2 * Fast Userspace Mutexes (which I call "Futexes!").
3 * (C) Rusty Russell, IBM 2002
5 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
8 * Removed page pinning, fix privately mapped COW pages and other cleanups
9 * (C) Copyright 2003, 2004 Jamie Lokier
11 * Robust futex support started by Ingo Molnar
12 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
15 * PI-futex support started by Ingo Molnar and Thomas Gleixner
16 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
19 * PRIVATE futexes by Eric Dumazet
20 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
22 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23 * Copyright (C) IBM Corporation, 2009
24 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
26 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27 * enough at me, Linus for the original (flawed) idea, Matthew
28 * Kirkwood for proof-of-concept implementation.
30 * "The futexes are also cursed."
31 * "But they come in a choice of three flavours!"
33 * This program is free software; you can redistribute it and/or modify
34 * it under the terms of the GNU General Public License as published by
35 * the Free Software Foundation; either version 2 of the License, or
36 * (at your option) any later version.
38 * This program is distributed in the hope that it will be useful,
39 * but WITHOUT ANY WARRANTY; without even the implied warranty of
40 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
41 * GNU General Public License for more details.
43 * You should have received a copy of the GNU General Public License
44 * along with this program; if not, write to the Free Software
45 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
47 #include <linux/slab.h>
48 #include <linux/poll.h>
50 #include <linux/file.h>
51 #include <linux/jhash.h>
52 #include <linux/init.h>
53 #include <linux/futex.h>
54 #include <linux/mount.h>
55 #include <linux/pagemap.h>
56 #include <linux/syscalls.h>
57 #include <linux/signal.h>
58 #include <linux/export.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62 #include <linux/ptrace.h>
63 #include <linux/sched/rt.h>
64 #include <linux/hugetlb.h>
65 #include <linux/freezer.h>
66 #include <linux/bootmem.h>
68 #include <asm/futex.h>
70 #include "locking/rtmutex_common.h"
72 int __read_mostly futex_cmpxchg_enabled;
75 * Futex flags used to encode options to functions and preserve them across
78 #define FLAGS_SHARED 0x01
79 #define FLAGS_CLOCKRT 0x02
80 #define FLAGS_HAS_TIMEOUT 0x04
83 * Priority Inheritance state:
85 struct futex_pi_state {
87 * list of 'owned' pi_state instances - these have to be
88 * cleaned up in do_exit() if the task exits prematurely:
90 struct list_head list;
95 struct rt_mutex pi_mutex;
97 struct task_struct *owner;
104 * struct futex_q - The hashed futex queue entry, one per waiting task
105 * @list: priority-sorted list of tasks waiting on this futex
106 * @task: the task waiting on the futex
107 * @lock_ptr: the hash bucket lock
108 * @key: the key the futex is hashed on
109 * @pi_state: optional priority inheritance state
110 * @rt_waiter: rt_waiter storage for use with requeue_pi
111 * @requeue_pi_key: the requeue_pi target futex key
112 * @bitset: bitset for the optional bitmasked wakeup
114 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
115 * we can wake only the relevant ones (hashed queues may be shared).
117 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
118 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
119 * The order of wakeup is always to make the first condition true, then
122 * PI futexes are typically woken before they are removed from the hash list via
123 * the rt_mutex code. See unqueue_me_pi().
126 struct plist_node list;
128 struct task_struct *task;
129 spinlock_t *lock_ptr;
131 struct futex_pi_state *pi_state;
132 struct rt_mutex_waiter *rt_waiter;
133 union futex_key *requeue_pi_key;
137 static const struct futex_q futex_q_init = {
138 /* list gets initialized in queue_me()*/
139 .key = FUTEX_KEY_INIT,
140 .bitset = FUTEX_BITSET_MATCH_ANY
144 * Hash buckets are shared by all the futex_keys that hash to the same
145 * location. Each key may have multiple futex_q structures, one for each task
146 * waiting on a futex.
148 struct futex_hash_bucket {
150 struct plist_head chain;
151 } ____cacheline_aligned_in_smp;
153 static unsigned long __read_mostly futex_hashsize;
155 static struct futex_hash_bucket *futex_queues;
158 * We hash on the keys returned from get_futex_key (see below).
160 static struct futex_hash_bucket *hash_futex(union futex_key *key)
162 u32 hash = jhash2((u32*)&key->both.word,
163 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
165 return &futex_queues[hash & (futex_hashsize - 1)];
169 * Return 1 if two futex_keys are equal, 0 otherwise.
171 static inline int match_futex(union futex_key *key1, union futex_key *key2)
174 && key1->both.word == key2->both.word
175 && key1->both.ptr == key2->both.ptr
176 && key1->both.offset == key2->both.offset);
180 * Take a reference to the resource addressed by a key.
181 * Can be called while holding spinlocks.
184 static void get_futex_key_refs(union futex_key *key)
189 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
191 ihold(key->shared.inode);
193 case FUT_OFF_MMSHARED:
194 atomic_inc(&key->private.mm->mm_count);
200 * Drop a reference to the resource addressed by a key.
201 * The hash bucket spinlock must not be held.
203 static void drop_futex_key_refs(union futex_key *key)
205 if (!key->both.ptr) {
206 /* If we're here then we tried to put a key we failed to get */
211 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
213 iput(key->shared.inode);
215 case FUT_OFF_MMSHARED:
216 mmdrop(key->private.mm);
222 * get_futex_key() - Get parameters which are the keys for a futex
223 * @uaddr: virtual address of the futex
224 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
225 * @key: address where result is stored.
226 * @rw: mapping needs to be read/write (values: VERIFY_READ,
229 * Return: a negative error code or 0
231 * The key words are stored in *key on success.
233 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
234 * offset_within_page). For private mappings, it's (uaddr, current->mm).
235 * We can usually work out the index without swapping in the page.
237 * lock_page() might sleep, the caller should not hold a spinlock.
240 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
242 unsigned long address = (unsigned long)uaddr;
243 struct mm_struct *mm = current->mm;
244 struct page *page, *page_head;
248 * The futex address must be "naturally" aligned.
250 key->both.offset = address % PAGE_SIZE;
251 if (unlikely((address % sizeof(u32)) != 0))
253 address -= key->both.offset;
255 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
259 * PROCESS_PRIVATE futexes are fast.
260 * As the mm cannot disappear under us and the 'key' only needs
261 * virtual address, we dont even have to find the underlying vma.
262 * Note : We do have to check 'uaddr' is a valid user address,
263 * but access_ok() should be faster than find_vma()
266 key->private.mm = mm;
267 key->private.address = address;
268 get_futex_key_refs(key);
273 err = get_user_pages_fast(address, 1, 1, &page);
275 * If write access is not required (eg. FUTEX_WAIT), try
276 * and get read-only access.
278 if (err == -EFAULT && rw == VERIFY_READ) {
279 err = get_user_pages_fast(address, 1, 0, &page);
287 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
289 if (unlikely(PageTail(page))) {
291 /* serialize against __split_huge_page_splitting() */
293 if (likely(__get_user_pages_fast(address, 1, !ro, &page) == 1)) {
294 page_head = compound_head(page);
296 * page_head is valid pointer but we must pin
297 * it before taking the PG_lock and/or
298 * PG_compound_lock. The moment we re-enable
299 * irqs __split_huge_page_splitting() can
300 * return and the head page can be freed from
301 * under us. We can't take the PG_lock and/or
302 * PG_compound_lock on a page that could be
303 * freed from under us.
