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>
66 #include <asm/futex.h>
68 #include "rtmutex_common.h"
70 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
71 int __read_mostly futex_cmpxchg_enabled
;
74 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
77 * Futex flags used to encode options to functions and preserve them across
80 #define FLAGS_SHARED 0x01
81 #define FLAGS_CLOCKRT 0x02
82 #define FLAGS_HAS_TIMEOUT 0x04
85 * Priority Inheritance state:
87 struct futex_pi_state
{
89 * list of 'owned' pi_state instances - these have to be
90 * cleaned up in do_exit() if the task exits prematurely:
92 struct list_head list
;
97 struct rt_mutex pi_mutex
;
99 struct task_struct
*owner
;
106 * struct futex_q - The hashed futex queue entry, one per waiting task
107 * @list: priority-sorted list of tasks waiting on this futex
108 * @task: the task waiting on the futex
109 * @lock_ptr: the hash bucket lock
110 * @key: the key the futex is hashed on
111 * @pi_state: optional priority inheritance state
112 * @rt_waiter: rt_waiter storage for use with requeue_pi
113 * @requeue_pi_key: the requeue_pi target futex key
114 * @bitset: bitset for the optional bitmasked wakeup
116 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
117 * we can wake only the relevant ones (hashed queues may be shared).
119 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
120 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
121 * The order of wakeup is always to make the first condition true, then
124 * PI futexes are typically woken before they are removed from the hash list via
125 * the rt_mutex code. See unqueue_me_pi().
128 struct plist_node list
;
130 struct task_struct
*task
;
131 spinlock_t
*lock_ptr
;
133 struct futex_pi_state
*pi_state
;
134 struct rt_mutex_waiter
*rt_waiter
;
135 union futex_key
*requeue_pi_key
;
139 static const struct futex_q futex_q_init
= {
140 /* list gets initialized in queue_me()*/
141 .key
= FUTEX_KEY_INIT
,
142 .bitset
= FUTEX_BITSET_MATCH_ANY
146 * Hash buckets are shared by all the futex_keys that hash to the same
147 * location. Each key may have multiple futex_q structures, one for each task
148 * waiting on a futex.
150 struct futex_hash_bucket
{
152 struct plist_head chain
;
155 static struct futex_hash_bucket futex_queues
[1<<FUTEX_HASHBITS
];
158 * We hash on the keys returned from get_futex_key (see below).
160 static struct futex_hash_bucket
*hash_futex(union futex_key
*key
)
162 u32 hash
= jhash2((u32
*)&key
->both
.word
,
163 (sizeof(key
->both
.word
)+sizeof(key
->both
.ptr
))/4,
165 return &futex_queues
[hash
& ((1 << FUTEX_HASHBITS
)-1)];
169 * Return 1 if two futex_keys are equal, 0 otherwise.
171 static inline int match_futex(union futex_key
*key1
, union futex_key
*key2
)
174 && key1
->both
.word
== key2
->both
.word
175 && key1
->both
.ptr
== key2
->both
.ptr
176 && key1
->both
.offset
== key2
->both
.offset
);
180 * Take a reference to the resource addressed by a key.
181 * Can be called while holding spinlocks.
184 static void get_futex_key_refs(union futex_key
*key
)
189 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
191 ihold(key
->shared
.inode
);
193 case FUT_OFF_MMSHARED
:
194 atomic_inc(&key
->private.mm
->mm_count
);
200 * Drop a reference to the resource addressed by a key.
201 * The hash bucket spinlock must not be held.
203 static void drop_futex_key_refs(union futex_key
*key
)
205 if (!key
->both
.ptr
) {
206 /* If we're here then we tried to put a key we failed to get */
211 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
213 iput(key
->shared
.inode
);
215 case FUT_OFF_MMSHARED
:
216 mmdrop(key
->private.mm
);
222 * get_futex_key() - Get parameters which are the keys for a futex
223 * @uaddr: virtual address of the futex
224 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
225 * @key: address where result is stored.
226 * @rw: mapping needs to be read/write (values: VERIFY_READ,
229 * Return: a negative error code or 0
231 * The key words are stored in *key on success.
233 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
234 * offset_within_page). For private mappings, it's (uaddr, current->mm).
235 * We can usually work out the index without swapping in the page.
237 * lock_page() might sleep, the caller should not hold a spinlock.
240 get_futex_key(u32 __user
*uaddr
, int fshared
, union futex_key
*key
, int rw
)
242 unsigned long address
= (unsigned long)uaddr
;
243 struct mm_struct
*mm
= current
->mm
;
244 struct page
*page
, *page_head
;
248 * The futex address must be "naturally" aligned.
250 key
->both
.offset
= address
% PAGE_SIZE
;
251 if (unlikely((address
% sizeof(u32
)) != 0))
253 address
-= key
->both
.offset
;
256 * PROCESS_PRIVATE futexes are fast.
257 * As the mm cannot disappear under us and the 'key' only needs
258 * virtual address, we dont even have to find the underlying vma.
259 * Note : We do have to check 'uaddr' is a valid user address,
260 * but access_ok() should be faster than find_vma()
263 if (unlikely(!access_ok(VERIFY_WRITE
, uaddr
, sizeof(u32
))))
265 key
->private.mm
= mm
;
266 key
->private.address
= address
;
267 get_futex_key_refs(key
);
272 err
= get_user_pages_fast(address
, 1, 1, &page
);
274 * If write access is not required (eg. FUTEX_WAIT), try
275 * and get read-only access.
277 if (err
== -EFAULT
&& rw
== VERIFY_READ
) {
278 err
= get_user_pages_fast(address
, 1, 0, &page
);
286 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
288 if (unlikely(PageTail(page
))) {
290 /* serialize against __split_huge_page_splitting() */
292 if (likely(__get_user_pages_fast(address
, 1, !ro
, &page
) == 1)) {
293 page_head
= compound_head(page
);
295 * page_head is valid pointer but we must pin
296 * it before taking the PG_lock and/or
297 * PG_compound_lock. The moment we re-enable
298 * irqs __split_huge_page_splitting() can
299 * return and the head page can be freed from
300 * under us. We can't take the PG_lock and/or
301 * PG_compound_lock on a page that could be
302 * freed from under us.
304 if (page
!= page_head
) {
315 page_head
= compound_head(page
);
316 if (page
!= page_head
) {
322 lock_page(page_head
);
325 * If page_head->mapping is NULL, then it cannot be a PageAnon
326 * page; but it might be the ZERO_PAGE or in the gate area or
327 * in a special mapping (all cases which we are happy to fail);
328 * or it may have been a good file page when get_user_pages_fast
329 * found it, but truncated or holepunched or subjected to
330 * invalidate_complete_page2 before we got the page lock (also
331 * cases which we are happy to fail). And we hold a reference,
332 * so refcount care in invalidate_complete_page's remove_mapping
333 * prevents drop_caches from setting mapping to NULL beneath us.
335 * The case we do have to guard against is when memory pressure made
336 * shmem_writepage move it from filecache to swapcache beneath us:
337 * an unlikely race, but we do need to retry for page_head->mapping.
339 if (!page_head
->mapping
) {
340 int shmem_swizzled
= PageSwapCache(page_head
);
341 unlock_page(page_head
);
349 * Private mappings are handled in a simple way.
351 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
352 * it's a read-only handle, it's expected that futexes attach to
353 * the object not the particular process.
355 if (PageAnon(page_head
)) {
357 * A RO anonymous page will never change and thus doesn't make
358 * sense for futex operations.
365 key
->both
.offset
|= FUT_OFF_MMSHARED
; /* ref taken on mm */
366 key
->private.mm
= mm
;
367 key
->private.address
= address
;
369 key
->both
.offset
|= FUT_OFF_INODE
; /* inode-based key */
370 key
->shared
.inode
= page_head
->mapping
->host
;
371 key
->shared
.pgoff
= basepage_index(page
);
374 get_futex_key_refs(key
);
377 unlock_page(page_head
);
382 static inline void put_futex_key(union futex_key
*key
)
384 drop_futex_key_refs(key
);
388 * fault_in_user_writeable() - Fault in user address and verify RW access
389 * @uaddr: pointer to faulting user space address
391 * Slow path to fixup the fault we just took in the atomic write
394 * We have no generic implementation of a non-destructive write to the
395 * user address. We know that we faulted in the atomic pagefault
396 * disabled section so we can as well avoid the #PF overhead by
397 * calling get_user_pages() right away.
399 static int fault_in_user_writeable(u32 __user
*uaddr
)
401 struct mm_struct
*mm
= current
->mm
;
404 down_read(&mm
->mmap_sem
);
405 ret
= fixup_user_fault(current
, mm
, (unsigned long)uaddr
,
407 up_read(&mm
->mmap_sem
);
409 return ret
< 0 ? ret
: 0;
413 * futex_top_waiter() - Return the highest priority waiter on a futex
414 * @hb: the hash bucket the futex_q's reside in
415 * @key: the futex key (to distinguish it from other futex futex_q's)
417 * Must be called with the hb lock held.
