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>
64 #include <asm/futex.h>
66 #include "rtmutex_common.h"
68 int __read_mostly futex_cmpxchg_enabled;
70 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
73 * Futex flags used to encode options to functions and preserve them across
76 #define FLAGS_SHARED 0x01
77 #define FLAGS_CLOCKRT 0x02
78 #define FLAGS_HAS_TIMEOUT 0x04
81 * Priority Inheritance state:
83 struct futex_pi_state {
85 * list of 'owned' pi_state instances - these have to be
86 * cleaned up in do_exit() if the task exits prematurely:
88 struct list_head list;
93 struct rt_mutex pi_mutex;
95 struct task_struct *owner;
102 * struct futex_q - The hashed futex queue entry, one per waiting task
103 * @list: priority-sorted list of tasks waiting on this futex
104 * @task: the task waiting on the futex
105 * @lock_ptr: the hash bucket lock
106 * @key: the key the futex is hashed on
107 * @pi_state: optional priority inheritance state
108 * @rt_waiter: rt_waiter storage for use with requeue_pi
109 * @requeue_pi_key: the requeue_pi target futex key
110 * @bitset: bitset for the optional bitmasked wakeup
112 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
113 * we can wake only the relevant ones (hashed queues may be shared).
115 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
116 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
117 * The order of wakeup is always to make the first condition true, then
120 * PI futexes are typically woken before they are removed from the hash list via
121 * the rt_mutex code. See unqueue_me_pi().
124 struct plist_node list;
126 struct task_struct *task;
127 spinlock_t *lock_ptr;
129 struct futex_pi_state *pi_state;
130 struct rt_mutex_waiter *rt_waiter;
131 union futex_key *requeue_pi_key;
135 static const struct futex_q futex_q_init = {
136 /* list gets initialized in queue_me()*/
137 .key = FUTEX_KEY_INIT,
138 .bitset = FUTEX_BITSET_MATCH_ANY
142 * Hash buckets are shared by all the futex_keys that hash to the same
143 * location. Each key may have multiple futex_q structures, one for each task
144 * waiting on a futex.
146 struct futex_hash_bucket {
148 struct plist_head chain;
151 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
154 * We hash on the keys returned from get_futex_key (see below).
156 static struct futex_hash_bucket *hash_futex(union futex_key *key)
158 u32 hash = jhash2((u32*)&key->both.word,
159 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
161 return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
165 * Return 1 if two futex_keys are equal, 0 otherwise.
167 static inline int match_futex(union futex_key *key1, union futex_key *key2)
170 && key1->both.word == key2->both.word
171 && key1->both.ptr == key2->both.ptr
172 && key1->both.offset == key2->both.offset);
176 * Take a reference to the resource addressed by a key.
177 * Can be called while holding spinlocks.
180 static void get_futex_key_refs(union futex_key *key)
185 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
187 ihold(key->shared.inode);
189 case FUT_OFF_MMSHARED:
190 atomic_inc(&key->private.mm->mm_count);
196 * Drop a reference to the resource addressed by a key.
197 * The hash bucket spinlock must not be held.
199 static void drop_futex_key_refs(union futex_key *key)
201 if (!key->both.ptr) {
202 /* If we're here then we tried to put a key we failed to get */
207 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
209 iput(key->shared.inode);
211 case FUT_OFF_MMSHARED:
212 mmdrop(key->private.mm);
218 * get_futex_key() - Get parameters which are the keys for a futex
219 * @uaddr: virtual address of the futex
220 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
221 * @key: address where result is stored.
222 * @rw: mapping needs to be read/write (values: VERIFY_READ,
225 * Returns a negative error code or 0
226 * The key words are stored in *key on success.
228 * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
229 * offset_within_page). For private mappings, it's (uaddr, current->mm).
230 * We can usually work out the index without swapping in the page.
232 * lock_page() might sleep, the caller should not hold a spinlock.
235 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
237 unsigned long address = (unsigned long)uaddr;
238 struct mm_struct *mm = current->mm;
239 struct page *page, *page_head;
243 * The futex address must be "naturally" aligned.
245 key->both.offset = address % PAGE_SIZE;
246 if (unlikely((address % sizeof(u32)) != 0))
248 address -= key->both.offset;
251 * PROCESS_PRIVATE futexes are fast.
252 * As the mm cannot disappear under us and the 'key' only needs
253 * virtual address, we dont even have to find the underlying vma.
254 * Note : We do have to check 'uaddr' is a valid user address,
255 * but access_ok() should be faster than find_vma()
258 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
260 key->private.mm = mm;
261 key->private.address = address;
262 get_futex_key_refs(key);
267 err = get_user_pages_fast(address, 1, 1, &page);
269 * If write access is not required (eg. FUTEX_WAIT), try
270 * and get read-only access.
272 if (err == -EFAULT && rw == VERIFY_READ) {
273 err = get_user_pages_fast(address, 1, 0, &page);
281 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
283 if (unlikely(PageTail(page))) {
285 /* serialize against __split_huge_page_splitting() */
287 if (likely(__get_user_pages_fast(address, 1, 1, &page) == 1)) {
288 page_head = compound_head(page);
290 * page_head is valid pointer but we must pin
291 * it before taking the PG_lock and/or
292 * PG_compound_lock. The moment we re-enable
293 * irqs __split_huge_page_splitting() can
294 * return and the head page can be freed from
295 * under us. We can't take the PG_lock and/or
296 * PG_compound_lock on a page that could be
297 * freed from under us.
299 if (page != page_head) {
310 page_head = compound_head(page);
311 if (page != page_head) {
317 lock_page(page_head);
320 * If page_head->mapping is NULL, then it cannot be a PageAnon
321 * page; but it might be the ZERO_PAGE or in the gate area or
322 * in a special mapping (all cases which we are happy to fail);
323 * or it may have been a good file page when get_user_pages_fast
324 * found it, but truncated or holepunched or subjected to
325 * invalidate_complete_page2 before we got the page lock (also
326 * cases which we are happy to fail). And we hold a reference,
327 * so refcount care in invalidate_complete_page's remove_mapping
328 * prevents drop_caches from setting mapping to NULL beneath us.
330 * The case we do have to guard against is when memory pressure made
331 * shmem_writepage move it from filecache to swapcache beneath us:
332 * an unlikely race, but we do need to retry for page_head->mapping.
334 if (!page_head->mapping) {
335 int shmem_swizzled = PageSwapCache(page_head);
336 unlock_page(page_head);
344 * Private mappings are handled in a simple way.
346 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
347 * it's a read-only handle, it's expected that futexes attach to
348 * the object not the particular process.
350 if (PageAnon(page_head)) {
352 * A RO anonymous page will never change and thus doesn't make
353 * sense for futex operations.
360 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
361 key->private.mm = mm;
362 key->private.address = address;
364 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
365 key->shared.inode = page_head->mapping->host;
366 key->shared.pgoff = page_head->index;
369 get_futex_key_refs(key);
372 unlock_page(page_head);
377 static inline void put_futex_key(union futex_key *key)
379 drop_futex_key_refs(key);
383 * fault_in_user_writeable() - Fault in user address and verify RW access
384 * @uaddr: pointer to faulting user space address
386 * Slow path to fixup the fault we just took in the atomic write
389 * We have no generic implementation of a non-destructive write to the
390 * user address. We know that we faulted in the atomic pagefault
391 * disabled section so we can as well avoid the #PF overhead by
392 * calling get_user_pages() right away.
