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 <linux/module.h>
52 #include <asm/futex.h>
54 #include "rtmutex_common.h"
56 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
59 * Priority Inheritance state:
61 struct futex_pi_state {
63 * list of 'owned' pi_state instances - these have to be
64 * cleaned up in do_exit() if the task exits prematurely:
66 struct list_head list;
71 struct rt_mutex pi_mutex;
73 struct task_struct *owner;
80 * We use this hashed waitqueue instead of a normal wait_queue_t, so
81 * we can wake only the relevant ones (hashed queues may be shared).
83 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
84 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
85 * The order of wakup is always to make the first condition true, then
86 * wake up q->waiters, then make the second condition true.
89 struct plist_node list;
90 wait_queue_head_t waiters;
92 /* Which hash list lock to use: */
95 /* Key which the futex is hashed on: */
98 /* For fd, sigio sent using these: */
102 /* Optional priority inheritance state: */
103 struct futex_pi_state *pi_state;
104 struct task_struct *task;
108 * Split the global futex_lock into every hash list lock.
110 struct futex_hash_bucket {
112 struct plist_head chain;
115 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
117 /* Futex-fs vfsmount entry: */
118 static struct vfsmount *futex_mnt;
121 * We hash on the keys returned from get_futex_key (see below).
123 static struct futex_hash_bucket *hash_futex(union futex_key *key)
125 u32 hash = jhash2((u32*)&key->both.word,
126 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
128 return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
132 * Return 1 if two futex_keys are equal, 0 otherwise.
134 static inline int match_futex(union futex_key *key1, union futex_key *key2)
136 return (key1->both.word == key2->both.word
137 && key1->both.ptr == key2->both.ptr
138 && key1->both.offset == key2->both.offset);
142 * Get parameters which are the keys for a futex.
144 * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
145 * offset_within_page). For private mappings, it's (uaddr, current->mm).
146 * We can usually work out the index without swapping in the page.
148 * Returns: 0, or negative error code.
149 * The key words are stored in *key on success.
151 * Should be called with ¤t->mm->mmap_sem but NOT any spinlocks.
153 int get_futex_key(u32 __user *uaddr, union futex_key *key)
155 unsigned long address = (unsigned long)uaddr;
156 struct mm_struct *mm = current->mm;
157 struct vm_area_struct *vma;
162 * The futex address must be "naturally" aligned.
164 key->both.offset = address % PAGE_SIZE;
165 if (unlikely((key->both.offset % sizeof(u32)) != 0))
167 address -= key->both.offset;
170 * The futex is hashed differently depending on whether
171 * it's in a shared or private mapping. So check vma first.
173 vma = find_extend_vma(mm, address);
180 if (unlikely((vma->vm_flags & (VM_IO|VM_READ)) != VM_READ))
181 return (vma->vm_flags & VM_IO) ? -EPERM : -EACCES;
184 * Private mappings are handled in a simple way.
186 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
187 * it's a read-only handle, it's expected that futexes attach to
188 * the object not the particular process. Therefore we use
189 * VM_MAYSHARE here, not VM_SHARED which is restricted to shared
190 * mappings of _writable_ handles.
192 if (likely(!(vma->vm_flags & VM_MAYSHARE))) {
193 key->private.mm = mm;
194 key->private.address = address;
199 * Linear file mappings are also simple.
201 key->shared.inode = vma->vm_file->f_path.dentry->d_inode;
202 key->both.offset++; /* Bit 0 of offset indicates inode-based key. */
203 if (likely(!(vma->vm_flags & VM_NONLINEAR))) {
204 key->shared.pgoff = (((address - vma->vm_start) >> PAGE_SHIFT)
210 * We could walk the page table to read the non-linear
211 * pte, and get the page index without fetching the page
212 * from swap. But that's a lot of code to duplicate here
213 * for a rare case, so we simply fetch the page.
215 err = get_user_pages(current, mm, address, 1, 0, 0, &page, NULL);
218 page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
224 EXPORT_SYMBOL_GPL(get_futex_key);
227 * Take a reference to the resource addressed by a key.
228 * Can be called while holding spinlocks.
230 * NOTE: mmap_sem MUST be held between get_futex_key() and calling this
231 * function, if it is called at all. mmap_sem keeps key->shared.inode valid.
233 inline void get_futex_key_refs(union futex_key *key)
235 if (key->both.ptr != 0) {
236 if (key->both.offset & 1)
237 atomic_inc(&key->shared.inode->i_count);
239 atomic_inc(&key->private.mm->mm_count);
242 EXPORT_SYMBOL_GPL(get_futex_key_refs);
245 * Drop a reference to the resource addressed by a key.
246 * The hash bucket spinlock must not be held.
248 void drop_futex_key_refs(union futex_key *key)
250 if (key->both.ptr != 0) {
251 if (key->both.offset & 1)
252 iput(key->shared.inode);
254 mmdrop(key->private.mm);
257 EXPORT_SYMBOL_GPL(drop_futex_key_refs);
259 static inline int get_futex_value_locked(u32 *dest, u32 __user *from)
264 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
267 return ret ? -EFAULT : 0;
271 * Fault handling. Called with current->mm->mmap_sem held.