419 static struct futex_q
*futex_top_waiter(struct futex_hash_bucket
*hb
,
420 union futex_key
*key
)
422 struct futex_q
*this;
424 plist_for_each_entry(this, &hb
->chain
, list
) {
425 if (match_futex(&this->key
, key
))
431 static int cmpxchg_futex_value_locked(u32
*curval
, u32 __user
*uaddr
,
432 u32 uval
, u32 newval
)
437 ret
= futex_atomic_cmpxchg_inatomic(curval
, uaddr
, uval
, newval
);
443 static int get_futex_value_locked(u32
*dest
, u32 __user
*from
)
448 ret
= __copy_from_user_inatomic(dest
, from
, sizeof(u32
));
451 return ret
? -EFAULT
: 0;
458 static int refill_pi_state_cache(void)
460 struct futex_pi_state
*pi_state
;
462 if (likely(current
->pi_state_cache
))
465 pi_state
= kzalloc(sizeof(*pi_state
), GFP_KERNEL
);
470 INIT_LIST_HEAD(&pi_state
->list
);
471 /* pi_mutex gets initialized later */
472 pi_state
->owner
= NULL
;
473 atomic_set(&pi_state
->refcount
, 1);
474 pi_state
->key
= FUTEX_KEY_INIT
;
476 current
->pi_state_cache
= pi_state
;
481 static struct futex_pi_state
* alloc_pi_state(void)
483 struct futex_pi_state
*pi_state
= current
->pi_state_cache
;
486 current
->pi_state_cache
= NULL
;
491 static void free_pi_state(struct futex_pi_state
*pi_state
)
493 if (!atomic_dec_and_test(&pi_state
->refcount
))
497 * If pi_state->owner is NULL, the owner is most probably dying
498 * and has cleaned up the pi_state already
500 if (pi_state
->owner
) {
501 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
502 list_del_init(&pi_state
->list
);
503 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
505 rt_mutex_proxy_unlock(&pi_state
->pi_mutex
, pi_state
->owner
);
508 if (current
->pi_state_cache
)
512 * pi_state->list is already empty.
513 * clear pi_state->owner.
514 * refcount is at 0 - put it back to 1.
516 pi_state
->owner
= NULL
;
517 atomic_set(&pi_state
->refcount
, 1);
518 current
->pi_state_cache
= pi_state
;
523 * Look up the task based on what TID userspace gave us.
526 static struct task_struct
* futex_find_get_task(pid_t pid
)
528 struct task_struct
*p
;
531 p
= find_task_by_vpid(pid
);
541 * This task is holding PI mutexes at exit time => bad.
542 * Kernel cleans up PI-state, but userspace is likely hosed.
543 * (Robust-futex cleanup is separate and might save the day for userspace.)
545 void exit_pi_state_list(struct task_struct
*curr
)
547 struct list_head
*next
, *head
= &curr
->pi_state_list
;
548 struct futex_pi_state
*pi_state
;
549 struct futex_hash_bucket
*hb
;
550 union futex_key key
= FUTEX_KEY_INIT
;
552 if (!futex_cmpxchg_enabled
)
555 * We are a ZOMBIE and nobody can enqueue itself on
556 * pi_state_list anymore, but we have to be careful
557 * versus waiters unqueueing themselves:
559 raw_spin_lock_irq(&curr
->pi_lock
);
560 while (!list_empty(head
)) {
563 pi_state
= list_entry(next
, struct futex_pi_state
, list
);
565 hb
= hash_futex(&key
);
566 raw_spin_unlock_irq(&curr
->pi_lock
);
568 spin_lock(&hb
->lock
);
570 raw_spin_lock_irq(&curr
->pi_lock
);
572 * We dropped the pi-lock, so re-check whether this
573 * task still owns the PI-state:
575 if (head
->next
!= next
) {
576 spin_unlock(&hb
->lock
);
580 WARN_ON(pi_state
->owner
!= curr
);
581 WARN_ON(list_empty(&pi_state
->list
));
582 list_del_init(&pi_state
->list
);
583 pi_state
->owner
= NULL
;
584 raw_spin_unlock_irq(&curr
->pi_lock
);
586 rt_mutex_unlock(&pi_state
->pi_mutex
);
588 spin_unlock(&hb
->lock
);
590 raw_spin_lock_irq(&curr
->pi_lock
);
592 raw_spin_unlock_irq(&curr
->pi_lock
);
596 lookup_pi_state(u32 uval
, struct futex_hash_bucket
*hb
,
597 union futex_key
*key
, struct futex_pi_state
**ps
,
598 struct task_struct
*task
)
600 struct futex_pi_state
*pi_state
= NULL
;
601 struct futex_q
*this, *next
;
602 struct plist_head
*head
;
603 struct task_struct
*p
;
604 pid_t pid
= uval
& FUTEX_TID_MASK
;
608 plist_for_each_entry_safe(this, next
, head
, list
) {
609 if (match_futex(&this->key
, key
)) {
611 * Another waiter already exists - bump up
612 * the refcount and return its pi_state:
614 pi_state
= this->pi_state
;
616 * Userspace might have messed up non-PI and PI futexes
618 if (unlikely(!pi_state
))
621 WARN_ON(!atomic_read(&pi_state
->refcount
));
624 * When pi_state->owner is NULL then the owner died
625 * and another waiter is on the fly. pi_state->owner
626 * is fixed up by the task which acquires
627 * pi_state->rt_mutex.
629 * We do not check for pid == 0 which can happen when
630 * the owner died and robust_list_exit() cleared the
633 if (pid
&& pi_state
->owner
) {
635 * Bail out if user space manipulated the
638 if (pid
!= task_pid_vnr(pi_state
->owner
))
643 * Protect against a corrupted uval. If uval
644 * is 0x80000000 then pid is 0 and the waiter
645 * bit is set. So the deadlock check in the
646 * calling code has failed and we did not fall
647 * into the check above due to !pid.
649 if (task
&& pi_state
->owner
== task
)
652 atomic_inc(&pi_state
->refcount
);
660 * We are the first waiter - try to look up the real owner and attach
661 * the new pi_state to it, but bail out when TID = 0
665 p
= futex_find_get_task(pid
);
675 * We need to look at the task state flags to figure out,
676 * whether the task is exiting. To protect against the do_exit
677 * change of the task flags, we do this protected by
680 raw_spin_lock_irq(&p
->pi_lock
);
681 if (unlikely(p
->flags
& PF_EXITING
)) {
683 * The task is on the way out. When PF_EXITPIDONE is
684 * set, we know that the task has finished the
687 int ret
= (p
->flags
& PF_EXITPIDONE
) ? -ESRCH
: -EAGAIN
;
689 raw_spin_unlock_irq(&p
->pi_lock
);
694 pi_state
= alloc_pi_state();
697 * Initialize the pi_mutex in locked state and make 'p'
700 rt_mutex_init_proxy_locked(&pi_state
->pi_mutex
, p
);
702 /* Store the key for possible exit cleanups: */
703 pi_state
->key
= *key
;
705 WARN_ON(!list_empty(&pi_state
->list
));
706 list_add(&pi_state
->list
, &p
->pi_state_list
);
708 raw_spin_unlock_irq(&p
->pi_lock
);
718 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
719 * @uaddr: the pi futex user address
720 * @hb: the pi futex hash bucket
721 * @key: the futex key associated with uaddr and hb
722 * @ps: the pi_state pointer where we store the result of the
724 * @task: the task to perform the atomic lock work for. This will
725 * be "current" except in the case of requeue pi.
726 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
730 * 1 - acquired the lock;
733 * The hb->lock and futex_key refs shall be held by the caller.
735 static int futex_lock_pi_atomic(u32 __user
*uaddr
, struct futex_hash_bucket
*hb
,
736 union futex_key
*key
,
737 struct futex_pi_state
**ps
,
738 struct task_struct
*task
, int set_waiters
)
740 int lock_taken
, ret
, force_take
= 0;
741 u32 uval
, newval
, curval
, vpid
= task_pid_vnr(task
);
744 ret
= lock_taken
= 0;
747 * To avoid races, we attempt to take the lock here again
748 * (by doing a 0 -> TID atomic cmpxchg), while holding all
749 * the locks. It will most likely not succeed.
753 newval
|= FUTEX_WAITERS
;
755 if (unlikely(cmpxchg_futex_value_locked(&curval
, uaddr
, 0, newval
)))
761 if ((unlikely((curval
& FUTEX_TID_MASK
) == vpid
)))
765 * Surprise - we got the lock, but we do not trust user space at all.
