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/hugetlb.h>
65 #include <asm/futex.h>
67 #include "rtmutex_common.h"
69 int __read_mostly futex_cmpxchg_enabled;
71 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
74 * Futex flags used to encode options to functions and preserve them across
77 #define FLAGS_SHARED 0x01
78 #define FLAGS_CLOCKRT 0x02
79 #define FLAGS_HAS_TIMEOUT 0x04
82 * Priority Inheritance state:
84 struct futex_pi_state {
86 * list of 'owned' pi_state instances - these have to be
87 * cleaned up in do_exit() if the task exits prematurely:
89 struct list_head list;
94 struct rt_mutex pi_mutex;
96 struct task_struct *owner;
103 * struct futex_q - The hashed futex queue entry, one per waiting task
104 * @list: priority-sorted list of tasks waiting on this futex
105 * @task: the task waiting on the futex
106 * @lock_ptr: the hash bucket lock
107 * @key: the key the futex is hashed on
108 * @pi_state: optional priority inheritance state
109 * @rt_waiter: rt_waiter storage for use with requeue_pi
110 * @requeue_pi_key: the requeue_pi target futex key
111 * @bitset: bitset for the optional bitmasked wakeup
113 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
114 * we can wake only the relevant ones (hashed queues may be shared).
116 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
117 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
118 * The order of wakeup is always to make the first condition true, then
121 * PI futexes are typically woken before they are removed from the hash list via
122 * the rt_mutex code. See unqueue_me_pi().
125 struct plist_node list;
127 struct task_struct *task;
128 spinlock_t *lock_ptr;
130 struct futex_pi_state *pi_state;
131 struct rt_mutex_waiter *rt_waiter;
132 union futex_key *requeue_pi_key;
136 static const struct futex_q futex_q_init = {
137 /* list gets initialized in queue_me()*/
138 .key = FUTEX_KEY_INIT,
139 .bitset = FUTEX_BITSET_MATCH_ANY
143 * Hash buckets are shared by all the futex_keys that hash to the same
144 * location. Each key may have multiple futex_q structures, one for each task
145 * waiting on a futex.
147 struct futex_hash_bucket {
149 struct plist_head chain;
152 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
155 * We hash on the keys returned from get_futex_key (see below).
157 static struct futex_hash_bucket *hash_futex(union futex_key *key)
159 u32 hash = jhash2((u32*)&key->both.word,
160 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
162 return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
166 * Return 1 if two futex_keys are equal, 0 otherwise.
168 static inline int match_futex(union futex_key *key1, union futex_key *key2)
171 && key1->both.word == key2->both.word
172 && key1->both.ptr == key2->both.ptr
173 && key1->both.offset == key2->both.offset);
177 * Take a reference to the resource addressed by a key.
178 * Can be called while holding spinlocks.
181 static void get_futex_key_refs(union futex_key *key)
186 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
188 ihold(key->shared.inode);
190 case FUT_OFF_MMSHARED:
191 atomic_inc(&key->private.mm->mm_count);
197 * Drop a reference to the resource addressed by a key.
198 * The hash bucket spinlock must not be held.
200 static void drop_futex_key_refs(union futex_key *key)
202 if (!key->both.ptr) {
203 /* If we're here then we tried to put a key we failed to get */
208 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
210 iput(key->shared.inode);
212 case FUT_OFF_MMSHARED:
213 mmdrop(key->private.mm);
219 * get_futex_key() - Get parameters which are the keys for a futex
220 * @uaddr: virtual address of the futex
221 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
222 * @key: address where result is stored.
223 * @rw: mapping needs to be read/write (values: VERIFY_READ,
226 * Returns a negative error code or 0
227 * The key words are stored in *key on success.
229 * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
230 * offset_within_page). For private mappings, it's (uaddr, current->mm).
231 * We can usually work out the index without swapping in the page.
233 * lock_page() might sleep, the caller should not hold a spinlock.
236 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
238 unsigned long address = (unsigned long)uaddr;
239 struct mm_struct *mm = current->mm;
240 struct page *page, *page_head;
244 * The futex address must be "naturally" aligned.
246 key->both.offset = address % PAGE_SIZE;
247 if (unlikely((address % sizeof(u32)) != 0))
249 address -= key->both.offset;
252 * PROCESS_PRIVATE futexes are fast.
253 * As the mm cannot disappear under us and the 'key' only needs
254 * virtual address, we dont even have to find the underlying vma.
255 * Note : We do have to check 'uaddr' is a valid user address,
256 * but access_ok() should be faster than find_vma()
259 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
261 key->private.mm = mm;
262 key->private.address = address;
263 get_futex_key_refs(key);
268 err = get_user_pages_fast(address, 1, 1, &page);
270 * If write access is not required (eg. FUTEX_WAIT), try
271 * and get read-only access.
273 if (err == -EFAULT && rw == VERIFY_READ) {
274 err = get_user_pages_fast(address, 1, 0, &page);
282 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
284 if (unlikely(PageTail(page))) {
286 /* serialize against __split_huge_page_splitting() */
288 if (likely(__get_user_pages_fast(address, 1, !ro, &page) == 1)) {
289 page_head = compound_head(page);
291 * page_head is valid pointer but we must pin
292 * it before taking the PG_lock and/or
293 * PG_compound_lock. The moment we re-enable
294 * irqs __split_huge_page_splitting() can
295 * return and the head page can be freed from
296 * under us. We can't take the PG_lock and/or
297 * PG_compound_lock on a page that could be
298 * freed from under us.
300 if (page != page_head) {
311 page_head = compound_head(page);
312 if (page != page_head) {
318 lock_page(page_head);
321 * If page_head->mapping is NULL, then it cannot be a PageAnon
322 * page; but it might be the ZERO_PAGE or in the gate area or
323 * in a special mapping (all cases which we are happy to fail);
324 * or it may have been a good file page when get_user_pages_fast
325 * found it, but truncated or holepunched or subjected to
326 * invalidate_complete_page2 before we got the page lock (also
327 * cases which we are happy to fail). And we hold a reference,
328 * so refcount care in invalidate_complete_page's remove_mapping
329 * prevents drop_caches from setting mapping to NULL beneath us.
331 * The case we do have to guard against is when memory pressure made
332 * shmem_writepage move it from filecache to swapcache beneath us:
333 * an unlikely race, but we do need to retry for page_head->mapping.
335 if (!page_head->mapping) {
336 int shmem_swizzled = PageSwapCache(page_head);
337 unlock_page(page_head);
345 * Private mappings are handled in a simple way.
347 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
348 * it's a read-only handle, it's expected that futexes attach to
349 * the object not the particular process.
351 if (PageAnon(page_head)) {
353 * A RO anonymous page will never change and thus doesn't make
354 * sense for futex operations.
361 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
362 key->private.mm = mm;
363 key->private.address = address;
365 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
366 key->shared.inode = page_head->mapping->host;
367 key->shared.pgoff = basepage_index(page);
370 get_futex_key_refs(key);
373 unlock_page(page_head);
378 static inline void put_futex_key(union futex_key *key)
380 drop_futex_key_refs(key);
384 * fault_in_user_writeable() - Fault in user address and verify RW access
385 * @uaddr: pointer to faulting user space address
387 * Slow path to fixup the fault we just took in the atomic write
390 * We have no generic implementation of a non-destructive write to the
391 * user address. We know that we faulted in the atomic pagefault
392 * disabled section so we can as well avoid the #PF overhead by
393 * calling get_user_pages() right away.
