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 * PRIVATE futexes by Eric Dumazet
20 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
22 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
23 * enough at me, Linus for the original (flawed) idea, Matthew
24 * Kirkwood for proof-of-concept implementation.
26 * "The futexes are also cursed."
27 * "But they come in a choice of three flavours!"
29 * This program is free software; you can redistribute it and/or modify
30 * it under the terms of the GNU General Public License as published by
31 * the Free Software Foundation; either version 2 of the License, or
32 * (at your option) any later version.
34 * This program is distributed in the hope that it will be useful,
35 * but WITHOUT ANY WARRANTY; without even the implied warranty of
36 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
37 * GNU General Public License for more details.
39 * You should have received a copy of the GNU General Public License
40 * along with this program; if not, write to the Free Software
41 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
43 #include <linux/slab.h>
44 #include <linux/poll.h>
46 #include <linux/file.h>
47 #include <linux/jhash.h>
48 #include <linux/init.h>
49 #include <linux/futex.h>
50 #include <linux/mount.h>
51 #include <linux/pagemap.h>
52 #include <linux/syscalls.h>
53 #include <linux/signal.h>
54 #include <linux/module.h>
55 #include <linux/magic.h>
56 #include <linux/pid.h>
57 #include <linux/nsproxy.h>
59 #include <asm/futex.h>
61 #include "rtmutex_common.h"
63 int __read_mostly futex_cmpxchg_enabled;
65 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
68 * Priority Inheritance state:
70 struct futex_pi_state {
72 * list of 'owned' pi_state instances - these have to be
73 * cleaned up in do_exit() if the task exits prematurely:
75 struct list_head list;
80 struct rt_mutex pi_mutex;
82 struct task_struct *owner;
89 * We use this hashed waitqueue instead of a normal wait_queue_t, so
90 * we can wake only the relevant ones (hashed queues may be shared).
92 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
93 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
94 * The order of wakup is always to make the first condition true, then
95 * wake up q->waiters, then make the second condition true.
98 struct plist_node list;
99 wait_queue_head_t waiters;
101 /* Which hash list lock to use: */
102 spinlock_t *lock_ptr;
104 /* Key which the futex is hashed on: */
107 /* Optional priority inheritance state: */
108 struct futex_pi_state *pi_state;
109 struct task_struct *task;
111 /* Bitset for the optional bitmasked wakeup */
116 * Split the global futex_lock into every hash list lock.
118 struct futex_hash_bucket {
120 struct plist_head chain;
123 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
126 * We hash on the keys returned from get_futex_key (see below).
128 static struct futex_hash_bucket *hash_futex(union futex_key *key)
130 u32 hash = jhash2((u32*)&key->both.word,
131 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
133 return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
137 * Return 1 if two futex_keys are equal, 0 otherwise.
139 static inline int match_futex(union futex_key *key1, union futex_key *key2)
141 return (key1->both.word == key2->both.word
142 && key1->both.ptr == key2->both.ptr
143 && key1->both.offset == key2->both.offset);
147 * Take a reference to the resource addressed by a key.
148 * Can be called while holding spinlocks.
151 static void get_futex_key_refs(union futex_key *key)
156 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
158 atomic_inc(&key->shared.inode->i_count);
160 case FUT_OFF_MMSHARED:
161 atomic_inc(&key->private.mm->mm_count);
167 * Drop a reference to the resource addressed by a key.
168 * The hash bucket spinlock must not be held.
170 static void drop_futex_key_refs(union futex_key *key)
175 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
177 iput(key->shared.inode);
179 case FUT_OFF_MMSHARED:
180 mmdrop(key->private.mm);
186 * get_futex_key - Get parameters which are the keys for a futex.
187 * @uaddr: virtual address of the futex
188 * @shared: NULL for a PROCESS_PRIVATE futex,
189 * ¤t->mm->mmap_sem for a PROCESS_SHARED futex
190 * @key: address where result is stored.
192 * Returns a negative error code or 0
193 * The key words are stored in *key on success.
195 * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
196 * offset_within_page). For private mappings, it's (uaddr, current->mm).
197 * We can usually work out the index without swapping in the page.
199 * fshared is NULL for PROCESS_PRIVATE futexes
200 * For other futexes, it points to ¤t->mm->mmap_sem and
201 * caller must have taken the reader lock. but NOT any spinlocks.
203 static int get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key)
205 unsigned long address = (unsigned long)uaddr;
206 struct mm_struct *mm = current->mm;
211 * The futex address must be "naturally" aligned.
213 key->both.offset = address % PAGE_SIZE;
214 if (unlikely((address % sizeof(u32)) != 0))
216 address -= key->both.offset;
219 * PROCESS_PRIVATE futexes are fast.
220 * As the mm cannot disappear under us and the 'key' only needs
221 * virtual address, we dont even have to find the underlying vma.
222 * Note : We do have to check 'uaddr' is a valid user address,
223 * but access_ok() should be faster than find_vma()
226 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
228 key->private.mm = mm;
229 key->private.address = address;
234 err = get_user_pages_fast(address, 1, 0, &page);
239 if (!page->mapping) {
246 * Private mappings are handled in a simple way.
248 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
249 * it's a read-only handle, it's expected that futexes attach to
250 * the object not the particular process.
