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 * READ this before attempting to hack on futexes!
75 * Basic futex operation and ordering guarantees
76 * =============================================
78 * The waiter reads the futex value in user space and calls
79 * futex_wait(). This function computes the hash bucket and acquires
80 * the hash bucket lock. After that it reads the futex user space value
81 * again and verifies that the data has not changed. If it has not changed
82 * it enqueues itself into the hash bucket, releases the hash bucket lock
85 * The waker side modifies the user space value of the futex and calls
86 * futex_wake(). This function computes the hash bucket and acquires the
87 * hash bucket lock. Then it looks for waiters on that futex in the hash
88 * bucket and wakes them.
90 * In futex wake up scenarios where no tasks are blocked on a futex, taking
91 * the hb spinlock can be avoided and simply return. In order for this
92 * optimization to work, ordering guarantees must exist so that the waiter
93 * being added to the list is acknowledged when the list is concurrently being
94 * checked by the waker, avoiding scenarios like the following:
98 * sys_futex(WAIT, futex, val);
99 * futex_wait(futex, val);
102 * sys_futex(WAKE, futex);
107 * lock(hash_bucket(futex));
109 * unlock(hash_bucket(futex));
112 * This would cause the waiter on CPU 0 to wait forever because it
113 * missed the transition of the user space value from val to newval
114 * and the waker did not find the waiter in the hash bucket queue.
116 * The correct serialization ensures that a waiter either observes
117 * the changed user space value before blocking or is woken by a
122 * sys_futex(WAIT, futex, val);
123 * futex_wait(futex, val);
126 * mb(); (A) <-- paired with -.
128 * lock(hash_bucket(futex)); |
132 * | sys_futex(WAKE, futex);
133 * | futex_wake(futex);
135 * `-------> mb(); (B)
138 * unlock(hash_bucket(futex));
139 * schedule(); if (waiters)
140 * lock(hash_bucket(futex));
141 * else wake_waiters(futex);
142 * waiters--; (b) unlock(hash_bucket(futex));
144 * Where (A) orders the waiters increment and the futex value read through
145 * atomic operations (see hb_waiters_inc) and where (B) orders the write
146 * to futex and the waiters read -- this is done by the barriers in
147 * get_futex_key_refs(), through either ihold or atomic_inc, depending on the
150 * This yields the following case (where X:=waiters, Y:=futex):
158 * Which guarantees that x==0 && y==0 is impossible; which translates back into
159 * the guarantee that we cannot both miss the futex variable change and the
162 * Note that a new waiter is accounted for in (a) even when it is possible that
163 * the wait call can return error, in which case we backtrack from it in (b).
164 * Refer to the comment in queue_lock().
166 * Similarly, in order to account for waiters being requeued on another
167 * address we always increment the waiters for the destination bucket before
168 * acquiring the lock. It then decrements them again after releasing it -
169 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
170 * will do the additional required waiter count housekeeping. This is done for
171 * double_lock_hb() and double_unlock_hb(), respectively.
174 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
175 int __read_mostly futex_cmpxchg_enabled
;
179 * Futex flags used to encode options to functions and preserve them across
182 #define FLAGS_SHARED 0x01
183 #define FLAGS_CLOCKRT 0x02
184 #define FLAGS_HAS_TIMEOUT 0x04
187 * Priority Inheritance state:
189 struct futex_pi_state
{
191 * list of 'owned' pi_state instances - these have to be
192 * cleaned up in do_exit() if the task exits prematurely:
194 struct list_head list
;
199 struct rt_mutex pi_mutex
;
201 struct task_struct
*owner
;
208 * struct futex_q - The hashed futex queue entry, one per waiting task
209 * @list: priority-sorted list of tasks waiting on this futex
210 * @task: the task waiting on the futex
211 * @lock_ptr: the hash bucket lock
212 * @key: the key the futex is hashed on
213 * @pi_state: optional priority inheritance state
214 * @rt_waiter: rt_waiter storage for use with requeue_pi
215 * @requeue_pi_key: the requeue_pi target futex key
216 * @bitset: bitset for the optional bitmasked wakeup
218 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
219 * we can wake only the relevant ones (hashed queues may be shared).
221 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
222 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
223 * The order of wakeup is always to make the first condition true, then
226 * PI futexes are typically woken before they are removed from the hash list via
227 * the rt_mutex code. See unqueue_me_pi().
230 struct plist_node list
;
232 struct task_struct
*task
;
233 spinlock_t
*lock_ptr
;
235 struct futex_pi_state
*pi_state
;
236 struct rt_mutex_waiter
*rt_waiter
;
237 union futex_key
*requeue_pi_key
;
241 static const struct futex_q futex_q_init
= {
242 /* list gets initialized in queue_me()*/
243 .key
= FUTEX_KEY_INIT
,
244 .bitset
= FUTEX_BITSET_MATCH_ANY
248 * Hash buckets are shared by all the futex_keys that hash to the same
249 * location. Each key may have multiple futex_q structures, one for each task
250 * waiting on a futex.
252 struct futex_hash_bucket
{
255 struct plist_head chain
;
256 } ____cacheline_aligned_in_smp
;
258 static unsigned long __read_mostly futex_hashsize
;
260 static struct futex_hash_bucket
*futex_queues
;
262 static inline void futex_get_mm(union futex_key
*key
)
264 atomic_inc(&key
->private.mm
->mm_count
);
266 * Ensure futex_get_mm() implies a full barrier such that
267 * get_futex_key() implies a full barrier. This is relied upon
268 * as full barrier (B), see the ordering comment above.
270 smp_mb__after_atomic_inc();
274 * Reflects a new waiter being added to the waitqueue.
276 static inline void hb_waiters_inc(struct futex_hash_bucket
*hb
)
279 atomic_inc(&hb
->waiters
);
281 * Full barrier (A), see the ordering comment above.
283 smp_mb__after_atomic_inc();
288 * Reflects a waiter being removed from the waitqueue by wakeup
291 static inline void hb_waiters_dec(struct futex_hash_bucket
*hb
)
294 atomic_dec(&hb
->waiters
);
298 static inline int hb_waiters_pending(struct futex_hash_bucket
*hb
)
301 return atomic_read(&hb
->waiters
);
308 * We hash on the keys returned from get_futex_key (see below).
310 static struct futex_hash_bucket
*hash_futex(union futex_key
*key
)
312 u32 hash
= jhash2((u32
*)&key
->both
.word
,
313 (sizeof(key
->both
.word
)+sizeof(key
->both
.ptr
))/4,
315 return &futex_queues
[hash
& (futex_hashsize
- 1)];
319 * Return 1 if two futex_keys are equal, 0 otherwise.
321 static inline int match_futex(union futex_key
*key1
, union futex_key
*key2
)
324 && key1
->both
.word
== key2
->both
.word
325 && key1
->both
.ptr
== key2
->both
.ptr
326 && key1
->both
.offset
== key2
->both
.offset
);
330 * Take a reference to the resource addressed by a key.
331 * Can be called while holding spinlocks.
334 static void get_futex_key_refs(union futex_key
*key
)
339 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
341 ihold(key
->shared
.inode
); /* implies MB (B) */
343 case FUT_OFF_MMSHARED
:
344 futex_get_mm(key
); /* implies MB (B) */
350 * Drop a reference to the resource addressed by a key.
351 * The hash bucket spinlock must not be held.
353 static void drop_futex_key_refs(union futex_key
*key
)
355 if (!key
->both
.ptr
) {
356 /* If we're here then we tried to put a key we failed to get */
361 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
363 iput(key
->shared
.inode
);
365 case FUT_OFF_MMSHARED
:
366 mmdrop(key
->private.mm
);
372 * get_futex_key() - Get parameters which are the keys for a futex
373 * @uaddr: virtual address of the futex
374 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
375 * @key: address where result is stored.
376 * @rw: mapping needs to be read/write (values: VERIFY_READ,
379 * Return: a negative error code or 0
381 * The key words are stored in *key on success.
383 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
384 * offset_within_page). For private mappings, it's (uaddr, current->mm).
385 * We can usually work out the index without swapping in the page.
387 * lock_page() might sleep, the caller should not hold a spinlock.
390 get_futex_key(u32 __user
*uaddr
, int fshared
, union futex_key
*key
, int rw
)
392 unsigned long address
= (unsigned long)uaddr
;
393 struct mm_struct
*mm
= current
->mm
;
394 struct page
*page
, *page_head
;
398 * The futex address must be "naturally" aligned.
400 key
->both
.offset
= address
% PAGE_SIZE
;
401 if (unlikely((address
% sizeof(u32
)) != 0))
403 address
-= key
->both
.offset
;
405 if (unlikely(!access_ok(rw
, uaddr
, sizeof(u32
))))
409 * PROCESS_PRIVATE futexes are fast.
410 * As the mm cannot disappear under us and the 'key' only needs
411 * virtual address, we dont even have to find the underlying vma.
412 * Note : We do have to check 'uaddr' is a valid user address,
413 * but access_ok() should be faster than find_vma()
416 key
->private.mm
= mm
;
417 key
->private.address
= address
;
418 get_futex_key_refs(key
); /* implies MB (B) */
423 err
= get_user_pages_fast(address
, 1, 1, &page
);
425 * If write access is not required (eg. FUTEX_WAIT), try
426 * and get read-only access.
