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();
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();
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 * We need to check the following states:
749 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
751 * [1] NULL | --- | --- | 0 | 0/1 | Valid
752 * [2] NULL | --- | --- | >0 | 0/1 | Valid
754 * [3] Found | NULL | -- | Any | 0/1 | Invalid
756 * [4] Found | Found | NULL | 0 | 1 | Valid
757 * [5] Found | Found | NULL | >0 | 1 | Invalid
759 * [6] Found | Found | task | 0 | 1 | Valid
761 * [7] Found | Found | NULL | Any | 0 | Invalid
763 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
764 * [9] Found | Found | task | 0 | 0 | Invalid
765 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
767 * [1] Indicates that the kernel can acquire the futex atomically. We
768 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
770 * [2] Valid, if TID does not belong to a kernel thread. If no matching
771 * thread is found then it indicates that the owner TID has died.
773 * [3] Invalid. The waiter is queued on a non PI futex
775 * [4] Valid state after exit_robust_list(), which sets the user space
776 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
778 * [5] The user space value got manipulated between exit_robust_list()
779 * and exit_pi_state_list()
781 * [6] Valid state after exit_pi_state_list() which sets the new owner in
782 * the pi_state but cannot access the user space value.
784 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
786 * [8] Owner and user space value match
788 * [9] There is no transient state which sets the user space TID to 0
789 * except exit_robust_list(), but this is indicated by the
790 * FUTEX_OWNER_DIED bit. See [4]
792 * [10] There is no transient state which leaves owner and user space
797 * Validate that the existing waiter has a pi_state and sanity check
798 * the pi_state against the user space value. If correct, attach to
801 static int attach_to_pi_state(u32 uval
, struct futex_pi_state
*pi_state
,
802 struct futex_pi_state
**ps
)
804 pid_t pid
= uval
& FUTEX_TID_MASK
;
807 * Userspace might have messed up non-PI and PI futexes [3]
809 if (unlikely(!pi_state
))
812 WARN_ON(!atomic_read(&pi_state
->refcount
));
815 * Handle the owner died case:
817 if (uval
& FUTEX_OWNER_DIED
) {
819 * exit_pi_state_list sets owner to NULL and wakes the
820 * topmost waiter. The task which acquires the
821 * pi_state->rt_mutex will fixup owner.
823 if (!pi_state
->owner
) {
825 * No pi state owner, but the user space TID
826 * is not 0. Inconsistent state. [5]
831 * Take a ref on the state and return success. [4]
837 * If TID is 0, then either the dying owner has not
838 * yet executed exit_pi_state_list() or some waiter
839 * acquired the rtmutex in the pi state, but did not
840 * yet fixup the TID in user space.
842 * Take a ref on the state and return success. [6]
848 * If the owner died bit is not set, then the pi_state
849 * must have an owner. [7]
851 if (!pi_state
->owner
)
856 * Bail out if user space manipulated the futex value. If pi
857 * state exists then the owner TID must be the same as the
858 * user space TID. [9/10]
860 if (pid
!= task_pid_vnr(pi_state
->owner
))
863 atomic_inc(&pi_state
->refcount
);
869 lookup_pi_state(u32 uval
, struct futex_hash_bucket
*hb
,
870 union futex_key
*key
, struct futex_pi_state
**ps
)
872 struct futex_q
*match
= futex_top_waiter(hb
, key
);
873 struct futex_pi_state
*pi_state
= NULL
;
874 struct task_struct
*p
;
875 pid_t pid
= uval
& FUTEX_TID_MASK
;
878 * If there is a waiter on that futex, validate it and
879 * attach to the pi_state when the validation succeeds.
882 return attach_to_pi_state(uval
, match
->pi_state
, ps
);
885 * We are the first waiter - try to look up the real owner and attach
886 * the new pi_state to it, but bail out when TID = 0 [1]
890 p
= futex_find_get_task(pid
);
900 * We need to look at the task state flags to figure out,
901 * whether the task is exiting. To protect against the do_exit
902 * change of the task flags, we do this protected by
905 raw_spin_lock_irq(&p
->pi_lock
);
906 if (unlikely(p
->flags
& PF_EXITING
)) {
908 * The task is on the way out. When PF_EXITPIDONE is
909 * set, we know that the task has finished the
912 int ret
= (p
->flags
& PF_EXITPIDONE
) ? -ESRCH
: -EAGAIN
;
914 raw_spin_unlock_irq(&p
->pi_lock
);
920 * No existing pi state. First waiter. [2]
922 pi_state
= alloc_pi_state();
925 * Initialize the pi_mutex in locked state and make 'p'
928 rt_mutex_init_proxy_locked(&pi_state
->pi_mutex
, p
);
930 /* Store the key for possible exit cleanups: */
931 pi_state
->key
= *key
;
933 WARN_ON(!list_empty(&pi_state
->list
));
934 list_add(&pi_state
->list
, &p
->pi_state_list
);
936 raw_spin_unlock_irq(&p
->pi_lock
);
946 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
947 * @uaddr: the pi futex user address
948 * @hb: the pi futex hash bucket
949 * @key: the futex key associated with uaddr and hb
950 * @ps: the pi_state pointer where we store the result of the
952 * @task: the task to perform the atomic lock work for. This will
953 * be "current" except in the case of requeue pi.
954 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
958 * 1 - acquired the lock;
961 * The hb->lock and futex_key refs shall be held by the caller.
963 static int futex_lock_pi_atomic(u32 __user
*uaddr
, struct futex_hash_bucket
*hb
,
964 union futex_key
*key
,
965 struct futex_pi_state
**ps
,
966 struct task_struct
*task
, int set_waiters
)
968 int lock_taken
, ret
, force_take
= 0;
969 u32 uval
, newval
, curval
, vpid
= task_pid_vnr(task
);
972 ret
= lock_taken
= 0;
975 * To avoid races, we attempt to take the lock here again
976 * (by doing a 0 -> TID atomic cmpxchg), while holding all
977 * the locks. It will most likely not succeed.
981 newval
|= FUTEX_WAITERS
;
983 if (unlikely(cmpxchg_futex_value_locked(&curval
, uaddr
, 0, newval
)))
989 if ((unlikely((curval
& FUTEX_TID_MASK
) == vpid
)))
993 * Surprise - we got the lock, but we do not trust user space at all.
995 if (unlikely(!curval
)) {
997 * We verify whether there is kernel state for this
998 * futex. If not, we can safely assume, that the 0 ->
999 * TID transition is correct. If state exists, we do
1000 * not bother to fixup the user space state as it was
1001 * corrupted already.
1003 return futex_top_waiter(hb
, key
) ? -EINVAL
: 1;
1009 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
1010 * to wake at the next unlock.
1012 newval
= curval
| FUTEX_WAITERS
;
1015 * Should we force take the futex? See below.
1017 if (unlikely(force_take
)) {
1019 * Keep the OWNER_DIED and the WAITERS bit and set the
1022 newval
= (curval
& ~FUTEX_TID_MASK
) | vpid
;
1027 if (unlikely(cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
)))
1029 if (unlikely(curval
!= uval
))
1033 * We took the lock due to forced take over.
1035 if (unlikely(lock_taken
))
1039 * We dont have the lock. Look up the PI state (or create it if
1040 * we are the first waiter):
1042 ret
= lookup_pi_state(uval
, hb
, key
, ps
);
1044 if (unlikely(ret
)) {
1048 * We failed to find an owner for this
1049 * futex. So we have no pi_state to block
1050 * on. This can happen in two cases:
1053 * 2) A stale FUTEX_WAITERS bit
1055 * Re-read the futex value.
