2 * Fast Userspace Mutexes (which I call "Futexes!").
3 * (C) Rusty Russell, IBM 2002
5 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
8 * Removed page pinning, fix privately mapped COW pages and other cleanups
9 * (C) Copyright 2003, 2004 Jamie Lokier
11 * Robust futex support started by Ingo Molnar
12 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
15 * PI-futex support started by Ingo Molnar and Thomas Gleixner
16 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
19 * PRIVATE futexes by Eric Dumazet
20 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
22 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
23 * enough at me, Linus for the original (flawed) idea, Matthew
24 * Kirkwood for proof-of-concept implementation.
26 * "The futexes are also cursed."
27 * "But they come in a choice of three flavours!"
29 * This program is free software; you can redistribute it and/or modify
30 * it under the terms of the GNU General Public License as published by
31 * the Free Software Foundation; either version 2 of the License, or
32 * (at your option) any later version.
34 * This program is distributed in the hope that it will be useful,
35 * but WITHOUT ANY WARRANTY; without even the implied warranty of
36 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
37 * GNU General Public License for more details.
39 * You should have received a copy of the GNU General Public License
40 * along with this program; if not, write to the Free Software
41 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
43 #include <linux/slab.h>
44 #include <linux/poll.h>
46 #include <linux/file.h>
47 #include <linux/jhash.h>
48 #include <linux/init.h>
49 #include <linux/futex.h>
50 #include <linux/mount.h>
51 #include <linux/pagemap.h>
52 #include <linux/syscalls.h>
53 #include <linux/signal.h>
54 #include <linux/module.h>
55 #include <linux/magic.h>
56 #include <linux/pid.h>
57 #include <linux/nsproxy.h>
59 #include <asm/futex.h>
61 #include "rtmutex_common.h"
63 int __read_mostly futex_cmpxchg_enabled;
65 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
68 * Priority Inheritance state:
70 struct futex_pi_state {
72 * list of 'owned' pi_state instances - these have to be
73 * cleaned up in do_exit() if the task exits prematurely:
75 struct list_head list;
80 struct rt_mutex pi_mutex;
82 struct task_struct *owner;
89 * We use this hashed waitqueue instead of a normal wait_queue_t, so
90 * we can wake only the relevant ones (hashed queues may be shared).
92 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
93 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
94 * The order of wakup is always to make the first condition true, then
95 * wake up q->waiters, then make the second condition true.
98 struct plist_node list;
99 wait_queue_head_t waiters;
101 /* Which hash list lock to use: */
102 spinlock_t *lock_ptr;
104 /* Key which the futex is hashed on: */
107 /* Optional priority inheritance state: */
108 struct futex_pi_state *pi_state;
109 struct task_struct *task;
111 /* Bitset for the optional bitmasked wakeup */
116 * Split the global futex_lock into every hash list lock.
118 struct futex_hash_bucket {
120 struct plist_head chain;
123 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
126 * Take mm->mmap_sem, when futex is shared
128 static inline void futex_lock_mm(struct rw_semaphore *fshared)
135 * Release mm->mmap_sem, when the futex is shared
137 static inline void futex_unlock_mm(struct rw_semaphore *fshared)
144 * We hash on the keys returned from get_futex_key (see below).
146 static struct futex_hash_bucket *hash_futex(union futex_key *key)
148 u32 hash = jhash2((u32*)&key->both.word,
149 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
151 return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
155 * Return 1 if two futex_keys are equal, 0 otherwise.
157 static inline int match_futex(union futex_key *key1, union futex_key *key2)
159 return (key1->both.word == key2->both.word
160 && key1->both.ptr == key2->both.ptr
161 && key1->both.offset == key2->both.offset);
165 * get_futex_key - Get parameters which are the keys for a futex.
166 * @uaddr: virtual address of the futex
167 * @shared: NULL for a PROCESS_PRIVATE futex,
168 * ¤t->mm->mmap_sem for a PROCESS_SHARED futex
169 * @key: address where result is stored.
171 * Returns a negative error code or 0
172 * The key words are stored in *key on success.
174 * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
175 * offset_within_page). For private mappings, it's (uaddr, current->mm).
176 * We can usually work out the index without swapping in the page.
178 * fshared is NULL for PROCESS_PRIVATE futexes
179 * For other futexes, it points to ¤t->mm->mmap_sem and
180 * caller must have taken the reader lock. but NOT any spinlocks.
182 static int get_futex_key(u32 __user *uaddr, struct rw_semaphore *fshared,
183 union futex_key *key)
185 unsigned long address = (unsigned long)uaddr;
186 struct mm_struct *mm = current->mm;
187 struct vm_area_struct *vma;
192 * The futex address must be "naturally" aligned.
194 key->both.offset = address % PAGE_SIZE;
195 if (unlikely((address % sizeof(u32)) != 0))
197 address -= key->both.offset;
200 * PROCESS_PRIVATE futexes are fast.
201 * As the mm cannot disappear under us and the 'key' only needs
202 * virtual address, we dont even have to find the underlying vma.
203 * Note : We do have to check 'uaddr' is a valid user address,
204 * but access_ok() should be faster than find_vma()
207 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
209 key->private.mm = mm;
210 key->private.address = address;
214 * The futex is hashed differently depending on whether
215 * it's in a shared or private mapping. So check vma first.
217 vma = find_extend_vma(mm, address);
224 if (unlikely((vma->vm_flags & (VM_IO|VM_READ)) != VM_READ))
225 return (vma->vm_flags & VM_IO) ? -EPERM : -EACCES;
228 * Private mappings are handled in a simple way.
230 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
231 * it's a read-only handle, it's expected that futexes attach to
232 * the object not the particular process. Therefore we use
233 * VM_MAYSHARE here, not VM_SHARED which is restricted to shared
234 * mappings of _writable_ handles.
