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
796 lookup_pi_state(u32 uval
, struct futex_hash_bucket
*hb
,
797 union futex_key
*key
, struct futex_pi_state
**ps
)
799 struct futex_q
*match
= futex_top_waiter(hb
, key
);
800 struct futex_pi_state
*pi_state
= NULL
;
801 struct task_struct
*p
;
802 pid_t pid
= uval
& FUTEX_TID_MASK
;
806 * Sanity check the waiter before increasing the
807 * refcount and attaching to it.
809 pi_state
= match
->pi_state
;
811 * Userspace might have messed up non-PI and PI
814 if (unlikely(!pi_state
))
817 WARN_ON(!atomic_read(&pi_state
->refcount
));
820 * Handle the owner died case:
822 if (uval
& FUTEX_OWNER_DIED
) {
824 * exit_pi_state_list sets owner to NULL and
825 * wakes the topmost waiter. The task which
826 * acquires the pi_state->rt_mutex will fixup
829 if (!pi_state
->owner
) {
831 * No pi state owner, but the user
832 * space TID is not 0. Inconsistent
838 * Take a ref on the state and
845 * If TID is 0, then either the dying owner
846 * has not yet executed exit_pi_state_list()
847 * or some waiter acquired the rtmutex in the
848 * pi state, but did not yet fixup the TID in
851 * Take a ref on the state and return. [6]
857 * If the owner died bit is not set,
858 * then the pi_state must have an
861 if (!pi_state
->owner
)
866 * Bail out if user space manipulated the
867 * futex value. If pi state exists then the
868 * owner TID must be the same as the user
871 if (pid
!= task_pid_vnr(pi_state
->owner
))
875 atomic_inc(&pi_state
->refcount
);
881 * We are the first waiter - try to look up the real owner and attach
882 * the new pi_state to it, but bail out when TID = 0 [1]
886 p
= futex_find_get_task(pid
);
896 * We need to look at the task state flags to figure out,
897 * whether the task is exiting. To protect against the do_exit
898 * change of the task flags, we do this protected by
901 raw_spin_lock_irq(&p
->pi_lock
);
902 if (unlikely(p
->flags
& PF_EXITING
)) {
904 * The task is on the way out. When PF_EXITPIDONE is
905 * set, we know that the task has finished the
908 int ret
= (p
->flags
& PF_EXITPIDONE
) ? -ESRCH
: -EAGAIN
;
910 raw_spin_unlock_irq(&p
->pi_lock
);
916 * No existing pi state. First waiter. [2]
918 pi_state
= alloc_pi_state();
921 * Initialize the pi_mutex in locked state and make 'p'
924 rt_mutex_init_proxy_locked(&pi_state
->pi_mutex
, p
);
926 /* Store the key for possible exit cleanups: */
927 pi_state
->key
= *key
;
929 WARN_ON(!list_empty(&pi_state
->list
));
930 list_add(&pi_state
->list
, &p
->pi_state_list
);
932 raw_spin_unlock_irq(&p
->pi_lock
);
942 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
943 * @uaddr: the pi futex user address
944 * @hb: the pi futex hash bucket
945 * @key: the futex key associated with uaddr and hb
946 * @ps: the pi_state pointer where we store the result of the
948 * @task: the task to perform the atomic lock work for. This will
949 * be "current" except in the case of requeue pi.
950 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
954 * 1 - acquired the lock;
957 * The hb->lock and futex_key refs shall be held by the caller.
959 static int futex_lock_pi_atomic(u32 __user
*uaddr
, struct futex_hash_bucket
*hb
,
960 union futex_key
*key
,
961 struct futex_pi_state
**ps
,
962 struct task_struct
*task
, int set_waiters
)
964 int lock_taken
, ret
, force_take
= 0;
965 u32 uval
, newval
, curval
, vpid
= task_pid_vnr(task
);
968 ret
= lock_taken
= 0;
971 * To avoid races, we attempt to take the lock here again
972 * (by doing a 0 -> TID atomic cmpxchg), while holding all
973 * the locks. It will most likely not succeed.
977 newval
|= FUTEX_WAITERS
;
979 if (unlikely(cmpxchg_futex_value_locked(&curval
, uaddr
, 0, newval
)))
985 if ((unlikely((curval
& FUTEX_TID_MASK
) == vpid
)))
989 * Surprise - we got the lock, but we do not trust user space at all.
991 if (unlikely(!curval
)) {
993 * We verify whether there is kernel state for this
994 * futex. If not, we can safely assume, that the 0 ->
995 * TID transition is correct. If state exists, we do
996 * not bother to fixup the user space state as it was
999 return futex_top_waiter(hb
, key
) ? -EINVAL
: 1;
1005 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
1006 * to wake at the next unlock.
1008 newval
= curval
| FUTEX_WAITERS
;
1011 * Should we force take the futex? See below.