305 if (page != page_head) {
316 page_head = compound_head(page);
317 if (page != page_head) {
323 lock_page(page_head);
326 * If page_head->mapping is NULL, then it cannot be a PageAnon
327 * page; but it might be the ZERO_PAGE or in the gate area or
328 * in a special mapping (all cases which we are happy to fail);
329 * or it may have been a good file page when get_user_pages_fast
330 * found it, but truncated or holepunched or subjected to
331 * invalidate_complete_page2 before we got the page lock (also
332 * cases which we are happy to fail). And we hold a reference,
333 * so refcount care in invalidate_complete_page's remove_mapping
334 * prevents drop_caches from setting mapping to NULL beneath us.
336 * The case we do have to guard against is when memory pressure made
337 * shmem_writepage move it from filecache to swapcache beneath us:
338 * an unlikely race, but we do need to retry for page_head->mapping.
340 if (!page_head->mapping) {
341 int shmem_swizzled = PageSwapCache(page_head);
342 unlock_page(page_head);
350 * Private mappings are handled in a simple way.
352 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
353 * it's a read-only handle, it's expected that futexes attach to
354 * the object not the particular process.
356 if (PageAnon(page_head)) {
358 * A RO anonymous page will never change and thus doesn't make
359 * sense for futex operations.
366 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
367 key->private.mm = mm;
368 key->private.address = address;
370 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
371 key->shared.inode = page_head->mapping->host;
372 key->shared.pgoff = basepage_index(page);
375 get_futex_key_refs(key);
378 unlock_page(page_head);
383 static inline void put_futex_key(union futex_key *key)
385 drop_futex_key_refs(key);
389 * fault_in_user_writeable() - Fault in user address and verify RW access
390 * @uaddr: pointer to faulting user space address
392 * Slow path to fixup the fault we just took in the atomic write
395 * We have no generic implementation of a non-destructive write to the
396 * user address. We know that we faulted in the atomic pagefault
397 * disabled section so we can as well avoid the #PF overhead by
398 * calling get_user_pages() right away.
400 static int fault_in_user_writeable(u32 __user *uaddr)
402 struct mm_struct *mm = current->mm;
405 down_read(&mm->mmap_sem);
406 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
408 up_read(&mm->mmap_sem);
410 return ret < 0 ? ret : 0;
414 * futex_top_waiter() - Return the highest priority waiter on a futex
415 * @hb: the hash bucket the futex_q's reside in
416 * @key: the futex key (to distinguish it from other futex futex_q's)
418 * Must be called with the hb lock held.
420 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
421 union futex_key *key)
423 struct futex_q *this;
425 plist_for_each_entry(this, &hb->chain, list) {
426 if (match_futex(&this->key, key))
432 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
433 u32 uval, u32 newval)
438 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
444 static int get_futex_value_locked(u32 *dest, u32 __user *from)
449 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
452 return ret ? -EFAULT : 0;
459 static int refill_pi_state_cache(void)
461 struct futex_pi_state *pi_state;
463 if (likely(current->pi_state_cache))
466 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
471 INIT_LIST_HEAD(&pi_state->list);
472 /* pi_mutex gets initialized later */
473 pi_state->owner = NULL;
474 atomic_set(&pi_state->refcount, 1);
475 pi_state->key = FUTEX_KEY_INIT;
477 current->pi_state_cache = pi_state;
482 static struct futex_pi_state * alloc_pi_state(void)
484 struct futex_pi_state *pi_state = current->pi_state_cache;
487 current->pi_state_cache = NULL;
492 static void free_pi_state(struct futex_pi_state *pi_state)
494 if (!atomic_dec_and_test(&pi_state->refcount))
498 * If pi_state->owner is NULL, the owner is most probably dying
499 * and has cleaned up the pi_state already
501 if (pi_state->owner) {
502 raw_spin_lock_irq(&pi_state->owner->pi_lock);
503 list_del_init(&pi_state->list);
504 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
506 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
509 if (current->pi_state_cache)
513 * pi_state->list is already empty.
514 * clear pi_state->owner.
515 * refcount is at 0 - put it back to 1.
517 pi_state->owner = NULL;
518 atomic_set(&pi_state->refcount, 1);
519 current->pi_state_cache = pi_state;
524 * Look up the task based on what TID userspace gave us.
527 static struct task_struct * futex_find_get_task(pid_t pid)
529 struct task_struct *p;
532 p = find_task_by_vpid(pid);
542 * This task is holding PI mutexes at exit time => bad.
543 * Kernel cleans up PI-state, but userspace is likely hosed.
544 * (Robust-futex cleanup is separate and might save the day for userspace.)
546 void exit_pi_state_list(struct task_struct *curr)
548 struct list_head *next, *head = &curr->pi_state_list;
549 struct futex_pi_state *pi_state;
550 struct futex_hash_bucket *hb;
551 union futex_key key = FUTEX_KEY_INIT;
553 if (!futex_cmpxchg_enabled)
556 * We are a ZOMBIE and nobody can enqueue itself on
557 * pi_state_list anymore, but we have to be careful
558 * versus waiters unqueueing themselves:
560 raw_spin_lock_irq(&curr->pi_lock);
561 while (!list_empty(head)) {
564 pi_state = list_entry(next, struct futex_pi_state, list);
566 hb = hash_futex(&key);
567 raw_spin_unlock_irq(&curr->pi_lock);
569 spin_lock(&hb->lock);
571 raw_spin_lock_irq(&curr->pi_lock);
573 * We dropped the pi-lock, so re-check whether this
574 * task still owns the PI-state:
576 if (head->next != next) {
577 spin_unlock(&hb->lock);
581 WARN_ON(pi_state->owner != curr);
582 WARN_ON(list_empty(&pi_state->list));
583 list_del_init(&pi_state->list);
584 pi_state->owner = NULL;
585 raw_spin_unlock_irq(&curr->pi_lock);
587 rt_mutex_unlock(&pi_state->pi_mutex);
589 spin_unlock(&hb->lock);
591 raw_spin_lock_irq(&curr->pi_lock);
593 raw_spin_unlock_irq(&curr->pi_lock);
597 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
598 union futex_key *key, struct futex_pi_state **ps)
600 struct futex_pi_state *pi_state = NULL;
601 struct futex_q *this, *next;
602 struct task_struct *p;
603 pid_t pid = uval & FUTEX_TID_MASK;
605 plist_for_each_entry_safe(this, next, &hb->chain, list) {
606 if (match_futex(&this->key, key)) {
608 * Another waiter already exists - bump up
609 * the refcount and return its pi_state:
611 pi_state = this->pi_state;
613 * Userspace might have messed up non-PI and PI futexes
615 if (unlikely(!pi_state))
618 WARN_ON(!atomic_read(&pi_state->refcount));
621 * When pi_state->owner is NULL then the owner died
622 * and another waiter is on the fly. pi_state->owner
623 * is fixed up by the task which acquires
624 * pi_state->rt_mutex.