252 if (PageAnon(page)) {
253 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
254 key->private.mm = mm;
255 key->private.address = address;
257 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
258 key->shared.inode = page->mapping->host;
259 key->shared.pgoff = page->index;
262 get_futex_key_refs(key);
270 void put_futex_key(int fshared, union futex_key *key)
272 drop_futex_key_refs(key);
275 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
280 curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
286 static int get_futex_value_locked(u32 *dest, u32 __user *from)
291 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
294 return ret ? -EFAULT : 0;
300 static int futex_handle_fault(unsigned long address, int attempt)
302 struct vm_area_struct * vma;
303 struct mm_struct *mm = current->mm;
309 down_read(&mm->mmap_sem);
310 vma = find_vma(mm, address);
311 if (vma && address >= vma->vm_start &&
312 (vma->vm_flags & VM_WRITE)) {
314 fault = handle_mm_fault(mm, vma, address, 1);
315 if (unlikely((fault & VM_FAULT_ERROR))) {
317 /* XXX: let's do this when we verify it is OK */
318 if (ret & VM_FAULT_OOM)
323 if (fault & VM_FAULT_MAJOR)
329 up_read(&mm->mmap_sem);
336 static int refill_pi_state_cache(void)
338 struct futex_pi_state *pi_state;
340 if (likely(current->pi_state_cache))
343 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
348 INIT_LIST_HEAD(&pi_state->list);
349 /* pi_mutex gets initialized later */
350 pi_state->owner = NULL;
351 atomic_set(&pi_state->refcount, 1);
352 pi_state->key = FUTEX_KEY_INIT;
354 current->pi_state_cache = pi_state;
359 static struct futex_pi_state * alloc_pi_state(void)
361 struct futex_pi_state *pi_state = current->pi_state_cache;
364 current->pi_state_cache = NULL;
369 static void free_pi_state(struct futex_pi_state *pi_state)
371 if (!atomic_dec_and_test(&pi_state->refcount))
375 * If pi_state->owner is NULL, the owner is most probably dying
376 * and has cleaned up the pi_state already
378 if (pi_state->owner) {
379 spin_lock_irq(&pi_state->owner->pi_lock);
380 list_del_init(&pi_state->list);
381 spin_unlock_irq(&pi_state->owner->pi_lock);
383 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
386 if (current->pi_state_cache)
390 * pi_state->list is already empty.
391 * clear pi_state->owner.
392 * refcount is at 0 - put it back to 1.
394 pi_state->owner = NULL;
395 atomic_set(&pi_state->refcount, 1);
396 current->pi_state_cache = pi_state;
401 * Look up the task based on what TID userspace gave us.
404 static struct task_struct * futex_find_get_task(pid_t pid)
406 struct task_struct *p;
409 p = find_task_by_vpid(pid);
410 if (!p || ((current->euid != p->euid) && (current->euid != p->uid)))
421 * This task is holding PI mutexes at exit time => bad.
422 * Kernel cleans up PI-state, but userspace is likely hosed.
423 * (Robust-futex cleanup is separate and might save the day for userspace.)
425 void exit_pi_state_list(struct task_struct *curr)
427 struct list_head *next, *head = &curr->pi_state_list;
428 struct futex_pi_state *pi_state;
429 struct futex_hash_bucket *hb;
430 union futex_key key = FUTEX_KEY_INIT;
432 if (!futex_cmpxchg_enabled)
435 * We are a ZOMBIE and nobody can enqueue itself on
436 * pi_state_list anymore, but we have to be careful
437 * versus waiters unqueueing themselves:
439 spin_lock_irq(&curr->pi_lock);
440 while (!list_empty(head)) {
443 pi_state = list_entry(next, struct futex_pi_state, list);
445 hb = hash_futex(&key);
446 spin_unlock_irq(&curr->pi_lock);
448 spin_lock(&hb->lock);
450 spin_lock_irq(&curr->pi_lock);
452 * We dropped the pi-lock, so re-check whether this
453 * task still owns the PI-state:
455 if (head->next != next) {
456 spin_unlock(&hb->lock);
460 WARN_ON(pi_state->owner != curr);
461 WARN_ON(list_empty(&pi_state->list));
462 list_del_init(&pi_state->list);
463 pi_state->owner = NULL;
464 spin_unlock_irq(&curr->pi_lock);
466 rt_mutex_unlock(&pi_state->pi_mutex);
468 spin_unlock(&hb->lock);
470 spin_lock_irq(&curr->pi_lock);
472 spin_unlock_irq(&curr->pi_lock);
476 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
477 union futex_key *key, struct futex_pi_state **ps)
479 struct futex_pi_state *pi_state = NULL;
480 struct futex_q *this, *next;
481 struct plist_head *head;
482 struct task_struct *p;
483 pid_t pid = uval & FUTEX_TID_MASK;
487 plist_for_each_entry_safe(this, next, head, list) {
488 if (match_futex(&this->key, key)) {
490 * Another waiter already exists - bump up
491 * the refcount and return its pi_state:
493 pi_state = this->pi_state;
495 * Userspace might have messed up non PI and PI futexes
497 if (unlikely(!pi_state))
500 WARN_ON(!atomic_read(&pi_state->refcount));
501 WARN_ON(pid && pi_state->owner &&
502 pi_state->owner->pid != pid);
504 atomic_inc(&pi_state->refcount);
512 * We are the first waiter - try to look up the real owner and attach
513 * the new pi_state to it, but bail out when TID = 0
517 p = futex_find_get_task(pid);
522 * We need to look at the task state flags to figure out,
523 * whether the task is exiting. To protect against the do_exit
524 * change of the task flags, we do this protected by
527 spin_lock_irq(&p->pi_lock);
528 if (unlikely(p->flags & PF_EXITING)) {
530 * The task is on the way out. When PF_EXITPIDONE is
531 * set, we know that the task has finished the
534 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
536 spin_unlock_irq(&p->pi_lock);
541 pi_state = alloc_pi_state();
544 * Initialize the pi_mutex in locked state and make 'p'
547 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
549 /* Store the key for possible exit cleanups: */
550 pi_state->key = *key;
552 WARN_ON(!list_empty(&pi_state->list));
553 list_add(&pi_state->list, &p->pi_state_list);
555 spin_unlock_irq(&p->pi_lock);
565 * The hash bucket lock must be held when this is called.
