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"
73 * Basic futex operation and ordering guarantees:
75 * The waiter reads the futex value in user space and calls
76 * futex_wait(). This function computes the hash bucket and acquires
77 * the hash bucket lock. After that it reads the futex user space value
78 * again and verifies that the data has not changed. If it has not changed
79 * it enqueues itself into the hash bucket, releases the hash bucket lock
82 * The waker side modifies the user space value of the futex and calls
83 * futex_wake(). This function computes the hash bucket and acquires the
84 * hash bucket lock. Then it looks for waiters on that futex in the hash
85 * bucket and wakes them.
87 * In futex wake up scenarios where no tasks are blocked on a futex, taking
88 * the hb spinlock can be avoided and simply return. In order for this
89 * optimization to work, ordering guarantees must exist so that the waiter
90 * being added to the list is acknowledged when the list is concurrently being
91 * checked by the waker, avoiding scenarios like the following:
95 * sys_futex(WAIT, futex, val);
96 * futex_wait(futex, val);
99 * sys_futex(WAKE, futex);
104 * lock(hash_bucket(futex));
106 * unlock(hash_bucket(futex));
109 * This would cause the waiter on CPU 0 to wait forever because it
110 * missed the transition of the user space value from val to newval
111 * and the waker did not find the waiter in the hash bucket queue.
113 * The correct serialization ensures that a waiter either observes
114 * the changed user space value before blocking or is woken by a
119 * sys_futex(WAIT, futex, val);
120 * futex_wait(futex, val);
123 * mb(); (A) <-- paired with -.
125 * lock(hash_bucket(futex)); |
129 * | sys_futex(WAKE, futex);
130 * | futex_wake(futex);
132 * `-------> mb(); (B)
135 * unlock(hash_bucket(futex));
136 * schedule(); if (waiters)
137 * lock(hash_bucket(futex));
138 * wake_waiters(futex);
139 * unlock(hash_bucket(futex));
141 * Where (A) orders the waiters increment and the futex value read -- this
142 * is guaranteed by the head counter in the hb spinlock; and where (B)
143 * orders the write to futex and the waiters read -- this is done by the
144 * barriers in get_futex_key_refs(), through either ihold or atomic_inc,
145 * depending on the futex type.
147 * This yields the following case (where X:=waiters, Y:=futex):
155 * Which guarantees that x==0 && y==0 is impossible; which translates back into
156 * the guarantee that we cannot both miss the futex variable change and the
160 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
161 int __read_mostly futex_cmpxchg_enabled
;
165 * Futex flags used to encode options to functions and preserve them across
168 #define FLAGS_SHARED 0x01
169 #define FLAGS_CLOCKRT 0x02
170 #define FLAGS_HAS_TIMEOUT 0x04
173 * Priority Inheritance state:
175 struct futex_pi_state
{
177 * list of 'owned' pi_state instances - these have to be
178 * cleaned up in do_exit() if the task exits prematurely:
180 struct list_head list
;
185 struct rt_mutex pi_mutex
;
187 struct task_struct
*owner
;
194 * struct futex_q - The hashed futex queue entry, one per waiting task
195 * @list: priority-sorted list of tasks waiting on this futex
196 * @task: the task waiting on the futex
197 * @lock_ptr: the hash bucket lock
198 * @key: the key the futex is hashed on
199 * @pi_state: optional priority inheritance state
200 * @rt_waiter: rt_waiter storage for use with requeue_pi
201 * @requeue_pi_key: the requeue_pi target futex key
202 * @bitset: bitset for the optional bitmasked wakeup
204 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
205 * we can wake only the relevant ones (hashed queues may be shared).
207 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
208 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
209 * The order of wakeup is always to make the first condition true, then
212 * PI futexes are typically woken before they are removed from the hash list via
213 * the rt_mutex code. See unqueue_me_pi().
216 struct plist_node list
;
218 struct task_struct
*task
;
219 spinlock_t
*lock_ptr
;
221 struct futex_pi_state
*pi_state
;
222 struct rt_mutex_waiter
*rt_waiter
;
223 union futex_key
*requeue_pi_key
;
227 static const struct futex_q futex_q_init
= {
228 /* list gets initialized in queue_me()*/
229 .key
= FUTEX_KEY_INIT
,
230 .bitset
= FUTEX_BITSET_MATCH_ANY
234 * Hash buckets are shared by all the futex_keys that hash to the same
235 * location. Each key may have multiple futex_q structures, one for each task
236 * waiting on a futex.
238 struct futex_hash_bucket
{
240 struct plist_head chain
;
241 } ____cacheline_aligned_in_smp
;
243 static unsigned long __read_mostly futex_hashsize
;
245 static struct futex_hash_bucket
*futex_queues
;
247 static inline void futex_get_mm(union futex_key
*key
)
249 atomic_inc(&key
->private.mm
->mm_count
);
251 * Ensure futex_get_mm() implies a full barrier such that
252 * get_futex_key() implies a full barrier. This is relied upon
253 * as full barrier (B), see the ordering comment above.
255 smp_mb__after_atomic_inc();
258 static inline bool hb_waiters_pending(struct futex_hash_bucket
*hb
)
262 * Tasks trying to enter the critical region are most likely
263 * potential waiters that will be added to the plist. Ensure
264 * that wakers won't miss to-be-slept tasks in the window between
265 * the wait call and the actual plist_add.
267 if (spin_is_locked(&hb
->lock
))
269 smp_rmb(); /* Make sure we check the lock state first */
271 return !plist_head_empty(&hb
->chain
);
278 * We hash on the keys returned from get_futex_key (see below).
280 static struct futex_hash_bucket
*hash_futex(union futex_key
*key
)
282 u32 hash
= jhash2((u32
*)&key
->both
.word
,
283 (sizeof(key
->both
.word
)+sizeof(key
->both
.ptr
))/4,
285 return &futex_queues
[hash
& (futex_hashsize
- 1)];
289 * Return 1 if two futex_keys are equal, 0 otherwise.
291 static inline int match_futex(union futex_key
*key1
, union futex_key
*key2
)
294 && key1
->both
.word
== key2
->both
.word
295 && key1
->both
.ptr
== key2
->both
.ptr
296 && key1
->both
.offset
== key2
->both
.offset
);
300 * Take a reference to the resource addressed by a key.
301 * Can be called while holding spinlocks.
304 static void get_futex_key_refs(union futex_key
*key
)
309 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
311 ihold(key
->shared
.inode
); /* implies MB (B) */
313 case FUT_OFF_MMSHARED
:
314 futex_get_mm(key
); /* implies MB (B) */
320 * Drop a reference to the resource addressed by a key.
321 * The hash bucket spinlock must not be held.
323 static void drop_futex_key_refs(union futex_key
*key
)
325 if (!key
->both
.ptr
) {
326 /* If we're here then we tried to put a key we failed to get */
331 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
333 iput(key
->shared
.inode
);
335 case FUT_OFF_MMSHARED
:
336 mmdrop(key
->private.mm
);
342 * get_futex_key() - Get parameters which are the keys for a futex
343 * @uaddr: virtual address of the futex
344 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
345 * @key: address where result is stored.
346 * @rw: mapping needs to be read/write (values: VERIFY_READ,
349 * Return: a negative error code or 0
351 * The key words are stored in *key on success.
353 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
354 * offset_within_page). For private mappings, it's (uaddr, current->mm).
355 * We can usually work out the index without swapping in the page.
357 * lock_page() might sleep, the caller should not hold a spinlock.
360 get_futex_key(u32 __user
*uaddr
, int fshared
, union futex_key
*key
, int rw
)
362 unsigned long address
= (unsigned long)uaddr
;
363 struct mm_struct
*mm
= current
->mm
;
364 struct page
*page
, *page_head
;
368 * The futex address must be "naturally" aligned.
370 key
->both
.offset
= address
% PAGE_SIZE
;
371 if (unlikely((address
% sizeof(u32
)) != 0))
373 address
-= key
->both
.offset
;
375 if (unlikely(!access_ok(rw
, uaddr
, sizeof(u32
))))
379 * PROCESS_PRIVATE futexes are fast.
380 * As the mm cannot disappear under us and the 'key' only needs
381 * virtual address, we dont even have to find the underlying vma.
382 * Note : We do have to check 'uaddr' is a valid user address,
383 * but access_ok() should be faster than find_vma()
386 key
->private.mm
= mm
;
387 key
->private.address
= address
;
388 get_futex_key_refs(key
); /* implies MB (B) */
393 err
= get_user_pages_fast(address
, 1, 1, &page
);
395 * If write access is not required (eg. FUTEX_WAIT), try
396 * and get read-only access.
398 if (err
== -EFAULT
&& rw
== VERIFY_READ
) {
399 err
= get_user_pages_fast(address
, 1, 0, &page
);
407 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
409 if (unlikely(PageTail(page
))) {
411 /* serialize against __split_huge_page_splitting() */
413 if (likely(__get_user_pages_fast(address
, 1, !ro
, &page
) == 1)) {
414 page_head
= compound_head(page
);
416 * page_head is valid pointer but we must pin
417 * it before taking the PG_lock and/or
418 * PG_compound_lock. The moment we re-enable
419 * irqs __split_huge_page_splitting() can
420 * return and the head page can be freed from
421 * under us. We can't take the PG_lock and/or
422 * PG_compound_lock on a page that could be
423 * freed from under us.