626 * We do not check for pid == 0 which can happen when
627 * the owner died and robust_list_exit() cleared the
630 if (pid && pi_state->owner) {
632 * Bail out if user space manipulated the
635 if (pid != task_pid_vnr(pi_state->owner))
639 atomic_inc(&pi_state->refcount);
647 * We are the first waiter - try to look up the real owner and attach
648 * the new pi_state to it, but bail out when TID = 0
652 p = futex_find_get_task(pid);
657 * We need to look at the task state flags to figure out,
658 * whether the task is exiting. To protect against the do_exit
659 * change of the task flags, we do this protected by
662 raw_spin_lock_irq(&p->pi_lock);
663 if (unlikely(p->flags & PF_EXITING)) {
665 * The task is on the way out. When PF_EXITPIDONE is
666 * set, we know that the task has finished the
669 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
671 raw_spin_unlock_irq(&p->pi_lock);
676 pi_state = alloc_pi_state();
679 * Initialize the pi_mutex in locked state and make 'p'
682 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
684 /* Store the key for possible exit cleanups: */
685 pi_state->key = *key;
687 WARN_ON(!list_empty(&pi_state->list));
688 list_add(&pi_state->list, &p->pi_state_list);
690 raw_spin_unlock_irq(&p->pi_lock);
700 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
701 * @uaddr: the pi futex user address
702 * @hb: the pi futex hash bucket
703 * @key: the futex key associated with uaddr and hb
704 * @ps: the pi_state pointer where we store the result of the
706 * @task: the task to perform the atomic lock work for. This will
707 * be "current" except in the case of requeue pi.
708 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
712 * 1 - acquired the lock;
715 * The hb->lock and futex_key refs shall be held by the caller.
717 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
718 union futex_key *key,
719 struct futex_pi_state **ps,
720 struct task_struct *task, int set_waiters)
722 int lock_taken, ret, force_take = 0;
723 u32 uval, newval, curval, vpid = task_pid_vnr(task);
726 ret = lock_taken = 0;
729 * To avoid races, we attempt to take the lock here again
730 * (by doing a 0 -> TID atomic cmpxchg), while holding all
731 * the locks. It will most likely not succeed.
735 newval |= FUTEX_WAITERS;
737 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, 0, newval)))
743 if ((unlikely((curval & FUTEX_TID_MASK) == vpid)))
747 * Surprise - we got the lock. Just return to userspace:
749 if (unlikely(!curval))
755 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
756 * to wake at the next unlock.
758 newval = curval | FUTEX_WAITERS;
761 * Should we force take the futex? See below.
763 if (unlikely(force_take)) {
765 * Keep the OWNER_DIED and the WAITERS bit and set the
768 newval = (curval & ~FUTEX_TID_MASK) | vpid;
773 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
775 if (unlikely(curval != uval))
779 * We took the lock due to forced take over.
781 if (unlikely(lock_taken))
785 * We dont have the lock. Look up the PI state (or create it if
786 * we are the first waiter):
788 ret = lookup_pi_state(uval, hb, key, ps);
794 * We failed to find an owner for this
795 * futex. So we have no pi_state to block
796 * on. This can happen in two cases:
799 * 2) A stale FUTEX_WAITERS bit
801 * Re-read the futex value.
803 if (get_futex_value_locked(&curval, uaddr))
807 * If the owner died or we have a stale
808 * WAITERS bit the owner TID in the user space
811 if (!(curval & FUTEX_TID_MASK)) {
824 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
825 * @q: The futex_q to unqueue
827 * The q->lock_ptr must not be NULL and must be held by the caller.
829 static void __unqueue_futex(struct futex_q *q)
831 struct futex_hash_bucket *hb;
833 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
834 || WARN_ON(plist_node_empty(&q->list)))
837 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
838 plist_del(&q->list, &hb->chain);
842 * The hash bucket lock must be held when this is called.
843 * Afterwards, the futex_q must not be accessed.
845 static void wake_futex(struct futex_q *q)
847 struct task_struct *p = q->task;
849 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
853 * We set q->lock_ptr = NULL _before_ we wake up the task. If
854 * a non-futex wake up happens on another CPU then the task
855 * might exit and p would dereference a non-existing task
856 * struct. Prevent this by holding a reference on p across the
863 * The waiting task can free the futex_q as soon as
864 * q->lock_ptr = NULL is written, without taking any locks. A
865 * memory barrier is required here to prevent the following
866 * store to lock_ptr from getting ahead of the plist_del.
871 wake_up_state(p, TASK_NORMAL);
875 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
877 struct task_struct *new_owner;
878 struct futex_pi_state *pi_state = this->pi_state;
879 u32 uninitialized_var(curval), newval;
885 * If current does not own the pi_state then the futex is
886 * inconsistent and user space fiddled with the futex value.
888 if (pi_state->owner != current)
891 raw_spin_lock(&pi_state->pi_mutex.wait_lock);
892 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
895 * It is possible that the next waiter (the one that brought
896 * this owner to the kernel) timed out and is no longer
897 * waiting on the lock.
900 new_owner = this->task;
903 * We pass it to the next owner. (The WAITERS bit is always
904 * kept enabled while there is PI state around. We must also
905 * preserve the owner died bit.)
907 if (!(uval & FUTEX_OWNER_DIED)) {
910 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
912 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
914 else if (curval != uval)
917 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
922 raw_spin_lock_irq(&pi_state->owner->pi_lock);
923 WARN_ON(list_empty(&pi_state->list));
924 list_del_init(&pi_state->list);
925 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
927 raw_spin_lock_irq(&new_owner->pi_lock);
928 WARN_ON(!list_empty(&pi_state->list));
929 list_add(&pi_state->list, &new_owner->pi_state_list);
930 pi_state->owner = new_owner;
931 raw_spin_unlock_irq(&new_owner->pi_lock);
933 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
934 rt_mutex_unlock(&pi_state->pi_mutex);
939 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
941 u32 uninitialized_var(oldval);
944 * There is no waiter, so we unlock the futex. The owner died
945 * bit has not to be preserved here. We are the owner:
947 if (cmpxchg_futex_value_locked(&oldval, uaddr, uval, 0))
956 * Express the locking dependencies for lockdep:
959 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
962 spin_lock(&hb1->lock);
964 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
965 } else { /* hb1 > hb2 */
966 spin_lock(&hb2->lock);
967 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
972 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
974 spin_unlock(&hb1->lock);
976 spin_unlock(&hb2->lock);
980 * Wake up waiters matching bitset queued on this futex (uaddr).
983 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
985 struct futex_hash_bucket *hb;
986 struct futex_q *this, *next;
987 union futex_key key = FUTEX_KEY_INIT;
993 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
994 if (unlikely(ret != 0))
997 hb = hash_futex(&key);
998 spin_lock(&hb->lock);
1000 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1001 if (match_futex (&this->key, &key)) {
1002 if (this->pi_state || this->rt_waiter) {
1007 /* Check if one of the bits is set in both bitsets */
1008 if (!(this->bitset & bitset))
1012 if (++ret >= nr_wake)
1017 spin_unlock(&hb->lock);
1018 put_futex_key(&key);
1024 * Wake up all waiters hashed on the physical page that is mapped
1025 * to this virtual address:
1028 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1029 int nr_wake, int nr_wake2, int op)
1031 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1032 struct futex_hash_bucket *hb1, *hb2;
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);
1080 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1081 if (match_futex (&this->key, &key1)) {
1082 if (this->pi_state || this->rt_waiter) {
1087 if (++ret >= nr_wake)
1094 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1095 if (match_futex (&this->key, &key2)) {
1096 if (this->pi_state || this->rt_waiter) {
1101 if (++op_ret >= nr_wake2)
1109 double_unlock_hb(hb1, hb2);
1111 put_futex_key(&key2);
1113 put_futex_key(&key1);
1119 * requeue_futex() - Requeue a futex_q from one hb to another
1120 * @q: the futex_q to requeue
1121 * @hb1: the source hash_bucket
1122 * @hb2: the target hash_bucket
1123 * @key2: the new key for the requeued futex_q
1126 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1127 struct futex_hash_bucket *hb2, union futex_key *key2)
1131 * If key1 and key2 hash to the same bucket, no need to
1134 if (likely(&hb1->chain != &hb2->chain)) {
1135 plist_del(&q->list, &hb1->chain);
1136 plist_add(&q->list, &hb2->chain);
1137 q->lock_ptr = &hb2->lock;
1139 get_futex_key_refs(key2);
1144 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1146 * @key: the key of the requeue target futex
1147 * @hb: the hash_bucket of the requeue target futex
1149 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1150 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1151 * to the requeue target futex so the waiter can detect the wakeup on the right
1152 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1153 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1154 * to protect access to the pi_state to fixup the owner later. Must be called
1155 * with both q->lock_ptr and hb->lock held.