428 if (err
== -EFAULT
&& rw
== VERIFY_READ
) {
429 err
= get_user_pages_fast(address
, 1, 0, &page
);
437 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
439 if (unlikely(PageTail(page
))) {
441 /* serialize against __split_huge_page_splitting() */
443 if (likely(__get_user_pages_fast(address
, 1, !ro
, &page
) == 1)) {
444 page_head
= compound_head(page
);
446 * page_head is valid pointer but we must pin
447 * it before taking the PG_lock and/or
448 * PG_compound_lock. The moment we re-enable
449 * irqs __split_huge_page_splitting() can
450 * return and the head page can be freed from
451 * under us. We can't take the PG_lock and/or
452 * PG_compound_lock on a page that could be
453 * freed from under us.
455 if (page
!= page_head
) {
466 page_head
= compound_head(page
);
467 if (page
!= page_head
) {
473 lock_page(page_head
);
476 * If page_head->mapping is NULL, then it cannot be a PageAnon
477 * page; but it might be the ZERO_PAGE or in the gate area or
478 * in a special mapping (all cases which we are happy to fail);
479 * or it may have been a good file page when get_user_pages_fast
480 * found it, but truncated or holepunched or subjected to
481 * invalidate_complete_page2 before we got the page lock (also
482 * cases which we are happy to fail). And we hold a reference,
483 * so refcount care in invalidate_complete_page's remove_mapping
484 * prevents drop_caches from setting mapping to NULL beneath us.
486 * The case we do have to guard against is when memory pressure made
487 * shmem_writepage move it from filecache to swapcache beneath us:
488 * an unlikely race, but we do need to retry for page_head->mapping.
490 if (!page_head
->mapping
) {
491 int shmem_swizzled
= PageSwapCache(page_head
);
492 unlock_page(page_head
);
500 * Private mappings are handled in a simple way.
502 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
503 * it's a read-only handle, it's expected that futexes attach to
504 * the object not the particular process.
506 if (PageAnon(page_head
)) {
508 * A RO anonymous page will never change and thus doesn't make
509 * sense for futex operations.
516 key
->both
.offset
|= FUT_OFF_MMSHARED
; /* ref taken on mm */
517 key
->private.mm
= mm
;
518 key
->private.address
= address
;
520 key
->both
.offset
|= FUT_OFF_INODE
; /* inode-based key */
521 key
->shared
.inode
= page_head
->mapping
->host
;
522 key
->shared
.pgoff
= basepage_index(page
);
525 get_futex_key_refs(key
); /* implies MB (B) */
528 unlock_page(page_head
);
533 static inline void put_futex_key(union futex_key
*key
)
535 drop_futex_key_refs(key
);
539 * fault_in_user_writeable() - Fault in user address and verify RW access
540 * @uaddr: pointer to faulting user space address
542 * Slow path to fixup the fault we just took in the atomic write
545 * We have no generic implementation of a non-destructive write to the
546 * user address. We know that we faulted in the atomic pagefault
547 * disabled section so we can as well avoid the #PF overhead by
548 * calling get_user_pages() right away.
550 static int fault_in_user_writeable(u32 __user
*uaddr
)
552 struct mm_struct
*mm
= current
->mm
;
555 down_read(&mm
->mmap_sem
);
556 ret
= fixup_user_fault(current
, mm
, (unsigned long)uaddr
,
558 up_read(&mm
->mmap_sem
);
560 return ret
< 0 ? ret
: 0;
564 * futex_top_waiter() - Return the highest priority waiter on a futex
565 * @hb: the hash bucket the futex_q's reside in
566 * @key: the futex key (to distinguish it from other futex futex_q's)
568 * Must be called with the hb lock held.
570 static struct futex_q
*futex_top_waiter(struct futex_hash_bucket
*hb
,
571 union futex_key
*key
)
573 struct futex_q
*this;
575 plist_for_each_entry(this, &hb
->chain
, list
) {
576 if (match_futex(&this->key
, key
))
582 static int cmpxchg_futex_value_locked(u32
*curval
, u32 __user
*uaddr
,
583 u32 uval
, u32 newval
)
588 ret
= futex_atomic_cmpxchg_inatomic(curval
, uaddr
, uval
, newval
);
594 static int get_futex_value_locked(u32
*dest
, u32 __user
*from
)
599 ret
= __copy_from_user_inatomic(dest
, from
, sizeof(u32
));
602 return ret
? -EFAULT
: 0;
609 static int refill_pi_state_cache(void)
611 struct futex_pi_state
*pi_state
;
613 if (likely(current
->pi_state_cache
))
616 pi_state
= kzalloc(sizeof(*pi_state
), GFP_KERNEL
);
621 INIT_LIST_HEAD(&pi_state
->list
);
622 /* pi_mutex gets initialized later */
623 pi_state
->owner
= NULL
;
624 atomic_set(&pi_state
->refcount
, 1);
625 pi_state
->key
= FUTEX_KEY_INIT
;
627 current
->pi_state_cache
= pi_state
;
632 static struct futex_pi_state
* alloc_pi_state(void)
634 struct futex_pi_state
*pi_state
= current
->pi_state_cache
;
637 current
->pi_state_cache
= NULL
;
642 static void free_pi_state(struct futex_pi_state
*pi_state
)
644 if (!atomic_dec_and_test(&pi_state
->refcount
))
648 * If pi_state->owner is NULL, the owner is most probably dying
649 * and has cleaned up the pi_state already
651 if (pi_state
->owner
) {
652 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
653 list_del_init(&pi_state
->list
);
654 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
656 rt_mutex_proxy_unlock(&pi_state
->pi_mutex
, pi_state
->owner
);
659 if (current
->pi_state_cache
)
663 * pi_state->list is already empty.
664 * clear pi_state->owner.
665 * refcount is at 0 - put it back to 1.
667 pi_state
->owner
= NULL
;
668 atomic_set(&pi_state
->refcount
, 1);
669 current
->pi_state_cache
= pi_state
;
674 * Look up the task based on what TID userspace gave us.
677 static struct task_struct
* futex_find_get_task(pid_t pid
)
679 struct task_struct
*p
;
682 p
= find_task_by_vpid(pid
);
692 * This task is holding PI mutexes at exit time => bad.
693 * Kernel cleans up PI-state, but userspace is likely hosed.
694 * (Robust-futex cleanup is separate and might save the day for userspace.)
696 void exit_pi_state_list(struct task_struct
*curr
)
698 struct list_head
*next
, *head
= &curr
->pi_state_list
;
699 struct futex_pi_state
*pi_state
;
700 struct futex_hash_bucket
*hb
;
701 union futex_key key
= FUTEX_KEY_INIT
;
703 if (!futex_cmpxchg_enabled
)
706 * We are a ZOMBIE and nobody can enqueue itself on
707 * pi_state_list anymore, but we have to be careful
708 * versus waiters unqueueing themselves:
710 raw_spin_lock_irq(&curr
->pi_lock
);
711 while (!list_empty(head
)) {
714 pi_state
= list_entry(next
, struct futex_pi_state
, list
);
716 hb
= hash_futex(&key
);
717 raw_spin_unlock_irq(&curr
->pi_lock
);
719 spin_lock(&hb
->lock
);
721 raw_spin_lock_irq(&curr
->pi_lock
);
723 * We dropped the pi-lock, so re-check whether this
724 * task still owns the PI-state:
726 if (head
->next
!= next
) {
727 spin_unlock(&hb
->lock
);
731 WARN_ON(pi_state
->owner
!= curr
);
732 WARN_ON(list_empty(&pi_state
->list
));
733 list_del_init(&pi_state
->list
);
734 pi_state
->owner
= NULL
;
735 raw_spin_unlock_irq(&curr
->pi_lock
);
737 rt_mutex_unlock(&pi_state
->pi_mutex
);
739 spin_unlock(&hb
->lock
);
741 raw_spin_lock_irq(&curr
->pi_lock
);
743 raw_spin_unlock_irq(&curr
->pi_lock
);
747 lookup_pi_state(u32 uval
, struct futex_hash_bucket
*hb
,
748 union futex_key
*key
, struct futex_pi_state
**ps
,
749 struct task_struct
*task
)
751 struct futex_pi_state
*pi_state
= NULL
;
752 struct futex_q
*this, *next
;
753 struct task_struct
*p
;
754 pid_t pid
= uval
& FUTEX_TID_MASK
;
756 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
757 if (match_futex(&this->key
, key
)) {
759 * Another waiter already exists - bump up
760 * the refcount and return its pi_state:
762 pi_state
= this->pi_state
;
764 * Userspace might have messed up non-PI and PI futexes
766 if (unlikely(!pi_state
))
769 WARN_ON(!atomic_read(&pi_state
->refcount
));
772 * When pi_state->owner is NULL then the owner died
773 * and another waiter is on the fly. pi_state->owner
774 * is fixed up by the task which acquires
775 * pi_state->rt_mutex.
777 * We do not check for pid == 0 which can happen when
778 * the owner died and robust_list_exit() cleared the
781 if (pid
&& pi_state
->owner
) {
783 * Bail out if user space manipulated the
786 if (pid
!= task_pid_vnr(pi_state
->owner
))
791 * Protect against a corrupted uval. If uval
792 * is 0x80000000 then pid is 0 and the waiter
793 * bit is set. So the deadlock check in the
794 * calling code has failed and we did not fall
795 * into the check above due to !pid.