395 static int fault_in_user_writeable(u32 __user *uaddr)
397 struct mm_struct *mm = current->mm;
400 down_read(&mm->mmap_sem);
401 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
403 up_read(&mm->mmap_sem);
405 return ret < 0 ? ret : 0;
409 * futex_top_waiter() - Return the highest priority waiter on a futex
410 * @hb: the hash bucket the futex_q's reside in
411 * @key: the futex key (to distinguish it from other futex futex_q's)
413 * Must be called with the hb lock held.
415 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
416 union futex_key *key)
418 struct futex_q *this;
420 plist_for_each_entry(this, &hb->chain, list) {
421 if (match_futex(&this->key, key))
427 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
428 u32 uval, u32 newval)
433 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
439 static int get_futex_value_locked(u32 *dest, u32 __user *from)
444 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
447 return ret ? -EFAULT : 0;
454 static int refill_pi_state_cache(void)
456 struct futex_pi_state *pi_state;
458 if (likely(current->pi_state_cache))
461 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
466 INIT_LIST_HEAD(&pi_state->list);
467 /* pi_mutex gets initialized later */
468 pi_state->owner = NULL;
469 atomic_set(&pi_state->refcount, 1);
470 pi_state->key = FUTEX_KEY_INIT;
472 current->pi_state_cache = pi_state;
477 static struct futex_pi_state * alloc_pi_state(void)
479 struct futex_pi_state *pi_state = current->pi_state_cache;
482 current->pi_state_cache = NULL;
487 static void free_pi_state(struct futex_pi_state *pi_state)
489 if (!atomic_dec_and_test(&pi_state->refcount))
493 * If pi_state->owner is NULL, the owner is most probably dying
494 * and has cleaned up the pi_state already
496 if (pi_state->owner) {
497 raw_spin_lock_irq(&pi_state->owner->pi_lock);
498 list_del_init(&pi_state->list);
499 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
501 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
504 if (current->pi_state_cache)
508 * pi_state->list is already empty.
509 * clear pi_state->owner.
510 * refcount is at 0 - put it back to 1.
512 pi_state->owner = NULL;
513 atomic_set(&pi_state->refcount, 1);
514 current->pi_state_cache = pi_state;
519 * Look up the task based on what TID userspace gave us.
522 static struct task_struct * futex_find_get_task(pid_t pid)
524 struct task_struct *p;
527 p = find_task_by_vpid(pid);
537 * This task is holding PI mutexes at exit time => bad.
538 * Kernel cleans up PI-state, but userspace is likely hosed.
539 * (Robust-futex cleanup is separate and might save the day for userspace.)
541 void exit_pi_state_list(struct task_struct *curr)
543 struct list_head *next, *head = &curr->pi_state_list;
544 struct futex_pi_state *pi_state;
545 struct futex_hash_bucket *hb;
546 union futex_key key = FUTEX_KEY_INIT;
548 if (!futex_cmpxchg_enabled)
551 * We are a ZOMBIE and nobody can enqueue itself on
552 * pi_state_list anymore, but we have to be careful
553 * versus waiters unqueueing themselves:
555 raw_spin_lock_irq(&curr->pi_lock);
556 while (!list_empty(head)) {
559 pi_state = list_entry(next, struct futex_pi_state, list);
561 hb = hash_futex(&key);
562 raw_spin_unlock_irq(&curr->pi_lock);
564 spin_lock(&hb->lock);
566 raw_spin_lock_irq(&curr->pi_lock);
568 * We dropped the pi-lock, so re-check whether this
569 * task still owns the PI-state:
571 if (head->next != next) {
572 spin_unlock(&hb->lock);
576 WARN_ON(pi_state->owner != curr);
577 WARN_ON(list_empty(&pi_state->list));
578 list_del_init(&pi_state->list);
579 pi_state->owner = NULL;
580 raw_spin_unlock_irq(&curr->pi_lock);
582 rt_mutex_unlock(&pi_state->pi_mutex);
584 spin_unlock(&hb->lock);
586 raw_spin_lock_irq(&curr->pi_lock);
588 raw_spin_unlock_irq(&curr->pi_lock);
592 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
593 union futex_key *key, struct futex_pi_state **ps,
594 struct task_struct *task)
596 struct futex_pi_state *pi_state = NULL;
597 struct futex_q *this, *next;
598 struct plist_head *head;
599 struct task_struct *p;
600 pid_t pid = uval & FUTEX_TID_MASK;
604 plist_for_each_entry_safe(this, next, head, 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))
639 * Protect against a corrupted uval. If uval
640 * is 0x80000000 then pid is 0 and the waiter
641 * bit is set. So the deadlock check in the
642 * calling code has failed and we did not fall
643 * into the check above due to !pid.
645 if (task && pi_state->owner == task)
648 atomic_inc(&pi_state->refcount);
656 * We are the first waiter - try to look up the real owner and attach
657 * the new pi_state to it, but bail out when TID = 0
661 p = futex_find_get_task(pid);
671 * We need to look at the task state flags to figure out,
672 * whether the task is exiting. To protect against the do_exit
673 * change of the task flags, we do this protected by
676 raw_spin_lock_irq(&p->pi_lock);
677 if (unlikely(p->flags & PF_EXITING)) {
679 * The task is on the way out. When PF_EXITPIDONE is
680 * set, we know that the task has finished the
683 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
685 raw_spin_unlock_irq(&p->pi_lock);
690 pi_state = alloc_pi_state();
693 * Initialize the pi_mutex in locked state and make 'p'
696 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
698 /* Store the key for possible exit cleanups: */
699 pi_state->key = *key;
701 WARN_ON(!list_empty(&pi_state->list));
702 list_add(&pi_state->list, &p->pi_state_list);
704 raw_spin_unlock_irq(&p->pi_lock);
714 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
715 * @uaddr: the pi futex user address
716 * @hb: the pi futex hash bucket
717 * @key: the futex key associated with uaddr and hb
718 * @ps: the pi_state pointer where we store the result of the
720 * @task: the task to perform the atomic lock work for. This will
721 * be "current" except in the case of requeue pi.
722 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
726 * 1 - acquired the lock
729 * The hb->lock and futex_key refs shall be held by the caller.
731 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
732 union futex_key *key,
733 struct futex_pi_state **ps,
734 struct task_struct *task, int set_waiters)
736 int lock_taken, ret, force_take = 0;
737 u32 uval, newval, curval, vpid = task_pid_vnr(task);
740 ret = lock_taken = 0;
743 * To avoid races, we attempt to take the lock here again
744 * (by doing a 0 -> TID atomic cmpxchg), while holding all
745 * the locks. It will most likely not succeed.
749 newval |= FUTEX_WAITERS;
751 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, 0, newval)))
757 if ((unlikely((curval & FUTEX_TID_MASK) == vpid)))
761 * Surprise - we got the lock. Just return to userspace:
763 if (unlikely(!curval))
769 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
770 * to wake at the next unlock.
772 newval = curval | FUTEX_WAITERS;
775 * Should we force take the futex? See below.
777 if (unlikely(force_take)) {
779 * Keep the OWNER_DIED and the WAITERS bit and set the
782 newval = (curval & ~FUTEX_TID_MASK) | vpid;
787 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
789 if (unlikely(curval != uval))
793 * We took the lock due to forced take over.