425 if (page
!= page_head
) {
436 page_head
= compound_head(page
);
437 if (page
!= page_head
) {
443 lock_page(page_head
);
446 * If page_head->mapping is NULL, then it cannot be a PageAnon
447 * page; but it might be the ZERO_PAGE or in the gate area or
448 * in a special mapping (all cases which we are happy to fail);
449 * or it may have been a good file page when get_user_pages_fast
450 * found it, but truncated or holepunched or subjected to
451 * invalidate_complete_page2 before we got the page lock (also
452 * cases which we are happy to fail). And we hold a reference,
453 * so refcount care in invalidate_complete_page's remove_mapping
454 * prevents drop_caches from setting mapping to NULL beneath us.
456 * The case we do have to guard against is when memory pressure made
457 * shmem_writepage move it from filecache to swapcache beneath us:
458 * an unlikely race, but we do need to retry for page_head->mapping.
460 if (!page_head
->mapping
) {
461 int shmem_swizzled
= PageSwapCache(page_head
);
462 unlock_page(page_head
);
470 * Private mappings are handled in a simple way.
472 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
473 * it's a read-only handle, it's expected that futexes attach to
474 * the object not the particular process.
476 if (PageAnon(page_head
)) {
478 * A RO anonymous page will never change and thus doesn't make
479 * sense for futex operations.
486 key
->both
.offset
|= FUT_OFF_MMSHARED
; /* ref taken on mm */
487 key
->private.mm
= mm
;
488 key
->private.address
= address
;
490 key
->both
.offset
|= FUT_OFF_INODE
; /* inode-based key */
491 key
->shared
.inode
= page_head
->mapping
->host
;
492 key
->shared
.pgoff
= basepage_index(page
);
495 get_futex_key_refs(key
); /* implies MB (B) */
498 unlock_page(page_head
);
503 static inline void put_futex_key(union futex_key
*key
)
505 drop_futex_key_refs(key
);
509 * fault_in_user_writeable() - Fault in user address and verify RW access
510 * @uaddr: pointer to faulting user space address
512 * Slow path to fixup the fault we just took in the atomic write
515 * We have no generic implementation of a non-destructive write to the
516 * user address. We know that we faulted in the atomic pagefault
517 * disabled section so we can as well avoid the #PF overhead by
518 * calling get_user_pages() right away.
520 static int fault_in_user_writeable(u32 __user
*uaddr
)
522 struct mm_struct
*mm
= current
->mm
;
525 down_read(&mm
->mmap_sem
);
526 ret
= fixup_user_fault(current
, mm
, (unsigned long)uaddr
,
528 up_read(&mm
->mmap_sem
);
530 return ret
< 0 ? ret
: 0;
534 * futex_top_waiter() - Return the highest priority waiter on a futex
535 * @hb: the hash bucket the futex_q's reside in
536 * @key: the futex key (to distinguish it from other futex futex_q's)
538 * Must be called with the hb lock held.
540 static struct futex_q
*futex_top_waiter(struct futex_hash_bucket
*hb
,
541 union futex_key
*key
)
543 struct futex_q
*this;
545 plist_for_each_entry(this, &hb
->chain
, list
) {
546 if (match_futex(&this->key
, key
))
552 static int cmpxchg_futex_value_locked(u32
*curval
, u32 __user
*uaddr
,
553 u32 uval
, u32 newval
)
558 ret
= futex_atomic_cmpxchg_inatomic(curval
, uaddr
, uval
, newval
);
564 static int get_futex_value_locked(u32
*dest
, u32 __user
*from
)
569 ret
= __copy_from_user_inatomic(dest
, from
, sizeof(u32
));
572 return ret
? -EFAULT
: 0;
579 static int refill_pi_state_cache(void)
581 struct futex_pi_state
*pi_state
;
583 if (likely(current
->pi_state_cache
))
586 pi_state
= kzalloc(sizeof(*pi_state
), GFP_KERNEL
);
591 INIT_LIST_HEAD(&pi_state
->list
);
592 /* pi_mutex gets initialized later */
593 pi_state
->owner
= NULL
;
594 atomic_set(&pi_state
->refcount
, 1);
595 pi_state
->key
= FUTEX_KEY_INIT
;
597 current
->pi_state_cache
= pi_state
;
602 static struct futex_pi_state
* alloc_pi_state(void)
604 struct futex_pi_state
*pi_state
= current
->pi_state_cache
;
607 current
->pi_state_cache
= NULL
;
612 static void free_pi_state(struct futex_pi_state
*pi_state
)
614 if (!atomic_dec_and_test(&pi_state
->refcount
))
618 * If pi_state->owner is NULL, the owner is most probably dying
619 * and has cleaned up the pi_state already
621 if (pi_state
->owner
) {
622 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
623 list_del_init(&pi_state
->list
);
624 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
626 rt_mutex_proxy_unlock(&pi_state
->pi_mutex
, pi_state
->owner
);
629 if (current
->pi_state_cache
)
633 * pi_state->list is already empty.
634 * clear pi_state->owner.
635 * refcount is at 0 - put it back to 1.
637 pi_state
->owner
= NULL
;
638 atomic_set(&pi_state
->refcount
, 1);
639 current
->pi_state_cache
= pi_state
;
644 * Look up the task based on what TID userspace gave us.
647 static struct task_struct
* futex_find_get_task(pid_t pid
)
649 struct task_struct
*p
;
652 p
= find_task_by_vpid(pid
);
662 * This task is holding PI mutexes at exit time => bad.
663 * Kernel cleans up PI-state, but userspace is likely hosed.
664 * (Robust-futex cleanup is separate and might save the day for userspace.)
666 void exit_pi_state_list(struct task_struct
*curr
)
668 struct list_head
*next
, *head
= &curr
->pi_state_list
;
669 struct futex_pi_state
*pi_state
;
670 struct futex_hash_bucket
*hb
;
671 union futex_key key
= FUTEX_KEY_INIT
;
673 if (!futex_cmpxchg_enabled
)
676 * We are a ZOMBIE and nobody can enqueue itself on
677 * pi_state_list anymore, but we have to be careful
678 * versus waiters unqueueing themselves:
680 raw_spin_lock_irq(&curr
->pi_lock
);
681 while (!list_empty(head
)) {
684 pi_state
= list_entry(next
, struct futex_pi_state
, list
);
686 hb
= hash_futex(&key
);
687 raw_spin_unlock_irq(&curr
->pi_lock
);
689 spin_lock(&hb
->lock
);
691 raw_spin_lock_irq(&curr
->pi_lock
);
693 * We dropped the pi-lock, so re-check whether this
694 * task still owns the PI-state:
696 if (head
->next
!= next
) {
697 spin_unlock(&hb
->lock
);
701 WARN_ON(pi_state
->owner
!= curr
);
702 WARN_ON(list_empty(&pi_state
->list
));
703 list_del_init(&pi_state
->list
);
704 pi_state
->owner
= NULL
;
705 raw_spin_unlock_irq(&curr
->pi_lock
);
707 rt_mutex_unlock(&pi_state
->pi_mutex
);
709 spin_unlock(&hb
->lock
);
711 raw_spin_lock_irq(&curr
->pi_lock
);
713 raw_spin_unlock_irq(&curr
->pi_lock
);
717 lookup_pi_state(u32 uval
, struct futex_hash_bucket
*hb
,
718 union futex_key
*key
, struct futex_pi_state
**ps
)
720 struct futex_pi_state
*pi_state
= NULL
;
721 struct futex_q
*this, *next
;
722 struct task_struct
*p
;
723 pid_t pid
= uval
& FUTEX_TID_MASK
;
725 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
726 if (match_futex(&this->key
, key
)) {
728 * Another waiter already exists - bump up
729 * the refcount and return its pi_state:
731 pi_state
= this->pi_state
;
733 * Userspace might have messed up non-PI and PI futexes
735 if (unlikely(!pi_state
))
738 WARN_ON(!atomic_read(&pi_state
->refcount
));
741 * When pi_state->owner is NULL then the owner died
742 * and another waiter is on the fly. pi_state->owner
743 * is fixed up by the task which acquires
744 * pi_state->rt_mutex.