273 static int futex_handle_fault(unsigned long address, int attempt)
275 struct vm_area_struct * vma;
276 struct mm_struct *mm = current->mm;
278 if (attempt > 2 || !(vma = find_vma(mm, address)) ||
279 vma->vm_start > address || !(vma->vm_flags & VM_WRITE))
282 switch (handle_mm_fault(mm, vma, address, 1)) {
298 static int refill_pi_state_cache(void)
300 struct futex_pi_state *pi_state;
302 if (likely(current->pi_state_cache))
305 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
310 INIT_LIST_HEAD(&pi_state->list);
311 /* pi_mutex gets initialized later */
312 pi_state->owner = NULL;
313 atomic_set(&pi_state->refcount, 1);
315 current->pi_state_cache = pi_state;
320 static struct futex_pi_state * alloc_pi_state(void)
322 struct futex_pi_state *pi_state = current->pi_state_cache;
325 current->pi_state_cache = NULL;
330 static void free_pi_state(struct futex_pi_state *pi_state)
332 if (!atomic_dec_and_test(&pi_state->refcount))
336 * If pi_state->owner is NULL, the owner is most probably dying
337 * and has cleaned up the pi_state already
339 if (pi_state->owner) {
340 spin_lock_irq(&pi_state->owner->pi_lock);
341 list_del_init(&pi_state->list);
342 spin_unlock_irq(&pi_state->owner->pi_lock);
344 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
347 if (current->pi_state_cache)
351 * pi_state->list is already empty.
352 * clear pi_state->owner.
353 * refcount is at 0 - put it back to 1.
355 pi_state->owner = NULL;
356 atomic_set(&pi_state->refcount, 1);
357 current->pi_state_cache = pi_state;
362 * Look up the task based on what TID userspace gave us.
365 static struct task_struct * futex_find_get_task(pid_t pid)
367 struct task_struct *p;
370 p = find_task_by_pid(pid);
373 if ((current->euid != p->euid) && (current->euid != p->uid)) {
377 if (p->exit_state != 0) {
389 * This task is holding PI mutexes at exit time => bad.
390 * Kernel cleans up PI-state, but userspace is likely hosed.
391 * (Robust-futex cleanup is separate and might save the day for userspace.)
393 void exit_pi_state_list(struct task_struct *curr)
395 struct list_head *next, *head = &curr->pi_state_list;
396 struct futex_pi_state *pi_state;
397 struct futex_hash_bucket *hb;
401 * We are a ZOMBIE and nobody can enqueue itself on
402 * pi_state_list anymore, but we have to be careful
403 * versus waiters unqueueing themselves:
405 spin_lock_irq(&curr->pi_lock);
406 while (!list_empty(head)) {
409 pi_state = list_entry(next, struct futex_pi_state, list);
411 hb = hash_futex(&key);
412 spin_unlock_irq(&curr->pi_lock);
414 spin_lock(&hb->lock);
416 spin_lock_irq(&curr->pi_lock);
418 * We dropped the pi-lock, so re-check whether this
419 * task still owns the PI-state:
421 if (head->next != next) {
422 spin_unlock(&hb->lock);
426 WARN_ON(pi_state->owner != curr);
427 WARN_ON(list_empty(&pi_state->list));
428 list_del_init(&pi_state->list);
429 pi_state->owner = NULL;
430 spin_unlock_irq(&curr->pi_lock);
432 rt_mutex_unlock(&pi_state->pi_mutex);
434 spin_unlock(&hb->lock);
436 spin_lock_irq(&curr->pi_lock);
438 spin_unlock_irq(&curr->pi_lock);
442 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb, struct futex_q *me)
444 struct futex_pi_state *pi_state = NULL;
445 struct futex_q *this, *next;
446 struct plist_head *head;
447 struct task_struct *p;
452 plist_for_each_entry_safe(this, next, head, list) {
453 if (match_futex(&this->key, &me->key)) {
455 * Another waiter already exists - bump up
456 * the refcount and return its pi_state:
458 pi_state = this->pi_state;
460 * Userspace might have messed up non PI and PI futexes
462 if (unlikely(!pi_state))
465 WARN_ON(!atomic_read(&pi_state->refcount));
467 atomic_inc(&pi_state->refcount);
468 me->pi_state = pi_state;
475 * We are the first waiter - try to look up the real owner and attach
476 * the new pi_state to it, but bail out when the owner died bit is set
479 pid = uval & FUTEX_TID_MASK;
480 if (!pid && (uval & FUTEX_OWNER_DIED))
482 p = futex_find_get_task(pid);
486 pi_state = alloc_pi_state();
489 * Initialize the pi_mutex in locked state and make 'p'
492 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
494 /* Store the key for possible exit cleanups: */
495 pi_state->key = me->key;
497 spin_lock_irq(&p->pi_lock);
498 WARN_ON(!list_empty(&pi_state->list));
499 list_add(&pi_state->list, &p->pi_state_list);
501 spin_unlock_irq(&p->pi_lock);
505 me->pi_state = pi_state;
511 * The hash bucket lock must be held when this is called.
512 * Afterwards, the futex_q must not be accessed.
514 static void wake_futex(struct futex_q *q)
516 plist_del(&q->list, &q->list.plist);
518 send_sigio(&q->filp->f_owner, q->fd, POLL_IN);
520 * The lock in wake_up_all() is a crucial memory barrier after the
521 * plist_del() and also before assigning to q->lock_ptr.
523 wake_up_all(&q->waiters);
525 * The waiting task can free the futex_q as soon as this is written,
526 * without taking any locks. This must come last.