1057 if (get_futex_value_locked(&curval
, uaddr
))
1061 * If the owner died or we have a stale
1062 * WAITERS bit the owner TID in the user space
1065 if (!(curval
& FUTEX_TID_MASK
)) {
1078 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1079 * @q: The futex_q to unqueue
1081 * The q->lock_ptr must not be NULL and must be held by the caller.
1083 static void __unqueue_futex(struct futex_q
*q
)
1085 struct futex_hash_bucket
*hb
;
1087 if (WARN_ON_SMP(!q
->lock_ptr
|| !spin_is_locked(q
->lock_ptr
))
1088 || WARN_ON(plist_node_empty(&q
->list
)))
1091 hb
= container_of(q
->lock_ptr
, struct futex_hash_bucket
, lock
);
1092 plist_del(&q
->list
, &hb
->chain
);
1097 * The hash bucket lock must be held when this is called.
1098 * Afterwards, the futex_q must not be accessed.
1100 static void wake_futex(struct futex_q
*q
)
1102 struct task_struct
*p
= q
->task
;
1104 if (WARN(q
->pi_state
|| q
->rt_waiter
, "refusing to wake PI futex\n"))
1108 * We set q->lock_ptr = NULL _before_ we wake up the task. If
1109 * a non-futex wake up happens on another CPU then the task
1110 * might exit and p would dereference a non-existing task
1111 * struct. Prevent this by holding a reference on p across the
1118 * The waiting task can free the futex_q as soon as
1119 * q->lock_ptr = NULL is written, without taking any locks. A
1120 * memory barrier is required here to prevent the following
1121 * store to lock_ptr from getting ahead of the plist_del.
1126 wake_up_state(p
, TASK_NORMAL
);
1130 static int wake_futex_pi(u32 __user
*uaddr
, u32 uval
, struct futex_q
*this)
1132 struct task_struct
*new_owner
;
1133 struct futex_pi_state
*pi_state
= this->pi_state
;
1134 u32
uninitialized_var(curval
), newval
;
1141 * If current does not own the pi_state then the futex is
1142 * inconsistent and user space fiddled with the futex value.
1144 if (pi_state
->owner
!= current
)
1147 raw_spin_lock(&pi_state
->pi_mutex
.wait_lock
);
1148 new_owner
= rt_mutex_next_owner(&pi_state
->pi_mutex
);
1151 * It is possible that the next waiter (the one that brought
1152 * this owner to the kernel) timed out and is no longer
1153 * waiting on the lock.
1156 new_owner
= this->task
;
1159 * We pass it to the next owner. The WAITERS bit is always
1160 * kept enabled while there is PI state around. We cleanup the
1161 * owner died bit, because we are the owner.
1163 newval
= FUTEX_WAITERS
| task_pid_vnr(new_owner
);
1165 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
))
1167 else if (curval
!= uval
)
1170 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
1174 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
1175 WARN_ON(list_empty(&pi_state
->list
));
1176 list_del_init(&pi_state
->list
);
1177 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
1179 raw_spin_lock_irq(&new_owner
->pi_lock
);
1180 WARN_ON(!list_empty(&pi_state
->list
));
1181 list_add(&pi_state
->list
, &new_owner
->pi_state_list
);
1182 pi_state
->owner
= new_owner
;
1183 raw_spin_unlock_irq(&new_owner
->pi_lock
);
1185 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
1186 rt_mutex_unlock(&pi_state
->pi_mutex
);
1192 * Express the locking dependencies for lockdep:
1195 double_lock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1198 spin_lock(&hb1
->lock
);
1200 spin_lock_nested(&hb2
->lock
, SINGLE_DEPTH_NESTING
);
1201 } else { /* hb1 > hb2 */
1202 spin_lock(&hb2
->lock
);
1203 spin_lock_nested(&hb1
->lock
, SINGLE_DEPTH_NESTING
);
1208 double_unlock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1210 spin_unlock(&hb1
->lock
);
1212 spin_unlock(&hb2
->lock
);
1216 * Wake up waiters matching bitset queued on this futex (uaddr).
1219 futex_wake(u32 __user
*uaddr
, unsigned int flags
, int nr_wake
, u32 bitset
)
1221 struct futex_hash_bucket
*hb
;
1222 struct futex_q
*this, *next
;
1223 union futex_key key
= FUTEX_KEY_INIT
;
1229 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_READ
);
1230 if (unlikely(ret
!= 0))
1233 hb
= hash_futex(&key
);
1235 /* Make sure we really have tasks to wakeup */
1236 if (!hb_waiters_pending(hb
))
1239 spin_lock(&hb
->lock
);
1241 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
1242 if (match_futex (&this->key
, &key
)) {
1243 if (this->pi_state
|| this->rt_waiter
) {
1248 /* Check if one of the bits is set in both bitsets */
1249 if (!(this->bitset
& bitset
))
1253 if (++ret
>= nr_wake
)
1258 spin_unlock(&hb
->lock
);
1260 put_futex_key(&key
);
1266 * Wake up all waiters hashed on the physical page that is mapped
1267 * to this virtual address:
1270 futex_wake_op(u32 __user
*uaddr1
, unsigned int flags
, u32 __user
*uaddr2
,
1271 int nr_wake
, int nr_wake2
, int op
)
1273 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1274 struct futex_hash_bucket
*hb1
, *hb2
;
1275 struct futex_q
*this, *next
;
1279 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1280 if (unlikely(ret
!= 0))
1282 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
1283 if (unlikely(ret
!= 0))
1286 hb1
= hash_futex(&key1
);
1287 hb2
= hash_futex(&key2
);
1290 double_lock_hb(hb1
, hb2
);
1291 op_ret
= futex_atomic_op_inuser(op
, uaddr2
);
1292 if (unlikely(op_ret
< 0)) {
1294 double_unlock_hb(hb1
, hb2
);
1298 * we don't get EFAULT from MMU faults if we don't have an MMU,
1299 * but we might get them from range checking
1305 if (unlikely(op_ret
!= -EFAULT
)) {
1310 ret
= fault_in_user_writeable(uaddr2
);
1314 if (!(flags
& FLAGS_SHARED
))
1317 put_futex_key(&key2
);
1318 put_futex_key(&key1
);
1322 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1323 if (match_futex (&this->key
, &key1
)) {
1324 if (this->pi_state
|| this->rt_waiter
) {
1329 if (++ret
>= nr_wake
)
1336 plist_for_each_entry_safe(this, next
, &hb2
->chain
, list
) {
1337 if (match_futex (&this->key
, &key2
)) {
1338 if (this->pi_state
|| this->rt_waiter
) {
1343 if (++op_ret
>= nr_wake2
)
1351 double_unlock_hb(hb1
, hb2
);
1353 put_futex_key(&key2
);
1355 put_futex_key(&key1
);
1361 * requeue_futex() - Requeue a futex_q from one hb to another
1362 * @q: the futex_q to requeue
1363 * @hb1: the source hash_bucket
1364 * @hb2: the target hash_bucket
1365 * @key2: the new key for the requeued futex_q
1368 void requeue_futex(struct futex_q
*q
, struct futex_hash_bucket
*hb1
,
1369 struct futex_hash_bucket
*hb2
, union futex_key
*key2
)
1373 * If key1 and key2 hash to the same bucket, no need to
1376 if (likely(&hb1
->chain
!= &hb2
->chain
)) {
1377 plist_del(&q
->list
, &hb1
->chain
);
1378 hb_waiters_dec(hb1
);
1379 plist_add(&q
->list
, &hb2
->chain
);
1380 hb_waiters_inc(hb2
);
1381 q
->lock_ptr
= &hb2
->lock
;
1383 get_futex_key_refs(key2
);
1388 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1390 * @key: the key of the requeue target futex
1391 * @hb: the hash_bucket of the requeue target futex
1393 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1394 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1395 * to the requeue target futex so the waiter can detect the wakeup on the right
1396 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1397 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1398 * to protect access to the pi_state to fixup the owner later. Must be called
1399 * with both q->lock_ptr and hb->lock held.