394 static int fault_in_user_writeable(u32 __user *uaddr)
396 struct mm_struct *mm = current->mm;
399 down_read(&mm->mmap_sem);
400 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
402 up_read(&mm->mmap_sem);
404 return ret < 0 ? ret : 0;
408 * futex_top_waiter() - Return the highest priority waiter on a futex
409 * @hb: the hash bucket the futex_q's reside in
410 * @key: the futex key (to distinguish it from other futex futex_q's)
412 * Must be called with the hb lock held.
414 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
415 union futex_key *key)
417 struct futex_q *this;
419 plist_for_each_entry(this, &hb->chain, list) {
420 if (match_futex(&this->key, key))
426 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
427 u32 uval, u32 newval)
432 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
438 static int get_futex_value_locked(u32 *dest, u32 __user *from)
443 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
446 return ret ? -EFAULT : 0;
453 static int refill_pi_state_cache(void)
455 struct futex_pi_state *pi_state;
457 if (likely(current->pi_state_cache))
460 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
465 INIT_LIST_HEAD(&pi_state->list);
466 /* pi_mutex gets initialized later */
467 pi_state->owner = NULL;
468 atomic_set(&pi_state->refcount, 1);
469 pi_state->key = FUTEX_KEY_INIT;
471 current->pi_state_cache = pi_state;
476 static struct futex_pi_state * alloc_pi_state(void)
478 struct futex_pi_state *pi_state = current->pi_state_cache;
481 current->pi_state_cache = NULL;
486 static void free_pi_state(struct futex_pi_state *pi_state)
488 if (!atomic_dec_and_test(&pi_state->refcount))
492 * If pi_state->owner is NULL, the owner is most probably dying
493 * and has cleaned up the pi_state already
495 if (pi_state->owner) {
496 raw_spin_lock_irq(&pi_state->owner->pi_lock);
497 list_del_init(&pi_state->list);
498 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
500 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
503 if (current->pi_state_cache)
507 * pi_state->list is already empty.
508 * clear pi_state->owner.
509 * refcount is at 0 - put it back to 1.
511 pi_state->owner = NULL;
512 atomic_set(&pi_state->refcount, 1);
513 current->pi_state_cache = pi_state;
518 * Look up the task based on what TID userspace gave us.
521 static struct task_struct * futex_find_get_task(pid_t pid)
523 struct task_struct *p;
526 p = find_task_by_vpid(pid);
536 * This task is holding PI mutexes at exit time => bad.
537 * Kernel cleans up PI-state, but userspace is likely hosed.
538 * (Robust-futex cleanup is separate and might save the day for userspace.)
540 void exit_pi_state_list(struct task_struct *curr)
542 struct list_head *next, *head = &curr->pi_state_list;
543 struct futex_pi_state *pi_state;
544 struct futex_hash_bucket *hb;
545 union futex_key key = FUTEX_KEY_INIT;
547 if (!futex_cmpxchg_enabled)
550 * We are a ZOMBIE and nobody can enqueue itself on
551 * pi_state_list anymore, but we have to be careful
552 * versus waiters unqueueing themselves:
554 raw_spin_lock_irq(&curr->pi_lock);
555 while (!list_empty(head)) {
558 pi_state = list_entry(next, struct futex_pi_state, list);
560 hb = hash_futex(&key);
561 raw_spin_unlock_irq(&curr->pi_lock);
563 spin_lock(&hb->lock);
565 raw_spin_lock_irq(&curr->pi_lock);
567 * We dropped the pi-lock, so re-check whether this
568 * task still owns the PI-state:
570 if (head->next != next) {
571 spin_unlock(&hb->lock);
575 WARN_ON(pi_state->owner != curr);
576 WARN_ON(list_empty(&pi_state->list));
577 list_del_init(&pi_state->list);
578 pi_state->owner = NULL;
579 raw_spin_unlock_irq(&curr->pi_lock);
581 rt_mutex_unlock(&pi_state->pi_mutex);
583 spin_unlock(&hb->lock);
585 raw_spin_lock_irq(&curr->pi_lock);
587 raw_spin_unlock_irq(&curr->pi_lock);
591 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
592 union futex_key *key, struct futex_pi_state **ps)
594 struct futex_pi_state *pi_state = NULL;
595 struct futex_q *this, *next;
596 struct plist_head *head;
597 struct task_struct *p;
598 pid_t pid = uval & FUTEX_TID_MASK;
602 plist_for_each_entry_safe(this, next, head, list) {
603 if (match_futex(&this->key, key)) {
605 * Another waiter already exists - bump up
606 * the refcount and return its pi_state:
608 pi_state = this->pi_state;
610 * Userspace might have messed up non-PI and PI futexes
612 if (unlikely(!pi_state))
615 WARN_ON(!atomic_read(&pi_state->refcount));
618 * When pi_state->owner is NULL then the owner died
619 * and another waiter is on the fly. pi_state->owner
620 * is fixed up by the task which acquires
621 * pi_state->rt_mutex.
623 * We do not check for pid == 0 which can happen when
624 * the owner died and robust_list_exit() cleared the
627 if (pid && pi_state->owner) {
629 * Bail out if user space manipulated the
632 if (pid != task_pid_vnr(pi_state->owner))
636 atomic_inc(&pi_state->refcount);
644 * We are the first waiter - try to look up the real owner and attach
645 * the new pi_state to it, but bail out when TID = 0
649 p = futex_find_get_task(pid);
654 * We need to look at the task state flags to figure out,
655 * whether the task is exiting. To protect against the do_exit
656 * change of the task flags, we do this protected by
659 raw_spin_lock_irq(&p->pi_lock);
660 if (unlikely(p->flags & PF_EXITING)) {
662 * The task is on the way out. When PF_EXITPIDONE is
663 * set, we know that the task has finished the
666 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
668 raw_spin_unlock_irq(&p->pi_lock);
673 pi_state = alloc_pi_state();
676 * Initialize the pi_mutex in locked state and make 'p'
679 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
681 /* Store the key for possible exit cleanups: */
682 pi_state->key = *key;
684 WARN_ON(!list_empty(&pi_state->list));
685 list_add(&pi_state->list, &p->pi_state_list);
687 raw_spin_unlock_irq(&p->pi_lock);
697 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
698 * @uaddr: the pi futex user address
699 * @hb: the pi futex hash bucket
700 * @key: the futex key associated with uaddr and hb
701 * @ps: the pi_state pointer where we store the result of the
703 * @task: the task to perform the atomic lock work for. This will
704 * be "current" except in the case of requeue pi.
705 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
709 * 1 - acquired the lock
712 * The hb->lock and futex_key refs shall be held by the caller.
714 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
715 union futex_key *key,
716 struct futex_pi_state **ps,
717 struct task_struct *task, int set_waiters)
719 int lock_taken, ret, force_take = 0;
720 u32 uval, newval, curval, vpid = task_pid_vnr(task);
723 ret = lock_taken = 0;
726 * To avoid races, we attempt to take the lock here again
727 * (by doing a 0 -> TID atomic cmpxchg), while holding all
728 * the locks. It will most likely not succeed.
