2 * Fast Userspace Mutexes (which I call "Futexes!").
3 * (C) Rusty Russell, IBM 2002
5 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
8 * Removed page pinning, fix privately mapped COW pages and other cleanups
9 * (C) Copyright 2003, 2004 Jamie Lokier
11 * Robust futex support started by Ingo Molnar
12 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
15 * PI-futex support started by Ingo Molnar and Thomas Gleixner
16 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
19 * PRIVATE futexes by Eric Dumazet
20 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
22 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23 * Copyright (C) IBM Corporation, 2009
24 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
26 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27 * enough at me, Linus for the original (flawed) idea, Matthew
28 * Kirkwood for proof-of-concept implementation.
30 * "The futexes are also cursed."
31 * "But they come in a choice of three flavours!"
33 * This program is free software; you can redistribute it and/or modify
34 * it under the terms of the GNU General Public License as published by
35 * the Free Software Foundation; either version 2 of the License, or
36 * (at your option) any later version.
38 * This program is distributed in the hope that it will be useful,
39 * but WITHOUT ANY WARRANTY; without even the implied warranty of
40 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
41 * GNU General Public License for more details.
43 * You should have received a copy of the GNU General Public License
44 * along with this program; if not, write to the Free Software
45 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
47 #include <linux/slab.h>
48 #include <linux/poll.h>
50 #include <linux/file.h>
51 #include <linux/jhash.h>
52 #include <linux/init.h>
53 #include <linux/futex.h>
54 #include <linux/mount.h>
55 #include <linux/pagemap.h>
56 #include <linux/syscalls.h>
57 #include <linux/signal.h>
58 #include <linux/export.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62 #include <linux/ptrace.h>
63 #include <linux/sched/rt.h>
64 #include <linux/hugetlb.h>
65 #include <linux/freezer.h>
66 #include <linux/bootmem.h>
68 #include <asm/futex.h>
70 #include "locking/rtmutex_common.h"
73 * Basic futex operation and ordering guarantees:
75 * The waiter reads the futex value in user space and calls
76 * futex_wait(). This function computes the hash bucket and acquires
77 * the hash bucket lock. After that it reads the futex user space value
78 * again and verifies that the data has not changed. If it has not changed
79 * it enqueues itself into the hash bucket, releases the hash bucket lock
82 * The waker side modifies the user space value of the futex and calls
83 * futex_wake(). This function computes the hash bucket and acquires the
84 * hash bucket lock. Then it looks for waiters on that futex in the hash
85 * bucket and wakes them.
87 * In futex wake up scenarios where no tasks are blocked on a futex, taking
88 * the hb spinlock can be avoided and simply return. In order for this
89 * optimization to work, ordering guarantees must exist so that the waiter
90 * being added to the list is acknowledged when the list is concurrently being
91 * checked by the waker, avoiding scenarios like the following:
95 * sys_futex(WAIT, futex, val);
96 * futex_wait(futex, val);
99 * sys_futex(WAKE, futex);
104 * lock(hash_bucket(futex));
106 * unlock(hash_bucket(futex));
109 * This would cause the waiter on CPU 0 to wait forever because it
110 * missed the transition of the user space value from val to newval
111 * and the waker did not find the waiter in the hash bucket queue.
113 * The correct serialization ensures that a waiter either observes
114 * the changed user space value before blocking or is woken by a
119 * sys_futex(WAIT, futex, val);
120 * futex_wait(futex, val);
123 * mb(); (A) <-- paired with -.
125 * lock(hash_bucket(futex)); |
129 * | sys_futex(WAKE, futex);
130 * | futex_wake(futex);
132 * `-------> mb(); (B)
135 * unlock(hash_bucket(futex));
136 * schedule(); if (waiters)
137 * lock(hash_bucket(futex));
138 * wake_waiters(futex);
139 * unlock(hash_bucket(futex));
141 * Where (A) orders the waiters increment and the futex value read -- this
142 * is guaranteed by the head counter in the hb spinlock; and where (B)
143 * orders the write to futex and the waiters read -- this is done by the
144 * barriers in get_futex_key_refs(), through either ihold or atomic_inc,
145 * depending on the futex type.
147 * This yields the following case (where X:=waiters, Y:=futex):
155 * Which guarantees that x==0 && y==0 is impossible; which translates back into
156 * the guarantee that we cannot both miss the futex variable change and the
160 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
161 int __read_mostly futex_cmpxchg_enabled
;
165 * Futex flags used to encode options to functions and preserve them across
168 #define FLAGS_SHARED 0x01
169 #define FLAGS_CLOCKRT 0x02
170 #define FLAGS_HAS_TIMEOUT 0x04
173 * Priority Inheritance state:
175 struct futex_pi_state
{
177 * list of 'owned' pi_state instances - these have to be
178 * cleaned up in do_exit() if the task exits prematurely:
180 struct list_head list
;
185 struct rt_mutex pi_mutex
;
187 struct task_struct
*owner
;
194 * struct futex_q - The hashed futex queue entry, one per waiting task
195 * @list: priority-sorted list of tasks waiting on this futex
196 * @task: the task waiting on the futex
197 * @lock_ptr: the hash bucket lock
198 * @key: the key the futex is hashed on
199 * @pi_state: optional priority inheritance state
200 * @rt_waiter: rt_waiter storage for use with requeue_pi
201 * @requeue_pi_key: the requeue_pi target futex key
202 * @bitset: bitset for the optional bitmasked wakeup
204 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
205 * we can wake only the relevant ones (hashed queues may be shared).
207 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
208 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
209 * The order of wakeup is always to make the first condition true, then
212 * PI futexes are typically woken before they are removed from the hash list via
213 * the rt_mutex code. See unqueue_me_pi().
216 struct plist_node list
;
218 struct task_struct
*task
;
219 spinlock_t
*lock_ptr
;
221 struct futex_pi_state
*pi_state
;
222 struct rt_mutex_waiter
*rt_waiter
;
223 union futex_key
*requeue_pi_key
;
227 static const struct futex_q futex_q_init
= {
228 /* list gets initialized in queue_me()*/
229 .key
= FUTEX_KEY_INIT
,
230 .bitset
= FUTEX_BITSET_MATCH_ANY
234 * Hash buckets are shared by all the futex_keys that hash to the same
235 * location. Each key may have multiple futex_q structures, one for each task
236 * waiting on a futex.
238 struct futex_hash_bucket
{
241 struct plist_head chain
;
242 } ____cacheline_aligned_in_smp
;
244 static unsigned long __read_mostly futex_hashsize
;
246 static struct futex_hash_bucket
*futex_queues
;
248 static inline void futex_get_mm(union futex_key
*key
)
250 atomic_inc(&key
->private.mm
->mm_count
);
252 * Ensure futex_get_mm() implies a full barrier such that
253 * get_futex_key() implies a full barrier. This is relied upon
254 * as full barrier (B), see the ordering comment above.
256 smp_mb__after_atomic_inc();
260 * Reflects a new waiter being added to the waitqueue.
262 static inline void hb_waiters_inc(struct futex_hash_bucket
*hb
)
265 atomic_inc(&hb
->waiters
);
267 * Full barrier (A), see the ordering comment above.
269 smp_mb__after_atomic_inc();
274 * Reflects a waiter being removed from the waitqueue by wakeup
277 static inline void hb_waiters_dec(struct futex_hash_bucket
*hb
)
280 atomic_dec(&hb
->waiters
);
284 static inline int hb_waiters_pending(struct futex_hash_bucket
*hb
)
287 return atomic_read(&hb
->waiters
);
294 * We hash on the keys returned from get_futex_key (see below).
296 static struct futex_hash_bucket
*hash_futex(union futex_key
*key
)
298 u32 hash
= jhash2((u32
*)&key
->both
.word
,
299 (sizeof(key
->both
.word
)+sizeof(key
->both
.ptr
))/4,
301 return &futex_queues
[hash
& (futex_hashsize
- 1)];
305 * Return 1 if two futex_keys are equal, 0 otherwise.
307 static inline int match_futex(union futex_key
*key1
, union futex_key
*key2
)
310 && key1
->both
.word
== key2
->both
.word
311 && key1
->both
.ptr
== key2
->both
.ptr
312 && key1
->both
.offset
== key2
->both
.offset
);
316 * Take a reference to the resource addressed by a key.
317 * Can be called while holding spinlocks.
320 static void get_futex_key_refs(union futex_key
*key
)
325 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
327 ihold(key
->shared
.inode
); /* implies MB (B) */
329 case FUT_OFF_MMSHARED
:
330 futex_get_mm(key
); /* implies MB (B) */
336 * Drop a reference to the resource addressed by a key.
337 * The hash bucket spinlock must not be held.
339 static void drop_futex_key_refs(union futex_key
*key
)
341 if (!key
->both
.ptr
) {
342 /* If we're here then we tried to put a key we failed to get */
347 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
349 iput(key
->shared
.inode
);
351 case FUT_OFF_MMSHARED
:
352 mmdrop(key
->private.mm
);
358 * get_futex_key() - Get parameters which are the keys for a futex
359 * @uaddr: virtual address of the futex
360 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
361 * @key: address where result is stored.
362 * @rw: mapping needs to be read/write (values: VERIFY_READ,
365 * Return: a negative error code or 0
367 * The key words are stored in *key on success.
369 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
370 * offset_within_page). For private mappings, it's (uaddr, current->mm).
371 * We can usually work out the index without swapping in the page.