1158 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1159 struct futex_hash_bucket *hb)
1161 get_futex_key_refs(key);
1166 WARN_ON(!q->rt_waiter);
1167 q->rt_waiter = NULL;
1169 q->lock_ptr = &hb->lock;
1171 wake_up_state(q->task, TASK_NORMAL);
1175 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1176 * @pifutex: the user address of the to futex
1177 * @hb1: the from futex hash bucket, must be locked by the caller
1178 * @hb2: the to futex hash bucket, must be locked by the caller
1179 * @key1: the from futex key
1180 * @key2: the to futex key
1181 * @ps: address to store the pi_state pointer
1182 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1184 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1185 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1186 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1187 * hb1 and hb2 must be held by the caller.
1190 * 0 - failed to acquire the lock atomically;
1191 * 1 - acquired the lock;
1194 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1195 struct futex_hash_bucket *hb1,
1196 struct futex_hash_bucket *hb2,
1197 union futex_key *key1, union futex_key *key2,
1198 struct futex_pi_state **ps, int set_waiters)
1200 struct futex_q *top_waiter = NULL;
1204 if (get_futex_value_locked(&curval, pifutex))
1208 * Find the top_waiter and determine if there are additional waiters.
1209 * If the caller intends to requeue more than 1 waiter to pifutex,
1210 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1211 * as we have means to handle the possible fault. If not, don't set
1212 * the bit unecessarily as it will force the subsequent unlock to enter
1215 top_waiter = futex_top_waiter(hb1, key1);
1217 /* There are no waiters, nothing for us to do. */
1221 /* Ensure we requeue to the expected futex. */
1222 if (!match_futex(top_waiter->requeue_pi_key, key2))
1226 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1227 * the contended case or if set_waiters is 1. The pi_state is returned
1228 * in ps in contended cases.
1230 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1233 requeue_pi_wake_futex(top_waiter, key2, hb2);
1239 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1240 * @uaddr1: source futex user address
1241 * @flags: futex flags (FLAGS_SHARED, etc.)
1242 * @uaddr2: target futex user address
1243 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1244 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1245 * @cmpval: @uaddr1 expected value (or %NULL)
1246 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1247 * pi futex (pi to pi requeue is not supported)
1249 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1250 * uaddr2 atomically on behalf of the top waiter.
1253 * >=0 - on success, the number of tasks requeued or woken;
1256 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1257 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1258 u32 *cmpval, int requeue_pi)
1260 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1261 int drop_count = 0, task_count = 0, ret;
1262 struct futex_pi_state *pi_state = NULL;
1263 struct futex_hash_bucket *hb1, *hb2;
1264 struct futex_q *this, *next;
1269 * requeue_pi requires a pi_state, try to allocate it now
1270 * without any locks in case it fails.
1272 if (refill_pi_state_cache())
1275 * requeue_pi must wake as many tasks as it can, up to nr_wake
1276 * + nr_requeue, since it acquires the rt_mutex prior to
1277 * returning to userspace, so as to not leave the rt_mutex with
1278 * waiters and no owner. However, second and third wake-ups
1279 * cannot be predicted as they involve race conditions with the
1280 * first wake and a fault while looking up the pi_state. Both
1281 * pthread_cond_signal() and pthread_cond_broadcast() should
1289 if (pi_state != NULL) {
1291 * We will have to lookup the pi_state again, so free this one
1292 * to keep the accounting correct.
1294 free_pi_state(pi_state);
1298 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1299 if (unlikely(ret != 0))
1301 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1302 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1303 if (unlikely(ret != 0))
1306 hb1 = hash_futex(&key1);
1307 hb2 = hash_futex(&key2);
1310 double_lock_hb(hb1, hb2);
1312 if (likely(cmpval != NULL)) {
1315 ret = get_futex_value_locked(&curval, uaddr1);
1317 if (unlikely(ret)) {
1318 double_unlock_hb(hb1, hb2);
1320 ret = get_user(curval, uaddr1);
1324 if (!(flags & FLAGS_SHARED))
1327 put_futex_key(&key2);
1328 put_futex_key(&key1);
1331 if (curval != *cmpval) {
1337 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1339 * Attempt to acquire uaddr2 and wake the top waiter. If we
1340 * intend to requeue waiters, force setting the FUTEX_WAITERS
1341 * bit. We force this here where we are able to easily handle
1342 * faults rather in the requeue loop below.
1344 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1345 &key2, &pi_state, nr_requeue);
1348 * At this point the top_waiter has either taken uaddr2 or is
1349 * waiting on it. If the former, then the pi_state will not
1350 * exist yet, look it up one more time to ensure we have a
1357 ret = get_futex_value_locked(&curval2, uaddr2);
1359 ret = lookup_pi_state(curval2, hb2, &key2,
1367 double_unlock_hb(hb1, hb2);
1368 put_futex_key(&key2);
1369 put_futex_key(&key1);
1370 ret = fault_in_user_writeable(uaddr2);
1375 /* The owner was exiting, try again. */
1376 double_unlock_hb(hb1, hb2);
1377 put_futex_key(&key2);
1378 put_futex_key(&key1);
1386 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1387 if (task_count - nr_wake >= nr_requeue)
1390 if (!match_futex(&this->key, &key1))
1394 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1395 * be paired with each other and no other futex ops.
1397 * We should never be requeueing a futex_q with a pi_state,
1398 * which is awaiting a futex_unlock_pi().
1400 if ((requeue_pi && !this->rt_waiter) ||
1401 (!requeue_pi && this->rt_waiter) ||
1408 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1409 * lock, we already woke the top_waiter. If not, it will be
1410 * woken by futex_unlock_pi().
1412 if (++task_count <= nr_wake && !requeue_pi) {
1417 /* Ensure we requeue to the expected futex for requeue_pi. */
1418 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1424 * Requeue nr_requeue waiters and possibly one more in the case
1425 * of requeue_pi if we couldn't acquire the lock atomically.
1428 /* Prepare the waiter to take the rt_mutex. */
1429 atomic_inc(&pi_state->refcount);
1430 this->pi_state = pi_state;
1431 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1435 /* We got the lock. */
1436 requeue_pi_wake_futex(this, &key2, hb2);
1441 this->pi_state = NULL;
1442 free_pi_state(pi_state);
1446 requeue_futex(this, hb1, hb2, &key2);
1451 double_unlock_hb(hb1, hb2);
1454 * drop_futex_key_refs() must be called outside the spinlocks. During
1455 * the requeue we moved futex_q's from the hash bucket at key1 to the
1456 * one at key2 and updated their key pointer. We no longer need to
1457 * hold the references to key1.