797 if (task
&& pi_state
->owner
== task
)
800 atomic_inc(&pi_state
->refcount
);
808 * We are the first waiter - try to look up the real owner and attach
809 * the new pi_state to it, but bail out when TID = 0
813 p
= futex_find_get_task(pid
);
818 * We need to look at the task state flags to figure out,
819 * whether the task is exiting. To protect against the do_exit
820 * change of the task flags, we do this protected by
823 raw_spin_lock_irq(&p
->pi_lock
);
824 if (unlikely(p
->flags
& PF_EXITING
)) {
826 * The task is on the way out. When PF_EXITPIDONE is
827 * set, we know that the task has finished the
830 int ret
= (p
->flags
& PF_EXITPIDONE
) ? -ESRCH
: -EAGAIN
;
832 raw_spin_unlock_irq(&p
->pi_lock
);
837 pi_state
= alloc_pi_state();
840 * Initialize the pi_mutex in locked state and make 'p'
843 rt_mutex_init_proxy_locked(&pi_state
->pi_mutex
, p
);
845 /* Store the key for possible exit cleanups: */
846 pi_state
->key
= *key
;
848 WARN_ON(!list_empty(&pi_state
->list
));
849 list_add(&pi_state
->list
, &p
->pi_state_list
);
851 raw_spin_unlock_irq(&p
->pi_lock
);
861 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
862 * @uaddr: the pi futex user address
863 * @hb: the pi futex hash bucket
864 * @key: the futex key associated with uaddr and hb
865 * @ps: the pi_state pointer where we store the result of the
867 * @task: the task to perform the atomic lock work for. This will
868 * be "current" except in the case of requeue pi.
869 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
873 * 1 - acquired the lock;
876 * The hb->lock and futex_key refs shall be held by the caller.
878 static int futex_lock_pi_atomic(u32 __user
*uaddr
, struct futex_hash_bucket
*hb
,
879 union futex_key
*key
,
880 struct futex_pi_state
**ps
,
881 struct task_struct
*task
, int set_waiters
)
883 int lock_taken
, ret
, force_take
= 0;
884 u32 uval
, newval
, curval
, vpid
= task_pid_vnr(task
);
887 ret
= lock_taken
= 0;
890 * To avoid races, we attempt to take the lock here again
891 * (by doing a 0 -> TID atomic cmpxchg), while holding all
892 * the locks. It will most likely not succeed.
896 newval
|= FUTEX_WAITERS
;
898 if (unlikely(cmpxchg_futex_value_locked(&curval
, uaddr
, 0, newval
)))
904 if ((unlikely((curval
& FUTEX_TID_MASK
) == vpid
)))
908 * Surprise - we got the lock. Just return to userspace:
910 if (unlikely(!curval
))
916 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
917 * to wake at the next unlock.
919 newval
= curval
| FUTEX_WAITERS
;
922 * Should we force take the futex? See below.
924 if (unlikely(force_take
)) {
926 * Keep the OWNER_DIED and the WAITERS bit and set the
929 newval
= (curval
& ~FUTEX_TID_MASK
) | vpid
;
934 if (unlikely(cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
)))
936 if (unlikely(curval
!= uval
))
940 * We took the lock due to forced take over.
942 if (unlikely(lock_taken
))
946 * We dont have the lock. Look up the PI state (or create it if
947 * we are the first waiter):
949 ret
= lookup_pi_state(uval
, hb
, key
, ps
, task
);
955 * We failed to find an owner for this
956 * futex. So we have no pi_state to block
957 * on. This can happen in two cases:
960 * 2) A stale FUTEX_WAITERS bit
962 * Re-read the futex value.
964 if (get_futex_value_locked(&curval
, uaddr
))
968 * If the owner died or we have a stale
969 * WAITERS bit the owner TID in the user space
972 if (!(curval
& FUTEX_TID_MASK
)) {
985 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
986 * @q: The futex_q to unqueue
988 * The q->lock_ptr must not be NULL and must be held by the caller.
990 static void __unqueue_futex(struct futex_q
*q
)
992 struct futex_hash_bucket
*hb
;
994 if (WARN_ON_SMP(!q
->lock_ptr
|| !spin_is_locked(q
->lock_ptr
))
995 || WARN_ON(plist_node_empty(&q
->list
)))
998 hb
= container_of(q
->lock_ptr
, struct futex_hash_bucket
, lock
);
999 plist_del(&q
->list
, &hb
->chain
);
1004 * The hash bucket lock must be held when this is called.
1005 * Afterwards, the futex_q must not be accessed.
1007 static void wake_futex(struct futex_q
*q
)
1009 struct task_struct
*p
= q
->task
;
1011 if (WARN(q
->pi_state
|| q
->rt_waiter
, "refusing to wake PI futex\n"))
1015 * We set q->lock_ptr = NULL _before_ we wake up the task. If
1016 * a non-futex wake up happens on another CPU then the task
1017 * might exit and p would dereference a non-existing task
1018 * struct. Prevent this by holding a reference on p across the
1025 * The waiting task can free the futex_q as soon as
1026 * q->lock_ptr = NULL is written, without taking any locks. A
1027 * memory barrier is required here to prevent the following
1028 * store to lock_ptr from getting ahead of the plist_del.
1033 wake_up_state(p
, TASK_NORMAL
);
1037 static int wake_futex_pi(u32 __user
*uaddr
, u32 uval
, struct futex_q
*this)
1039 struct task_struct
*new_owner
;
1040 struct futex_pi_state
*pi_state
= this->pi_state
;
1041 u32
uninitialized_var(curval
), newval
;
1047 * If current does not own the pi_state then the futex is
1048 * inconsistent and user space fiddled with the futex value.
1050 if (pi_state
->owner
!= current
)
1053 raw_spin_lock(&pi_state
->pi_mutex
.wait_lock
);
1054 new_owner
= rt_mutex_next_owner(&pi_state
->pi_mutex
);
1057 * It is possible that the next waiter (the one that brought
1058 * this owner to the kernel) timed out and is no longer
1059 * waiting on the lock.
1062 new_owner
= this->task
;
1065 * We pass it to the next owner. (The WAITERS bit is always
1066 * kept enabled while there is PI state around. We must also
1067 * preserve the owner died bit.)
1069 if (!(uval
& FUTEX_OWNER_DIED
)) {
1072 newval
= FUTEX_WAITERS
| task_pid_vnr(new_owner
);
1074 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
))
1076 else if (curval
!= uval
)
1079 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
1084 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
1085 WARN_ON(list_empty(&pi_state
->list
));
1086 list_del_init(&pi_state
->list
);
1087 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
1089 raw_spin_lock_irq(&new_owner
->pi_lock
);
1090 WARN_ON(!list_empty(&pi_state
->list
));
1091 list_add(&pi_state
->list
, &new_owner
->pi_state_list
);
1092 pi_state
->owner
= new_owner
;
1093 raw_spin_unlock_irq(&new_owner
->pi_lock
);
1095 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
1096 rt_mutex_unlock(&pi_state
->pi_mutex
);
1101 static int unlock_futex_pi(u32 __user
*uaddr
, u32 uval
)
1103 u32
uninitialized_var(oldval
);
1106 * There is no waiter, so we unlock the futex. The owner died
1107 * bit has not to be preserved here. We are the owner:
1109 if (cmpxchg_futex_value_locked(&oldval
, uaddr
, uval
, 0))
1118 * Express the locking dependencies for lockdep:
1121 double_lock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1124 spin_lock(&hb1
->lock
);
1126 spin_lock_nested(&hb2
->lock
, SINGLE_DEPTH_NESTING
);
1127 } else { /* hb1 > hb2 */
1128 spin_lock(&hb2
->lock
);
1129 spin_lock_nested(&hb1
->lock
, SINGLE_DEPTH_NESTING
);
1134 double_unlock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1136 spin_unlock(&hb1
->lock
);
1138 spin_unlock(&hb2
->lock
);
1142 * Wake up waiters matching bitset queued on this futex (uaddr).
