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 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23 * Copyright (C) IBM Corporation, 2009
24 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
26 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27 * enough at me, Linus for the original (flawed) idea, Matthew
28 * Kirkwood for proof-of-concept implementation.
30 * "The futexes are also cursed."
31 * "But they come in a choice of three flavours!"
33 * This program is free software; you can redistribute it and/or modify
34 * it under the terms of the GNU General Public License as published by
35 * the Free Software Foundation; either version 2 of the License, or
36 * (at your option) any later version.
38 * This program is distributed in the hope that it will be useful,
39 * but WITHOUT ANY WARRANTY; without even the implied warranty of
40 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
41 * GNU General Public License for more details.
43 * You should have received a copy of the GNU General Public License
44 * along with this program; if not, write to the Free Software
45 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
47 #include <linux/slab.h>
48 #include <linux/poll.h>
50 #include <linux/file.h>
51 #include <linux/jhash.h>
52 #include <linux/init.h>
53 #include <linux/futex.h>
54 #include <linux/mount.h>
55 #include <linux/pagemap.h>
56 #include <linux/syscalls.h>
57 #include <linux/signal.h>
58 #include <linux/export.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62 #include <linux/ptrace.h>
63 #include <linux/sched/rt.h>
64 #include <linux/hugetlb.h>
65 #include <linux/freezer.h>
67 #include <asm/futex.h>
69 #include "locking/rtmutex_common.h"
71 int __read_mostly futex_cmpxchg_enabled;
73 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
76 * Futex flags used to encode options to functions and preserve them across
79 #define FLAGS_SHARED 0x01
80 #define FLAGS_CLOCKRT 0x02
81 #define FLAGS_HAS_TIMEOUT 0x04
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 * struct futex_q - The hashed futex queue entry, one per waiting task
106 * @list: priority-sorted list of tasks waiting on this futex
107 * @task: the task waiting on the futex
108 * @lock_ptr: the hash bucket lock
109 * @key: the key the futex is hashed on
110 * @pi_state: optional priority inheritance state
111 * @rt_waiter: rt_waiter storage for use with requeue_pi
112 * @requeue_pi_key: the requeue_pi target futex key
113 * @bitset: bitset for the optional bitmasked wakeup
115 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
116 * we can wake only the relevant ones (hashed queues may be shared).
118 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
119 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
120 * The order of wakeup is always to make the first condition true, then
123 * PI futexes are typically woken before they are removed from the hash list via
124 * the rt_mutex code. See unqueue_me_pi().
127 struct plist_node list;
129 struct task_struct *task;
130 spinlock_t *lock_ptr;
132 struct futex_pi_state *pi_state;
133 struct rt_mutex_waiter *rt_waiter;
134 union futex_key *requeue_pi_key;
138 static const struct futex_q futex_q_init = {
139 /* list gets initialized in queue_me()*/
140 .key = FUTEX_KEY_INIT,
141 .bitset = FUTEX_BITSET_MATCH_ANY
145 * Hash buckets are shared by all the futex_keys that hash to the same
146 * location. Each key may have multiple futex_q structures, one for each task
147 * waiting on a futex.
149 struct futex_hash_bucket {
151 struct plist_head chain;
154 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
157 * We hash on the keys returned from get_futex_key (see below).
159 static struct futex_hash_bucket *hash_futex(union futex_key *key)
161 u32 hash = jhash2((u32*)&key->both.word,
162 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
164 return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
168 * Return 1 if two futex_keys are equal, 0 otherwise.
170 static inline int match_futex(union futex_key *key1, union futex_key *key2)
173 && key1->both.word == key2->both.word
174 && key1->both.ptr == key2->both.ptr
175 && key1->both.offset == key2->both.offset);
179 * Take a reference to the resource addressed by a key.
180 * Can be called while holding spinlocks.
183 static void get_futex_key_refs(union futex_key *key)
188 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
190 ihold(key->shared.inode);
192 case FUT_OFF_MMSHARED:
193 atomic_inc(&key->private.mm->mm_count);
199 * Drop a reference to the resource addressed by a key.
200 * The hash bucket spinlock must not be held.
202 static void drop_futex_key_refs(union futex_key *key)
204 if (!key->both.ptr) {
205 /* If we're here then we tried to put a key we failed to get */
210 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
212 iput(key->shared.inode);
214 case FUT_OFF_MMSHARED:
215 mmdrop(key->private.mm);
221 * get_futex_key() - Get parameters which are the keys for a futex
222 * @uaddr: virtual address of the futex
223 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
224 * @key: address where result is stored.
225 * @rw: mapping needs to be read/write (values: VERIFY_READ,
228 * Return: a negative error code or 0
230 * The key words are stored in *key on success.
232 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
233 * offset_within_page). For private mappings, it's (uaddr, current->mm).
234 * We can usually work out the index without swapping in the page.
236 * lock_page() might sleep, the caller should not hold a spinlock.
239 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
241 unsigned long address = (unsigned long)uaddr;
242 struct mm_struct *mm = current->mm;
243 struct page *page, *page_head;
247 * The futex address must be "naturally" aligned.
249 key->both.offset = address % PAGE_SIZE;
250 if (unlikely((address % sizeof(u32)) != 0))
252 address -= key->both.offset;
254 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
258 * PROCESS_PRIVATE futexes are fast.
259 * As the mm cannot disappear under us and the 'key' only needs
260 * virtual address, we dont even have to find the underlying vma.
261 * Note : We do have to check 'uaddr' is a valid user address,
262 * but access_ok() should be faster than find_vma()
265 key->private.mm = mm;
266 key->private.address = address;
267 get_futex_key_refs(key);
272 err = get_user_pages_fast(address, 1, 1, &page);
274 * If write access is not required (eg. FUTEX_WAIT), try
275 * and get read-only access.
277 if (err == -EFAULT && rw == VERIFY_READ) {
278 err = get_user_pages_fast(address, 1, 0, &page);
286 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
288 if (unlikely(PageTail(page))) {
290 /* serialize against __split_huge_page_splitting() */
292 if (likely(__get_user_pages_fast(address, 1, !ro, &page) == 1)) {
293 page_head = compound_head(page);
295 * page_head is valid pointer but we must pin
296 * it before taking the PG_lock and/or
297 * PG_compound_lock. The moment we re-enable
298 * irqs __split_huge_page_splitting() can
299 * return and the head page can be freed from
300 * under us. We can't take the PG_lock and/or
301 * PG_compound_lock on a page that could be
302 * freed from under us.
304 if (page != page_head) {
315 page_head = compound_head(page);
316 if (page != page_head) {
322 lock_page(page_head);
325 * If page_head->mapping is NULL, then it cannot be a PageAnon
326 * page; but it might be the ZERO_PAGE or in the gate area or
327 * in a special mapping (all cases which we are happy to fail);
328 * or it may have been a good file page when get_user_pages_fast
329 * found it, but truncated or holepunched or subjected to
330 * invalidate_complete_page2 before we got the page lock (also
331 * cases which we are happy to fail). And we hold a reference,
332 * so refcount care in invalidate_complete_page's remove_mapping
333 * prevents drop_caches from setting mapping to NULL beneath us.
335 * The case we do have to guard against is when memory pressure made
336 * shmem_writepage move it from filecache to swapcache beneath us:
337 * an unlikely race, but we do need to retry for page_head->mapping.