1013 if (unlikely(force_take
)) {
1015 * Keep the OWNER_DIED and the WAITERS bit and set the
1018 newval
= (curval
& ~FUTEX_TID_MASK
) | vpid
;
1023 if (unlikely(cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
)))
1025 if (unlikely(curval
!= uval
))
1029 * We took the lock due to forced take over.
1031 if (unlikely(lock_taken
))
1035 * We dont have the lock. Look up the PI state (or create it if
1036 * we are the first waiter):
1038 ret
= lookup_pi_state(uval
, hb
, key
, ps
);
1040 if (unlikely(ret
)) {
1044 * We failed to find an owner for this
1045 * futex. So we have no pi_state to block
1046 * on. This can happen in two cases:
1049 * 2) A stale FUTEX_WAITERS bit
1051 * Re-read the futex value.
1053 if (get_futex_value_locked(&curval
, uaddr
))
1057 * If the owner died or we have a stale
1058 * WAITERS bit the owner TID in the user space
1061 if (!(curval
& FUTEX_TID_MASK
)) {
1074 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1075 * @q: The futex_q to unqueue
1077 * The q->lock_ptr must not be NULL and must be held by the caller.
1079 static void __unqueue_futex(struct futex_q
*q
)
1081 struct futex_hash_bucket
*hb
;
1083 if (WARN_ON_SMP(!q
->lock_ptr
|| !spin_is_locked(q
->lock_ptr
))
1084 || WARN_ON(plist_node_empty(&q
->list
)))
1087 hb
= container_of(q
->lock_ptr
, struct futex_hash_bucket
, lock
);
1088 plist_del(&q
->list
, &hb
->chain
);
1093 * The hash bucket lock must be held when this is called.
1094 * Afterwards, the futex_q must not be accessed.
1096 static void wake_futex(struct futex_q
*q
)
1098 struct task_struct
*p
= q
->task
;
1100 if (WARN(q
->pi_state
|| q
->rt_waiter
, "refusing to wake PI futex\n"))
1104 * We set q->lock_ptr = NULL _before_ we wake up the task. If
1105 * a non-futex wake up happens on another CPU then the task
1106 * might exit and p would dereference a non-existing task
1107 * struct. Prevent this by holding a reference on p across the
1114 * The waiting task can free the futex_q as soon as
1115 * q->lock_ptr = NULL is written, without taking any locks. A
1116 * memory barrier is required here to prevent the following
1117 * store to lock_ptr from getting ahead of the plist_del.
1122 wake_up_state(p
, TASK_NORMAL
);
1126 static int wake_futex_pi(u32 __user
*uaddr
, u32 uval
, struct futex_q
*this)
1128 struct task_struct
*new_owner
;
1129 struct futex_pi_state
*pi_state
= this->pi_state
;
1130 u32
uninitialized_var(curval
), newval
;
1137 * If current does not own the pi_state then the futex is
1138 * inconsistent and user space fiddled with the futex value.
1140 if (pi_state
->owner
!= current
)
1143 raw_spin_lock(&pi_state
->pi_mutex
.wait_lock
);
1144 new_owner
= rt_mutex_next_owner(&pi_state
->pi_mutex
);
1147 * It is possible that the next waiter (the one that brought
1148 * this owner to the kernel) timed out and is no longer
1149 * waiting on the lock.
1152 new_owner
= this->task
;
1155 * We pass it to the next owner. The WAITERS bit is always
1156 * kept enabled while there is PI state around. We cleanup the
1157 * owner died bit, because we are the owner.
1159 newval
= FUTEX_WAITERS
| task_pid_vnr(new_owner
);
1161 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
))
1163 else if (curval
!= uval
)
1166 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
1170 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
1171 WARN_ON(list_empty(&pi_state
->list
));
1172 list_del_init(&pi_state
->list
);
1173 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
1175 raw_spin_lock_irq(&new_owner
->pi_lock
);
1176 WARN_ON(!list_empty(&pi_state
->list
));
1177 list_add(&pi_state
->list
, &new_owner
->pi_state_list
);
1178 pi_state
->owner
= new_owner
;
1179 raw_spin_unlock_irq(&new_owner
->pi_lock
);
1181 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
1182 rt_mutex_unlock(&pi_state
->pi_mutex
);
1188 * Express the locking dependencies for lockdep:
1191 double_lock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1194 spin_lock(&hb1
->lock
);
1196 spin_lock_nested(&hb2
->lock
, SINGLE_DEPTH_NESTING
);
1197 } else { /* hb1 > hb2 */
1198 spin_lock(&hb2
->lock
);
1199 spin_lock_nested(&hb1
->lock
, SINGLE_DEPTH_NESTING
);
1204 double_unlock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1206 spin_unlock(&hb1
->lock
);
1208 spin_unlock(&hb2
->lock
);
1212 * Wake up waiters matching bitset queued on this futex (uaddr).