732 newval |= FUTEX_WAITERS;
734 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, 0, newval)))
740 if ((unlikely((curval & FUTEX_TID_MASK) == vpid)))
744 * Surprise - we got the lock. Just return to userspace:
746 if (unlikely(!curval))
752 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
753 * to wake at the next unlock.
755 newval = curval | FUTEX_WAITERS;
758 * Should we force take the futex? See below.
760 if (unlikely(force_take)) {
762 * Keep the OWNER_DIED and the WAITERS bit and set the
765 newval = (curval & ~FUTEX_TID_MASK) | vpid;
770 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
772 if (unlikely(curval != uval))
776 * We took the lock due to forced take over.
778 if (unlikely(lock_taken))
782 * We dont have the lock. Look up the PI state (or create it if
783 * we are the first waiter):
785 ret = lookup_pi_state(uval, hb, key, ps);
791 * We failed to find an owner for this
792 * futex. So we have no pi_state to block
793 * on. This can happen in two cases:
796 * 2) A stale FUTEX_WAITERS bit
798 * Re-read the futex value.
800 if (get_futex_value_locked(&curval, uaddr))
804 * If the owner died or we have a stale
805 * WAITERS bit the owner TID in the user space
808 if (!(curval & FUTEX_TID_MASK)) {
821 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
822 * @q: The futex_q to unqueue
824 * The q->lock_ptr must not be NULL and must be held by the caller.
826 static void __unqueue_futex(struct futex_q *q)
828 struct futex_hash_bucket *hb;
830 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
831 || WARN_ON(plist_node_empty(&q->list)))
834 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
835 plist_del(&q->list, &hb->chain);
839 * The hash bucket lock must be held when this is called.
840 * Afterwards, the futex_q must not be accessed.
842 static void wake_futex(struct futex_q *q)
844 struct task_struct *p = q->task;
846 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
850 * We set q->lock_ptr = NULL _before_ we wake up the task. If
851 * a non-futex wake up happens on another CPU then the task
852 * might exit and p would dereference a non-existing task
853 * struct. Prevent this by holding a reference on p across the
860 * The waiting task can free the futex_q as soon as
861 * q->lock_ptr = NULL is written, without taking any locks. A
862 * memory barrier is required here to prevent the following
863 * store to lock_ptr from getting ahead of the plist_del.
868 wake_up_state(p, TASK_NORMAL);
872 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
874 struct task_struct *new_owner;
875 struct futex_pi_state *pi_state = this->pi_state;
876 u32 uninitialized_var(curval), newval;
882 * If current does not own the pi_state then the futex is
883 * inconsistent and user space fiddled with the futex value.
885 if (pi_state->owner != current)
888 raw_spin_lock(&pi_state->pi_mutex.wait_lock);
889 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
892 * It is possible that the next waiter (the one that brought
893 * this owner to the kernel) timed out and is no longer
894 * waiting on the lock.
897 new_owner = this->task;
900 * We pass it to the next owner. (The WAITERS bit is always
901 * kept enabled while there is PI state around. We must also
902 * preserve the owner died bit.)
904 if (!(uval & FUTEX_OWNER_DIED)) {
907 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
909 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
911 else if (curval != uval)
914 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
919 raw_spin_lock_irq(&pi_state->owner->pi_lock);
920 WARN_ON(list_empty(&pi_state->list));
921 list_del_init(&pi_state->list);
922 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
924 raw_spin_lock_irq(&new_owner->pi_lock);
925 WARN_ON(!list_empty(&pi_state->list));
926 list_add(&pi_state->list, &new_owner->pi_state_list);
927 pi_state->owner = new_owner;
928 raw_spin_unlock_irq(&new_owner->pi_lock);
930 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
931 rt_mutex_unlock(&pi_state->pi_mutex);
936 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
938 u32 uninitialized_var(oldval);
941 * There is no waiter, so we unlock the futex. The owner died
942 * bit has not to be preserved here. We are the owner:
944 if (cmpxchg_futex_value_locked(&oldval, uaddr, uval, 0))
953 * Express the locking dependencies for lockdep:
956 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
959 spin_lock(&hb1->lock);
961 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
962 } else { /* hb1 > hb2 */
963 spin_lock(&hb2->lock);
964 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
969 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
971 spin_unlock(&hb1->lock);
973 spin_unlock(&hb2->lock);
977 * Wake up waiters matching bitset queued on this futex (uaddr).
980 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
982 struct futex_hash_bucket *hb;
983 struct futex_q *this, *next;
984 struct plist_head *head;
985 union futex_key key = FUTEX_KEY_INIT;
991 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
992 if (unlikely(ret != 0))
995 hb = hash_futex(&key);
996 spin_lock(&hb->lock);
999 plist_for_each_entry_safe(this, next, head, 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 plist_head *head;
1033 struct futex_q *this, *next;
1037 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1038 if (unlikely(ret != 0))
1040 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1041 if (unlikely(ret != 0))
1044 hb1 = hash_futex(&key1);
1045 hb2 = hash_futex(&key2);
1048 double_lock_hb(hb1, hb2);
1049 op_ret = futex_atomic_op_inuser(op, uaddr2);
1050 if (unlikely(op_ret < 0)) {
1052 double_unlock_hb(hb1, hb2);
1056 * we don't get EFAULT from MMU faults if we don't have an MMU,
1057 * but we might get them from range checking
1063 if (unlikely(op_ret != -EFAULT)) {
1068 ret = fault_in_user_writeable(uaddr2);
1072 if (!(flags & FLAGS_SHARED))
1075 put_futex_key(&key2);
1076 put_futex_key(&key1);
1082 plist_for_each_entry_safe(this, next, head, list) {
1083 if (match_futex (&this->key, &key1)) {
1084 if (this->pi_state || this->rt_waiter) {
1089 if (++ret >= nr_wake)
1098 plist_for_each_entry_safe(this, next, head, list) {
1099 if (match_futex (&this->key, &key2)) {
1100 if (this->pi_state || this->rt_waiter) {
1105 if (++op_ret >= nr_wake2)
1113 double_unlock_hb(hb1, hb2);
1115 put_futex_key(&key2);
1117 put_futex_key(&key1);
1123 * requeue_futex() - Requeue a futex_q from one hb to another
1124 * @q: the futex_q to requeue
1125 * @hb1: the source hash_bucket
1126 * @hb2: the target hash_bucket
1127 * @key2: the new key for the requeued futex_q
1130 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1131 struct futex_hash_bucket *hb2, union futex_key *key2)
1135 * If key1 and key2 hash to the same bucket, no need to
1138 if (likely(&hb1->chain != &hb2->chain)) {
1139 plist_del(&q->list, &hb1->chain);
1140 plist_add(&q->list, &hb2->chain);
1141 q->lock_ptr = &hb2->lock;
1143 get_futex_key_refs(key2);
1148 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1150 * @key: the key of the requeue target futex
1151 * @hb: the hash_bucket of the requeue target futex
1153 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1154 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1155 * to the requeue target futex so the waiter can detect the wakeup on the right
1156 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1157 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1158 * to protect access to the pi_state to fixup the owner later. Must be called
1159 * with both q->lock_ptr and hb->lock held.