795 if (unlikely(lock_taken))
799 * We dont have the lock. Look up the PI state (or create it if
800 * we are the first waiter):
802 ret = lookup_pi_state(uval, hb, key, ps, task);
808 * We failed to find an owner for this
809 * futex. So we have no pi_state to block
810 * on. This can happen in two cases:
813 * 2) A stale FUTEX_WAITERS bit
815 * Re-read the futex value.
817 if (get_futex_value_locked(&curval, uaddr))
821 * If the owner died or we have a stale
822 * WAITERS bit the owner TID in the user space
825 if (!(curval & FUTEX_TID_MASK)) {
838 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
839 * @q: The futex_q to unqueue
841 * The q->lock_ptr must not be NULL and must be held by the caller.
843 static void __unqueue_futex(struct futex_q *q)
845 struct futex_hash_bucket *hb;
847 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
848 || WARN_ON(plist_node_empty(&q->list)))
851 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
852 plist_del(&q->list, &hb->chain);
856 * The hash bucket lock must be held when this is called.
857 * Afterwards, the futex_q must not be accessed.
859 static void wake_futex(struct futex_q *q)
861 struct task_struct *p = q->task;
863 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
867 * We set q->lock_ptr = NULL _before_ we wake up the task. If
868 * a non-futex wake up happens on another CPU then the task
869 * might exit and p would dereference a non-existing task
870 * struct. Prevent this by holding a reference on p across the
877 * The waiting task can free the futex_q as soon as
878 * q->lock_ptr = NULL is written, without taking any locks. A
879 * memory barrier is required here to prevent the following
880 * store to lock_ptr from getting ahead of the plist_del.
885 wake_up_state(p, TASK_NORMAL);
889 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
891 struct task_struct *new_owner;
892 struct futex_pi_state *pi_state = this->pi_state;
893 u32 uninitialized_var(curval), newval;
899 * If current does not own the pi_state then the futex is
900 * inconsistent and user space fiddled with the futex value.
902 if (pi_state->owner != current)
905 raw_spin_lock(&pi_state->pi_mutex.wait_lock);
906 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
909 * It is possible that the next waiter (the one that brought
910 * this owner to the kernel) timed out and is no longer
911 * waiting on the lock.
914 new_owner = this->task;
917 * We pass it to the next owner. (The WAITERS bit is always
918 * kept enabled while there is PI state around. We must also
919 * preserve the owner died bit.)
921 if (!(uval & FUTEX_OWNER_DIED)) {
924 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
926 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
928 else if (curval != uval)
931 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
936 raw_spin_lock_irq(&pi_state->owner->pi_lock);
937 WARN_ON(list_empty(&pi_state->list));
938 list_del_init(&pi_state->list);
939 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
941 raw_spin_lock_irq(&new_owner->pi_lock);
942 WARN_ON(!list_empty(&pi_state->list));
943 list_add(&pi_state->list, &new_owner->pi_state_list);
944 pi_state->owner = new_owner;
945 raw_spin_unlock_irq(&new_owner->pi_lock);
947 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
948 rt_mutex_unlock(&pi_state->pi_mutex);
953 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
955 u32 uninitialized_var(oldval);
958 * There is no waiter, so we unlock the futex. The owner died
959 * bit has not to be preserved here. We are the owner:
961 if (cmpxchg_futex_value_locked(&oldval, uaddr, uval, 0))
970 * Express the locking dependencies for lockdep:
973 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
976 spin_lock(&hb1->lock);
978 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
979 } else { /* hb1 > hb2 */
980 spin_lock(&hb2->lock);
981 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
986 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
988 spin_unlock(&hb1->lock);
990 spin_unlock(&hb2->lock);
994 * Wake up waiters matching bitset queued on this futex (uaddr).
997 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
999 struct futex_hash_bucket *hb;
1000 struct futex_q *this, *next;
1001 struct plist_head *head;
1002 union futex_key key = FUTEX_KEY_INIT;
1008 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1009 if (unlikely(ret != 0))
1012 hb = hash_futex(&key);
1013 spin_lock(&hb->lock);
1016 plist_for_each_entry_safe(this, next, head, list) {
1017 if (match_futex (&this->key, &key)) {
1018 if (this->pi_state || this->rt_waiter) {
1023 /* Check if one of the bits is set in both bitsets */
1024 if (!(this->bitset & bitset))
1028 if (++ret >= nr_wake)
1033 spin_unlock(&hb->lock);
1034 put_futex_key(&key);
1040 * Wake up all waiters hashed on the physical page that is mapped
1041 * to this virtual address:
1044 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1045 int nr_wake, int nr_wake2, int op)
1047 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1048 struct futex_hash_bucket *hb1, *hb2;
1049 struct plist_head *head;
1050 struct futex_q *this, *next;
1054 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1055 if (unlikely(ret != 0))
1057 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1058 if (unlikely(ret != 0))
1061 hb1 = hash_futex(&key1);
1062 hb2 = hash_futex(&key2);
1065 double_lock_hb(hb1, hb2);
1066 op_ret = futex_atomic_op_inuser(op, uaddr2);
1067 if (unlikely(op_ret < 0)) {
1069 double_unlock_hb(hb1, hb2);
1073 * we don't get EFAULT from MMU faults if we don't have an MMU,
1074 * but we might get them from range checking
1080 if (unlikely(op_ret != -EFAULT)) {
1085 ret = fault_in_user_writeable(uaddr2);
1089 if (!(flags & FLAGS_SHARED))
1092 put_futex_key(&key2);
1093 put_futex_key(&key1);
1099 plist_for_each_entry_safe(this, next, head, list) {
1100 if (match_futex (&this->key, &key1)) {
1101 if (this->pi_state || this->rt_waiter) {
1106 if (++ret >= nr_wake)
1115 plist_for_each_entry_safe(this, next, head, list) {
1116 if (match_futex (&this->key, &key2)) {
1117 if (this->pi_state || this->rt_waiter) {
1122 if (++op_ret >= nr_wake2)
1130 double_unlock_hb(hb1, hb2);
1132 put_futex_key(&key2);
1134 put_futex_key(&key1);
1140 * requeue_futex() - Requeue a futex_q from one hb to another
1141 * @q: the futex_q to requeue
1142 * @hb1: the source hash_bucket
1143 * @hb2: the target hash_bucket
1144 * @key2: the new key for the requeued futex_q
1147 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1148 struct futex_hash_bucket *hb2, union futex_key *key2)
1152 * If key1 and key2 hash to the same bucket, no need to
1155 if (likely(&hb1->chain != &hb2->chain)) {
1156 plist_del(&q->list, &hb1->chain);
1157 plist_add(&q->list, &hb2->chain);
1158 q->lock_ptr = &hb2->lock;
1160 get_futex_key_refs(key2);
1165 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1167 * @key: the key of the requeue target futex
1168 * @hb: the hash_bucket of the requeue target futex
1170 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1171 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1172 * to the requeue target futex so the waiter can detect the wakeup on the right
1173 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1174 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1175 * to protect access to the pi_state to fixup the owner later. Must be called
1176 * with both q->lock_ptr and hb->lock held.