339 if (!page_head->mapping) {
340 int shmem_swizzled = PageSwapCache(page_head);
341 unlock_page(page_head);
349 * Private mappings are handled in a simple way.
351 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
352 * it's a read-only handle, it's expected that futexes attach to
353 * the object not the particular process.
355 if (PageAnon(page_head)) {
357 * A RO anonymous page will never change and thus doesn't make
358 * sense for futex operations.
365 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
366 key->private.mm = mm;
367 key->private.address = address;
369 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
370 key->shared.inode = page_head->mapping->host;
371 key->shared.pgoff = basepage_index(page);
374 get_futex_key_refs(key);
377 unlock_page(page_head);
382 static inline void put_futex_key(union futex_key *key)
384 drop_futex_key_refs(key);
388 * fault_in_user_writeable() - Fault in user address and verify RW access
389 * @uaddr: pointer to faulting user space address
391 * Slow path to fixup the fault we just took in the atomic write
394 * We have no generic implementation of a non-destructive write to the
395 * user address. We know that we faulted in the atomic pagefault
396 * disabled section so we can as well avoid the #PF overhead by
397 * calling get_user_pages() right away.
399 static int fault_in_user_writeable(u32 __user *uaddr)
401 struct mm_struct *mm = current->mm;
404 down_read(&mm->mmap_sem);
405 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
407 up_read(&mm->mmap_sem);
409 return ret < 0 ? ret : 0;
413 * futex_top_waiter() - Return the highest priority waiter on a futex
414 * @hb: the hash bucket the futex_q's reside in
415 * @key: the futex key (to distinguish it from other futex futex_q's)
417 * Must be called with the hb lock held.
419 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
420 union futex_key *key)
422 struct futex_q *this;
424 plist_for_each_entry(this, &hb->chain, list) {
425 if (match_futex(&this->key, key))
431 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
432 u32 uval, u32 newval)
437 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
443 static int get_futex_value_locked(u32 *dest, u32 __user *from)
448 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
451 return ret ? -EFAULT : 0;
458 static int refill_pi_state_cache(void)
460 struct futex_pi_state *pi_state;
462 if (likely(current->pi_state_cache))
465 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
470 INIT_LIST_HEAD(&pi_state->list);
471 /* pi_mutex gets initialized later */
472 pi_state->owner = NULL;
473 atomic_set(&pi_state->refcount, 1);
474 pi_state->key = FUTEX_KEY_INIT;
476 current->pi_state_cache = pi_state;
481 static struct futex_pi_state * alloc_pi_state(void)
483 struct futex_pi_state *pi_state = current->pi_state_cache;
486 current->pi_state_cache = NULL;
491 static void free_pi_state(struct futex_pi_state *pi_state)
493 if (!atomic_dec_and_test(&pi_state->refcount))
497 * If pi_state->owner is NULL, the owner is most probably dying
498 * and has cleaned up the pi_state already
500 if (pi_state->owner) {
501 raw_spin_lock_irq(&pi_state->owner->pi_lock);
502 list_del_init(&pi_state->list);
503 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
505 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
508 if (current->pi_state_cache)
512 * pi_state->list is already empty.
513 * clear pi_state->owner.
514 * refcount is at 0 - put it back to 1.
516 pi_state->owner = NULL;
517 atomic_set(&pi_state->refcount, 1);
518 current->pi_state_cache = pi_state;
523 * Look up the task based on what TID userspace gave us.
526 static struct task_struct * futex_find_get_task(pid_t pid)
528 struct task_struct *p;
531 p = find_task_by_vpid(pid);
541 * This task is holding PI mutexes at exit time => bad.
542 * Kernel cleans up PI-state, but userspace is likely hosed.
543 * (Robust-futex cleanup is separate and might save the day for userspace.)
545 void exit_pi_state_list(struct task_struct *curr)
547 struct list_head *next, *head = &curr->pi_state_list;
548 struct futex_pi_state *pi_state;
549 struct futex_hash_bucket *hb;
550 union futex_key key = FUTEX_KEY_INIT;
552 if (!futex_cmpxchg_enabled)
555 * We are a ZOMBIE and nobody can enqueue itself on
556 * pi_state_list anymore, but we have to be careful
557 * versus waiters unqueueing themselves:
559 raw_spin_lock_irq(&curr->pi_lock);
560 while (!list_empty(head)) {
563 pi_state = list_entry(next, struct futex_pi_state, list);
565 hb = hash_futex(&key);
566 raw_spin_unlock_irq(&curr->pi_lock);
568 spin_lock(&hb->lock);
570 raw_spin_lock_irq(&curr->pi_lock);
572 * We dropped the pi-lock, so re-check whether this
573 * task still owns the PI-state:
575 if (head->next != next) {
576 spin_unlock(&hb->lock);
580 WARN_ON(pi_state->owner != curr);
581 WARN_ON(list_empty(&pi_state->list));
582 list_del_init(&pi_state->list);
583 pi_state->owner = NULL;
584 raw_spin_unlock_irq(&curr->pi_lock);
586 rt_mutex_unlock(&pi_state->pi_mutex);
588 spin_unlock(&hb->lock);
590 raw_spin_lock_irq(&curr->pi_lock);
592 raw_spin_unlock_irq(&curr->pi_lock);
596 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
597 union futex_key *key, struct futex_pi_state **ps)
599 struct futex_pi_state *pi_state = NULL;
600 struct futex_q *this, *next;
601 struct task_struct *p;
602 pid_t pid = uval & FUTEX_TID_MASK;
604 plist_for_each_entry_safe(this, next, &hb->chain, list) {
605 if (match_futex(&this->key, key)) {
607 * Another waiter already exists - bump up
608 * the refcount and return its pi_state:
610 pi_state = this->pi_state;
612 * Userspace might have messed up non-PI and PI futexes
614 if (unlikely(!pi_state))
617 WARN_ON(!atomic_read(&pi_state->refcount));
620 * When pi_state->owner is NULL then the owner died
621 * and another waiter is on the fly. pi_state->owner
622 * is fixed up by the task which acquires
623 * pi_state->rt_mutex.
625 * We do not check for pid == 0 which can happen when
626 * the owner died and robust_list_exit() cleared the
629 if (pid && pi_state->owner) {
631 * Bail out if user space manipulated the
634 if (pid != task_pid_vnr(pi_state->owner))
638 atomic_inc(&pi_state->refcount);
646 * We are the first waiter - try to look up the real owner and attach
647 * the new pi_state to it, but bail out when TID = 0
651 p = futex_find_get_task(pid);
656 * We need to look at the task state flags to figure out,
657 * whether the task is exiting. To protect against the do_exit
658 * change of the task flags, we do this protected by
661 raw_spin_lock_irq(&p->pi_lock);
662 if (unlikely(p->flags & PF_EXITING)) {
664 * The task is on the way out. When PF_EXITPIDONE is
665 * set, we know that the task has finished the
668 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
670 raw_spin_unlock_irq(&p->pi_lock);
675 pi_state = alloc_pi_state();
678 * Initialize the pi_mutex in locked state and make 'p'
681 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
683 /* Store the key for possible exit cleanups: */
684 pi_state->key = *key;
686 WARN_ON(!list_empty(&pi_state->list));
687 list_add(&pi_state->list, &p->pi_state_list);
689 raw_spin_unlock_irq(&p->pi_lock);
699 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
700 * @uaddr: the pi futex user address
701 * @hb: the pi futex hash bucket
702 * @key: the futex key associated with uaddr and hb
703 * @ps: the pi_state pointer where we store the result of the
705 * @task: the task to perform the atomic lock work for. This will
706 * be "current" except in the case of requeue pi.