1162 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1163 struct futex_hash_bucket *hb)
1165 get_futex_key_refs(key);
1170 WARN_ON(!q->rt_waiter);
1171 q->rt_waiter = NULL;
1173 q->lock_ptr = &hb->lock;
1175 wake_up_state(q->task, TASK_NORMAL);
1179 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1180 * @pifutex: the user address of the to futex
1181 * @hb1: the from futex hash bucket, must be locked by the caller
1182 * @hb2: the to futex hash bucket, must be locked by the caller
1183 * @key1: the from futex key
1184 * @key2: the to futex key
1185 * @ps: address to store the pi_state pointer
1186 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1188 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1189 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1190 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1191 * hb1 and hb2 must be held by the caller.
1194 * 0 - failed to acquire the lock atomicly
1195 * 1 - acquired the lock
1198 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1199 struct futex_hash_bucket *hb1,
1200 struct futex_hash_bucket *hb2,
1201 union futex_key *key1, union futex_key *key2,
1202 struct futex_pi_state **ps, int set_waiters)
1204 struct futex_q *top_waiter = NULL;
1208 if (get_futex_value_locked(&curval, pifutex))
1212 * Find the top_waiter and determine if there are additional waiters.
1213 * If the caller intends to requeue more than 1 waiter to pifutex,
1214 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1215 * as we have means to handle the possible fault. If not, don't set
1216 * the bit unecessarily as it will force the subsequent unlock to enter
1219 top_waiter = futex_top_waiter(hb1, key1);
1221 /* There are no waiters, nothing for us to do. */
1225 /* Ensure we requeue to the expected futex. */
1226 if (!match_futex(top_waiter->requeue_pi_key, key2))
1230 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1231 * the contended case or if set_waiters is 1. The pi_state is returned
1232 * in ps in contended cases.
1234 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1237 requeue_pi_wake_futex(top_waiter, key2, hb2);
1243 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1244 * @uaddr1: source futex user address
1245 * @flags: futex flags (FLAGS_SHARED, etc.)
1246 * @uaddr2: target futex user address
1247 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1248 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1249 * @cmpval: @uaddr1 expected value (or %NULL)
1250 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1251 * pi futex (pi to pi requeue is not supported)
1253 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1254 * uaddr2 atomically on behalf of the top waiter.
1257 * >=0 - on success, the number of tasks requeued or woken
1260 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1261 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1262 u32 *cmpval, int requeue_pi)
1264 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1265 int drop_count = 0, task_count = 0, ret;
1266 struct futex_pi_state *pi_state = NULL;
1267 struct futex_hash_bucket *hb1, *hb2;
1268 struct plist_head *head1;
1269 struct futex_q *this, *next;
1274 * requeue_pi requires a pi_state, try to allocate it now
1275 * without any locks in case it fails.
1277 if (refill_pi_state_cache())
1280 * requeue_pi must wake as many tasks as it can, up to nr_wake
1281 * + nr_requeue, since it acquires the rt_mutex prior to
1282 * returning to userspace, so as to not leave the rt_mutex with
1283 * waiters and no owner. However, second and third wake-ups
1284 * cannot be predicted as they involve race conditions with the
1285 * first wake and a fault while looking up the pi_state. Both
1286 * pthread_cond_signal() and pthread_cond_broadcast() should
1294 if (pi_state != NULL) {
1296 * We will have to lookup the pi_state again, so free this one
1297 * to keep the accounting correct.
1299 free_pi_state(pi_state);
1303 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1304 if (unlikely(ret != 0))
1306 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1307 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1308 if (unlikely(ret != 0))
1311 hb1 = hash_futex(&key1);
1312 hb2 = hash_futex(&key2);
1315 double_lock_hb(hb1, hb2);
1317 if (likely(cmpval != NULL)) {
1320 ret = get_futex_value_locked(&curval, uaddr1);
1322 if (unlikely(ret)) {
1323 double_unlock_hb(hb1, hb2);
1325 ret = get_user(curval, uaddr1);
1329 if (!(flags & FLAGS_SHARED))
1332 put_futex_key(&key2);
1333 put_futex_key(&key1);
1336 if (curval != *cmpval) {
1342 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1344 * Attempt to acquire uaddr2 and wake the top waiter. If we
1345 * intend to requeue waiters, force setting the FUTEX_WAITERS
1346 * bit. We force this here where we are able to easily handle
1347 * faults rather in the requeue loop below.
1349 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1350 &key2, &pi_state, nr_requeue);
1353 * At this point the top_waiter has either taken uaddr2 or is
1354 * waiting on it. If the former, then the pi_state will not
1355 * exist yet, look it up one more time to ensure we have a
1362 ret = get_futex_value_locked(&curval2, uaddr2);
1364 ret = lookup_pi_state(curval2, hb2, &key2,
1372 double_unlock_hb(hb1, hb2);
1373 put_futex_key(&key2);
1374 put_futex_key(&key1);
1375 ret = fault_in_user_writeable(uaddr2);
1380 /* The owner was exiting, try again. */
1381 double_unlock_hb(hb1, hb2);
1382 put_futex_key(&key2);
1383 put_futex_key(&key1);
1391 head1 = &hb1->chain;
1392 plist_for_each_entry_safe(this, next, head1, list) {
1393 if (task_count - nr_wake >= nr_requeue)
1396 if (!match_futex(&this->key, &key1))
1400 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1401 * be paired with each other and no other futex ops.
1403 * We should never be requeueing a futex_q with a pi_state,
1404 * which is awaiting a futex_unlock_pi().
1406 if ((requeue_pi && !this->rt_waiter) ||
1407 (!requeue_pi && this->rt_waiter) ||
1414 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1415 * lock, we already woke the top_waiter. If not, it will be
1416 * woken by futex_unlock_pi().
1418 if (++task_count <= nr_wake && !requeue_pi) {
1423 /* Ensure we requeue to the expected futex for requeue_pi. */
1424 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1430 * Requeue nr_requeue waiters and possibly one more in the case
1431 * of requeue_pi if we couldn't acquire the lock atomically.
1434 /* Prepare the waiter to take the rt_mutex. */
1435 atomic_inc(&pi_state->refcount);
1436 this->pi_state = pi_state;
1437 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1441 /* We got the lock. */
1442 requeue_pi_wake_futex(this, &key2, hb2);
1447 this->pi_state = NULL;
1448 free_pi_state(pi_state);
1452 requeue_futex(this, hb1, hb2, &key2);
1457 double_unlock_hb(hb1, hb2);
1460 * drop_futex_key_refs() must be called outside the spinlocks. During
1461 * the requeue we moved futex_q's from the hash bucket at key1 to the
1462 * one at key2 and updated their key pointer. We no longer need to
1463 * hold the references to key1.