1145 futex_wake(u32 __user
*uaddr
, unsigned int flags
, int nr_wake
, u32 bitset
)
1147 struct futex_hash_bucket
*hb
;
1148 struct futex_q
*this, *next
;
1149 union futex_key key
= FUTEX_KEY_INIT
;
1155 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_READ
);
1156 if (unlikely(ret
!= 0))
1159 hb
= hash_futex(&key
);
1161 /* Make sure we really have tasks to wakeup */
1162 if (!hb_waiters_pending(hb
))
1165 spin_lock(&hb
->lock
);
1167 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
1168 if (match_futex (&this->key
, &key
)) {
1169 if (this->pi_state
|| this->rt_waiter
) {
1174 /* Check if one of the bits is set in both bitsets */
1175 if (!(this->bitset
& bitset
))
1179 if (++ret
>= nr_wake
)
1184 spin_unlock(&hb
->lock
);
1186 put_futex_key(&key
);
1192 * Wake up all waiters hashed on the physical page that is mapped
1193 * to this virtual address:
1196 futex_wake_op(u32 __user
*uaddr1
, unsigned int flags
, u32 __user
*uaddr2
,
1197 int nr_wake
, int nr_wake2
, int op
)
1199 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1200 struct futex_hash_bucket
*hb1
, *hb2
;
1201 struct futex_q
*this, *next
;
1205 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1206 if (unlikely(ret
!= 0))
1208 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
1209 if (unlikely(ret
!= 0))
1212 hb1
= hash_futex(&key1
);
1213 hb2
= hash_futex(&key2
);
1216 double_lock_hb(hb1
, hb2
);
1217 op_ret
= futex_atomic_op_inuser(op
, uaddr2
);
1218 if (unlikely(op_ret
< 0)) {
1220 double_unlock_hb(hb1
, hb2
);
1224 * we don't get EFAULT from MMU faults if we don't have an MMU,
1225 * but we might get them from range checking
1231 if (unlikely(op_ret
!= -EFAULT
)) {
1236 ret
= fault_in_user_writeable(uaddr2
);
1240 if (!(flags
& FLAGS_SHARED
))
1243 put_futex_key(&key2
);
1244 put_futex_key(&key1
);
1248 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1249 if (match_futex (&this->key
, &key1
)) {
1250 if (this->pi_state
|| this->rt_waiter
) {
1255 if (++ret
>= nr_wake
)
1262 plist_for_each_entry_safe(this, next
, &hb2
->chain
, list
) {
1263 if (match_futex (&this->key
, &key2
)) {
1264 if (this->pi_state
|| this->rt_waiter
) {
1269 if (++op_ret
>= nr_wake2
)
1277 double_unlock_hb(hb1
, hb2
);
1279 put_futex_key(&key2
);
1281 put_futex_key(&key1
);
1287 * requeue_futex() - Requeue a futex_q from one hb to another
1288 * @q: the futex_q to requeue
1289 * @hb1: the source hash_bucket
1290 * @hb2: the target hash_bucket
1291 * @key2: the new key for the requeued futex_q
1294 void requeue_futex(struct futex_q
*q
, struct futex_hash_bucket
*hb1
,
1295 struct futex_hash_bucket
*hb2
, union futex_key
*key2
)
1299 * If key1 and key2 hash to the same bucket, no need to
1302 if (likely(&hb1
->chain
!= &hb2
->chain
)) {
1303 plist_del(&q
->list
, &hb1
->chain
);
1304 hb_waiters_dec(hb1
);
1305 plist_add(&q
->list
, &hb2
->chain
);
1306 hb_waiters_inc(hb2
);
1307 q
->lock_ptr
= &hb2
->lock
;
1309 get_futex_key_refs(key2
);
1314 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1316 * @key: the key of the requeue target futex
1317 * @hb: the hash_bucket of the requeue target futex
1319 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1320 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1321 * to the requeue target futex so the waiter can detect the wakeup on the right
1322 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1323 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1324 * to protect access to the pi_state to fixup the owner later. Must be called
1325 * with both q->lock_ptr and hb->lock held.
1328 void requeue_pi_wake_futex(struct futex_q
*q
, union futex_key
*key
,
1329 struct futex_hash_bucket
*hb
)
1331 get_futex_key_refs(key
);
1336 WARN_ON(!q
->rt_waiter
);
1337 q
->rt_waiter
= NULL
;
1339 q
->lock_ptr
= &hb
->lock
;
1341 wake_up_state(q
->task
, TASK_NORMAL
);
1345 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1346 * @pifutex: the user address of the to futex
1347 * @hb1: the from futex hash bucket, must be locked by the caller
1348 * @hb2: the to futex hash bucket, must be locked by the caller
1349 * @key1: the from futex key
1350 * @key2: the to futex key
1351 * @ps: address to store the pi_state pointer
1352 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1354 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1355 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1356 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1357 * hb1 and hb2 must be held by the caller.
1360 * 0 - failed to acquire the lock atomically;
1361 * >0 - acquired the lock, return value is vpid of the top_waiter
1364 static int futex_proxy_trylock_atomic(u32 __user
*pifutex
,
1365 struct futex_hash_bucket
*hb1
,
1366 struct futex_hash_bucket
*hb2
,
1367 union futex_key
*key1
, union futex_key
*key2
,
1368 struct futex_pi_state
**ps
, int set_waiters
)
1370 struct futex_q
*top_waiter
= NULL
;
1374 if (get_futex_value_locked(&curval
, pifutex
))
1378 * Find the top_waiter and determine if there are additional waiters.
1379 * If the caller intends to requeue more than 1 waiter to pifutex,
1380 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1381 * as we have means to handle the possible fault. If not, don't set
1382 * the bit unecessarily as it will force the subsequent unlock to enter
1385 top_waiter
= futex_top_waiter(hb1
, key1
);
1387 /* There are no waiters, nothing for us to do. */
1391 /* Ensure we requeue to the expected futex. */
1392 if (!match_futex(top_waiter
->requeue_pi_key
, key2
))
1396 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1397 * the contended case or if set_waiters is 1. The pi_state is returned
1398 * in ps in contended cases.
1400 vpid
= task_pid_vnr(top_waiter
->task
);
1401 ret
= futex_lock_pi_atomic(pifutex
, hb2
, key2
, ps
, top_waiter
->task
,
1404 requeue_pi_wake_futex(top_waiter
, key2
, hb2
);
1411 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1412 * @uaddr1: source futex user address
1413 * @flags: futex flags (FLAGS_SHARED, etc.)
1414 * @uaddr2: target futex user address
1415 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1416 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1417 * @cmpval: @uaddr1 expected value (or %NULL)
1418 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1419 * pi futex (pi to pi requeue is not supported)
1421 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1422 * uaddr2 atomically on behalf of the top waiter.
1425 * >=0 - on success, the number of tasks requeued or woken;
1428 static int futex_requeue(u32 __user
*uaddr1
, unsigned int flags
,
1429 u32 __user
*uaddr2
, int nr_wake
, int nr_requeue
,
1430 u32
*cmpval
, int requeue_pi
)
1432 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1433 int drop_count
= 0, task_count
= 0, ret
;
1434 struct futex_pi_state
*pi_state
= NULL
;
1435 struct futex_hash_bucket
*hb1
, *hb2
;
1436 struct futex_q
*this, *next
;
1440 * requeue_pi requires a pi_state, try to allocate it now
1441 * without any locks in case it fails.
1443 if (refill_pi_state_cache())
1446 * requeue_pi must wake as many tasks as it can, up to nr_wake
1447 * + nr_requeue, since it acquires the rt_mutex prior to
1448 * returning to userspace, so as to not leave the rt_mutex with
1449 * waiters and no owner. However, second and third wake-ups
1450 * cannot be predicted as they involve race conditions with the
1451 * first wake and a fault while looking up the pi_state. Both
1452 * pthread_cond_signal() and pthread_cond_broadcast() should
1460 if (pi_state
!= NULL
) {
1462 * We will have to lookup the pi_state again, so free this one
1463 * to keep the accounting correct.
1465 free_pi_state(pi_state
);
1469 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1470 if (unlikely(ret
!= 0))
1472 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
,
1473 requeue_pi
? VERIFY_WRITE
: VERIFY_READ
);
1474 if (unlikely(ret
!= 0))
1477 hb1
= hash_futex(&key1
);
1478 hb2
= hash_futex(&key2
);
1481 hb_waiters_inc(hb2
);
1482 double_lock_hb(hb1
, hb2
);
1484 if (likely(cmpval
!= NULL
)) {
1487 ret
= get_futex_value_locked(&curval
, uaddr1
);
1489 if (unlikely(ret
)) {
1490 double_unlock_hb(hb1
, hb2
);
1491 hb_waiters_dec(hb2
);
1493 ret
= get_user(curval
, uaddr1
);
1497 if (!(flags
& FLAGS_SHARED
))
1500 put_futex_key(&key2
);
1501 put_futex_key(&key1
);
1504 if (curval
!= *cmpval
) {
1510 if (requeue_pi
&& (task_count
- nr_wake
< nr_requeue
)) {
1512 * Attempt to acquire uaddr2 and wake the top waiter. If we
1513 * intend to requeue waiters, force setting the FUTEX_WAITERS
1514 * bit. We force this here where we are able to easily handle
1515 * faults rather in the requeue loop below.
1517 ret
= futex_proxy_trylock_atomic(uaddr2
, hb1
, hb2
, &key1
,
1518 &key2
, &pi_state
, nr_requeue
);
1521 * At this point the top_waiter has either taken uaddr2 or is
1522 * waiting on it. If the former, then the pi_state will not
1523 * exist yet, look it up one more time to ensure we have a
1524 * reference to it. If the lock was taken, ret contains the
1525 * vpid of the top waiter task.
1532 * If we acquired the lock, then the user
1533 * space value of uaddr2 should be vpid. It
1534 * cannot be changed by the top waiter as it
1535 * is blocked on hb2 lock if it tries to do
1536 * so. If something fiddled with it behind our
1537 * back the pi state lookup might unearth
1538 * it. So we rather use the known value than
1539 * rereading and handing potential crap to
1542 ret
= lookup_pi_state(ret
, hb2
, &key2
, &pi_state
, NULL
);
1549 double_unlock_hb(hb1
, hb2
);
1550 hb_waiters_dec(hb2
);
1551 put_futex_key(&key2
);
1552 put_futex_key(&key1
);
1553 ret
= fault_in_user_writeable(uaddr2
);
1558 /* The owner was exiting, try again. */
1559 double_unlock_hb(hb1
, hb2
);
1560 hb_waiters_dec(hb2
);
1561 put_futex_key(&key2
);
1562 put_futex_key(&key1
);
1570 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1571 if (task_count
- nr_wake
>= nr_requeue
)
1574 if (!match_futex(&this->key
, &key1
))
1578 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1579 * be paired with each other and no other futex ops.