707 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
711 * 1 - acquired the lock;
714 * The hb->lock and futex_key refs shall be held by the caller.
716 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
717 union futex_key *key,
718 struct futex_pi_state **ps,
719 struct task_struct *task, int set_waiters)
721 int lock_taken, ret, force_take = 0;
722 u32 uval, newval, curval, vpid = task_pid_vnr(task);
725 ret = lock_taken = 0;
728 * To avoid races, we attempt to take the lock here again
729 * (by doing a 0 -> TID atomic cmpxchg), while holding all
730 * the locks. It will most likely not succeed.
734 newval |= FUTEX_WAITERS;
736 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, 0, newval)))
742 if ((unlikely((curval & FUTEX_TID_MASK) == vpid)))
746 * Surprise - we got the lock. Just return to userspace:
748 if (unlikely(!curval))
754 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
755 * to wake at the next unlock.
757 newval = curval | FUTEX_WAITERS;
760 * Should we force take the futex? See below.
762 if (unlikely(force_take)) {
764 * Keep the OWNER_DIED and the WAITERS bit and set the
767 newval = (curval & ~FUTEX_TID_MASK) | vpid;
772 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
774 if (unlikely(curval != uval))
778 * We took the lock due to forced take over.
780 if (unlikely(lock_taken))
784 * We dont have the lock. Look up the PI state (or create it if
785 * we are the first waiter):
787 ret = lookup_pi_state(uval, hb, key, ps);
793 * We failed to find an owner for this
794 * futex. So we have no pi_state to block
795 * on. This can happen in two cases:
798 * 2) A stale FUTEX_WAITERS bit
800 * Re-read the futex value.
802 if (get_futex_value_locked(&curval, uaddr))
806 * If the owner died or we have a stale
807 * WAITERS bit the owner TID in the user space
810 if (!(curval & FUTEX_TID_MASK)) {
823 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
824 * @q: The futex_q to unqueue
826 * The q->lock_ptr must not be NULL and must be held by the caller.
828 static void __unqueue_futex(struct futex_q *q)
830 struct futex_hash_bucket *hb;
832 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
833 || WARN_ON(plist_node_empty(&q->list)))
836 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
837 plist_del(&q->list, &hb->chain);
841 * The hash bucket lock must be held when this is called.
842 * Afterwards, the futex_q must not be accessed.
844 static void wake_futex(struct futex_q *q)
846 struct task_struct *p = q->task;
848 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
852 * We set q->lock_ptr = NULL _before_ we wake up the task. If
853 * a non-futex wake up happens on another CPU then the task
854 * might exit and p would dereference a non-existing task
855 * struct. Prevent this by holding a reference on p across the
862 * The waiting task can free the futex_q as soon as
863 * q->lock_ptr = NULL is written, without taking any locks. A
864 * memory barrier is required here to prevent the following
865 * store to lock_ptr from getting ahead of the plist_del.
870 wake_up_state(p, TASK_NORMAL);
874 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
876 struct task_struct *new_owner;
877 struct futex_pi_state *pi_state = this->pi_state;
878 u32 uninitialized_var(curval), newval;
884 * If current does not own the pi_state then the futex is
885 * inconsistent and user space fiddled with the futex value.
887 if (pi_state->owner != current)
890 raw_spin_lock(&pi_state->pi_mutex.wait_lock);
891 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
894 * It is possible that the next waiter (the one that brought
895 * this owner to the kernel) timed out and is no longer
896 * waiting on the lock.
899 new_owner = this->task;
902 * We pass it to the next owner. (The WAITERS bit is always
903 * kept enabled while there is PI state around. We must also
904 * preserve the owner died bit.)
906 if (!(uval & FUTEX_OWNER_DIED)) {
909 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
911 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
913 else if (curval != uval)
916 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
921 raw_spin_lock_irq(&pi_state->owner->pi_lock);
922 WARN_ON(list_empty(&pi_state->list));
923 list_del_init(&pi_state->list);
924 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
926 raw_spin_lock_irq(&new_owner->pi_lock);
927 WARN_ON(!list_empty(&pi_state->list));
928 list_add(&pi_state->list, &new_owner->pi_state_list);
929 pi_state->owner = new_owner;
930 raw_spin_unlock_irq(&new_owner->pi_lock);
932 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
933 rt_mutex_unlock(&pi_state->pi_mutex);
938 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
940 u32 uninitialized_var(oldval);
943 * There is no waiter, so we unlock the futex. The owner died
944 * bit has not to be preserved here. We are the owner:
946 if (cmpxchg_futex_value_locked(&oldval, uaddr, uval, 0))
955 * Express the locking dependencies for lockdep:
958 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
961 spin_lock(&hb1->lock);
963 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
964 } else { /* hb1 > hb2 */
965 spin_lock(&hb2->lock);
966 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
971 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
973 spin_unlock(&hb1->lock);
975 spin_unlock(&hb2->lock);
979 * Wake up waiters matching bitset queued on this futex (uaddr).
982 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
984 struct futex_hash_bucket *hb;
985 struct futex_q *this, *next;
986 union futex_key key = FUTEX_KEY_INIT;
992 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
993 if (unlikely(ret != 0))
996 hb = hash_futex(&key);
997 spin_lock(&hb->lock);
999 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1000 if (match_futex (&this->key, &key)) {
1001 if (this->pi_state || this->rt_waiter) {
1006 /* Check if one of the bits is set in both bitsets */
1007 if (!(this->bitset & bitset))
1011 if (++ret >= nr_wake)
1016 spin_unlock(&hb->lock);
1017 put_futex_key(&key);
1023 * Wake up all waiters hashed on the physical page that is mapped
1024 * to this virtual address:
1027 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1028 int nr_wake, int nr_wake2, int op)
1030 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1031 struct futex_hash_bucket *hb1, *hb2;
1032 struct futex_q *this, *next;
1036 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1037 if (unlikely(ret != 0))
1039 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1040 if (unlikely(ret != 0))
1043 hb1 = hash_futex(&key1);
1044 hb2 = hash_futex(&key2);
1047 double_lock_hb(hb1, hb2);
1048 op_ret = futex_atomic_op_inuser(op, uaddr2);
1049 if (unlikely(op_ret < 0)) {
1051 double_unlock_hb(hb1, hb2);
1055 * we don't get EFAULT from MMU faults if we don't have an MMU,
1056 * but we might get them from range checking
1062 if (unlikely(op_ret != -EFAULT)) {
1067 ret = fault_in_user_writeable(uaddr2);
1071 if (!(flags & FLAGS_SHARED))
1074 put_futex_key(&key2);
1075 put_futex_key(&key1);
1079 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1080 if (match_futex (&this->key, &key1)) {
1081 if (this->pi_state || this->rt_waiter) {
1086 if (++ret >= nr_wake)
1093 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1094 if (match_futex (&this->key, &key2)) {
1095 if (this->pi_state || this->rt_waiter) {
1100 if (++op_ret >= nr_wake2)
1108 double_unlock_hb(hb1, hb2);
1110 put_futex_key(&key2);
1112 put_futex_key(&key1);
1118 * requeue_futex() - Requeue a futex_q from one hb to another
1119 * @q: the futex_q to requeue
1120 * @hb1: the source hash_bucket
1121 * @hb2: the target hash_bucket
1122 * @key2: the new key for the requeued futex_q
1125 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1126 struct futex_hash_bucket *hb2, union futex_key *key2)
1130 * If key1 and key2 hash to the same bucket, no need to
1133 if (likely(&hb1->chain != &hb2->chain)) {
1134 plist_del(&q->list, &hb1->chain);
1135 plist_add(&q->list, &hb2->chain);
1136 q->lock_ptr = &hb2->lock;
1138 get_futex_key_refs(key2);
1143 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1145 * @key: the key of the requeue target futex
1146 * @hb: the hash_bucket of the requeue target futex
1148 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1149 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1150 * to the requeue target futex so the waiter can detect the wakeup on the right
1151 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1152 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1153 * to protect access to the pi_state to fixup the owner later. Must be called
1154 * with both q->lock_ptr and hb->lock held.