1465 while (--drop_count >= 0)
1466 drop_futex_key_refs(&key1);
1469 put_futex_key(&key2);
1471 put_futex_key(&key1);
1473 if (pi_state != NULL)
1474 free_pi_state(pi_state);
1475 return ret ? ret : task_count;
1478 /* The key must be already stored in q->key. */
1479 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1480 __acquires(&hb->lock)
1482 struct futex_hash_bucket *hb;
1484 hb = hash_futex(&q->key);
1485 q->lock_ptr = &hb->lock;
1487 spin_lock(&hb->lock);
1492 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1493 __releases(&hb->lock)
1495 spin_unlock(&hb->lock);
1499 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1500 * @q: The futex_q to enqueue
1501 * @hb: The destination hash bucket
1503 * The hb->lock must be held by the caller, and is released here. A call to
1504 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1505 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1506 * or nothing if the unqueue is done as part of the wake process and the unqueue
1507 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1510 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1511 __releases(&hb->lock)
1516 * The priority used to register this element is
1517 * - either the real thread-priority for the real-time threads
1518 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1519 * - or MAX_RT_PRIO for non-RT threads.
1520 * Thus, all RT-threads are woken first in priority order, and
1521 * the others are woken last, in FIFO order.
1523 prio = min(current->normal_prio, MAX_RT_PRIO);
1525 plist_node_init(&q->list, prio);
1526 plist_add(&q->list, &hb->chain);
1528 spin_unlock(&hb->lock);
1532 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1533 * @q: The futex_q to unqueue
1535 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1536 * be paired with exactly one earlier call to queue_me().
1539 * 1 - if the futex_q was still queued (and we removed unqueued it)
1540 * 0 - if the futex_q was already removed by the waking thread
1542 static int unqueue_me(struct futex_q *q)
1544 spinlock_t *lock_ptr;
1547 /* In the common case we don't take the spinlock, which is nice. */
1549 lock_ptr = q->lock_ptr;
1551 if (lock_ptr != NULL) {
1552 spin_lock(lock_ptr);
1554 * q->lock_ptr can change between reading it and
1555 * spin_lock(), causing us to take the wrong lock. This
1556 * corrects the race condition.
1558 * Reasoning goes like this: if we have the wrong lock,
1559 * q->lock_ptr must have changed (maybe several times)
1560 * between reading it and the spin_lock(). It can
1561 * change again after the spin_lock() but only if it was
1562 * already changed before the spin_lock(). It cannot,
1563 * however, change back to the original value. Therefore
1564 * we can detect whether we acquired the correct lock.
1566 if (unlikely(lock_ptr != q->lock_ptr)) {
1567 spin_unlock(lock_ptr);
1572 BUG_ON(q->pi_state);
1574 spin_unlock(lock_ptr);
1578 drop_futex_key_refs(&q->key);
1583 * PI futexes can not be requeued and must remove themself from the
1584 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1587 static void unqueue_me_pi(struct futex_q *q)
1588 __releases(q->lock_ptr)
1592 BUG_ON(!q->pi_state);
1593 free_pi_state(q->pi_state);
1596 spin_unlock(q->lock_ptr);
1600 * Fixup the pi_state owner with the new owner.
1602 * Must be called with hash bucket lock held and mm->sem held for non
1605 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1606 struct task_struct *newowner)
1608 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1609 struct futex_pi_state *pi_state = q->pi_state;
1610 struct task_struct *oldowner = pi_state->owner;
1611 u32 uval, uninitialized_var(curval), newval;
1615 if (!pi_state->owner)
1616 newtid |= FUTEX_OWNER_DIED;
1619 * We are here either because we stole the rtmutex from the
1620 * previous highest priority waiter or we are the highest priority
1621 * waiter but failed to get the rtmutex the first time.
1622 * We have to replace the newowner TID in the user space variable.
1623 * This must be atomic as we have to preserve the owner died bit here.
1625 * Note: We write the user space value _before_ changing the pi_state
1626 * because we can fault here. Imagine swapped out pages or a fork
1627 * that marked all the anonymous memory readonly for cow.
1629 * Modifying pi_state _before_ the user space value would
1630 * leave the pi_state in an inconsistent state when we fault
1631 * here, because we need to drop the hash bucket lock to
1632 * handle the fault. This might be observed in the PID check
1633 * in lookup_pi_state.
1636 if (get_futex_value_locked(&uval, uaddr))
1640 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1642 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1650 * We fixed up user space. Now we need to fix the pi_state
1653 if (pi_state->owner != NULL) {
1654 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1655 WARN_ON(list_empty(&pi_state->list));
1656 list_del_init(&pi_state->list);
1657 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1660 pi_state->owner = newowner;
1662 raw_spin_lock_irq(&newowner->pi_lock);
1663 WARN_ON(!list_empty(&pi_state->list));
1664 list_add(&pi_state->list, &newowner->pi_state_list);
1665 raw_spin_unlock_irq(&newowner->pi_lock);
1669 * To handle the page fault we need to drop the hash bucket
1670 * lock here. That gives the other task (either the highest priority
1671 * waiter itself or the task which stole the rtmutex) the
1672 * chance to try the fixup of the pi_state. So once we are
1673 * back from handling the fault we need to check the pi_state
1674 * after reacquiring the hash bucket lock and before trying to
1675 * do another fixup. When the fixup has been done already we
1679 spin_unlock(q->lock_ptr);
1681 ret = fault_in_user_writeable(uaddr);
1683 spin_lock(q->lock_ptr);
1686 * Check if someone else fixed it for us:
1688 if (pi_state->owner != oldowner)
1697 static long futex_wait_restart(struct restart_block *restart);
1700 * fixup_owner() - Post lock pi_state and corner case management
1701 * @uaddr: user address of the futex
1702 * @q: futex_q (contains pi_state and access to the rt_mutex)
1703 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1705 * After attempting to lock an rt_mutex, this function is called to cleanup
1706 * the pi_state owner as well as handle race conditions that may allow us to
1707 * acquire the lock. Must be called with the hb lock held.
1710 * 1 - success, lock taken
1711 * 0 - success, lock not taken
1712 * <0 - on error (-EFAULT)
1714 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
1716 struct task_struct *owner;
1721 * Got the lock. We might not be the anticipated owner if we
1722 * did a lock-steal - fix up the PI-state in that case:
1724 if (q->pi_state->owner != current)
1725 ret = fixup_pi_state_owner(uaddr, q, current);
1730 * Catch the rare case, where the lock was released when we were on the
1731 * way back before we locked the hash bucket.
1733 if (q->pi_state->owner == current) {
1735 * Try to get the rt_mutex now. This might fail as some other
1736 * task acquired the rt_mutex after we removed ourself from the
1737 * rt_mutex waiters list.
1739 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1745 * pi_state is incorrect, some other task did a lock steal and
1746 * we returned due to timeout or signal without taking the
1747 * rt_mutex. Too late.
1749 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
1750 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1752 owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
1753 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
1754 ret = fixup_pi_state_owner(uaddr, q, owner);
1759 * Paranoia check. If we did not take the lock, then we should not be
1760 * the owner of the rt_mutex.
1762 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1763 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1764 "pi-state %p\n", ret,
1765 q->pi_state->pi_mutex.owner,
1766 q->pi_state->owner);
1769 return ret ? ret : locked;
1773 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1774 * @hb: the futex hash bucket, must be locked by the caller
1775 * @q: the futex_q to queue up on
1776 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1778 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1779 struct hrtimer_sleeper *timeout)
1782 * The task state is guaranteed to be set before another task can
1783 * wake it. set_current_state() is implemented using set_mb() and
1784 * queue_me() calls spin_unlock() upon completion, both serializing
1785 * access to the hash list and forcing another memory barrier.