373 * lock_page() might sleep, the caller should not hold a spinlock.
376 get_futex_key(u32 __user
*uaddr
, int fshared
, union futex_key
*key
, int rw
)
378 unsigned long address
= (unsigned long)uaddr
;
379 struct mm_struct
*mm
= current
->mm
;
380 struct page
*page
, *page_head
;
384 * The futex address must be "naturally" aligned.
386 key
->both
.offset
= address
% PAGE_SIZE
;
387 if (unlikely((address
% sizeof(u32
)) != 0))
389 address
-= key
->both
.offset
;
391 if (unlikely(!access_ok(rw
, uaddr
, sizeof(u32
))))
395 * PROCESS_PRIVATE futexes are fast.
396 * As the mm cannot disappear under us and the 'key' only needs
397 * virtual address, we dont even have to find the underlying vma.
398 * Note : We do have to check 'uaddr' is a valid user address,
399 * but access_ok() should be faster than find_vma()
402 key
->private.mm
= mm
;
403 key
->private.address
= address
;
404 get_futex_key_refs(key
); /* implies MB (B) */
409 err
= get_user_pages_fast(address
, 1, 1, &page
);
411 * If write access is not required (eg. FUTEX_WAIT), try
412 * and get read-only access.
414 if (err
== -EFAULT
&& rw
== VERIFY_READ
) {
415 err
= get_user_pages_fast(address
, 1, 0, &page
);
423 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
425 if (unlikely(PageTail(page
))) {
427 /* serialize against __split_huge_page_splitting() */
429 if (likely(__get_user_pages_fast(address
, 1, !ro
, &page
) == 1)) {
430 page_head
= compound_head(page
);
432 * page_head is valid pointer but we must pin
433 * it before taking the PG_lock and/or
434 * PG_compound_lock. The moment we re-enable
435 * irqs __split_huge_page_splitting() can
436 * return and the head page can be freed from
437 * under us. We can't take the PG_lock and/or
438 * PG_compound_lock on a page that could be
439 * freed from under us.
441 if (page
!= page_head
) {
452 page_head
= compound_head(page
);
453 if (page
!= page_head
) {
459 lock_page(page_head
);
462 * If page_head->mapping is NULL, then it cannot be a PageAnon
463 * page; but it might be the ZERO_PAGE or in the gate area or
464 * in a special mapping (all cases which we are happy to fail);
465 * or it may have been a good file page when get_user_pages_fast
466 * found it, but truncated or holepunched or subjected to
467 * invalidate_complete_page2 before we got the page lock (also
468 * cases which we are happy to fail). And we hold a reference,
469 * so refcount care in invalidate_complete_page's remove_mapping
470 * prevents drop_caches from setting mapping to NULL beneath us.
472 * The case we do have to guard against is when memory pressure made
473 * shmem_writepage move it from filecache to swapcache beneath us:
474 * an unlikely race, but we do need to retry for page_head->mapping.
476 if (!page_head
->mapping
) {
477 int shmem_swizzled
= PageSwapCache(page_head
);
478 unlock_page(page_head
);
486 * Private mappings are handled in a simple way.
488 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
489 * it's a read-only handle, it's expected that futexes attach to
490 * the object not the particular process.
492 if (PageAnon(page_head
)) {
494 * A RO anonymous page will never change and thus doesn't make
495 * sense for futex operations.
502 key
->both
.offset
|= FUT_OFF_MMSHARED
; /* ref taken on mm */
503 key
->private.mm
= mm
;
504 key
->private.address
= address
;
506 key
->both
.offset
|= FUT_OFF_INODE
; /* inode-based key */
507 key
->shared
.inode
= page_head
->mapping
->host
;
508 key
->shared
.pgoff
= basepage_index(page
);
511 get_futex_key_refs(key
); /* implies MB (B) */
514 unlock_page(page_head
);
519 static inline void put_futex_key(union futex_key
*key
)
521 drop_futex_key_refs(key
);
525 * fault_in_user_writeable() - Fault in user address and verify RW access
526 * @uaddr: pointer to faulting user space address
528 * Slow path to fixup the fault we just took in the atomic write
531 * We have no generic implementation of a non-destructive write to the
532 * user address. We know that we faulted in the atomic pagefault
533 * disabled section so we can as well avoid the #PF overhead by
534 * calling get_user_pages() right away.
536 static int fault_in_user_writeable(u32 __user
*uaddr
)
538 struct mm_struct
*mm
= current
->mm
;
541 down_read(&mm
->mmap_sem
);
542 ret
= fixup_user_fault(current
, mm
, (unsigned long)uaddr
,
544 up_read(&mm
->mmap_sem
);
546 return ret
< 0 ? ret
: 0;
550 * futex_top_waiter() - Return the highest priority waiter on a futex
551 * @hb: the hash bucket the futex_q's reside in
552 * @key: the futex key (to distinguish it from other futex futex_q's)
554 * Must be called with the hb lock held.
556 static struct futex_q
*futex_top_waiter(struct futex_hash_bucket
*hb
,
557 union futex_key
*key
)
559 struct futex_q
*this;
561 plist_for_each_entry(this, &hb
->chain
, list
) {
562 if (match_futex(&this->key
, key
))
568 static int cmpxchg_futex_value_locked(u32
*curval
, u32 __user
*uaddr
,
569 u32 uval
, u32 newval
)
574 ret
= futex_atomic_cmpxchg_inatomic(curval
, uaddr
, uval
, newval
);
580 static int get_futex_value_locked(u32
*dest
, u32 __user
*from
)
585 ret
= __copy_from_user_inatomic(dest
, from
, sizeof(u32
));
588 return ret
? -EFAULT
: 0;
595 static int refill_pi_state_cache(void)
597 struct futex_pi_state
*pi_state
;
599 if (likely(current
->pi_state_cache
))
602 pi_state
= kzalloc(sizeof(*pi_state
), GFP_KERNEL
);
607 INIT_LIST_HEAD(&pi_state
->list
);
608 /* pi_mutex gets initialized later */
609 pi_state
->owner
= NULL
;
610 atomic_set(&pi_state
->refcount
, 1);
611 pi_state
->key
= FUTEX_KEY_INIT
;
613 current
->pi_state_cache
= pi_state
;
618 static struct futex_pi_state
* alloc_pi_state(void)
620 struct futex_pi_state
*pi_state
= current
->pi_state_cache
;
623 current
->pi_state_cache
= NULL
;
628 static void free_pi_state(struct futex_pi_state
*pi_state
)
630 if (!atomic_dec_and_test(&pi_state
->refcount
))
634 * If pi_state->owner is NULL, the owner is most probably dying
635 * and has cleaned up the pi_state already
637 if (pi_state
->owner
) {
638 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
639 list_del_init(&pi_state
->list
);
640 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
642 rt_mutex_proxy_unlock(&pi_state
->pi_mutex
, pi_state
->owner
);
645 if (current
->pi_state_cache
)
649 * pi_state->list is already empty.
650 * clear pi_state->owner.
651 * refcount is at 0 - put it back to 1.
653 pi_state
->owner
= NULL
;
654 atomic_set(&pi_state
->refcount
, 1);
655 current
->pi_state_cache
= pi_state
;
660 * Look up the task based on what TID userspace gave us.
663 static struct task_struct
* futex_find_get_task(pid_t pid
)
665 struct task_struct
*p
;
668 p
= find_task_by_vpid(pid
);
678 * This task is holding PI mutexes at exit time => bad.
679 * Kernel cleans up PI-state, but userspace is likely hosed.
680 * (Robust-futex cleanup is separate and might save the day for userspace.)
682 void exit_pi_state_list(struct task_struct
*curr
)
684 struct list_head
*next
, *head
= &curr
->pi_state_list
;
685 struct futex_pi_state
*pi_state
;
686 struct futex_hash_bucket
*hb
;
687 union futex_key key
= FUTEX_KEY_INIT
;
689 if (!futex_cmpxchg_enabled
)
692 * We are a ZOMBIE and nobody can enqueue itself on
693 * pi_state_list anymore, but we have to be careful
694 * versus waiters unqueueing themselves:
696 raw_spin_lock_irq(&curr
->pi_lock
);
697 while (!list_empty(head
)) {
700 pi_state
= list_entry(next
, struct futex_pi_state
, list
);
702 hb
= hash_futex(&key
);
703 raw_spin_unlock_irq(&curr
->pi_lock
);
705 spin_lock(&hb
->lock
);
707 raw_spin_lock_irq(&curr
->pi_lock
);
709 * We dropped the pi-lock, so re-check whether this
710 * task still owns the PI-state:
712 if (head
->next
!= next
) {
713 spin_unlock(&hb
->lock
);
717 WARN_ON(pi_state
->owner
!= curr
);
718 WARN_ON(list_empty(&pi_state
->list
));
719 list_del_init(&pi_state
->list
);
720 pi_state
->owner
= NULL
;
721 raw_spin_unlock_irq(&curr
->pi_lock
);
723 rt_mutex_unlock(&pi_state
->pi_mutex
);
725 spin_unlock(&hb
->lock
);
727 raw_spin_lock_irq(&curr
->pi_lock
);
729 raw_spin_unlock_irq(&curr
->pi_lock
);
733 lookup_pi_state(u32 uval
, struct futex_hash_bucket
*hb
,
734 union futex_key
*key
, struct futex_pi_state
**ps
)
736 struct futex_pi_state
*pi_state
= NULL
;
737 struct futex_q
*this, *next
;
738 struct task_struct
*p
;
739 pid_t pid
= uval
& FUTEX_TID_MASK
;
741 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
742 if (match_futex(&this->key
, key
)) {
744 * Another waiter already exists - bump up
745 * the refcount and return its pi_state:
747 pi_state
= this->pi_state
;
749 * Userspace might have messed up non-PI and PI futexes
751 if (unlikely(!pi_state
))
754 WARN_ON(!atomic_read(&pi_state
->refcount
));
757 * When pi_state->owner is NULL then the owner died
758 * and another waiter is on the fly. pi_state->owner
759 * is fixed up by the task which acquires
760 * pi_state->rt_mutex.