236 if (likely(!(vma->vm_flags & VM_MAYSHARE))) {
237 key->both.offset |= FUT_OFF_MMSHARED; /* reference taken on mm */
238 key->private.mm = mm;
239 key->private.address = address;
244 * Linear file mappings are also simple.
246 key->shared.inode = vma->vm_file->f_path.dentry->d_inode;
247 key->both.offset |= FUT_OFF_INODE; /* inode-based key. */
248 if (likely(!(vma->vm_flags & VM_NONLINEAR))) {
249 key->shared.pgoff = (((address - vma->vm_start) >> PAGE_SHIFT)
255 * We could walk the page table to read the non-linear
256 * pte, and get the page index without fetching the page
257 * from swap. But that's a lot of code to duplicate here
258 * for a rare case, so we simply fetch the page.
260 err = get_user_pages(current, mm, address, 1, 0, 0, &page, NULL);
263 page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
271 * Take a reference to the resource addressed by a key.
272 * Can be called while holding spinlocks.
275 static void get_futex_key_refs(union futex_key *key)
277 if (key->both.ptr == NULL)
279 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
281 atomic_inc(&key->shared.inode->i_count);
283 case FUT_OFF_MMSHARED:
284 atomic_inc(&key->private.mm->mm_count);
290 * Drop a reference to the resource addressed by a key.
291 * The hash bucket spinlock must not be held.
293 static void drop_futex_key_refs(union futex_key *key)
297 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
299 iput(key->shared.inode);
301 case FUT_OFF_MMSHARED:
302 mmdrop(key->private.mm);
307 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
312 curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
318 static int get_futex_value_locked(u32 *dest, u32 __user *from)
323 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
326 return ret ? -EFAULT : 0;
331 * if fshared is non NULL, current->mm->mmap_sem is already held
333 static int futex_handle_fault(unsigned long address,
334 struct rw_semaphore *fshared, int attempt)
336 struct vm_area_struct * vma;
337 struct mm_struct *mm = current->mm;
344 down_read(&mm->mmap_sem);
345 vma = find_vma(mm, address);
346 if (vma && address >= vma->vm_start &&
347 (vma->vm_flags & VM_WRITE)) {
349 fault = handle_mm_fault(mm, vma, address, 1);
350 if (unlikely((fault & VM_FAULT_ERROR))) {
352 /* XXX: let's do this when we verify it is OK */
353 if (ret & VM_FAULT_OOM)
358 if (fault & VM_FAULT_MAJOR)
365 up_read(&mm->mmap_sem);
372 static int refill_pi_state_cache(void)
374 struct futex_pi_state *pi_state;
376 if (likely(current->pi_state_cache))
379 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
384 INIT_LIST_HEAD(&pi_state->list);
385 /* pi_mutex gets initialized later */
386 pi_state->owner = NULL;
387 atomic_set(&pi_state->refcount, 1);
389 current->pi_state_cache = pi_state;
394 static struct futex_pi_state * alloc_pi_state(void)
396 struct futex_pi_state *pi_state = current->pi_state_cache;
399 current->pi_state_cache = NULL;
404 static void free_pi_state(struct futex_pi_state *pi_state)
406 if (!atomic_dec_and_test(&pi_state->refcount))
410 * If pi_state->owner is NULL, the owner is most probably dying
411 * and has cleaned up the pi_state already
413 if (pi_state->owner) {
414 spin_lock_irq(&pi_state->owner->pi_lock);
415 list_del_init(&pi_state->list);
416 spin_unlock_irq(&pi_state->owner->pi_lock);
418 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
421 if (current->pi_state_cache)
425 * pi_state->list is already empty.
426 * clear pi_state->owner.
427 * refcount is at 0 - put it back to 1.
429 pi_state->owner = NULL;
430 atomic_set(&pi_state->refcount, 1);
431 current->pi_state_cache = pi_state;
436 * Look up the task based on what TID userspace gave us.
439 static struct task_struct * futex_find_get_task(pid_t pid)
441 struct task_struct *p;
444 p = find_task_by_vpid(pid);
445 if (!p || ((current->euid != p->euid) && (current->euid != p->uid)))
456 * This task is holding PI mutexes at exit time => bad.
457 * Kernel cleans up PI-state, but userspace is likely hosed.
458 * (Robust-futex cleanup is separate and might save the day for userspace.)
460 void exit_pi_state_list(struct task_struct *curr)
462 struct list_head *next, *head = &curr->pi_state_list;
463 struct futex_pi_state *pi_state;
464 struct futex_hash_bucket *hb;
467 if (!futex_cmpxchg_enabled)
470 * We are a ZOMBIE and nobody can enqueue itself on
471 * pi_state_list anymore, but we have to be careful
472 * versus waiters unqueueing themselves:
474 spin_lock_irq(&curr->pi_lock);
475 while (!list_empty(head)) {
478 pi_state = list_entry(next, struct futex_pi_state, list);
480 hb = hash_futex(&key);
481 spin_unlock_irq(&curr->pi_lock);
483 spin_lock(&hb->lock);
485 spin_lock_irq(&curr->pi_lock);
487 * We dropped the pi-lock, so re-check whether this
488 * task still owns the PI-state:
490 if (head->next != next) {
491 spin_unlock(&hb->lock);
495 WARN_ON(pi_state->owner != curr);
496 WARN_ON(list_empty(&pi_state->list));
497 list_del_init(&pi_state->list);
498 pi_state->owner = NULL;
499 spin_unlock_irq(&curr->pi_lock);
501 rt_mutex_unlock(&pi_state->pi_mutex);
503 spin_unlock(&hb->lock);
505 spin_lock_irq(&curr->pi_lock);
507 spin_unlock_irq(&curr->pi_lock);
511 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
512 union futex_key *key, struct futex_pi_state **ps)
514 struct futex_pi_state *pi_state = NULL;
515 struct futex_q *this, *next;
516 struct plist_head *head;
517 struct task_struct *p;
518 pid_t pid = uval & FUTEX_TID_MASK;
522 plist_for_each_entry_safe(this, next, head, list) {
523 if (match_futex(&this->key, key)) {
525 * Another waiter already exists - bump up
526 * the refcount and return its pi_state:
528 pi_state = this->pi_state;
530 * Userspace might have messed up non PI and PI futexes
532 if (unlikely(!pi_state))
535 WARN_ON(!atomic_read(&pi_state->refcount));
536 WARN_ON(pid && pi_state->owner &&
537 pi_state->owner->pid != pid);
539 atomic_inc(&pi_state->refcount);
547 * We are the first waiter - try to look up the real owner and attach
548 * the new pi_state to it, but bail out when TID = 0
552 p = futex_find_get_task(pid);
557 * We need to look at the task state flags to figure out,
558 * whether the task is exiting. To protect against the do_exit
559 * change of the task flags, we do this protected by
562 spin_lock_irq(&p->pi_lock);
563 if (unlikely(p->flags & PF_EXITING)) {
565 * The task is on the way out. When PF_EXITPIDONE is
566 * set, we know that the task has finished the
569 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
571 spin_unlock_irq(&p->pi_lock);
576 pi_state = alloc_pi_state();
579 * Initialize the pi_mutex in locked state and make 'p'
582 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
584 /* Store the key for possible exit cleanups: */
585 pi_state->key = *key;
587 WARN_ON(!list_empty(&pi_state->list));
588 list_add(&pi_state->list, &p->pi_state_list);
590 spin_unlock_irq(&p->pi_lock);
600 * The hash bucket lock must be held when this is called.