1157 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1158 struct futex_hash_bucket *hb)
1160 get_futex_key_refs(key);
1165 WARN_ON(!q->rt_waiter);
1166 q->rt_waiter = NULL;
1168 q->lock_ptr = &hb->lock;
1170 wake_up_state(q->task, TASK_NORMAL);
1174 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1175 * @pifutex: the user address of the to futex
1176 * @hb1: the from futex hash bucket, must be locked by the caller
1177 * @hb2: the to futex hash bucket, must be locked by the caller
1178 * @key1: the from futex key
1179 * @key2: the to futex key
1180 * @ps: address to store the pi_state pointer
1181 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1183 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1184 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1185 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1186 * hb1 and hb2 must be held by the caller.
1189 * 0 - failed to acquire the lock atomically;
1190 * 1 - acquired the lock;
1193 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1194 struct futex_hash_bucket *hb1,
1195 struct futex_hash_bucket *hb2,
1196 union futex_key *key1, union futex_key *key2,
1197 struct futex_pi_state **ps, int set_waiters)
1199 struct futex_q *top_waiter = NULL;
1203 if (get_futex_value_locked(&curval, pifutex))
1207 * Find the top_waiter and determine if there are additional waiters.
1208 * If the caller intends to requeue more than 1 waiter to pifutex,
1209 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1210 * as we have means to handle the possible fault. If not, don't set
1211 * the bit unecessarily as it will force the subsequent unlock to enter
1214 top_waiter = futex_top_waiter(hb1, key1);
1216 /* There are no waiters, nothing for us to do. */
1220 /* Ensure we requeue to the expected futex. */
1221 if (!match_futex(top_waiter->requeue_pi_key, key2))
1225 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1226 * the contended case or if set_waiters is 1. The pi_state is returned
1227 * in ps in contended cases.
1229 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1232 requeue_pi_wake_futex(top_waiter, key2, hb2);
1238 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1239 * @uaddr1: source futex user address
1240 * @flags: futex flags (FLAGS_SHARED, etc.)
1241 * @uaddr2: target futex user address
1242 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1243 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1244 * @cmpval: @uaddr1 expected value (or %NULL)
1245 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1246 * pi futex (pi to pi requeue is not supported)
1248 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1249 * uaddr2 atomically on behalf of the top waiter.
1252 * >=0 - on success, the number of tasks requeued or woken;
1255 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1256 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1257 u32 *cmpval, int requeue_pi)
1259 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1260 int drop_count = 0, task_count = 0, ret;
1261 struct futex_pi_state *pi_state = NULL;
1262 struct futex_hash_bucket *hb1, *hb2;
1263 struct futex_q *this, *next;
1268 * requeue_pi requires a pi_state, try to allocate it now
1269 * without any locks in case it fails.
1271 if (refill_pi_state_cache())
1274 * requeue_pi must wake as many tasks as it can, up to nr_wake
1275 * + nr_requeue, since it acquires the rt_mutex prior to
1276 * returning to userspace, so as to not leave the rt_mutex with
1277 * waiters and no owner. However, second and third wake-ups
1278 * cannot be predicted as they involve race conditions with the
1279 * first wake and a fault while looking up the pi_state. Both
1280 * pthread_cond_signal() and pthread_cond_broadcast() should
1288 if (pi_state != NULL) {
1290 * We will have to lookup the pi_state again, so free this one
1291 * to keep the accounting correct.
1293 free_pi_state(pi_state);
1297 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1298 if (unlikely(ret != 0))
1300 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1301 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1302 if (unlikely(ret != 0))
1305 hb1 = hash_futex(&key1);
1306 hb2 = hash_futex(&key2);
1309 double_lock_hb(hb1, hb2);
1311 if (likely(cmpval != NULL)) {
1314 ret = get_futex_value_locked(&curval, uaddr1);
1316 if (unlikely(ret)) {
1317 double_unlock_hb(hb1, hb2);
1319 ret = get_user(curval, uaddr1);
1323 if (!(flags & FLAGS_SHARED))
1326 put_futex_key(&key2);
1327 put_futex_key(&key1);
1330 if (curval != *cmpval) {
1336 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1338 * Attempt to acquire uaddr2 and wake the top waiter. If we
1339 * intend to requeue waiters, force setting the FUTEX_WAITERS
1340 * bit. We force this here where we are able to easily handle
1341 * faults rather in the requeue loop below.
1343 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1344 &key2, &pi_state, nr_requeue);
1347 * At this point the top_waiter has either taken uaddr2 or is
1348 * waiting on it. If the former, then the pi_state will not
1349 * exist yet, look it up one more time to ensure we have a
1356 ret = get_futex_value_locked(&curval2, uaddr2);
1358 ret = lookup_pi_state(curval2, hb2, &key2,
1366 double_unlock_hb(hb1, hb2);
1367 put_futex_key(&key2);
1368 put_futex_key(&key1);
1369 ret = fault_in_user_writeable(uaddr2);
1374 /* The owner was exiting, try again. */
1375 double_unlock_hb(hb1, hb2);
1376 put_futex_key(&key2);
1377 put_futex_key(&key1);
1385 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1386 if (task_count - nr_wake >= nr_requeue)
1389 if (!match_futex(&this->key, &key1))
1393 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1394 * be paired with each other and no other futex ops.
1396 * We should never be requeueing a futex_q with a pi_state,
1397 * which is awaiting a futex_unlock_pi().
1399 if ((requeue_pi && !this->rt_waiter) ||
1400 (!requeue_pi && this->rt_waiter) ||
1407 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1408 * lock, we already woke the top_waiter. If not, it will be
1409 * woken by futex_unlock_pi().
1411 if (++task_count <= nr_wake && !requeue_pi) {
1416 /* Ensure we requeue to the expected futex for requeue_pi. */
1417 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1423 * Requeue nr_requeue waiters and possibly one more in the case
1424 * of requeue_pi if we couldn't acquire the lock atomically.