762 * We do not check for pid == 0 which can happen when
763 * the owner died and robust_list_exit() cleared the
766 if (pid
&& pi_state
->owner
) {
768 * Bail out if user space manipulated the
771 if (pid
!= task_pid_vnr(pi_state
->owner
))
775 atomic_inc(&pi_state
->refcount
);
783 * We are the first waiter - try to look up the real owner and attach
784 * the new pi_state to it, but bail out when TID = 0
788 p
= futex_find_get_task(pid
);
793 * We need to look at the task state flags to figure out,
794 * whether the task is exiting. To protect against the do_exit
795 * change of the task flags, we do this protected by
798 raw_spin_lock_irq(&p
->pi_lock
);
799 if (unlikely(p
->flags
& PF_EXITING
)) {
801 * The task is on the way out. When PF_EXITPIDONE is
802 * set, we know that the task has finished the
805 int ret
= (p
->flags
& PF_EXITPIDONE
) ? -ESRCH
: -EAGAIN
;
807 raw_spin_unlock_irq(&p
->pi_lock
);
812 pi_state
= alloc_pi_state();
815 * Initialize the pi_mutex in locked state and make 'p'
818 rt_mutex_init_proxy_locked(&pi_state
->pi_mutex
, p
);
820 /* Store the key for possible exit cleanups: */
821 pi_state
->key
= *key
;
823 WARN_ON(!list_empty(&pi_state
->list
));
824 list_add(&pi_state
->list
, &p
->pi_state_list
);
826 raw_spin_unlock_irq(&p
->pi_lock
);
836 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
837 * @uaddr: the pi futex user address
838 * @hb: the pi futex hash bucket
839 * @key: the futex key associated with uaddr and hb
840 * @ps: the pi_state pointer where we store the result of the
842 * @task: the task to perform the atomic lock work for. This will
843 * be "current" except in the case of requeue pi.
844 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
848 * 1 - acquired the lock;
851 * The hb->lock and futex_key refs shall be held by the caller.
853 static int futex_lock_pi_atomic(u32 __user
*uaddr
, struct futex_hash_bucket
*hb
,
854 union futex_key
*key
,
855 struct futex_pi_state
**ps
,
856 struct task_struct
*task
, int set_waiters
)
858 int lock_taken
, ret
, force_take
= 0;
859 u32 uval
, newval
, curval
, vpid
= task_pid_vnr(task
);
862 ret
= lock_taken
= 0;
865 * To avoid races, we attempt to take the lock here again
866 * (by doing a 0 -> TID atomic cmpxchg), while holding all
867 * the locks. It will most likely not succeed.
871 newval
|= FUTEX_WAITERS
;
873 if (unlikely(cmpxchg_futex_value_locked(&curval
, uaddr
, 0, newval
)))
879 if ((unlikely((curval
& FUTEX_TID_MASK
) == vpid
)))
883 * Surprise - we got the lock. Just return to userspace:
885 if (unlikely(!curval
))
891 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
892 * to wake at the next unlock.
894 newval
= curval
| FUTEX_WAITERS
;
897 * Should we force take the futex? See below.
899 if (unlikely(force_take
)) {
901 * Keep the OWNER_DIED and the WAITERS bit and set the
904 newval
= (curval
& ~FUTEX_TID_MASK
) | vpid
;
909 if (unlikely(cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
)))
911 if (unlikely(curval
!= uval
))
915 * We took the lock due to forced take over.
917 if (unlikely(lock_taken
))
921 * We dont have the lock. Look up the PI state (or create it if
922 * we are the first waiter):
924 ret
= lookup_pi_state(uval
, hb
, key
, ps
);
930 * We failed to find an owner for this
931 * futex. So we have no pi_state to block
932 * on. This can happen in two cases:
935 * 2) A stale FUTEX_WAITERS bit
937 * Re-read the futex value.
939 if (get_futex_value_locked(&curval
, uaddr
))
943 * If the owner died or we have a stale
944 * WAITERS bit the owner TID in the user space
947 if (!(curval
& FUTEX_TID_MASK
)) {
960 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
961 * @q: The futex_q to unqueue
963 * The q->lock_ptr must not be NULL and must be held by the caller.
965 static void __unqueue_futex(struct futex_q
*q
)
967 struct futex_hash_bucket
*hb
;
969 if (WARN_ON_SMP(!q
->lock_ptr
|| !spin_is_locked(q
->lock_ptr
))
970 || WARN_ON(plist_node_empty(&q
->list
)))
973 hb
= container_of(q
->lock_ptr
, struct futex_hash_bucket
, lock
);
974 plist_del(&q
->list
, &hb
->chain
);
979 * The hash bucket lock must be held when this is called.
980 * Afterwards, the futex_q must not be accessed.
982 static void wake_futex(struct futex_q
*q
)
984 struct task_struct
*p
= q
->task
;
986 if (WARN(q
->pi_state
|| q
->rt_waiter
, "refusing to wake PI futex\n"))
990 * We set q->lock_ptr = NULL _before_ we wake up the task. If
991 * a non-futex wake up happens on another CPU then the task
992 * might exit and p would dereference a non-existing task
993 * struct. Prevent this by holding a reference on p across the
1000 * The waiting task can free the futex_q as soon as
1001 * q->lock_ptr = NULL is written, without taking any locks. A
1002 * memory barrier is required here to prevent the following
1003 * store to lock_ptr from getting ahead of the plist_del.
1008 wake_up_state(p
, TASK_NORMAL
);
1012 static int wake_futex_pi(u32 __user
*uaddr
, u32 uval
, struct futex_q
*this)
1014 struct task_struct
*new_owner
;
1015 struct futex_pi_state
*pi_state
= this->pi_state
;
1016 u32
uninitialized_var(curval
), newval
;
1022 * If current does not own the pi_state then the futex is
1023 * inconsistent and user space fiddled with the futex value.
1025 if (pi_state
->owner
!= current
)
1028 raw_spin_lock(&pi_state
->pi_mutex
.wait_lock
);
1029 new_owner
= rt_mutex_next_owner(&pi_state
->pi_mutex
);
1032 * It is possible that the next waiter (the one that brought
1033 * this owner to the kernel) timed out and is no longer
1034 * waiting on the lock.
1037 new_owner
= this->task
;
1040 * We pass it to the next owner. (The WAITERS bit is always
1041 * kept enabled while there is PI state around. We must also
1042 * preserve the owner died bit.)
1044 if (!(uval
& FUTEX_OWNER_DIED
)) {
1047 newval
= FUTEX_WAITERS
| task_pid_vnr(new_owner
);
1049 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
))
1051 else if (curval
!= uval
)
1054 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
1059 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
1060 WARN_ON(list_empty(&pi_state
->list
));
1061 list_del_init(&pi_state
->list
);
1062 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
1064 raw_spin_lock_irq(&new_owner
->pi_lock
);
1065 WARN_ON(!list_empty(&pi_state
->list
));
1066 list_add(&pi_state
->list
, &new_owner
->pi_state_list
);
1067 pi_state
->owner
= new_owner
;
1068 raw_spin_unlock_irq(&new_owner
->pi_lock
);
1070 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
1071 rt_mutex_unlock(&pi_state
->pi_mutex
);
1076 static int unlock_futex_pi(u32 __user
*uaddr
, u32 uval
)
1078 u32
uninitialized_var(oldval
);
1081 * There is no waiter, so we unlock the futex. The owner died
1082 * bit has not to be preserved here. We are the owner:
1084 if (cmpxchg_futex_value_locked(&oldval
, uaddr
, uval
, 0))
1093 * Express the locking dependencies for lockdep:
1096 double_lock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1099 spin_lock(&hb1
->lock
);
1101 spin_lock_nested(&hb2
->lock
, SINGLE_DEPTH_NESTING
);
1102 } else { /* hb1 > hb2 */
1103 spin_lock(&hb2
->lock
);
1104 spin_lock_nested(&hb1
->lock
, SINGLE_DEPTH_NESTING
);
1109 double_unlock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1111 spin_unlock(&hb1
->lock
);
1113 spin_unlock(&hb2
->lock
);
1117 * Wake up waiters matching bitset queued on this futex (uaddr).