1179 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1180 struct futex_hash_bucket *hb)
1182 get_futex_key_refs(key);
1187 WARN_ON(!q->rt_waiter);
1188 q->rt_waiter = NULL;
1190 q->lock_ptr = &hb->lock;
1192 wake_up_state(q->task, TASK_NORMAL);
1196 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1197 * @pifutex: the user address of the to futex
1198 * @hb1: the from futex hash bucket, must be locked by the caller
1199 * @hb2: the to futex hash bucket, must be locked by the caller
1200 * @key1: the from futex key
1201 * @key2: the to futex key
1202 * @ps: address to store the pi_state pointer
1203 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1205 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1206 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1207 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1208 * hb1 and hb2 must be held by the caller.
1211 * 0 - failed to acquire the lock atomicly
1212 * >0 - acquired the lock, return value is vpid of the top_waiter
1215 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1216 struct futex_hash_bucket *hb1,
1217 struct futex_hash_bucket *hb2,
1218 union futex_key *key1, union futex_key *key2,
1219 struct futex_pi_state **ps, int set_waiters)
1221 struct futex_q *top_waiter = NULL;
1225 if (get_futex_value_locked(&curval, pifutex))
1229 * Find the top_waiter and determine if there are additional waiters.
1230 * If the caller intends to requeue more than 1 waiter to pifutex,
1231 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1232 * as we have means to handle the possible fault. If not, don't set
1233 * the bit unecessarily as it will force the subsequent unlock to enter
1236 top_waiter = futex_top_waiter(hb1, key1);
1238 /* There are no waiters, nothing for us to do. */
1242 /* Ensure we requeue to the expected futex. */
1243 if (!match_futex(top_waiter->requeue_pi_key, key2))
1247 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1248 * the contended case or if set_waiters is 1. The pi_state is returned
1249 * in ps in contended cases.
1251 vpid = task_pid_vnr(top_waiter->task);
1252 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1255 requeue_pi_wake_futex(top_waiter, key2, hb2);
1262 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1263 * @uaddr1: source futex user address
1264 * @flags: futex flags (FLAGS_SHARED, etc.)
1265 * @uaddr2: target futex user address
1266 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1267 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1268 * @cmpval: @uaddr1 expected value (or %NULL)
1269 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1270 * pi futex (pi to pi requeue is not supported)
1272 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1273 * uaddr2 atomically on behalf of the top waiter.
1276 * >=0 - on success, the number of tasks requeued or woken
1279 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1280 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1281 u32 *cmpval, int requeue_pi)
1283 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1284 int drop_count = 0, task_count = 0, ret;
1285 struct futex_pi_state *pi_state = NULL;
1286 struct futex_hash_bucket *hb1, *hb2;
1287 struct plist_head *head1;
1288 struct futex_q *this, *next;
1292 * Requeue PI only works on two distinct uaddrs. This
1293 * check is only valid for private futexes. See below.
1295 if (uaddr1 == uaddr2)
1299 * requeue_pi requires a pi_state, try to allocate it now
1300 * without any locks in case it fails.
1302 if (refill_pi_state_cache())
1305 * requeue_pi must wake as many tasks as it can, up to nr_wake
1306 * + nr_requeue, since it acquires the rt_mutex prior to
1307 * returning to userspace, so as to not leave the rt_mutex with
1308 * waiters and no owner. However, second and third wake-ups
1309 * cannot be predicted as they involve race conditions with the
1310 * first wake and a fault while looking up the pi_state. Both
1311 * pthread_cond_signal() and pthread_cond_broadcast() should
1319 if (pi_state != NULL) {
1321 * We will have to lookup the pi_state again, so free this one
1322 * to keep the accounting correct.
1324 free_pi_state(pi_state);
1328 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1329 if (unlikely(ret != 0))
1331 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1332 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1333 if (unlikely(ret != 0))
1337 * The check above which compares uaddrs is not sufficient for
1338 * shared futexes. We need to compare the keys:
1340 if (requeue_pi && match_futex(&key1, &key2)) {
1345 hb1 = hash_futex(&key1);
1346 hb2 = hash_futex(&key2);
1349 double_lock_hb(hb1, hb2);
1351 if (likely(cmpval != NULL)) {
1354 ret = get_futex_value_locked(&curval, uaddr1);
1356 if (unlikely(ret)) {
1357 double_unlock_hb(hb1, hb2);
1359 ret = get_user(curval, uaddr1);
1363 if (!(flags & FLAGS_SHARED))
1366 put_futex_key(&key2);
1367 put_futex_key(&key1);
1370 if (curval != *cmpval) {
1376 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1378 * Attempt to acquire uaddr2 and wake the top waiter. If we
1379 * intend to requeue waiters, force setting the FUTEX_WAITERS
1380 * bit. We force this here where we are able to easily handle
1381 * faults rather in the requeue loop below.
1383 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1384 &key2, &pi_state, nr_requeue);
1387 * At this point the top_waiter has either taken uaddr2 or is
1388 * waiting on it. If the former, then the pi_state will not
1389 * exist yet, look it up one more time to ensure we have a
1390 * reference to it. If the lock was taken, ret contains the
1391 * vpid of the top waiter task.
1398 * If we acquired the lock, then the user
1399 * space value of uaddr2 should be vpid. It
1400 * cannot be changed by the top waiter as it
1401 * is blocked on hb2 lock if it tries to do
1402 * so. If something fiddled with it behind our
1403 * back the pi state lookup might unearth
1404 * it. So we rather use the known value than
1405 * rereading and handing potential crap to
1408 ret = lookup_pi_state(ret, hb2, &key2, &pi_state, NULL);
1415 double_unlock_hb(hb1, hb2);
1416 put_futex_key(&key2);
1417 put_futex_key(&key1);
1418 ret = fault_in_user_writeable(uaddr2);
1423 /* The owner was exiting, try again. */
1424 double_unlock_hb(hb1, hb2);
1425 put_futex_key(&key2);
1426 put_futex_key(&key1);
1434 head1 = &hb1->chain;
1435 plist_for_each_entry_safe(this, next, head1, list) {
1436 if (task_count - nr_wake >= nr_requeue)
1439 if (!match_futex(&this->key, &key1))
1443 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1444 * be paired with each other and no other futex ops.
1446 * We should never be requeueing a futex_q with a pi_state,
1447 * which is awaiting a futex_unlock_pi().
1449 if ((requeue_pi && !this->rt_waiter) ||
1450 (!requeue_pi && this->rt_waiter) ||
1457 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1458 * lock, we already woke the top_waiter. If not, it will be
1459 * woken by futex_unlock_pi().
1461 if (++task_count <= nr_wake && !requeue_pi) {
1466 /* Ensure we requeue to the expected futex for requeue_pi. */
1467 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1473 * Requeue nr_requeue waiters and possibly one more in the case
1474 * of requeue_pi if we couldn't acquire the lock atomically.
1477 /* Prepare the waiter to take the rt_mutex. */
1478 atomic_inc(&pi_state->refcount);
1479 this->pi_state = pi_state;
1480 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1484 /* We got the lock. */
1485 requeue_pi_wake_futex(this, &key2, hb2);
1490 this->pi_state = NULL;
1491 free_pi_state(pi_state);
1495 requeue_futex(this, hb1, hb2, &key2);
1500 double_unlock_hb(hb1, hb2);
1503 * drop_futex_key_refs() must be called outside the spinlocks. During
1504 * the requeue we moved futex_q's from the hash bucket at key1 to the
1505 * one at key2 and updated their key pointer. We no longer need to
1506 * hold the references to key1.