1402 void requeue_pi_wake_futex(struct futex_q
*q
, union futex_key
*key
,
1403 struct futex_hash_bucket
*hb
)
1405 get_futex_key_refs(key
);
1410 WARN_ON(!q
->rt_waiter
);
1411 q
->rt_waiter
= NULL
;
1413 q
->lock_ptr
= &hb
->lock
;
1415 wake_up_state(q
->task
, TASK_NORMAL
);
1419 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1420 * @pifutex: the user address of the to futex
1421 * @hb1: the from futex hash bucket, must be locked by the caller
1422 * @hb2: the to futex hash bucket, must be locked by the caller
1423 * @key1: the from futex key
1424 * @key2: the to futex key
1425 * @ps: address to store the pi_state pointer
1426 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1428 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1429 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1430 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1431 * hb1 and hb2 must be held by the caller.
1434 * 0 - failed to acquire the lock atomically;
1435 * >0 - acquired the lock, return value is vpid of the top_waiter
1438 static int futex_proxy_trylock_atomic(u32 __user
*pifutex
,
1439 struct futex_hash_bucket
*hb1
,
1440 struct futex_hash_bucket
*hb2
,
1441 union futex_key
*key1
, union futex_key
*key2
,
1442 struct futex_pi_state
**ps
, int set_waiters
)
1444 struct futex_q
*top_waiter
= NULL
;
1448 if (get_futex_value_locked(&curval
, pifutex
))
1452 * Find the top_waiter and determine if there are additional waiters.
1453 * If the caller intends to requeue more than 1 waiter to pifutex,
1454 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1455 * as we have means to handle the possible fault. If not, don't set
1456 * the bit unecessarily as it will force the subsequent unlock to enter
1459 top_waiter
= futex_top_waiter(hb1
, key1
);
1461 /* There are no waiters, nothing for us to do. */
1465 /* Ensure we requeue to the expected futex. */
1466 if (!match_futex(top_waiter
->requeue_pi_key
, key2
))
1470 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1471 * the contended case or if set_waiters is 1. The pi_state is returned
1472 * in ps in contended cases.
1474 vpid
= task_pid_vnr(top_waiter
->task
);
1475 ret
= futex_lock_pi_atomic(pifutex
, hb2
, key2
, ps
, top_waiter
->task
,
1478 requeue_pi_wake_futex(top_waiter
, key2
, hb2
);
1485 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1486 * @uaddr1: source futex user address
1487 * @flags: futex flags (FLAGS_SHARED, etc.)
1488 * @uaddr2: target futex user address
1489 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1490 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1491 * @cmpval: @uaddr1 expected value (or %NULL)
1492 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1493 * pi futex (pi to pi requeue is not supported)
1495 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1496 * uaddr2 atomically on behalf of the top waiter.
1499 * >=0 - on success, the number of tasks requeued or woken;
1502 static int futex_requeue(u32 __user
*uaddr1
, unsigned int flags
,
1503 u32 __user
*uaddr2
, int nr_wake
, int nr_requeue
,
1504 u32
*cmpval
, int requeue_pi
)
1506 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1507 int drop_count
= 0, task_count
= 0, ret
;
1508 struct futex_pi_state
*pi_state
= NULL
;
1509 struct futex_hash_bucket
*hb1
, *hb2
;
1510 struct futex_q
*this, *next
;
1514 * Requeue PI only works on two distinct uaddrs. This
1515 * check is only valid for private futexes. See below.
1517 if (uaddr1
== uaddr2
)
1521 * requeue_pi requires a pi_state, try to allocate it now
1522 * without any locks in case it fails.
1524 if (refill_pi_state_cache())
1527 * requeue_pi must wake as many tasks as it can, up to nr_wake
1528 * + nr_requeue, since it acquires the rt_mutex prior to
1529 * returning to userspace, so as to not leave the rt_mutex with
1530 * waiters and no owner. However, second and third wake-ups
1531 * cannot be predicted as they involve race conditions with the
1532 * first wake and a fault while looking up the pi_state. Both
1533 * pthread_cond_signal() and pthread_cond_broadcast() should
1541 if (pi_state
!= NULL
) {
1543 * We will have to lookup the pi_state again, so free this one
1544 * to keep the accounting correct.
1546 free_pi_state(pi_state
);
1550 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1551 if (unlikely(ret
!= 0))
1553 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
,
1554 requeue_pi
? VERIFY_WRITE
: VERIFY_READ
);
1555 if (unlikely(ret
!= 0))
1559 * The check above which compares uaddrs is not sufficient for
1560 * shared futexes. We need to compare the keys:
1562 if (requeue_pi
&& match_futex(&key1
, &key2
)) {
1567 hb1
= hash_futex(&key1
);
1568 hb2
= hash_futex(&key2
);
1571 hb_waiters_inc(hb2
);
1572 double_lock_hb(hb1
, hb2
);
1574 if (likely(cmpval
!= NULL
)) {
1577 ret
= get_futex_value_locked(&curval
, uaddr1
);
1579 if (unlikely(ret
)) {
1580 double_unlock_hb(hb1
, hb2
);
1581 hb_waiters_dec(hb2
);
1583 ret
= get_user(curval
, uaddr1
);
1587 if (!(flags
& FLAGS_SHARED
))
1590 put_futex_key(&key2
);
1591 put_futex_key(&key1
);
1594 if (curval
!= *cmpval
) {
1600 if (requeue_pi
&& (task_count
- nr_wake
< nr_requeue
)) {
1602 * Attempt to acquire uaddr2 and wake the top waiter. If we
1603 * intend to requeue waiters, force setting the FUTEX_WAITERS
1604 * bit. We force this here where we are able to easily handle
1605 * faults rather in the requeue loop below.
1607 ret
= futex_proxy_trylock_atomic(uaddr2
, hb1
, hb2
, &key1
,
1608 &key2
, &pi_state
, nr_requeue
);
1611 * At this point the top_waiter has either taken uaddr2 or is
1612 * waiting on it. If the former, then the pi_state will not
1613 * exist yet, look it up one more time to ensure we have a
1614 * reference to it. If the lock was taken, ret contains the
1615 * vpid of the top waiter task.
1622 * If we acquired the lock, then the user
1623 * space value of uaddr2 should be vpid. It
1624 * cannot be changed by the top waiter as it
1625 * is blocked on hb2 lock if it tries to do
1626 * so. If something fiddled with it behind our
1627 * back the pi state lookup might unearth
1628 * it. So we rather use the known value than
1629 * rereading and handing potential crap to
1632 ret
= lookup_pi_state(ret
, hb2
, &key2
, &pi_state
);
1639 double_unlock_hb(hb1
, hb2
);
1640 hb_waiters_dec(hb2
);
1641 put_futex_key(&key2
);
1642 put_futex_key(&key1
);
1643 ret
= fault_in_user_writeable(uaddr2
);
1648 /* The owner was exiting, try again. */
1649 double_unlock_hb(hb1
, hb2
);
1650 hb_waiters_dec(hb2
);
1651 put_futex_key(&key2
);
1652 put_futex_key(&key1
);
1660 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1661 if (task_count
- nr_wake
>= nr_requeue
)
1664 if (!match_futex(&this->key
, &key1
))
1668 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1669 * be paired with each other and no other futex ops.