767 if (unlikely(!curval
)) {
769 * We verify whether there is kernel state for this
770 * futex. If not, we can safely assume, that the 0 ->
771 * TID transition is correct. If state exists, we do
772 * not bother to fixup the user space state as it was
775 return futex_top_waiter(hb
, key
) ? -EINVAL
: 1;
781 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
782 * to wake at the next unlock.
784 newval
= curval
| FUTEX_WAITERS
;
787 * Should we force take the futex? See below.
789 if (unlikely(force_take
)) {
791 * Keep the OWNER_DIED and the WAITERS bit and set the
794 newval
= (curval
& ~FUTEX_TID_MASK
) | vpid
;
799 if (unlikely(cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
)))
801 if (unlikely(curval
!= uval
))
805 * We took the lock due to forced take over.
807 if (unlikely(lock_taken
))
811 * We dont have the lock. Look up the PI state (or create it if
812 * we are the first waiter):
814 ret
= lookup_pi_state(uval
, hb
, key
, ps
, task
);
820 * We failed to find an owner for this
821 * futex. So we have no pi_state to block
822 * on. This can happen in two cases:
825 * 2) A stale FUTEX_WAITERS bit
827 * Re-read the futex value.
829 if (get_futex_value_locked(&curval
, uaddr
))
833 * If the owner died or we have a stale
834 * WAITERS bit the owner TID in the user space
837 if (!(curval
& FUTEX_TID_MASK
)) {
850 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
851 * @q: The futex_q to unqueue
853 * The q->lock_ptr must not be NULL and must be held by the caller.
855 static void __unqueue_futex(struct futex_q
*q
)
857 struct futex_hash_bucket
*hb
;
859 if (WARN_ON_SMP(!q
->lock_ptr
|| !spin_is_locked(q
->lock_ptr
))
860 || WARN_ON(plist_node_empty(&q
->list
)))
863 hb
= container_of(q
->lock_ptr
, struct futex_hash_bucket
, lock
);
864 plist_del(&q
->list
, &hb
->chain
);
868 * The hash bucket lock must be held when this is called.
869 * Afterwards, the futex_q must not be accessed.
871 static void wake_futex(struct futex_q
*q
)
873 struct task_struct
*p
= q
->task
;
875 if (WARN(q
->pi_state
|| q
->rt_waiter
, "refusing to wake PI futex\n"))
879 * We set q->lock_ptr = NULL _before_ we wake up the task. If
880 * a non-futex wake up happens on another CPU then the task
881 * might exit and p would dereference a non-existing task
882 * struct. Prevent this by holding a reference on p across the
889 * The waiting task can free the futex_q as soon as
890 * q->lock_ptr = NULL is written, without taking any locks. A
891 * memory barrier is required here to prevent the following
892 * store to lock_ptr from getting ahead of the plist_del.
897 wake_up_state(p
, TASK_NORMAL
);
901 static int wake_futex_pi(u32 __user
*uaddr
, u32 uval
, struct futex_q
*this)
903 struct task_struct
*new_owner
;
904 struct futex_pi_state
*pi_state
= this->pi_state
;
905 u32
uninitialized_var(curval
), newval
;
911 * If current does not own the pi_state then the futex is
912 * inconsistent and user space fiddled with the futex value.
914 if (pi_state
->owner
!= current
)
917 raw_spin_lock(&pi_state
->pi_mutex
.wait_lock
);
918 new_owner
= rt_mutex_next_owner(&pi_state
->pi_mutex
);
921 * It is possible that the next waiter (the one that brought
922 * this owner to the kernel) timed out and is no longer
923 * waiting on the lock.
926 new_owner
= this->task
;
929 * We pass it to the next owner. (The WAITERS bit is always
930 * kept enabled while there is PI state around. We must also
931 * preserve the owner died bit.)
933 if (!(uval
& FUTEX_OWNER_DIED
)) {
936 newval
= FUTEX_WAITERS
| task_pid_vnr(new_owner
);
938 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
))
940 else if (curval
!= uval
)
943 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
948 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
949 WARN_ON(list_empty(&pi_state
->list
));
950 list_del_init(&pi_state
->list
);
951 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
953 raw_spin_lock_irq(&new_owner
->pi_lock
);
954 WARN_ON(!list_empty(&pi_state
->list
));
955 list_add(&pi_state
->list
, &new_owner
->pi_state_list
);
956 pi_state
->owner
= new_owner
;
957 raw_spin_unlock_irq(&new_owner
->pi_lock
);
959 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
960 rt_mutex_unlock(&pi_state
->pi_mutex
);
965 static int unlock_futex_pi(u32 __user
*uaddr
, u32 uval
)
967 u32
uninitialized_var(oldval
);
970 * There is no waiter, so we unlock the futex. The owner died
971 * bit has not to be preserved here. We are the owner:
973 if (cmpxchg_futex_value_locked(&oldval
, uaddr
, uval
, 0))
982 * Express the locking dependencies for lockdep:
985 double_lock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
988 spin_lock(&hb1
->lock
);
990 spin_lock_nested(&hb2
->lock
, SINGLE_DEPTH_NESTING
);
991 } else { /* hb1 > hb2 */
992 spin_lock(&hb2
->lock
);
993 spin_lock_nested(&hb1
->lock
, SINGLE_DEPTH_NESTING
);
998 double_unlock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1000 spin_unlock(&hb1
->lock
);
1002 spin_unlock(&hb2
->lock
);
1006 * Wake up waiters matching bitset queued on this futex (uaddr).
1009 futex_wake(u32 __user
*uaddr
, unsigned int flags
, int nr_wake
, u32 bitset
)
1011 struct futex_hash_bucket
*hb
;
1012 struct futex_q
*this, *next
;
1013 struct plist_head
*head
;
1014 union futex_key key
= FUTEX_KEY_INIT
;
1020 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_READ
);
1021 if (unlikely(ret
!= 0))
1024 hb
= hash_futex(&key
);
1025 spin_lock(&hb
->lock
);
1028 plist_for_each_entry_safe(this, next
, head
, list
) {
1029 if (match_futex (&this->key
, &key
)) {
1030 if (this->pi_state
|| this->rt_waiter
) {
1035 /* Check if one of the bits is set in both bitsets */
1036 if (!(this->bitset
& bitset
))
1040 if (++ret
>= nr_wake
)
1045 spin_unlock(&hb
->lock
);
1046 put_futex_key(&key
);
1052 * Wake up all waiters hashed on the physical page that is mapped
1053 * to this virtual address:
1056 futex_wake_op(u32 __user
*uaddr1
, unsigned int flags
, u32 __user
*uaddr2
,
1057 int nr_wake
, int nr_wake2
, int op
)
1059 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1060 struct futex_hash_bucket
*hb1
, *hb2
;
1061 struct plist_head
*head
;
1062 struct futex_q
*this, *next
;
1066 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1067 if (unlikely(ret
!= 0))
1069 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
1070 if (unlikely(ret
!= 0))
1073 hb1
= hash_futex(&key1
);
1074 hb2
= hash_futex(&key2
);
1077 double_lock_hb(hb1
, hb2
);
1078 op_ret
= futex_atomic_op_inuser(op
, uaddr2
);
1079 if (unlikely(op_ret
< 0)) {
1081 double_unlock_hb(hb1
, hb2
);
1085 * we don't get EFAULT from MMU faults if we don't have an MMU,
1086 * but we might get them from range checking
1092 if (unlikely(op_ret
!= -EFAULT
)) {
1097 ret
= fault_in_user_writeable(uaddr2
);
1101 if (!(flags
& FLAGS_SHARED
))
1104 put_futex_key(&key2
);
1105 put_futex_key(&key1
);
1111 plist_for_each_entry_safe(this, next
, head
, list
) {
1112 if (match_futex (&this->key
, &key1
)) {
1113 if (this->pi_state
|| this->rt_waiter
) {
1118 if (++ret
>= nr_wake
)
1127 plist_for_each_entry_safe(this, next
, head
, list
) {
1128 if (match_futex (&this->key
, &key2
)) {
1129 if (this->pi_state
|| this->rt_waiter
) {
1134 if (++op_ret
>= nr_wake2
)
1142 double_unlock_hb(hb1
, hb2
);
1144 put_futex_key(&key2
);
1146 put_futex_key(&key1
);
1152 * requeue_futex() - Requeue a futex_q from one hb to another
1153 * @q: the futex_q to requeue
1154 * @hb1: the source hash_bucket
1155 * @hb2: the target hash_bucket
1156 * @key2: the new key for the requeued futex_q
1159 void requeue_futex(struct futex_q
*q
, struct futex_hash_bucket
*hb1
,
1160 struct futex_hash_bucket
*hb2
, union futex_key
*key2
)
1164 * If key1 and key2 hash to the same bucket, no need to
1167 if (likely(&hb1
->chain
!= &hb2
->chain
)) {
1168 plist_del(&q
->list
, &hb1
->chain
);
1169 plist_add(&q
->list
, &hb2
->chain
);
1170 q
->lock_ptr
= &hb2
->lock
;
1172 get_futex_key_refs(key2
);
1177 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1179 * @key: the key of the requeue target futex
1180 * @hb: the hash_bucket of the requeue target futex
1182 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1183 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1184 * to the requeue target futex so the waiter can detect the wakeup on the right
1185 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1186 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1187 * to protect access to the pi_state to fixup the owner later. Must be called
1188 * with both q->lock_ptr and hb->lock held.