1427 /* Prepare the waiter to take the rt_mutex. */
1428 atomic_inc(&pi_state->refcount);
1429 this->pi_state = pi_state;
1430 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1434 /* We got the lock. */
1435 requeue_pi_wake_futex(this, &key2, hb2);
1440 this->pi_state = NULL;
1441 free_pi_state(pi_state);
1445 requeue_futex(this, hb1, hb2, &key2);
1450 double_unlock_hb(hb1, hb2);
1453 * drop_futex_key_refs() must be called outside the spinlocks. During
1454 * the requeue we moved futex_q's from the hash bucket at key1 to the
1455 * one at key2 and updated their key pointer. We no longer need to
1456 * hold the references to key1.
1458 while (--drop_count >= 0)
1459 drop_futex_key_refs(&key1);
1462 put_futex_key(&key2);
1464 put_futex_key(&key1);
1466 if (pi_state != NULL)
1467 free_pi_state(pi_state);
1468 return ret ? ret : task_count;
1471 /* The key must be already stored in q->key. */
1472 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1473 __acquires(&hb->lock)
1475 struct futex_hash_bucket *hb;
1477 hb = hash_futex(&q->key);
1478 q->lock_ptr = &hb->lock;
1480 spin_lock(&hb->lock);
1485 queue_unlock(struct futex_hash_bucket *hb)
1486 __releases(&hb->lock)
1488 spin_unlock(&hb->lock);
1492 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1493 * @q: The futex_q to enqueue
1494 * @hb: The destination hash bucket
1496 * The hb->lock must be held by the caller, and is released here. A call to
1497 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1498 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1499 * or nothing if the unqueue is done as part of the wake process and the unqueue
1500 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1503 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1504 __releases(&hb->lock)
1509 * The priority used to register this element is
1510 * - either the real thread-priority for the real-time threads
1511 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1512 * - or MAX_RT_PRIO for non-RT threads.
1513 * Thus, all RT-threads are woken first in priority order, and
1514 * the others are woken last, in FIFO order.
1516 prio = min(current->normal_prio, MAX_RT_PRIO);
1518 plist_node_init(&q->list, prio);
1519 plist_add(&q->list, &hb->chain);
1521 spin_unlock(&hb->lock);
1525 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1526 * @q: The futex_q to unqueue
1528 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1529 * be paired with exactly one earlier call to queue_me().
1532 * 1 - if the futex_q was still queued (and we removed unqueued it);
1533 * 0 - if the futex_q was already removed by the waking thread
1535 static int unqueue_me(struct futex_q *q)
1537 spinlock_t *lock_ptr;
1540 /* In the common case we don't take the spinlock, which is nice. */
1542 lock_ptr = q->lock_ptr;
1544 if (lock_ptr != NULL) {
1545 spin_lock(lock_ptr);
1547 * q->lock_ptr can change between reading it and
1548 * spin_lock(), causing us to take the wrong lock. This
1549 * corrects the race condition.
1551 * Reasoning goes like this: if we have the wrong lock,
1552 * q->lock_ptr must have changed (maybe several times)
1553 * between reading it and the spin_lock(). It can
1554 * change again after the spin_lock() but only if it was
1555 * already changed before the spin_lock(). It cannot,
1556 * however, change back to the original value. Therefore
1557 * we can detect whether we acquired the correct lock.
1559 if (unlikely(lock_ptr != q->lock_ptr)) {
1560 spin_unlock(lock_ptr);
1565 BUG_ON(q->pi_state);
1567 spin_unlock(lock_ptr);
1571 drop_futex_key_refs(&q->key);
1576 * PI futexes can not be requeued and must remove themself from the
1577 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1580 static void unqueue_me_pi(struct futex_q *q)
1581 __releases(q->lock_ptr)
1585 BUG_ON(!q->pi_state);
1586 free_pi_state(q->pi_state);
1589 spin_unlock(q->lock_ptr);
1593 * Fixup the pi_state owner with the new owner.
1595 * Must be called with hash bucket lock held and mm->sem held for non
1598 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1599 struct task_struct *newowner)
1601 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1602 struct futex_pi_state *pi_state = q->pi_state;
1603 struct task_struct *oldowner = pi_state->owner;
1604 u32 uval, uninitialized_var(curval), newval;
1608 if (!pi_state->owner)
1609 newtid |= FUTEX_OWNER_DIED;
1612 * We are here either because we stole the rtmutex from the
1613 * previous highest priority waiter or we are the highest priority
1614 * waiter but failed to get the rtmutex the first time.
1615 * We have to replace the newowner TID in the user space variable.
1616 * This must be atomic as we have to preserve the owner died bit here.
1618 * Note: We write the user space value _before_ changing the pi_state
1619 * because we can fault here. Imagine swapped out pages or a fork
1620 * that marked all the anonymous memory readonly for cow.
1622 * Modifying pi_state _before_ the user space value would
1623 * leave the pi_state in an inconsistent state when we fault
1624 * here, because we need to drop the hash bucket lock to
1625 * handle the fault. This might be observed in the PID check
1626 * in lookup_pi_state.
1629 if (get_futex_value_locked(&uval, uaddr))
1633 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1635 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1643 * We fixed up user space. Now we need to fix the pi_state
1646 if (pi_state->owner != NULL) {
1647 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1648 WARN_ON(list_empty(&pi_state->list));
1649 list_del_init(&pi_state->list);
1650 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1653 pi_state->owner = newowner;
1655 raw_spin_lock_irq(&newowner->pi_lock);
1656 WARN_ON(!list_empty(&pi_state->list));
1657 list_add(&pi_state->list, &newowner->pi_state_list);
1658 raw_spin_unlock_irq(&newowner->pi_lock);
1662 * To handle the page fault we need to drop the hash bucket
1663 * lock here. That gives the other task (either the highest priority
1664 * waiter itself or the task which stole the rtmutex) the
1665 * chance to try the fixup of the pi_state. So once we are
1666 * back from handling the fault we need to check the pi_state
1667 * after reacquiring the hash bucket lock and before trying to
1668 * do another fixup. When the fixup has been done already we
1672 spin_unlock(q->lock_ptr);
1674 ret = fault_in_user_writeable(uaddr);
1676 spin_lock(q->lock_ptr);
1679 * Check if someone else fixed it for us:
1681 if (pi_state->owner != oldowner)
1690 static long futex_wait_restart(struct restart_block *restart);
1693 * fixup_owner() - Post lock pi_state and corner case management
1694 * @uaddr: user address of the futex
1695 * @q: futex_q (contains pi_state and access to the rt_mutex)
1696 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1698 * After attempting to lock an rt_mutex, this function is called to cleanup
1699 * the pi_state owner as well as handle race conditions that may allow us to
1700 * acquire the lock. Must be called with the hb lock held.
1703 * 1 - success, lock taken;
1704 * 0 - success, lock not taken;
1705 * <0 - on error (-EFAULT)
1707 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
1709 struct task_struct *owner;
1714 * Got the lock. We might not be the anticipated owner if we
1715 * did a lock-steal - fix up the PI-state in that case:
1717 if (q->pi_state->owner != current)
1718 ret = fixup_pi_state_owner(uaddr, q, current);
1723 * Catch the rare case, where the lock was released when we were on the
1724 * way back before we locked the hash bucket.
1726 if (q->pi_state->owner == current) {
1728 * Try to get the rt_mutex now. This might fail as some other
1729 * task acquired the rt_mutex after we removed ourself from the
1730 * rt_mutex waiters list.
1732 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1738 * pi_state is incorrect, some other task did a lock steal and
1739 * we returned due to timeout or signal without taking the
1740 * rt_mutex. Too late.