1671 * We should never be requeueing a futex_q with a pi_state,
1672 * which is awaiting a futex_unlock_pi().
1674 if ((requeue_pi
&& !this->rt_waiter
) ||
1675 (!requeue_pi
&& this->rt_waiter
) ||
1682 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1683 * lock, we already woke the top_waiter. If not, it will be
1684 * woken by futex_unlock_pi().
1686 if (++task_count
<= nr_wake
&& !requeue_pi
) {
1691 /* Ensure we requeue to the expected futex for requeue_pi. */
1692 if (requeue_pi
&& !match_futex(this->requeue_pi_key
, &key2
)) {
1698 * Requeue nr_requeue waiters and possibly one more in the case
1699 * of requeue_pi if we couldn't acquire the lock atomically.
1702 /* Prepare the waiter to take the rt_mutex. */
1703 atomic_inc(&pi_state
->refcount
);
1704 this->pi_state
= pi_state
;
1705 ret
= rt_mutex_start_proxy_lock(&pi_state
->pi_mutex
,
1709 /* We got the lock. */
1710 requeue_pi_wake_futex(this, &key2
, hb2
);
1715 this->pi_state
= NULL
;
1716 free_pi_state(pi_state
);
1720 requeue_futex(this, hb1
, hb2
, &key2
);
1725 double_unlock_hb(hb1
, hb2
);
1726 hb_waiters_dec(hb2
);
1729 * drop_futex_key_refs() must be called outside the spinlocks. During
1730 * the requeue we moved futex_q's from the hash bucket at key1 to the
1731 * one at key2 and updated their key pointer. We no longer need to
1732 * hold the references to key1.
1734 while (--drop_count
>= 0)
1735 drop_futex_key_refs(&key1
);
1738 put_futex_key(&key2
);
1740 put_futex_key(&key1
);
1742 if (pi_state
!= NULL
)
1743 free_pi_state(pi_state
);
1744 return ret
? ret
: task_count
;
1747 /* The key must be already stored in q->key. */
1748 static inline struct futex_hash_bucket
*queue_lock(struct futex_q
*q
)
1749 __acquires(&hb
->lock
)
1751 struct futex_hash_bucket
*hb
;
1753 hb
= hash_futex(&q
->key
);
1756 * Increment the counter before taking the lock so that
1757 * a potential waker won't miss a to-be-slept task that is
1758 * waiting for the spinlock. This is safe as all queue_lock()
1759 * users end up calling queue_me(). Similarly, for housekeeping,
1760 * decrement the counter at queue_unlock() when some error has
1761 * occurred and we don't end up adding the task to the list.
1765 q
->lock_ptr
= &hb
->lock
;
1767 spin_lock(&hb
->lock
); /* implies MB (A) */
1772 queue_unlock(struct futex_hash_bucket
*hb
)
1773 __releases(&hb
->lock
)
1775 spin_unlock(&hb
->lock
);
1780 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1781 * @q: The futex_q to enqueue
1782 * @hb: The destination hash bucket
1784 * The hb->lock must be held by the caller, and is released here. A call to
1785 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1786 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1787 * or nothing if the unqueue is done as part of the wake process and the unqueue
1788 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1791 static inline void queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
1792 __releases(&hb
->lock
)
1797 * The priority used to register this element is
1798 * - either the real thread-priority for the real-time threads
1799 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1800 * - or MAX_RT_PRIO for non-RT threads.
1801 * Thus, all RT-threads are woken first in priority order, and
1802 * the others are woken last, in FIFO order.
1804 prio
= min(current
->normal_prio
, MAX_RT_PRIO
);
1806 plist_node_init(&q
->list
, prio
);
1807 plist_add(&q
->list
, &hb
->chain
);
1809 spin_unlock(&hb
->lock
);
1813 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1814 * @q: The futex_q to unqueue
1816 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1817 * be paired with exactly one earlier call to queue_me().
1820 * 1 - if the futex_q was still queued (and we removed unqueued it);
1821 * 0 - if the futex_q was already removed by the waking thread
1823 static int unqueue_me(struct futex_q
*q
)
1825 spinlock_t
*lock_ptr
;
1828 /* In the common case we don't take the spinlock, which is nice. */
1830 lock_ptr
= q
->lock_ptr
;
1832 if (lock_ptr
!= NULL
) {
1833 spin_lock(lock_ptr
);
1835 * q->lock_ptr can change between reading it and
1836 * spin_lock(), causing us to take the wrong lock. This
1837 * corrects the race condition.
1839 * Reasoning goes like this: if we have the wrong lock,
1840 * q->lock_ptr must have changed (maybe several times)
1841 * between reading it and the spin_lock(). It can
1842 * change again after the spin_lock() but only if it was
1843 * already changed before the spin_lock(). It cannot,
1844 * however, change back to the original value. Therefore
1845 * we can detect whether we acquired the correct lock.
1847 if (unlikely(lock_ptr
!= q
->lock_ptr
)) {
1848 spin_unlock(lock_ptr
);
1853 BUG_ON(q
->pi_state
);
1855 spin_unlock(lock_ptr
);
1859 drop_futex_key_refs(&q
->key
);
1864 * PI futexes can not be requeued and must remove themself from the
1865 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1868 static void unqueue_me_pi(struct futex_q
*q
)
1869 __releases(q
->lock_ptr
)
1873 BUG_ON(!q
->pi_state
);
1874 free_pi_state(q
->pi_state
);
1877 spin_unlock(q
->lock_ptr
);
1881 * Fixup the pi_state owner with the new owner.
1883 * Must be called with hash bucket lock held and mm->sem held for non
1886 static int fixup_pi_state_owner(u32 __user
*uaddr
, struct futex_q
*q
,
1887 struct task_struct
*newowner
)
1889 u32 newtid
= task_pid_vnr(newowner
) | FUTEX_WAITERS
;
1890 struct futex_pi_state
*pi_state
= q
->pi_state
;
1891 struct task_struct
*oldowner
= pi_state
->owner
;
1892 u32 uval
, uninitialized_var(curval
), newval
;
1896 if (!pi_state
->owner
)
1897 newtid
|= FUTEX_OWNER_DIED
;
1900 * We are here either because we stole the rtmutex from the
1901 * previous highest priority waiter or we are the highest priority
1902 * waiter but failed to get the rtmutex the first time.
1903 * We have to replace the newowner TID in the user space variable.
1904 * This must be atomic as we have to preserve the owner died bit here.
1906 * Note: We write the user space value _before_ changing the pi_state
1907 * because we can fault here. Imagine swapped out pages or a fork
1908 * that marked all the anonymous memory readonly for cow.
1910 * Modifying pi_state _before_ the user space value would
1911 * leave the pi_state in an inconsistent state when we fault
1912 * here, because we need to drop the hash bucket lock to
1913 * handle the fault. This might be observed in the PID check
1914 * in lookup_pi_state.