528 * A memory barrier is required here to prevent the following store
529 * to lock_ptr from getting ahead of the wakeup. Clearing the lock
530 * at the end of wake_up_all() does not prevent this store from
537 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
539 struct task_struct *new_owner;
540 struct futex_pi_state *pi_state = this->pi_state;
546 spin_lock(&pi_state->pi_mutex.wait_lock);
547 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
550 * This happens when we have stolen the lock and the original
551 * pending owner did not enqueue itself back on the rt_mutex.
552 * Thats not a tragedy. We know that way, that a lock waiter
553 * is on the fly. We make the futex_q waiter the pending owner.
556 new_owner = this->task;
559 * We pass it to the next owner. (The WAITERS bit is always
560 * kept enabled while there is PI state around. We must also
561 * preserve the owner died bit.)
563 if (!(uval & FUTEX_OWNER_DIED)) {
564 newval = FUTEX_WAITERS | new_owner->pid;
567 curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
569 if (curval == -EFAULT)
575 spin_lock_irq(&pi_state->owner->pi_lock);
576 WARN_ON(list_empty(&pi_state->list));
577 list_del_init(&pi_state->list);
578 spin_unlock_irq(&pi_state->owner->pi_lock);
580 spin_lock_irq(&new_owner->pi_lock);
581 WARN_ON(!list_empty(&pi_state->list));
582 list_add(&pi_state->list, &new_owner->pi_state_list);
583 pi_state->owner = new_owner;
584 spin_unlock_irq(&new_owner->pi_lock);
586 spin_unlock(&pi_state->pi_mutex.wait_lock);
587 rt_mutex_unlock(&pi_state->pi_mutex);
592 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
597 * There is no waiter, so we unlock the futex. The owner died
598 * bit has not to be preserved here. We are the owner:
601 oldval = futex_atomic_cmpxchg_inatomic(uaddr, uval, 0);
604 if (oldval == -EFAULT)
613 * Express the locking dependencies for lockdep:
616 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
619 spin_lock(&hb1->lock);
621 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
622 } else { /* hb1 > hb2 */
623 spin_lock(&hb2->lock);
624 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
629 * Wake up all waiters hashed on the physical page that is mapped
630 * to this virtual address:
632 static int futex_wake(u32 __user *uaddr, int nr_wake)
634 struct futex_hash_bucket *hb;
635 struct futex_q *this, *next;
636 struct plist_head *head;
640 down_read(¤t->mm->mmap_sem);
642 ret = get_futex_key(uaddr, &key);
643 if (unlikely(ret != 0))
646 hb = hash_futex(&key);
647 spin_lock(&hb->lock);
650 plist_for_each_entry_safe(this, next, head, list) {
651 if (match_futex (&this->key, &key)) {
652 if (this->pi_state) {
657 if (++ret >= nr_wake)
662 spin_unlock(&hb->lock);
664 up_read(¤t->mm->mmap_sem);
669 * Wake up all waiters hashed on the physical page that is mapped
670 * to this virtual address:
673 futex_wake_op(u32 __user *uaddr1, u32 __user *uaddr2,
674 int nr_wake, int nr_wake2, int op)
676 union futex_key key1, key2;
677 struct futex_hash_bucket *hb1, *hb2;
678 struct plist_head *head;
679 struct futex_q *this, *next;
680 int ret, op_ret, attempt = 0;
683 down_read(¤t->mm->mmap_sem);
685 ret = get_futex_key(uaddr1, &key1);
686 if (unlikely(ret != 0))
688 ret = get_futex_key(uaddr2, &key2);
689 if (unlikely(ret != 0))
692 hb1 = hash_futex(&key1);
693 hb2 = hash_futex(&key2);
696 double_lock_hb(hb1, hb2);
698 op_ret = futex_atomic_op_inuser(op, uaddr2);
699 if (unlikely(op_ret < 0)) {
702 spin_unlock(&hb1->lock);
704 spin_unlock(&hb2->lock);
708 * we don't get EFAULT from MMU faults if we don't have an MMU,
709 * but we might get them from range checking
715 if (unlikely(op_ret != -EFAULT)) {
721 * futex_atomic_op_inuser needs to both read and write
722 * *(int __user *)uaddr2, but we can't modify it
723 * non-atomically. Therefore, if get_user below is not
724 * enough, we need to handle the fault ourselves, while
725 * still holding the mmap_sem.
728 if (futex_handle_fault((unsigned long)uaddr2,
737 * If we would have faulted, release mmap_sem,
738 * fault it in and start all over again.
740 up_read(¤t->mm->mmap_sem);
742 ret = get_user(dummy, uaddr2);
751 plist_for_each_entry_safe(this, next, head, list) {
752 if (match_futex (&this->key, &key1)) {
754 if (++ret >= nr_wake)
763 plist_for_each_entry_safe(this, next, head, list) {
764 if (match_futex (&this->key, &key2)) {
766 if (++op_ret >= nr_wake2)
773 spin_unlock(&hb1->lock);
775 spin_unlock(&hb2->lock);
777 up_read(¤t->mm->mmap_sem);
782 * Requeue all waiters hashed on one physical page to another
785 static int futex_requeue(u32 __user *uaddr1, u32 __user *uaddr2,
786 int nr_wake, int nr_requeue, u32 *cmpval)
788 union futex_key key1, key2;
789 struct futex_hash_bucket *hb1, *hb2;
790 struct plist_head *head1;
791 struct futex_q *this, *next;
792 int ret, drop_count = 0;
795 down_read(¤t->mm->mmap_sem);
797 ret = get_futex_key(uaddr1, &key1);
798 if (unlikely(ret != 0))
800 ret = get_futex_key(uaddr2, &key2);
801 if (unlikely(ret != 0))
804 hb1 = hash_futex(&key1);
805 hb2 = hash_futex(&key2);
807 double_lock_hb(hb1, hb2);
809 if (likely(cmpval != NULL)) {
812 ret = get_futex_value_locked(&curval, uaddr1);
815 spin_unlock(&hb1->lock);
817 spin_unlock(&hb2->lock);
820 * If we would have faulted, release mmap_sem, fault
821 * it in and start all over again.