566 * Afterwards, the futex_q must not be accessed.
568 static void wake_futex(struct futex_q *q)
570 plist_del(&q->list, &q->list.plist);
572 * The lock in wake_up_all() is a crucial memory barrier after the
573 * plist_del() and also before assigning to q->lock_ptr.
575 wake_up_all(&q->waiters);
577 * The waiting task can free the futex_q as soon as this is written,
578 * without taking any locks. This must come last.
580 * A memory barrier is required here to prevent the following store
581 * to lock_ptr from getting ahead of the wakeup. Clearing the lock
582 * at the end of wake_up_all() does not prevent this store from
589 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
591 struct task_struct *new_owner;
592 struct futex_pi_state *pi_state = this->pi_state;
598 spin_lock(&pi_state->pi_mutex.wait_lock);
599 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
602 * This happens when we have stolen the lock and the original
603 * pending owner did not enqueue itself back on the rt_mutex.
604 * Thats not a tragedy. We know that way, that a lock waiter
605 * is on the fly. We make the futex_q waiter the pending owner.
608 new_owner = this->task;
611 * We pass it to the next owner. (The WAITERS bit is always
612 * kept enabled while there is PI state around. We must also
613 * preserve the owner died bit.)
615 if (!(uval & FUTEX_OWNER_DIED)) {
618 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
620 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
622 if (curval == -EFAULT)
624 else if (curval != uval)
627 spin_unlock(&pi_state->pi_mutex.wait_lock);
632 spin_lock_irq(&pi_state->owner->pi_lock);
633 WARN_ON(list_empty(&pi_state->list));
634 list_del_init(&pi_state->list);
635 spin_unlock_irq(&pi_state->owner->pi_lock);
637 spin_lock_irq(&new_owner->pi_lock);
638 WARN_ON(!list_empty(&pi_state->list));
639 list_add(&pi_state->list, &new_owner->pi_state_list);
640 pi_state->owner = new_owner;
641 spin_unlock_irq(&new_owner->pi_lock);
643 spin_unlock(&pi_state->pi_mutex.wait_lock);
644 rt_mutex_unlock(&pi_state->pi_mutex);
649 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
654 * There is no waiter, so we unlock the futex. The owner died
655 * bit has not to be preserved here. We are the owner:
657 oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
659 if (oldval == -EFAULT)
668 * Express the locking dependencies for lockdep:
671 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
674 spin_lock(&hb1->lock);
676 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
677 } else { /* hb1 > hb2 */
678 spin_lock(&hb2->lock);
679 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
684 * Wake up all waiters hashed on the physical page that is mapped
685 * to this virtual address:
687 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
689 struct futex_hash_bucket *hb;
690 struct futex_q *this, *next;
691 struct plist_head *head;
692 union futex_key key = FUTEX_KEY_INIT;
698 ret = get_futex_key(uaddr, fshared, &key);
699 if (unlikely(ret != 0))
702 hb = hash_futex(&key);
703 spin_lock(&hb->lock);
706 plist_for_each_entry_safe(this, next, head, list) {
707 if (match_futex (&this->key, &key)) {
708 if (this->pi_state) {
713 /* Check if one of the bits is set in both bitsets */
714 if (!(this->bitset & bitset))
718 if (++ret >= nr_wake)
723 spin_unlock(&hb->lock);
725 put_futex_key(fshared, &key);
730 * Wake up all waiters hashed on the physical page that is mapped
731 * to this virtual address:
734 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
735 int nr_wake, int nr_wake2, int op)
737 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
738 struct futex_hash_bucket *hb1, *hb2;
739 struct plist_head *head;
740 struct futex_q *this, *next;
741 int ret, op_ret, attempt = 0;
744 ret = get_futex_key(uaddr1, fshared, &key1);
745 if (unlikely(ret != 0))
747 ret = get_futex_key(uaddr2, fshared, &key2);
748 if (unlikely(ret != 0))
751 hb1 = hash_futex(&key1);
752 hb2 = hash_futex(&key2);
755 double_lock_hb(hb1, hb2);
757 op_ret = futex_atomic_op_inuser(op, uaddr2);
758 if (unlikely(op_ret < 0)) {
761 spin_unlock(&hb1->lock);
763 spin_unlock(&hb2->lock);
767 * we don't get EFAULT from MMU faults if we don't have an MMU,
768 * but we might get them from range checking
774 if (unlikely(op_ret != -EFAULT)) {
780 * futex_atomic_op_inuser needs to both read and write
781 * *(int __user *)uaddr2, but we can't modify it
782 * non-atomically. Therefore, if get_user below is not
783 * enough, we need to handle the fault ourselves, while
784 * still holding the mmap_sem.