1917 if (get_futex_value_locked(&uval
, uaddr
))
1921 newval
= (uval
& FUTEX_OWNER_DIED
) | newtid
;
1923 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
))
1931 * We fixed up user space. Now we need to fix the pi_state
1934 if (pi_state
->owner
!= NULL
) {
1935 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
1936 WARN_ON(list_empty(&pi_state
->list
));
1937 list_del_init(&pi_state
->list
);
1938 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
1941 pi_state
->owner
= newowner
;
1943 raw_spin_lock_irq(&newowner
->pi_lock
);
1944 WARN_ON(!list_empty(&pi_state
->list
));
1945 list_add(&pi_state
->list
, &newowner
->pi_state_list
);
1946 raw_spin_unlock_irq(&newowner
->pi_lock
);
1950 * To handle the page fault we need to drop the hash bucket
1951 * lock here. That gives the other task (either the highest priority
1952 * waiter itself or the task which stole the rtmutex) the
1953 * chance to try the fixup of the pi_state. So once we are
1954 * back from handling the fault we need to check the pi_state
1955 * after reacquiring the hash bucket lock and before trying to
1956 * do another fixup. When the fixup has been done already we
1960 spin_unlock(q
->lock_ptr
);
1962 ret
= fault_in_user_writeable(uaddr
);
1964 spin_lock(q
->lock_ptr
);
1967 * Check if someone else fixed it for us:
1969 if (pi_state
->owner
!= oldowner
)
1978 static long futex_wait_restart(struct restart_block
*restart
);
1981 * fixup_owner() - Post lock pi_state and corner case management
1982 * @uaddr: user address of the futex
1983 * @q: futex_q (contains pi_state and access to the rt_mutex)
1984 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1986 * After attempting to lock an rt_mutex, this function is called to cleanup
1987 * the pi_state owner as well as handle race conditions that may allow us to
1988 * acquire the lock. Must be called with the hb lock held.
1991 * 1 - success, lock taken;
1992 * 0 - success, lock not taken;
1993 * <0 - on error (-EFAULT)
1995 static int fixup_owner(u32 __user
*uaddr
, struct futex_q
*q
, int locked
)
1997 struct task_struct
*owner
;
2002 * Got the lock. We might not be the anticipated owner if we
2003 * did a lock-steal - fix up the PI-state in that case:
2005 if (q
->pi_state
->owner
!= current
)
2006 ret
= fixup_pi_state_owner(uaddr
, q
, current
);
2011 * Catch the rare case, where the lock was released when we were on the
2012 * way back before we locked the hash bucket.
2014 if (q
->pi_state
->owner
== current
) {
2016 * Try to get the rt_mutex now. This might fail as some other
2017 * task acquired the rt_mutex after we removed ourself from the
2018 * rt_mutex waiters list.
2020 if (rt_mutex_trylock(&q
->pi_state
->pi_mutex
)) {
2026 * pi_state is incorrect, some other task did a lock steal and
2027 * we returned due to timeout or signal without taking the
2028 * rt_mutex. Too late.
2030 raw_spin_lock(&q
->pi_state
->pi_mutex
.wait_lock
);
2031 owner
= rt_mutex_owner(&q
->pi_state
->pi_mutex
);
2033 owner
= rt_mutex_next_owner(&q
->pi_state
->pi_mutex
);
2034 raw_spin_unlock(&q
->pi_state
->pi_mutex
.wait_lock
);
2035 ret
= fixup_pi_state_owner(uaddr
, q
, owner
);
2040 * Paranoia check. If we did not take the lock, then we should not be
2041 * the owner of the rt_mutex.
2043 if (rt_mutex_owner(&q
->pi_state
->pi_mutex
) == current
)
2044 printk(KERN_ERR
"fixup_owner: ret = %d pi-mutex: %p "
2045 "pi-state %p\n", ret
,
2046 q
->pi_state
->pi_mutex
.owner
,
2047 q
->pi_state
->owner
);
2050 return ret
? ret
: locked
;
2054 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2055 * @hb: the futex hash bucket, must be locked by the caller
2056 * @q: the futex_q to queue up on
2057 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2059 static void futex_wait_queue_me(struct futex_hash_bucket
*hb
, struct futex_q
*q
,
2060 struct hrtimer_sleeper
*timeout
)
2063 * The task state is guaranteed to be set before another task can
2064 * wake it. set_current_state() is implemented using set_mb() and
2065 * queue_me() calls spin_unlock() upon completion, both serializing
2066 * access to the hash list and forcing another memory barrier.
2068 set_current_state(TASK_INTERRUPTIBLE
);
2073 hrtimer_start_expires(&timeout
->timer
, HRTIMER_MODE_ABS
);
2074 if (!hrtimer_active(&timeout
->timer
))
2075 timeout
->task
= NULL
;
2079 * If we have been removed from the hash list, then another task
2080 * has tried to wake us, and we can skip the call to schedule().
2082 if (likely(!plist_node_empty(&q
->list
))) {
2084 * If the timer has already expired, current will already be
2085 * flagged for rescheduling. Only call schedule if there
2086 * is no timeout, or if it has yet to expire.
2088 if (!timeout
|| timeout
->task
)
2089 freezable_schedule();
2091 __set_current_state(TASK_RUNNING
);
2095 * futex_wait_setup() - Prepare to wait on a futex
2096 * @uaddr: the futex userspace address
2097 * @val: the expected value
2098 * @flags: futex flags (FLAGS_SHARED, etc.)
2099 * @q: the associated futex_q
2100 * @hb: storage for hash_bucket pointer to be returned to caller
2102 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2103 * compare it with the expected value. Handle atomic faults internally.
2104 * Return with the hb lock held and a q.key reference on success, and unlocked
2105 * with no q.key reference on failure.
2108 * 0 - uaddr contains val and hb has been locked;
2109 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2111 static int futex_wait_setup(u32 __user
*uaddr
, u32 val
, unsigned int flags
,
2112 struct futex_q
*q
, struct futex_hash_bucket
**hb
)
2118 * Access the page AFTER the hash-bucket is locked.
2119 * Order is important:
2121 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2122 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2124 * The basic logical guarantee of a futex is that it blocks ONLY
2125 * if cond(var) is known to be true at the time of blocking, for
2126 * any cond. If we locked the hash-bucket after testing *uaddr, that
2127 * would open a race condition where we could block indefinitely with
2128 * cond(var) false, which would violate the guarantee.
2130 * On the other hand, we insert q and release the hash-bucket only
2131 * after testing *uaddr. This guarantees that futex_wait() will NOT
2132 * absorb a wakeup if *uaddr does not match the desired values
2133 * while the syscall executes.
2136 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
->key
, VERIFY_READ
);
2137 if (unlikely(ret
!= 0))
2141 *hb
= queue_lock(q
);
2143 ret
= get_futex_value_locked(&uval
, uaddr
);
2148 ret
= get_user(uval
, uaddr
);
2152 if (!(flags
& FLAGS_SHARED
))
2155 put_futex_key(&q
->key
);
2166 put_futex_key(&q
->key
);
2170 static int futex_wait(u32 __user
*uaddr
, unsigned int flags
, u32 val
,
2171 ktime_t
*abs_time
, u32 bitset
)
2173 struct hrtimer_sleeper timeout
, *to
= NULL
;
2174 struct restart_block
*restart
;
2175 struct futex_hash_bucket
*hb
;
2176 struct futex_q q
= futex_q_init
;
2186 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2187 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2189 hrtimer_init_sleeper(to
, current
);
2190 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2191 current
->timer_slack_ns
);
2196 * Prepare to wait on uaddr. On success, holds hb lock and increments
2199 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2203 /* queue_me and wait for wakeup, timeout, or a signal. */
2204 futex_wait_queue_me(hb
, &q
, to
);
2206 /* If we were woken (and unqueued), we succeeded, whatever. */
2208 /* unqueue_me() drops q.key ref */
2209 if (!unqueue_me(&q
))
2212 if (to
&& !to
->task
)
2216 * We expect signal_pending(current), but we might be the
2217 * victim of a spurious wakeup as well.