1459 while (--drop_count >= 0)
1460 drop_futex_key_refs(&key1);
1463 put_futex_key(&key2);
1465 put_futex_key(&key1);
1467 if (pi_state != NULL)
1468 free_pi_state(pi_state);
1469 return ret ? ret : task_count;
1472 /* The key must be already stored in q->key. */
1473 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1474 __acquires(&hb->lock)
1476 struct futex_hash_bucket *hb;
1478 hb = hash_futex(&q->key);
1479 q->lock_ptr = &hb->lock;
1481 spin_lock(&hb->lock);
1486 queue_unlock(struct futex_hash_bucket *hb)
1487 __releases(&hb->lock)
1489 spin_unlock(&hb->lock);
1493 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1494 * @q: The futex_q to enqueue
1495 * @hb: The destination hash bucket
1497 * The hb->lock must be held by the caller, and is released here. A call to
1498 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1499 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1500 * or nothing if the unqueue is done as part of the wake process and the unqueue
1501 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1504 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1505 __releases(&hb->lock)
1510 * The priority used to register this element is
1511 * - either the real thread-priority for the real-time threads
1512 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1513 * - or MAX_RT_PRIO for non-RT threads.
1514 * Thus, all RT-threads are woken first in priority order, and
1515 * the others are woken last, in FIFO order.
1517 prio = min(current->normal_prio, MAX_RT_PRIO);
1519 plist_node_init(&q->list, prio);
1520 plist_add(&q->list, &hb->chain);
1522 spin_unlock(&hb->lock);
1526 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1527 * @q: The futex_q to unqueue
1529 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1530 * be paired with exactly one earlier call to queue_me().
1533 * 1 - if the futex_q was still queued (and we removed unqueued it);
1534 * 0 - if the futex_q was already removed by the waking thread
1536 static int unqueue_me(struct futex_q *q)
1538 spinlock_t *lock_ptr;
1541 /* In the common case we don't take the spinlock, which is nice. */
1543 lock_ptr = q->lock_ptr;
1545 if (lock_ptr != NULL) {
1546 spin_lock(lock_ptr);
1548 * q->lock_ptr can change between reading it and
1549 * spin_lock(), causing us to take the wrong lock. This
1550 * corrects the race condition.
1552 * Reasoning goes like this: if we have the wrong lock,
1553 * q->lock_ptr must have changed (maybe several times)
1554 * between reading it and the spin_lock(). It can
1555 * change again after the spin_lock() but only if it was
1556 * already changed before the spin_lock(). It cannot,
1557 * however, change back to the original value. Therefore
1558 * we can detect whether we acquired the correct lock.
1560 if (unlikely(lock_ptr != q->lock_ptr)) {
1561 spin_unlock(lock_ptr);
1566 BUG_ON(q->pi_state);
1568 spin_unlock(lock_ptr);
1572 drop_futex_key_refs(&q->key);
1577 * PI futexes can not be requeued and must remove themself from the
1578 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1581 static void unqueue_me_pi(struct futex_q *q)
1582 __releases(q->lock_ptr)
1586 BUG_ON(!q->pi_state);
1587 free_pi_state(q->pi_state);
1590 spin_unlock(q->lock_ptr);
1594 * Fixup the pi_state owner with the new owner.
1596 * Must be called with hash bucket lock held and mm->sem held for non
1599 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1600 struct task_struct *newowner)
1602 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1603 struct futex_pi_state *pi_state = q->pi_state;
1604 struct task_struct *oldowner = pi_state->owner;
1605 u32 uval, uninitialized_var(curval), newval;
1609 if (!pi_state->owner)
1610 newtid |= FUTEX_OWNER_DIED;
1613 * We are here either because we stole the rtmutex from the
1614 * previous highest priority waiter or we are the highest priority
1615 * waiter but failed to get the rtmutex the first time.
1616 * We have to replace the newowner TID in the user space variable.
1617 * This must be atomic as we have to preserve the owner died bit here.
1619 * Note: We write the user space value _before_ changing the pi_state
1620 * because we can fault here. Imagine swapped out pages or a fork
1621 * that marked all the anonymous memory readonly for cow.
1623 * Modifying pi_state _before_ the user space value would
1624 * leave the pi_state in an inconsistent state when we fault
1625 * here, because we need to drop the hash bucket lock to
1626 * handle the fault. This might be observed in the PID check
1627 * in lookup_pi_state.
1630 if (get_futex_value_locked(&uval, uaddr))
1634 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1636 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1644 * We fixed up user space. Now we need to fix the pi_state
1647 if (pi_state->owner != NULL) {
1648 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1649 WARN_ON(list_empty(&pi_state->list));
1650 list_del_init(&pi_state->list);
1651 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1654 pi_state->owner = newowner;
1656 raw_spin_lock_irq(&newowner->pi_lock);
1657 WARN_ON(!list_empty(&pi_state->list));
1658 list_add(&pi_state->list, &newowner->pi_state_list);
1659 raw_spin_unlock_irq(&newowner->pi_lock);
1663 * To handle the page fault we need to drop the hash bucket
1664 * lock here. That gives the other task (either the highest priority
1665 * waiter itself or the task which stole the rtmutex) the
1666 * chance to try the fixup of the pi_state. So once we are
1667 * back from handling the fault we need to check the pi_state
1668 * after reacquiring the hash bucket lock and before trying to
1669 * do another fixup. When the fixup has been done already we
1673 spin_unlock(q->lock_ptr);
1675 ret = fault_in_user_writeable(uaddr);
1677 spin_lock(q->lock_ptr);
1680 * Check if someone else fixed it for us:
1682 if (pi_state->owner != oldowner)
1691 static long futex_wait_restart(struct restart_block *restart);
1694 * fixup_owner() - Post lock pi_state and corner case management
1695 * @uaddr: user address of the futex
1696 * @q: futex_q (contains pi_state and access to the rt_mutex)
1697 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1699 * After attempting to lock an rt_mutex, this function is called to cleanup
1700 * the pi_state owner as well as handle race conditions that may allow us to
1701 * acquire the lock. Must be called with the hb lock held.
1704 * 1 - success, lock taken;
1705 * 0 - success, lock not taken;
1706 * <0 - on error (-EFAULT)
1708 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
1710 struct task_struct *owner;
1715 * Got the lock. We might not be the anticipated owner if we
1716 * did a lock-steal - fix up the PI-state in that case:
1718 if (q->pi_state->owner != current)
1719 ret = fixup_pi_state_owner(uaddr, q, current);
1724 * Catch the rare case, where the lock was released when we were on the
1725 * way back before we locked the hash bucket.
1727 if (q->pi_state->owner == current) {
1729 * Try to get the rt_mutex now. This might fail as some other
1730 * task acquired the rt_mutex after we removed ourself from the
1731 * rt_mutex waiters list.
1733 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1739 * pi_state is incorrect, some other task did a lock steal and
1740 * we returned due to timeout or signal without taking the
1741 * rt_mutex. Too late.