787 ret = futex_handle_fault((unsigned long)uaddr2,
794 ret = get_user(dummy, uaddr2);
803 plist_for_each_entry_safe(this, next, head, list) {
804 if (match_futex (&this->key, &key1)) {
806 if (++ret >= nr_wake)
815 plist_for_each_entry_safe(this, next, head, list) {
816 if (match_futex (&this->key, &key2)) {
818 if (++op_ret >= nr_wake2)
825 spin_unlock(&hb1->lock);
827 spin_unlock(&hb2->lock);
829 put_futex_key(fshared, &key2);
830 put_futex_key(fshared, &key1);
836 * Requeue all waiters hashed on one physical page to another
839 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
840 int nr_wake, int nr_requeue, u32 *cmpval)
842 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
843 struct futex_hash_bucket *hb1, *hb2;
844 struct plist_head *head1;
845 struct futex_q *this, *next;
846 int ret, drop_count = 0;
849 ret = get_futex_key(uaddr1, fshared, &key1);
850 if (unlikely(ret != 0))
852 ret = get_futex_key(uaddr2, fshared, &key2);
853 if (unlikely(ret != 0))
856 hb1 = hash_futex(&key1);
857 hb2 = hash_futex(&key2);
859 double_lock_hb(hb1, hb2);
861 if (likely(cmpval != NULL)) {
864 ret = get_futex_value_locked(&curval, uaddr1);
867 spin_unlock(&hb1->lock);
869 spin_unlock(&hb2->lock);
871 ret = get_user(curval, uaddr1);
878 if (curval != *cmpval) {
885 plist_for_each_entry_safe(this, next, head1, list) {
886 if (!match_futex (&this->key, &key1))
888 if (++ret <= nr_wake) {
892 * If key1 and key2 hash to the same bucket, no need to
895 if (likely(head1 != &hb2->chain)) {
896 plist_del(&this->list, &hb1->chain);
897 plist_add(&this->list, &hb2->chain);
898 this->lock_ptr = &hb2->lock;
899 #ifdef CONFIG_DEBUG_PI_LIST
900 this->list.plist.lock = &hb2->lock;
904 get_futex_key_refs(&key2);
907 if (ret - nr_wake >= nr_requeue)
913 spin_unlock(&hb1->lock);
915 spin_unlock(&hb2->lock);
917 /* drop_futex_key_refs() must be called outside the spinlocks. */
918 while (--drop_count >= 0)
919 drop_futex_key_refs(&key1);
922 put_futex_key(fshared, &key2);
923 put_futex_key(fshared, &key1);
927 /* The key must be already stored in q->key. */
928 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
930 struct futex_hash_bucket *hb;
932 init_waitqueue_head(&q->waiters);
934 get_futex_key_refs(&q->key);
935 hb = hash_futex(&q->key);
936 q->lock_ptr = &hb->lock;
938 spin_lock(&hb->lock);
942 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
947 * The priority used to register this element is
948 * - either the real thread-priority for the real-time threads
949 * (i.e. threads with a priority lower than MAX_RT_PRIO)
950 * - or MAX_RT_PRIO for non-RT threads.
951 * Thus, all RT-threads are woken first in priority order, and
952 * the others are woken last, in FIFO order.
954 prio = min(current->normal_prio, MAX_RT_PRIO);
956 plist_node_init(&q->list, prio);
957 #ifdef CONFIG_DEBUG_PI_LIST
958 q->list.plist.lock = &hb->lock;
960 plist_add(&q->list, &hb->chain);
962 spin_unlock(&hb->lock);
966 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
968 spin_unlock(&hb->lock);
969 drop_futex_key_refs(&q->key);
973 * queue_me and unqueue_me must be called as a pair, each
974 * exactly once. They are called with the hashed spinlock held.
977 /* Return 1 if we were still queued (ie. 0 means we were woken) */
978 static int unqueue_me(struct futex_q *q)
980 spinlock_t *lock_ptr;
983 /* In the common case we don't take the spinlock, which is nice. */
985 lock_ptr = q->lock_ptr;
987 if (lock_ptr != NULL) {
990 * q->lock_ptr can change between reading it and
991 * spin_lock(), causing us to take the wrong lock. This
992 * corrects the race condition.
994 * Reasoning goes like this: if we have the wrong lock,
995 * q->lock_ptr must have changed (maybe several times)
996 * between reading it and the spin_lock(). It can
997 * change again after the spin_lock() but only if it was
998 * already changed before the spin_lock(). It cannot,
999 * however, change back to the original value. Therefore
1000 * we can detect whether we acquired the correct lock.
1002 if (unlikely(lock_ptr != q->lock_ptr)) {
1003 spin_unlock(lock_ptr);
1006 WARN_ON(plist_node_empty(&q->list));
1007 plist_del(&q->list, &q->list.plist);
1009 BUG_ON(q->pi_state);
1011 spin_unlock(lock_ptr);
1015 drop_futex_key_refs(&q->key);
1020 * PI futexes can not be requeued and must remove themself from the
1021 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1024 static void unqueue_me_pi(struct futex_q *q)
1026 WARN_ON(plist_node_empty(&q->list));
1027 plist_del(&q->list, &q->list.plist);
1029 BUG_ON(!q->pi_state);
1030 free_pi_state(q->pi_state);
1033 spin_unlock(q->lock_ptr);
1035 drop_futex_key_refs(&q->key);
1039 * Fixup the pi_state owner with the new owner.