2219 if (!signal_pending(current
))
2226 restart
= ¤t_thread_info()->restart_block
;
2227 restart
->fn
= futex_wait_restart
;
2228 restart
->futex
.uaddr
= uaddr
;
2229 restart
->futex
.val
= val
;
2230 restart
->futex
.time
= abs_time
->tv64
;
2231 restart
->futex
.bitset
= bitset
;
2232 restart
->futex
.flags
= flags
| FLAGS_HAS_TIMEOUT
;
2234 ret
= -ERESTART_RESTARTBLOCK
;
2238 hrtimer_cancel(&to
->timer
);
2239 destroy_hrtimer_on_stack(&to
->timer
);
2245 static long futex_wait_restart(struct restart_block
*restart
)
2247 u32 __user
*uaddr
= restart
->futex
.uaddr
;
2248 ktime_t t
, *tp
= NULL
;
2250 if (restart
->futex
.flags
& FLAGS_HAS_TIMEOUT
) {
2251 t
.tv64
= restart
->futex
.time
;
2254 restart
->fn
= do_no_restart_syscall
;
2256 return (long)futex_wait(uaddr
, restart
->futex
.flags
,
2257 restart
->futex
.val
, tp
, restart
->futex
.bitset
);
2262 * Userspace tried a 0 -> TID atomic transition of the futex value
2263 * and failed. The kernel side here does the whole locking operation:
2264 * if there are waiters then it will block, it does PI, etc. (Due to
2265 * races the kernel might see a 0 value of the futex too.)
2267 static int futex_lock_pi(u32 __user
*uaddr
, unsigned int flags
, int detect
,
2268 ktime_t
*time
, int trylock
)
2270 struct hrtimer_sleeper timeout
, *to
= NULL
;
2271 struct futex_hash_bucket
*hb
;
2272 struct futex_q q
= futex_q_init
;
2275 if (refill_pi_state_cache())
2280 hrtimer_init_on_stack(&to
->timer
, CLOCK_REALTIME
,
2282 hrtimer_init_sleeper(to
, current
);
2283 hrtimer_set_expires(&to
->timer
, *time
);
2287 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
.key
, VERIFY_WRITE
);
2288 if (unlikely(ret
!= 0))
2292 hb
= queue_lock(&q
);
2294 ret
= futex_lock_pi_atomic(uaddr
, hb
, &q
.key
, &q
.pi_state
, current
, 0);
2295 if (unlikely(ret
)) {
2298 /* We got the lock. */
2300 goto out_unlock_put_key
;
2305 * Task is exiting and we just wait for the
2309 put_futex_key(&q
.key
);
2313 goto out_unlock_put_key
;
2318 * Only actually queue now that the atomic ops are done:
2322 WARN_ON(!q
.pi_state
);
2324 * Block on the PI mutex:
2327 ret
= rt_mutex_timed_futex_lock(&q
.pi_state
->pi_mutex
, to
);
2329 ret
= rt_mutex_trylock(&q
.pi_state
->pi_mutex
);
2330 /* Fixup the trylock return value: */
2331 ret
= ret
? 0 : -EWOULDBLOCK
;
2334 spin_lock(q
.lock_ptr
);
2336 * Fixup the pi_state owner and possibly acquire the lock if we
2339 res
= fixup_owner(uaddr
, &q
, !ret
);
2341 * If fixup_owner() returned an error, proprogate that. If it acquired
2342 * the lock, clear our -ETIMEDOUT or -EINTR.
2345 ret
= (res
< 0) ? res
: 0;
2348 * If fixup_owner() faulted and was unable to handle the fault, unlock
2349 * it and return the fault to userspace.
2351 if (ret
&& (rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
))
2352 rt_mutex_unlock(&q
.pi_state
->pi_mutex
);
2354 /* Unqueue and drop the lock */
2363 put_futex_key(&q
.key
);
2366 destroy_hrtimer_on_stack(&to
->timer
);
2367 return ret
!= -EINTR
? ret
: -ERESTARTNOINTR
;
2372 ret
= fault_in_user_writeable(uaddr
);
2376 if (!(flags
& FLAGS_SHARED
))
2379 put_futex_key(&q
.key
);
2384 * Userspace attempted a TID -> 0 atomic transition, and failed.
2385 * This is the in-kernel slowpath: we look up the PI state (if any),
2386 * and do the rt-mutex unlock.
2388 static int futex_unlock_pi(u32 __user
*uaddr
, unsigned int flags
)
2390 u32
uninitialized_var(curval
), uval
, vpid
= task_pid_vnr(current
);
2391 union futex_key key
= FUTEX_KEY_INIT
;
2392 struct futex_hash_bucket
*hb
;
2393 struct futex_q
*match
;
2397 if (get_user(uval
, uaddr
))
2400 * We release only a lock we actually own:
2402 if ((uval
& FUTEX_TID_MASK
) != vpid
)
2405 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_WRITE
);
2409 hb
= hash_futex(&key
);
2410 spin_lock(&hb
->lock
);
2413 * Check waiters first. We do not trust user space values at
2414 * all and we at least want to know if user space fiddled
2415 * with the futex value instead of blindly unlocking.
2417 match
= futex_top_waiter(hb
, &key
);
2419 ret
= wake_futex_pi(uaddr
, uval
, match
);
2421 * The atomic access to the futex value generated a
2422 * pagefault, so retry the user-access and the wakeup:
2430 * We have no kernel internal state, i.e. no waiters in the
2431 * kernel. Waiters which are about to queue themselves are stuck
2432 * on hb->lock. So we can safely ignore them. We do neither
2433 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2436 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, 0))
2440 * If uval has changed, let user space handle it.
2442 ret
= (curval
== uval
) ? 0 : -EAGAIN
;
2445 spin_unlock(&hb
->lock
);
2446 put_futex_key(&key
);
2450 spin_unlock(&hb
->lock
);
2451 put_futex_key(&key
);
2453 ret
= fault_in_user_writeable(uaddr
);
2461 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2462 * @hb: the hash_bucket futex_q was original enqueued on
2463 * @q: the futex_q woken while waiting to be requeued
2464 * @key2: the futex_key of the requeue target futex
2465 * @timeout: the timeout associated with the wait (NULL if none)
2467 * Detect if the task was woken on the initial futex as opposed to the requeue
2468 * target futex. If so, determine if it was a timeout or a signal that caused
2469 * the wakeup and return the appropriate error code to the caller. Must be
2470 * called with the hb lock held.
2473 * 0 = no early wakeup detected;
2474 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2477 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket
*hb
,
2478 struct futex_q
*q
, union futex_key
*key2
,
2479 struct hrtimer_sleeper
*timeout
)
2484 * With the hb lock held, we avoid races while we process the wakeup.
2485 * We only need to hold hb (and not hb2) to ensure atomicity as the
2486 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2487 * It can't be requeued from uaddr2 to something else since we don't
2488 * support a PI aware source futex for requeue.
2490 if (!match_futex(&q
->key
, key2
)) {
2491 WARN_ON(q
->lock_ptr
&& (&hb
->lock
!= q
->lock_ptr
));
2493 * We were woken prior to requeue by a timeout or a signal.
2494 * Unqueue the futex_q and determine which it was.