746 * We do not check for pid == 0 which can happen when
747 * the owner died and robust_list_exit() cleared the
750 if (pid
&& pi_state
->owner
) {
752 * Bail out if user space manipulated the
755 if (pid
!= task_pid_vnr(pi_state
->owner
))
759 atomic_inc(&pi_state
->refcount
);
767 * We are the first waiter - try to look up the real owner and attach
768 * the new pi_state to it, but bail out when TID = 0
772 p
= futex_find_get_task(pid
);
777 * We need to look at the task state flags to figure out,
778 * whether the task is exiting. To protect against the do_exit
779 * change of the task flags, we do this protected by
782 raw_spin_lock_irq(&p
->pi_lock
);
783 if (unlikely(p
->flags
& PF_EXITING
)) {
785 * The task is on the way out. When PF_EXITPIDONE is
786 * set, we know that the task has finished the
789 int ret
= (p
->flags
& PF_EXITPIDONE
) ? -ESRCH
: -EAGAIN
;
791 raw_spin_unlock_irq(&p
->pi_lock
);
796 pi_state
= alloc_pi_state();
799 * Initialize the pi_mutex in locked state and make 'p'
802 rt_mutex_init_proxy_locked(&pi_state
->pi_mutex
, p
);
804 /* Store the key for possible exit cleanups: */
805 pi_state
->key
= *key
;
807 WARN_ON(!list_empty(&pi_state
->list
));
808 list_add(&pi_state
->list
, &p
->pi_state_list
);
810 raw_spin_unlock_irq(&p
->pi_lock
);
820 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
821 * @uaddr: the pi futex user address
822 * @hb: the pi futex hash bucket
823 * @key: the futex key associated with uaddr and hb
824 * @ps: the pi_state pointer where we store the result of the
826 * @task: the task to perform the atomic lock work for. This will
827 * be "current" except in the case of requeue pi.
828 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
832 * 1 - acquired the lock;
835 * The hb->lock and futex_key refs shall be held by the caller.
837 static int futex_lock_pi_atomic(u32 __user
*uaddr
, struct futex_hash_bucket
*hb
,
838 union futex_key
*key
,
839 struct futex_pi_state
**ps
,
840 struct task_struct
*task
, int set_waiters
)
842 int lock_taken
, ret
, force_take
= 0;
843 u32 uval
, newval
, curval
, vpid
= task_pid_vnr(task
);
846 ret
= lock_taken
= 0;
849 * To avoid races, we attempt to take the lock here again
850 * (by doing a 0 -> TID atomic cmpxchg), while holding all
851 * the locks. It will most likely not succeed.
855 newval
|= FUTEX_WAITERS
;
857 if (unlikely(cmpxchg_futex_value_locked(&curval
, uaddr
, 0, newval
)))
863 if ((unlikely((curval
& FUTEX_TID_MASK
) == vpid
)))
867 * Surprise - we got the lock. Just return to userspace:
869 if (unlikely(!curval
))
875 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
876 * to wake at the next unlock.
878 newval
= curval
| FUTEX_WAITERS
;
881 * Should we force take the futex? See below.
883 if (unlikely(force_take
)) {
885 * Keep the OWNER_DIED and the WAITERS bit and set the
888 newval
= (curval
& ~FUTEX_TID_MASK
) | vpid
;
893 if (unlikely(cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
)))
895 if (unlikely(curval
!= uval
))
899 * We took the lock due to forced take over.
901 if (unlikely(lock_taken
))
905 * We dont have the lock. Look up the PI state (or create it if
906 * we are the first waiter):
908 ret
= lookup_pi_state(uval
, hb
, key
, ps
);
914 * We failed to find an owner for this
915 * futex. So we have no pi_state to block
916 * on. This can happen in two cases:
919 * 2) A stale FUTEX_WAITERS bit
921 * Re-read the futex value.
923 if (get_futex_value_locked(&curval
, uaddr
))
927 * If the owner died or we have a stale
928 * WAITERS bit the owner TID in the user space
931 if (!(curval
& FUTEX_TID_MASK
)) {
944 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
945 * @q: The futex_q to unqueue
947 * The q->lock_ptr must not be NULL and must be held by the caller.
949 static void __unqueue_futex(struct futex_q
*q
)
951 struct futex_hash_bucket
*hb
;
953 if (WARN_ON_SMP(!q
->lock_ptr
|| !spin_is_locked(q
->lock_ptr
))
954 || WARN_ON(plist_node_empty(&q
->list
)))
957 hb
= container_of(q
->lock_ptr
, struct futex_hash_bucket
, lock
);
958 plist_del(&q
->list
, &hb
->chain
);
962 * The hash bucket lock must be held when this is called.
963 * Afterwards, the futex_q must not be accessed.
965 static void wake_futex(struct futex_q
*q
)
967 struct task_struct
*p
= q
->task
;
969 if (WARN(q
->pi_state
|| q
->rt_waiter
, "refusing to wake PI futex\n"))
973 * We set q->lock_ptr = NULL _before_ we wake up the task. If
974 * a non-futex wake up happens on another CPU then the task
975 * might exit and p would dereference a non-existing task
976 * struct. Prevent this by holding a reference on p across the
983 * The waiting task can free the futex_q as soon as
984 * q->lock_ptr = NULL is written, without taking any locks. A
985 * memory barrier is required here to prevent the following
986 * store to lock_ptr from getting ahead of the plist_del.
991 wake_up_state(p
, TASK_NORMAL
);
995 static int wake_futex_pi(u32 __user
*uaddr
, u32 uval
, struct futex_q
*this)
997 struct task_struct
*new_owner
;
998 struct futex_pi_state
*pi_state
= this->pi_state
;
999 u32
uninitialized_var(curval
), newval
;
1005 * If current does not own the pi_state then the futex is
1006 * inconsistent and user space fiddled with the futex value.
1008 if (pi_state
->owner
!= current
)
1011 raw_spin_lock(&pi_state
->pi_mutex
.wait_lock
);
1012 new_owner
= rt_mutex_next_owner(&pi_state
->pi_mutex
);
1015 * It is possible that the next waiter (the one that brought
1016 * this owner to the kernel) timed out and is no longer
1017 * waiting on the lock.
1020 new_owner
= this->task
;
1023 * We pass it to the next owner. (The WAITERS bit is always
1024 * kept enabled while there is PI state around. We must also
1025 * preserve the owner died bit.)
1027 if (!(uval
& FUTEX_OWNER_DIED
)) {
1030 newval
= FUTEX_WAITERS
| task_pid_vnr(new_owner
);
1032 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
))
1034 else if (curval
!= uval
)
1037 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
1042 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
1043 WARN_ON(list_empty(&pi_state
->list
));
1044 list_del_init(&pi_state
->list
);
1045 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
1047 raw_spin_lock_irq(&new_owner
->pi_lock
);
1048 WARN_ON(!list_empty(&pi_state
->list
));
1049 list_add(&pi_state
->list
, &new_owner
->pi_state_list
);
1050 pi_state
->owner
= new_owner
;
1051 raw_spin_unlock_irq(&new_owner
->pi_lock
);
1053 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
1054 rt_mutex_unlock(&pi_state
->pi_mutex
);
1059 static int unlock_futex_pi(u32 __user
*uaddr
, u32 uval
)
1061 u32
uninitialized_var(oldval
);
1064 * There is no waiter, so we unlock the futex. The owner died
1065 * bit has not to be preserved here. We are the owner:
1067 if (cmpxchg_futex_value_locked(&oldval
, uaddr
, uval
, 0))
1076 * Express the locking dependencies for lockdep:
1079 double_lock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1082 spin_lock(&hb1
->lock
);
1084 spin_lock_nested(&hb2
->lock
, SINGLE_DEPTH_NESTING
);
1085 } else { /* hb1 > hb2 */
1086 spin_lock(&hb2
->lock
);
1087 spin_lock_nested(&hb1
->lock
, SINGLE_DEPTH_NESTING
);
1092 double_unlock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1094 spin_unlock(&hb1
->lock
);
1096 spin_unlock(&hb2
->lock
);
1100 * Wake up waiters matching bitset queued on this futex (uaddr).
