entry in the list:
1745 if (get_user(entry, &entry->next))
1748 * Avoid excessively long or circular lists:
1757 long do_futex(u32 __user *uaddr, int op, u32 val, unsigned long timeout,
1758 u32 __user *uaddr2, u32 val2, u32 val3)
1764 ret = futex_wait(uaddr, val, timeout);
1767 ret = futex_wake(uaddr, val);
1770 /* non-zero val means F_SETOWN(getpid()) & F_SETSIG(val) */
1771 ret = futex_fd(uaddr, val);
1774 ret = futex_requeue(uaddr, uaddr2, val, val2, NULL);
1776 case FUTEX_CMP_REQUEUE:
1777 ret = futex_requeue(uaddr, uaddr2, val, val2, &val3);
1780 ret = futex_wake_op(uaddr, uaddr2, val, val2, val3);
1783 ret = futex_lock_pi(uaddr, val, timeout, val2, 0);
1785 case FUTEX_UNLOCK_PI:
1786 ret = futex_unlock_pi(uaddr);
1788 case FUTEX_TRYLOCK_PI:
1789 ret = futex_lock_pi(uaddr, 0, timeout, val2, 1);
1798 asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
1799 struct timespec __user *utime, u32 __user *uaddr2,
1803 unsigned long timeout = MAX_SCHEDULE_TIMEOUT;
1806 if (utime && (op == FUTEX_WAIT || op == FUTEX_LOCK_PI)) {
1807 if (copy_from_user(&t, utime, sizeof(t)) != 0)
1809 if (!timespec_valid(&t))
1811 if (op == FUTEX_WAIT)
1812 timeout = timespec_to_jiffies(&t) + 1;
1819 * requeue parameter in 'utime' if op == FUTEX_REQUEUE.
1821 if (op == FUTEX_REQUEUE || op == FUTEX_CMP_REQUEUE)
1822 val2 = (u32) (unsigned long) utime;
1824 return do_futex(uaddr, op, val, timeout, uaddr2, val2, val3);
1827 static int futexfs_get_sb(struct file_system_type *fs_type,
1828 int flags, const char *dev_name, void *data,
1829 struct vfsmount *mnt)
1831 return get_sb_pseudo(fs_type, "futex", NULL, 0xBAD1DEA, mnt);
1834 static struct file_system_type futex_fs_type = {
1836 .get_sb = futexfs_get_sb,
1837 .kill_sb = kill_anon_super,
1840 static int __init init(void)
1844 register_filesystem(&futex_fs_type);
1845 futex_mnt = kern_mount(&futex_fs_type);
1847 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
1848 INIT_LIST_HEAD(&futex_queues[i].chain);
1849 spin_lock_init(&futex_queues[i].lock);