2496 plist_del(&q
->list
, &hb
->chain
);
2499 /* Handle spurious wakeups gracefully */
2501 if (timeout
&& !timeout
->task
)
2503 else if (signal_pending(current
))
2504 ret
= -ERESTARTNOINTR
;
2510 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2511 * @uaddr: the futex we initially wait on (non-pi)
2512 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2513 * the same type, no requeueing from private to shared, etc.
2514 * @val: the expected value of uaddr
2515 * @abs_time: absolute timeout
2516 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2517 * @uaddr2: the pi futex we will take prior to returning to user-space
2519 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2520 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2521 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2522 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2523 * without one, the pi logic would not know which task to boost/deboost, if
2524 * there was a need to.
2526 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2527 * via the following--
2528 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2529 * 2) wakeup on uaddr2 after a requeue
2533 * If 3, cleanup and return -ERESTARTNOINTR.
2535 * If 2, we may then block on trying to take the rt_mutex and return via:
2536 * 5) successful lock
2539 * 8) other lock acquisition failure
2541 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2543 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2549 static int futex_wait_requeue_pi(u32 __user
*uaddr
, unsigned int flags
,
2550 u32 val
, ktime_t
*abs_time
, u32 bitset
,
2553 struct hrtimer_sleeper timeout
, *to
= NULL
;
2554 struct rt_mutex_waiter rt_waiter
;
2555 struct rt_mutex
*pi_mutex
= NULL
;
2556 struct futex_hash_bucket
*hb
;
2557 union futex_key key2
= FUTEX_KEY_INIT
;
2558 struct futex_q q
= futex_q_init
;
2561 if (uaddr
== uaddr2
)
2569 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2570 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2572 hrtimer_init_sleeper(to
, current
);
2573 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2574 current
->timer_slack_ns
);
2578 * The waiter is allocated on our stack, manipulated by the requeue
2579 * code while we sleep on uaddr.
2581 debug_rt_mutex_init_waiter(&rt_waiter
);
2582 RB_CLEAR_NODE(&rt_waiter
.pi_tree_entry
);
2583 RB_CLEAR_NODE(&rt_waiter
.tree_entry
);
2584 rt_waiter
.task
= NULL
;
2586 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
2587 if (unlikely(ret
!= 0))
2591 q
.rt_waiter
= &rt_waiter
;
2592 q
.requeue_pi_key
= &key2
;
2595 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2598 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2603 * The check above which compares uaddrs is not sufficient for
2604 * shared futexes. We need to compare the keys:
2606 if (match_futex(&q
.key
, &key2
)) {
2611 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2612 futex_wait_queue_me(hb
, &q
, to
);
2614 spin_lock(&hb
->lock
);
2615 ret
= handle_early_requeue_pi_wakeup(hb
, &q
, &key2
, to
);
2616 spin_unlock(&hb
->lock
);
2621 * In order for us to be here, we know our q.key == key2, and since
2622 * we took the hb->lock above, we also know that futex_requeue() has
2623 * completed and we no longer have to concern ourselves with a wakeup
2624 * race with the atomic proxy lock acquisition by the requeue code. The
2625 * futex_requeue dropped our key1 reference and incremented our key2
2629 /* Check if the requeue code acquired the second futex for us. */
2632 * Got the lock. We might not be the anticipated owner if we
2633 * did a lock-steal - fix up the PI-state in that case.
2635 if (q
.pi_state
&& (q
.pi_state
->owner
!= current
)) {
2636 spin_lock(q
.lock_ptr
);
2637 ret
= fixup_pi_state_owner(uaddr2
, &q
, current
);
2638 spin_unlock(q
.lock_ptr
);
2642 * We have been woken up by futex_unlock_pi(), a timeout, or a
2643 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2646 WARN_ON(!q
.pi_state
);
2647 pi_mutex
= &q
.pi_state
->pi_mutex
;
2648 ret
= rt_mutex_finish_proxy_lock(pi_mutex
, to
, &rt_waiter
);
2649 debug_rt_mutex_free_waiter(&rt_waiter
);
2651 spin_lock(q
.lock_ptr
);
2653 * Fixup the pi_state owner and possibly acquire the lock if we
2656 res
= fixup_owner(uaddr2
, &q
, !ret
);
2658 * If fixup_owner() returned an error, proprogate that. If it
2659 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2662 ret
= (res
< 0) ? res
: 0;
2664 /* Unqueue and drop the lock. */
2669 * If fixup_pi_state_owner() faulted and was unable to handle the
2670 * fault, unlock the rt_mutex and return the fault to userspace.
2672 if (ret
== -EFAULT
) {
2673 if (pi_mutex
&& rt_mutex_owner(pi_mutex
) == current
)
2674 rt_mutex_unlock(pi_mutex
);
2675 } else if (ret
== -EINTR
) {
2677 * We've already been requeued, but cannot restart by calling
2678 * futex_lock_pi() directly. We could restart this syscall, but
2679 * it would detect that the user space "val" changed and return
2680 * -EWOULDBLOCK. Save the overhead of the restart and return
2681 * -EWOULDBLOCK directly.
2687 put_futex_key(&q
.key
);
2689 put_futex_key(&key2
);
2693 hrtimer_cancel(&to
->timer
);
2694 destroy_hrtimer_on_stack(&to
->timer
);
2700 * Support for robust futexes: the kernel cleans up held futexes at
2703 * Implementation: user-space maintains a per-thread list of locks it
2704 * is holding. Upon do_exit(), the kernel carefully walks this list,
2705 * and marks all locks that are owned by this thread with the
2706 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2707 * always manipulated with the lock held, so the list is private and
2708 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2709 * field, to allow the kernel to clean up if the thread dies after
2710 * acquiring the lock, but just before it could have added itself to
2711 * the list. There can only be one such pending lock.
2715 * sys_set_robust_list() - Set the robust-futex list head of a task
2716 * @head: pointer to the list-head
2717 * @len: length of the list-head, as userspace expects
2719 SYSCALL_DEFINE2(set_robust_list
, struct robust_list_head __user
*, head
,
2722 if (!futex_cmpxchg_enabled
)
2725 * The kernel knows only one size for now:
2727 if (unlikely(len
!= sizeof(*head
)))
2730 current
->robust_list
= head
;
2736 * sys_get_robust_list() - Get the robust-futex list head of a task
2737 * @pid: pid of the process [zero for current task]
2738 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2739 * @len_ptr: pointer to a length field, the kernel fills in the header size
2741 SYSCALL_DEFINE3(get_robust_list
, int, pid
,
2742 struct robust_list_head __user
* __user
*, head_ptr
,
2743 size_t __user
*, len_ptr
)
2745 struct robust_list_head __user
*head
;
2747 struct task_struct
*p
;
2749 if (!futex_cmpxchg_enabled
)
2758 p
= find_task_by_vpid(pid
);
2764 if (!ptrace_may_access(p
, PTRACE_MODE_READ
))
2767 head
= p
->robust_list
;
2770 if (put_user(sizeof(*head
), len_ptr
))
2772 return put_user(head
, head_ptr
);
2781 * Process a futex-list entry, check whether it's owned by the
2782 * dying task, and do notification if so:
2784 int handle_futex_death(u32 __user
*uaddr
, struct task_struct
*curr
, int pi
)
2786 u32 uval
, uninitialized_var(nval
), mval
;
2789 if (get_user(uval
, uaddr
))
2792 if ((uval
& FUTEX_TID_MASK
) == task_pid_vnr(curr
)) {
2794 * Ok, this dying thread is truly holding a futex
2795 * of interest. Set the OWNER_DIED bit atomically
2796 * via cmpxchg, and if the value had FUTEX_WAITERS
2797 * set, wake up a waiter (if any). (We have to do a
2798 * futex_wake() even if OWNER_DIED is already set -
2799 * to handle the rare but possible case of recursive
2800 * thread-death.) The rest of the cleanup is done in
2803 mval
= (uval
& FUTEX_WAITERS
) | FUTEX_OWNER_DIED
;
2805 * We are not holding a lock here, but we want to have
2806 * the pagefault_disable/enable() protection because
2807 * we want to handle the fault gracefully. If the
2808 * access fails we try to fault in the futex with R/W
2809 * verification via get_user_pages. get_user() above
2810 * does not guarantee R/W access. If that fails we
2811 * give up and leave the futex locked.