1041 * Must be called with hash bucket lock held and mm->sem held for non
1044 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1045 struct task_struct *newowner, int fshared)
1047 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1048 struct futex_pi_state *pi_state = q->pi_state;
1049 struct task_struct *oldowner = pi_state->owner;
1050 u32 uval, curval, newval;
1051 int ret, attempt = 0;
1054 if (!pi_state->owner)
1055 newtid |= FUTEX_OWNER_DIED;
1058 * We are here either because we stole the rtmutex from the
1059 * pending owner or we are the pending owner which failed to
1060 * get the rtmutex. We have to replace the pending owner TID
1061 * in the user space variable. This must be atomic as we have
1062 * to preserve the owner died bit here.
1064 * Note: We write the user space value _before_ changing the
1065 * pi_state because we can fault here. Imagine swapped out
1066 * pages or a fork, which was running right before we acquired
1067 * mmap_sem, that marked all the anonymous memory readonly for
1070 * Modifying pi_state _before_ the user space value would
1071 * leave the pi_state in an inconsistent state when we fault
1072 * here, because we need to drop the hash bucket lock to
1073 * handle the fault. This might be observed in the PID check
1074 * in lookup_pi_state.
1077 if (get_futex_value_locked(&uval, uaddr))
1081 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1083 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1085 if (curval == -EFAULT)
1093 * We fixed up user space. Now we need to fix the pi_state
1096 if (pi_state->owner != NULL) {
1097 spin_lock_irq(&pi_state->owner->pi_lock);
1098 WARN_ON(list_empty(&pi_state->list));
1099 list_del_init(&pi_state->list);
1100 spin_unlock_irq(&pi_state->owner->pi_lock);
1103 pi_state->owner = newowner;
1105 spin_lock_irq(&newowner->pi_lock);
1106 WARN_ON(!list_empty(&pi_state->list));
1107 list_add(&pi_state->list, &newowner->pi_state_list);
1108 spin_unlock_irq(&newowner->pi_lock);
1112 * To handle the page fault we need to drop the hash bucket
1113 * lock here. That gives the other task (either the pending
1114 * owner itself or the task which stole the rtmutex) the
1115 * chance to try the fixup of the pi_state. So once we are
1116 * back from handling the fault we need to check the pi_state
1117 * after reacquiring the hash bucket lock and before trying to
1118 * do another fixup. When the fixup has been done already we
1122 spin_unlock(q->lock_ptr);
1124 ret = futex_handle_fault((unsigned long)uaddr, attempt++);
1126 spin_lock(q->lock_ptr);
1129 * Check if someone else fixed it for us:
1131 if (pi_state->owner != oldowner)
1141 * In case we must use restart_block to restart a futex_wait,
1142 * we encode in the 'flags' shared capability
1144 #define FLAGS_SHARED 1
1146 static long futex_wait_restart(struct restart_block *restart);
1148 static int futex_wait(u32 __user *uaddr, int fshared,
1149 u32 val, ktime_t *abs_time, u32 bitset)
1151 struct task_struct *curr = current;
1152 DECLARE_WAITQUEUE(wait, curr);
1153 struct futex_hash_bucket *hb;
1157 struct hrtimer_sleeper t;
1166 q.key = FUTEX_KEY_INIT;
1167 ret = get_futex_key(uaddr, fshared, &q.key);
1168 if (unlikely(ret != 0))
1169 goto out_release_sem;
1171 hb = queue_lock(&q);
1174 * Access the page AFTER the futex is queued.
1175 * Order is important:
1177 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1178 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1180 * The basic logical guarantee of a futex is that it blocks ONLY
1181 * if cond(var) is known to be true at the time of blocking, for
1182 * any cond. If we queued after testing *uaddr, that would open
1183 * a race condition where we could block indefinitely with
1184 * cond(var) false, which would violate the guarantee.
1186 * A consequence is that futex_wait() can return zero and absorb
1187 * a wakeup when *uaddr != val on entry to the syscall. This is
1190 * for shared futexes, we hold the mmap semaphore, so the mapping
1191 * cannot have changed since we looked it up in get_futex_key.
1193 ret = get_futex_value_locked(&uval, uaddr);
1195 if (unlikely(ret)) {
1196 queue_unlock(&q, hb);
1198 ret = get_user(uval, uaddr);
1206 goto out_unlock_release_sem;
1208 /* Only actually queue if *uaddr contained val. */
1212 * There might have been scheduling since the queue_me(), as we
1213 * cannot hold a spinlock across the get_user() in case it
1214 * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1215 * queueing ourselves into the futex hash. This code thus has to
1216 * rely on the futex_wake() code removing us from hash when it
1220 /* add_wait_queue is the barrier after __set_current_state. */
1221 __set_current_state(TASK_INTERRUPTIBLE);
1222 add_wait_queue(&q.waiters, &wait);
1224 * !plist_node_empty() is safe here without any lock.
1225 * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1227 if (likely(!plist_node_empty(&q.list))) {
1231 hrtimer_init_on_stack(&t.timer, CLOCK_MONOTONIC,
1233 hrtimer_init_sleeper(&t, current);
1234 t.timer.expires = *abs_time;
1236 hrtimer_start(&t.timer, t.timer.expires,
1238 if (!hrtimer_active(&t.timer))
1242 * the timer could have already expired, in which
1243 * case current would be flagged for rescheduling.
1244 * Don't bother calling schedule.
1249 hrtimer_cancel(&t.timer);
1251 /* Flag if a timeout occured */
1252 rem = (t.task == NULL);
1254 destroy_hrtimer_on_stack(&t.timer);
1257 __set_current_state(TASK_RUNNING);
1260 * NOTE: we don't remove ourselves from the waitqueue because
1261 * we are the only user of it.