1787 set_current_state(TASK_INTERRUPTIBLE);
1792 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1793 if (!hrtimer_active(&timeout->timer))
1794 timeout->task = NULL;
1798 * If we have been removed from the hash list, then another task
1799 * has tried to wake us, and we can skip the call to schedule().
1801 if (likely(!plist_node_empty(&q->list))) {
1803 * If the timer has already expired, current will already be
1804 * flagged for rescheduling. Only call schedule if there
1805 * is no timeout, or if it has yet to expire.
1807 if (!timeout || timeout->task)
1810 __set_current_state(TASK_RUNNING);
1814 * futex_wait_setup() - Prepare to wait on a futex
1815 * @uaddr: the futex userspace address
1816 * @val: the expected value
1817 * @flags: futex flags (FLAGS_SHARED, etc.)
1818 * @q: the associated futex_q
1819 * @hb: storage for hash_bucket pointer to be returned to caller
1821 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1822 * compare it with the expected value. Handle atomic faults internally.
1823 * Return with the hb lock held and a q.key reference on success, and unlocked
1824 * with no q.key reference on failure.
1827 * 0 - uaddr contains val and hb has been locked
1828 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
1830 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
1831 struct futex_q *q, struct futex_hash_bucket **hb)
1837 * Access the page AFTER the hash-bucket is locked.
1838 * Order is important:
1840 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1841 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1843 * The basic logical guarantee of a futex is that it blocks ONLY
1844 * if cond(var) is known to be true at the time of blocking, for
1845 * any cond. If we locked the hash-bucket after testing *uaddr, that
1846 * would open a race condition where we could block indefinitely with
1847 * cond(var) false, which would violate the guarantee.
1849 * On the other hand, we insert q and release the hash-bucket only
1850 * after testing *uaddr. This guarantees that futex_wait() will NOT
1851 * absorb a wakeup if *uaddr does not match the desired values
1852 * while the syscall executes.
1855 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
1856 if (unlikely(ret != 0))
1860 *hb = queue_lock(q);
1862 ret = get_futex_value_locked(&uval, uaddr);
1865 queue_unlock(q, *hb);
1867 ret = get_user(uval, uaddr);
1871 if (!(flags & FLAGS_SHARED))
1874 put_futex_key(&q->key);
1879 queue_unlock(q, *hb);
1885 put_futex_key(&q->key);
1889 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
1890 ktime_t *abs_time, u32 bitset)
1892 struct hrtimer_sleeper timeout, *to = NULL;
1893 struct restart_block *restart;
1894 struct futex_hash_bucket *hb;
1895 struct futex_q q = futex_q_init;
1905 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
1906 CLOCK_REALTIME : CLOCK_MONOTONIC,
1908 hrtimer_init_sleeper(to, current);
1909 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1910 current->timer_slack_ns);
1915 * Prepare to wait on uaddr. On success, holds hb lock and increments
1918 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
1922 /* queue_me and wait for wakeup, timeout, or a signal. */
1923 futex_wait_queue_me(hb, &q, to);
1925 /* If we were woken (and unqueued), we succeeded, whatever. */
1927 /* unqueue_me() drops q.key ref */
1928 if (!unqueue_me(&q))
1931 if (to && !to->task)
1935 * We expect signal_pending(current), but we might be the
1936 * victim of a spurious wakeup as well.
1938 if (!signal_pending(current))
1945 restart = ¤t_thread_info()->restart_block;
1946 restart->fn = futex_wait_restart;
1947 restart->futex.uaddr = uaddr;
1948 restart->futex.val = val;
1949 restart->futex.time = abs_time->tv64;
1950 restart->futex.bitset = bitset;
1951 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
1953 ret = -ERESTART_RESTARTBLOCK;
1957 hrtimer_cancel(&to->timer);
1958 destroy_hrtimer_on_stack(&to->timer);
1964 static long futex_wait_restart(struct restart_block *restart)
1966 u32 __user *uaddr = restart->futex.uaddr;
1967 ktime_t t, *tp = NULL;
1969 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1970 t.tv64 = restart->futex.time;
1973 restart->fn = do_no_restart_syscall;
1975 return (long)futex_wait(uaddr, restart->futex.flags,
1976 restart->futex.val, tp, restart->futex.bitset);
1981 * Userspace tried a 0 -> TID atomic transition of the futex value
1982 * and failed. The kernel side here does the whole locking operation:
1983 * if there are waiters then it will block, it does PI, etc. (Due to
1984 * races the kernel might see a 0 value of the futex too.)
1986 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
1987 ktime_t *time, int trylock)
1989 struct hrtimer_sleeper timeout, *to = NULL;
1990 struct futex_hash_bucket *hb;
1991 struct futex_q q = futex_q_init;
1994 if (refill_pi_state_cache())
1999 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2001 hrtimer_init_sleeper(to, current);
2002 hrtimer_set_expires(&to->timer, *time);
2006 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2007 if (unlikely(ret != 0))
2011 hb = queue_lock(&q);
2013 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2014 if (unlikely(ret)) {
2017 /* We got the lock. */
2019 goto out_unlock_put_key;
2024 * Task is exiting and we just wait for the
2027 queue_unlock(&q, hb);
2028 put_futex_key(&q.key);
2032 goto out_unlock_put_key;
2037 * Only actually queue now that the atomic ops are done:
2041 WARN_ON(!q.pi_state);
2043 * Block on the PI mutex:
2046 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
2048 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2049 /* Fixup the trylock return value: */
2050 ret = ret ? 0 : -EWOULDBLOCK;
2053 spin_lock(q.lock_ptr);
2055 * Fixup the pi_state owner and possibly acquire the lock if we
2058 res = fixup_owner(uaddr, &q, !ret);
2060 * If fixup_owner() returned an error, proprogate that. If it acquired
2061 * the lock, clear our -ETIMEDOUT or -EINTR.
2064 ret = (res < 0) ? res : 0;
2067 * If fixup_owner() faulted and was unable to handle the fault, unlock
2068 * it and return the fault to userspace.
2070 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2071 rt_mutex_unlock(&q.pi_state->pi_mutex);
2073 /* Unqueue and drop the lock */
2079 queue_unlock(&q, hb);
2082 put_futex_key(&q.key);
2085 destroy_hrtimer_on_stack(&to->timer);
2086 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2089 queue_unlock(&q, hb);
2091 ret = fault_in_user_writeable(uaddr);
2095 if (!(flags & FLAGS_SHARED))
2098 put_futex_key(&q.key);
2103 * Userspace attempted a TID -> 0 atomic transition, and failed.
2104 * This is the in-kernel slowpath: we look up the PI state (if any),
2105 * and do the rt-mutex unlock.