1191 void requeue_pi_wake_futex(struct futex_q
*q
, union futex_key
*key
,
1192 struct futex_hash_bucket
*hb
)
1194 get_futex_key_refs(key
);
1199 WARN_ON(!q
->rt_waiter
);
1200 q
->rt_waiter
= NULL
;
1202 q
->lock_ptr
= &hb
->lock
;
1204 wake_up_state(q
->task
, TASK_NORMAL
);
1208 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1209 * @pifutex: the user address of the to futex
1210 * @hb1: the from futex hash bucket, must be locked by the caller
1211 * @hb2: the to futex hash bucket, must be locked by the caller
1212 * @key1: the from futex key
1213 * @key2: the to futex key
1214 * @ps: address to store the pi_state pointer
1215 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1217 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1218 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1219 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1220 * hb1 and hb2 must be held by the caller.
1223 * 0 - failed to acquire the lock atomically;
1224 * >0 - acquired the lock, return value is vpid of the top_waiter
1227 static int futex_proxy_trylock_atomic(u32 __user
*pifutex
,
1228 struct futex_hash_bucket
*hb1
,
1229 struct futex_hash_bucket
*hb2
,
1230 union futex_key
*key1
, union futex_key
*key2
,
1231 struct futex_pi_state
**ps
, int set_waiters
)
1233 struct futex_q
*top_waiter
= NULL
;
1237 if (get_futex_value_locked(&curval
, pifutex
))
1241 * Find the top_waiter and determine if there are additional waiters.
1242 * If the caller intends to requeue more than 1 waiter to pifutex,
1243 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1244 * as we have means to handle the possible fault. If not, don't set
1245 * the bit unecessarily as it will force the subsequent unlock to enter
1248 top_waiter
= futex_top_waiter(hb1
, key1
);
1250 /* There are no waiters, nothing for us to do. */
1254 /* Ensure we requeue to the expected futex. */
1255 if (!match_futex(top_waiter
->requeue_pi_key
, key2
))
1259 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1260 * the contended case or if set_waiters is 1. The pi_state is returned
1261 * in ps in contended cases.
1263 vpid
= task_pid_vnr(top_waiter
->task
);
1264 ret
= futex_lock_pi_atomic(pifutex
, hb2
, key2
, ps
, top_waiter
->task
,
1267 requeue_pi_wake_futex(top_waiter
, key2
, hb2
);
1274 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1275 * @uaddr1: source futex user address
1276 * @flags: futex flags (FLAGS_SHARED, etc.)
1277 * @uaddr2: target futex user address
1278 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1279 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1280 * @cmpval: @uaddr1 expected value (or %NULL)
1281 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1282 * pi futex (pi to pi requeue is not supported)
1284 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1285 * uaddr2 atomically on behalf of the top waiter.
1288 * >=0 - on success, the number of tasks requeued or woken;
1291 static int futex_requeue(u32 __user
*uaddr1
, unsigned int flags
,
1292 u32 __user
*uaddr2
, int nr_wake
, int nr_requeue
,
1293 u32
*cmpval
, int requeue_pi
)
1295 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1296 int drop_count
= 0, task_count
= 0, ret
;
1297 struct futex_pi_state
*pi_state
= NULL
;
1298 struct futex_hash_bucket
*hb1
, *hb2
;
1299 struct plist_head
*head1
;
1300 struct futex_q
*this, *next
;
1304 * Requeue PI only works on two distinct uaddrs. This
1305 * check is only valid for private futexes. See below.
1307 if (uaddr1
== uaddr2
)
1311 * requeue_pi requires a pi_state, try to allocate it now
1312 * without any locks in case it fails.
1314 if (refill_pi_state_cache())
1317 * requeue_pi must wake as many tasks as it can, up to nr_wake
1318 * + nr_requeue, since it acquires the rt_mutex prior to
1319 * returning to userspace, so as to not leave the rt_mutex with
1320 * waiters and no owner. However, second and third wake-ups
1321 * cannot be predicted as they involve race conditions with the
1322 * first wake and a fault while looking up the pi_state. Both
1323 * pthread_cond_signal() and pthread_cond_broadcast() should
1331 if (pi_state
!= NULL
) {
1333 * We will have to lookup the pi_state again, so free this one
1334 * to keep the accounting correct.
1336 free_pi_state(pi_state
);
1340 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1341 if (unlikely(ret
!= 0))
1343 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
,
1344 requeue_pi
? VERIFY_WRITE
: VERIFY_READ
);
1345 if (unlikely(ret
!= 0))
1349 * The check above which compares uaddrs is not sufficient for
1350 * shared futexes. We need to compare the keys:
1352 if (requeue_pi
&& match_futex(&key1
, &key2
)) {
1357 hb1
= hash_futex(&key1
);
1358 hb2
= hash_futex(&key2
);
1361 double_lock_hb(hb1
, hb2
);
1363 if (likely(cmpval
!= NULL
)) {
1366 ret
= get_futex_value_locked(&curval
, uaddr1
);
1368 if (unlikely(ret
)) {
1369 double_unlock_hb(hb1
, hb2
);
1371 ret
= get_user(curval
, uaddr1
);
1375 if (!(flags
& FLAGS_SHARED
))
1378 put_futex_key(&key2
);
1379 put_futex_key(&key1
);
1382 if (curval
!= *cmpval
) {
1388 if (requeue_pi
&& (task_count
- nr_wake
< nr_requeue
)) {
1390 * Attempt to acquire uaddr2 and wake the top waiter. If we
1391 * intend to requeue waiters, force setting the FUTEX_WAITERS
1392 * bit. We force this here where we are able to easily handle
1393 * faults rather in the requeue loop below.
1395 ret
= futex_proxy_trylock_atomic(uaddr2
, hb1
, hb2
, &key1
,
1396 &key2
, &pi_state
, nr_requeue
);
1399 * At this point the top_waiter has either taken uaddr2 or is
1400 * waiting on it. If the former, then the pi_state will not
1401 * exist yet, look it up one more time to ensure we have a
1402 * reference to it. If the lock was taken, ret contains the
1403 * vpid of the top waiter task.
1410 * If we acquired the lock, then the user
1411 * space value of uaddr2 should be vpid. It
1412 * cannot be changed by the top waiter as it
1413 * is blocked on hb2 lock if it tries to do
1414 * so. If something fiddled with it behind our
1415 * back the pi state lookup might unearth
1416 * it. So we rather use the known value than
1417 * rereading and handing potential crap to
1420 ret
= lookup_pi_state(ret
, hb2
, &key2
, &pi_state
, NULL
);
1427 double_unlock_hb(hb1
, hb2
);
1428 put_futex_key(&key2
);
1429 put_futex_key(&key1
);
1430 ret
= fault_in_user_writeable(uaddr2
);
1435 /* The owner was exiting, try again. */
1436 double_unlock_hb(hb1
, hb2
);
1437 put_futex_key(&key2
);
1438 put_futex_key(&key1
);
1446 head1
= &hb1
->chain
;
1447 plist_for_each_entry_safe(this, next
, head1
, list
) {
1448 if (task_count
- nr_wake
>= nr_requeue
)
1451 if (!match_futex(&this->key
, &key1
))
1455 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1456 * be paired with each other and no other futex ops.
1458 * We should never be requeueing a futex_q with a pi_state,
1459 * which is awaiting a futex_unlock_pi().
1461 if ((requeue_pi
&& !this->rt_waiter
) ||
1462 (!requeue_pi
&& this->rt_waiter
) ||
1469 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1470 * lock, we already woke the top_waiter. If not, it will be
1471 * woken by futex_unlock_pi().
1473 if (++task_count
<= nr_wake
&& !requeue_pi
) {
1478 /* Ensure we requeue to the expected futex for requeue_pi. */
1479 if (requeue_pi
&& !match_futex(this->requeue_pi_key
, &key2
)) {
1485 * Requeue nr_requeue waiters and possibly one more in the case
1486 * of requeue_pi if we couldn't acquire the lock atomically.
1489 /* Prepare the waiter to take the rt_mutex. */
1490 atomic_inc(&pi_state
->refcount
);
1491 this->pi_state
= pi_state
;
1492 ret
= rt_mutex_start_proxy_lock(&pi_state
->pi_mutex
,
1496 /* We got the lock. */
1497 requeue_pi_wake_futex(this, &key2
, hb2
);
1502 this->pi_state
= NULL
;
1503 free_pi_state(pi_state
);
1507 requeue_futex(this, hb1
, hb2
, &key2
);
1512 double_unlock_hb(hb1
, hb2
);
1515 * drop_futex_key_refs() must be called outside the spinlocks. During
1516 * the requeue we moved futex_q's from the hash bucket at key1 to the
1517 * one at key2 and updated their key pointer. We no longer need to
1518 * hold the references to key1.