1103 futex_wake(u32 __user
*uaddr
, unsigned int flags
, int nr_wake
, u32 bitset
)
1105 struct futex_hash_bucket
*hb
;
1106 struct futex_q
*this, *next
;
1107 union futex_key key
= FUTEX_KEY_INIT
;
1113 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_READ
);
1114 if (unlikely(ret
!= 0))
1117 hb
= hash_futex(&key
);
1119 /* Make sure we really have tasks to wakeup */
1120 if (!hb_waiters_pending(hb
))
1123 spin_lock(&hb
->lock
);
1125 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
1126 if (match_futex (&this->key
, &key
)) {
1127 if (this->pi_state
|| this->rt_waiter
) {
1132 /* Check if one of the bits is set in both bitsets */
1133 if (!(this->bitset
& bitset
))
1137 if (++ret
>= nr_wake
)
1142 spin_unlock(&hb
->lock
);
1144 put_futex_key(&key
);
1150 * Wake up all waiters hashed on the physical page that is mapped
1151 * to this virtual address:
1154 futex_wake_op(u32 __user
*uaddr1
, unsigned int flags
, u32 __user
*uaddr2
,
1155 int nr_wake
, int nr_wake2
, int op
)
1157 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1158 struct futex_hash_bucket
*hb1
, *hb2
;
1159 struct futex_q
*this, *next
;
1163 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1164 if (unlikely(ret
!= 0))
1166 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
1167 if (unlikely(ret
!= 0))
1170 hb1
= hash_futex(&key1
);
1171 hb2
= hash_futex(&key2
);
1174 double_lock_hb(hb1
, hb2
);
1175 op_ret
= futex_atomic_op_inuser(op
, uaddr2
);
1176 if (unlikely(op_ret
< 0)) {
1178 double_unlock_hb(hb1
, hb2
);
1182 * we don't get EFAULT from MMU faults if we don't have an MMU,
1183 * but we might get them from range checking
1189 if (unlikely(op_ret
!= -EFAULT
)) {
1194 ret
= fault_in_user_writeable(uaddr2
);
1198 if (!(flags
& FLAGS_SHARED
))
1201 put_futex_key(&key2
);
1202 put_futex_key(&key1
);
1206 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1207 if (match_futex (&this->key
, &key1
)) {
1208 if (this->pi_state
|| this->rt_waiter
) {
1213 if (++ret
>= nr_wake
)
1220 plist_for_each_entry_safe(this, next
, &hb2
->chain
, list
) {
1221 if (match_futex (&this->key
, &key2
)) {
1222 if (this->pi_state
|| this->rt_waiter
) {
1227 if (++op_ret
>= nr_wake2
)
1235 double_unlock_hb(hb1
, hb2
);
1237 put_futex_key(&key2
);
1239 put_futex_key(&key1
);
1245 * requeue_futex() - Requeue a futex_q from one hb to another
1246 * @q: the futex_q to requeue
1247 * @hb1: the source hash_bucket
1248 * @hb2: the target hash_bucket
1249 * @key2: the new key for the requeued futex_q
1252 void requeue_futex(struct futex_q
*q
, struct futex_hash_bucket
*hb1
,
1253 struct futex_hash_bucket
*hb2
, union futex_key
*key2
)
1257 * If key1 and key2 hash to the same bucket, no need to
1260 if (likely(&hb1
->chain
!= &hb2
->chain
)) {
1261 plist_del(&q
->list
, &hb1
->chain
);
1262 plist_add(&q
->list
, &hb2
->chain
);
1263 q
->lock_ptr
= &hb2
->lock
;
1265 get_futex_key_refs(key2
);
1270 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1272 * @key: the key of the requeue target futex
1273 * @hb: the hash_bucket of the requeue target futex
1275 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1276 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1277 * to the requeue target futex so the waiter can detect the wakeup on the right
1278 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1279 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1280 * to protect access to the pi_state to fixup the owner later. Must be called
1281 * with both q->lock_ptr and hb->lock held.
1284 void requeue_pi_wake_futex(struct futex_q
*q
, union futex_key
*key
,
1285 struct futex_hash_bucket
*hb
)
1287 get_futex_key_refs(key
);
1292 WARN_ON(!q
->rt_waiter
);
1293 q
->rt_waiter
= NULL
;
1295 q
->lock_ptr
= &hb
->lock
;
1297 wake_up_state(q
->task
, TASK_NORMAL
);
1301 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1302 * @pifutex: the user address of the to futex
1303 * @hb1: the from futex hash bucket, must be locked by the caller
1304 * @hb2: the to futex hash bucket, must be locked by the caller
1305 * @key1: the from futex key
1306 * @key2: the to futex key
1307 * @ps: address to store the pi_state pointer
1308 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1310 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1311 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1312 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1313 * hb1 and hb2 must be held by the caller.
1316 * 0 - failed to acquire the lock atomically;
1317 * 1 - acquired the lock;
1320 static int futex_proxy_trylock_atomic(u32 __user
*pifutex
,
1321 struct futex_hash_bucket
*hb1
,
1322 struct futex_hash_bucket
*hb2
,
1323 union futex_key
*key1
, union futex_key
*key2
,
1324 struct futex_pi_state
**ps
, int set_waiters
)
1326 struct futex_q
*top_waiter
= NULL
;
1330 if (get_futex_value_locked(&curval
, pifutex
))
1334 * Find the top_waiter and determine if there are additional waiters.
1335 * If the caller intends to requeue more than 1 waiter to pifutex,
1336 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1337 * as we have means to handle the possible fault. If not, don't set
1338 * the bit unecessarily as it will force the subsequent unlock to enter
1341 top_waiter
= futex_top_waiter(hb1
, key1
);
1343 /* There are no waiters, nothing for us to do. */
1347 /* Ensure we requeue to the expected futex. */
1348 if (!match_futex(top_waiter
->requeue_pi_key
, key2
))
1352 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1353 * the contended case or if set_waiters is 1. The pi_state is returned
1354 * in ps in contended cases.
1356 ret
= futex_lock_pi_atomic(pifutex
, hb2
, key2
, ps
, top_waiter
->task
,
1359 requeue_pi_wake_futex(top_waiter
, key2
, hb2
);
1365 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1366 * @uaddr1: source futex user address
1367 * @flags: futex flags (FLAGS_SHARED, etc.)
1368 * @uaddr2: target futex user address
1369 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1370 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1371 * @cmpval: @uaddr1 expected value (or %NULL)
1372 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1373 * pi futex (pi to pi requeue is not supported)
1375 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1376 * uaddr2 atomically on behalf of the top waiter.
1379 * >=0 - on success, the number of tasks requeued or woken;
1382 static int futex_requeue(u32 __user
*uaddr1
, unsigned int flags
,
1383 u32 __user
*uaddr2
, int nr_wake
, int nr_requeue
,
1384 u32
*cmpval
, int requeue_pi
)
1386 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1387 int drop_count
= 0, task_count
= 0, ret
;
1388 struct futex_pi_state
*pi_state
= NULL
;
1389 struct futex_hash_bucket
*hb1
, *hb2
;
1390 struct futex_q
*this, *next
;
1395 * requeue_pi requires a pi_state, try to allocate it now
1396 * without any locks in case it fails.
1398 if (refill_pi_state_cache())
1401 * requeue_pi must wake as many tasks as it can, up to nr_wake
1402 * + nr_requeue, since it acquires the rt_mutex prior to
1403 * returning to userspace, so as to not leave the rt_mutex with
1404 * waiters and no owner. However, second and third wake-ups
1405 * cannot be predicted as they involve race conditions with the
1406 * first wake and a fault while looking up the pi_state. Both
1407 * pthread_cond_signal() and pthread_cond_broadcast() should
1415 if (pi_state
!= NULL
) {
1417 * We will have to lookup the pi_state again, so free this one
1418 * to keep the accounting correct.
1420 free_pi_state(pi_state
);
1424 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1425 if (unlikely(ret
!= 0))
1427 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
,
1428 requeue_pi
? VERIFY_WRITE
: VERIFY_READ
);
1429 if (unlikely(ret
!= 0))
1432 hb1
= hash_futex(&key1
);
1433 hb2
= hash_futex(&key2
);
1436 double_lock_hb(hb1
, hb2
);
1438 if (likely(cmpval
!= NULL
)) {
1441 ret
= get_futex_value_locked(&curval
, uaddr1
);
1443 if (unlikely(ret
)) {
1444 double_unlock_hb(hb1
, hb2
);
1446 ret
= get_user(curval
, uaddr1
);
1450 if (!(flags
& FLAGS_SHARED
))
1453 put_futex_key(&key2
);
1454 put_futex_key(&key1
);
1457 if (curval
!= *cmpval
) {
1463 if (requeue_pi
&& (task_count
- nr_wake
< nr_requeue
)) {
1465 * Attempt to acquire uaddr2 and wake the top waiter. If we
1466 * intend to requeue waiters, force setting the FUTEX_WAITERS
1467 * bit. We force this here where we are able to easily handle
1468 * faults rather in the requeue loop below.
1470 ret
= futex_proxy_trylock_atomic(uaddr2
, hb1
, hb2
, &key1
,
1471 &key2
, &pi_state
, nr_requeue
);
1474 * At this point the top_waiter has either taken uaddr2 or is
1475 * waiting on it. If the former, then the pi_state will not
1476 * exist yet, look it up one more time to ensure we have a
1483 ret
= get_futex_value_locked(&curval2
, uaddr2
);
1485 ret
= lookup_pi_state(curval2
, hb2
, &key2
,
1493 double_unlock_hb(hb1
, hb2
);
1494 put_futex_key(&key2
);
1495 put_futex_key(&key1
);
1496 ret
= fault_in_user_writeable(uaddr2
);
1501 /* The owner was exiting, try again. */
1502 double_unlock_hb(hb1
, hb2
);
1503 put_futex_key(&key2
);
1504 put_futex_key(&key1
);
1512 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1513 if (task_count
- nr_wake
>= nr_requeue
)
1516 if (!match_futex(&this->key
, &key1
))
1520 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1521 * be paired with each other and no other futex ops.
1523 * We should never be requeueing a futex_q with a pi_state,
1524 * which is awaiting a futex_unlock_pi().
1526 if ((requeue_pi
&& !this->rt_waiter
) ||
1527 (!requeue_pi
&& this->rt_waiter
) ||
1534 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1535 * lock, we already woke the top_waiter. If not, it will be
1536 * woken by futex_unlock_pi().