1120 futex_wake(u32 __user
*uaddr
, unsigned int flags
, int nr_wake
, u32 bitset
)
1122 struct futex_hash_bucket
*hb
;
1123 struct futex_q
*this, *next
;
1124 union futex_key key
= FUTEX_KEY_INIT
;
1130 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_READ
);
1131 if (unlikely(ret
!= 0))
1134 hb
= hash_futex(&key
);
1136 /* Make sure we really have tasks to wakeup */
1137 if (!hb_waiters_pending(hb
))
1140 spin_lock(&hb
->lock
);
1142 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
1143 if (match_futex (&this->key
, &key
)) {
1144 if (this->pi_state
|| this->rt_waiter
) {
1149 /* Check if one of the bits is set in both bitsets */
1150 if (!(this->bitset
& bitset
))
1154 if (++ret
>= nr_wake
)
1159 spin_unlock(&hb
->lock
);
1161 put_futex_key(&key
);
1167 * Wake up all waiters hashed on the physical page that is mapped
1168 * to this virtual address:
1171 futex_wake_op(u32 __user
*uaddr1
, unsigned int flags
, u32 __user
*uaddr2
,
1172 int nr_wake
, int nr_wake2
, int op
)
1174 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1175 struct futex_hash_bucket
*hb1
, *hb2
;
1176 struct futex_q
*this, *next
;
1180 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1181 if (unlikely(ret
!= 0))
1183 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
1184 if (unlikely(ret
!= 0))
1187 hb1
= hash_futex(&key1
);
1188 hb2
= hash_futex(&key2
);
1191 double_lock_hb(hb1
, hb2
);
1192 op_ret
= futex_atomic_op_inuser(op
, uaddr2
);
1193 if (unlikely(op_ret
< 0)) {
1195 double_unlock_hb(hb1
, hb2
);
1199 * we don't get EFAULT from MMU faults if we don't have an MMU,
1200 * but we might get them from range checking
1206 if (unlikely(op_ret
!= -EFAULT
)) {
1211 ret
= fault_in_user_writeable(uaddr2
);
1215 if (!(flags
& FLAGS_SHARED
))
1218 put_futex_key(&key2
);
1219 put_futex_key(&key1
);
1223 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1224 if (match_futex (&this->key
, &key1
)) {
1225 if (this->pi_state
|| this->rt_waiter
) {
1230 if (++ret
>= nr_wake
)
1237 plist_for_each_entry_safe(this, next
, &hb2
->chain
, list
) {
1238 if (match_futex (&this->key
, &key2
)) {
1239 if (this->pi_state
|| this->rt_waiter
) {
1244 if (++op_ret
>= nr_wake2
)
1252 double_unlock_hb(hb1
, hb2
);
1254 put_futex_key(&key2
);
1256 put_futex_key(&key1
);
1262 * requeue_futex() - Requeue a futex_q from one hb to another
1263 * @q: the futex_q to requeue
1264 * @hb1: the source hash_bucket
1265 * @hb2: the target hash_bucket
1266 * @key2: the new key for the requeued futex_q
1269 void requeue_futex(struct futex_q
*q
, struct futex_hash_bucket
*hb1
,
1270 struct futex_hash_bucket
*hb2
, union futex_key
*key2
)
1274 * If key1 and key2 hash to the same bucket, no need to
1277 if (likely(&hb1
->chain
!= &hb2
->chain
)) {
1278 plist_del(&q
->list
, &hb1
->chain
);
1279 hb_waiters_dec(hb1
);
1280 plist_add(&q
->list
, &hb2
->chain
);
1281 hb_waiters_inc(hb2
);
1282 q
->lock_ptr
= &hb2
->lock
;
1284 get_futex_key_refs(key2
);
1289 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1291 * @key: the key of the requeue target futex
1292 * @hb: the hash_bucket of the requeue target futex
1294 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1295 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1296 * to the requeue target futex so the waiter can detect the wakeup on the right
1297 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1298 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1299 * to protect access to the pi_state to fixup the owner later. Must be called
1300 * with both q->lock_ptr and hb->lock held.
1303 void requeue_pi_wake_futex(struct futex_q
*q
, union futex_key
*key
,
1304 struct futex_hash_bucket
*hb
)
1306 get_futex_key_refs(key
);
1311 WARN_ON(!q
->rt_waiter
);
1312 q
->rt_waiter
= NULL
;
1314 q
->lock_ptr
= &hb
->lock
;
1316 wake_up_state(q
->task
, TASK_NORMAL
);
1320 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1321 * @pifutex: the user address of the to futex
1322 * @hb1: the from futex hash bucket, must be locked by the caller
1323 * @hb2: the to futex hash bucket, must be locked by the caller
1324 * @key1: the from futex key
1325 * @key2: the to futex key
1326 * @ps: address to store the pi_state pointer
1327 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1329 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1330 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1331 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1332 * hb1 and hb2 must be held by the caller.
1335 * 0 - failed to acquire the lock atomically;
1336 * 1 - acquired the lock;
1339 static int futex_proxy_trylock_atomic(u32 __user
*pifutex
,
1340 struct futex_hash_bucket
*hb1
,
1341 struct futex_hash_bucket
*hb2
,
1342 union futex_key
*key1
, union futex_key
*key2
,
1343 struct futex_pi_state
**ps
, int set_waiters
)
1345 struct futex_q
*top_waiter
= NULL
;
1349 if (get_futex_value_locked(&curval
, pifutex
))
1353 * Find the top_waiter and determine if there are additional waiters.
1354 * If the caller intends to requeue more than 1 waiter to pifutex,
1355 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1356 * as we have means to handle the possible fault. If not, don't set
1357 * the bit unecessarily as it will force the subsequent unlock to enter
1360 top_waiter
= futex_top_waiter(hb1
, key1
);
1362 /* There are no waiters, nothing for us to do. */
1366 /* Ensure we requeue to the expected futex. */
1367 if (!match_futex(top_waiter
->requeue_pi_key
, key2
))
1371 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1372 * the contended case or if set_waiters is 1. The pi_state is returned
1373 * in ps in contended cases.
1375 ret
= futex_lock_pi_atomic(pifutex
, hb2
, key2
, ps
, top_waiter
->task
,
1378 requeue_pi_wake_futex(top_waiter
, key2
, hb2
);
1384 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1385 * @uaddr1: source futex user address
1386 * @flags: futex flags (FLAGS_SHARED, etc.)
1387 * @uaddr2: target futex user address
1388 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1389 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1390 * @cmpval: @uaddr1 expected value (or %NULL)
1391 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1392 * pi futex (pi to pi requeue is not supported)
1394 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1395 * uaddr2 atomically on behalf of the top waiter.
1398 * >=0 - on success, the number of tasks requeued or woken;
1401 static int futex_requeue(u32 __user
*uaddr1
, unsigned int flags
,
1402 u32 __user
*uaddr2
, int nr_wake
, int nr_requeue
,
1403 u32
*cmpval
, int requeue_pi
)
1405 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1406 int drop_count
= 0, task_count
= 0, ret
;
1407 struct futex_pi_state
*pi_state
= NULL
;
1408 struct futex_hash_bucket
*hb1
, *hb2
;
1409 struct futex_q
*this, *next
;
1414 * requeue_pi requires a pi_state, try to allocate it now
1415 * without any locks in case it fails.
1417 if (refill_pi_state_cache())
1420 * requeue_pi must wake as many tasks as it can, up to nr_wake
1421 * + nr_requeue, since it acquires the rt_mutex prior to
1422 * returning to userspace, so as to not leave the rt_mutex with
1423 * waiters and no owner. However, second and third wake-ups
1424 * cannot be predicted as they involve race conditions with the
1425 * first wake and a fault while looking up the pi_state. Both
1426 * pthread_cond_signal() and pthread_cond_broadcast() should
1434 if (pi_state
!= NULL
) {
1436 * We will have to lookup the pi_state again, so free this one
1437 * to keep the accounting correct.
1439 free_pi_state(pi_state
);
1443 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1444 if (unlikely(ret
!= 0))
1446 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
,
1447 requeue_pi
? VERIFY_WRITE
: VERIFY_READ
);
1448 if (unlikely(ret
!= 0))
1451 hb1
= hash_futex(&key1
);
1452 hb2
= hash_futex(&key2
);
1455 hb_waiters_inc(hb2
);
1456 double_lock_hb(hb1
, hb2
);
1458 if (likely(cmpval
!= NULL
)) {
1461 ret
= get_futex_value_locked(&curval
, uaddr1
);
1463 if (unlikely(ret
)) {
1464 double_unlock_hb(hb1
, hb2
);
1465 hb_waiters_dec(hb2
);
1467 ret
= get_user(curval
, uaddr1
);
1471 if (!(flags
& FLAGS_SHARED
))
1474 put_futex_key(&key2
);
1475 put_futex_key(&key1
);
1478 if (curval
!= *cmpval
) {
1484 if (requeue_pi
&& (task_count
- nr_wake
< nr_requeue
)) {
1486 * Attempt to acquire uaddr2 and wake the top waiter. If we
1487 * intend to requeue waiters, force setting the FUTEX_WAITERS
1488 * bit. We force this here where we are able to easily handle
1489 * faults rather in the requeue loop below.
1491 ret
= futex_proxy_trylock_atomic(uaddr2
, hb1
, hb2
, &key1
,
1492 &key2
, &pi_state
, nr_requeue
);
1495 * At this point the top_waiter has either taken uaddr2 or is
1496 * waiting on it. If the former, then the pi_state will not
1497 * exist yet, look it up one more time to ensure we have a
1504 ret
= get_futex_value_locked(&curval2
, uaddr2
);
1506 ret
= lookup_pi_state(curval2
, hb2
, &key2
,
1514 double_unlock_hb(hb1
, hb2
);
1515 hb_waiters_dec(hb2
);
1516 put_futex_key(&key2
);
1517 put_futex_key(&key1
);
1518 ret
= fault_in_user_writeable(uaddr2
);
1523 /* The owner was exiting, try again. */
1524 double_unlock_hb(hb1
, hb2
);
1525 hb_waiters_dec(hb2
);
1526 put_futex_key(&key2
);
1527 put_futex_key(&key1
);
1535 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1536 if (task_count
- nr_wake
>= nr_requeue
)
1539 if (!match_futex(&this->key
, &key1
))
1543 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1544 * be paired with each other and no other futex ops.