1520 while (--drop_count
>= 0)
1521 drop_futex_key_refs(&key1
);
1524 put_futex_key(&key2
);
1526 put_futex_key(&key1
);
1528 if (pi_state
!= NULL
)
1529 free_pi_state(pi_state
);
1530 return ret
? ret
: task_count
;
1533 /* The key must be already stored in q->key. */
1534 static inline struct futex_hash_bucket
*queue_lock(struct futex_q
*q
)
1535 __acquires(&hb
->lock
)
1537 struct futex_hash_bucket
*hb
;
1539 hb
= hash_futex(&q
->key
);
1540 q
->lock_ptr
= &hb
->lock
;
1542 spin_lock(&hb
->lock
);
1547 queue_unlock(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
1548 __releases(&hb
->lock
)
1550 spin_unlock(&hb
->lock
);
1554 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1555 * @q: The futex_q to enqueue
1556 * @hb: The destination hash bucket
1558 * The hb->lock must be held by the caller, and is released here. A call to
1559 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1560 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1561 * or nothing if the unqueue is done as part of the wake process and the unqueue
1562 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1565 static inline void queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
1566 __releases(&hb
->lock
)
1571 * The priority used to register this element is
1572 * - either the real thread-priority for the real-time threads
1573 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1574 * - or MAX_RT_PRIO for non-RT threads.
1575 * Thus, all RT-threads are woken first in priority order, and
1576 * the others are woken last, in FIFO order.
1578 prio
= min(current
->normal_prio
, MAX_RT_PRIO
);
1580 plist_node_init(&q
->list
, prio
);
1581 plist_add(&q
->list
, &hb
->chain
);
1583 spin_unlock(&hb
->lock
);
1587 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1588 * @q: The futex_q to unqueue
1590 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1591 * be paired with exactly one earlier call to queue_me().
1594 * 1 - if the futex_q was still queued (and we removed unqueued it);
1595 * 0 - if the futex_q was already removed by the waking thread
1597 static int unqueue_me(struct futex_q
*q
)
1599 spinlock_t
*lock_ptr
;
1602 /* In the common case we don't take the spinlock, which is nice. */
1604 lock_ptr
= q
->lock_ptr
;
1606 if (lock_ptr
!= NULL
) {
1607 spin_lock(lock_ptr
);
1609 * q->lock_ptr can change between reading it and
1610 * spin_lock(), causing us to take the wrong lock. This
1611 * corrects the race condition.
1613 * Reasoning goes like this: if we have the wrong lock,
1614 * q->lock_ptr must have changed (maybe several times)
1615 * between reading it and the spin_lock(). It can
1616 * change again after the spin_lock() but only if it was
1617 * already changed before the spin_lock(). It cannot,
1618 * however, change back to the original value. Therefore
1619 * we can detect whether we acquired the correct lock.
1621 if (unlikely(lock_ptr
!= q
->lock_ptr
)) {
1622 spin_unlock(lock_ptr
);
1627 BUG_ON(q
->pi_state
);
1629 spin_unlock(lock_ptr
);
1633 drop_futex_key_refs(&q
->key
);
1638 * PI futexes can not be requeued and must remove themself from the
1639 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1642 static void unqueue_me_pi(struct futex_q
*q
)
1643 __releases(q
->lock_ptr
)
1647 BUG_ON(!q
->pi_state
);
1648 free_pi_state(q
->pi_state
);
1651 spin_unlock(q
->lock_ptr
);
1655 * Fixup the pi_state owner with the new owner.
1657 * Must be called with hash bucket lock held and mm->sem held for non
1660 static int fixup_pi_state_owner(u32 __user
*uaddr
, struct futex_q
*q
,
1661 struct task_struct
*newowner
)
1663 u32 newtid
= task_pid_vnr(newowner
) | FUTEX_WAITERS
;
1664 struct futex_pi_state
*pi_state
= q
->pi_state
;
1665 struct task_struct
*oldowner
= pi_state
->owner
;
1666 u32 uval
, uninitialized_var(curval
), newval
;
1670 if (!pi_state
->owner
)
1671 newtid
|= FUTEX_OWNER_DIED
;
1674 * We are here either because we stole the rtmutex from the
1675 * previous highest priority waiter or we are the highest priority
1676 * waiter but failed to get the rtmutex the first time.
1677 * We have to replace the newowner TID in the user space variable.
1678 * This must be atomic as we have to preserve the owner died bit here.
1680 * Note: We write the user space value _before_ changing the pi_state
1681 * because we can fault here. Imagine swapped out pages or a fork
1682 * that marked all the anonymous memory readonly for cow.
1684 * Modifying pi_state _before_ the user space value would
1685 * leave the pi_state in an inconsistent state when we fault
1686 * here, because we need to drop the hash bucket lock to
1687 * handle the fault. This might be observed in the PID check
1688 * in lookup_pi_state.
1691 if (get_futex_value_locked(&uval
, uaddr
))
1695 newval
= (uval
& FUTEX_OWNER_DIED
) | newtid
;
1697 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
))
1705 * We fixed up user space. Now we need to fix the pi_state
1708 if (pi_state
->owner
!= NULL
) {
1709 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
1710 WARN_ON(list_empty(&pi_state
->list
));
1711 list_del_init(&pi_state
->list
);
1712 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
1715 pi_state
->owner
= newowner
;
1717 raw_spin_lock_irq(&newowner
->pi_lock
);
1718 WARN_ON(!list_empty(&pi_state
->list
));
1719 list_add(&pi_state
->list
, &newowner
->pi_state_list
);
1720 raw_spin_unlock_irq(&newowner
->pi_lock
);
1724 * To handle the page fault we need to drop the hash bucket
1725 * lock here. That gives the other task (either the highest priority
1726 * waiter itself or the task which stole the rtmutex) the
1727 * chance to try the fixup of the pi_state. So once we are
1728 * back from handling the fault we need to check the pi_state
1729 * after reacquiring the hash bucket lock and before trying to
1730 * do another fixup. When the fixup has been done already we
1734 spin_unlock(q
->lock_ptr
);
1736 ret
= fault_in_user_writeable(uaddr
);
1738 spin_lock(q
->lock_ptr
);
1741 * Check if someone else fixed it for us:
1743 if (pi_state
->owner
!= oldowner
)
1752 static long futex_wait_restart(struct restart_block
*restart
);
1755 * fixup_owner() - Post lock pi_state and corner case management
1756 * @uaddr: user address of the futex
1757 * @q: futex_q (contains pi_state and access to the rt_mutex)
1758 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1760 * After attempting to lock an rt_mutex, this function is called to cleanup
1761 * the pi_state owner as well as handle race conditions that may allow us to
1762 * acquire the lock. Must be called with the hb lock held.
1765 * 1 - success, lock taken;
1766 * 0 - success, lock not taken;
1767 * <0 - on error (-EFAULT)
1769 static int fixup_owner(u32 __user
*uaddr
, struct futex_q
*q
, int locked
)
1771 struct task_struct
*owner
;
1776 * Got the lock. We might not be the anticipated owner if we
1777 * did a lock-steal - fix up the PI-state in that case:
1779 if (q
->pi_state
->owner
!= current
)
1780 ret
= fixup_pi_state_owner(uaddr
, q
, current
);
1785 * Catch the rare case, where the lock was released when we were on the
1786 * way back before we locked the hash bucket.
1788 if (q
->pi_state
->owner
== current
) {
1790 * Try to get the rt_mutex now. This might fail as some other
1791 * task acquired the rt_mutex after we removed ourself from the
1792 * rt_mutex waiters list.
1794 if (rt_mutex_trylock(&q
->pi_state
->pi_mutex
)) {
1800 * pi_state is incorrect, some other task did a lock steal and
1801 * we returned due to timeout or signal without taking the
1802 * rt_mutex. Too late.
1804 raw_spin_lock(&q
->pi_state
->pi_mutex
.wait_lock
);
1805 owner
= rt_mutex_owner(&q
->pi_state
->pi_mutex
);
1807 owner
= rt_mutex_next_owner(&q
->pi_state
->pi_mutex
);
1808 raw_spin_unlock(&q
->pi_state
->pi_mutex
.wait_lock
);
1809 ret
= fixup_pi_state_owner(uaddr
, q
, owner
);
1814 * Paranoia check. If we did not take the lock, then we should not be
1815 * the owner of the rt_mutex.