2107 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2109 struct futex_hash_bucket *hb;
2110 struct futex_q *this, *next;
2111 struct plist_head *head;
2112 union futex_key key = FUTEX_KEY_INIT;
2113 u32 uval, vpid = task_pid_vnr(current);
2117 if (get_user(uval, uaddr))
2120 * We release only a lock we actually own:
2122 if ((uval & FUTEX_TID_MASK) != vpid)
2125 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2126 if (unlikely(ret != 0))
2129 hb = hash_futex(&key);
2130 spin_lock(&hb->lock);
2133 * To avoid races, try to do the TID -> 0 atomic transition
2134 * again. If it succeeds then we can return without waking
2137 if (!(uval & FUTEX_OWNER_DIED) &&
2138 cmpxchg_futex_value_locked(&uval, uaddr, vpid, 0))
2141 * Rare case: we managed to release the lock atomically,
2142 * no need to wake anyone else up:
2144 if (unlikely(uval == vpid))
2148 * Ok, other tasks may need to be woken up - check waiters
2149 * and do the wakeup if necessary:
2153 plist_for_each_entry_safe(this, next, head, list) {
2154 if (!match_futex (&this->key, &key))
2156 ret = wake_futex_pi(uaddr, uval, this);
2158 * The atomic access to the futex value
2159 * generated a pagefault, so retry the
2160 * user-access and the wakeup:
2167 * No waiters - kernel unlocks the futex:
2169 if (!(uval & FUTEX_OWNER_DIED)) {
2170 ret = unlock_futex_pi(uaddr, uval);
2176 spin_unlock(&hb->lock);
2177 put_futex_key(&key);
2183 spin_unlock(&hb->lock);
2184 put_futex_key(&key);
2186 ret = fault_in_user_writeable(uaddr);
2194 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2195 * @hb: the hash_bucket futex_q was original enqueued on
2196 * @q: the futex_q woken while waiting to be requeued
2197 * @key2: the futex_key of the requeue target futex
2198 * @timeout: the timeout associated with the wait (NULL if none)
2200 * Detect if the task was woken on the initial futex as opposed to the requeue
2201 * target futex. If so, determine if it was a timeout or a signal that caused
2202 * the wakeup and return the appropriate error code to the caller. Must be
2203 * called with the hb lock held.
2206 * 0 - no early wakeup detected
2207 * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2210 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2211 struct futex_q *q, union futex_key *key2,
2212 struct hrtimer_sleeper *timeout)
2217 * With the hb lock held, we avoid races while we process the wakeup.
2218 * We only need to hold hb (and not hb2) to ensure atomicity as the
2219 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2220 * It can't be requeued from uaddr2 to something else since we don't
2221 * support a PI aware source futex for requeue.
2223 if (!match_futex(&q->key, key2)) {
2224 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2226 * We were woken prior to requeue by a timeout or a signal.
2227 * Unqueue the futex_q and determine which it was.
2229 plist_del(&q->list, &hb->chain);
2231 /* Handle spurious wakeups gracefully */
2233 if (timeout && !timeout->task)
2235 else if (signal_pending(current))
2236 ret = -ERESTARTNOINTR;
2242 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2243 * @uaddr: the futex we initially wait on (non-pi)
2244 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2245 * the same type, no requeueing from private to shared, etc.
2246 * @val: the expected value of uaddr
2247 * @abs_time: absolute timeout
2248 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2249 * @clockrt: whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2250 * @uaddr2: the pi futex we will take prior to returning to user-space
2252 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2253 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2254 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2255 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2256 * without one, the pi logic would not know which task to boost/deboost, if
2257 * there was a need to.
2259 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2260 * via the following:
2261 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2262 * 2) wakeup on uaddr2 after a requeue
2266 * If 3, cleanup and return -ERESTARTNOINTR.
2268 * If 2, we may then block on trying to take the rt_mutex and return via:
2269 * 5) successful lock
2272 * 8) other lock acquisition failure
2274 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2276 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2282 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2283 u32 val, ktime_t *abs_time, u32 bitset,
2286 struct hrtimer_sleeper timeout, *to = NULL;
2287 struct rt_mutex_waiter rt_waiter;
2288 struct rt_mutex *pi_mutex = NULL;
2289 struct futex_hash_bucket *hb;
2290 union futex_key key2 = FUTEX_KEY_INIT;
2291 struct futex_q q = futex_q_init;
2294 if (uaddr == uaddr2)
2302 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2303 CLOCK_REALTIME : CLOCK_MONOTONIC,
2305 hrtimer_init_sleeper(to, current);
2306 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2307 current->timer_slack_ns);
2311 * The waiter is allocated on our stack, manipulated by the requeue
2312 * code while we sleep on uaddr.
2314 debug_rt_mutex_init_waiter(&rt_waiter);
2315 rt_waiter.task = NULL;
2317 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2318 if (unlikely(ret != 0))
2322 q.rt_waiter = &rt_waiter;
2323 q.requeue_pi_key = &key2;
2326 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2329 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2333 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2334 futex_wait_queue_me(hb, &q, to);
2336 spin_lock(&hb->lock);
2337 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2338 spin_unlock(&hb->lock);
2343 * In order for us to be here, we know our q.key == key2, and since
2344 * we took the hb->lock above, we also know that futex_requeue() has
2345 * completed and we no longer have to concern ourselves with a wakeup
2346 * race with the atomic proxy lock acquisition by the requeue code. The
2347 * futex_requeue dropped our key1 reference and incremented our key2
2351 /* Check if the requeue code acquired the second futex for us. */
2354 * Got the lock. We might not be the anticipated owner if we
2355 * did a lock-steal - fix up the PI-state in that case.
2357 if (q.pi_state && (q.pi_state->owner != current)) {
2358 spin_lock(q.lock_ptr);
2359 ret = fixup_pi_state_owner(uaddr2, &q, current);
2360 spin_unlock(q.lock_ptr);
2364 * We have been woken up by futex_unlock_pi(), a timeout, or a
2365 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2368 WARN_ON(!q.pi_state);
2369 pi_mutex = &q.pi_state->pi_mutex;
2370 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2371 debug_rt_mutex_free_waiter(&rt_waiter);
2373 spin_lock(q.lock_ptr);
2375 * Fixup the pi_state owner and possibly acquire the lock if we
2378 res = fixup_owner(uaddr2, &q, !ret);
2380 * If fixup_owner() returned an error, proprogate that. If it
2381 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2384 ret = (res < 0) ? res : 0;
2386 /* Unqueue and drop the lock. */
2391 * If fixup_pi_state_owner() faulted and was unable to handle the
2392 * fault, unlock the rt_mutex and return the fault to userspace.
2394 if (ret == -EFAULT) {
2395 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2396 rt_mutex_unlock(pi_mutex);
2397 } else if (ret == -EINTR) {
2399 * We've already been requeued, but cannot restart by calling
2400 * futex_lock_pi() directly. We could restart this syscall, but
2401 * it would detect that the user space "val" changed and return
2402 * -EWOULDBLOCK. Save the overhead of the restart and return
2403 * -EWOULDBLOCK directly.
2409 put_futex_key(&q.key);
2411 put_futex_key(&key2);
2415 hrtimer_cancel(&to->timer);
2416 destroy_hrtimer_on_stack(&to->timer);
2422 * Support for robust futexes: the kernel cleans up held futexes at
2425 * Implementation: user-space maintains a per-thread list of locks it
2426 * is holding. Upon do_exit(), the kernel carefully walks this list,
2427 * and marks all locks that are owned by this thread with the
2428 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2429 * always manipulated with the lock held, so the list is private and
2430 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2431 * field, to allow the kernel to clean up if the thread dies after
2432 * acquiring the lock, but just before it could have added itself to
2433 * the list. There can only be one such pending lock.