1546 * We should never be requeueing a futex_q with a pi_state,
1547 * which is awaiting a futex_unlock_pi().
1549 if ((requeue_pi
&& !this->rt_waiter
) ||
1550 (!requeue_pi
&& this->rt_waiter
) ||
1557 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1558 * lock, we already woke the top_waiter. If not, it will be
1559 * woken by futex_unlock_pi().
1561 if (++task_count
<= nr_wake
&& !requeue_pi
) {
1566 /* Ensure we requeue to the expected futex for requeue_pi. */
1567 if (requeue_pi
&& !match_futex(this->requeue_pi_key
, &key2
)) {
1573 * Requeue nr_requeue waiters and possibly one more in the case
1574 * of requeue_pi if we couldn't acquire the lock atomically.
1577 /* Prepare the waiter to take the rt_mutex. */
1578 atomic_inc(&pi_state
->refcount
);
1579 this->pi_state
= pi_state
;
1580 ret
= rt_mutex_start_proxy_lock(&pi_state
->pi_mutex
,
1584 /* We got the lock. */
1585 requeue_pi_wake_futex(this, &key2
, hb2
);
1590 this->pi_state
= NULL
;
1591 free_pi_state(pi_state
);
1595 requeue_futex(this, hb1
, hb2
, &key2
);
1600 double_unlock_hb(hb1
, hb2
);
1601 hb_waiters_dec(hb2
);
1604 * drop_futex_key_refs() must be called outside the spinlocks. During
1605 * the requeue we moved futex_q's from the hash bucket at key1 to the
1606 * one at key2 and updated their key pointer. We no longer need to
1607 * hold the references to key1.
1609 while (--drop_count
>= 0)
1610 drop_futex_key_refs(&key1
);
1613 put_futex_key(&key2
);
1615 put_futex_key(&key1
);
1617 if (pi_state
!= NULL
)
1618 free_pi_state(pi_state
);
1619 return ret
? ret
: task_count
;
1622 /* The key must be already stored in q->key. */
1623 static inline struct futex_hash_bucket
*queue_lock(struct futex_q
*q
)
1624 __acquires(&hb
->lock
)
1626 struct futex_hash_bucket
*hb
;
1628 hb
= hash_futex(&q
->key
);
1631 * Increment the counter before taking the lock so that
1632 * a potential waker won't miss a to-be-slept task that is
1633 * waiting for the spinlock. This is safe as all queue_lock()
1634 * users end up calling queue_me(). Similarly, for housekeeping,
1635 * decrement the counter at queue_unlock() when some error has
1636 * occurred and we don't end up adding the task to the list.
1640 q
->lock_ptr
= &hb
->lock
;
1642 spin_lock(&hb
->lock
); /* implies MB (A) */
1647 queue_unlock(struct futex_hash_bucket
*hb
)
1648 __releases(&hb
->lock
)
1650 spin_unlock(&hb
->lock
);
1655 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1656 * @q: The futex_q to enqueue
1657 * @hb: The destination hash bucket
1659 * The hb->lock must be held by the caller, and is released here. A call to
1660 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1661 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1662 * or nothing if the unqueue is done as part of the wake process and the unqueue
1663 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1666 static inline void queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
1667 __releases(&hb
->lock
)
1672 * The priority used to register this element is
1673 * - either the real thread-priority for the real-time threads
1674 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1675 * - or MAX_RT_PRIO for non-RT threads.
1676 * Thus, all RT-threads are woken first in priority order, and
1677 * the others are woken last, in FIFO order.
1679 prio
= min(current
->normal_prio
, MAX_RT_PRIO
);
1681 plist_node_init(&q
->list
, prio
);
1682 plist_add(&q
->list
, &hb
->chain
);
1684 spin_unlock(&hb
->lock
);
1688 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1689 * @q: The futex_q to unqueue
1691 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1692 * be paired with exactly one earlier call to queue_me().
1695 * 1 - if the futex_q was still queued (and we removed unqueued it);
1696 * 0 - if the futex_q was already removed by the waking thread
1698 static int unqueue_me(struct futex_q
*q
)
1700 spinlock_t
*lock_ptr
;
1703 /* In the common case we don't take the spinlock, which is nice. */
1705 lock_ptr
= q
->lock_ptr
;
1707 if (lock_ptr
!= NULL
) {
1708 spin_lock(lock_ptr
);
1710 * q->lock_ptr can change between reading it and
1711 * spin_lock(), causing us to take the wrong lock. This
1712 * corrects the race condition.
1714 * Reasoning goes like this: if we have the wrong lock,
1715 * q->lock_ptr must have changed (maybe several times)
1716 * between reading it and the spin_lock(). It can
1717 * change again after the spin_lock() but only if it was
1718 * already changed before the spin_lock(). It cannot,
1719 * however, change back to the original value. Therefore
1720 * we can detect whether we acquired the correct lock.
1722 if (unlikely(lock_ptr
!= q
->lock_ptr
)) {
1723 spin_unlock(lock_ptr
);
1728 BUG_ON(q
->pi_state
);
1730 spin_unlock(lock_ptr
);
1734 drop_futex_key_refs(&q
->key
);
1739 * PI futexes can not be requeued and must remove themself from the
1740 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1743 static void unqueue_me_pi(struct futex_q
*q
)
1744 __releases(q
->lock_ptr
)
1748 BUG_ON(!q
->pi_state
);
1749 free_pi_state(q
->pi_state
);
1752 spin_unlock(q
->lock_ptr
);
1756 * Fixup the pi_state owner with the new owner.
1758 * Must be called with hash bucket lock held and mm->sem held for non
1761 static int fixup_pi_state_owner(u32 __user
*uaddr
, struct futex_q
*q
,
1762 struct task_struct
*newowner
)
1764 u32 newtid
= task_pid_vnr(newowner
) | FUTEX_WAITERS
;
1765 struct futex_pi_state
*pi_state
= q
->pi_state
;
1766 struct task_struct
*oldowner
= pi_state
->owner
;
1767 u32 uval
, uninitialized_var(curval
), newval
;
1771 if (!pi_state
->owner
)
1772 newtid
|= FUTEX_OWNER_DIED
;
1775 * We are here either because we stole the rtmutex from the
1776 * previous highest priority waiter or we are the highest priority
1777 * waiter but failed to get the rtmutex the first time.
1778 * We have to replace the newowner TID in the user space variable.
1779 * This must be atomic as we have to preserve the owner died bit here.
1781 * Note: We write the user space value _before_ changing the pi_state
1782 * because we can fault here. Imagine swapped out pages or a fork
1783 * that marked all the anonymous memory readonly for cow.
1785 * Modifying pi_state _before_ the user space value would
1786 * leave the pi_state in an inconsistent state when we fault
1787 * here, because we need to drop the hash bucket lock to
1788 * handle the fault. This might be observed in the PID check
1789 * in lookup_pi_state.
1792 if (get_futex_value_locked(&uval
, uaddr
))
1796 newval
= (uval
& FUTEX_OWNER_DIED
) | newtid
;
1798 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
))
1806 * We fixed up user space. Now we need to fix the pi_state
1809 if (pi_state
->owner
!= NULL
) {
1810 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
1811 WARN_ON(list_empty(&pi_state
->list
));
1812 list_del_init(&pi_state
->list
);
1813 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
1816 pi_state
->owner
= newowner
;
1818 raw_spin_lock_irq(&newowner
->pi_lock
);
1819 WARN_ON(!list_empty(&pi_state
->list
));
1820 list_add(&pi_state
->list
, &newowner
->pi_state_list
);
1821 raw_spin_unlock_irq(&newowner
->pi_lock
);
1825 * To handle the page fault we need to drop the hash bucket
1826 * lock here. That gives the other task (either the highest priority
1827 * waiter itself or the task which stole the rtmutex) the
1828 * chance to try the fixup of the pi_state. So once we are
1829 * back from handling the fault we need to check the pi_state
1830 * after reacquiring the hash bucket lock and before trying to
1831 * do another fixup. When the fixup has been done already we
1835 spin_unlock(q
->lock_ptr
);
1837 ret
= fault_in_user_writeable(uaddr
);
1839 spin_lock(q
->lock_ptr
);
1842 * Check if someone else fixed it for us:
1844 if (pi_state
->owner
!= oldowner
)
1853 static long futex_wait_restart(struct restart_block
*restart
);
1856 * fixup_owner() - Post lock pi_state and corner case management
1857 * @uaddr: user address of the futex
1858 * @q: futex_q (contains pi_state and access to the rt_mutex)
1859 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1861 * After attempting to lock an rt_mutex, this function is called to cleanup
1862 * the pi_state owner as well as handle race conditions that may allow us to
1863 * acquire the lock. Must be called with the hb lock held.