1742 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
1743 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1745 owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
1746 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
1747 ret = fixup_pi_state_owner(uaddr, q, owner);
1752 * Paranoia check. If we did not take the lock, then we should not be
1753 * the owner of the rt_mutex.
1755 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1756 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1757 "pi-state %p\n", ret,
1758 q->pi_state->pi_mutex.owner,
1759 q->pi_state->owner);
1762 return ret ? ret : locked;
1766 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1767 * @hb: the futex hash bucket, must be locked by the caller
1768 * @q: the futex_q to queue up on
1769 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1771 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1772 struct hrtimer_sleeper *timeout)
1775 * The task state is guaranteed to be set before another task can
1776 * wake it. set_current_state() is implemented using set_mb() and
1777 * queue_me() calls spin_unlock() upon completion, both serializing
1778 * access to the hash list and forcing another memory barrier.
1780 set_current_state(TASK_INTERRUPTIBLE);
1785 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1786 if (!hrtimer_active(&timeout->timer))
1787 timeout->task = NULL;
1791 * If we have been removed from the hash list, then another task
1792 * has tried to wake us, and we can skip the call to schedule().
1794 if (likely(!plist_node_empty(&q->list))) {
1796 * If the timer has already expired, current will already be
1797 * flagged for rescheduling. Only call schedule if there
1798 * is no timeout, or if it has yet to expire.
1800 if (!timeout || timeout->task)
1801 freezable_schedule();
1803 __set_current_state(TASK_RUNNING);
1807 * futex_wait_setup() - Prepare to wait on a futex
1808 * @uaddr: the futex userspace address
1809 * @val: the expected value
1810 * @flags: futex flags (FLAGS_SHARED, etc.)
1811 * @q: the associated futex_q
1812 * @hb: storage for hash_bucket pointer to be returned to caller
1814 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1815 * compare it with the expected value. Handle atomic faults internally.
1816 * Return with the hb lock held and a q.key reference on success, and unlocked
1817 * with no q.key reference on failure.
1820 * 0 - uaddr contains val and hb has been locked;
1821 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
1823 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
1824 struct futex_q *q, struct futex_hash_bucket **hb)
1830 * Access the page AFTER the hash-bucket is locked.
1831 * Order is important:
1833 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1834 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1836 * The basic logical guarantee of a futex is that it blocks ONLY
1837 * if cond(var) is known to be true at the time of blocking, for
1838 * any cond. If we locked the hash-bucket after testing *uaddr, that
1839 * would open a race condition where we could block indefinitely with
1840 * cond(var) false, which would violate the guarantee.
1842 * On the other hand, we insert q and release the hash-bucket only
1843 * after testing *uaddr. This guarantees that futex_wait() will NOT
1844 * absorb a wakeup if *uaddr does not match the desired values
1845 * while the syscall executes.
1848 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
1849 if (unlikely(ret != 0))
1853 *hb = queue_lock(q);
1855 ret = get_futex_value_locked(&uval, uaddr);
1860 ret = get_user(uval, uaddr);
1864 if (!(flags & FLAGS_SHARED))
1867 put_futex_key(&q->key);
1878 put_futex_key(&q->key);
1882 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
1883 ktime_t *abs_time, u32 bitset)
1885 struct hrtimer_sleeper timeout, *to = NULL;
1886 struct restart_block *restart;
1887 struct futex_hash_bucket *hb;
1888 struct futex_q q = futex_q_init;
1898 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
1899 CLOCK_REALTIME : CLOCK_MONOTONIC,
1901 hrtimer_init_sleeper(to, current);
1902 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1903 current->timer_slack_ns);
1908 * Prepare to wait on uaddr. On success, holds hb lock and increments
1911 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
1915 /* queue_me and wait for wakeup, timeout, or a signal. */
1916 futex_wait_queue_me(hb, &q, to);
1918 /* If we were woken (and unqueued), we succeeded, whatever. */
1920 /* unqueue_me() drops q.key ref */
1921 if (!unqueue_me(&q))
1924 if (to && !to->task)
1928 * We expect signal_pending(current), but we might be the
1929 * victim of a spurious wakeup as well.
1931 if (!signal_pending(current))
1938 restart = ¤t_thread_info()->restart_block;
1939 restart->fn = futex_wait_restart;
1940 restart->futex.uaddr = uaddr;
1941 restart->futex.val = val;
1942 restart->futex.time = abs_time->tv64;
1943 restart->futex.bitset = bitset;
1944 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
1946 ret = -ERESTART_RESTARTBLOCK;
1950 hrtimer_cancel(&to->timer);
1951 destroy_hrtimer_on_stack(&to->timer);
1957 static long futex_wait_restart(struct restart_block *restart)
1959 u32 __user *uaddr = restart->futex.uaddr;
1960 ktime_t t, *tp = NULL;
1962 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1963 t.tv64 = restart->futex.time;
1966 restart->fn = do_no_restart_syscall;
1968 return (long)futex_wait(uaddr, restart->futex.flags,
1969 restart->futex.val, tp, restart->futex.bitset);
1974 * Userspace tried a 0 -> TID atomic transition of the futex value
1975 * and failed. The kernel side here does the whole locking operation:
1976 * if there are waiters then it will block, it does PI, etc. (Due to
1977 * races the kernel might see a 0 value of the futex too.)
1979 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
1980 ktime_t *time, int trylock)
1982 struct hrtimer_sleeper timeout, *to = NULL;
1983 struct futex_hash_bucket *hb;
1984 struct futex_q q = futex_q_init;
1987 if (refill_pi_state_cache())
1992 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1994 hrtimer_init_sleeper(to, current);
1995 hrtimer_set_expires(&to->timer, *time);
1999 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2000 if (unlikely(ret != 0))
2004 hb = queue_lock(&q);
2006 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2007 if (unlikely(ret)) {
2010 /* We got the lock. */
2012 goto out_unlock_put_key;
2017 * Task is exiting and we just wait for the
2021 put_futex_key(&q.key);
2025 goto out_unlock_put_key;
2030 * Only actually queue now that the atomic ops are done:
2034 WARN_ON(!q.pi_state);
2036 * Block on the PI mutex:
2039 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
2041 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2042 /* Fixup the trylock return value: */
2043 ret = ret ? 0 : -EWOULDBLOCK;
2046 spin_lock(q.lock_ptr);
2048 * Fixup the pi_state owner and possibly acquire the lock if we
2051 res = fixup_owner(uaddr, &q, !ret);
2053 * If fixup_owner() returned an error, proprogate that. If it acquired
2054 * the lock, clear our -ETIMEDOUT or -EINTR.
2057 ret = (res < 0) ? res : 0;
2060 * If fixup_owner() faulted and was unable to handle the fault, unlock
2061 * it and return the fault to userspace.
2063 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2064 rt_mutex_unlock(&q.pi_state->pi_mutex);
2066 /* Unqueue and drop the lock */
2075 put_futex_key(&q.key);
2078 destroy_hrtimer_on_stack(&to->timer);
2079 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2084 ret = fault_in_user_writeable(uaddr);
2088 if (!(flags & FLAGS_SHARED))
2091 put_futex_key(&q.key);
2096 * Userspace attempted a TID -> 0 atomic transition, and failed.
2097 * This is the in-kernel slowpath: we look up the PI state (if any),
2098 * and do the rt-mutex unlock.