1508 while (--drop_count >= 0)
1509 drop_futex_key_refs(&key1);
1512 put_futex_key(&key2);
1514 put_futex_key(&key1);
1516 if (pi_state != NULL)
1517 free_pi_state(pi_state);
1518 return ret ? ret : task_count;
1521 /* The key must be already stored in q->key. */
1522 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1523 __acquires(&hb->lock)
1525 struct futex_hash_bucket *hb;
1527 hb = hash_futex(&q->key);
1528 q->lock_ptr = &hb->lock;
1530 spin_lock(&hb->lock);
1535 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1536 __releases(&hb->lock)
1538 spin_unlock(&hb->lock);
1542 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1543 * @q: The futex_q to enqueue
1544 * @hb: The destination hash bucket
1546 * The hb->lock must be held by the caller, and is released here. A call to
1547 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1548 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1549 * or nothing if the unqueue is done as part of the wake process and the unqueue
1550 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1553 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1554 __releases(&hb->lock)
1559 * The priority used to register this element is
1560 * - either the real thread-priority for the real-time threads
1561 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1562 * - or MAX_RT_PRIO for non-RT threads.
1563 * Thus, all RT-threads are woken first in priority order, and
1564 * the others are woken last, in FIFO order.
1566 prio = min(current->normal_prio, MAX_RT_PRIO);
1568 plist_node_init(&q->list, prio);
1569 plist_add(&q->list, &hb->chain);
1571 spin_unlock(&hb->lock);
1575 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1576 * @q: The futex_q to unqueue
1578 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1579 * be paired with exactly one earlier call to queue_me().
1582 * 1 - if the futex_q was still queued (and we removed unqueued it)
1583 * 0 - if the futex_q was already removed by the waking thread
1585 static int unqueue_me(struct futex_q *q)
1587 spinlock_t *lock_ptr;
1590 /* In the common case we don't take the spinlock, which is nice. */
1592 lock_ptr = q->lock_ptr;
1594 if (lock_ptr != NULL) {
1595 spin_lock(lock_ptr);
1597 * q->lock_ptr can change between reading it and
1598 * spin_lock(), causing us to take the wrong lock. This
1599 * corrects the race condition.
1601 * Reasoning goes like this: if we have the wrong lock,
1602 * q->lock_ptr must have changed (maybe several times)
1603 * between reading it and the spin_lock(). It can
1604 * change again after the spin_lock() but only if it was
1605 * already changed before the spin_lock(). It cannot,
1606 * however, change back to the original value. Therefore
1607 * we can detect whether we acquired the correct lock.
1609 if (unlikely(lock_ptr != q->lock_ptr)) {
1610 spin_unlock(lock_ptr);
1615 BUG_ON(q->pi_state);
1617 spin_unlock(lock_ptr);
1621 drop_futex_key_refs(&q->key);
1626 * PI futexes can not be requeued and must remove themself from the
1627 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1630 static void unqueue_me_pi(struct futex_q *q)
1631 __releases(q->lock_ptr)
1635 BUG_ON(!q->pi_state);
1636 free_pi_state(q->pi_state);
1639 spin_unlock(q->lock_ptr);
1643 * Fixup the pi_state owner with the new owner.
1645 * Must be called with hash bucket lock held and mm->sem held for non
1648 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1649 struct task_struct *newowner)
1651 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1652 struct futex_pi_state *pi_state = q->pi_state;
1653 struct task_struct *oldowner = pi_state->owner;
1654 u32 uval, uninitialized_var(curval), newval;
1658 if (!pi_state->owner)
1659 newtid |= FUTEX_OWNER_DIED;
1662 * We are here either because we stole the rtmutex from the
1663 * previous highest priority waiter or we are the highest priority
1664 * waiter but failed to get the rtmutex the first time.
1665 * We have to replace the newowner TID in the user space variable.
1666 * This must be atomic as we have to preserve the owner died bit here.
1668 * Note: We write the user space value _before_ changing the pi_state
1669 * because we can fault here. Imagine swapped out pages or a fork
1670 * that marked all the anonymous memory readonly for cow.
1672 * Modifying pi_state _before_ the user space value would
1673 * leave the pi_state in an inconsistent state when we fault
1674 * here, because we need to drop the hash bucket lock to
1675 * handle the fault. This might be observed in the PID check
1676 * in lookup_pi_state.
1679 if (get_futex_value_locked(&uval, uaddr))
1683 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1685 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1693 * We fixed up user space. Now we need to fix the pi_state
1696 if (pi_state->owner != NULL) {
1697 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1698 WARN_ON(list_empty(&pi_state->list));
1699 list_del_init(&pi_state->list);
1700 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1703 pi_state->owner = newowner;
1705 raw_spin_lock_irq(&newowner->pi_lock);
1706 WARN_ON(!list_empty(&pi_state->list));
1707 list_add(&pi_state->list, &newowner->pi_state_list);
1708 raw_spin_unlock_irq(&newowner->pi_lock);
1712 * To handle the page fault we need to drop the hash bucket
1713 * lock here. That gives the other task (either the highest priority
1714 * waiter itself or the task which stole the rtmutex) the
1715 * chance to try the fixup of the pi_state. So once we are
1716 * back from handling the fault we need to check the pi_state
1717 * after reacquiring the hash bucket lock and before trying to
1718 * do another fixup. When the fixup has been done already we
1722 spin_unlock(q->lock_ptr);
1724 ret = fault_in_user_writeable(uaddr);
1726 spin_lock(q->lock_ptr);
1729 * Check if someone else fixed it for us:
1731 if (pi_state->owner != oldowner)
1740 static long futex_wait_restart(struct restart_block *restart);
1743 * fixup_owner() - Post lock pi_state and corner case management
1744 * @uaddr: user address of the futex
1745 * @q: futex_q (contains pi_state and access to the rt_mutex)
1746 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1748 * After attempting to lock an rt_mutex, this function is called to cleanup
1749 * the pi_state owner as well as handle race conditions that may allow us to
1750 * acquire the lock. Must be called with the hb lock held.
1753 * 1 - success, lock taken
1754 * 0 - success, lock not taken
1755 * <0 - on error (-EFAULT)
1757 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
1759 struct task_struct *owner;
1764 * Got the lock. We might not be the anticipated owner if we
1765 * did a lock-steal - fix up the PI-state in that case:
1767 if (q->pi_state->owner != current)
1768 ret = fixup_pi_state_owner(uaddr, q, current);
1773 * Catch the rare case, where the lock was released when we were on the
1774 * way back before we locked the hash bucket.
1776 if (q->pi_state->owner == current) {
1778 * Try to get the rt_mutex now. This might fail as some other
1779 * task acquired the rt_mutex after we removed ourself from the
1780 * rt_mutex waiters list.
1782 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1788 * pi_state is incorrect, some other task did a lock steal and
1789 * we returned due to timeout or signal without taking the
1790 * rt_mutex. Too late.
1792 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
1793 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1795 owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
1796 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
1797 ret = fixup_pi_state_owner(uaddr, q, owner);
1802 * Paranoia check. If we did not take the lock, then we should not be
1803 * the owner of the rt_mutex.