1866 * 1 - success, lock taken;
1867 * 0 - success, lock not taken;
1868 * <0 - on error (-EFAULT)
1870 static int fixup_owner(u32 __user
*uaddr
, struct futex_q
*q
, int locked
)
1872 struct task_struct
*owner
;
1877 * Got the lock. We might not be the anticipated owner if we
1878 * did a lock-steal - fix up the PI-state in that case:
1880 if (q
->pi_state
->owner
!= current
)
1881 ret
= fixup_pi_state_owner(uaddr
, q
, current
);
1886 * Catch the rare case, where the lock was released when we were on the
1887 * way back before we locked the hash bucket.
1889 if (q
->pi_state
->owner
== current
) {
1891 * Try to get the rt_mutex now. This might fail as some other
1892 * task acquired the rt_mutex after we removed ourself from the
1893 * rt_mutex waiters list.
1895 if (rt_mutex_trylock(&q
->pi_state
->pi_mutex
)) {
1901 * pi_state is incorrect, some other task did a lock steal and
1902 * we returned due to timeout or signal without taking the
1903 * rt_mutex. Too late.
1905 raw_spin_lock(&q
->pi_state
->pi_mutex
.wait_lock
);
1906 owner
= rt_mutex_owner(&q
->pi_state
->pi_mutex
);
1908 owner
= rt_mutex_next_owner(&q
->pi_state
->pi_mutex
);
1909 raw_spin_unlock(&q
->pi_state
->pi_mutex
.wait_lock
);
1910 ret
= fixup_pi_state_owner(uaddr
, q
, owner
);
1915 * Paranoia check. If we did not take the lock, then we should not be
1916 * the owner of the rt_mutex.
1918 if (rt_mutex_owner(&q
->pi_state
->pi_mutex
) == current
)
1919 printk(KERN_ERR
"fixup_owner: ret = %d pi-mutex: %p "
1920 "pi-state %p\n", ret
,
1921 q
->pi_state
->pi_mutex
.owner
,
1922 q
->pi_state
->owner
);
1925 return ret
? ret
: locked
;
1929 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1930 * @hb: the futex hash bucket, must be locked by the caller
1931 * @q: the futex_q to queue up on
1932 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1934 static void futex_wait_queue_me(struct futex_hash_bucket
*hb
, struct futex_q
*q
,
1935 struct hrtimer_sleeper
*timeout
)
1938 * The task state is guaranteed to be set before another task can
1939 * wake it. set_current_state() is implemented using set_mb() and
1940 * queue_me() calls spin_unlock() upon completion, both serializing
1941 * access to the hash list and forcing another memory barrier.
1943 set_current_state(TASK_INTERRUPTIBLE
);
1948 hrtimer_start_expires(&timeout
->timer
, HRTIMER_MODE_ABS
);
1949 if (!hrtimer_active(&timeout
->timer
))
1950 timeout
->task
= NULL
;
1954 * If we have been removed from the hash list, then another task
1955 * has tried to wake us, and we can skip the call to schedule().
1957 if (likely(!plist_node_empty(&q
->list
))) {
1959 * If the timer has already expired, current will already be
1960 * flagged for rescheduling. Only call schedule if there
1961 * is no timeout, or if it has yet to expire.
1963 if (!timeout
|| timeout
->task
)
1964 freezable_schedule();
1966 __set_current_state(TASK_RUNNING
);
1970 * futex_wait_setup() - Prepare to wait on a futex
1971 * @uaddr: the futex userspace address
1972 * @val: the expected value
1973 * @flags: futex flags (FLAGS_SHARED, etc.)
1974 * @q: the associated futex_q
1975 * @hb: storage for hash_bucket pointer to be returned to caller
1977 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1978 * compare it with the expected value. Handle atomic faults internally.
1979 * Return with the hb lock held and a q.key reference on success, and unlocked
1980 * with no q.key reference on failure.
1983 * 0 - uaddr contains val and hb has been locked;
1984 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
1986 static int futex_wait_setup(u32 __user
*uaddr
, u32 val
, unsigned int flags
,
1987 struct futex_q
*q
, struct futex_hash_bucket
**hb
)
1993 * Access the page AFTER the hash-bucket is locked.
1994 * Order is important:
1996 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1997 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1999 * The basic logical guarantee of a futex is that it blocks ONLY
2000 * if cond(var) is known to be true at the time of blocking, for
2001 * any cond. If we locked the hash-bucket after testing *uaddr, that
2002 * would open a race condition where we could block indefinitely with
2003 * cond(var) false, which would violate the guarantee.
2005 * On the other hand, we insert q and release the hash-bucket only
2006 * after testing *uaddr. This guarantees that futex_wait() will NOT
2007 * absorb a wakeup if *uaddr does not match the desired values
2008 * while the syscall executes.
2011 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
->key
, VERIFY_READ
);
2012 if (unlikely(ret
!= 0))
2016 *hb
= queue_lock(q
);
2018 ret
= get_futex_value_locked(&uval
, uaddr
);
2023 ret
= get_user(uval
, uaddr
);
2027 if (!(flags
& FLAGS_SHARED
))
2030 put_futex_key(&q
->key
);
2041 put_futex_key(&q
->key
);
2045 static int futex_wait(u32 __user
*uaddr
, unsigned int flags
, u32 val
,
2046 ktime_t
*abs_time
, u32 bitset
)
2048 struct hrtimer_sleeper timeout
, *to
= NULL
;
2049 struct restart_block
*restart
;
2050 struct futex_hash_bucket
*hb
;
2051 struct futex_q q
= futex_q_init
;
2061 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2062 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2064 hrtimer_init_sleeper(to
, current
);
2065 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2066 current
->timer_slack_ns
);
2071 * Prepare to wait on uaddr. On success, holds hb lock and increments
2074 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2078 /* queue_me and wait for wakeup, timeout, or a signal. */
2079 futex_wait_queue_me(hb
, &q
, to
);
2081 /* If we were woken (and unqueued), we succeeded, whatever. */
2083 /* unqueue_me() drops q.key ref */
2084 if (!unqueue_me(&q
))
2087 if (to
&& !to
->task
)
2091 * We expect signal_pending(current), but we might be the
2092 * victim of a spurious wakeup as well.
2094 if (!signal_pending(current
))
2101 restart
= ¤t_thread_info()->restart_block
;
2102 restart
->fn
= futex_wait_restart
;
2103 restart
->futex
.uaddr
= uaddr
;
2104 restart
->futex
.val
= val
;
2105 restart
->futex
.time
= abs_time
->tv64
;
2106 restart
->futex
.bitset
= bitset
;
2107 restart
->futex
.flags
= flags
| FLAGS_HAS_TIMEOUT
;
2109 ret
= -ERESTART_RESTARTBLOCK
;
2113 hrtimer_cancel(&to
->timer
);
2114 destroy_hrtimer_on_stack(&to
->timer
);
2120 static long futex_wait_restart(struct restart_block
*restart
)
2122 u32 __user
*uaddr
= restart
->futex
.uaddr
;
2123 ktime_t t
, *tp
= NULL
;
2125 if (restart
->futex
.flags
& FLAGS_HAS_TIMEOUT
) {
2126 t
.tv64
= restart
->futex
.time
;
2129 restart
->fn
= do_no_restart_syscall
;
2131 return (long)futex_wait(uaddr
, restart
->futex
.flags
,
2132 restart
->futex
.val
, tp
, restart
->futex
.bitset
);
2137 * Userspace tried a 0 -> TID atomic transition of the futex value
2138 * and failed. The kernel side here does the whole locking operation:
2139 * if there are waiters then it will block, it does PI, etc. (Due to
2140 * races the kernel might see a 0 value of the futex too.)
2142 static int futex_lock_pi(u32 __user
*uaddr
, unsigned int flags
, int detect
,
2143 ktime_t
*time
, int trylock
)
2145 struct hrtimer_sleeper timeout
, *to
= NULL
;
2146 struct futex_hash_bucket
*hb
;
2147 struct futex_q q
= futex_q_init
;
2150 if (refill_pi_state_cache())
2155 hrtimer_init_on_stack(&to
->timer
, CLOCK_REALTIME
,
2157 hrtimer_init_sleeper(to
, current
);
2158 hrtimer_set_expires(&to
->timer
, *time
);
2162 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
.key
, VERIFY_WRITE
);
2163 if (unlikely(ret
!= 0))
2167 hb
= queue_lock(&q
);
2169 ret
= futex_lock_pi_atomic(uaddr
, hb
, &q
.key
, &q
.pi_state
, current
, 0);
2170 if (unlikely(ret
)) {
2173 /* We got the lock. */
2175 goto out_unlock_put_key
;
2180 * Task is exiting and we just wait for the
2184 put_futex_key(&q
.key
);
2188 goto out_unlock_put_key
;
2193 * Only actually queue now that the atomic ops are done:
2197 WARN_ON(!q
.pi_state
);
2199 * Block on the PI mutex:
2202 ret
= rt_mutex_timed_lock(&q
.pi_state
->pi_mutex
, to
, 1);
2204 ret
= rt_mutex_trylock(&q
.pi_state
->pi_mutex
);
2205 /* Fixup the trylock return value: */
2206 ret
= ret
? 0 : -EWOULDBLOCK
;
2209 spin_lock(q
.lock_ptr
);
2211 * Fixup the pi_state owner and possibly acquire the lock if we
2214 res
= fixup_owner(uaddr
, &q
, !ret
);
2216 * If fixup_owner() returned an error, proprogate that. If it acquired
2217 * the lock, clear our -ETIMEDOUT or -EINTR.
2220 ret
= (res
< 0) ? res
: 0;
2223 * If fixup_owner() faulted and was unable to handle the fault, unlock
2224 * it and return the fault to userspace.