601 * Afterwards, the futex_q must not be accessed.
603 static void wake_futex(struct futex_q *q)
605 plist_del(&q->list, &q->list.plist);
607 * The lock in wake_up_all() is a crucial memory barrier after the
608 * plist_del() and also before assigning to q->lock_ptr.
610 wake_up_all(&q->waiters);
612 * The waiting task can free the futex_q as soon as this is written,
613 * without taking any locks. This must come last.
615 * A memory barrier is required here to prevent the following store
616 * to lock_ptr from getting ahead of the wakeup. Clearing the lock
617 * at the end of wake_up_all() does not prevent this store from
624 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
626 struct task_struct *new_owner;
627 struct futex_pi_state *pi_state = this->pi_state;
633 spin_lock(&pi_state->pi_mutex.wait_lock);
634 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
637 * This happens when we have stolen the lock and the original
638 * pending owner did not enqueue itself back on the rt_mutex.
639 * Thats not a tragedy. We know that way, that a lock waiter
640 * is on the fly. We make the futex_q waiter the pending owner.
643 new_owner = this->task;
646 * We pass it to the next owner. (The WAITERS bit is always
647 * kept enabled while there is PI state around. We must also
648 * preserve the owner died bit.)
650 if (!(uval & FUTEX_OWNER_DIED)) {
653 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
655 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
657 if (curval == -EFAULT)
659 else if (curval != uval)
662 spin_unlock(&pi_state->pi_mutex.wait_lock);
667 spin_lock_irq(&pi_state->owner->pi_lock);
668 WARN_ON(list_empty(&pi_state->list));
669 list_del_init(&pi_state->list);
670 spin_unlock_irq(&pi_state->owner->pi_lock);
672 spin_lock_irq(&new_owner->pi_lock);
673 WARN_ON(!list_empty(&pi_state->list));
674 list_add(&pi_state->list, &new_owner->pi_state_list);
675 pi_state->owner = new_owner;
676 spin_unlock_irq(&new_owner->pi_lock);
678 spin_unlock(&pi_state->pi_mutex.wait_lock);
679 rt_mutex_unlock(&pi_state->pi_mutex);
684 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
689 * There is no waiter, so we unlock the futex. The owner died
690 * bit has not to be preserved here. We are the owner:
692 oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
694 if (oldval == -EFAULT)
703 * Express the locking dependencies for lockdep:
706 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
709 spin_lock(&hb1->lock);
711 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
712 } else { /* hb1 > hb2 */
713 spin_lock(&hb2->lock);
714 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
719 * Wake up all waiters hashed on the physical page that is mapped
720 * to this virtual address:
722 static int futex_wake(u32 __user *uaddr, struct rw_semaphore *fshared,
723 int nr_wake, u32 bitset)
725 struct futex_hash_bucket *hb;
726 struct futex_q *this, *next;
727 struct plist_head *head;
734 futex_lock_mm(fshared);
736 ret = get_futex_key(uaddr, fshared, &key);
737 if (unlikely(ret != 0))
740 hb = hash_futex(&key);
741 spin_lock(&hb->lock);
744 plist_for_each_entry_safe(this, next, head, list) {
745 if (match_futex (&this->key, &key)) {
746 if (this->pi_state) {
751 /* Check if one of the bits is set in both bitsets */
752 if (!(this->bitset & bitset))
756 if (++ret >= nr_wake)
761 spin_unlock(&hb->lock);
763 futex_unlock_mm(fshared);
768 * Wake up all waiters hashed on the physical page that is mapped
769 * to this virtual address:
772 futex_wake_op(u32 __user *uaddr1, struct rw_semaphore *fshared,
774 int nr_wake, int nr_wake2, int op)
776 union futex_key key1, key2;
777 struct futex_hash_bucket *hb1, *hb2;
778 struct plist_head *head;
779 struct futex_q *this, *next;
780 int ret, op_ret, attempt = 0;
783 futex_lock_mm(fshared);
785 ret = get_futex_key(uaddr1, fshared, &key1);
786 if (unlikely(ret != 0))
788 ret = get_futex_key(uaddr2, fshared, &key2);
789 if (unlikely(ret != 0))
792 hb1 = hash_futex(&key1);
793 hb2 = hash_futex(&key2);
796 double_lock_hb(hb1, hb2);
798 op_ret = futex_atomic_op_inuser(op, uaddr2);
799 if (unlikely(op_ret < 0)) {
802 spin_unlock(&hb1->lock);
804 spin_unlock(&hb2->lock);
808 * we don't get EFAULT from MMU faults if we don't have an MMU,
809 * but we might get them from range checking
815 if (unlikely(op_ret != -EFAULT)) {
821 * futex_atomic_op_inuser needs to both read and write
822 * *(int __user *)uaddr2, but we can't modify it
823 * non-atomically. Therefore, if get_user below is not
824 * enough, we need to handle the fault ourselves, while
825 * still holding the mmap_sem.