1805 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1806 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1807 "pi-state %p\n", ret,
1808 q->pi_state->pi_mutex.owner,
1809 q->pi_state->owner);
1812 return ret ? ret : locked;
1816 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1817 * @hb: the futex hash bucket, must be locked by the caller
1818 * @q: the futex_q to queue up on
1819 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1821 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1822 struct hrtimer_sleeper *timeout)
1825 * The task state is guaranteed to be set before another task can
1826 * wake it. set_current_state() is implemented using set_mb() and
1827 * queue_me() calls spin_unlock() upon completion, both serializing
1828 * access to the hash list and forcing another memory barrier.
1830 set_current_state(TASK_INTERRUPTIBLE);
1835 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1836 if (!hrtimer_active(&timeout->timer))
1837 timeout->task = NULL;
1841 * If we have been removed from the hash list, then another task
1842 * has tried to wake us, and we can skip the call to schedule().
1844 if (likely(!plist_node_empty(&q->list))) {
1846 * If the timer has already expired, current will already be
1847 * flagged for rescheduling. Only call schedule if there
1848 * is no timeout, or if it has yet to expire.
1850 if (!timeout || timeout->task)
1853 __set_current_state(TASK_RUNNING);
1857 * futex_wait_setup() - Prepare to wait on a futex
1858 * @uaddr: the futex userspace address
1859 * @val: the expected value
1860 * @flags: futex flags (FLAGS_SHARED, etc.)
1861 * @q: the associated futex_q
1862 * @hb: storage for hash_bucket pointer to be returned to caller
1864 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1865 * compare it with the expected value. Handle atomic faults internally.
1866 * Return with the hb lock held and a q.key reference on success, and unlocked
1867 * with no q.key reference on failure.
1870 * 0 - uaddr contains val and hb has been locked
1871 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
1873 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
1874 struct futex_q *q, struct futex_hash_bucket **hb)
1880 * Access the page AFTER the hash-bucket is locked.
1881 * Order is important:
1883 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1884 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1886 * The basic logical guarantee of a futex is that it blocks ONLY
1887 * if cond(var) is known to be true at the time of blocking, for
1888 * any cond. If we locked the hash-bucket after testing *uaddr, that
1889 * would open a race condition where we could block indefinitely with
1890 * cond(var) false, which would violate the guarantee.
1892 * On the other hand, we insert q and release the hash-bucket only
1893 * after testing *uaddr. This guarantees that futex_wait() will NOT
1894 * absorb a wakeup if *uaddr does not match the desired values
1895 * while the syscall executes.
1898 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
1899 if (unlikely(ret != 0))
1903 *hb = queue_lock(q);
1905 ret = get_futex_value_locked(&uval, uaddr);
1908 queue_unlock(q, *hb);
1910 ret = get_user(uval, uaddr);
1914 if (!(flags & FLAGS_SHARED))
1917 put_futex_key(&q->key);
1922 queue_unlock(q, *hb);
1928 put_futex_key(&q->key);
1932 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
1933 ktime_t *abs_time, u32 bitset)
1935 struct hrtimer_sleeper timeout, *to = NULL;
1936 struct restart_block *restart;
1937 struct futex_hash_bucket *hb;
1938 struct futex_q q = futex_q_init;
1948 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
1949 CLOCK_REALTIME : CLOCK_MONOTONIC,
1951 hrtimer_init_sleeper(to, current);
1952 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1953 current->timer_slack_ns);
1958 * Prepare to wait on uaddr. On success, holds hb lock and increments
1961 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
1965 /* queue_me and wait for wakeup, timeout, or a signal. */
1966 futex_wait_queue_me(hb, &q, to);
1968 /* If we were woken (and unqueued), we succeeded, whatever. */
1970 /* unqueue_me() drops q.key ref */
1971 if (!unqueue_me(&q))
1974 if (to && !to->task)
1978 * We expect signal_pending(current), but we might be the
1979 * victim of a spurious wakeup as well.
1981 if (!signal_pending(current))
1988 restart = ¤t_thread_info()->restart_block;
1989 restart->fn = futex_wait_restart;
1990 restart->futex.uaddr = uaddr;
1991 restart->futex.val = val;
1992 restart->futex.time = abs_time->tv64;
1993 restart->futex.bitset = bitset;
1994 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
1996 ret = -ERESTART_RESTARTBLOCK;
2000 hrtimer_cancel(&to->timer);
2001 destroy_hrtimer_on_stack(&to->timer);
2007 static long futex_wait_restart(struct restart_block *restart)
2009 u32 __user *uaddr = restart->futex.uaddr;
2010 ktime_t t, *tp = NULL;
2012 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2013 t.tv64 = restart->futex.time;
2016 restart->fn = do_no_restart_syscall;
2018 return (long)futex_wait(uaddr, restart->futex.flags,
2019 restart->futex.val, tp, restart->futex.bitset);
2024 * Userspace tried a 0 -> TID atomic transition of the futex value
2025 * and failed. The kernel side here does the whole locking operation:
2026 * if there are waiters then it will block, it does PI, etc. (Due to
2027 * races the kernel might see a 0 value of the futex too.)
2029 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
2030 ktime_t *time, int trylock)
2032 struct hrtimer_sleeper timeout, *to = NULL;
2033 struct futex_hash_bucket *hb;
2034 struct futex_q q = futex_q_init;
2037 if (refill_pi_state_cache())
2042 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2044 hrtimer_init_sleeper(to, current);
2045 hrtimer_set_expires(&to->timer, *time);
2049 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2050 if (unlikely(ret != 0))
2054 hb = queue_lock(&q);
2056 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2057 if (unlikely(ret)) {
2060 /* We got the lock. */
2062 goto out_unlock_put_key;
2067 * Task is exiting and we just wait for the
2070 queue_unlock(&q, hb);
2071 put_futex_key(&q.key);
2075 goto out_unlock_put_key;
2080 * Only actually queue now that the atomic ops are done:
2084 WARN_ON(!q.pi_state);
2086 * Block on the PI mutex:
2089 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
2091 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2092 /* Fixup the trylock return value: */
2093 ret = ret ? 0 : -EWOULDBLOCK;
2096 spin_lock(q.lock_ptr);
2098 * Fixup the pi_state owner and possibly acquire the lock if we
2101 res = fixup_owner(uaddr, &q, !ret);
2103 * If fixup_owner() returned an error, proprogate that. If it acquired
2104 * the lock, clear our -ETIMEDOUT or -EINTR.
2107 ret = (res < 0) ? res : 0;
2110 * If fixup_owner() faulted and was unable to handle the fault, unlock
2111 * it and return the fault to userspace.
2113 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2114 rt_mutex_unlock(&q.pi_state->pi_mutex);
2116 /* Unqueue and drop the lock */
2122 queue_unlock(&q, hb);
2125 put_futex_key(&q.key);
2128 destroy_hrtimer_on_stack(&to->timer);
2129 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2132 queue_unlock(&q, hb);
2134 ret = fault_in_user_writeable(uaddr);
2138 if (!(flags & FLAGS_SHARED))
2141 put_futex_key(&q.key);
2146 * Userspace attempted a TID -> 0 atomic transition, and failed.
2147 * This is the in-kernel slowpath: we look up the PI state (if any),
2148 * and do the rt-mutex unlock.