823 up_read(¤t->mm->mmap_sem);
825 ret = get_user(curval, uaddr1);
832 if (curval != *cmpval) {
839 plist_for_each_entry_safe(this, next, head1, list) {
840 if (!match_futex (&this->key, &key1))
842 if (++ret <= nr_wake) {
846 * If key1 and key2 hash to the same bucket, no need to
849 if (likely(head1 != &hb2->chain)) {
850 plist_del(&this->list, &hb1->chain);
851 plist_add(&this->list, &hb2->chain);
852 this->lock_ptr = &hb2->lock;
853 #ifdef CONFIG_DEBUG_PI_LIST
854 this->list.plist.lock = &hb2->lock;
858 get_futex_key_refs(&key2);
861 if (ret - nr_wake >= nr_requeue)
867 spin_unlock(&hb1->lock);
869 spin_unlock(&hb2->lock);
871 /* drop_futex_key_refs() must be called outside the spinlocks. */
872 while (--drop_count >= 0)
873 drop_futex_key_refs(&key1);
876 up_read(¤t->mm->mmap_sem);
880 /* The key must be already stored in q->key. */
881 static inline struct futex_hash_bucket *
882 queue_lock(struct futex_q *q, int fd, struct file *filp)
884 struct futex_hash_bucket *hb;
889 init_waitqueue_head(&q->waiters);
891 get_futex_key_refs(&q->key);
892 hb = hash_futex(&q->key);
893 q->lock_ptr = &hb->lock;
895 spin_lock(&hb->lock);
899 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
904 * The priority used to register this element is
905 * - either the real thread-priority for the real-time threads
906 * (i.e. threads with a priority lower than MAX_RT_PRIO)
907 * - or MAX_RT_PRIO for non-RT threads.
908 * Thus, all RT-threads are woken first in priority order, and
909 * the others are woken last, in FIFO order.
911 prio = min(current->normal_prio, MAX_RT_PRIO);
913 plist_node_init(&q->list, prio);
914 #ifdef CONFIG_DEBUG_PI_LIST
915 q->list.plist.lock = &hb->lock;
917 plist_add(&q->list, &hb->chain);
919 spin_unlock(&hb->lock);
923 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
925 spin_unlock(&hb->lock);
926 drop_futex_key_refs(&q->key);
930 * queue_me and unqueue_me must be called as a pair, each
931 * exactly once. They are called with the hashed spinlock held.
934 /* The key must be already stored in q->key. */
935 static void queue_me(struct futex_q *q, int fd, struct file *filp)
937 struct futex_hash_bucket *hb;
939 hb = queue_lock(q, fd, filp);
943 /* Return 1 if we were still queued (ie. 0 means we were woken) */
944 static int unqueue_me(struct futex_q *q)
946 spinlock_t *lock_ptr;
949 /* In the common case we don't take the spinlock, which is nice. */
951 lock_ptr = q->lock_ptr;
956 * q->lock_ptr can change between reading it and
957 * spin_lock(), causing us to take the wrong lock. This
958 * corrects the race condition.
960 * Reasoning goes like this: if we have the wrong lock,
961 * q->lock_ptr must have changed (maybe several times)
962 * between reading it and the spin_lock(). It can
963 * change again after the spin_lock() but only if it was
964 * already changed before the spin_lock(). It cannot,
965 * however, change back to the original value. Therefore
966 * we can detect whether we acquired the correct lock.
968 if (unlikely(lock_ptr != q->lock_ptr)) {
969 spin_unlock(lock_ptr);
972 WARN_ON(plist_node_empty(&q->list));
973 plist_del(&q->list, &q->list.plist);
977 spin_unlock(lock_ptr);
981 drop_futex_key_refs(&q->key);
986 * PI futexes can not be requeued and must remove themself from the
987 * hash bucket. The hash bucket lock is held on entry and dropped here.
989 static void unqueue_me_pi(struct futex_q *q, struct futex_hash_bucket *hb)
991 WARN_ON(plist_node_empty(&q->list));
992 plist_del(&q->list, &q->list.plist);
994 BUG_ON(!q->pi_state);
995 free_pi_state(q->pi_state);
998 spin_unlock(&hb->lock);
1000 drop_futex_key_refs(&q->key);
1003 static long futex_wait_restart(struct restart_block *restart);
1004 static int futex_wait_abstime(u32 __user *uaddr, u32 val,
1005 int timed, unsigned long abs_time)
1007 struct task_struct *curr = current;
1008 DECLARE_WAITQUEUE(wait, curr);
1009 struct futex_hash_bucket *hb;
1011 unsigned long time_left = 0;
1017 down_read(&curr->mm->mmap_sem);
1019 ret = get_futex_key(uaddr, &q.key);
1020 if (unlikely(ret != 0))
1021 goto out_release_sem;
1023 hb = queue_lock(&q, -1, NULL);
1026 * Access the page AFTER the futex is queued.