2100 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2102 struct futex_hash_bucket *hb;
2103 struct futex_q *this, *next;
2104 union futex_key key = FUTEX_KEY_INIT;
2105 u32 uval, vpid = task_pid_vnr(current);
2109 if (get_user(uval, uaddr))
2112 * We release only a lock we actually own:
2114 if ((uval & FUTEX_TID_MASK) != vpid)
2117 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2118 if (unlikely(ret != 0))
2121 hb = hash_futex(&key);
2122 spin_lock(&hb->lock);
2125 * To avoid races, try to do the TID -> 0 atomic transition
2126 * again. If it succeeds then we can return without waking
2129 if (!(uval & FUTEX_OWNER_DIED) &&
2130 cmpxchg_futex_value_locked(&uval, uaddr, vpid, 0))
2133 * Rare case: we managed to release the lock atomically,
2134 * no need to wake anyone else up:
2136 if (unlikely(uval == vpid))
2140 * Ok, other tasks may need to be woken up - check waiters
2141 * and do the wakeup if necessary:
2143 plist_for_each_entry_safe(this, next, &hb->chain, list) {
2144 if (!match_futex (&this->key, &key))
2146 ret = wake_futex_pi(uaddr, uval, this);
2148 * The atomic access to the futex value
2149 * generated a pagefault, so retry the
2150 * user-access and the wakeup:
2157 * No waiters - kernel unlocks the futex:
2159 if (!(uval & FUTEX_OWNER_DIED)) {
2160 ret = unlock_futex_pi(uaddr, uval);
2166 spin_unlock(&hb->lock);
2167 put_futex_key(&key);
2173 spin_unlock(&hb->lock);
2174 put_futex_key(&key);
2176 ret = fault_in_user_writeable(uaddr);
2184 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2185 * @hb: the hash_bucket futex_q was original enqueued on
2186 * @q: the futex_q woken while waiting to be requeued
2187 * @key2: the futex_key of the requeue target futex
2188 * @timeout: the timeout associated with the wait (NULL if none)
2190 * Detect if the task was woken on the initial futex as opposed to the requeue
2191 * target futex. If so, determine if it was a timeout or a signal that caused
2192 * the wakeup and return the appropriate error code to the caller. Must be
2193 * called with the hb lock held.
2196 * 0 = no early wakeup detected;
2197 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2200 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2201 struct futex_q *q, union futex_key *key2,
2202 struct hrtimer_sleeper *timeout)
2207 * With the hb lock held, we avoid races while we process the wakeup.
2208 * We only need to hold hb (and not hb2) to ensure atomicity as the
2209 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2210 * It can't be requeued from uaddr2 to something else since we don't
2211 * support a PI aware source futex for requeue.
2213 if (!match_futex(&q->key, key2)) {
2214 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2216 * We were woken prior to requeue by a timeout or a signal.
2217 * Unqueue the futex_q and determine which it was.
2219 plist_del(&q->list, &hb->chain);
2221 /* Handle spurious wakeups gracefully */
2223 if (timeout && !timeout->task)
2225 else if (signal_pending(current))
2226 ret = -ERESTARTNOINTR;
2232 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2233 * @uaddr: the futex we initially wait on (non-pi)
2234 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2235 * the same type, no requeueing from private to shared, etc.
2236 * @val: the expected value of uaddr
2237 * @abs_time: absolute timeout
2238 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2239 * @uaddr2: the pi futex we will take prior to returning to user-space
2241 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2242 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2243 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2244 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2245 * without one, the pi logic would not know which task to boost/deboost, if
2246 * there was a need to.
2248 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2249 * via the following--
2250 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2251 * 2) wakeup on uaddr2 after a requeue
2255 * If 3, cleanup and return -ERESTARTNOINTR.
2257 * If 2, we may then block on trying to take the rt_mutex and return via:
2258 * 5) successful lock
2261 * 8) other lock acquisition failure
2263 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2265 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2271 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2272 u32 val, ktime_t *abs_time, u32 bitset,
2275 struct hrtimer_sleeper timeout, *to = NULL;
2276 struct rt_mutex_waiter rt_waiter;
2277 struct rt_mutex *pi_mutex = NULL;
2278 struct futex_hash_bucket *hb;
2279 union futex_key key2 = FUTEX_KEY_INIT;
2280 struct futex_q q = futex_q_init;
2283 if (uaddr == uaddr2)
2291 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2292 CLOCK_REALTIME : CLOCK_MONOTONIC,
2294 hrtimer_init_sleeper(to, current);
2295 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2296 current->timer_slack_ns);
2300 * The waiter is allocated on our stack, manipulated by the requeue
2301 * code while we sleep on uaddr.
2303 debug_rt_mutex_init_waiter(&rt_waiter);
2304 rt_waiter.task = NULL;
2306 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2307 if (unlikely(ret != 0))
2311 q.rt_waiter = &rt_waiter;
2312 q.requeue_pi_key = &key2;
2315 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2318 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2322 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2323 futex_wait_queue_me(hb, &q, to);
2325 spin_lock(&hb->lock);
2326 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2327 spin_unlock(&hb->lock);
2332 * In order for us to be here, we know our q.key == key2, and since
2333 * we took the hb->lock above, we also know that futex_requeue() has
2334 * completed and we no longer have to concern ourselves with a wakeup
2335 * race with the atomic proxy lock acquisition by the requeue code. The
2336 * futex_requeue dropped our key1 reference and incremented our key2
2340 /* Check if the requeue code acquired the second futex for us. */
2343 * Got the lock. We might not be the anticipated owner if we
2344 * did a lock-steal - fix up the PI-state in that case.
2346 if (q.pi_state && (q.pi_state->owner != current)) {
2347 spin_lock(q.lock_ptr);
2348 ret = fixup_pi_state_owner(uaddr2, &q, current);
2349 spin_unlock(q.lock_ptr);
2353 * We have been woken up by futex_unlock_pi(), a timeout, or a
2354 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2357 WARN_ON(!q.pi_state);
2358 pi_mutex = &q.pi_state->pi_mutex;
2359 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2360 debug_rt_mutex_free_waiter(&rt_waiter);
2362 spin_lock(q.lock_ptr);
2364 * Fixup the pi_state owner and possibly acquire the lock if we
2367 res = fixup_owner(uaddr2, &q, !ret);
2369 * If fixup_owner() returned an error, proprogate that. If it
2370 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2373 ret = (res < 0) ? res : 0;
2375 /* Unqueue and drop the lock. */
2380 * If fixup_pi_state_owner() faulted and was unable to handle the
2381 * fault, unlock the rt_mutex and return the fault to userspace.
2383 if (ret == -EFAULT) {
2384 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2385 rt_mutex_unlock(pi_mutex);
2386 } else if (ret == -EINTR) {
2388 * We've already been requeued, but cannot restart by calling
2389 * futex_lock_pi() directly. We could restart this syscall, but
2390 * it would detect that the user space "val" changed and return
2391 * -EWOULDBLOCK. Save the overhead of the restart and return
2392 * -EWOULDBLOCK directly.