828 ret = futex_handle_fault((unsigned long)uaddr2,
836 * If we would have faulted, release mmap_sem,
837 * fault it in and start all over again.
839 futex_unlock_mm(fshared);
841 ret = get_user(dummy, uaddr2);
850 plist_for_each_entry_safe(this, next, head, list) {
851 if (match_futex (&this->key, &key1)) {
853 if (++ret >= nr_wake)
862 plist_for_each_entry_safe(this, next, head, list) {
863 if (match_futex (&this->key, &key2)) {
865 if (++op_ret >= nr_wake2)
872 spin_unlock(&hb1->lock);
874 spin_unlock(&hb2->lock);
876 futex_unlock_mm(fshared);
882 * Requeue all waiters hashed on one physical page to another
885 static int futex_requeue(u32 __user *uaddr1, struct rw_semaphore *fshared,
887 int nr_wake, int nr_requeue, u32 *cmpval)
889 union futex_key key1, key2;
890 struct futex_hash_bucket *hb1, *hb2;
891 struct plist_head *head1;
892 struct futex_q *this, *next;
893 int ret, drop_count = 0;
896 futex_lock_mm(fshared);
898 ret = get_futex_key(uaddr1, fshared, &key1);
899 if (unlikely(ret != 0))
901 ret = get_futex_key(uaddr2, fshared, &key2);
902 if (unlikely(ret != 0))
905 hb1 = hash_futex(&key1);
906 hb2 = hash_futex(&key2);
908 double_lock_hb(hb1, hb2);
910 if (likely(cmpval != NULL)) {
913 ret = get_futex_value_locked(&curval, uaddr1);
916 spin_unlock(&hb1->lock);
918 spin_unlock(&hb2->lock);
921 * If we would have faulted, release mmap_sem, fault
922 * it in and start all over again.
924 futex_unlock_mm(fshared);
926 ret = get_user(curval, uaddr1);
933 if (curval != *cmpval) {
940 plist_for_each_entry_safe(this, next, head1, list) {
941 if (!match_futex (&this->key, &key1))
943 if (++ret <= nr_wake) {
947 * If key1 and key2 hash to the same bucket, no need to
950 if (likely(head1 != &hb2->chain)) {
951 plist_del(&this->list, &hb1->chain);
952 plist_add(&this->list, &hb2->chain);
953 this->lock_ptr = &hb2->lock;
954 #ifdef CONFIG_DEBUG_PI_LIST
955 this->list.plist.lock = &hb2->lock;
959 get_futex_key_refs(&key2);
962 if (ret - nr_wake >= nr_requeue)
968 spin_unlock(&hb1->lock);
970 spin_unlock(&hb2->lock);
972 /* drop_futex_key_refs() must be called outside the spinlocks. */
973 while (--drop_count >= 0)
974 drop_futex_key_refs(&key1);
977 futex_unlock_mm(fshared);
981 /* The key must be already stored in q->key. */
982 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
984 struct futex_hash_bucket *hb;
986 init_waitqueue_head(&q->waiters);
988 get_futex_key_refs(&q->key);
989 hb = hash_futex(&q->key);
990 q->lock_ptr = &hb->lock;
992 spin_lock(&hb->lock);
996 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1001 * The priority used to register this element is
1002 * - either the real thread-priority for the real-time threads
1003 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1004 * - or MAX_RT_PRIO for non-RT threads.
1005 * Thus, all RT-threads are woken first in priority order, and
1006 * the others are woken last, in FIFO order.
1008 prio = min(current->normal_prio, MAX_RT_PRIO);
1010 plist_node_init(&q->list, prio);
1011 #ifdef CONFIG_DEBUG_PI_LIST
1012 q->list.plist.lock = &hb->lock;
1014 plist_add(&q->list, &hb->chain);
1016 spin_unlock(&hb->lock);
1020 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1022 spin_unlock(&hb->lock);
1023 drop_futex_key_refs(&q->key);
1027 * queue_me and unqueue_me must be called as a pair, each
1028 * exactly once. They are called with the hashed spinlock held.
1031 /* Return 1 if we were still queued (ie. 0 means we were woken) */
1032 static int unqueue_me(struct futex_q *q)
1034 spinlock_t *lock_ptr;
1037 /* In the common case we don't take the spinlock, which is nice. */
1039 lock_ptr = q->lock_ptr;
1041 if (lock_ptr != NULL) {
1042 spin_lock(lock_ptr);
1044 * q->lock_ptr can change between reading it and
1045 * spin_lock(), causing us to take the wrong lock. This
1046 * corrects the race condition.
1048 * Reasoning goes like this: if we have the wrong lock,
1049 * q->lock_ptr must have changed (maybe several times)
1050 * between reading it and the spin_lock(). It can
1051 * change again after the spin_lock() but only if it was
1052 * already changed before the spin_lock(). It cannot,
1053 * however, change back to the original value. Therefore
1054 * we can detect whether we acquired the correct lock.