1027 * Order is important:
1029 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1030 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1032 * The basic logical guarantee of a futex is that it blocks ONLY
1033 * if cond(var) is known to be true at the time of blocking, for
1034 * any cond. If we queued after testing *uaddr, that would open
1035 * a race condition where we could block indefinitely with
1036 * cond(var) false, which would violate the guarantee.
1038 * A consequence is that futex_wait() can return zero and absorb
1039 * a wakeup when *uaddr != val on entry to the syscall. This is
1042 * We hold the mmap semaphore, so the mapping cannot have changed
1043 * since we looked it up in get_futex_key.
1045 ret = get_futex_value_locked(&uval, uaddr);
1047 if (unlikely(ret)) {
1048 queue_unlock(&q, hb);
1051 * If we would have faulted, release mmap_sem, fault it in and
1052 * start all over again.
1054 up_read(&curr->mm->mmap_sem);
1056 ret = get_user(uval, uaddr);
1064 goto out_unlock_release_sem;
1066 /* Only actually queue if *uaddr contained val. */
1070 * Now the futex is queued and we have checked the data, we
1071 * don't want to hold mmap_sem while we sleep.
1073 up_read(&curr->mm->mmap_sem);
1076 * There might have been scheduling since the queue_me(), as we
1077 * cannot hold a spinlock across the get_user() in case it
1078 * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1079 * queueing ourselves into the futex hash. This code thus has to
1080 * rely on the futex_wake() code removing us from hash when it
1084 /* add_wait_queue is the barrier after __set_current_state. */
1085 __set_current_state(TASK_INTERRUPTIBLE);
1086 add_wait_queue(&q.waiters, &wait);
1088 * !plist_node_empty() is safe here without any lock.
1089 * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1092 if (likely(!plist_node_empty(&q.list))) {
1093 unsigned long rel_time;
1096 unsigned long now = jiffies;
1097 if (time_after(now, abs_time))
1100 rel_time = abs_time - now;
1102 rel_time = MAX_SCHEDULE_TIMEOUT;
1104 time_left = schedule_timeout(rel_time);
1106 __set_current_state(TASK_RUNNING);
1109 * NOTE: we don't remove ourselves from the waitqueue because
1110 * we are the only user of it.
1113 /* If we were woken (and unqueued), we succeeded, whatever. */
1114 if (!unqueue_me(&q))
1120 * We expect signal_pending(current), but another thread may
1121 * have handled it for us already.
1123 if (time_left == MAX_SCHEDULE_TIMEOUT)
1124 return -ERESTARTSYS;
1126 struct restart_block *restart;
1127 restart = ¤t_thread_info()->restart_block;
1128 restart->fn = futex_wait_restart;
1129 restart->arg0 = (unsigned long)uaddr;
1130 restart->arg1 = (unsigned long)val;
1131 restart->arg2 = (unsigned long)timed;
1132 restart->arg3 = abs_time;
1133 return -ERESTART_RESTARTBLOCK;
1136 out_unlock_release_sem:
1137 queue_unlock(&q, hb);
1140 up_read(&curr->mm->mmap_sem);
1144 static int futex_wait(u32 __user *uaddr, u32 val, unsigned long rel_time)
1146 int timed = (rel_time != MAX_SCHEDULE_TIMEOUT);
1147 return futex_wait_abstime(uaddr, val, timed, jiffies+rel_time);
1150 static long futex_wait_restart(struct restart_block *restart)
1152 u32 __user *uaddr = (u32 __user *)restart->arg0;
1153 u32 val = (u32)restart->arg1;
1154 int timed = (int)restart->arg2;
1155 unsigned long abs_time = restart->arg3;
1157 restart->fn = do_no_restart_syscall;
1158 return (long)futex_wait_abstime(uaddr, val, timed, abs_time);
1163 * Userspace tried a 0 -> TID atomic transition of the futex value
1164 * and failed. The kernel side here does the whole locking operation:
1165 * if there are waiters then it will block, it does PI, etc. (Due to
1166 * races the kernel might see a 0 value of the futex too.)
1168 static int futex_lock_pi(u32 __user *uaddr, int detect, unsigned long sec,
1169 long nsec, int trylock)
1171 struct hrtimer_sleeper timeout, *to = NULL;
1172 struct task_struct *curr = current;
1173 struct futex_hash_bucket *hb;
1174 u32 uval, newval, curval;
1176 int ret, attempt = 0;
1178 if (refill_pi_state_cache())
1181 if (sec != MAX_SCHEDULE_TIMEOUT) {
1183 hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_MODE_ABS);
1184 hrtimer_init_sleeper(to, current);
1185 to->timer.expires = ktime_set(sec, nsec);
1190 down_read(&curr->mm->mmap_sem);
1192 ret = get_futex_key(uaddr, &q.key);
1193 if (unlikely(ret != 0))
1194 goto out_release_sem;
1196 hb = queue_lock(&q, -1, NULL);
1200 * To avoid races, we attempt to take the lock here again
1201 * (by doing a 0 -> TID atomic cmpxchg), while holding all
1202 * the locks. It will most likely not succeed.