1743 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
1744 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1746 owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
1747 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
1748 ret = fixup_pi_state_owner(uaddr, q, owner);
1753 * Paranoia check. If we did not take the lock, then we should not be
1754 * the owner of the rt_mutex.
1756 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1757 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1758 "pi-state %p\n", ret,
1759 q->pi_state->pi_mutex.owner,
1760 q->pi_state->owner);
1763 return ret ? ret : locked;
1767 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1768 * @hb: the futex hash bucket, must be locked by the caller
1769 * @q: the futex_q to queue up on
1770 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1772 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1773 struct hrtimer_sleeper *timeout)
1776 * The task state is guaranteed to be set before another task can
1777 * wake it. set_current_state() is implemented using set_mb() and
1778 * queue_me() calls spin_unlock() upon completion, both serializing
1779 * access to the hash list and forcing another memory barrier.
1781 set_current_state(TASK_INTERRUPTIBLE);
1786 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1787 if (!hrtimer_active(&timeout->timer))
1788 timeout->task = NULL;
1792 * If we have been removed from the hash list, then another task
1793 * has tried to wake us, and we can skip the call to schedule().
1795 if (likely(!plist_node_empty(&q->list))) {
1797 * If the timer has already expired, current will already be
1798 * flagged for rescheduling. Only call schedule if there
1799 * is no timeout, or if it has yet to expire.
1801 if (!timeout || timeout->task)
1802 freezable_schedule();
1804 __set_current_state(TASK_RUNNING);
1808 * futex_wait_setup() - Prepare to wait on a futex
1809 * @uaddr: the futex userspace address
1810 * @val: the expected value
1811 * @flags: futex flags (FLAGS_SHARED, etc.)
1812 * @q: the associated futex_q
1813 * @hb: storage for hash_bucket pointer to be returned to caller
1815 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1816 * compare it with the expected value. Handle atomic faults internally.
1817 * Return with the hb lock held and a q.key reference on success, and unlocked
1818 * with no q.key reference on failure.
1821 * 0 - uaddr contains val and hb has been locked;
1822 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
1824 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
1825 struct futex_q *q, struct futex_hash_bucket **hb)
1831 * Access the page AFTER the hash-bucket is locked.
1832 * Order is important:
1834 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1835 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1837 * The basic logical guarantee of a futex is that it blocks ONLY
1838 * if cond(var) is known to be true at the time of blocking, for
1839 * any cond. If we locked the hash-bucket after testing *uaddr, that
1840 * would open a race condition where we could block indefinitely with
1841 * cond(var) false, which would violate the guarantee.
1843 * On the other hand, we insert q and release the hash-bucket only
1844 * after testing *uaddr. This guarantees that futex_wait() will NOT
1845 * absorb a wakeup if *uaddr does not match the desired values
1846 * while the syscall executes.
1849 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
1850 if (unlikely(ret != 0))
1854 *hb = queue_lock(q);
1856 ret = get_futex_value_locked(&uval, uaddr);
1861 ret = get_user(uval, uaddr);
1865 if (!(flags & FLAGS_SHARED))
1868 put_futex_key(&q->key);
1879 put_futex_key(&q->key);
1883 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
1884 ktime_t *abs_time, u32 bitset)
1886 struct hrtimer_sleeper timeout, *to = NULL;
1887 struct restart_block *restart;
1888 struct futex_hash_bucket *hb;
1889 struct futex_q q = futex_q_init;
1899 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
1900 CLOCK_REALTIME : CLOCK_MONOTONIC,
1902 hrtimer_init_sleeper(to, current);
1903 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1904 current->timer_slack_ns);
1909 * Prepare to wait on uaddr. On success, holds hb lock and increments
1912 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
1916 /* queue_me and wait for wakeup, timeout, or a signal. */
1917 futex_wait_queue_me(hb, &q, to);
1919 /* If we were woken (and unqueued), we succeeded, whatever. */
1921 /* unqueue_me() drops q.key ref */
1922 if (!unqueue_me(&q))
1925 if (to && !to->task)
1929 * We expect signal_pending(current), but we might be the
1930 * victim of a spurious wakeup as well.
1932 if (!signal_pending(current))
1939 restart = ¤t_thread_info()->restart_block;
1940 restart->fn = futex_wait_restart;
1941 restart->futex.uaddr = uaddr;
1942 restart->futex.val = val;
1943 restart->futex.time = abs_time->tv64;
1944 restart->futex.bitset = bitset;
1945 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
1947 ret = -ERESTART_RESTARTBLOCK;
1951 hrtimer_cancel(&to->timer);
1952 destroy_hrtimer_on_stack(&to->timer);
1958 static long futex_wait_restart(struct restart_block *restart)
1960 u32 __user *uaddr = restart->futex.uaddr;
1961 ktime_t t, *tp = NULL;
1963 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1964 t.tv64 = restart->futex.time;
1967 restart->fn = do_no_restart_syscall;
1969 return (long)futex_wait(uaddr, restart->futex.flags,
1970 restart->futex.val, tp, restart->futex.bitset);
1975 * Userspace tried a 0 -> TID atomic transition of the futex value
1976 * and failed. The kernel side here does the whole locking operation:
1977 * if there are waiters then it will block, it does PI, etc. (Due to
1978 * races the kernel might see a 0 value of the futex too.)
1980 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
1981 ktime_t *time, int trylock)
1983 struct hrtimer_sleeper timeout, *to = NULL;
1984 struct futex_hash_bucket *hb;
1985 struct futex_q q = futex_q_init;
1988 if (refill_pi_state_cache())
1993 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1995 hrtimer_init_sleeper(to, current);
1996 hrtimer_set_expires(&to->timer, *time);
2000 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2001 if (unlikely(ret != 0))
2005 hb = queue_lock(&q);
2007 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2008 if (unlikely(ret)) {
2011 /* We got the lock. */
2013 goto out_unlock_put_key;
2018 * Task is exiting and we just wait for the
2022 put_futex_key(&q.key);
2026 goto out_unlock_put_key;
2031 * Only actually queue now that the atomic ops are done:
2035 WARN_ON(!q.pi_state);
2037 * Block on the PI mutex:
2040 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
2042 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2043 /* Fixup the trylock return value: */
2044 ret = ret ? 0 : -EWOULDBLOCK;
2047 spin_lock(q.lock_ptr);
2049 * Fixup the pi_state owner and possibly acquire the lock if we
2052 res = fixup_owner(uaddr, &q, !ret);
2054 * If fixup_owner() returned an error, proprogate that. If it acquired
2055 * the lock, clear our -ETIMEDOUT or -EINTR.
2058 ret = (res < 0) ? res : 0;
2061 * If fixup_owner() faulted and was unable to handle the fault, unlock
2062 * it and return the fault to userspace.
2064 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2065 rt_mutex_unlock(&q.pi_state->pi_mutex);
2067 /* Unqueue and drop the lock */
2076 put_futex_key(&q.key);
2079 destroy_hrtimer_on_stack(&to->timer);
2080 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2085 ret = fault_in_user_writeable(uaddr);
2089 if (!(flags & FLAGS_SHARED))
2092 put_futex_key(&q.key);
2097 * Userspace attempted a TID -> 0 atomic transition, and failed.
2098 * This is the in-kernel slowpath: we look up the PI state (if any),
2099 * and do the rt-mutex unlock.