1264 /* If we were woken (and unqueued), we succeeded, whatever. */
1265 if (!unqueue_me(&q))
1271 * We expect signal_pending(current), but another thread may
1272 * have handled it for us already.
1275 return -ERESTARTSYS;
1277 struct restart_block *restart;
1278 restart = ¤t_thread_info()->restart_block;
1279 restart->fn = futex_wait_restart;
1280 restart->futex.uaddr = (u32 *)uaddr;
1281 restart->futex.val = val;
1282 restart->futex.time = abs_time->tv64;
1283 restart->futex.bitset = bitset;
1284 restart->futex.flags = 0;
1287 restart->futex.flags |= FLAGS_SHARED;
1288 return -ERESTART_RESTARTBLOCK;
1291 out_unlock_release_sem:
1292 queue_unlock(&q, hb);
1295 put_futex_key(fshared, &q.key);
1300 static long futex_wait_restart(struct restart_block *restart)
1302 u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1306 t.tv64 = restart->futex.time;
1307 restart->fn = do_no_restart_syscall;
1308 if (restart->futex.flags & FLAGS_SHARED)
1310 return (long)futex_wait(uaddr, fshared, restart->futex.val, &t,
1311 restart->futex.bitset);
1316 * Userspace tried a 0 -> TID atomic transition of the futex value
1317 * and failed. The kernel side here does the whole locking operation:
1318 * if there are waiters then it will block, it does PI, etc. (Due to
1319 * races the kernel might see a 0 value of the futex too.)
1321 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1322 int detect, ktime_t *time, int trylock)
1324 struct hrtimer_sleeper timeout, *to = NULL;
1325 struct task_struct *curr = current;
1326 struct futex_hash_bucket *hb;
1327 u32 uval, newval, curval;
1329 int ret, lock_taken, ownerdied = 0, attempt = 0;
1331 if (refill_pi_state_cache())
1336 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1338 hrtimer_init_sleeper(to, current);
1339 to->timer.expires = *time;
1344 q.key = FUTEX_KEY_INIT;
1345 ret = get_futex_key(uaddr, fshared, &q.key);
1346 if (unlikely(ret != 0))
1347 goto out_release_sem;
1350 hb = queue_lock(&q);
1353 ret = lock_taken = 0;
1356 * To avoid races, we attempt to take the lock here again
1357 * (by doing a 0 -> TID atomic cmpxchg), while holding all
1358 * the locks. It will most likely not succeed.
1360 newval = task_pid_vnr(current);
1362 curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
1364 if (unlikely(curval == -EFAULT))
1368 * Detect deadlocks. In case of REQUEUE_PI this is a valid
1369 * situation and we return success to user space.
1371 if (unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(current))) {
1373 goto out_unlock_release_sem;
1377 * Surprise - we got the lock. Just return to userspace:
1379 if (unlikely(!curval))
1380 goto out_unlock_release_sem;
1385 * Set the WAITERS flag, so the owner will know it has someone
1386 * to wake at next unlock
1388 newval = curval | FUTEX_WAITERS;
1391 * There are two cases, where a futex might have no owner (the
1392 * owner TID is 0): OWNER_DIED. We take over the futex in this
1393 * case. We also do an unconditional take over, when the owner
1394 * of the futex died.
1396 * This is safe as we are protected by the hash bucket lock !
1398 if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
1399 /* Keep the OWNER_DIED bit */
1400 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(current);
1405 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1407 if (unlikely(curval == -EFAULT))
1409 if (unlikely(curval != uval))
1413 * We took the lock due to owner died take over.
1415 if (unlikely(lock_taken))
1416 goto out_unlock_release_sem;
1419 * We dont have the lock. Look up the PI state (or create it if
1420 * we are the first waiter):
1422 ret = lookup_pi_state(uval, hb, &q.key, &q.pi_state);
1424 if (unlikely(ret)) {
1429 * Task is exiting and we just wait for the
1432 queue_unlock(&q, hb);
1438 * No owner found for this futex. Check if the
1439 * OWNER_DIED bit is set to figure out whether
1440 * this is a robust futex or not.
1442 if (get_futex_value_locked(&curval, uaddr))
1446 * We simply start over in case of a robust
1447 * futex. The code above will take the futex
1450 if (curval & FUTEX_OWNER_DIED) {
1455 goto out_unlock_release_sem;
1460 * Only actually queue now that the atomic ops are done:
1464 WARN_ON(!q.pi_state);
1466 * Block on the PI mutex:
1469 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1471 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1472 /* Fixup the trylock return value: */
1473 ret = ret ? 0 : -EWOULDBLOCK;
1476 spin_lock(q.lock_ptr);
1480 * Got the lock. We might not be the anticipated owner
1481 * if we did a lock-steal - fix up the PI-state in
1484 if (q.pi_state->owner != curr)
1485 ret = fixup_pi_state_owner(uaddr, &q, curr, fshared);
1488 * Catch the rare case, where the lock was released
1489 * when we were on the way back before we locked the
1492 if (q.pi_state->owner == curr) {
1494 * Try to get the rt_mutex now. This might
1495 * fail as some other task acquired the
1496 * rt_mutex after we removed ourself from the
1497 * rt_mutex waiters list.