1204 newval = current->pid;
1206 pagefault_disable();
1207 curval = futex_atomic_cmpxchg_inatomic(uaddr, 0, newval);
1210 if (unlikely(curval == -EFAULT))
1213 /* We own the lock already */
1214 if (unlikely((curval & FUTEX_TID_MASK) == current->pid)) {
1216 force_sig(SIGKILL, current);
1218 goto out_unlock_release_sem;
1222 * Surprise - we got the lock. Just return
1225 if (unlikely(!curval))
1226 goto out_unlock_release_sem;
1229 newval = uval | FUTEX_WAITERS;
1231 pagefault_disable();
1232 curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
1235 if (unlikely(curval == -EFAULT))
1237 if (unlikely(curval != uval))
1241 * We dont have the lock. Look up the PI state (or create it if
1242 * we are the first waiter):
1244 ret = lookup_pi_state(uval, hb, &q);
1246 if (unlikely(ret)) {
1248 * There were no waiters and the owner task lookup
1249 * failed. When the OWNER_DIED bit is set, then we
1250 * know that this is a robust futex and we actually
1251 * take the lock. This is safe as we are protected by
1252 * the hash bucket lock. We also set the waiters bit
1253 * unconditionally here, to simplify glibc handling of
1254 * multiple tasks racing to acquire the lock and
1255 * cleanup the problems which were left by the dead
1258 if (curval & FUTEX_OWNER_DIED) {
1260 newval = current->pid |
1261 FUTEX_OWNER_DIED | FUTEX_WAITERS;
1263 pagefault_disable();
1264 curval = futex_atomic_cmpxchg_inatomic(uaddr,
1268 if (unlikely(curval == -EFAULT))
1270 if (unlikely(curval != uval))
1274 goto out_unlock_release_sem;
1278 * Only actually queue now that the atomic ops are done:
1283 * Now the futex is queued and we have checked the data, we
1284 * don't want to hold mmap_sem while we sleep.
1286 up_read(&curr->mm->mmap_sem);
1288 WARN_ON(!q.pi_state);
1290 * Block on the PI mutex:
1293 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1295 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1296 /* Fixup the trylock return value: */
1297 ret = ret ? 0 : -EWOULDBLOCK;
1300 down_read(&curr->mm->mmap_sem);
1301 spin_lock(q.lock_ptr);
1304 * Got the lock. We might not be the anticipated owner if we
1305 * did a lock-steal - fix up the PI-state in that case.
1307 if (!ret && q.pi_state->owner != curr) {
1308 u32 newtid = current->pid | FUTEX_WAITERS;
1311 if (q.pi_state->owner != NULL) {
1312 spin_lock_irq(&q.pi_state->owner->pi_lock);
1313 WARN_ON(list_empty(&q.pi_state->list));
1314 list_del_init(&q.pi_state->list);
1315 spin_unlock_irq(&q.pi_state->owner->pi_lock);
1317 newtid |= FUTEX_OWNER_DIED;
1319 q.pi_state->owner = current;
1321 spin_lock_irq(¤t->pi_lock);
1322 WARN_ON(!list_empty(&q.pi_state->list));
1323 list_add(&q.pi_state->list, ¤t->pi_state_list);
1324 spin_unlock_irq(¤t->pi_lock);
1326 /* Unqueue and drop the lock */
1327 unqueue_me_pi(&q, hb);
1328 up_read(&curr->mm->mmap_sem);
1330 * We own it, so we have to replace the pending owner
1331 * TID. This must be atomic as we have preserve the
1332 * owner died bit here.
1334 ret = get_user(uval, uaddr);
1336 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1337 curval = futex_atomic_cmpxchg_inatomic(uaddr,
1339 if (curval == -EFAULT)
1347 * Catch the rare case, where the lock was released
1348 * when we were on the way back before we locked
1351 if (ret && q.pi_state->owner == curr) {
1352 if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1355 /* Unqueue and drop the lock */
1356 unqueue_me_pi(&q, hb);
1357 up_read(&curr->mm->mmap_sem);
1360 if (!detect && ret == -EDEADLK && 0)
1361 force_sig(SIGKILL, current);
1363 return ret != -EINTR ? ret : -ERESTARTNOINTR;
1365 out_unlock_release_sem:
1366 queue_unlock(&q, hb);
1369 up_read(&curr->mm->mmap_sem);
1374 * We have to r/w *(int __user *)uaddr, but we can't modify it
1375 * non-atomically. Therefore, if get_user below is not
1376 * enough, we need to handle the fault ourselves, while
1377 * still holding the mmap_sem.
1380 if (futex_handle_fault((unsigned long)uaddr, attempt)) {
1382 goto out_unlock_release_sem;
1387 queue_unlock(&q, hb);
1388 up_read(&curr->mm->mmap_sem);
1390 ret = get_user(uval, uaddr);
1391 if (!ret && (uval != -EFAULT))
1398 * Userspace attempted a TID -> 0 atomic transition, and failed.
1399 * This is the in-kernel slowpath: we look up the PI state (if any),
1400 * and do the rt-mutex unlock.