1817 if (rt_mutex_owner(&q
->pi_state
->pi_mutex
) == current
)
1818 printk(KERN_ERR
"fixup_owner: ret = %d pi-mutex: %p "
1819 "pi-state %p\n", ret
,
1820 q
->pi_state
->pi_mutex
.owner
,
1821 q
->pi_state
->owner
);
1824 return ret
? ret
: locked
;
1828 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1829 * @hb: the futex hash bucket, must be locked by the caller
1830 * @q: the futex_q to queue up on
1831 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1833 static void futex_wait_queue_me(struct futex_hash_bucket
*hb
, struct futex_q
*q
,
1834 struct hrtimer_sleeper
*timeout
)
1837 * The task state is guaranteed to be set before another task can
1838 * wake it. set_current_state() is implemented using set_mb() and
1839 * queue_me() calls spin_unlock() upon completion, both serializing
1840 * access to the hash list and forcing another memory barrier.
1842 set_current_state(TASK_INTERRUPTIBLE
);
1847 hrtimer_start_expires(&timeout
->timer
, HRTIMER_MODE_ABS
);
1848 if (!hrtimer_active(&timeout
->timer
))
1849 timeout
->task
= NULL
;
1853 * If we have been removed from the hash list, then another task
1854 * has tried to wake us, and we can skip the call to schedule().
1856 if (likely(!plist_node_empty(&q
->list
))) {
1858 * If the timer has already expired, current will already be
1859 * flagged for rescheduling. Only call schedule if there
1860 * is no timeout, or if it has yet to expire.
1862 if (!timeout
|| timeout
->task
)
1865 __set_current_state(TASK_RUNNING
);
1869 * futex_wait_setup() - Prepare to wait on a futex
1870 * @uaddr: the futex userspace address
1871 * @val: the expected value
1872 * @flags: futex flags (FLAGS_SHARED, etc.)
1873 * @q: the associated futex_q
1874 * @hb: storage for hash_bucket pointer to be returned to caller
1876 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1877 * compare it with the expected value. Handle atomic faults internally.
1878 * Return with the hb lock held and a q.key reference on success, and unlocked
1879 * with no q.key reference on failure.
1882 * 0 - uaddr contains val and hb has been locked;
1883 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
1885 static int futex_wait_setup(u32 __user
*uaddr
, u32 val
, unsigned int flags
,
1886 struct futex_q
*q
, struct futex_hash_bucket
**hb
)
1892 * Access the page AFTER the hash-bucket is locked.
1893 * Order is important:
1895 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1896 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1898 * The basic logical guarantee of a futex is that it blocks ONLY
1899 * if cond(var) is known to be true at the time of blocking, for
1900 * any cond. If we locked the hash-bucket after testing *uaddr, that
1901 * would open a race condition where we could block indefinitely with
1902 * cond(var) false, which would violate the guarantee.
1904 * On the other hand, we insert q and release the hash-bucket only
1905 * after testing *uaddr. This guarantees that futex_wait() will NOT
1906 * absorb a wakeup if *uaddr does not match the desired values
1907 * while the syscall executes.
1910 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
->key
, VERIFY_READ
);
1911 if (unlikely(ret
!= 0))
1915 *hb
= queue_lock(q
);
1917 ret
= get_futex_value_locked(&uval
, uaddr
);
1920 queue_unlock(q
, *hb
);
1922 ret
= get_user(uval
, uaddr
);
1926 if (!(flags
& FLAGS_SHARED
))
1929 put_futex_key(&q
->key
);
1934 queue_unlock(q
, *hb
);
1940 put_futex_key(&q
->key
);
1944 static int futex_wait(u32 __user
*uaddr
, unsigned int flags
, u32 val
,
1945 ktime_t
*abs_time
, u32 bitset
)
1947 struct hrtimer_sleeper timeout
, *to
= NULL
;
1948 struct restart_block
*restart
;
1949 struct futex_hash_bucket
*hb
;
1950 struct futex_q q
= futex_q_init
;
1960 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
1961 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
1963 hrtimer_init_sleeper(to
, current
);
1964 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
1965 current
->timer_slack_ns
);
1970 * Prepare to wait on uaddr. On success, holds hb lock and increments
1973 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
1977 /* queue_me and wait for wakeup, timeout, or a signal. */
1978 futex_wait_queue_me(hb
, &q
, to
);
1980 /* If we were woken (and unqueued), we succeeded, whatever. */
1982 /* unqueue_me() drops q.key ref */
1983 if (!unqueue_me(&q
))
1986 if (to
&& !to
->task
)
1990 * We expect signal_pending(current), but we might be the
1991 * victim of a spurious wakeup as well.
1993 if (!signal_pending(current
))
2000 restart
= ¤t_thread_info()->restart_block
;
2001 restart
->fn
= futex_wait_restart
;
2002 restart
->futex
.uaddr
= uaddr
;
2003 restart
->futex
.val
= val
;
2004 restart
->futex
.time
= abs_time
->tv64
;
2005 restart
->futex
.bitset
= bitset
;
2006 restart
->futex
.flags
= flags
| FLAGS_HAS_TIMEOUT
;
2008 ret
= -ERESTART_RESTARTBLOCK
;
2012 hrtimer_cancel(&to
->timer
);
2013 destroy_hrtimer_on_stack(&to
->timer
);
2019 static long futex_wait_restart(struct restart_block
*restart
)
2021 u32 __user
*uaddr
= restart
->futex
.uaddr
;
2022 ktime_t t
, *tp
= NULL
;
2024 if (restart
->futex
.flags
& FLAGS_HAS_TIMEOUT
) {
2025 t
.tv64
= restart
->futex
.time
;
2028 restart
->fn
= do_no_restart_syscall
;
2030 return (long)futex_wait(uaddr
, restart
->futex
.flags
,
2031 restart
->futex
.val
, tp
, restart
->futex
.bitset
);
2036 * Userspace tried a 0 -> TID atomic transition of the futex value
2037 * and failed. The kernel side here does the whole locking operation:
2038 * if there are waiters then it will block, it does PI, etc. (Due to
2039 * races the kernel might see a 0 value of the futex too.)
2041 static int futex_lock_pi(u32 __user
*uaddr
, unsigned int flags
, int detect
,
2042 ktime_t
*time
, int trylock
)
2044 struct hrtimer_sleeper timeout
, *to
= NULL
;
2045 struct futex_hash_bucket
*hb
;
2046 struct futex_q q
= futex_q_init
;
2049 if (refill_pi_state_cache())
2054 hrtimer_init_on_stack(&to
->timer
, CLOCK_REALTIME
,
2056 hrtimer_init_sleeper(to
, current
);
2057 hrtimer_set_expires(&to
->timer
, *time
);
2061 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
.key
, VERIFY_WRITE
);
2062 if (unlikely(ret
!= 0))
2066 hb
= queue_lock(&q
);
2068 ret
= futex_lock_pi_atomic(uaddr
, hb
, &q
.key
, &q
.pi_state
, current
, 0);
2069 if (unlikely(ret
)) {
2072 /* We got the lock. */
2074 goto out_unlock_put_key
;
2079 * Task is exiting and we just wait for the
2082 queue_unlock(&q
, hb
);
2083 put_futex_key(&q
.key
);
2087 goto out_unlock_put_key
;
2092 * Only actually queue now that the atomic ops are done:
2096 WARN_ON(!q
.pi_state
);
2098 * Block on the PI mutex:
2101 ret
= rt_mutex_timed_lock(&q
.pi_state
->pi_mutex
, to
, 1);
2103 ret
= rt_mutex_trylock(&q
.pi_state
->pi_mutex
);
2104 /* Fixup the trylock return value: */
2105 ret
= ret
? 0 : -EWOULDBLOCK
;
2108 spin_lock(q
.lock_ptr
);
2110 * Fixup the pi_state owner and possibly acquire the lock if we
2113 res
= fixup_owner(uaddr
, &q
, !ret
);
2115 * If fixup_owner() returned an error, proprogate that. If it acquired
2116 * the lock, clear our -ETIMEDOUT or -EINTR.
2119 ret
= (res
< 0) ? res
: 0;
2122 * If fixup_owner() faulted and was unable to handle the fault, unlock
2123 * it and return the fault to userspace.
2125 if (ret
&& (rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
))
2126 rt_mutex_unlock(&q
.pi_state
->pi_mutex
);
2128 /* Unqueue and drop the lock */
2134 queue_unlock(&q
, hb
);
2137 put_futex_key(&q
.key
);
2140 destroy_hrtimer_on_stack(&to
->timer
);
2141 return ret
!= -EINTR
? ret
: -ERESTARTNOINTR
;
2144 queue_unlock(&q
, hb
);
2146 ret
= fault_in_user_writeable(uaddr
);
2150 if (!(flags
& FLAGS_SHARED
))
2153 put_futex_key(&q
.key
);
2158 * Userspace attempted a TID -> 0 atomic transition, and failed.