1538 if (++task_count
<= nr_wake
&& !requeue_pi
) {
1543 /* Ensure we requeue to the expected futex for requeue_pi. */
1544 if (requeue_pi
&& !match_futex(this->requeue_pi_key
, &key2
)) {
1550 * Requeue nr_requeue waiters and possibly one more in the case
1551 * of requeue_pi if we couldn't acquire the lock atomically.
1554 /* Prepare the waiter to take the rt_mutex. */
1555 atomic_inc(&pi_state
->refcount
);
1556 this->pi_state
= pi_state
;
1557 ret
= rt_mutex_start_proxy_lock(&pi_state
->pi_mutex
,
1561 /* We got the lock. */
1562 requeue_pi_wake_futex(this, &key2
, hb2
);
1567 this->pi_state
= NULL
;
1568 free_pi_state(pi_state
);
1572 requeue_futex(this, hb1
, hb2
, &key2
);
1577 double_unlock_hb(hb1
, hb2
);
1580 * drop_futex_key_refs() must be called outside the spinlocks. During
1581 * the requeue we moved futex_q's from the hash bucket at key1 to the
1582 * one at key2 and updated their key pointer. We no longer need to
1583 * hold the references to key1.
1585 while (--drop_count
>= 0)
1586 drop_futex_key_refs(&key1
);
1589 put_futex_key(&key2
);
1591 put_futex_key(&key1
);
1593 if (pi_state
!= NULL
)
1594 free_pi_state(pi_state
);
1595 return ret
? ret
: task_count
;
1598 /* The key must be already stored in q->key. */
1599 static inline struct futex_hash_bucket
*queue_lock(struct futex_q
*q
)
1600 __acquires(&hb
->lock
)
1602 struct futex_hash_bucket
*hb
;
1604 hb
= hash_futex(&q
->key
);
1605 q
->lock_ptr
= &hb
->lock
;
1607 spin_lock(&hb
->lock
); /* implies MB (A) */
1612 queue_unlock(struct futex_hash_bucket
*hb
)
1613 __releases(&hb
->lock
)
1615 spin_unlock(&hb
->lock
);
1619 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1620 * @q: The futex_q to enqueue
1621 * @hb: The destination hash bucket
1623 * The hb->lock must be held by the caller, and is released here. A call to
1624 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1625 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1626 * or nothing if the unqueue is done as part of the wake process and the unqueue
1627 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1630 static inline void queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
1631 __releases(&hb
->lock
)
1636 * The priority used to register this element is
1637 * - either the real thread-priority for the real-time threads
1638 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1639 * - or MAX_RT_PRIO for non-RT threads.
1640 * Thus, all RT-threads are woken first in priority order, and
1641 * the others are woken last, in FIFO order.
1643 prio
= min(current
->normal_prio
, MAX_RT_PRIO
);
1645 plist_node_init(&q
->list
, prio
);
1646 plist_add(&q
->list
, &hb
->chain
);
1648 spin_unlock(&hb
->lock
);
1652 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1653 * @q: The futex_q to unqueue
1655 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1656 * be paired with exactly one earlier call to queue_me().
1659 * 1 - if the futex_q was still queued (and we removed unqueued it);
1660 * 0 - if the futex_q was already removed by the waking thread
1662 static int unqueue_me(struct futex_q
*q
)
1664 spinlock_t
*lock_ptr
;
1667 /* In the common case we don't take the spinlock, which is nice. */
1669 lock_ptr
= q
->lock_ptr
;
1671 if (lock_ptr
!= NULL
) {
1672 spin_lock(lock_ptr
);
1674 * q->lock_ptr can change between reading it and
1675 * spin_lock(), causing us to take the wrong lock. This
1676 * corrects the race condition.
1678 * Reasoning goes like this: if we have the wrong lock,
1679 * q->lock_ptr must have changed (maybe several times)
1680 * between reading it and the spin_lock(). It can
1681 * change again after the spin_lock() but only if it was
1682 * already changed before the spin_lock(). It cannot,
1683 * however, change back to the original value. Therefore
1684 * we can detect whether we acquired the correct lock.
1686 if (unlikely(lock_ptr
!= q
->lock_ptr
)) {
1687 spin_unlock(lock_ptr
);
1692 BUG_ON(q
->pi_state
);
1694 spin_unlock(lock_ptr
);
1698 drop_futex_key_refs(&q
->key
);
1703 * PI futexes can not be requeued and must remove themself from the
1704 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1707 static void unqueue_me_pi(struct futex_q
*q
)
1708 __releases(q
->lock_ptr
)
1712 BUG_ON(!q
->pi_state
);
1713 free_pi_state(q
->pi_state
);
1716 spin_unlock(q
->lock_ptr
);
1720 * Fixup the pi_state owner with the new owner.
1722 * Must be called with hash bucket lock held and mm->sem held for non
1725 static int fixup_pi_state_owner(u32 __user
*uaddr
, struct futex_q
*q
,
1726 struct task_struct
*newowner
)
1728 u32 newtid
= task_pid_vnr(newowner
) | FUTEX_WAITERS
;
1729 struct futex_pi_state
*pi_state
= q
->pi_state
;
1730 struct task_struct
*oldowner
= pi_state
->owner
;
1731 u32 uval
, uninitialized_var(curval
), newval
;
1735 if (!pi_state
->owner
)
1736 newtid
|= FUTEX_OWNER_DIED
;
1739 * We are here either because we stole the rtmutex from the
1740 * previous highest priority waiter or we are the highest priority
1741 * waiter but failed to get the rtmutex the first time.
1742 * We have to replace the newowner TID in the user space variable.
1743 * This must be atomic as we have to preserve the owner died bit here.
1745 * Note: We write the user space value _before_ changing the pi_state
1746 * because we can fault here. Imagine swapped out pages or a fork
1747 * that marked all the anonymous memory readonly for cow.
1749 * Modifying pi_state _before_ the user space value would
1750 * leave the pi_state in an inconsistent state when we fault
1751 * here, because we need to drop the hash bucket lock to
1752 * handle the fault. This might be observed in the PID check
1753 * in lookup_pi_state.
1756 if (get_futex_value_locked(&uval
, uaddr
))
1760 newval
= (uval
& FUTEX_OWNER_DIED
) | newtid
;
1762 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
))
1770 * We fixed up user space. Now we need to fix the pi_state
1773 if (pi_state
->owner
!= NULL
) {
1774 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
1775 WARN_ON(list_empty(&pi_state
->list
));
1776 list_del_init(&pi_state
->list
);
1777 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
1780 pi_state
->owner
= newowner
;
1782 raw_spin_lock_irq(&newowner
->pi_lock
);
1783 WARN_ON(!list_empty(&pi_state
->list
));
1784 list_add(&pi_state
->list
, &newowner
->pi_state_list
);
1785 raw_spin_unlock_irq(&newowner
->pi_lock
);
1789 * To handle the page fault we need to drop the hash bucket
1790 * lock here. That gives the other task (either the highest priority
1791 * waiter itself or the task which stole the rtmutex) the
1792 * chance to try the fixup of the pi_state. So once we are
1793 * back from handling the fault we need to check the pi_state
1794 * after reacquiring the hash bucket lock and before trying to
1795 * do another fixup. When the fixup has been done already we
1799 spin_unlock(q
->lock_ptr
);
1801 ret
= fault_in_user_writeable(uaddr
);
1803 spin_lock(q
->lock_ptr
);
1806 * Check if someone else fixed it for us:
1808 if (pi_state
->owner
!= oldowner
)
1817 static long futex_wait_restart(struct restart_block
*restart
);
1820 * fixup_owner() - Post lock pi_state and corner case management
1821 * @uaddr: user address of the futex
1822 * @q: futex_q (contains pi_state and access to the rt_mutex)
1823 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1825 * After attempting to lock an rt_mutex, this function is called to cleanup
1826 * the pi_state owner as well as handle race conditions that may allow us to
1827 * acquire the lock. Must be called with the hb lock held.
1830 * 1 - success, lock taken;
1831 * 0 - success, lock not taken;
1832 * <0 - on error (-EFAULT)
1834 static int fixup_owner(u32 __user
*uaddr
, struct futex_q
*q
, int locked
)
1836 struct task_struct
*owner
;
1841 * Got the lock. We might not be the anticipated owner if we
1842 * did a lock-steal - fix up the PI-state in that case:
1844 if (q
->pi_state
->owner
!= current
)
1845 ret
= fixup_pi_state_owner(uaddr
, q
, current
);
1850 * Catch the rare case, where the lock was released when we were on the
1851 * way back before we locked the hash bucket.
1853 if (q
->pi_state
->owner
== current
) {
1855 * Try to get the rt_mutex now. This might fail as some other
1856 * task acquired the rt_mutex after we removed ourself from the
1857 * rt_mutex waiters list.
1859 if (rt_mutex_trylock(&q
->pi_state
->pi_mutex
)) {
1865 * pi_state is incorrect, some other task did a lock steal and
1866 * we returned due to timeout or signal without taking the
1867 * rt_mutex. Too late.