1402 static int futex_unlock_pi(u32 __user *uaddr)
1404 struct futex_hash_bucket *hb;
1405 struct futex_q *this, *next;
1407 struct plist_head *head;
1408 union futex_key key;
1409 int ret, attempt = 0;
1412 if (get_user(uval, uaddr))
1415 * We release only a lock we actually own:
1417 if ((uval & FUTEX_TID_MASK) != current->pid)
1420 * First take all the futex related locks:
1422 down_read(¤t->mm->mmap_sem);
1424 ret = get_futex_key(uaddr, &key);
1425 if (unlikely(ret != 0))
1428 hb = hash_futex(&key);
1429 spin_lock(&hb->lock);
1433 * To avoid races, try to do the TID -> 0 atomic transition
1434 * again. If it succeeds then we can return without waking
1437 if (!(uval & FUTEX_OWNER_DIED)) {
1438 pagefault_disable();
1439 uval = futex_atomic_cmpxchg_inatomic(uaddr, current->pid, 0);
1443 if (unlikely(uval == -EFAULT))
1446 * Rare case: we managed to release the lock atomically,
1447 * no need to wake anyone else up:
1449 if (unlikely(uval == current->pid))
1453 * Ok, other tasks may need to be woken up - check waiters
1454 * and do the wakeup if necessary:
1458 plist_for_each_entry_safe(this, next, head, list) {
1459 if (!match_futex (&this->key, &key))
1461 ret = wake_futex_pi(uaddr, uval, this);
1463 * The atomic access to the futex value
1464 * generated a pagefault, so retry the
1465 * user-access and the wakeup:
1472 * No waiters - kernel unlocks the futex:
1474 if (!(uval & FUTEX_OWNER_DIED)) {
1475 ret = unlock_futex_pi(uaddr, uval);
1481 spin_unlock(&hb->lock);
1483 up_read(¤t->mm->mmap_sem);
1489 * We have to r/w *(int __user *)uaddr, but we can't modify it
1490 * non-atomically. Therefore, if get_user below is not
1491 * enough, we need to handle the fault ourselves, while
1492 * still holding the mmap_sem.
1495 if (futex_handle_fault((unsigned long)uaddr, attempt)) {
1502 spin_unlock(&hb->lock);
1503 up_read(¤t->mm->mmap_sem);
1505 ret = get_user(uval, uaddr);
1506 if (!ret && (uval != -EFAULT))
1512 static int futex_close(struct inode *inode, struct file *filp)
1514 struct futex_q *q = filp->private_data;
1522 /* This is one-shot: once it's gone off you need a new fd */
1523 static unsigned int futex_poll(struct file *filp,
1524 struct poll_table_struct *wait)
1526 struct futex_q *q = filp->private_data;
1529 poll_wait(filp, &q->waiters, wait);
1532 * plist_node_empty() is safe here without any lock.
1533 * q->lock_ptr != 0 is not safe, because of ordering against wakeup.
1535 if (plist_node_empty(&q->list))
1536 ret = POLLIN | POLLRDNORM;
1541 static const struct file_operations futex_fops = {
1542 .release = futex_close,
1547 * Signal allows caller to avoid the race which would occur if they
1548 * set the sigio stuff up afterwards.
1550 static int futex_fd(u32 __user *uaddr, int signal)
1555 static unsigned long printk_interval;
1557 if (printk_timed_ratelimit(&printk_interval, 60 * 60 * 1000)) {
1558 printk(KERN_WARNING "Process `%s' used FUTEX_FD, which "
1559 "will be removed from the kernel in June 2007\n",
1564 if (!valid_signal(signal))
1567 ret = get_unused_fd();
1570 filp = get_empty_filp();
1576 filp->f_op = &futex_fops;
1577 filp->f_path.mnt = mntget(futex_mnt);
1578 filp->f_path.dentry = dget(futex_mnt->mnt_root);
1579 filp->f_mapping = filp->f_path.dentry->d_inode->i_mapping;
1582 err = __f_setown(filp, task_pid(current), PIDTYPE_PID, 1);
1586 filp->f_owner.signum = signal;
1589 q = kmalloc(sizeof(*q), GFP_KERNEL);
1596 down_read(¤t->mm->mmap_sem);
1597 err = get_futex_key(uaddr, &q->key);
1599 if (unlikely(err != 0)) {
1600 up_read(¤t->mm->mmap_sem);
1606 * queue_me() must be called before releasing mmap_sem, because
1607 * key->shared.inode needs to be referenced while holding it.
1609 filp->private_data = q;
1611 queue_me(q, ret, filp);
1612 up_read(¤t->mm->mmap_sem);
1614 /* Now we map fd to filp, so userspace can access it */
1615 fd_install(ret, filp);
1626 * Support for robust futexes: the kernel cleans up held futexes at
1629 * Implementation: user-space maintains a per-thread list of locks it
1630 * is holding. Upon do_exit(), the kernel carefully walks this list,
1631 * and marks all locks that are owned by this thread with the
1632 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1633 * always manipulated with the lock held, so the list is private and
1634 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1635 * field, to allow the kernel to clean up if the thread dies after
1636 * acquiring the lock, but just before it could have added itself to
1637 * the list. There can only be one such pending lock.