1056 if (unlikely(lock_ptr != q->lock_ptr)) {
1057 spin_unlock(lock_ptr);
1060 WARN_ON(plist_node_empty(&q->list));
1061 plist_del(&q->list, &q->list.plist);
1063 BUG_ON(q->pi_state);
1065 spin_unlock(lock_ptr);
1069 drop_futex_key_refs(&q->key);
1074 * PI futexes can not be requeued and must remove themself from the
1075 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1078 static void unqueue_me_pi(struct futex_q *q)
1080 WARN_ON(plist_node_empty(&q->list));
1081 plist_del(&q->list, &q->list.plist);
1083 BUG_ON(!q->pi_state);
1084 free_pi_state(q->pi_state);
1087 spin_unlock(q->lock_ptr);
1089 drop_futex_key_refs(&q->key);
1093 * Fixup the pi_state owner with the new owner.
1095 * Must be called with hash bucket lock held and mm->sem held for non
1098 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1099 struct task_struct *newowner,
1100 struct rw_semaphore *fshared)
1102 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1103 struct futex_pi_state *pi_state = q->pi_state;
1104 struct task_struct *oldowner = pi_state->owner;
1105 u32 uval, curval, newval;
1106 int ret, attempt = 0;
1109 if (!pi_state->owner)
1110 newtid |= FUTEX_OWNER_DIED;
1113 * We are here either because we stole the rtmutex from the
1114 * pending owner or we are the pending owner which failed to
1115 * get the rtmutex. We have to replace the pending owner TID
1116 * in the user space variable. This must be atomic as we have
1117 * to preserve the owner died bit here.
1119 * Note: We write the user space value _before_ changing the
1120 * pi_state because we can fault here. Imagine swapped out
1121 * pages or a fork, which was running right before we acquired
1122 * mmap_sem, that marked all the anonymous memory readonly for
1125 * Modifying pi_state _before_ the user space value would
1126 * leave the pi_state in an inconsistent state when we fault
1127 * here, because we need to drop the hash bucket lock to
1128 * handle the fault. This might be observed in the PID check
1129 * in lookup_pi_state.
1132 if (get_futex_value_locked(&uval, uaddr))
1136 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1138 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1140 if (curval == -EFAULT)
1148 * We fixed up user space. Now we need to fix the pi_state
1151 if (pi_state->owner != NULL) {
1152 spin_lock_irq(&pi_state->owner->pi_lock);
1153 WARN_ON(list_empty(&pi_state->list));
1154 list_del_init(&pi_state->list);
1155 spin_unlock_irq(&pi_state->owner->pi_lock);
1158 pi_state->owner = newowner;
1160 spin_lock_irq(&newowner->pi_lock);
1161 WARN_ON(!list_empty(&pi_state->list));
1162 list_add(&pi_state->list, &newowner->pi_state_list);
1163 spin_unlock_irq(&newowner->pi_lock);
1167 * To handle the page fault we need to drop the hash bucket
1168 * lock here. That gives the other task (either the pending
1169 * owner itself or the task which stole the rtmutex) the
1170 * chance to try the fixup of the pi_state. So once we are
1171 * back from handling the fault we need to check the pi_state
1172 * after reacquiring the hash bucket lock and before trying to
1173 * do another fixup. When the fixup has been done already we
1177 spin_unlock(q->lock_ptr);
1179 ret = futex_handle_fault((unsigned long)uaddr, fshared, attempt++);
1181 spin_lock(q->lock_ptr);
1184 * Check if someone else fixed it for us:
1186 if (pi_state->owner != oldowner)
1196 * In case we must use restart_block to restart a futex_wait,
1197 * we encode in the 'flags' shared capability
1199 #define FLAGS_SHARED 1
1201 static long futex_wait_restart(struct restart_block *restart);
1203 static int futex_wait(u32 __user *uaddr, struct rw_semaphore *fshared,
1204 u32 val, ktime_t *abs_time, u32 bitset)
1206 struct task_struct *curr = current;
1207 DECLARE_WAITQUEUE(wait, curr);
1208 struct futex_hash_bucket *hb;
1212 struct hrtimer_sleeper t;
1221 futex_lock_mm(fshared);
1223 ret = get_futex_key(uaddr, fshared, &q.key);
1224 if (unlikely(ret != 0))
1225 goto out_release_sem;
1227 hb = queue_lock(&q);
1230 * Access the page AFTER the futex is queued.
1231 * Order is important:
1233 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1234 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1236 * The basic logical guarantee of a futex is that it blocks ONLY
1237 * if cond(var) is known to be true at the time of blocking, for
1238 * any cond. If we queued after testing *uaddr, that would open
1239 * a race condition where we could block indefinitely with
1240 * cond(var) false, which would violate the guarantee.
1242 * A consequence is that futex_wait() can return zero and absorb
1243 * a wakeup when *uaddr != val on entry to the syscall. This is
1246 * for shared futexes, we hold the mmap semaphore, so the mapping
1247 * cannot have changed since we looked it up in get_futex_key.
1249 ret = get_futex_value_locked(&uval, uaddr);
1251 if (unlikely(ret)) {
1252 queue_unlock(&q, hb);
1255 * If we would have faulted, release mmap_sem, fault it in and
1256 * start all over again.
1258 futex_unlock_mm(fshared);
1260 ret = get_user(uval, uaddr);
1268 goto out_unlock_release_sem;
1270 /* Only actually queue if *uaddr contained val. */
1274 * Now the futex is queued and we have checked the data, we
1275 * don't want to hold mmap_sem while we sleep.
1277 futex_unlock_mm(fshared);
1280 * There might have been scheduling since the queue_me(), as we
1281 * cannot hold a spinlock across the get_user() in case it
1282 * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1283 * queueing ourselves into the futex hash. This code thus has to
1284 * rely on the futex_wake() code removing us from hash when it
1288 /* add_wait_queue is the barrier after __set_current_state. */
1289 __set_current_state(TASK_INTERRUPTIBLE);
1290 add_wait_queue(&q.waiters, &wait);
1292 * !plist_node_empty() is safe here without any lock.