2226 if (ret
&& (rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
))
2227 rt_mutex_unlock(&q
.pi_state
->pi_mutex
);
2229 /* Unqueue and drop the lock */
2238 put_futex_key(&q
.key
);
2241 destroy_hrtimer_on_stack(&to
->timer
);
2242 return ret
!= -EINTR
? ret
: -ERESTARTNOINTR
;
2247 ret
= fault_in_user_writeable(uaddr
);
2251 if (!(flags
& FLAGS_SHARED
))
2254 put_futex_key(&q
.key
);
2259 * Userspace attempted a TID -> 0 atomic transition, and failed.
2260 * This is the in-kernel slowpath: we look up the PI state (if any),
2261 * and do the rt-mutex unlock.
2263 static int futex_unlock_pi(u32 __user
*uaddr
, unsigned int flags
)
2265 struct futex_hash_bucket
*hb
;
2266 struct futex_q
*this, *next
;
2267 union futex_key key
= FUTEX_KEY_INIT
;
2268 u32 uval
, vpid
= task_pid_vnr(current
);
2272 if (get_user(uval
, uaddr
))
2275 * We release only a lock we actually own:
2277 if ((uval
& FUTEX_TID_MASK
) != vpid
)
2280 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_WRITE
);
2281 if (unlikely(ret
!= 0))
2284 hb
= hash_futex(&key
);
2285 spin_lock(&hb
->lock
);
2288 * To avoid races, try to do the TID -> 0 atomic transition
2289 * again. If it succeeds then we can return without waking
2292 if (!(uval
& FUTEX_OWNER_DIED
) &&
2293 cmpxchg_futex_value_locked(&uval
, uaddr
, vpid
, 0))
2296 * Rare case: we managed to release the lock atomically,
2297 * no need to wake anyone else up:
2299 if (unlikely(uval
== vpid
))
2303 * Ok, other tasks may need to be woken up - check waiters
2304 * and do the wakeup if necessary:
2306 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
2307 if (!match_futex (&this->key
, &key
))
2309 ret
= wake_futex_pi(uaddr
, uval
, this);
2311 * The atomic access to the futex value
2312 * generated a pagefault, so retry the
2313 * user-access and the wakeup:
2320 * No waiters - kernel unlocks the futex:
2322 if (!(uval
& FUTEX_OWNER_DIED
)) {
2323 ret
= unlock_futex_pi(uaddr
, uval
);
2329 spin_unlock(&hb
->lock
);
2330 put_futex_key(&key
);
2336 spin_unlock(&hb
->lock
);
2337 put_futex_key(&key
);
2339 ret
= fault_in_user_writeable(uaddr
);
2347 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2348 * @hb: the hash_bucket futex_q was original enqueued on
2349 * @q: the futex_q woken while waiting to be requeued
2350 * @key2: the futex_key of the requeue target futex
2351 * @timeout: the timeout associated with the wait (NULL if none)
2353 * Detect if the task was woken on the initial futex as opposed to the requeue
2354 * target futex. If so, determine if it was a timeout or a signal that caused
2355 * the wakeup and return the appropriate error code to the caller. Must be
2356 * called with the hb lock held.
2359 * 0 = no early wakeup detected;
2360 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2363 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket
*hb
,
2364 struct futex_q
*q
, union futex_key
*key2
,
2365 struct hrtimer_sleeper
*timeout
)
2370 * With the hb lock held, we avoid races while we process the wakeup.
2371 * We only need to hold hb (and not hb2) to ensure atomicity as the
2372 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2373 * It can't be requeued from uaddr2 to something else since we don't
2374 * support a PI aware source futex for requeue.
2376 if (!match_futex(&q
->key
, key2
)) {
2377 WARN_ON(q
->lock_ptr
&& (&hb
->lock
!= q
->lock_ptr
));
2379 * We were woken prior to requeue by a timeout or a signal.
2380 * Unqueue the futex_q and determine which it was.
2382 plist_del(&q
->list
, &hb
->chain
);
2385 /* Handle spurious wakeups gracefully */
2387 if (timeout
&& !timeout
->task
)
2389 else if (signal_pending(current
))
2390 ret
= -ERESTARTNOINTR
;
2396 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2397 * @uaddr: the futex we initially wait on (non-pi)
2398 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2399 * the same type, no requeueing from private to shared, etc.
2400 * @val: the expected value of uaddr
2401 * @abs_time: absolute timeout
2402 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2403 * @uaddr2: the pi futex we will take prior to returning to user-space
2405 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2406 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2407 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2408 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2409 * without one, the pi logic would not know which task to boost/deboost, if
2410 * there was a need to.
2412 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2413 * via the following--
2414 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2415 * 2) wakeup on uaddr2 after a requeue
2419 * If 3, cleanup and return -ERESTARTNOINTR.
2421 * If 2, we may then block on trying to take the rt_mutex and return via:
2422 * 5) successful lock
2425 * 8) other lock acquisition failure
2427 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2429 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2435 static int futex_wait_requeue_pi(u32 __user
*uaddr
, unsigned int flags
,
2436 u32 val
, ktime_t
*abs_time
, u32 bitset
,
2439 struct hrtimer_sleeper timeout
, *to
= NULL
;
2440 struct rt_mutex_waiter rt_waiter
;
2441 struct rt_mutex
*pi_mutex
= NULL
;
2442 struct futex_hash_bucket
*hb
;
2443 union futex_key key2
= FUTEX_KEY_INIT
;
2444 struct futex_q q
= futex_q_init
;
2447 if (uaddr
== uaddr2
)
2455 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2456 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2458 hrtimer_init_sleeper(to
, current
);
2459 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2460 current
->timer_slack_ns
);
2464 * The waiter is allocated on our stack, manipulated by the requeue
2465 * code while we sleep on uaddr.
2467 debug_rt_mutex_init_waiter(&rt_waiter
);
2468 RB_CLEAR_NODE(&rt_waiter
.pi_tree_entry
);
2469 RB_CLEAR_NODE(&rt_waiter
.tree_entry
);
2470 rt_waiter
.task
= NULL
;
2472 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
2473 if (unlikely(ret
!= 0))
2477 q
.rt_waiter
= &rt_waiter
;
2478 q
.requeue_pi_key
= &key2
;
2481 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2484 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2488 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2489 futex_wait_queue_me(hb
, &q
, to
);
2491 spin_lock(&hb
->lock
);
2492 ret
= handle_early_requeue_pi_wakeup(hb
, &q
, &key2
, to
);
2493 spin_unlock(&hb
->lock
);
2498 * In order for us to be here, we know our q.key == key2, and since
2499 * we took the hb->lock above, we also know that futex_requeue() has
2500 * completed and we no longer have to concern ourselves with a wakeup
2501 * race with the atomic proxy lock acquisition by the requeue code. The
2502 * futex_requeue dropped our key1 reference and incremented our key2
2506 /* Check if the requeue code acquired the second futex for us. */
2509 * Got the lock. We might not be the anticipated owner if we
2510 * did a lock-steal - fix up the PI-state in that case.
2512 if (q
.pi_state
&& (q
.pi_state
->owner
!= current
)) {
2513 spin_lock(q
.lock_ptr
);
2514 ret
= fixup_pi_state_owner(uaddr2
, &q
, current
);
2515 spin_unlock(q
.lock_ptr
);
2519 * We have been woken up by futex_unlock_pi(), a timeout, or a
2520 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2523 WARN_ON(!q
.pi_state
);
2524 pi_mutex
= &q
.pi_state
->pi_mutex
;
2525 ret
= rt_mutex_finish_proxy_lock(pi_mutex
, to
, &rt_waiter
, 1);
2526 debug_rt_mutex_free_waiter(&rt_waiter
);
2528 spin_lock(q
.lock_ptr
);
2530 * Fixup the pi_state owner and possibly acquire the lock if we
2533 res
= fixup_owner(uaddr2
, &q
, !ret
);
2535 * If fixup_owner() returned an error, proprogate that. If it
2536 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2539 ret
= (res
< 0) ? res
: 0;
2541 /* Unqueue and drop the lock. */
2546 * If fixup_pi_state_owner() faulted and was unable to handle the
2547 * fault, unlock the rt_mutex and return the fault to userspace.
2549 if (ret
== -EFAULT
) {
2550 if (pi_mutex
&& rt_mutex_owner(pi_mutex
) == current
)
2551 rt_mutex_unlock(pi_mutex
);
2552 } else if (ret
== -EINTR
) {
2554 * We've already been requeued, but cannot restart by calling
2555 * futex_lock_pi() directly. We could restart this syscall, but
2556 * it would detect that the user space "val" changed and return
2557 * -EWOULDBLOCK. Save the overhead of the restart and return
2558 * -EWOULDBLOCK directly.
2564 put_futex_key(&q
.key
);
2566 put_futex_key(&key2
);
2570 hrtimer_cancel(&to
->timer
);
2571 destroy_hrtimer_on_stack(&to
->timer
);
2577 * Support for robust futexes: the kernel cleans up held futexes at
2580 * Implementation: user-space maintains a per-thread list of locks it
2581 * is holding. Upon do_exit(), the kernel carefully walks this list,
2582 * and marks all locks that are owned by this thread with the
2583 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2584 * always manipulated with the lock held, so the list is private and
2585 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2586 * field, to allow the kernel to clean up if the thread dies after
2587 * acquiring the lock, but just before it could have added itself to
2588 * the list. There can only be one such pending lock.