2813 if (cmpxchg_futex_value_locked(&nval
, uaddr
, uval
, mval
)) {
2814 if (fault_in_user_writeable(uaddr
))
2822 * Wake robust non-PI futexes here. The wakeup of
2823 * PI futexes happens in exit_pi_state():
2825 if (!pi
&& (uval
& FUTEX_WAITERS
))
2826 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
2832 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2834 static inline int fetch_robust_entry(struct robust_list __user
**entry
,
2835 struct robust_list __user
* __user
*head
,
2838 unsigned long uentry
;
2840 if (get_user(uentry
, (unsigned long __user
*)head
))
2843 *entry
= (void __user
*)(uentry
& ~1UL);
2850 * Walk curr->robust_list (very carefully, it's a userspace list!)
2851 * and mark any locks found there dead, and notify any waiters.
2853 * We silently return on any sign of list-walking problem.
2855 void exit_robust_list(struct task_struct
*curr
)
2857 struct robust_list_head __user
*head
= curr
->robust_list
;
2858 struct robust_list __user
*entry
, *next_entry
, *pending
;
2859 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
2860 unsigned int uninitialized_var(next_pi
);
2861 unsigned long futex_offset
;
2864 if (!futex_cmpxchg_enabled
)
2868 * Fetch the list head (which was registered earlier, via
2869 * sys_set_robust_list()):
2871 if (fetch_robust_entry(&entry
, &head
->list
.next
, &pi
))
2874 * Fetch the relative futex offset:
2876 if (get_user(futex_offset
, &head
->futex_offset
))
2879 * Fetch any possibly pending lock-add first, and handle it
2882 if (fetch_robust_entry(&pending
, &head
->list_op_pending
, &pip
))
2885 next_entry
= NULL
; /* avoid warning with gcc */
2886 while (entry
!= &head
->list
) {
2888 * Fetch the next entry in the list before calling
2889 * handle_futex_death:
2891 rc
= fetch_robust_entry(&next_entry
, &entry
->next
, &next_pi
);
2893 * A pending lock might already be on the list, so
2894 * don't process it twice:
2896 if (entry
!= pending
)
2897 if (handle_futex_death((void __user
*)entry
+ futex_offset
,
2905 * Avoid excessively long or circular lists:
2914 handle_futex_death((void __user
*)pending
+ futex_offset
,
2918 long do_futex(u32 __user
*uaddr
, int op
, u32 val
, ktime_t
*timeout
,
2919 u32 __user
*uaddr2
, u32 val2
, u32 val3
)
2921 int cmd
= op
& FUTEX_CMD_MASK
;
2922 unsigned int flags
= 0;
2924 if (!(op
& FUTEX_PRIVATE_FLAG
))
2925 flags
|= FLAGS_SHARED
;
2927 if (op
& FUTEX_CLOCK_REALTIME
) {
2928 flags
|= FLAGS_CLOCKRT
;
2929 if (cmd
!= FUTEX_WAIT_BITSET
&& cmd
!= FUTEX_WAIT_REQUEUE_PI
)
2935 case FUTEX_UNLOCK_PI
:
2936 case FUTEX_TRYLOCK_PI
:
2937 case FUTEX_WAIT_REQUEUE_PI
:
2938 case FUTEX_CMP_REQUEUE_PI
:
2939 if (!futex_cmpxchg_enabled
)
2945 val3
= FUTEX_BITSET_MATCH_ANY
;
2946 case FUTEX_WAIT_BITSET
:
2947 return futex_wait(uaddr
, flags
, val
, timeout
, val3
);
2949 val3
= FUTEX_BITSET_MATCH_ANY
;
2950 case FUTEX_WAKE_BITSET
:
2951 return futex_wake(uaddr
, flags
, val
, val3
);
2953 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, NULL
, 0);
2954 case FUTEX_CMP_REQUEUE
:
2955 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 0);
2957 return futex_wake_op(uaddr
, flags
, uaddr2
, val
, val2
, val3
);
2959 return futex_lock_pi(uaddr
, flags
, val
, timeout
, 0);
2960 case FUTEX_UNLOCK_PI
:
2961 return futex_unlock_pi(uaddr
, flags
);
2962 case FUTEX_TRYLOCK_PI
:
2963 return futex_lock_pi(uaddr
, flags
, 0, timeout
, 1);
2964 case FUTEX_WAIT_REQUEUE_PI
:
2965 val3
= FUTEX_BITSET_MATCH_ANY
;
2966 return futex_wait_requeue_pi(uaddr
, flags
, val
, timeout
, val3
,
2968 case FUTEX_CMP_REQUEUE_PI
:
2969 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 1);
2975 SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
2976 struct timespec __user
*, utime
, u32 __user
*, uaddr2
,
2980 ktime_t t
, *tp
= NULL
;
2982 int cmd
= op
& FUTEX_CMD_MASK
;
2984 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
2985 cmd
== FUTEX_WAIT_BITSET
||
2986 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
2987 if (copy_from_user(&ts
, utime
, sizeof(ts
)) != 0)
2989 if (!timespec_valid(&ts
))
2992 t
= timespec_to_ktime(ts
);
2993 if (cmd
== FUTEX_WAIT
)
2994 t
= ktime_add_safe(ktime_get(), t
);
2998 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2999 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3001 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
3002 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
3003 val2
= (u32
) (unsigned long) utime
;
3005 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
3008 static void __init
futex_detect_cmpxchg(void)
3010 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3014 * This will fail and we want it. Some arch implementations do
3015 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3016 * functionality. We want to know that before we call in any
3017 * of the complex code paths. Also we want to prevent
3018 * registration of robust lists in that case. NULL is
3019 * guaranteed to fault and we get -EFAULT on functional
3020 * implementation, the non-functional ones will return
3023 if (cmpxchg_futex_value_locked(&curval
, NULL
, 0, 0) == -EFAULT
)
3024 futex_cmpxchg_enabled
= 1;
3028 static int __init
futex_init(void)
3030 unsigned int futex_shift
;
3033 #if CONFIG_BASE_SMALL
3034 futex_hashsize
= 16;
3036 futex_hashsize
= roundup_pow_of_two(256 * num_possible_cpus());
3039 futex_queues
= alloc_large_system_hash("futex", sizeof(*futex_queues
),
3041 futex_hashsize
< 256 ? HASH_SMALL
: 0,
3043 futex_hashsize
, futex_hashsize
);
3044 futex_hashsize
= 1UL << futex_shift
;
3046 futex_detect_cmpxchg();
3048 for (i
= 0; i
< futex_hashsize
; i
++) {
3049 atomic_set(&futex_queues
[i
].waiters
, 0);
3050 plist_head_init(&futex_queues
[i
].chain
);
3051 spin_lock_init(&futex_queues
[i
].lock
);
3056 __initcall(futex_init
);