1581 * We should never be requeueing a futex_q with a pi_state,
1582 * which is awaiting a futex_unlock_pi().
1584 if ((requeue_pi
&& !this->rt_waiter
) ||
1585 (!requeue_pi
&& this->rt_waiter
) ||
1592 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1593 * lock, we already woke the top_waiter. If not, it will be
1594 * woken by futex_unlock_pi().
1596 if (++task_count
<= nr_wake
&& !requeue_pi
) {
1601 /* Ensure we requeue to the expected futex for requeue_pi. */
1602 if (requeue_pi
&& !match_futex(this->requeue_pi_key
, &key2
)) {
1608 * Requeue nr_requeue waiters and possibly one more in the case
1609 * of requeue_pi if we couldn't acquire the lock atomically.
1612 /* Prepare the waiter to take the rt_mutex. */
1613 atomic_inc(&pi_state
->refcount
);
1614 this->pi_state
= pi_state
;
1615 ret
= rt_mutex_start_proxy_lock(&pi_state
->pi_mutex
,
1619 /* We got the lock. */
1620 requeue_pi_wake_futex(this, &key2
, hb2
);
1625 this->pi_state
= NULL
;
1626 free_pi_state(pi_state
);
1630 requeue_futex(this, hb1
, hb2
, &key2
);
1635 double_unlock_hb(hb1
, hb2
);
1636 hb_waiters_dec(hb2
);
1639 * drop_futex_key_refs() must be called outside the spinlocks. During
1640 * the requeue we moved futex_q's from the hash bucket at key1 to the
1641 * one at key2 and updated their key pointer. We no longer need to
1642 * hold the references to key1.
1644 while (--drop_count
>= 0)
1645 drop_futex_key_refs(&key1
);
1648 put_futex_key(&key2
);
1650 put_futex_key(&key1
);
1652 if (pi_state
!= NULL
)
1653 free_pi_state(pi_state
);
1654 return ret
? ret
: task_count
;
1657 /* The key must be already stored in q->key. */
1658 static inline struct futex_hash_bucket
*queue_lock(struct futex_q
*q
)
1659 __acquires(&hb
->lock
)
1661 struct futex_hash_bucket
*hb
;
1663 hb
= hash_futex(&q
->key
);
1666 * Increment the counter before taking the lock so that
1667 * a potential waker won't miss a to-be-slept task that is
1668 * waiting for the spinlock. This is safe as all queue_lock()
1669 * users end up calling queue_me(). Similarly, for housekeeping,
1670 * decrement the counter at queue_unlock() when some error has
1671 * occurred and we don't end up adding the task to the list.
1675 q
->lock_ptr
= &hb
->lock
;
1677 spin_lock(&hb
->lock
); /* implies MB (A) */
1682 queue_unlock(struct futex_hash_bucket
*hb
)
1683 __releases(&hb
->lock
)
1685 spin_unlock(&hb
->lock
);
1690 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1691 * @q: The futex_q to enqueue
1692 * @hb: The destination hash bucket
1694 * The hb->lock must be held by the caller, and is released here. A call to
1695 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1696 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1697 * or nothing if the unqueue is done as part of the wake process and the unqueue
1698 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1701 static inline void queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
1702 __releases(&hb
->lock
)
1707 * The priority used to register this element is
1708 * - either the real thread-priority for the real-time threads
1709 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1710 * - or MAX_RT_PRIO for non-RT threads.
1711 * Thus, all RT-threads are woken first in priority order, and
1712 * the others are woken last, in FIFO order.
1714 prio
= min(current
->normal_prio
, MAX_RT_PRIO
);
1716 plist_node_init(&q
->list
, prio
);
1717 plist_add(&q
->list
, &hb
->chain
);
1719 spin_unlock(&hb
->lock
);
1723 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1724 * @q: The futex_q to unqueue
1726 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1727 * be paired with exactly one earlier call to queue_me().
1730 * 1 - if the futex_q was still queued (and we removed unqueued it);
1731 * 0 - if the futex_q was already removed by the waking thread
1733 static int unqueue_me(struct futex_q
*q
)
1735 spinlock_t
*lock_ptr
;
1738 /* In the common case we don't take the spinlock, which is nice. */
1740 lock_ptr
= q
->lock_ptr
;
1742 if (lock_ptr
!= NULL
) {
1743 spin_lock(lock_ptr
);
1745 * q->lock_ptr can change between reading it and
1746 * spin_lock(), causing us to take the wrong lock. This
1747 * corrects the race condition.
1749 * Reasoning goes like this: if we have the wrong lock,
1750 * q->lock_ptr must have changed (maybe several times)
1751 * between reading it and the spin_lock(). It can
1752 * change again after the spin_lock() but only if it was
1753 * already changed before the spin_lock(). It cannot,
1754 * however, change back to the original value. Therefore
1755 * we can detect whether we acquired the correct lock.
1757 if (unlikely(lock_ptr
!= q
->lock_ptr
)) {
1758 spin_unlock(lock_ptr
);
1763 BUG_ON(q
->pi_state
);
1765 spin_unlock(lock_ptr
);
1769 drop_futex_key_refs(&q
->key
);
1774 * PI futexes can not be requeued and must remove themself from the
1775 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1778 static void unqueue_me_pi(struct futex_q
*q
)
1779 __releases(q
->lock_ptr
)
1783 BUG_ON(!q
->pi_state
);
1784 free_pi_state(q
->pi_state
);
1787 spin_unlock(q
->lock_ptr
);
1791 * Fixup the pi_state owner with the new owner.
1793 * Must be called with hash bucket lock held and mm->sem held for non
1796 static int fixup_pi_state_owner(u32 __user
*uaddr
, struct futex_q
*q
,
1797 struct task_struct
*newowner
)
1799 u32 newtid
= task_pid_vnr(newowner
) | FUTEX_WAITERS
;
1800 struct futex_pi_state
*pi_state
= q
->pi_state
;
1801 struct task_struct
*oldowner
= pi_state
->owner
;
1802 u32 uval
, uninitialized_var(curval
), newval
;
1806 if (!pi_state
->owner
)
1807 newtid
|= FUTEX_OWNER_DIED
;
1810 * We are here either because we stole the rtmutex from the
1811 * previous highest priority waiter or we are the highest priority
1812 * waiter but failed to get the rtmutex the first time.
1813 * We have to replace the newowner TID in the user space variable.
1814 * This must be atomic as we have to preserve the owner died bit here.
1816 * Note: We write the user space value _before_ changing the pi_state
1817 * because we can fault here. Imagine swapped out pages or a fork
1818 * that marked all the anonymous memory readonly for cow.
1820 * Modifying pi_state _before_ the user space value would
1821 * leave the pi_state in an inconsistent state when we fault
1822 * here, because we need to drop the hash bucket lock to
1823 * handle the fault. This might be observed in the PID check
1824 * in lookup_pi_state.
1827 if (get_futex_value_locked(&uval
, uaddr
))
1831 newval
= (uval
& FUTEX_OWNER_DIED
) | newtid
;
1833 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
))
1841 * We fixed up user space. Now we need to fix the pi_state
1844 if (pi_state
->owner
!= NULL
) {
1845 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
1846 WARN_ON(list_empty(&pi_state
->list
));
1847 list_del_init(&pi_state
->list
);
1848 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
1851 pi_state
->owner
= newowner
;
1853 raw_spin_lock_irq(&newowner
->pi_lock
);
1854 WARN_ON(!list_empty(&pi_state
->list
));
1855 list_add(&pi_state
->list
, &newowner
->pi_state_list
);
1856 raw_spin_unlock_irq(&newowner
->pi_lock
);
1860 * To handle the page fault we need to drop the hash bucket
1861 * lock here. That gives the other task (either the highest priority
1862 * waiter itself or the task which stole the rtmutex) the
1863 * chance to try the fixup of the pi_state. So once we are
1864 * back from handling the fault we need to check the pi_state
1865 * after reacquiring the hash bucket lock and before trying to
1866 * do another fixup. When the fixup has been done already we
1870 spin_unlock(q
->lock_ptr
);
1872 ret
= fault_in_user_writeable(uaddr
);
1874 spin_lock(q
->lock_ptr
);
1877 * Check if someone else fixed it for us:
1879 if (pi_state
->owner
!= oldowner
)
1888 static long futex_wait_restart(struct restart_block
*restart
);
1891 * fixup_owner() - Post lock pi_state and corner case management
1892 * @uaddr: user address of the futex
1893 * @q: futex_q (contains pi_state and access to the rt_mutex)
1894 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1896 * After attempting to lock an rt_mutex, this function is called to cleanup
1897 * the pi_state owner as well as handle race conditions that may allow us to
1898 * acquire the lock. Must be called with the hb lock held.
1901 * 1 - success, lock taken;
1902 * 0 - success, lock not taken;
1903 * <0 - on error (-EFAULT)
1905 static int fixup_owner(u32 __user
*uaddr
, struct futex_q
*q
, int locked
)
1907 struct task_struct
*owner
;
1912 * Got the lock. We might not be the anticipated owner if we
1913 * did a lock-steal - fix up the PI-state in that case:
1915 if (q
->pi_state
->owner
!= current
)
1916 ret
= fixup_pi_state_owner(uaddr
, q
, current
);
1921 * Catch the rare case, where the lock was released when we were on the
1922 * way back before we locked the hash bucket.