1293 * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1295 if (likely(!plist_node_empty(&q.list))) {
1299 hrtimer_init_on_stack(&t.timer, CLOCK_MONOTONIC,
1301 hrtimer_init_sleeper(&t, current);
1302 t.timer.expires = *abs_time;
1304 hrtimer_start(&t.timer, t.timer.expires,
1306 if (!hrtimer_active(&t.timer))
1310 * the timer could have already expired, in which
1311 * case current would be flagged for rescheduling.
1312 * Don't bother calling schedule.
1317 hrtimer_cancel(&t.timer);
1319 /* Flag if a timeout occured */
1320 rem = (t.task == NULL);
1322 destroy_hrtimer_on_stack(&t.timer);
1325 __set_current_state(TASK_RUNNING);
1328 * NOTE: we don't remove ourselves from the waitqueue because
1329 * we are the only user of it.
1332 /* If we were woken (and unqueued), we succeeded, whatever. */
1333 if (!unqueue_me(&q))
1339 * We expect signal_pending(current), but another thread may
1340 * have handled it for us already.
1343 return -ERESTARTSYS;
1345 struct restart_block *restart;
1346 restart = ¤t_thread_info()->restart_block;
1347 restart->fn = futex_wait_restart;
1348 restart->futex.uaddr = (u32 *)uaddr;
1349 restart->futex.val = val;
1350 restart->futex.time = abs_time->tv64;
1351 restart->futex.bitset = bitset;
1352 restart->futex.flags = 0;
1355 restart->futex.flags |= FLAGS_SHARED;
1356 return -ERESTART_RESTARTBLOCK;
1359 out_unlock_release_sem:
1360 queue_unlock(&q, hb);
1363 futex_unlock_mm(fshared);
1368 static long futex_wait_restart(struct restart_block *restart)
1370 u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1371 struct rw_semaphore *fshared = NULL;
1374 t.tv64 = restart->futex.time;
1375 restart->fn = do_no_restart_syscall;
1376 if (restart->futex.flags & FLAGS_SHARED)
1377 fshared = ¤t->mm->mmap_sem;
1378 return (long)futex_wait(uaddr, fshared, restart->futex.val, &t,
1379 restart->futex.bitset);
1384 * Userspace tried a 0 -> TID atomic transition of the futex value
1385 * and failed. The kernel side here does the whole locking operation:
1386 * if there are waiters then it will block, it does PI, etc. (Due to
1387 * races the kernel might see a 0 value of the futex too.)
1389 static int futex_lock_pi(u32 __user *uaddr, struct rw_semaphore *fshared,
1390 int detect, ktime_t *time, int trylock)
1392 struct hrtimer_sleeper timeout, *to = NULL;
1393 struct task_struct *curr = current;
1394 struct futex_hash_bucket *hb;
1395 u32 uval, newval, curval;
1397 int ret, lock_taken, ownerdied = 0, attempt = 0;
1399 if (refill_pi_state_cache())
1404 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1406 hrtimer_init_sleeper(to, current);
1407 to->timer.expires = *time;
1412 futex_lock_mm(fshared);
1414 ret = get_futex_key(uaddr, fshared, &q.key);
1415 if (unlikely(ret != 0))
1416 goto out_release_sem;
1419 hb = queue_lock(&q);
1422 ret = lock_taken = 0;
1425 * To avoid races, we attempt to take the lock here again
1426 * (by doing a 0 -> TID atomic cmpxchg), while holding all
1427 * the locks. It will most likely not succeed.
1429 newval = task_pid_vnr(current);
1431 curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
1433 if (unlikely(curval == -EFAULT))
1437 * Detect deadlocks. In case of REQUEUE_PI this is a valid
1438 * situation and we return success to user space.
1440 if (unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(current))) {
1442 goto out_unlock_release_sem;
1446 * Surprise - we got the lock. Just return to userspace:
1448 if (unlikely(!curval))
1449 goto out_unlock_release_sem;
1454 * Set the WAITERS flag, so the owner will know it has someone
1455 * to wake at next unlock
1457 newval = curval | FUTEX_WAITERS;
1460 * There are two cases, where a futex might have no owner (the
1461 * owner TID is 0): OWNER_DIED. We take over the futex in this
1462 * case. We also do an unconditional take over, when the owner
1463 * of the futex died.
1465 * This is safe as we are protected by the hash bucket lock !
1467 if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
1468 /* Keep the OWNER_DIED bit */
1469 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(current);
1474 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1476 if (unlikely(curval == -EFAULT))
1478 if (unlikely(curval != uval))
1482 * We took the lock due to owner died take over.
1484 if (unlikely(lock_taken))
1485 goto out_unlock_release_sem;
1488 * We dont have the lock. Look up the PI state (or create it if
1489 * we are the first waiter):
1491 ret = lookup_pi_state(uval, hb, &q.key, &q.pi_state);
1493 if (unlikely(ret)) {
1498 * Task is exiting and we just wait for the
1501 queue_unlock(&q, hb);
1502 futex_unlock_mm(fshared);
1508 * No owner found for this futex. Check if the
1509 * OWNER_DIED bit is set to figure out whether
1510 * this is a robust futex or not.
1512 if (get_futex_value_locked(&curval, uaddr))
1516 * We simply start over in case of a robust
1517 * futex. The code above will take the futex
1520 if (curval & FUTEX_OWNER_DIED) {
1525 goto out_unlock_release_sem;
1530 * Only actually queue now that the atomic ops are done:
1535 * Now the futex is queued and we have checked the data, we
1536 * don't want to hold mmap_sem while we sleep.
1538 futex_unlock_mm(fshared);
1540 WARN_ON(!q.pi_state);
1542 * Block on the PI mutex:
1545 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1547 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1548 /* Fixup the trylock return value: */
1549 ret = ret ? 0 : -EWOULDBLOCK;
1552 futex_lock_mm(fshared);
1553 spin_lock(q.lock_ptr);
1557 * Got the lock. We might not be the anticipated owner
1558 * if we did a lock-steal - fix up the PI-state in
1561 if (q.pi_state->owner != curr)
1562 ret = fixup_pi_state_owner(uaddr, &q, curr, fshared);
1565 * Catch the rare case, where the lock was released
1566 * when we were on the way back before we locked the
1569 if (q.pi_state->owner == curr) {
1571 * Try to get the rt_mutex now. This might
1572 * fail as some other task acquired the
1573 * rt_mutex after we removed ourself from the
1574 * rt_mutex waiters list.