1641 * sys_set_robust_list - set the robust-futex list head of a task
1642 * @head: pointer to the list-head
1643 * @len: length of the list-head, as userspace expects
1646 sys_set_robust_list(struct robust_list_head __user *head,
1650 * The kernel knows only one size for now:
1652 if (unlikely(len != sizeof(*head)))
1655 current->robust_list = head;
1661 * sys_get_robust_list - get the robust-futex list head of a task
1662 * @pid: pid of the process [zero for current task]
1663 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1664 * @len_ptr: pointer to a length field, the kernel fills in the header size
1667 sys_get_robust_list(int pid, struct robust_list_head __user * __user *head_ptr,
1668 size_t __user *len_ptr)
1670 struct robust_list_head __user *head;
1674 head = current->robust_list;
1676 struct task_struct *p;
1680 p = find_task_by_pid(pid);
1684 if ((current->euid != p->euid) && (current->euid != p->uid) &&
1685 !capable(CAP_SYS_PTRACE))
1687 head = p->robust_list;
1691 if (put_user(sizeof(*head), len_ptr))
1693 return put_user(head, head_ptr);
1702 * Process a futex-list entry, check whether it's owned by the
1703 * dying task, and do notification if so:
1705 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
1707 u32 uval, nval, mval;
1710 if (get_user(uval, uaddr))
1713 if ((uval & FUTEX_TID_MASK) == curr->pid) {
1715 * Ok, this dying thread is truly holding a futex
1716 * of interest. Set the OWNER_DIED bit atomically
1717 * via cmpxchg, and if the value had FUTEX_WAITERS
1718 * set, wake up a waiter (if any). (We have to do a
1719 * futex_wake() even if OWNER_DIED is already set -
1720 * to handle the rare but possible case of recursive
1721 * thread-death.) The rest of the cleanup is done in
1724 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1725 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
1727 if (nval == -EFAULT)
1734 * Wake robust non-PI futexes here. The wakeup of
1735 * PI futexes happens in exit_pi_state():
1738 if (uval & FUTEX_WAITERS)
1739 futex_wake(uaddr, 1);
1746 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1748 static inline int fetch_robust_entry(struct robust_list __user **entry,
1749 struct robust_list __user * __user *head,
1752 unsigned long uentry;
1754 if (get_user(uentry, (unsigned long __user *)head))
1757 *entry = (void __user *)(uentry & ~1UL);
1764 * Walk curr->robust_list (very carefully, it's a userspace list!)
1765 * and mark any locks found there dead, and notify any waiters.
1767 * We silently return on any sign of list-walking problem.
1769 void exit_robust_list(struct task_struct *curr)
1771 struct robust_list_head __user *head = curr->robust_list;
1772 struct robust_list __user *entry, *pending;
1773 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
1774 unsigned long futex_offset;
1777 * Fetch the list head (which was registered earlier, via
1778 * sys_set_robust_list()):
1780 if (fetch_robust_entry(&entry, &head->list.next, &pi))
1783 * Fetch the relative futex offset:
1785 if (get_user(futex_offset, &head->futex_offset))
1788 * Fetch any possibly pending lock-add first, and handle it
1791 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
1795 handle_futex_death((void __user *)pending + futex_offset, curr, pip);
1797 while (entry != &head->list) {
1799 * A pending lock might already be on the list, so
1800 * don't process it twice:
1802 if (entry != pending)
1803 if (handle_futex_death((void __user *)entry + futex_offset,
1807 * Fetch the next entry in the list:
1809 if (fetch_robust_entry(&entry, &entry->next, &pi))
1812 * Avoid excessively long or circular lists:
1821 long do_futex(u32 __user *uaddr, int op, u32 val, unsigned long timeout,
1822 u32 __user *uaddr2, u32 val2, u32 val3)
1828 ret = futex_wait(uaddr, val, timeout);
1831 ret = futex_wake(uaddr, val);
1834 /* non-zero val means F_SETOWN(getpid()) & F_SETSIG(val) */
1835 ret = futex_fd(uaddr, val);
1838 ret = futex_requeue(uaddr, uaddr2, val, val2, NULL);
1840 case FUTEX_CMP_REQUEUE:
1841 ret = futex_requeue(uaddr, uaddr2, val, val2, &val3);
1844 ret = futex_wake_op(uaddr, uaddr2, val, val2, val3);
1847 ret = futex_lock_pi(uaddr, val, timeout, val2, 0);
1849 case FUTEX_UNLOCK_PI:
1850 ret = futex_unlock_pi(uaddr);
1852 case FUTEX_TRYLOCK_PI:
1853 ret = futex_lock_pi(uaddr, 0, timeout, val2, 1);
1862 asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
1863 struct timespec __user *utime, u32 __user *uaddr2,
1867 unsigned long timeout = MAX_SCHEDULE_TIMEOUT;
1870 if (utime && (op == FUTEX_WAIT || op == FUTEX_LOCK_PI)) {
1871 if (copy_from_user(&t, utime, sizeof(t)) != 0)
1873 if (!timespec_valid(&t))
1875 if (op == FUTEX_WAIT)
1876 timeout = timespec_to_jiffies(&t) + 1;
1883 * requeue parameter in 'utime' if op == FUTEX_REQUEUE.
1885 if (op == FUTEX_REQUEUE || op == FUTEX_CMP_REQUEUE)
1886 val2 = (u32) (unsigned long) utime;
1888 return do_futex(uaddr, op, val, timeout, uaddr2, val2, val3);
1891 static int futexfs_get_sb(struct file_system_type *fs_type,
1892 int flags, const char *dev_name, void *data,
1893 struct vfsmount *mnt)
1895 return get_sb_pseudo(fs_type, "futex", NULL, 0xBAD1DEA, mnt);
1898 static struct file_system_type futex_fs_type = {
1900 .get_sb = futexfs_get_sb,
1901 .kill_sb = kill_anon_super,
1904 static int __init init(void)
1906 int i = register_filesystem(&futex_fs_type);
1911 futex_mnt = kern_mount(&futex_fs_type);
1912 if (IS_ERR(futex_mnt)) {
1913 unregister_filesystem(&futex_fs_type);
1914 return PTR_ERR(futex_mnt);
1917 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
1918 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
1919 spin_lock_init(&futex_queues[i].lock);