1924 if (q
->pi_state
->owner
== current
) {
1926 * Try to get the rt_mutex now. This might fail as some other
1927 * task acquired the rt_mutex after we removed ourself from the
1928 * rt_mutex waiters list.
1930 if (rt_mutex_trylock(&q
->pi_state
->pi_mutex
)) {
1936 * pi_state is incorrect, some other task did a lock steal and
1937 * we returned due to timeout or signal without taking the
1938 * rt_mutex. Too late.
1940 raw_spin_lock(&q
->pi_state
->pi_mutex
.wait_lock
);
1941 owner
= rt_mutex_owner(&q
->pi_state
->pi_mutex
);
1943 owner
= rt_mutex_next_owner(&q
->pi_state
->pi_mutex
);
1944 raw_spin_unlock(&q
->pi_state
->pi_mutex
.wait_lock
);
1945 ret
= fixup_pi_state_owner(uaddr
, q
, owner
);
1950 * Paranoia check. If we did not take the lock, then we should not be
1951 * the owner of the rt_mutex.
1953 if (rt_mutex_owner(&q
->pi_state
->pi_mutex
) == current
)
1954 printk(KERN_ERR
"fixup_owner: ret = %d pi-mutex: %p "
1955 "pi-state %p\n", ret
,
1956 q
->pi_state
->pi_mutex
.owner
,
1957 q
->pi_state
->owner
);
1960 return ret
? ret
: locked
;
1964 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1965 * @hb: the futex hash bucket, must be locked by the caller
1966 * @q: the futex_q to queue up on
1967 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1969 static void futex_wait_queue_me(struct futex_hash_bucket
*hb
, struct futex_q
*q
,
1970 struct hrtimer_sleeper
*timeout
)
1973 * The task state is guaranteed to be set before another task can
1974 * wake it. set_current_state() is implemented using set_mb() and
1975 * queue_me() calls spin_unlock() upon completion, both serializing
1976 * access to the hash list and forcing another memory barrier.
1978 set_current_state(TASK_INTERRUPTIBLE
);
1983 hrtimer_start_expires(&timeout
->timer
, HRTIMER_MODE_ABS
);
1984 if (!hrtimer_active(&timeout
->timer
))
1985 timeout
->task
= NULL
;
1989 * If we have been removed from the hash list, then another task
1990 * has tried to wake us, and we can skip the call to schedule().
1992 if (likely(!plist_node_empty(&q
->list
))) {
1994 * If the timer has already expired, current will already be
1995 * flagged for rescheduling. Only call schedule if there
1996 * is no timeout, or if it has yet to expire.
1998 if (!timeout
|| timeout
->task
)
1999 freezable_schedule();
2001 __set_current_state(TASK_RUNNING
);
2005 * futex_wait_setup() - Prepare to wait on a futex
2006 * @uaddr: the futex userspace address
2007 * @val: the expected value
2008 * @flags: futex flags (FLAGS_SHARED, etc.)
2009 * @q: the associated futex_q
2010 * @hb: storage for hash_bucket pointer to be returned to caller
2012 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2013 * compare it with the expected value. Handle atomic faults internally.
2014 * Return with the hb lock held and a q.key reference on success, and unlocked
2015 * with no q.key reference on failure.
2018 * 0 - uaddr contains val and hb has been locked;
2019 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2021 static int futex_wait_setup(u32 __user
*uaddr
, u32 val
, unsigned int flags
,
2022 struct futex_q
*q
, struct futex_hash_bucket
**hb
)
2028 * Access the page AFTER the hash-bucket is locked.
2029 * Order is important:
2031 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2032 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2034 * The basic logical guarantee of a futex is that it blocks ONLY
2035 * if cond(var) is known to be true at the time of blocking, for
2036 * any cond. If we locked the hash-bucket after testing *uaddr, that
2037 * would open a race condition where we could block indefinitely with
2038 * cond(var) false, which would violate the guarantee.
2040 * On the other hand, we insert q and release the hash-bucket only
2041 * after testing *uaddr. This guarantees that futex_wait() will NOT
2042 * absorb a wakeup if *uaddr does not match the desired values
2043 * while the syscall executes.
2046 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
->key
, VERIFY_READ
);
2047 if (unlikely(ret
!= 0))
2051 *hb
= queue_lock(q
);
2053 ret
= get_futex_value_locked(&uval
, uaddr
);
2058 ret
= get_user(uval
, uaddr
);
2062 if (!(flags
& FLAGS_SHARED
))
2065 put_futex_key(&q
->key
);
2076 put_futex_key(&q
->key
);
2080 static int futex_wait(u32 __user
*uaddr
, unsigned int flags
, u32 val
,
2081 ktime_t
*abs_time
, u32 bitset
)
2083 struct hrtimer_sleeper timeout
, *to
= NULL
;
2084 struct restart_block
*restart
;
2085 struct futex_hash_bucket
*hb
;
2086 struct futex_q q
= futex_q_init
;
2096 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2097 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2099 hrtimer_init_sleeper(to
, current
);
2100 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2101 current
->timer_slack_ns
);
2106 * Prepare to wait on uaddr. On success, holds hb lock and increments
2109 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2113 /* queue_me and wait for wakeup, timeout, or a signal. */
2114 futex_wait_queue_me(hb
, &q
, to
);
2116 /* If we were woken (and unqueued), we succeeded, whatever. */
2118 /* unqueue_me() drops q.key ref */
2119 if (!unqueue_me(&q
))
2122 if (to
&& !to
->task
)
2126 * We expect signal_pending(current), but we might be the
2127 * victim of a spurious wakeup as well.
2129 if (!signal_pending(current
))
2136 restart
= ¤t_thread_info()->restart_block
;
2137 restart
->fn
= futex_wait_restart
;
2138 restart
->futex
.uaddr
= uaddr
;
2139 restart
->futex
.val
= val
;
2140 restart
->futex
.time
= abs_time
->tv64
;
2141 restart
->futex
.bitset
= bitset
;
2142 restart
->futex
.flags
= flags
| FLAGS_HAS_TIMEOUT
;
2144 ret
= -ERESTART_RESTARTBLOCK
;
2148 hrtimer_cancel(&to
->timer
);
2149 destroy_hrtimer_on_stack(&to
->timer
);
2155 static long futex_wait_restart(struct restart_block
*restart
)
2157 u32 __user
*uaddr
= restart
->futex
.uaddr
;
2158 ktime_t t
, *tp
= NULL
;
2160 if (restart
->futex
.flags
& FLAGS_HAS_TIMEOUT
) {
2161 t
.tv64
= restart
->futex
.time
;
2164 restart
->fn
= do_no_restart_syscall
;
2166 return (long)futex_wait(uaddr
, restart
->futex
.flags
,
2167 restart
->futex
.val
, tp
, restart
->futex
.bitset
);
2172 * Userspace tried a 0 -> TID atomic transition of the futex value
2173 * and failed. The kernel side here does the whole locking operation:
2174 * if there are waiters then it will block, it does PI, etc. (Due to
2175 * races the kernel might see a 0 value of the futex too.)
2177 static int futex_lock_pi(u32 __user
*uaddr
, unsigned int flags
, int detect
,
2178 ktime_t
*time
, int trylock
)
2180 struct hrtimer_sleeper timeout
, *to
= NULL
;
2181 struct futex_hash_bucket
*hb
;
2182 struct futex_q q
= futex_q_init
;
2185 if (refill_pi_state_cache())
2190 hrtimer_init_on_stack(&to
->timer
, CLOCK_REALTIME
,
2192 hrtimer_init_sleeper(to
, current
);
2193 hrtimer_set_expires(&to
->timer
, *time
);
2197 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
.key
, VERIFY_WRITE
);
2198 if (unlikely(ret
!= 0))
2202 hb
= queue_lock(&q
);
2204 ret
= futex_lock_pi_atomic(uaddr
, hb
, &q
.key
, &q
.pi_state
, current
, 0);
2205 if (unlikely(ret
)) {
2208 /* We got the lock. */
2210 goto out_unlock_put_key
;
2215 * Task is exiting and we just wait for the
2219 put_futex_key(&q
.key
);
2223 goto out_unlock_put_key
;
2228 * Only actually queue now that the atomic ops are done:
2232 WARN_ON(!q
.pi_state
);
2234 * Block on the PI mutex:
2237 ret
= rt_mutex_timed_lock(&q
.pi_state
->pi_mutex
, to
, 1);
2239 ret
= rt_mutex_trylock(&q
.pi_state
->pi_mutex
);
2240 /* Fixup the trylock return value: */
2241 ret
= ret
? 0 : -EWOULDBLOCK
;
2244 spin_lock(q
.lock_ptr
);
2246 * Fixup the pi_state owner and possibly acquire the lock if we
2249 res
= fixup_owner(uaddr
, &q
, !ret
);
2251 * If fixup_owner() returned an error, proprogate that. If it acquired
2252 * the lock, clear our -ETIMEDOUT or -EINTR.
2255 ret
= (res
< 0) ? res
: 0;
2258 * If fixup_owner() faulted and was unable to handle the fault, unlock
2259 * it and return the fault to userspace.