2592 * sys_set_robust_list() - Set the robust-futex list head of a task
2593 * @head: pointer to the list-head
2594 * @len: length of the list-head, as userspace expects
2596 SYSCALL_DEFINE2(set_robust_list
, struct robust_list_head __user
*, head
,
2599 if (!futex_cmpxchg_enabled
)
2602 * The kernel knows only one size for now:
2604 if (unlikely(len
!= sizeof(*head
)))
2607 current
->robust_list
= head
;
2613 * sys_get_robust_list() - Get the robust-futex list head of a task
2614 * @pid: pid of the process [zero for current task]
2615 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2616 * @len_ptr: pointer to a length field, the kernel fills in the header size
2618 SYSCALL_DEFINE3(get_robust_list
, int, pid
,
2619 struct robust_list_head __user
* __user
*, head_ptr
,
2620 size_t __user
*, len_ptr
)
2622 struct robust_list_head __user
*head
;
2624 struct task_struct
*p
;
2626 if (!futex_cmpxchg_enabled
)
2635 p
= find_task_by_vpid(pid
);
2641 if (!ptrace_may_access(p
, PTRACE_MODE_READ
))
2644 head
= p
->robust_list
;
2647 if (put_user(sizeof(*head
), len_ptr
))
2649 return put_user(head
, head_ptr
);
2658 * Process a futex-list entry, check whether it's owned by the
2659 * dying task, and do notification if so:
2661 int handle_futex_death(u32 __user
*uaddr
, struct task_struct
*curr
, int pi
)
2663 u32 uval
, uninitialized_var(nval
), mval
;
2666 if (get_user(uval
, uaddr
))
2669 if ((uval
& FUTEX_TID_MASK
) == task_pid_vnr(curr
)) {
2671 * Ok, this dying thread is truly holding a futex
2672 * of interest. Set the OWNER_DIED bit atomically
2673 * via cmpxchg, and if the value had FUTEX_WAITERS
2674 * set, wake up a waiter (if any). (We have to do a
2675 * futex_wake() even if OWNER_DIED is already set -
2676 * to handle the rare but possible case of recursive
2677 * thread-death.) The rest of the cleanup is done in
2680 mval
= (uval
& FUTEX_WAITERS
) | FUTEX_OWNER_DIED
;
2682 * We are not holding a lock here, but we want to have
2683 * the pagefault_disable/enable() protection because
2684 * we want to handle the fault gracefully. If the
2685 * access fails we try to fault in the futex with R/W
2686 * verification via get_user_pages. get_user() above
2687 * does not guarantee R/W access. If that fails we
2688 * give up and leave the futex locked.
2690 if (cmpxchg_futex_value_locked(&nval
, uaddr
, uval
, mval
)) {
2691 if (fault_in_user_writeable(uaddr
))
2699 * Wake robust non-PI futexes here. The wakeup of
2700 * PI futexes happens in exit_pi_state():
2702 if (!pi
&& (uval
& FUTEX_WAITERS
))
2703 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
2709 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2711 static inline int fetch_robust_entry(struct robust_list __user
**entry
,
2712 struct robust_list __user
* __user
*head
,
2715 unsigned long uentry
;
2717 if (get_user(uentry
, (unsigned long __user
*)head
))
2720 *entry
= (void __user
*)(uentry
& ~1UL);
2727 * Walk curr->robust_list (very carefully, it's a userspace list!)
2728 * and mark any locks found there dead, and notify any waiters.
2730 * We silently return on any sign of list-walking problem.
2732 void exit_robust_list(struct task_struct
*curr
)
2734 struct robust_list_head __user
*head
= curr
->robust_list
;
2735 struct robust_list __user
*entry
, *next_entry
, *pending
;
2736 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
2737 unsigned int uninitialized_var(next_pi
);
2738 unsigned long futex_offset
;
2741 if (!futex_cmpxchg_enabled
)
2745 * Fetch the list head (which was registered earlier, via
2746 * sys_set_robust_list()):
2748 if (fetch_robust_entry(&entry
, &head
->list
.next
, &pi
))
2751 * Fetch the relative futex offset:
2753 if (get_user(futex_offset
, &head
->futex_offset
))
2756 * Fetch any possibly pending lock-add first, and handle it
2759 if (fetch_robust_entry(&pending
, &head
->list_op_pending
, &pip
))
2762 next_entry
= NULL
; /* avoid warning with gcc */
2763 while (entry
!= &head
->list
) {
2765 * Fetch the next entry in the list before calling
2766 * handle_futex_death:
2768 rc
= fetch_robust_entry(&next_entry
, &entry
->next
, &next_pi
);
2770 * A pending lock might already be on the list, so
2771 * don't process it twice:
2773 if (entry
!= pending
)
2774 if (handle_futex_death((void __user
*)entry
+ futex_offset
,
2782 * Avoid excessively long or circular lists:
2791 handle_futex_death((void __user
*)pending
+ futex_offset
,
2795 long do_futex(u32 __user
*uaddr
, int op
, u32 val
, ktime_t
*timeout
,
2796 u32 __user
*uaddr2
, u32 val2
, u32 val3
)
2798 int cmd
= op
& FUTEX_CMD_MASK
;
2799 unsigned int flags
= 0;
2801 if (!(op
& FUTEX_PRIVATE_FLAG
))
2802 flags
|= FLAGS_SHARED
;
2804 if (op
& FUTEX_CLOCK_REALTIME
) {
2805 flags
|= FLAGS_CLOCKRT
;
2806 if (cmd
!= FUTEX_WAIT_BITSET
&& cmd
!= FUTEX_WAIT_REQUEUE_PI
)
2812 case FUTEX_UNLOCK_PI
:
2813 case FUTEX_TRYLOCK_PI
:
2814 case FUTEX_WAIT_REQUEUE_PI
:
2815 case FUTEX_CMP_REQUEUE_PI
:
2816 if (!futex_cmpxchg_enabled
)
2822 val3
= FUTEX_BITSET_MATCH_ANY
;
2823 case FUTEX_WAIT_BITSET
:
2824 return futex_wait(uaddr
, flags
, val
, timeout
, val3
);
2826 val3
= FUTEX_BITSET_MATCH_ANY
;
2827 case FUTEX_WAKE_BITSET
:
2828 return futex_wake(uaddr
, flags
, val
, val3
);
2830 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, NULL
, 0);
2831 case FUTEX_CMP_REQUEUE
:
2832 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 0);
2834 return futex_wake_op(uaddr
, flags
, uaddr2
, val
, val2
, val3
);
2836 return futex_lock_pi(uaddr
, flags
, val
, timeout
, 0);
2837 case FUTEX_UNLOCK_PI
:
2838 return futex_unlock_pi(uaddr
, flags
);
2839 case FUTEX_TRYLOCK_PI
:
2840 return futex_lock_pi(uaddr
, flags
, 0, timeout
, 1);
2841 case FUTEX_WAIT_REQUEUE_PI
:
2842 val3
= FUTEX_BITSET_MATCH_ANY
;
2843 return futex_wait_requeue_pi(uaddr
, flags
, val
, timeout
, val3
,
2845 case FUTEX_CMP_REQUEUE_PI
:
2846 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 1);
2852 SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
2853 struct timespec __user
*, utime
, u32 __user
*, uaddr2
,
2857 ktime_t t
, *tp
= NULL
;
2859 int cmd
= op
& FUTEX_CMD_MASK
;
2861 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
2862 cmd
== FUTEX_WAIT_BITSET
||
2863 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
2864 if (copy_from_user(&ts
, utime
, sizeof(ts
)) != 0)
2866 if (!timespec_valid(&ts
))
2869 t
= timespec_to_ktime(ts
);
2870 if (cmd
== FUTEX_WAIT
)
2871 t
= ktime_add_safe(ktime_get(), t
);
2875 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2876 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2878 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
2879 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
2880 val2
= (u32
) (unsigned long) utime
;
2882 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
2885 static void __init
futex_detect_cmpxchg(void)
2887 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
2891 * This will fail and we want it. Some arch implementations do
2892 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2893 * functionality. We want to know that before we call in any
2894 * of the complex code paths. Also we want to prevent
2895 * registration of robust lists in that case. NULL is
2896 * guaranteed to fault and we get -EFAULT on functional
2897 * implementation, the non-functional ones will return
2900 if (cmpxchg_futex_value_locked(&curval
, NULL
, 0, 0) == -EFAULT
)
2901 futex_cmpxchg_enabled
= 1;
2905 static int __init
futex_init(void)
2907 unsigned int futex_shift
;
2910 #if CONFIG_BASE_SMALL
2911 futex_hashsize
= 16;
2913 futex_hashsize
= roundup_pow_of_two(256 * num_possible_cpus());
2916 futex_queues
= alloc_large_system_hash("futex", sizeof(*futex_queues
),
2918 futex_hashsize
< 256 ? HASH_SMALL
: 0,
2920 futex_hashsize
, futex_hashsize
);
2921 futex_hashsize
= 1UL << futex_shift
;
2923 futex_detect_cmpxchg();
2925 for (i
= 0; i
< futex_hashsize
; i
++) {
2926 atomic_set(&futex_queues
[i
].waiters
, 0);
2927 plist_head_init(&futex_queues
[i
].chain
);
2928 spin_lock_init(&futex_queues
[i
].lock
);
2933 __initcall(futex_init
);