2150 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2152 struct futex_hash_bucket *hb;
2153 struct futex_q *this, *next;
2154 struct plist_head *head;
2155 union futex_key key = FUTEX_KEY_INIT;
2156 u32 uval, vpid = task_pid_vnr(current);
2160 if (get_user(uval, uaddr))
2163 * We release only a lock we actually own:
2165 if ((uval & FUTEX_TID_MASK) != vpid)
2168 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2169 if (unlikely(ret != 0))
2172 hb = hash_futex(&key);
2173 spin_lock(&hb->lock);
2176 * To avoid races, try to do the TID -> 0 atomic transition
2177 * again. If it succeeds then we can return without waking
2180 if (!(uval & FUTEX_OWNER_DIED) &&
2181 cmpxchg_futex_value_locked(&uval, uaddr, vpid, 0))
2184 * Rare case: we managed to release the lock atomically,
2185 * no need to wake anyone else up:
2187 if (unlikely(uval == vpid))
2191 * Ok, other tasks may need to be woken up - check waiters
2192 * and do the wakeup if necessary:
2196 plist_for_each_entry_safe(this, next, head, list) {
2197 if (!match_futex (&this->key, &key))
2199 ret = wake_futex_pi(uaddr, uval, this);
2201 * The atomic access to the futex value
2202 * generated a pagefault, so retry the
2203 * user-access and the wakeup:
2210 * No waiters - kernel unlocks the futex:
2212 if (!(uval & FUTEX_OWNER_DIED)) {
2213 ret = unlock_futex_pi(uaddr, uval);
2219 spin_unlock(&hb->lock);
2220 put_futex_key(&key);
2226 spin_unlock(&hb->lock);
2227 put_futex_key(&key);
2229 ret = fault_in_user_writeable(uaddr);
2237 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2238 * @hb: the hash_bucket futex_q was original enqueued on
2239 * @q: the futex_q woken while waiting to be requeued
2240 * @key2: the futex_key of the requeue target futex
2241 * @timeout: the timeout associated with the wait (NULL if none)
2243 * Detect if the task was woken on the initial futex as opposed to the requeue
2244 * target futex. If so, determine if it was a timeout or a signal that caused
2245 * the wakeup and return the appropriate error code to the caller. Must be
2246 * called with the hb lock held.
2249 * 0 - no early wakeup detected
2250 * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2253 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2254 struct futex_q *q, union futex_key *key2,
2255 struct hrtimer_sleeper *timeout)
2260 * With the hb lock held, we avoid races while we process the wakeup.
2261 * We only need to hold hb (and not hb2) to ensure atomicity as the
2262 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2263 * It can't be requeued from uaddr2 to something else since we don't
2264 * support a PI aware source futex for requeue.
2266 if (!match_futex(&q->key, key2)) {
2267 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2269 * We were woken prior to requeue by a timeout or a signal.
2270 * Unqueue the futex_q and determine which it was.
2272 plist_del(&q->list, &hb->chain);
2274 /* Handle spurious wakeups gracefully */
2276 if (timeout && !timeout->task)
2278 else if (signal_pending(current))
2279 ret = -ERESTARTNOINTR;
2285 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2286 * @uaddr: the futex we initially wait on (non-pi)
2287 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2288 * the same type, no requeueing from private to shared, etc.
2289 * @val: the expected value of uaddr
2290 * @abs_time: absolute timeout
2291 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2292 * @clockrt: whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2293 * @uaddr2: the pi futex we will take prior to returning to user-space
2295 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2296 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2297 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2298 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2299 * without one, the pi logic would not know which task to boost/deboost, if
2300 * there was a need to.
2302 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2303 * via the following:
2304 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2305 * 2) wakeup on uaddr2 after a requeue
2309 * If 3, cleanup and return -ERESTARTNOINTR.
2311 * If 2, we may then block on trying to take the rt_mutex and return via:
2312 * 5) successful lock
2315 * 8) other lock acquisition failure
2317 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2319 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2325 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2326 u32 val, ktime_t *abs_time, u32 bitset,
2329 struct hrtimer_sleeper timeout, *to = NULL;
2330 struct rt_mutex_waiter rt_waiter;
2331 struct rt_mutex *pi_mutex = NULL;
2332 struct futex_hash_bucket *hb;
2333 union futex_key key2 = FUTEX_KEY_INIT;
2334 struct futex_q q = futex_q_init;
2337 if (uaddr == uaddr2)
2345 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2346 CLOCK_REALTIME : CLOCK_MONOTONIC,
2348 hrtimer_init_sleeper(to, current);
2349 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2350 current->timer_slack_ns);
2354 * The waiter is allocated on our stack, manipulated by the requeue
2355 * code while we sleep on uaddr.
2357 debug_rt_mutex_init_waiter(&rt_waiter);
2358 rt_waiter.task = NULL;
2360 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2361 if (unlikely(ret != 0))
2365 q.rt_waiter = &rt_waiter;
2366 q.requeue_pi_key = &key2;
2369 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2372 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2377 * The check above which compares uaddrs is not sufficient for
2378 * shared futexes. We need to compare the keys:
2380 if (match_futex(&q.key, &key2)) {
2385 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2386 futex_wait_queue_me(hb, &q, to);
2388 spin_lock(&hb->lock);
2389 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2390 spin_unlock(&hb->lock);
2395 * In order for us to be here, we know our q.key == key2, and since
2396 * we took the hb->lock above, we also know that futex_requeue() has
2397 * completed and we no longer have to concern ourselves with a wakeup
2398 * race with the atomic proxy lock acquisition by the requeue code. The
2399 * futex_requeue dropped our key1 reference and incremented our key2
2403 /* Check if the requeue code acquired the second futex for us. */
2406 * Got the lock. We might not be the anticipated owner if we
2407 * did a lock-steal - fix up the PI-state in that case.
2409 if (q.pi_state && (q.pi_state->owner != current)) {
2410 spin_lock(q.lock_ptr);
2411 ret = fixup_pi_state_owner(uaddr2, &q, current);
2412 spin_unlock(q.lock_ptr);
2416 * We have been woken up by futex_unlock_pi(), a timeout, or a
2417 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2420 WARN_ON(!q.pi_state);
2421 pi_mutex = &q.pi_state->pi_mutex;
2422 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2423 debug_rt_mutex_free_waiter(&rt_waiter);
2425 spin_lock(q.lock_ptr);
2427 * Fixup the pi_state owner and possibly acquire the lock if we
2430 res = fixup_owner(uaddr2, &q, !ret);
2432 * If fixup_owner() returned an error, proprogate that. If it
2433 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2436 ret = (res < 0) ? res : 0;
2438 /* Unqueue and drop the lock. */
2443 * If fixup_pi_state_owner() faulted and was unable to handle the
2444 * fault, unlock the rt_mutex and return the fault to userspace.
2446 if (ret == -EFAULT) {
2447 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2448 rt_mutex_unlock(pi_mutex);
2449 } else if (ret == -EINTR) {
2451 * We've already been requeued, but cannot restart by calling
2452 * futex_lock_pi() directly. We could restart this syscall, but
2453 * it would detect that the user space "val" changed and return
2454 * -EWOULDBLOCK. Save the overhead of the restart and return
2455 * -EWOULDBLOCK directly.