2159 * This is the in-kernel slowpath: we look up the PI state (if any),
2160 * and do the rt-mutex unlock.
2162 static int futex_unlock_pi(u32 __user
*uaddr
, unsigned int flags
)
2164 struct futex_hash_bucket
*hb
;
2165 struct futex_q
*this, *next
;
2166 struct plist_head
*head
;
2167 union futex_key key
= FUTEX_KEY_INIT
;
2168 u32 uval
, vpid
= task_pid_vnr(current
);
2172 if (get_user(uval
, uaddr
))
2175 * We release only a lock we actually own:
2177 if ((uval
& FUTEX_TID_MASK
) != vpid
)
2180 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_WRITE
);
2181 if (unlikely(ret
!= 0))
2184 hb
= hash_futex(&key
);
2185 spin_lock(&hb
->lock
);
2188 * To avoid races, try to do the TID -> 0 atomic transition
2189 * again. If it succeeds then we can return without waking
2192 if (!(uval
& FUTEX_OWNER_DIED
) &&
2193 cmpxchg_futex_value_locked(&uval
, uaddr
, vpid
, 0))
2196 * Rare case: we managed to release the lock atomically,
2197 * no need to wake anyone else up:
2199 if (unlikely(uval
== vpid
))
2203 * Ok, other tasks may need to be woken up - check waiters
2204 * and do the wakeup if necessary:
2208 plist_for_each_entry_safe(this, next
, head
, list
) {
2209 if (!match_futex (&this->key
, &key
))
2211 ret
= wake_futex_pi(uaddr
, uval
, this);
2213 * The atomic access to the futex value
2214 * generated a pagefault, so retry the
2215 * user-access and the wakeup:
2222 * No waiters - kernel unlocks the futex:
2224 if (!(uval
& FUTEX_OWNER_DIED
)) {
2225 ret
= unlock_futex_pi(uaddr
, uval
);
2231 spin_unlock(&hb
->lock
);
2232 put_futex_key(&key
);
2238 spin_unlock(&hb
->lock
);
2239 put_futex_key(&key
);
2241 ret
= fault_in_user_writeable(uaddr
);
2249 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2250 * @hb: the hash_bucket futex_q was original enqueued on
2251 * @q: the futex_q woken while waiting to be requeued
2252 * @key2: the futex_key of the requeue target futex
2253 * @timeout: the timeout associated with the wait (NULL if none)
2255 * Detect if the task was woken on the initial futex as opposed to the requeue
2256 * target futex. If so, determine if it was a timeout or a signal that caused
2257 * the wakeup and return the appropriate error code to the caller. Must be
2258 * called with the hb lock held.
2261 * 0 = no early wakeup detected;
2262 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2265 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket
*hb
,
2266 struct futex_q
*q
, union futex_key
*key2
,
2267 struct hrtimer_sleeper
*timeout
)
2272 * With the hb lock held, we avoid races while we process the wakeup.
2273 * We only need to hold hb (and not hb2) to ensure atomicity as the
2274 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2275 * It can't be requeued from uaddr2 to something else since we don't
2276 * support a PI aware source futex for requeue.
2278 if (!match_futex(&q
->key
, key2
)) {
2279 WARN_ON(q
->lock_ptr
&& (&hb
->lock
!= q
->lock_ptr
));
2281 * We were woken prior to requeue by a timeout or a signal.
2282 * Unqueue the futex_q and determine which it was.
2284 plist_del(&q
->list
, &hb
->chain
);
2286 /* Handle spurious wakeups gracefully */
2288 if (timeout
&& !timeout
->task
)
2290 else if (signal_pending(current
))
2291 ret
= -ERESTARTNOINTR
;
2297 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2298 * @uaddr: the futex we initially wait on (non-pi)
2299 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2300 * the same type, no requeueing from private to shared, etc.
2301 * @val: the expected value of uaddr
2302 * @abs_time: absolute timeout
2303 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2304 * @uaddr2: the pi futex we will take prior to returning to user-space
2306 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2307 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2308 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2309 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2310 * without one, the pi logic would not know which task to boost/deboost, if
2311 * there was a need to.
2313 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2314 * via the following--
2315 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2316 * 2) wakeup on uaddr2 after a requeue
2320 * If 3, cleanup and return -ERESTARTNOINTR.
2322 * If 2, we may then block on trying to take the rt_mutex and return via:
2323 * 5) successful lock
2326 * 8) other lock acquisition failure
2328 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2330 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2336 static int futex_wait_requeue_pi(u32 __user
*uaddr
, unsigned int flags
,
2337 u32 val
, ktime_t
*abs_time
, u32 bitset
,
2340 struct hrtimer_sleeper timeout
, *to
= NULL
;
2341 struct rt_mutex_waiter rt_waiter
;
2342 struct rt_mutex
*pi_mutex
= NULL
;
2343 struct futex_hash_bucket
*hb
;
2344 union futex_key key2
= FUTEX_KEY_INIT
;
2345 struct futex_q q
= futex_q_init
;
2348 if (uaddr
== uaddr2
)
2356 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2357 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2359 hrtimer_init_sleeper(to
, current
);
2360 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2361 current
->timer_slack_ns
);
2365 * The waiter is allocated on our stack, manipulated by the requeue
2366 * code while we sleep on uaddr.
2368 debug_rt_mutex_init_waiter(&rt_waiter
);
2369 rt_waiter
.task
= NULL
;
2371 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
2372 if (unlikely(ret
!= 0))
2376 q
.rt_waiter
= &rt_waiter
;
2377 q
.requeue_pi_key
= &key2
;
2380 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2383 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2388 * The check above which compares uaddrs is not sufficient for
2389 * shared futexes. We need to compare the keys:
2391 if (match_futex(&q
.key
, &key2
)) {
2396 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2397 futex_wait_queue_me(hb
, &q
, to
);
2399 spin_lock(&hb
->lock
);
2400 ret
= handle_early_requeue_pi_wakeup(hb
, &q
, &key2
, to
);
2401 spin_unlock(&hb
->lock
);
2406 * In order for us to be here, we know our q.key == key2, and since
2407 * we took the hb->lock above, we also know that futex_requeue() has
2408 * completed and we no longer have to concern ourselves with a wakeup
2409 * race with the atomic proxy lock acquisition by the requeue code. The
2410 * futex_requeue dropped our key1 reference and incremented our key2
2414 /* Check if the requeue code acquired the second futex for us. */
2417 * Got the lock. We might not be the anticipated owner if we
2418 * did a lock-steal - fix up the PI-state in that case.
2420 if (q
.pi_state
&& (q
.pi_state
->owner
!= current
)) {
2421 spin_lock(q
.lock_ptr
);
2422 ret
= fixup_pi_state_owner(uaddr2
, &q
, current
);
2423 spin_unlock(q
.lock_ptr
);
2427 * We have been woken up by futex_unlock_pi(), a timeout, or a
2428 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2431 WARN_ON(!q
.pi_state
);
2432 pi_mutex
= &q
.pi_state
->pi_mutex
;
2433 ret
= rt_mutex_finish_proxy_lock(pi_mutex
, to
, &rt_waiter
, 1);
2434 debug_rt_mutex_free_waiter(&rt_waiter
);
2436 spin_lock(q
.lock_ptr
);
2438 * Fixup the pi_state owner and possibly acquire the lock if we
2441 res
= fixup_owner(uaddr2
, &q
, !ret
);
2443 * If fixup_owner() returned an error, proprogate that. If it
2444 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2447 ret
= (res
< 0) ? res
: 0;
2449 /* Unqueue and drop the lock. */
2454 * If fixup_pi_state_owner() faulted and was unable to handle the
2455 * fault, unlock the rt_mutex and return the fault to userspace.
2457 if (ret
== -EFAULT
) {
2458 if (pi_mutex
&& rt_mutex_owner(pi_mutex
) == current
)
2459 rt_mutex_unlock(pi_mutex
);
2460 } else if (ret
== -EINTR
) {
2462 * We've already been requeued, but cannot restart by calling
2463 * futex_lock_pi() directly. We could restart this syscall, but
2464 * it would detect that the user space "val" changed and return
2465 * -EWOULDBLOCK. Save the overhead of the restart and return
2466 * -EWOULDBLOCK directly.
2472 put_futex_key(&q
.key
);
2474 put_futex_key(&key2
);
2478 hrtimer_cancel(&to
->timer
);
2479 destroy_hrtimer_on_stack(&to
->timer
);
2485 * Support for robust futexes: the kernel cleans up held futexes at
2488 * Implementation: user-space maintains a per-thread list of locks it
2489 * is holding. Upon do_exit(), the kernel carefully walks this list,
2490 * and marks all locks that are owned by this thread with the
2491 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2492 * always manipulated with the lock held, so the list is private and
2493 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2494 * field, to allow the kernel to clean up if the thread dies after
2495 * acquiring the lock, but just before it could have added itself to
2496 * the list. There can only be one such pending lock.