2101 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2103 struct futex_hash_bucket *hb;
2104 struct futex_q *this, *next;
2105 union futex_key key = FUTEX_KEY_INIT;
2106 u32 uval, vpid = task_pid_vnr(current);
2110 if (get_user(uval, uaddr))
2113 * We release only a lock we actually own:
2115 if ((uval & FUTEX_TID_MASK) != vpid)
2118 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2119 if (unlikely(ret != 0))
2122 hb = hash_futex(&key);
2123 spin_lock(&hb->lock);
2126 * To avoid races, try to do the TID -> 0 atomic transition
2127 * again. If it succeeds then we can return without waking
2130 if (!(uval & FUTEX_OWNER_DIED) &&
2131 cmpxchg_futex_value_locked(&uval, uaddr, vpid, 0))
2134 * Rare case: we managed to release the lock atomically,
2135 * no need to wake anyone else up:
2137 if (unlikely(uval == vpid))
2141 * Ok, other tasks may need to be woken up - check waiters
2142 * and do the wakeup if necessary:
2144 plist_for_each_entry_safe(this, next, &hb->chain, list) {
2145 if (!match_futex (&this->key, &key))
2147 ret = wake_futex_pi(uaddr, uval, this);
2149 * The atomic access to the futex value
2150 * generated a pagefault, so retry the
2151 * user-access and the wakeup:
2158 * No waiters - kernel unlocks the futex:
2160 if (!(uval & FUTEX_OWNER_DIED)) {
2161 ret = unlock_futex_pi(uaddr, uval);
2167 spin_unlock(&hb->lock);
2168 put_futex_key(&key);
2174 spin_unlock(&hb->lock);
2175 put_futex_key(&key);
2177 ret = fault_in_user_writeable(uaddr);
2185 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2186 * @hb: the hash_bucket futex_q was original enqueued on
2187 * @q: the futex_q woken while waiting to be requeued
2188 * @key2: the futex_key of the requeue target futex
2189 * @timeout: the timeout associated with the wait (NULL if none)
2191 * Detect if the task was woken on the initial futex as opposed to the requeue
2192 * target futex. If so, determine if it was a timeout or a signal that caused
2193 * the wakeup and return the appropriate error code to the caller. Must be
2194 * called with the hb lock held.
2197 * 0 = no early wakeup detected;
2198 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2201 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2202 struct futex_q *q, union futex_key *key2,
2203 struct hrtimer_sleeper *timeout)
2208 * With the hb lock held, we avoid races while we process the wakeup.
2209 * We only need to hold hb (and not hb2) to ensure atomicity as the
2210 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2211 * It can't be requeued from uaddr2 to something else since we don't
2212 * support a PI aware source futex for requeue.
2214 if (!match_futex(&q->key, key2)) {
2215 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2217 * We were woken prior to requeue by a timeout or a signal.
2218 * Unqueue the futex_q and determine which it was.
2220 plist_del(&q->list, &hb->chain);
2222 /* Handle spurious wakeups gracefully */
2224 if (timeout && !timeout->task)
2226 else if (signal_pending(current))
2227 ret = -ERESTARTNOINTR;
2233 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2234 * @uaddr: the futex we initially wait on (non-pi)
2235 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2236 * the same type, no requeueing from private to shared, etc.
2237 * @val: the expected value of uaddr
2238 * @abs_time: absolute timeout
2239 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2240 * @uaddr2: the pi futex we will take prior to returning to user-space
2242 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2243 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2244 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2245 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2246 * without one, the pi logic would not know which task to boost/deboost, if
2247 * there was a need to.
2249 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2250 * via the following--
2251 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2252 * 2) wakeup on uaddr2 after a requeue
2256 * If 3, cleanup and return -ERESTARTNOINTR.
2258 * If 2, we may then block on trying to take the rt_mutex and return via:
2259 * 5) successful lock
2262 * 8) other lock acquisition failure
2264 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2266 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2272 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2273 u32 val, ktime_t *abs_time, u32 bitset,
2276 struct hrtimer_sleeper timeout, *to = NULL;
2277 struct rt_mutex_waiter rt_waiter;
2278 struct rt_mutex *pi_mutex = NULL;
2279 struct futex_hash_bucket *hb;
2280 union futex_key key2 = FUTEX_KEY_INIT;
2281 struct futex_q q = futex_q_init;
2284 if (uaddr == uaddr2)
2292 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2293 CLOCK_REALTIME : CLOCK_MONOTONIC,
2295 hrtimer_init_sleeper(to, current);
2296 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2297 current->timer_slack_ns);
2301 * The waiter is allocated on our stack, manipulated by the requeue
2302 * code while we sleep on uaddr.
2304 debug_rt_mutex_init_waiter(&rt_waiter);
2305 rt_waiter.task = NULL;
2307 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2308 if (unlikely(ret != 0))
2312 q.rt_waiter = &rt_waiter;
2313 q.requeue_pi_key = &key2;
2316 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2319 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2323 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2324 futex_wait_queue_me(hb, &q, to);
2326 spin_lock(&hb->lock);
2327 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2328 spin_unlock(&hb->lock);
2333 * In order for us to be here, we know our q.key == key2, and since
2334 * we took the hb->lock above, we also know that futex_requeue() has
2335 * completed and we no longer have to concern ourselves with a wakeup
2336 * race with the atomic proxy lock acquisition by the requeue code. The
2337 * futex_requeue dropped our key1 reference and incremented our key2
2341 /* Check if the requeue code acquired the second futex for us. */
2344 * Got the lock. We might not be the anticipated owner if we
2345 * did a lock-steal - fix up the PI-state in that case.
2347 if (q.pi_state && (q.pi_state->owner != current)) {
2348 spin_lock(q.lock_ptr);
2349 ret = fixup_pi_state_owner(uaddr2, &q, current);
2350 spin_unlock(q.lock_ptr);
2354 * We have been woken up by futex_unlock_pi(), a timeout, or a
2355 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2358 WARN_ON(!q.pi_state);
2359 pi_mutex = &q.pi_state->pi_mutex;
2360 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2361 debug_rt_mutex_free_waiter(&rt_waiter);
2363 spin_lock(q.lock_ptr);
2365 * Fixup the pi_state owner and possibly acquire the lock if we
2368 res = fixup_owner(uaddr2, &q, !ret);
2370 * If fixup_owner() returned an error, proprogate that. If it
2371 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2374 ret = (res < 0) ? res : 0;
2376 /* Unqueue and drop the lock. */
2381 * If fixup_pi_state_owner() faulted and was unable to handle the
2382 * fault, unlock the rt_mutex and return the fault to userspace.
2384 if (ret == -EFAULT) {
2385 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2386 rt_mutex_unlock(pi_mutex);
2387 } else if (ret == -EINTR) {
2389 * We've already been requeued, but cannot restart by calling
2390 * futex_lock_pi() directly. We could restart this syscall, but
2391 * it would detect that the user space "val" changed and return
2392 * -EWOULDBLOCK. Save the overhead of the restart and return
2393 * -EWOULDBLOCK directly.
2399 put_futex_key(&q.key);
2401 put_futex_key(&key2);
2405 hrtimer_cancel(&to->timer);
2406 destroy_hrtimer_on_stack(&to->timer);
2412 * Support for robust futexes: the kernel cleans up held futexes at
2415 * Implementation: user-space maintains a per-thread list of locks it
2416 * is holding. Upon do_exit(), the kernel carefully walks this list,
2417 * and marks all locks that are owned by this thread with the
2418 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2419 * always manipulated with the lock held, so the list is private and
2420 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2421 * field, to allow the kernel to clean up if the thread dies after
2422 * acquiring the lock, but just before it could have added itself to
2423 * the list. There can only be one such pending lock.