2437 * sys_set_robust_list() - Set the robust-futex list head of a task
2438 * @head: pointer to the list-head
2439 * @len: length of the list-head, as userspace expects
2441 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2444 if (!futex_cmpxchg_enabled)
2447 * The kernel knows only one size for now:
2449 if (unlikely(len != sizeof(*head)))
2452 current->robust_list = head;
2458 * sys_get_robust_list() - Get the robust-futex list head of a task
2459 * @pid: pid of the process [zero for current task]
2460 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2461 * @len_ptr: pointer to a length field, the kernel fills in the header size
2463 SYSCALL_DEFINE3(get_robust_list, int, pid,
2464 struct robust_list_head __user * __user *, head_ptr,
2465 size_t __user *, len_ptr)
2467 struct robust_list_head __user *head;
2469 struct task_struct *p;
2471 if (!futex_cmpxchg_enabled)
2474 WARN_ONCE(1, "deprecated: get_robust_list will be deleted in 2013.\n");
2482 p = find_task_by_vpid(pid);
2488 if (!ptrace_may_access(p, PTRACE_MODE_READ))
2491 head = p->robust_list;
2494 if (put_user(sizeof(*head), len_ptr))
2496 return put_user(head, head_ptr);
2505 * Process a futex-list entry, check whether it's owned by the
2506 * dying task, and do notification if so:
2508 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2510 u32 uval, uninitialized_var(nval), mval;
2513 if (get_user(uval, uaddr))
2516 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2518 * Ok, this dying thread is truly holding a futex
2519 * of interest. Set the OWNER_DIED bit atomically
2520 * via cmpxchg, and if the value had FUTEX_WAITERS
2521 * set, wake up a waiter (if any). (We have to do a
2522 * futex_wake() even if OWNER_DIED is already set -
2523 * to handle the rare but possible case of recursive
2524 * thread-death.) The rest of the cleanup is done in
2527 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2529 * We are not holding a lock here, but we want to have
2530 * the pagefault_disable/enable() protection because
2531 * we want to handle the fault gracefully. If the
2532 * access fails we try to fault in the futex with R/W
2533 * verification via get_user_pages. get_user() above
2534 * does not guarantee R/W access. If that fails we
2535 * give up and leave the futex locked.
2537 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2538 if (fault_in_user_writeable(uaddr))
2546 * Wake robust non-PI futexes here. The wakeup of
2547 * PI futexes happens in exit_pi_state():
2549 if (!pi && (uval & FUTEX_WAITERS))
2550 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2556 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2558 static inline int fetch_robust_entry(struct robust_list __user **entry,
2559 struct robust_list __user * __user *head,
2562 unsigned long uentry;
2564 if (get_user(uentry, (unsigned long __user *)head))
2567 *entry = (void __user *)(uentry & ~1UL);
2574 * Walk curr->robust_list (very carefully, it's a userspace list!)
2575 * and mark any locks found there dead, and notify any waiters.
2577 * We silently return on any sign of list-walking problem.
2579 void exit_robust_list(struct task_struct *curr)
2581 struct robust_list_head __user *head = curr->robust_list;
2582 struct robust_list __user *entry, *next_entry, *pending;
2583 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2584 unsigned int uninitialized_var(next_pi);
2585 unsigned long futex_offset;
2588 if (!futex_cmpxchg_enabled)
2592 * Fetch the list head (which was registered earlier, via
2593 * sys_set_robust_list()):
2595 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2598 * Fetch the relative futex offset:
2600 if (get_user(futex_offset, &head->futex_offset))
2603 * Fetch any possibly pending lock-add first, and handle it
2606 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2609 next_entry = NULL; /* avoid warning with gcc */
2610 while (entry != &head->list) {
2612 * Fetch the next entry in the list before calling
2613 * handle_futex_death:
2615 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2617 * A pending lock might already be on the list, so
2618 * don't process it twice:
2620 if (entry != pending)
2621 if (handle_futex_death((void __user *)entry + futex_offset,
2629 * Avoid excessively long or circular lists:
2638 handle_futex_death((void __user *)pending + futex_offset,
2642 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2643 u32 __user *uaddr2, u32 val2, u32 val3)
2645 int cmd = op & FUTEX_CMD_MASK;
2646 unsigned int flags = 0;
2648 if (!(op & FUTEX_PRIVATE_FLAG))
2649 flags |= FLAGS_SHARED;
2651 if (op & FUTEX_CLOCK_REALTIME) {
2652 flags |= FLAGS_CLOCKRT;
2653 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2659 case FUTEX_UNLOCK_PI:
2660 case FUTEX_TRYLOCK_PI:
2661 case FUTEX_WAIT_REQUEUE_PI:
2662 case FUTEX_CMP_REQUEUE_PI:
2663 if (!futex_cmpxchg_enabled)
2669 val3 = FUTEX_BITSET_MATCH_ANY;
2670 case FUTEX_WAIT_BITSET:
2671 return futex_wait(uaddr, flags, val, timeout, val3);
2673 val3 = FUTEX_BITSET_MATCH_ANY;
2674 case FUTEX_WAKE_BITSET:
2675 return futex_wake(uaddr, flags, val, val3);
2677 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2678 case FUTEX_CMP_REQUEUE:
2679 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2681 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2683 return futex_lock_pi(uaddr, flags, val, timeout, 0);
2684 case FUTEX_UNLOCK_PI:
2685 return futex_unlock_pi(uaddr, flags);
2686 case FUTEX_TRYLOCK_PI:
2687 return futex_lock_pi(uaddr, flags, 0, timeout, 1);
2688 case FUTEX_WAIT_REQUEUE_PI:
2689 val3 = FUTEX_BITSET_MATCH_ANY;
2690 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2692 case FUTEX_CMP_REQUEUE_PI:
2693 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2699 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2700 struct timespec __user *, utime, u32 __user *, uaddr2,
2704 ktime_t t, *tp = NULL;
2706 int cmd = op & FUTEX_CMD_MASK;
2708 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2709 cmd == FUTEX_WAIT_BITSET ||
2710 cmd == FUTEX_WAIT_REQUEUE_PI)) {
2711 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2713 if (!timespec_valid(&ts))
2716 t = timespec_to_ktime(ts);
2717 if (cmd == FUTEX_WAIT)
2718 t = ktime_add_safe(ktime_get(), t);
2722 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2723 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2725 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2726 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2727 val2 = (u32) (unsigned long) utime;
2729 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2732 static int __init futex_init(void)
2738 * This will fail and we want it. Some arch implementations do
2739 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2740 * functionality. We want to know that before we call in any
2741 * of the complex code paths. Also we want to prevent
2742 * registration of robust lists in that case. NULL is
2743 * guaranteed to fault and we get -EFAULT on functional
2744 * implementation, the non-functional ones will return
2747 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
2748 futex_cmpxchg_enabled = 1;
2750 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2751 plist_head_init(&futex_queues[i].chain);
2752 spin_lock_init(&futex_queues[i].lock);
2757 __initcall(futex_init);