2461 put_futex_key(&q.key);
2463 put_futex_key(&key2);
2467 hrtimer_cancel(&to->timer);
2468 destroy_hrtimer_on_stack(&to->timer);
2474 * Support for robust futexes: the kernel cleans up held futexes at
2477 * Implementation: user-space maintains a per-thread list of locks it
2478 * is holding. Upon do_exit(), the kernel carefully walks this list,
2479 * and marks all locks that are owned by this thread with the
2480 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2481 * always manipulated with the lock held, so the list is private and
2482 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2483 * field, to allow the kernel to clean up if the thread dies after
2484 * acquiring the lock, but just before it could have added itself to
2485 * the list. There can only be one such pending lock.
2489 * sys_set_robust_list() - Set the robust-futex list head of a task
2490 * @head: pointer to the list-head
2491 * @len: length of the list-head, as userspace expects
2493 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2496 if (!futex_cmpxchg_enabled)
2499 * The kernel knows only one size for now:
2501 if (unlikely(len != sizeof(*head)))
2504 current->robust_list = head;
2510 * sys_get_robust_list() - Get the robust-futex list head of a task
2511 * @pid: pid of the process [zero for current task]
2512 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2513 * @len_ptr: pointer to a length field, the kernel fills in the header size
2515 SYSCALL_DEFINE3(get_robust_list, int, pid,
2516 struct robust_list_head __user * __user *, head_ptr,
2517 size_t __user *, len_ptr)
2519 struct robust_list_head __user *head;
2521 struct task_struct *p;
2523 if (!futex_cmpxchg_enabled)
2532 p = find_task_by_vpid(pid);
2538 if (!ptrace_may_access(p, PTRACE_MODE_READ))
2541 head = p->robust_list;
2544 if (put_user(sizeof(*head), len_ptr))
2546 return put_user(head, head_ptr);
2555 * Process a futex-list entry, check whether it's owned by the
2556 * dying task, and do notification if so:
2558 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2560 u32 uval, uninitialized_var(nval), mval;
2563 if (get_user(uval, uaddr))
2566 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2568 * Ok, this dying thread is truly holding a futex
2569 * of interest. Set the OWNER_DIED bit atomically
2570 * via cmpxchg, and if the value had FUTEX_WAITERS
2571 * set, wake up a waiter (if any). (We have to do a
2572 * futex_wake() even if OWNER_DIED is already set -
2573 * to handle the rare but possible case of recursive
2574 * thread-death.) The rest of the cleanup is done in
2577 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2579 * We are not holding a lock here, but we want to have
2580 * the pagefault_disable/enable() protection because
2581 * we want to handle the fault gracefully. If the
2582 * access fails we try to fault in the futex with R/W
2583 * verification via get_user_pages. get_user() above
2584 * does not guarantee R/W access. If that fails we
2585 * give up and leave the futex locked.
2587 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2588 if (fault_in_user_writeable(uaddr))
2596 * Wake robust non-PI futexes here. The wakeup of
2597 * PI futexes happens in exit_pi_state():
2599 if (!pi && (uval & FUTEX_WAITERS))
2600 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2606 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2608 static inline int fetch_robust_entry(struct robust_list __user **entry,
2609 struct robust_list __user * __user *head,
2612 unsigned long uentry;
2614 if (get_user(uentry, (unsigned long __user *)head))
2617 *entry = (void __user *)(uentry & ~1UL);
2624 * Walk curr->robust_list (very carefully, it's a userspace list!)
2625 * and mark any locks found there dead, and notify any waiters.
2627 * We silently return on any sign of list-walking problem.
2629 void exit_robust_list(struct task_struct *curr)
2631 struct robust_list_head __user *head = curr->robust_list;
2632 struct robust_list __user *entry, *next_entry, *pending;
2633 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2634 unsigned int uninitialized_var(next_pi);
2635 unsigned long futex_offset;
2638 if (!futex_cmpxchg_enabled)
2642 * Fetch the list head (which was registered earlier, via
2643 * sys_set_robust_list()):
2645 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2648 * Fetch the relative futex offset:
2650 if (get_user(futex_offset, &head->futex_offset))
2653 * Fetch any possibly pending lock-add first, and handle it
2656 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2659 next_entry = NULL; /* avoid warning with gcc */
2660 while (entry != &head->list) {
2662 * Fetch the next entry in the list before calling
2663 * handle_futex_death:
2665 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2667 * A pending lock might already be on the list, so
2668 * don't process it twice:
2670 if (entry != pending)
2671 if (handle_futex_death((void __user *)entry + futex_offset,
2679 * Avoid excessively long or circular lists:
2688 handle_futex_death((void __user *)pending + futex_offset,
2692 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2693 u32 __user *uaddr2, u32 val2, u32 val3)
2695 int ret = -ENOSYS, cmd = op & FUTEX_CMD_MASK;
2696 unsigned int flags = 0;
2698 if (!(op & FUTEX_PRIVATE_FLAG))
2699 flags |= FLAGS_SHARED;
2701 if (op & FUTEX_CLOCK_REALTIME) {
2702 flags |= FLAGS_CLOCKRT;
2703 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2709 case FUTEX_UNLOCK_PI:
2710 case FUTEX_TRYLOCK_PI:
2711 case FUTEX_WAIT_REQUEUE_PI:
2712 case FUTEX_CMP_REQUEUE_PI:
2713 if (!futex_cmpxchg_enabled)
2719 val3 = FUTEX_BITSET_MATCH_ANY;
2720 case FUTEX_WAIT_BITSET:
2721 ret = futex_wait(uaddr, flags, val, timeout, val3);
2724 val3 = FUTEX_BITSET_MATCH_ANY;
2725 case FUTEX_WAKE_BITSET:
2726 ret = futex_wake(uaddr, flags, val, val3);
2729 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2731 case FUTEX_CMP_REQUEUE:
2732 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2735 ret = futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2738 ret = futex_lock_pi(uaddr, flags, val, timeout, 0);
2740 case FUTEX_UNLOCK_PI:
2741 ret = futex_unlock_pi(uaddr, flags);
2743 case FUTEX_TRYLOCK_PI:
2744 ret = futex_lock_pi(uaddr, flags, 0, timeout, 1);
2746 case FUTEX_WAIT_REQUEUE_PI:
2747 val3 = FUTEX_BITSET_MATCH_ANY;
2748 ret = futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2751 case FUTEX_CMP_REQUEUE_PI:
2752 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2761 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2762 struct timespec __user *, utime, u32 __user *, uaddr2,
2766 ktime_t t, *tp = NULL;
2768 int cmd = op & FUTEX_CMD_MASK;
2770 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2771 cmd == FUTEX_WAIT_BITSET ||
2772 cmd == FUTEX_WAIT_REQUEUE_PI)) {
2773 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2775 if (!timespec_valid(&ts))
2778 t = timespec_to_ktime(ts);
2779 if (cmd == FUTEX_WAIT)
2780 t = ktime_add_safe(ktime_get(), t);
2784 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2785 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2787 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2788 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2789 val2 = (u32) (unsigned long) utime;
2791 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2794 static int __init futex_init(void)
2800 * This will fail and we want it. Some arch implementations do
2801 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2802 * functionality. We want to know that before we call in any
2803 * of the complex code paths. Also we want to prevent
2804 * registration of robust lists in that case. NULL is
2805 * guaranteed to fault and we get -EFAULT on functional
2806 * implementation, the non-functional ones will return
2809 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
2810 futex_cmpxchg_enabled = 1;
2812 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2813 plist_head_init(&futex_queues[i].chain);
2814 spin_lock_init(&futex_queues[i].lock);
2819 __initcall(futex_init);