1499 if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1503 * pi_state is incorrect, some other
1504 * task did a lock steal and we
1505 * returned due to timeout or signal
1506 * without taking the rt_mutex. Too
1507 * late. We can access the
1508 * rt_mutex_owner without locking, as
1509 * the other task is now blocked on
1510 * the hash bucket lock. Fix the state
1513 struct task_struct *owner;
1516 owner = rt_mutex_owner(&q.pi_state->pi_mutex);
1517 res = fixup_pi_state_owner(uaddr, &q, owner,
1520 /* propagate -EFAULT, if the fixup failed */
1526 * Paranoia check. If we did not take the lock
1527 * in the trylock above, then we should not be
1528 * the owner of the rtmutex, neither the real
1529 * nor the pending one:
1531 if (rt_mutex_owner(&q.pi_state->pi_mutex) == curr)
1532 printk(KERN_ERR "futex_lock_pi: ret = %d "
1533 "pi-mutex: %p pi-state %p\n", ret,
1534 q.pi_state->pi_mutex.owner,
1539 /* Unqueue and drop the lock */
1543 destroy_hrtimer_on_stack(&to->timer);
1544 return ret != -EINTR ? ret : -ERESTARTNOINTR;
1546 out_unlock_release_sem:
1547 queue_unlock(&q, hb);
1550 put_futex_key(fshared, &q.key);
1552 destroy_hrtimer_on_stack(&to->timer);
1557 * We have to r/w *(int __user *)uaddr, but we can't modify it
1558 * non-atomically. Therefore, if get_user below is not
1559 * enough, we need to handle the fault ourselves, while
1560 * still holding the mmap_sem.
1562 * ... and hb->lock. :-) --ANK
1564 queue_unlock(&q, hb);
1567 ret = futex_handle_fault((unsigned long)uaddr, attempt);
1569 goto out_release_sem;
1570 goto retry_unlocked;
1573 ret = get_user(uval, uaddr);
1574 if (!ret && (uval != -EFAULT))
1578 destroy_hrtimer_on_stack(&to->timer);
1583 * Userspace attempted a TID -> 0 atomic transition, and failed.
1584 * This is the in-kernel slowpath: we look up the PI state (if any),
1585 * and do the rt-mutex unlock.
1587 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
1589 struct futex_hash_bucket *hb;
1590 struct futex_q *this, *next;
1592 struct plist_head *head;
1593 union futex_key key = FUTEX_KEY_INIT;
1594 int ret, attempt = 0;
1597 if (get_user(uval, uaddr))
1600 * We release only a lock we actually own:
1602 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
1605 ret = get_futex_key(uaddr, fshared, &key);
1606 if (unlikely(ret != 0))
1609 hb = hash_futex(&key);
1611 spin_lock(&hb->lock);
1614 * To avoid races, try to do the TID -> 0 atomic transition
1615 * again. If it succeeds then we can return without waking
1618 if (!(uval & FUTEX_OWNER_DIED))
1619 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
1622 if (unlikely(uval == -EFAULT))
1625 * Rare case: we managed to release the lock atomically,
1626 * no need to wake anyone else up:
1628 if (unlikely(uval == task_pid_vnr(current)))
1632 * Ok, other tasks may need to be woken up - check waiters
1633 * and do the wakeup if necessary:
1637 plist_for_each_entry_safe(this, next, head, list) {
1638 if (!match_futex (&this->key, &key))
1640 ret = wake_futex_pi(uaddr, uval, this);
1642 * The atomic access to the futex value
1643 * generated a pagefault, so retry the
1644 * user-access and the wakeup:
1651 * No waiters - kernel unlocks the futex:
1653 if (!(uval & FUTEX_OWNER_DIED)) {
1654 ret = unlock_futex_pi(uaddr, uval);
1660 spin_unlock(&hb->lock);
1662 put_futex_key(fshared, &key);
1668 * We have to r/w *(int __user *)uaddr, but we can't modify it
1669 * non-atomically. Therefore, if get_user below is not
1670 * enough, we need to handle the fault ourselves, while
1671 * still holding the mmap_sem.
1673 * ... and hb->lock. --ANK
1675 spin_unlock(&hb->lock);
1678 ret = futex_handle_fault((unsigned long)uaddr, attempt);
1682 goto retry_unlocked;
1685 ret = get_user(uval, uaddr);
1686 if (!ret && (uval != -EFAULT))
1693 * Support for robust futexes: the kernel cleans up held futexes at
1696 * Implementation: user-space maintains a per-thread list of locks it
1697 * is holding. Upon do_exit(), the kernel carefully walks this list,
1698 * and marks all locks that are owned by this thread with the
1699 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1700 * always manipulated with the lock held, so the list is private and
1701 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1702 * field, to allow the kernel to clean up if the thread dies after
1703 * acquiring the lock, but just before it could have added itself to
1704 * the list. There can only be one such pending lock.