2261 if (ret
&& (rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
))
2262 rt_mutex_unlock(&q
.pi_state
->pi_mutex
);
2264 /* Unqueue and drop the lock */
2273 put_futex_key(&q
.key
);
2276 destroy_hrtimer_on_stack(&to
->timer
);
2277 return ret
!= -EINTR
? ret
: -ERESTARTNOINTR
;
2282 ret
= fault_in_user_writeable(uaddr
);
2286 if (!(flags
& FLAGS_SHARED
))
2289 put_futex_key(&q
.key
);
2294 * Userspace attempted a TID -> 0 atomic transition, and failed.
2295 * This is the in-kernel slowpath: we look up the PI state (if any),
2296 * and do the rt-mutex unlock.
2298 static int futex_unlock_pi(u32 __user
*uaddr
, unsigned int flags
)
2300 struct futex_hash_bucket
*hb
;
2301 struct futex_q
*this, *next
;
2302 union futex_key key
= FUTEX_KEY_INIT
;
2303 u32 uval
, vpid
= task_pid_vnr(current
);
2307 if (get_user(uval
, uaddr
))
2310 * We release only a lock we actually own:
2312 if ((uval
& FUTEX_TID_MASK
) != vpid
)
2315 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_WRITE
);
2316 if (unlikely(ret
!= 0))
2319 hb
= hash_futex(&key
);
2320 spin_lock(&hb
->lock
);
2323 * To avoid races, try to do the TID -> 0 atomic transition
2324 * again. If it succeeds then we can return without waking
2327 if (!(uval
& FUTEX_OWNER_DIED
) &&
2328 cmpxchg_futex_value_locked(&uval
, uaddr
, vpid
, 0))
2331 * Rare case: we managed to release the lock atomically,
2332 * no need to wake anyone else up:
2334 if (unlikely(uval
== vpid
))
2338 * Ok, other tasks may need to be woken up - check waiters
2339 * and do the wakeup if necessary:
2341 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
2342 if (!match_futex (&this->key
, &key
))
2344 ret
= wake_futex_pi(uaddr
, uval
, this);
2346 * The atomic access to the futex value
2347 * generated a pagefault, so retry the
2348 * user-access and the wakeup:
2355 * No waiters - kernel unlocks the futex:
2357 if (!(uval
& FUTEX_OWNER_DIED
)) {
2358 ret
= unlock_futex_pi(uaddr
, uval
);
2364 spin_unlock(&hb
->lock
);
2365 put_futex_key(&key
);
2371 spin_unlock(&hb
->lock
);
2372 put_futex_key(&key
);
2374 ret
= fault_in_user_writeable(uaddr
);
2382 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2383 * @hb: the hash_bucket futex_q was original enqueued on
2384 * @q: the futex_q woken while waiting to be requeued
2385 * @key2: the futex_key of the requeue target futex
2386 * @timeout: the timeout associated with the wait (NULL if none)
2388 * Detect if the task was woken on the initial futex as opposed to the requeue
2389 * target futex. If so, determine if it was a timeout or a signal that caused
2390 * the wakeup and return the appropriate error code to the caller. Must be
2391 * called with the hb lock held.
2394 * 0 = no early wakeup detected;
2395 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2398 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket
*hb
,
2399 struct futex_q
*q
, union futex_key
*key2
,
2400 struct hrtimer_sleeper
*timeout
)
2405 * With the hb lock held, we avoid races while we process the wakeup.
2406 * We only need to hold hb (and not hb2) to ensure atomicity as the
2407 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2408 * It can't be requeued from uaddr2 to something else since we don't
2409 * support a PI aware source futex for requeue.
2411 if (!match_futex(&q
->key
, key2
)) {
2412 WARN_ON(q
->lock_ptr
&& (&hb
->lock
!= q
->lock_ptr
));
2414 * We were woken prior to requeue by a timeout or a signal.
2415 * Unqueue the futex_q and determine which it was.
2417 plist_del(&q
->list
, &hb
->chain
);
2420 /* Handle spurious wakeups gracefully */
2422 if (timeout
&& !timeout
->task
)
2424 else if (signal_pending(current
))
2425 ret
= -ERESTARTNOINTR
;
2431 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2432 * @uaddr: the futex we initially wait on (non-pi)
2433 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2434 * the same type, no requeueing from private to shared, etc.
2435 * @val: the expected value of uaddr
2436 * @abs_time: absolute timeout
2437 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2438 * @uaddr2: the pi futex we will take prior to returning to user-space
2440 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2441 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2442 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2443 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2444 * without one, the pi logic would not know which task to boost/deboost, if
2445 * there was a need to.
2447 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2448 * via the following--
2449 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2450 * 2) wakeup on uaddr2 after a requeue
2454 * If 3, cleanup and return -ERESTARTNOINTR.
2456 * If 2, we may then block on trying to take the rt_mutex and return via:
2457 * 5) successful lock
2460 * 8) other lock acquisition failure
2462 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2464 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2470 static int futex_wait_requeue_pi(u32 __user
*uaddr
, unsigned int flags
,
2471 u32 val
, ktime_t
*abs_time
, u32 bitset
,
2474 struct hrtimer_sleeper timeout
, *to
= NULL
;
2475 struct rt_mutex_waiter rt_waiter
;
2476 struct rt_mutex
*pi_mutex
= NULL
;
2477 struct futex_hash_bucket
*hb
;
2478 union futex_key key2
= FUTEX_KEY_INIT
;
2479 struct futex_q q
= futex_q_init
;
2482 if (uaddr
== uaddr2
)
2490 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2491 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2493 hrtimer_init_sleeper(to
, current
);
2494 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2495 current
->timer_slack_ns
);
2499 * The waiter is allocated on our stack, manipulated by the requeue
2500 * code while we sleep on uaddr.
2502 debug_rt_mutex_init_waiter(&rt_waiter
);
2503 RB_CLEAR_NODE(&rt_waiter
.pi_tree_entry
);
2504 RB_CLEAR_NODE(&rt_waiter
.tree_entry
);
2505 rt_waiter
.task
= NULL
;
2507 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
2508 if (unlikely(ret
!= 0))
2512 q
.rt_waiter
= &rt_waiter
;
2513 q
.requeue_pi_key
= &key2
;
2516 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2519 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2523 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2524 futex_wait_queue_me(hb
, &q
, to
);
2526 spin_lock(&hb
->lock
);
2527 ret
= handle_early_requeue_pi_wakeup(hb
, &q
, &key2
, to
);
2528 spin_unlock(&hb
->lock
);
2533 * In order for us to be here, we know our q.key == key2, and since
2534 * we took the hb->lock above, we also know that futex_requeue() has
2535 * completed and we no longer have to concern ourselves with a wakeup
2536 * race with the atomic proxy lock acquisition by the requeue code. The
2537 * futex_requeue dropped our key1 reference and incremented our key2
2541 /* Check if the requeue code acquired the second futex for us. */
2544 * Got the lock. We might not be the anticipated owner if we
2545 * did a lock-steal - fix up the PI-state in that case.
2547 if (q
.pi_state
&& (q
.pi_state
->owner
!= current
)) {
2548 spin_lock(q
.lock_ptr
);
2549 ret
= fixup_pi_state_owner(uaddr2
, &q
, current
);
2550 spin_unlock(q
.lock_ptr
);
2554 * We have been woken up by futex_unlock_pi(), a timeout, or a
2555 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2558 WARN_ON(!q
.pi_state
);
2559 pi_mutex
= &q
.pi_state
->pi_mutex
;
2560 ret
= rt_mutex_finish_proxy_lock(pi_mutex
, to
, &rt_waiter
, 1);
2561 debug_rt_mutex_free_waiter(&rt_waiter
);
2563 spin_lock(q
.lock_ptr
);
2565 * Fixup the pi_state owner and possibly acquire the lock if we
2568 res
= fixup_owner(uaddr2
, &q
, !ret
);
2570 * If fixup_owner() returned an error, proprogate that. If it
2571 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2574 ret
= (res
< 0) ? res
: 0;
2576 /* Unqueue and drop the lock. */
2581 * If fixup_pi_state_owner() faulted and was unable to handle the
2582 * fault, unlock the rt_mutex and return the fault to userspace.
2584 if (ret
== -EFAULT
) {
2585 if (pi_mutex
&& rt_mutex_owner(pi_mutex
) == current
)
2586 rt_mutex_unlock(pi_mutex
);
2587 } else if (ret
== -EINTR
) {
2589 * We've already been requeued, but cannot restart by calling
2590 * futex_lock_pi() directly. We could restart this syscall, but
2591 * it would detect that the user space "val" changed and return
2592 * -EWOULDBLOCK. Save the overhead of the restart and return
2593 * -EWOULDBLOCK directly.
2599 put_futex_key(&q
.key
);
2601 put_futex_key(&key2
);
2605 hrtimer_cancel(&to
->timer
);
2606 destroy_hrtimer_on_stack(&to
->timer
);
2612 * Support for robust futexes: the kernel cleans up held futexes at
2615 * Implementation: user-space maintains a per-thread list of locks it
2616 * is holding. Upon do_exit(), the kernel carefully walks this list,
2617 * and marks all locks that are owned by this thread with the
2618 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2619 * always manipulated with the lock held, so the list is private and
2620 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2621 * field, to allow the kernel to clean up if the thread dies after
2622 * acquiring the lock, but just before it could have added itself to
2623 * the list. There can only be one such pending lock.