1869 raw_spin_lock(&q
->pi_state
->pi_mutex
.wait_lock
);
1870 owner
= rt_mutex_owner(&q
->pi_state
->pi_mutex
);
1872 owner
= rt_mutex_next_owner(&q
->pi_state
->pi_mutex
);
1873 raw_spin_unlock(&q
->pi_state
->pi_mutex
.wait_lock
);
1874 ret
= fixup_pi_state_owner(uaddr
, q
, owner
);
1879 * Paranoia check. If we did not take the lock, then we should not be
1880 * the owner of the rt_mutex.
1882 if (rt_mutex_owner(&q
->pi_state
->pi_mutex
) == current
)
1883 printk(KERN_ERR
"fixup_owner: ret = %d pi-mutex: %p "
1884 "pi-state %p\n", ret
,
1885 q
->pi_state
->pi_mutex
.owner
,
1886 q
->pi_state
->owner
);
1889 return ret
? ret
: locked
;
1893 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1894 * @hb: the futex hash bucket, must be locked by the caller
1895 * @q: the futex_q to queue up on
1896 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1898 static void futex_wait_queue_me(struct futex_hash_bucket
*hb
, struct futex_q
*q
,
1899 struct hrtimer_sleeper
*timeout
)
1902 * The task state is guaranteed to be set before another task can
1903 * wake it. set_current_state() is implemented using set_mb() and
1904 * queue_me() calls spin_unlock() upon completion, both serializing
1905 * access to the hash list and forcing another memory barrier.
1907 set_current_state(TASK_INTERRUPTIBLE
);
1912 hrtimer_start_expires(&timeout
->timer
, HRTIMER_MODE_ABS
);
1913 if (!hrtimer_active(&timeout
->timer
))
1914 timeout
->task
= NULL
;
1918 * If we have been removed from the hash list, then another task
1919 * has tried to wake us, and we can skip the call to schedule().
1921 if (likely(!plist_node_empty(&q
->list
))) {
1923 * If the timer has already expired, current will already be
1924 * flagged for rescheduling. Only call schedule if there
1925 * is no timeout, or if it has yet to expire.
1927 if (!timeout
|| timeout
->task
)
1928 freezable_schedule();
1930 __set_current_state(TASK_RUNNING
);
1934 * futex_wait_setup() - Prepare to wait on a futex
1935 * @uaddr: the futex userspace address
1936 * @val: the expected value
1937 * @flags: futex flags (FLAGS_SHARED, etc.)
1938 * @q: the associated futex_q
1939 * @hb: storage for hash_bucket pointer to be returned to caller
1941 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1942 * compare it with the expected value. Handle atomic faults internally.
1943 * Return with the hb lock held and a q.key reference on success, and unlocked
1944 * with no q.key reference on failure.
1947 * 0 - uaddr contains val and hb has been locked;
1948 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
1950 static int futex_wait_setup(u32 __user
*uaddr
, u32 val
, unsigned int flags
,
1951 struct futex_q
*q
, struct futex_hash_bucket
**hb
)
1957 * Access the page AFTER the hash-bucket is locked.
1958 * Order is important:
1960 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1961 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1963 * The basic logical guarantee of a futex is that it blocks ONLY
1964 * if cond(var) is known to be true at the time of blocking, for
1965 * any cond. If we locked the hash-bucket after testing *uaddr, that
1966 * would open a race condition where we could block indefinitely with
1967 * cond(var) false, which would violate the guarantee.
1969 * On the other hand, we insert q and release the hash-bucket only
1970 * after testing *uaddr. This guarantees that futex_wait() will NOT
1971 * absorb a wakeup if *uaddr does not match the desired values
1972 * while the syscall executes.
1975 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
->key
, VERIFY_READ
);
1976 if (unlikely(ret
!= 0))
1980 *hb
= queue_lock(q
);
1982 ret
= get_futex_value_locked(&uval
, uaddr
);
1987 ret
= get_user(uval
, uaddr
);
1991 if (!(flags
& FLAGS_SHARED
))
1994 put_futex_key(&q
->key
);
2005 put_futex_key(&q
->key
);
2009 static int futex_wait(u32 __user
*uaddr
, unsigned int flags
, u32 val
,
2010 ktime_t
*abs_time
, u32 bitset
)
2012 struct hrtimer_sleeper timeout
, *to
= NULL
;
2013 struct restart_block
*restart
;
2014 struct futex_hash_bucket
*hb
;
2015 struct futex_q q
= futex_q_init
;
2025 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2026 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2028 hrtimer_init_sleeper(to
, current
);
2029 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2030 current
->timer_slack_ns
);
2035 * Prepare to wait on uaddr. On success, holds hb lock and increments
2038 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2042 /* queue_me and wait for wakeup, timeout, or a signal. */
2043 futex_wait_queue_me(hb
, &q
, to
);
2045 /* If we were woken (and unqueued), we succeeded, whatever. */
2047 /* unqueue_me() drops q.key ref */
2048 if (!unqueue_me(&q
))
2051 if (to
&& !to
->task
)
2055 * We expect signal_pending(current), but we might be the
2056 * victim of a spurious wakeup as well.
2058 if (!signal_pending(current
))
2065 restart
= ¤t_thread_info()->restart_block
;
2066 restart
->fn
= futex_wait_restart
;
2067 restart
->futex
.uaddr
= uaddr
;
2068 restart
->futex
.val
= val
;
2069 restart
->futex
.time
= abs_time
->tv64
;
2070 restart
->futex
.bitset
= bitset
;
2071 restart
->futex
.flags
= flags
| FLAGS_HAS_TIMEOUT
;
2073 ret
= -ERESTART_RESTARTBLOCK
;
2077 hrtimer_cancel(&to
->timer
);
2078 destroy_hrtimer_on_stack(&to
->timer
);
2084 static long futex_wait_restart(struct restart_block
*restart
)
2086 u32 __user
*uaddr
= restart
->futex
.uaddr
;
2087 ktime_t t
, *tp
= NULL
;
2089 if (restart
->futex
.flags
& FLAGS_HAS_TIMEOUT
) {
2090 t
.tv64
= restart
->futex
.time
;
2093 restart
->fn
= do_no_restart_syscall
;
2095 return (long)futex_wait(uaddr
, restart
->futex
.flags
,
2096 restart
->futex
.val
, tp
, restart
->futex
.bitset
);
2101 * Userspace tried a 0 -> TID atomic transition of the futex value
2102 * and failed. The kernel side here does the whole locking operation:
2103 * if there are waiters then it will block, it does PI, etc. (Due to
2104 * races the kernel might see a 0 value of the futex too.)
2106 static int futex_lock_pi(u32 __user
*uaddr
, unsigned int flags
, int detect
,
2107 ktime_t
*time
, int trylock
)
2109 struct hrtimer_sleeper timeout
, *to
= NULL
;
2110 struct futex_hash_bucket
*hb
;
2111 struct futex_q q
= futex_q_init
;
2114 if (refill_pi_state_cache())
2119 hrtimer_init_on_stack(&to
->timer
, CLOCK_REALTIME
,
2121 hrtimer_init_sleeper(to
, current
);
2122 hrtimer_set_expires(&to
->timer
, *time
);
2126 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
.key
, VERIFY_WRITE
);
2127 if (unlikely(ret
!= 0))
2131 hb
= queue_lock(&q
);
2133 ret
= futex_lock_pi_atomic(uaddr
, hb
, &q
.key
, &q
.pi_state
, current
, 0);
2134 if (unlikely(ret
)) {
2137 /* We got the lock. */
2139 goto out_unlock_put_key
;
2144 * Task is exiting and we just wait for the
2148 put_futex_key(&q
.key
);
2152 goto out_unlock_put_key
;
2157 * Only actually queue now that the atomic ops are done:
2161 WARN_ON(!q
.pi_state
);
2163 * Block on the PI mutex:
2166 ret
= rt_mutex_timed_lock(&q
.pi_state
->pi_mutex
, to
, 1);
2168 ret
= rt_mutex_trylock(&q
.pi_state
->pi_mutex
);
2169 /* Fixup the trylock return value: */
2170 ret
= ret
? 0 : -EWOULDBLOCK
;
2173 spin_lock(q
.lock_ptr
);
2175 * Fixup the pi_state owner and possibly acquire the lock if we
2178 res
= fixup_owner(uaddr
, &q
, !ret
);
2180 * If fixup_owner() returned an error, proprogate that. If it acquired
2181 * the lock, clear our -ETIMEDOUT or -EINTR.
2184 ret
= (res
< 0) ? res
: 0;
2187 * If fixup_owner() faulted and was unable to handle the fault, unlock
2188 * it and return the fault to userspace.
2190 if (ret
&& (rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
))
2191 rt_mutex_unlock(&q
.pi_state
->pi_mutex
);
2193 /* Unqueue and drop the lock */
2202 put_futex_key(&q
.key
);
2205 destroy_hrtimer_on_stack(&to
->timer
);
2206 return ret
!= -EINTR
? ret
: -ERESTARTNOINTR
;
2211 ret
= fault_in_user_writeable(uaddr
);
2215 if (!(flags
& FLAGS_SHARED
))
2218 put_futex_key(&q
.key
);
2223 * Userspace attempted a TID -> 0 atomic transition, and failed.