1708 * sys_set_robust_list - set the robust-futex list head of a task
1709 * @head: pointer to the list-head
1710 * @len: length of the list-head, as userspace expects
1713 sys_set_robust_list(struct robust_list_head __user *head,
1716 if (!futex_cmpxchg_enabled)
1719 * The kernel knows only one size for now:
1721 if (unlikely(len != sizeof(*head)))
1724 current->robust_list = head;
1730 * sys_get_robust_list - get the robust-futex list head of a task
1731 * @pid: pid of the process [zero for current task]
1732 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1733 * @len_ptr: pointer to a length field, the kernel fills in the header size
1736 sys_get_robust_list(int pid, struct robust_list_head __user * __user *head_ptr,
1737 size_t __user *len_ptr)
1739 struct robust_list_head __user *head;
1742 if (!futex_cmpxchg_enabled)
1746 head = current->robust_list;
1748 struct task_struct *p;
1752 p = find_task_by_vpid(pid);
1756 if ((current->euid != p->euid) && (current->euid != p->uid) &&
1757 !capable(CAP_SYS_PTRACE))
1759 head = p->robust_list;
1763 if (put_user(sizeof(*head), len_ptr))
1765 return put_user(head, head_ptr);
1774 * Process a futex-list entry, check whether it's owned by the
1775 * dying task, and do notification if so:
1777 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
1779 u32 uval, nval, mval;
1782 if (get_user(uval, uaddr))
1785 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
1787 * Ok, this dying thread is truly holding a futex
1788 * of interest. Set the OWNER_DIED bit atomically
1789 * via cmpxchg, and if the value had FUTEX_WAITERS
1790 * set, wake up a waiter (if any). (We have to do a
1791 * futex_wake() even if OWNER_DIED is already set -
1792 * to handle the rare but possible case of recursive
1793 * thread-death.) The rest of the cleanup is done in
1796 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1797 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
1799 if (nval == -EFAULT)
1806 * Wake robust non-PI futexes here. The wakeup of
1807 * PI futexes happens in exit_pi_state():
1809 if (!pi && (uval & FUTEX_WAITERS))
1810 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
1816 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1818 static inline int fetch_robust_entry(struct robust_list __user **entry,
1819 struct robust_list __user * __user *head,
1822 unsigned long uentry;
1824 if (get_user(uentry, (unsigned long __user *)head))
1827 *entry = (void __user *)(uentry & ~1UL);
1834 * Walk curr->robust_list (very carefully, it's a userspace list!)
1835 * and mark any locks found there dead, and notify any waiters.
1837 * We silently return on any sign of list-walking problem.
1839 void exit_robust_list(struct task_struct *curr)
1841 struct robust_list_head __user *head = curr->robust_list;
1842 struct robust_list __user *entry, *next_entry, *pending;
1843 unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
1844 unsigned long futex_offset;
1847 if (!futex_cmpxchg_enabled)
1851 * Fetch the list head (which was registered earlier, via
1852 * sys_set_robust_list()):
1854 if (fetch_robust_entry(&entry, &head->list.next, &pi))
1857 * Fetch the relative futex offset:
1859 if (get_user(futex_offset, &head->futex_offset))
1862 * Fetch any possibly pending lock-add first, and handle it
1865 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
1868 next_entry = NULL; /* avoid warning with gcc */
1869 while (entry != &head->list) {
1871 * Fetch the next entry in the list before calling
1872 * handle_futex_death:
1874 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
1876 * A pending lock might already be on the list, so
1877 * don't process it twice:
1879 if (entry != pending)
1880 if (handle_futex_death((void __user *)entry + futex_offset,
1888 * Avoid excessively long or circular lists:
1897 handle_futex_death((void __user *)pending + futex_offset,
1901 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
1902 u32 __user *uaddr2, u32 val2, u32 val3)
1905 int cmd = op & FUTEX_CMD_MASK;
1908 if (!(op & FUTEX_PRIVATE_FLAG))
1913 val3 = FUTEX_BITSET_MATCH_ANY;
1914 case FUTEX_WAIT_BITSET:
1915 ret = futex_wait(uaddr, fshared, val, timeout, val3);
1918 val3 = FUTEX_BITSET_MATCH_ANY;
1919 case FUTEX_WAKE_BITSET:
1920 ret = futex_wake(uaddr, fshared, val, val3);
1923 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL);
1925 case FUTEX_CMP_REQUEUE:
1926 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3);
1929 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
1932 if (futex_cmpxchg_enabled)
1933 ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
1935 case FUTEX_UNLOCK_PI:
1936 if (futex_cmpxchg_enabled)
1937 ret = futex_unlock_pi(uaddr, fshared);
1939 case FUTEX_TRYLOCK_PI:
1940 if (futex_cmpxchg_enabled)
1941 ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
1950 asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
1951 struct timespec __user *utime, u32 __user *uaddr2,
1955 ktime_t t, *tp = NULL;
1957 int cmd = op & FUTEX_CMD_MASK;
1959 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
1960 cmd == FUTEX_WAIT_BITSET)) {
1961 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
1963 if (!timespec_valid(&ts))
1966 t = timespec_to_ktime(ts);
1967 if (cmd == FUTEX_WAIT)
1968 t = ktime_add_safe(ktime_get(), t);
1972 * requeue parameter in 'utime' if cmd == FUTEX_REQUEUE.
1973 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
1975 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
1976 cmd == FUTEX_WAKE_OP)
1977 val2 = (u32) (unsigned long) utime;
1979 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
1982 static int __init futex_init(void)
1988 * This will fail and we want it. Some arch implementations do
1989 * runtime detection of the futex_atomic_cmpxchg_inatomic()
1990 * functionality. We want to know that before we call in any
1991 * of the complex code paths. Also we want to prevent
1992 * registration of robust lists in that case. NULL is
1993 * guaranteed to fault and we get -EFAULT on functional
1994 * implementation, the non functional ones will return
1997 curval = cmpxchg_futex_value_locked(NULL, 0, 0);
1998 if (curval == -EFAULT)
1999 futex_cmpxchg_enabled = 1;
2001 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2002 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2003 spin_lock_init(&futex_queues[i].lock);
2008 __initcall(futex_init);