2427 * sys_set_robust_list() - Set the robust-futex list head of a task
2428 * @head: pointer to the list-head
2429 * @len: length of the list-head, as userspace expects
2431 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2434 if (!futex_cmpxchg_enabled)
2437 * The kernel knows only one size for now:
2439 if (unlikely(len != sizeof(*head)))
2442 current->robust_list = head;
2448 * sys_get_robust_list() - Get the robust-futex list head of a task
2449 * @pid: pid of the process [zero for current task]
2450 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2451 * @len_ptr: pointer to a length field, the kernel fills in the header size
2453 SYSCALL_DEFINE3(get_robust_list, int, pid,
2454 struct robust_list_head __user * __user *, head_ptr,
2455 size_t __user *, len_ptr)
2457 struct robust_list_head __user *head;
2459 struct task_struct *p;
2461 if (!futex_cmpxchg_enabled)
2470 p = find_task_by_vpid(pid);
2476 if (!ptrace_may_access(p, PTRACE_MODE_READ))
2479 head = p->robust_list;
2482 if (put_user(sizeof(*head), len_ptr))
2484 return put_user(head, head_ptr);
2493 * Process a futex-list entry, check whether it's owned by the
2494 * dying task, and do notification if so:
2496 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2498 u32 uval, uninitialized_var(nval), mval;
2501 if (get_user(uval, uaddr))
2504 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2506 * Ok, this dying thread is truly holding a futex
2507 * of interest. Set the OWNER_DIED bit atomically
2508 * via cmpxchg, and if the value had FUTEX_WAITERS
2509 * set, wake up a waiter (if any). (We have to do a
2510 * futex_wake() even if OWNER_DIED is already set -
2511 * to handle the rare but possible case of recursive
2512 * thread-death.) The rest of the cleanup is done in
2515 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2517 * We are not holding a lock here, but we want to have
2518 * the pagefault_disable/enable() protection because
2519 * we want to handle the fault gracefully. If the
2520 * access fails we try to fault in the futex with R/W
2521 * verification via get_user_pages. get_user() above
2522 * does not guarantee R/W access. If that fails we
2523 * give up and leave the futex locked.
2525 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2526 if (fault_in_user_writeable(uaddr))
2534 * Wake robust non-PI futexes here. The wakeup of
2535 * PI futexes happens in exit_pi_state():
2537 if (!pi && (uval & FUTEX_WAITERS))
2538 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2544 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2546 static inline int fetch_robust_entry(struct robust_list __user **entry,
2547 struct robust_list __user * __user *head,
2550 unsigned long uentry;
2552 if (get_user(uentry, (unsigned long __user *)head))
2555 *entry = (void __user *)(uentry & ~1UL);
2562 * Walk curr->robust_list (very carefully, it's a userspace list!)
2563 * and mark any locks found there dead, and notify any waiters.
2565 * We silently return on any sign of list-walking problem.
2567 void exit_robust_list(struct task_struct *curr)
2569 struct robust_list_head __user *head = curr->robust_list;
2570 struct robust_list __user *entry, *next_entry, *pending;
2571 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2572 unsigned int uninitialized_var(next_pi);
2573 unsigned long futex_offset;
2576 if (!futex_cmpxchg_enabled)
2580 * Fetch the list head (which was registered earlier, via
2581 * sys_set_robust_list()):
2583 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2586 * Fetch the relative futex offset:
2588 if (get_user(futex_offset, &head->futex_offset))
2591 * Fetch any possibly pending lock-add first, and handle it
2594 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2597 next_entry = NULL; /* avoid warning with gcc */
2598 while (entry != &head->list) {
2600 * Fetch the next entry in the list before calling
2601 * handle_futex_death:
2603 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2605 * A pending lock might already be on the list, so
2606 * don't process it twice:
2608 if (entry != pending)
2609 if (handle_futex_death((void __user *)entry + futex_offset,
2617 * Avoid excessively long or circular lists:
2626 handle_futex_death((void __user *)pending + futex_offset,
2630 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2631 u32 __user *uaddr2, u32 val2, u32 val3)
2633 int cmd = op & FUTEX_CMD_MASK;
2634 unsigned int flags = 0;
2636 if (!(op & FUTEX_PRIVATE_FLAG))
2637 flags |= FLAGS_SHARED;
2639 if (op & FUTEX_CLOCK_REALTIME) {
2640 flags |= FLAGS_CLOCKRT;
2641 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2647 case FUTEX_UNLOCK_PI:
2648 case FUTEX_TRYLOCK_PI:
2649 case FUTEX_WAIT_REQUEUE_PI:
2650 case FUTEX_CMP_REQUEUE_PI:
2651 if (!futex_cmpxchg_enabled)
2657 val3 = FUTEX_BITSET_MATCH_ANY;
2658 case FUTEX_WAIT_BITSET:
2659 return futex_wait(uaddr, flags, val, timeout, val3);
2661 val3 = FUTEX_BITSET_MATCH_ANY;
2662 case FUTEX_WAKE_BITSET:
2663 return futex_wake(uaddr, flags, val, val3);
2665 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2666 case FUTEX_CMP_REQUEUE:
2667 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2669 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2671 return futex_lock_pi(uaddr, flags, val, timeout, 0);
2672 case FUTEX_UNLOCK_PI:
2673 return futex_unlock_pi(uaddr, flags);
2674 case FUTEX_TRYLOCK_PI:
2675 return futex_lock_pi(uaddr, flags, 0, timeout, 1);
2676 case FUTEX_WAIT_REQUEUE_PI:
2677 val3 = FUTEX_BITSET_MATCH_ANY;
2678 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2680 case FUTEX_CMP_REQUEUE_PI:
2681 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2687 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2688 struct timespec __user *, utime, u32 __user *, uaddr2,
2692 ktime_t t, *tp = NULL;
2694 int cmd = op & FUTEX_CMD_MASK;
2696 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2697 cmd == FUTEX_WAIT_BITSET ||
2698 cmd == FUTEX_WAIT_REQUEUE_PI)) {
2699 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2701 if (!timespec_valid(&ts))
2704 t = timespec_to_ktime(ts);
2705 if (cmd == FUTEX_WAIT)
2706 t = ktime_add_safe(ktime_get(), t);
2710 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2711 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2713 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2714 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2715 val2 = (u32) (unsigned long) utime;
2717 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2720 static int __init futex_init(void)
2725 #if CONFIG_BASE_SMALL
2726 futex_hashsize = 16;
2728 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
2731 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
2733 futex_hashsize < 256 ? HASH_SMALL : 0,
2734 NULL, NULL, futex_hashsize, futex_hashsize);
2737 * This will fail and we want it. Some arch implementations do
2738 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2739 * functionality. We want to know that before we call in any
2740 * of the complex code paths. Also we want to prevent
2741 * registration of robust lists in that case. NULL is
2742 * guaranteed to fault and we get -EFAULT on functional
2743 * implementation, the non-functional ones will return
2746 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
2747 futex_cmpxchg_enabled = 1;
2749 for (i = 0; i < futex_hashsize; i++) {
2750 plist_head_init(&futex_queues[i].chain);
2751 spin_lock_init(&futex_queues[i].lock);
2756 __initcall(futex_init);