2224 * This is the in-kernel slowpath: we look up the PI state (if any),
2225 * and do the rt-mutex unlock.
2227 static int futex_unlock_pi(u32 __user
*uaddr
, unsigned int flags
)
2229 struct futex_hash_bucket
*hb
;
2230 struct futex_q
*this, *next
;
2231 union futex_key key
= FUTEX_KEY_INIT
;
2232 u32 uval
, vpid
= task_pid_vnr(current
);
2236 if (get_user(uval
, uaddr
))
2239 * We release only a lock we actually own:
2241 if ((uval
& FUTEX_TID_MASK
) != vpid
)
2244 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_WRITE
);
2245 if (unlikely(ret
!= 0))
2248 hb
= hash_futex(&key
);
2249 spin_lock(&hb
->lock
);
2252 * To avoid races, try to do the TID -> 0 atomic transition
2253 * again. If it succeeds then we can return without waking
2256 if (!(uval
& FUTEX_OWNER_DIED
) &&
2257 cmpxchg_futex_value_locked(&uval
, uaddr
, vpid
, 0))
2260 * Rare case: we managed to release the lock atomically,
2261 * no need to wake anyone else up:
2263 if (unlikely(uval
== vpid
))
2267 * Ok, other tasks may need to be woken up - check waiters
2268 * and do the wakeup if necessary:
2270 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
2271 if (!match_futex (&this->key
, &key
))
2273 ret
= wake_futex_pi(uaddr
, uval
, this);
2275 * The atomic access to the futex value
2276 * generated a pagefault, so retry the
2277 * user-access and the wakeup:
2284 * No waiters - kernel unlocks the futex:
2286 if (!(uval
& FUTEX_OWNER_DIED
)) {
2287 ret
= unlock_futex_pi(uaddr
, uval
);
2293 spin_unlock(&hb
->lock
);
2294 put_futex_key(&key
);
2300 spin_unlock(&hb
->lock
);
2301 put_futex_key(&key
);
2303 ret
= fault_in_user_writeable(uaddr
);
2311 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2312 * @hb: the hash_bucket futex_q was original enqueued on
2313 * @q: the futex_q woken while waiting to be requeued
2314 * @key2: the futex_key of the requeue target futex
2315 * @timeout: the timeout associated with the wait (NULL if none)
2317 * Detect if the task was woken on the initial futex as opposed to the requeue
2318 * target futex. If so, determine if it was a timeout or a signal that caused
2319 * the wakeup and return the appropriate error code to the caller. Must be
2320 * called with the hb lock held.
2323 * 0 = no early wakeup detected;
2324 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2327 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket
*hb
,
2328 struct futex_q
*q
, union futex_key
*key2
,
2329 struct hrtimer_sleeper
*timeout
)
2334 * With the hb lock held, we avoid races while we process the wakeup.
2335 * We only need to hold hb (and not hb2) to ensure atomicity as the
2336 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2337 * It can't be requeued from uaddr2 to something else since we don't
2338 * support a PI aware source futex for requeue.
2340 if (!match_futex(&q
->key
, key2
)) {
2341 WARN_ON(q
->lock_ptr
&& (&hb
->lock
!= q
->lock_ptr
));
2343 * We were woken prior to requeue by a timeout or a signal.
2344 * Unqueue the futex_q and determine which it was.
2346 plist_del(&q
->list
, &hb
->chain
);
2348 /* Handle spurious wakeups gracefully */
2350 if (timeout
&& !timeout
->task
)
2352 else if (signal_pending(current
))
2353 ret
= -ERESTARTNOINTR
;
2359 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2360 * @uaddr: the futex we initially wait on (non-pi)
2361 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2362 * the same type, no requeueing from private to shared, etc.
2363 * @val: the expected value of uaddr
2364 * @abs_time: absolute timeout
2365 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2366 * @uaddr2: the pi futex we will take prior to returning to user-space
2368 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2369 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2370 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2371 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2372 * without one, the pi logic would not know which task to boost/deboost, if
2373 * there was a need to.
2375 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2376 * via the following--
2377 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2378 * 2) wakeup on uaddr2 after a requeue
2382 * If 3, cleanup and return -ERESTARTNOINTR.
2384 * If 2, we may then block on trying to take the rt_mutex and return via:
2385 * 5) successful lock
2388 * 8) other lock acquisition failure
2390 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2392 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2398 static int futex_wait_requeue_pi(u32 __user
*uaddr
, unsigned int flags
,
2399 u32 val
, ktime_t
*abs_time
, u32 bitset
,
2402 struct hrtimer_sleeper timeout
, *to
= NULL
;
2403 struct rt_mutex_waiter rt_waiter
;
2404 struct rt_mutex
*pi_mutex
= NULL
;
2405 struct futex_hash_bucket
*hb
;
2406 union futex_key key2
= FUTEX_KEY_INIT
;
2407 struct futex_q q
= futex_q_init
;
2410 if (uaddr
== uaddr2
)
2418 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2419 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2421 hrtimer_init_sleeper(to
, current
);
2422 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2423 current
->timer_slack_ns
);
2427 * The waiter is allocated on our stack, manipulated by the requeue
2428 * code while we sleep on uaddr.
2430 debug_rt_mutex_init_waiter(&rt_waiter
);
2431 RB_CLEAR_NODE(&rt_waiter
.pi_tree_entry
);
2432 RB_CLEAR_NODE(&rt_waiter
.tree_entry
);
2433 rt_waiter
.task
= NULL
;
2435 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
2436 if (unlikely(ret
!= 0))
2440 q
.rt_waiter
= &rt_waiter
;
2441 q
.requeue_pi_key
= &key2
;
2444 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2447 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2451 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2452 futex_wait_queue_me(hb
, &q
, to
);
2454 spin_lock(&hb
->lock
);
2455 ret
= handle_early_requeue_pi_wakeup(hb
, &q
, &key2
, to
);
2456 spin_unlock(&hb
->lock
);
2461 * In order for us to be here, we know our q.key == key2, and since
2462 * we took the hb->lock above, we also know that futex_requeue() has
2463 * completed and we no longer have to concern ourselves with a wakeup
2464 * race with the atomic proxy lock acquisition by the requeue code. The
2465 * futex_requeue dropped our key1 reference and incremented our key2
2469 /* Check if the requeue code acquired the second futex for us. */
2472 * Got the lock. We might not be the anticipated owner if we
2473 * did a lock-steal - fix up the PI-state in that case.
2475 if (q
.pi_state
&& (q
.pi_state
->owner
!= current
)) {
2476 spin_lock(q
.lock_ptr
);
2477 ret
= fixup_pi_state_owner(uaddr2
, &q
, current
);
2478 spin_unlock(q
.lock_ptr
);
2482 * We have been woken up by futex_unlock_pi(), a timeout, or a
2483 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2486 WARN_ON(!q
.pi_state
);
2487 pi_mutex
= &q
.pi_state
->pi_mutex
;
2488 ret
= rt_mutex_finish_proxy_lock(pi_mutex
, to
, &rt_waiter
, 1);
2489 debug_rt_mutex_free_waiter(&rt_waiter
);
2491 spin_lock(q
.lock_ptr
);
2493 * Fixup the pi_state owner and possibly acquire the lock if we
2496 res
= fixup_owner(uaddr2
, &q
, !ret
);
2498 * If fixup_owner() returned an error, proprogate that. If it
2499 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2502 ret
= (res
< 0) ? res
: 0;
2504 /* Unqueue and drop the lock. */
2509 * If fixup_pi_state_owner() faulted and was unable to handle the
2510 * fault, unlock the rt_mutex and return the fault to userspace.
2512 if (ret
== -EFAULT
) {
2513 if (pi_mutex
&& rt_mutex_owner(pi_mutex
) == current
)
2514 rt_mutex_unlock(pi_mutex
);
2515 } else if (ret
== -EINTR
) {
2517 * We've already been requeued, but cannot restart by calling
2518 * futex_lock_pi() directly. We could restart this syscall, but
2519 * it would detect that the user space "val" changed and return
2520 * -EWOULDBLOCK. Save the overhead of the restart and return
2521 * -EWOULDBLOCK directly.
2527 put_futex_key(&q
.key
);
2529 put_futex_key(&key2
);
2533 hrtimer_cancel(&to
->timer
);
2534 destroy_hrtimer_on_stack(&to
->timer
);
2540 * Support for robust futexes: the kernel cleans up held futexes at
2543 * Implementation: user-space maintains a per-thread list of locks it
2544 * is holding. Upon do_exit(), the kernel carefully walks this list,
2545 * and marks all locks that are owned by this thread with the
2546 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2547 * always manipulated with the lock held, so the list is private and
2548 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2549 * field, to allow the kernel to clean up if the thread dies after
2550 * acquiring the lock, but just before it could have added itself to
2551 * the list. There can only be one such pending lock.