2627 * sys_set_robust_list() - Set the robust-futex list head of a task
2628 * @head: pointer to the list-head
2629 * @len: length of the list-head, as userspace expects
2631 SYSCALL_DEFINE2(set_robust_list
, struct robust_list_head __user
*, head
,
2634 if (!futex_cmpxchg_enabled
)
2637 * The kernel knows only one size for now:
2639 if (unlikely(len
!= sizeof(*head
)))
2642 current
->robust_list
= head
;
2648 * sys_get_robust_list() - Get the robust-futex list head of a task
2649 * @pid: pid of the process [zero for current task]
2650 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2651 * @len_ptr: pointer to a length field, the kernel fills in the header size
2653 SYSCALL_DEFINE3(get_robust_list
, int, pid
,
2654 struct robust_list_head __user
* __user
*, head_ptr
,
2655 size_t __user
*, len_ptr
)
2657 struct robust_list_head __user
*head
;
2659 struct task_struct
*p
;
2661 if (!futex_cmpxchg_enabled
)
2670 p
= find_task_by_vpid(pid
);
2676 if (!ptrace_may_access(p
, PTRACE_MODE_READ
))
2679 head
= p
->robust_list
;
2682 if (put_user(sizeof(*head
), len_ptr
))
2684 return put_user(head
, head_ptr
);
2693 * Process a futex-list entry, check whether it's owned by the
2694 * dying task, and do notification if so:
2696 int handle_futex_death(u32 __user
*uaddr
, struct task_struct
*curr
, int pi
)
2698 u32 uval
, uninitialized_var(nval
), mval
;
2701 if (get_user(uval
, uaddr
))
2704 if ((uval
& FUTEX_TID_MASK
) == task_pid_vnr(curr
)) {
2706 * Ok, this dying thread is truly holding a futex
2707 * of interest. Set the OWNER_DIED bit atomically
2708 * via cmpxchg, and if the value had FUTEX_WAITERS
2709 * set, wake up a waiter (if any). (We have to do a
2710 * futex_wake() even if OWNER_DIED is already set -
2711 * to handle the rare but possible case of recursive
2712 * thread-death.) The rest of the cleanup is done in
2715 mval
= (uval
& FUTEX_WAITERS
) | FUTEX_OWNER_DIED
;
2717 * We are not holding a lock here, but we want to have
2718 * the pagefault_disable/enable() protection because
2719 * we want to handle the fault gracefully. If the
2720 * access fails we try to fault in the futex with R/W
2721 * verification via get_user_pages. get_user() above
2722 * does not guarantee R/W access. If that fails we
2723 * give up and leave the futex locked.
2725 if (cmpxchg_futex_value_locked(&nval
, uaddr
, uval
, mval
)) {
2726 if (fault_in_user_writeable(uaddr
))
2734 * Wake robust non-PI futexes here. The wakeup of
2735 * PI futexes happens in exit_pi_state():
2737 if (!pi
&& (uval
& FUTEX_WAITERS
))
2738 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
2744 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2746 static inline int fetch_robust_entry(struct robust_list __user
**entry
,
2747 struct robust_list __user
* __user
*head
,
2750 unsigned long uentry
;
2752 if (get_user(uentry
, (unsigned long __user
*)head
))
2755 *entry
= (void __user
*)(uentry
& ~1UL);
2762 * Walk curr->robust_list (very carefully, it's a userspace list!)
2763 * and mark any locks found there dead, and notify any waiters.
2765 * We silently return on any sign of list-walking problem.
2767 void exit_robust_list(struct task_struct
*curr
)
2769 struct robust_list_head __user
*head
= curr
->robust_list
;
2770 struct robust_list __user
*entry
, *next_entry
, *pending
;
2771 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
2772 unsigned int uninitialized_var(next_pi
);
2773 unsigned long futex_offset
;
2776 if (!futex_cmpxchg_enabled
)
2780 * Fetch the list head (which was registered earlier, via
2781 * sys_set_robust_list()):
2783 if (fetch_robust_entry(&entry
, &head
->list
.next
, &pi
))
2786 * Fetch the relative futex offset:
2788 if (get_user(futex_offset
, &head
->futex_offset
))
2791 * Fetch any possibly pending lock-add first, and handle it
2794 if (fetch_robust_entry(&pending
, &head
->list_op_pending
, &pip
))
2797 next_entry
= NULL
; /* avoid warning with gcc */
2798 while (entry
!= &head
->list
) {
2800 * Fetch the next entry in the list before calling
2801 * handle_futex_death:
2803 rc
= fetch_robust_entry(&next_entry
, &entry
->next
, &next_pi
);
2805 * A pending lock might already be on the list, so
2806 * don't process it twice:
2808 if (entry
!= pending
)
2809 if (handle_futex_death((void __user
*)entry
+ futex_offset
,
2817 * Avoid excessively long or circular lists:
2826 handle_futex_death((void __user
*)pending
+ futex_offset
,
2830 long do_futex(u32 __user
*uaddr
, int op
, u32 val
, ktime_t
*timeout
,
2831 u32 __user
*uaddr2
, u32 val2
, u32 val3
)
2833 int cmd
= op
& FUTEX_CMD_MASK
;
2834 unsigned int flags
= 0;
2836 if (!(op
& FUTEX_PRIVATE_FLAG
))
2837 flags
|= FLAGS_SHARED
;
2839 if (op
& FUTEX_CLOCK_REALTIME
) {
2840 flags
|= FLAGS_CLOCKRT
;
2841 if (cmd
!= FUTEX_WAIT_BITSET
&& cmd
!= FUTEX_WAIT_REQUEUE_PI
)
2847 case FUTEX_UNLOCK_PI
:
2848 case FUTEX_TRYLOCK_PI
:
2849 case FUTEX_WAIT_REQUEUE_PI
:
2850 case FUTEX_CMP_REQUEUE_PI
:
2851 if (!futex_cmpxchg_enabled
)
2857 val3
= FUTEX_BITSET_MATCH_ANY
;
2858 case FUTEX_WAIT_BITSET
:
2859 return futex_wait(uaddr
, flags
, val
, timeout
, val3
);
2861 val3
= FUTEX_BITSET_MATCH_ANY
;
2862 case FUTEX_WAKE_BITSET
:
2863 return futex_wake(uaddr
, flags
, val
, val3
);
2865 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, NULL
, 0);
2866 case FUTEX_CMP_REQUEUE
:
2867 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 0);
2869 return futex_wake_op(uaddr
, flags
, uaddr2
, val
, val2
, val3
);
2871 return futex_lock_pi(uaddr
, flags
, val
, timeout
, 0);
2872 case FUTEX_UNLOCK_PI
:
2873 return futex_unlock_pi(uaddr
, flags
);
2874 case FUTEX_TRYLOCK_PI
:
2875 return futex_lock_pi(uaddr
, flags
, 0, timeout
, 1);
2876 case FUTEX_WAIT_REQUEUE_PI
:
2877 val3
= FUTEX_BITSET_MATCH_ANY
;
2878 return futex_wait_requeue_pi(uaddr
, flags
, val
, timeout
, val3
,
2880 case FUTEX_CMP_REQUEUE_PI
:
2881 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 1);
2887 SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
2888 struct timespec __user
*, utime
, u32 __user
*, uaddr2
,
2892 ktime_t t
, *tp
= NULL
;
2894 int cmd
= op
& FUTEX_CMD_MASK
;
2896 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
2897 cmd
== FUTEX_WAIT_BITSET
||
2898 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
2899 if (copy_from_user(&ts
, utime
, sizeof(ts
)) != 0)
2901 if (!timespec_valid(&ts
))
2904 t
= timespec_to_ktime(ts
);
2905 if (cmd
== FUTEX_WAIT
)
2906 t
= ktime_add_safe(ktime_get(), t
);
2910 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2911 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2913 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
2914 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
2915 val2
= (u32
) (unsigned long) utime
;
2917 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
2920 static void __init
futex_detect_cmpxchg(void)
2922 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
2926 * This will fail and we want it. Some arch implementations do
2927 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2928 * functionality. We want to know that before we call in any
2929 * of the complex code paths. Also we want to prevent
2930 * registration of robust lists in that case. NULL is
2931 * guaranteed to fault and we get -EFAULT on functional
2932 * implementation, the non-functional ones will return
2935 if (cmpxchg_futex_value_locked(&curval
, NULL
, 0, 0) == -EFAULT
)
2936 futex_cmpxchg_enabled
= 1;
2940 static int __init
futex_init(void)
2942 unsigned int futex_shift
;
2945 #if CONFIG_BASE_SMALL
2946 futex_hashsize
= 16;
2948 futex_hashsize
= roundup_pow_of_two(256 * num_possible_cpus());
2951 futex_queues
= alloc_large_system_hash("futex", sizeof(*futex_queues
),
2953 futex_hashsize
< 256 ? HASH_SMALL
: 0,
2955 futex_hashsize
, futex_hashsize
);
2956 futex_hashsize
= 1UL << futex_shift
;
2958 futex_detect_cmpxchg();
2960 for (i
= 0; i
< futex_hashsize
; i
++) {
2961 atomic_set(&futex_queues
[i
].waiters
, 0);
2962 plist_head_init(&futex_queues
[i
].chain
);
2963 spin_lock_init(&futex_queues
[i
].lock
);
2968 __initcall(futex_init
);