2398 put_futex_key(&q.key);
2400 put_futex_key(&key2);
2404 hrtimer_cancel(&to->timer);
2405 destroy_hrtimer_on_stack(&to->timer);
2411 * Support for robust futexes: the kernel cleans up held futexes at
2414 * Implementation: user-space maintains a per-thread list of locks it
2415 * is holding. Upon do_exit(), the kernel carefully walks this list,
2416 * and marks all locks that are owned by this thread with the
2417 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2418 * always manipulated with the lock held, so the list is private and
2419 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2420 * field, to allow the kernel to clean up if the thread dies after
2421 * acquiring the lock, but just before it could have added itself to
2422 * the list. There can only be one such pending lock.
2426 * sys_set_robust_list() - Set the robust-futex list head of a task
2427 * @head: pointer to the list-head
2428 * @len: length of the list-head, as userspace expects
2430 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2433 if (!futex_cmpxchg_enabled)
2436 * The kernel knows only one size for now:
2438 if (unlikely(len != sizeof(*head)))
2441 current->robust_list = head;
2447 * sys_get_robust_list() - Get the robust-futex list head of a task
2448 * @pid: pid of the process [zero for current task]
2449 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2450 * @len_ptr: pointer to a length field, the kernel fills in the header size
2452 SYSCALL_DEFINE3(get_robust_list, int, pid,
2453 struct robust_list_head __user * __user *, head_ptr,
2454 size_t __user *, len_ptr)
2456 struct robust_list_head __user *head;
2458 struct task_struct *p;
2460 if (!futex_cmpxchg_enabled)
2469 p = find_task_by_vpid(pid);
2475 if (!ptrace_may_access(p, PTRACE_MODE_READ))
2478 head = p->robust_list;
2481 if (put_user(sizeof(*head), len_ptr))
2483 return put_user(head, head_ptr);
2492 * Process a futex-list entry, check whether it's owned by the
2493 * dying task, and do notification if so:
2495 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2497 u32 uval, uninitialized_var(nval), mval;
2500 if (get_user(uval, uaddr))
2503 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2505 * Ok, this dying thread is truly holding a futex
2506 * of interest. Set the OWNER_DIED bit atomically
2507 * via cmpxchg, and if the value had FUTEX_WAITERS
2508 * set, wake up a waiter (if any). (We have to do a
2509 * futex_wake() even if OWNER_DIED is already set -
2510 * to handle the rare but possible case of recursive
2511 * thread-death.) The rest of the cleanup is done in
2514 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2516 * We are not holding a lock here, but we want to have
2517 * the pagefault_disable/enable() protection because
2518 * we want to handle the fault gracefully. If the
2519 * access fails we try to fault in the futex with R/W
2520 * verification via get_user_pages. get_user() above
2521 * does not guarantee R/W access. If that fails we
2522 * give up and leave the futex locked.
2524 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2525 if (fault_in_user_writeable(uaddr))
2533 * Wake robust non-PI futexes here. The wakeup of
2534 * PI futexes happens in exit_pi_state():
2536 if (!pi && (uval & FUTEX_WAITERS))
2537 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2543 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2545 static inline int fetch_robust_entry(struct robust_list __user **entry,
2546 struct robust_list __user * __user *head,
2549 unsigned long uentry;
2551 if (get_user(uentry, (unsigned long __user *)head))
2554 *entry = (void __user *)(uentry & ~1UL);
2561 * Walk curr->robust_list (very carefully, it's a userspace list!)
2562 * and mark any locks found there dead, and notify any waiters.
2564 * We silently return on any sign of list-walking problem.
2566 void exit_robust_list(struct task_struct *curr)
2568 struct robust_list_head __user *head = curr->robust_list;
2569 struct robust_list __user *entry, *next_entry, *pending;
2570 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2571 unsigned int uninitialized_var(next_pi);
2572 unsigned long futex_offset;
2575 if (!futex_cmpxchg_enabled)
2579 * Fetch the list head (which was registered earlier, via
2580 * sys_set_robust_list()):
2582 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2585 * Fetch the relative futex offset:
2587 if (get_user(futex_offset, &head->futex_offset))
2590 * Fetch any possibly pending lock-add first, and handle it
2593 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2596 next_entry = NULL; /* avoid warning with gcc */
2597 while (entry != &head->list) {
2599 * Fetch the next entry in the list before calling
2600 * handle_futex_death:
2602 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2604 * A pending lock might already be on the list, so
2605 * don't process it twice:
2607 if (entry != pending)
2608 if (handle_futex_death((void __user *)entry + futex_offset,
2616 * Avoid excessively long or circular lists:
2625 handle_futex_death((void __user *)pending + futex_offset,
2629 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2630 u32 __user *uaddr2, u32 val2, u32 val3)
2632 int cmd = op & FUTEX_CMD_MASK;
2633 unsigned int flags = 0;
2635 if (!(op & FUTEX_PRIVATE_FLAG))
2636 flags |= FLAGS_SHARED;
2638 if (op & FUTEX_CLOCK_REALTIME) {
2639 flags |= FLAGS_CLOCKRT;
2640 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2646 case FUTEX_UNLOCK_PI:
2647 case FUTEX_TRYLOCK_PI:
2648 case FUTEX_WAIT_REQUEUE_PI:
2649 case FUTEX_CMP_REQUEUE_PI:
2650 if (!futex_cmpxchg_enabled)
2656 val3 = FUTEX_BITSET_MATCH_ANY;
2657 case FUTEX_WAIT_BITSET:
2658 return futex_wait(uaddr, flags, val, timeout, val3);
2660 val3 = FUTEX_BITSET_MATCH_ANY;
2661 case FUTEX_WAKE_BITSET:
2662 return futex_wake(uaddr, flags, val, val3);
2664 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2665 case FUTEX_CMP_REQUEUE:
2666 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2668 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2670 return futex_lock_pi(uaddr, flags, val, timeout, 0);
2671 case FUTEX_UNLOCK_PI:
2672 return futex_unlock_pi(uaddr, flags);
2673 case FUTEX_TRYLOCK_PI:
2674 return futex_lock_pi(uaddr, flags, 0, timeout, 1);
2675 case FUTEX_WAIT_REQUEUE_PI:
2676 val3 = FUTEX_BITSET_MATCH_ANY;
2677 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2679 case FUTEX_CMP_REQUEUE_PI:
2680 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2686 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2687 struct timespec __user *, utime, u32 __user *, uaddr2,
2691 ktime_t t, *tp = NULL;
2693 int cmd = op & FUTEX_CMD_MASK;
2695 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2696 cmd == FUTEX_WAIT_BITSET ||
2697 cmd == FUTEX_WAIT_REQUEUE_PI)) {
2698 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2700 if (!timespec_valid(&ts))
2703 t = timespec_to_ktime(ts);
2704 if (cmd == FUTEX_WAIT)
2705 t = ktime_add_safe(ktime_get(), t);
2709 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2710 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2712 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2713 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2714 val2 = (u32) (unsigned long) utime;
2716 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2719 static int __init futex_init(void)
2725 * This will fail and we want it. Some arch implementations do
2726 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2727 * functionality. We want to know that before we call in any
2728 * of the complex code paths. Also we want to prevent
2729 * registration of robust lists in that case. NULL is
2730 * guaranteed to fault and we get -EFAULT on functional
2731 * implementation, the non-functional ones will return
2734 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
2735 futex_cmpxchg_enabled = 1;
2737 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2738 plist_head_init(&futex_queues[i].chain);
2739 spin_lock_init(&futex_queues[i].lock);
2744 __initcall(futex_init);