2500 * sys_set_robust_list() - Set the robust-futex list head of a task
2501 * @head: pointer to the list-head
2502 * @len: length of the list-head, as userspace expects
2504 SYSCALL_DEFINE2(set_robust_list
, struct robust_list_head __user
*, head
,
2507 if (!futex_cmpxchg_enabled
)
2510 * The kernel knows only one size for now:
2512 if (unlikely(len
!= sizeof(*head
)))
2515 current
->robust_list
= head
;
2521 * sys_get_robust_list() - Get the robust-futex list head of a task
2522 * @pid: pid of the process [zero for current task]
2523 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2524 * @len_ptr: pointer to a length field, the kernel fills in the header size
2526 SYSCALL_DEFINE3(get_robust_list
, int, pid
,
2527 struct robust_list_head __user
* __user
*, head_ptr
,
2528 size_t __user
*, len_ptr
)
2530 struct robust_list_head __user
*head
;
2532 struct task_struct
*p
;
2534 if (!futex_cmpxchg_enabled
)
2543 p
= find_task_by_vpid(pid
);
2549 if (!ptrace_may_access(p
, PTRACE_MODE_READ
))
2552 head
= p
->robust_list
;
2555 if (put_user(sizeof(*head
), len_ptr
))
2557 return put_user(head
, head_ptr
);
2566 * Process a futex-list entry, check whether it's owned by the
2567 * dying task, and do notification if so:
2569 int handle_futex_death(u32 __user
*uaddr
, struct task_struct
*curr
, int pi
)
2571 u32 uval
, uninitialized_var(nval
), mval
;
2574 if (get_user(uval
, uaddr
))
2577 if ((uval
& FUTEX_TID_MASK
) == task_pid_vnr(curr
)) {
2579 * Ok, this dying thread is truly holding a futex
2580 * of interest. Set the OWNER_DIED bit atomically
2581 * via cmpxchg, and if the value had FUTEX_WAITERS
2582 * set, wake up a waiter (if any). (We have to do a
2583 * futex_wake() even if OWNER_DIED is already set -
2584 * to handle the rare but possible case of recursive
2585 * thread-death.) The rest of the cleanup is done in
2588 mval
= (uval
& FUTEX_WAITERS
) | FUTEX_OWNER_DIED
;
2590 * We are not holding a lock here, but we want to have
2591 * the pagefault_disable/enable() protection because
2592 * we want to handle the fault gracefully. If the
2593 * access fails we try to fault in the futex with R/W
2594 * verification via get_user_pages. get_user() above
2595 * does not guarantee R/W access. If that fails we
2596 * give up and leave the futex locked.
2598 if (cmpxchg_futex_value_locked(&nval
, uaddr
, uval
, mval
)) {
2599 if (fault_in_user_writeable(uaddr
))
2607 * Wake robust non-PI futexes here. The wakeup of
2608 * PI futexes happens in exit_pi_state():
2610 if (!pi
&& (uval
& FUTEX_WAITERS
))
2611 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
2617 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2619 static inline int fetch_robust_entry(struct robust_list __user
**entry
,
2620 struct robust_list __user
* __user
*head
,
2623 unsigned long uentry
;
2625 if (get_user(uentry
, (unsigned long __user
*)head
))
2628 *entry
= (void __user
*)(uentry
& ~1UL);
2635 * Walk curr->robust_list (very carefully, it's a userspace list!)
2636 * and mark any locks found there dead, and notify any waiters.
2638 * We silently return on any sign of list-walking problem.
2640 void exit_robust_list(struct task_struct
*curr
)
2642 struct robust_list_head __user
*head
= curr
->robust_list
;
2643 struct robust_list __user
*entry
, *next_entry
, *pending
;
2644 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
2645 unsigned int uninitialized_var(next_pi
);
2646 unsigned long futex_offset
;
2649 if (!futex_cmpxchg_enabled
)
2653 * Fetch the list head (which was registered earlier, via
2654 * sys_set_robust_list()):
2656 if (fetch_robust_entry(&entry
, &head
->list
.next
, &pi
))
2659 * Fetch the relative futex offset:
2661 if (get_user(futex_offset
, &head
->futex_offset
))
2664 * Fetch any possibly pending lock-add first, and handle it
2667 if (fetch_robust_entry(&pending
, &head
->list_op_pending
, &pip
))
2670 next_entry
= NULL
; /* avoid warning with gcc */
2671 while (entry
!= &head
->list
) {
2673 * Fetch the next entry in the list before calling
2674 * handle_futex_death:
2676 rc
= fetch_robust_entry(&next_entry
, &entry
->next
, &next_pi
);
2678 * A pending lock might already be on the list, so
2679 * don't process it twice:
2681 if (entry
!= pending
)
2682 if (handle_futex_death((void __user
*)entry
+ futex_offset
,
2690 * Avoid excessively long or circular lists:
2699 handle_futex_death((void __user
*)pending
+ futex_offset
,
2703 long do_futex(u32 __user
*uaddr
, int op
, u32 val
, ktime_t
*timeout
,
2704 u32 __user
*uaddr2
, u32 val2
, u32 val3
)
2706 int cmd
= op
& FUTEX_CMD_MASK
;
2707 unsigned int flags
= 0;
2709 if (!(op
& FUTEX_PRIVATE_FLAG
))
2710 flags
|= FLAGS_SHARED
;
2712 if (op
& FUTEX_CLOCK_REALTIME
) {
2713 flags
|= FLAGS_CLOCKRT
;
2714 if (cmd
!= FUTEX_WAIT_BITSET
&& cmd
!= FUTEX_WAIT_REQUEUE_PI
)
2720 case FUTEX_UNLOCK_PI
:
2721 case FUTEX_TRYLOCK_PI
:
2722 case FUTEX_WAIT_REQUEUE_PI
:
2723 case FUTEX_CMP_REQUEUE_PI
:
2724 if (!futex_cmpxchg_enabled
)
2730 val3
= FUTEX_BITSET_MATCH_ANY
;
2731 case FUTEX_WAIT_BITSET
:
2732 return futex_wait(uaddr
, flags
, val
, timeout
, val3
);
2734 val3
= FUTEX_BITSET_MATCH_ANY
;
2735 case FUTEX_WAKE_BITSET
:
2736 return futex_wake(uaddr
, flags
, val
, val3
);
2738 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, NULL
, 0);
2739 case FUTEX_CMP_REQUEUE
:
2740 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 0);
2742 return futex_wake_op(uaddr
, flags
, uaddr2
, val
, val2
, val3
);
2744 return futex_lock_pi(uaddr
, flags
, val
, timeout
, 0);
2745 case FUTEX_UNLOCK_PI
:
2746 return futex_unlock_pi(uaddr
, flags
);
2747 case FUTEX_TRYLOCK_PI
:
2748 return futex_lock_pi(uaddr
, flags
, 0, timeout
, 1);
2749 case FUTEX_WAIT_REQUEUE_PI
:
2750 val3
= FUTEX_BITSET_MATCH_ANY
;
2751 return futex_wait_requeue_pi(uaddr
, flags
, val
, timeout
, val3
,
2753 case FUTEX_CMP_REQUEUE_PI
:
2754 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 1);
2760 SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
2761 struct timespec __user
*, utime
, u32 __user
*, uaddr2
,
2765 ktime_t t
, *tp
= NULL
;
2767 int cmd
= op
& FUTEX_CMD_MASK
;
2769 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
2770 cmd
== FUTEX_WAIT_BITSET
||
2771 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
2772 if (copy_from_user(&ts
, utime
, sizeof(ts
)) != 0)
2774 if (!timespec_valid(&ts
))
2777 t
= timespec_to_ktime(ts
);
2778 if (cmd
== FUTEX_WAIT
)
2779 t
= ktime_add_safe(ktime_get(), t
);
2783 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2784 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2786 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
2787 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
2788 val2
= (u32
) (unsigned long) utime
;
2790 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
2793 static void __init
futex_detect_cmpxchg(void)
2795 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
2799 * This will fail and we want it. Some arch implementations do
2800 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2801 * functionality. We want to know that before we call in any
2802 * of the complex code paths. Also we want to prevent
2803 * registration of robust lists in that case. NULL is
2804 * guaranteed to fault and we get -EFAULT on functional
2805 * implementation, the non-functional ones will return
2808 if (cmpxchg_futex_value_locked(&curval
, NULL
, 0, 0) == -EFAULT
)
2809 futex_cmpxchg_enabled
= 1;
2813 static int __init
futex_init(void)
2817 futex_detect_cmpxchg();
2819 for (i
= 0; i
< ARRAY_SIZE(futex_queues
); i
++) {
2820 plist_head_init(&futex_queues
[i
].chain
);
2821 spin_lock_init(&futex_queues
[i
].lock
);
2826 __initcall(futex_init
);