2555 * sys_set_robust_list() - Set the robust-futex list head of a task
2556 * @head: pointer to the list-head
2557 * @len: length of the list-head, as userspace expects
2559 SYSCALL_DEFINE2(set_robust_list
, struct robust_list_head __user
*, head
,
2562 if (!futex_cmpxchg_enabled
)
2565 * The kernel knows only one size for now:
2567 if (unlikely(len
!= sizeof(*head
)))
2570 current
->robust_list
= head
;
2576 * sys_get_robust_list() - Get the robust-futex list head of a task
2577 * @pid: pid of the process [zero for current task]
2578 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2579 * @len_ptr: pointer to a length field, the kernel fills in the header size
2581 SYSCALL_DEFINE3(get_robust_list
, int, pid
,
2582 struct robust_list_head __user
* __user
*, head_ptr
,
2583 size_t __user
*, len_ptr
)
2585 struct robust_list_head __user
*head
;
2587 struct task_struct
*p
;
2589 if (!futex_cmpxchg_enabled
)
2598 p
= find_task_by_vpid(pid
);
2604 if (!ptrace_may_access(p
, PTRACE_MODE_READ
))
2607 head
= p
->robust_list
;
2610 if (put_user(sizeof(*head
), len_ptr
))
2612 return put_user(head
, head_ptr
);
2621 * Process a futex-list entry, check whether it's owned by the
2622 * dying task, and do notification if so:
2624 int handle_futex_death(u32 __user
*uaddr
, struct task_struct
*curr
, int pi
)
2626 u32 uval
, uninitialized_var(nval
), mval
;
2629 if (get_user(uval
, uaddr
))
2632 if ((uval
& FUTEX_TID_MASK
) == task_pid_vnr(curr
)) {
2634 * Ok, this dying thread is truly holding a futex
2635 * of interest. Set the OWNER_DIED bit atomically
2636 * via cmpxchg, and if the value had FUTEX_WAITERS
2637 * set, wake up a waiter (if any). (We have to do a
2638 * futex_wake() even if OWNER_DIED is already set -
2639 * to handle the rare but possible case of recursive
2640 * thread-death.) The rest of the cleanup is done in
2643 mval
= (uval
& FUTEX_WAITERS
) | FUTEX_OWNER_DIED
;
2645 * We are not holding a lock here, but we want to have
2646 * the pagefault_disable/enable() protection because
2647 * we want to handle the fault gracefully. If the
2648 * access fails we try to fault in the futex with R/W
2649 * verification via get_user_pages. get_user() above
2650 * does not guarantee R/W access. If that fails we
2651 * give up and leave the futex locked.
2653 if (cmpxchg_futex_value_locked(&nval
, uaddr
, uval
, mval
)) {
2654 if (fault_in_user_writeable(uaddr
))
2662 * Wake robust non-PI futexes here. The wakeup of
2663 * PI futexes happens in exit_pi_state():
2665 if (!pi
&& (uval
& FUTEX_WAITERS
))
2666 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
2672 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2674 static inline int fetch_robust_entry(struct robust_list __user
**entry
,
2675 struct robust_list __user
* __user
*head
,
2678 unsigned long uentry
;
2680 if (get_user(uentry
, (unsigned long __user
*)head
))
2683 *entry
= (void __user
*)(uentry
& ~1UL);
2690 * Walk curr->robust_list (very carefully, it's a userspace list!)
2691 * and mark any locks found there dead, and notify any waiters.
2693 * We silently return on any sign of list-walking problem.
2695 void exit_robust_list(struct task_struct
*curr
)
2697 struct robust_list_head __user
*head
= curr
->robust_list
;
2698 struct robust_list __user
*entry
, *next_entry
, *pending
;
2699 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
2700 unsigned int uninitialized_var(next_pi
);
2701 unsigned long futex_offset
;
2704 if (!futex_cmpxchg_enabled
)
2708 * Fetch the list head (which was registered earlier, via
2709 * sys_set_robust_list()):
2711 if (fetch_robust_entry(&entry
, &head
->list
.next
, &pi
))
2714 * Fetch the relative futex offset:
2716 if (get_user(futex_offset
, &head
->futex_offset
))
2719 * Fetch any possibly pending lock-add first, and handle it
2722 if (fetch_robust_entry(&pending
, &head
->list_op_pending
, &pip
))
2725 next_entry
= NULL
; /* avoid warning with gcc */
2726 while (entry
!= &head
->list
) {
2728 * Fetch the next entry in the list before calling
2729 * handle_futex_death:
2731 rc
= fetch_robust_entry(&next_entry
, &entry
->next
, &next_pi
);
2733 * A pending lock might already be on the list, so
2734 * don't process it twice:
2736 if (entry
!= pending
)
2737 if (handle_futex_death((void __user
*)entry
+ futex_offset
,
2745 * Avoid excessively long or circular lists:
2754 handle_futex_death((void __user
*)pending
+ futex_offset
,
2758 long do_futex(u32 __user
*uaddr
, int op
, u32 val
, ktime_t
*timeout
,
2759 u32 __user
*uaddr2
, u32 val2
, u32 val3
)
2761 int cmd
= op
& FUTEX_CMD_MASK
;
2762 unsigned int flags
= 0;
2764 if (!(op
& FUTEX_PRIVATE_FLAG
))
2765 flags
|= FLAGS_SHARED
;
2767 if (op
& FUTEX_CLOCK_REALTIME
) {
2768 flags
|= FLAGS_CLOCKRT
;
2769 if (cmd
!= FUTEX_WAIT_BITSET
&& cmd
!= FUTEX_WAIT_REQUEUE_PI
)
2775 case FUTEX_UNLOCK_PI
:
2776 case FUTEX_TRYLOCK_PI
:
2777 case FUTEX_WAIT_REQUEUE_PI
:
2778 case FUTEX_CMP_REQUEUE_PI
:
2779 if (!futex_cmpxchg_enabled
)
2785 val3
= FUTEX_BITSET_MATCH_ANY
;
2786 case FUTEX_WAIT_BITSET
:
2787 return futex_wait(uaddr
, flags
, val
, timeout
, val3
);
2789 val3
= FUTEX_BITSET_MATCH_ANY
;
2790 case FUTEX_WAKE_BITSET
:
2791 return futex_wake(uaddr
, flags
, val
, val3
);
2793 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, NULL
, 0);
2794 case FUTEX_CMP_REQUEUE
:
2795 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 0);
2797 return futex_wake_op(uaddr
, flags
, uaddr2
, val
, val2
, val3
);
2799 return futex_lock_pi(uaddr
, flags
, val
, timeout
, 0);
2800 case FUTEX_UNLOCK_PI
:
2801 return futex_unlock_pi(uaddr
, flags
);
2802 case FUTEX_TRYLOCK_PI
:
2803 return futex_lock_pi(uaddr
, flags
, 0, timeout
, 1);
2804 case FUTEX_WAIT_REQUEUE_PI
:
2805 val3
= FUTEX_BITSET_MATCH_ANY
;
2806 return futex_wait_requeue_pi(uaddr
, flags
, val
, timeout
, val3
,
2808 case FUTEX_CMP_REQUEUE_PI
:
2809 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 1);
2815 SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
2816 struct timespec __user
*, utime
, u32 __user
*, uaddr2
,
2820 ktime_t t
, *tp
= NULL
;
2822 int cmd
= op
& FUTEX_CMD_MASK
;
2824 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
2825 cmd
== FUTEX_WAIT_BITSET
||
2826 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
2827 if (copy_from_user(&ts
, utime
, sizeof(ts
)) != 0)
2829 if (!timespec_valid(&ts
))
2832 t
= timespec_to_ktime(ts
);
2833 if (cmd
== FUTEX_WAIT
)
2834 t
= ktime_add_safe(ktime_get(), t
);
2838 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2839 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2841 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
2842 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
2843 val2
= (u32
) (unsigned long) utime
;
2845 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
2848 static void __init
futex_detect_cmpxchg(void)
2850 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
2854 * This will fail and we want it. Some arch implementations do
2855 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2856 * functionality. We want to know that before we call in any
2857 * of the complex code paths. Also we want to prevent
2858 * registration of robust lists in that case. NULL is
2859 * guaranteed to fault and we get -EFAULT on functional
2860 * implementation, the non-functional ones will return
2863 if (cmpxchg_futex_value_locked(&curval
, NULL
, 0, 0) == -EFAULT
)
2864 futex_cmpxchg_enabled
= 1;
2868 static int __init
futex_init(void)
2870 unsigned int futex_shift
;
2873 #if CONFIG_BASE_SMALL
2874 futex_hashsize
= 16;
2876 futex_hashsize
= roundup_pow_of_two(256 * num_possible_cpus());
2879 futex_queues
= alloc_large_system_hash("futex", sizeof(*futex_queues
),
2881 futex_hashsize
< 256 ? HASH_SMALL
: 0,
2883 futex_hashsize
, futex_hashsize
);
2884 futex_hashsize
= 1UL << futex_shift
;
2886 futex_detect_cmpxchg();
2888 for (i
= 0; i
< futex_hashsize
; i
++) {
2889 plist_head_init(&futex_queues
[i
].chain
);
2890 spin_lock_init(&futex_queues
[i
].lock
);
2895 __initcall(futex_init
);