1576 if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1580 * pi_state is incorrect, some other
1581 * task did a lock steal and we
1582 * returned due to timeout or signal
1583 * without taking the rt_mutex. Too
1584 * late. We can access the
1585 * rt_mutex_owner without locking, as
1586 * the other task is now blocked on
1587 * the hash bucket lock. Fix the state
1590 struct task_struct *owner;
1593 owner = rt_mutex_owner(&q.pi_state->pi_mutex);
1594 res = fixup_pi_state_owner(uaddr, &q, owner,
1597 /* propagate -EFAULT, if the fixup failed */
1603 * Paranoia check. If we did not take the lock
1604 * in the trylock above, then we should not be
1605 * the owner of the rtmutex, neither the real
1606 * nor the pending one:
1608 if (rt_mutex_owner(&q.pi_state->pi_mutex) == curr)
1609 printk(KERN_ERR "futex_lock_pi: ret = %d "
1610 "pi-mutex: %p pi-state %p\n", ret,
1611 q.pi_state->pi_mutex.owner,
1616 /* Unqueue and drop the lock */
1618 futex_unlock_mm(fshared);
1621 destroy_hrtimer_on_stack(&to->timer);
1622 return ret != -EINTR ? ret : -ERESTARTNOINTR;
1624 out_unlock_release_sem:
1625 queue_unlock(&q, hb);
1628 futex_unlock_mm(fshared);
1630 destroy_hrtimer_on_stack(&to->timer);
1635 * We have to r/w *(int __user *)uaddr, but we can't modify it
1636 * non-atomically. Therefore, if get_user below is not
1637 * enough, we need to handle the fault ourselves, while
1638 * still holding the mmap_sem.
1640 * ... and hb->lock. :-) --ANK
1642 queue_unlock(&q, hb);
1645 ret = futex_handle_fault((unsigned long)uaddr, fshared,
1648 goto out_release_sem;
1649 goto retry_unlocked;
1652 futex_unlock_mm(fshared);
1654 ret = get_user(uval, uaddr);
1655 if (!ret && (uval != -EFAULT))
1659 destroy_hrtimer_on_stack(&to->timer);
1664 * Userspace attempted a TID -> 0 atomic transition, and failed.
1665 * This is the in-kernel slowpath: we look up the PI state (if any),
1666 * and do the rt-mutex unlock.
1668 static int futex_unlock_pi(u32 __user *uaddr, struct rw_semaphore *fshared)
1670 struct futex_hash_bucket *hb;
1671 struct futex_q *this, *next;
1673 struct plist_head *head;
1674 union futex_key key;
1675 int ret, attempt = 0;
1678 if (get_user(uval, uaddr))
1681 * We release only a lock we actually own:
1683 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
1686 * First take all the futex related locks:
1688 futex_lock_mm(fshared);
1690 ret = get_futex_key(uaddr, fshared, &key);
1691 if (unlikely(ret != 0))
1694 hb = hash_futex(&key);
1696 spin_lock(&hb->lock);
1699 * To avoid races, try to do the TID -> 0 atomic transition
1700 * again. If it succeeds then we can return without waking
1703 if (!(uval & FUTEX_OWNER_DIED))
1704 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
1707 if (unlikely(uval == -EFAULT))
1710 * Rare case: we managed to release the lock atomically,
1711 * no need to wake anyone else up:
1713 if (unlikely(uval == task_pid_vnr(current)))
1717 * Ok, other tasks may need to be woken up - check waiters
1718 * and do the wakeup if necessary:
1722 plist_for_each_entry_safe(this, next, head, list) {
1723 if (!match_futex (&this->key, &key))
1725 ret = wake_futex_pi(uaddr, uval, this);
1727 * The atomic access to the futex value
1728 * generated a pagefault, so retry the
1729 * user-access and the wakeup:
1736 * No waiters - kernel unlocks the futex:
1738 if (!(uval & FUTEX_OWNER_DIED)) {
1739 ret = unlock_futex_pi(uaddr, uval);
1745 spin_unlock(&hb->lock);
1747 futex_unlock_mm(fshared);
1753 * We have to r/w *(int __user *)uaddr, but we can't modify it
1754 * non-atomically. Therefore, if get_user below is not
1755 * enough, we need to handle the fault ourselves, while
1756 * still holding the mmap_sem.
1758 * ... and hb->lock. --ANK
1760 spin_unlock(&hb->lock);
1763 ret = futex_handle_fault((unsigned long)uaddr, fshared,
1768 goto retry_unlocked;
1771 futex_unlock_mm(fshared);
1773 ret = get_user(uval, uaddr);
1774 if (!ret && (uval != -EFAULT))
1781 * Support for robust futexes: the kernel cleans up held futexes at
1784 * Implementation: user-space maintains a per-thread list of locks it
1785 * is holding. Upon do_exit(), the kernel carefully walks this list,
1786 * and marks all locks that are owned by this thread with the
1787 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1788 * always manipulated with the lock held, so the list is private and
1789 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1790 * field, to allow the kernel to clean up if the thread dies after
1791 * acquiring the lock, but just before it could have added itself to
1792 * the list. There can only be one such pending lock.
1796 * sys_set_robust_list - set the robust-futex list head of a task
1797 * @head: pointer to the list-head
1798 * @len: length of the list-head, as userspace expects
1801 sys_set_robust_list(struct robust_list_head __user *head,
1804 if (!futex_cmpxchg_enabled)
1807 * The kernel knows only one size for now:
1809 if (unlikely(len != sizeof(*head)))
1812 current->robust_list = head;
1818 * sys_get_robust_list - get the robust-futex list head of a task
1819 * @pid: pid of the process [zero for current task]
1820 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1821 * @len_ptr: pointer to a length field, the kernel fills in the header size
1824 sys_get_robust_list(int pid, struct robust_list_head __user * __user *head_ptr,
1825 size_t __user *len_ptr)
1827 struct robust_list_head __user *head;
1830 if (!futex_cmpxchg_enabled)
1834 head = current->robust_list;
1836 struct task_struct *p;
1840 p = find_task_by_vpid(pid);
1844 if ((current->euid != p->euid) && (current->euid != p->uid) &&
1845 !capable(CAP_SYS_PTRACE))
1847 head = p->robust_list;
1851 if (put_user(sizeof(*head), len_ptr))
1853 return put_user(head, head_ptr);
1862 * Process a futex-list entry, check whether it's owned by the
1863 * dying task, and do notification if so:
1865 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
1867 u32 uval, nval, mval;
1870 if (get_user(uval, uaddr))
1873 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
1875 * Ok, this dying thread is truly holding a futex
1876 * of interest. Set the OWNER_DIED bit atomically
1877 * via cmpxchg, and if the value had FUTEX_WAITERS
1878 * set, wake up a waiter (if any). (We have to do a
1879 * futex_wake() even if OWNER_DIED is already set -
1880 * to handle the rare but possible case of recursive
1881 * thread-death.) The rest of the cleanup is done in
1884 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1885 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
1887 if (nval == -EFAULT)
1894 * Wake robust non-PI futexes here. The wakeup of
1895 * PI futexes happens in exit_pi_state():
1897 if (!pi && (uval & FUTEX_WAITERS))
1898 futex_wake(uaddr, &curr->mm->mmap_sem, 1,
1899 FUTEX_BITSET_MATCH_ANY);
1905 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1907 static inline int fetch_robust_entry(struct robust_list __user **entry,
1908 struct robust_list __user * __user *head,
1911 unsigned long uentry;
1913 if (get_user(uentry, (unsigned long __user *)head))
1916 *entry = (void __user *)(uentry & ~1UL);
1923 * Walk curr->robust_list (very carefully, it's a userspace list!)
1924 * and mark any locks found there dead, and notify any waiters.
1926 * We silently return on any sign of list-walking problem.
1928 void exit_robust_list(struct task_struct *curr)
1930 struct robust_list_head __user *head = curr->robust_list;
1931 struct robust_list __user *entry, *next_entry, *pending;
1932 unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
1933 unsigned long futex_offset;
1936 if (!futex_cmpxchg_enabled)
1940 * Fetch the list head (which was registered earlier, via
1941 * sys_set_robust_list()):
1943 if (fetch_robust_entry(&entry, &head->list.next, &pi))
1946 * Fetch the relative futex offset:
1948 if (get_user(futex_offset, &head->futex_offset))
1951 * Fetch any possibly pending lock-add first, and handle it
1954 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
1957 next_entry = NULL; /* avoid warning with gcc */
1958 while (entry != &head->list) {
1960 * Fetch the next entry in the list before calling
1961 * handle_futex_death:
1963 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
1965 * A pending lock might already be on the list, so
1966 * don't process it twice:
1968 if (entry != pending)
1969 if (handle_futex_death((void __user *)entry + futex_offset,
1977 * Avoid excessively long or circular lists:
1986 handle_futex_death((void __user *)pending + futex_offset,
1990 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
1991 u32 __user *uaddr2, u32 val2, u32 val3)
1994 int cmd = op & FUTEX_CMD_MASK;
1995 struct rw_semaphore *fshared = NULL;
1997 if (!(op & FUTEX_PRIVATE_FLAG))
1998 fshared = ¤t->mm->mmap_sem;
2002 val3 = FUTEX_BITSET_MATCH_ANY;
2003 case FUTEX_WAIT_BITSET:
2004 ret = futex_wait(uaddr, fshared, val, timeout, val3);
2007 val3 = FUTEX_BITSET_MATCH_ANY;
2008 case FUTEX_WAKE_BITSET:
2009 ret = futex_wake(uaddr, fshared, val, val3);
2012 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL);
2014 case FUTEX_CMP_REQUEUE:
2015 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3);
2018 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2021 if (futex_cmpxchg_enabled)
2022 ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2024 case FUTEX_UNLOCK_PI:
2025 if (futex_cmpxchg_enabled)
2026 ret = futex_unlock_pi(uaddr, fshared);
2028 case FUTEX_TRYLOCK_PI:
2029 if (futex_cmpxchg_enabled)
2030 ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2039 asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
2040 struct timespec __user *utime, u32 __user *uaddr2,
2044 ktime_t t, *tp = NULL;
2046 int cmd = op & FUTEX_CMD_MASK;
2048 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2049 cmd == FUTEX_WAIT_BITSET)) {
2050 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2052 if (!timespec_valid(&ts))
2055 t = timespec_to_ktime(ts);
2056 if (cmd == FUTEX_WAIT)
2057 t = ktime_add_safe(ktime_get(), t);
2061 * requeue parameter in 'utime' if cmd == FUTEX_REQUEUE.
2062 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2064 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2065 cmd == FUTEX_WAKE_OP)
2066 val2 = (u32) (unsigned long) utime;
2068 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2071 static int __init futex_init(void)
2077 * This will fail and we want it. Some arch implementations do
2078 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2079 * functionality. We want to know that before we call in any
2080 * of the complex code paths. Also we want to prevent
2081 * registration of robust lists in that case. NULL is
2082 * guaranteed to fault and we get -EFAULT on functional
2083 * implementation, the non functional ones will return
2086 curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2087 if (curval == -EFAULT)
2088 futex_cmpxchg_enabled = 1;
2090 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2091 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2092 spin_lock_init(&futex_queues[i].lock);
2097 __initcall(futex_init);