1215 futex_wake(u32 __user
*uaddr
, unsigned int flags
, int nr_wake
, u32 bitset
)
1217 struct futex_hash_bucket
*hb
;
1218 struct futex_q
*this, *next
;
1219 union futex_key key
= FUTEX_KEY_INIT
;
1225 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_READ
);
1226 if (unlikely(ret
!= 0))
1229 hb
= hash_futex(&key
);
1231 /* Make sure we really have tasks to wakeup */
1232 if (!hb_waiters_pending(hb
))
1235 spin_lock(&hb
->lock
);
1237 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
1238 if (match_futex (&this->key
, &key
)) {
1239 if (this->pi_state
|| this->rt_waiter
) {
1244 /* Check if one of the bits is set in both bitsets */
1245 if (!(this->bitset
& bitset
))
1249 if (++ret
>= nr_wake
)
1254 spin_unlock(&hb
->lock
);
1256 put_futex_key(&key
);
1262 * Wake up all waiters hashed on the physical page that is mapped
1263 * to this virtual address:
1266 futex_wake_op(u32 __user
*uaddr1
, unsigned int flags
, u32 __user
*uaddr2
,
1267 int nr_wake
, int nr_wake2
, int op
)
1269 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1270 struct futex_hash_bucket
*hb1
, *hb2
;
1271 struct futex_q
*this, *next
;
1275 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1276 if (unlikely(ret
!= 0))
1278 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
1279 if (unlikely(ret
!= 0))
1282 hb1
= hash_futex(&key1
);
1283 hb2
= hash_futex(&key2
);
1286 double_lock_hb(hb1
, hb2
);
1287 op_ret
= futex_atomic_op_inuser(op
, uaddr2
);
1288 if (unlikely(op_ret
< 0)) {
1290 double_unlock_hb(hb1
, hb2
);
1294 * we don't get EFAULT from MMU faults if we don't have an MMU,
1295 * but we might get them from range checking
1301 if (unlikely(op_ret
!= -EFAULT
)) {
1306 ret
= fault_in_user_writeable(uaddr2
);
1310 if (!(flags
& FLAGS_SHARED
))
1313 put_futex_key(&key2
);
1314 put_futex_key(&key1
);
1318 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1319 if (match_futex (&this->key
, &key1
)) {
1320 if (this->pi_state
|| this->rt_waiter
) {
1325 if (++ret
>= nr_wake
)
1332 plist_for_each_entry_safe(this, next
, &hb2
->chain
, list
) {
1333 if (match_futex (&this->key
, &key2
)) {
1334 if (this->pi_state
|| this->rt_waiter
) {
1339 if (++op_ret
>= nr_wake2
)
1347 double_unlock_hb(hb1
, hb2
);
1349 put_futex_key(&key2
);
1351 put_futex_key(&key1
);
1357 * requeue_futex() - Requeue a futex_q from one hb to another
1358 * @q: the futex_q to requeue
1359 * @hb1: the source hash_bucket
1360 * @hb2: the target hash_bucket
1361 * @key2: the new key for the requeued futex_q
1364 void requeue_futex(struct futex_q
*q
, struct futex_hash_bucket
*hb1
,
1365 struct futex_hash_bucket
*hb2
, union futex_key
*key2
)
1369 * If key1 and key2 hash to the same bucket, no need to
1372 if (likely(&hb1
->chain
!= &hb2
->chain
)) {
1373 plist_del(&q
->list
, &hb1
->chain
);
1374 hb_waiters_dec(hb1
);
1375 plist_add(&q
->list
, &hb2
->chain
);
1376 hb_waiters_inc(hb2
);
1377 q
->lock_ptr
= &hb2
->lock
;
1379 get_futex_key_refs(key2
);
1384 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1386 * @key: the key of the requeue target futex
1387 * @hb: the hash_bucket of the requeue target futex
1389 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1390 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1391 * to the requeue target futex so the waiter can detect the wakeup on the right
1392 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1393 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1394 * to protect access to the pi_state to fixup the owner later. Must be called
1395 * with both q->lock_ptr and hb->lock held.
1398 void requeue_pi_wake_futex(struct futex_q
*q
, union futex_key
*key
,
1399 struct futex_hash_bucket
*hb
)
1401 get_futex_key_refs(key
);
1406 WARN_ON(!q
->rt_waiter
);
1407 q
->rt_waiter
= NULL
;
1409 q
->lock_ptr
= &hb
->lock
;
1411 wake_up_state(q
->task
, TASK_NORMAL
);
1415 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1416 * @pifutex: the user address of the to futex
1417 * @hb1: the from futex hash bucket, must be locked by the caller
1418 * @hb2: the to futex hash bucket, must be locked by the caller
1419 * @key1: the from futex key
1420 * @key2: the to futex key
1421 * @ps: address to store the pi_state pointer
1422 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1424 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1425 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1426 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1427 * hb1 and hb2 must be held by the caller.
1430 * 0 - failed to acquire the lock atomically;
1431 * >0 - acquired the lock, return value is vpid of the top_waiter
1434 static int futex_proxy_trylock_atomic(u32 __user
*pifutex
,
1435 struct futex_hash_bucket
*hb1
,
1436 struct futex_hash_bucket
*hb2
,
1437 union futex_key
*key1
, union futex_key
*key2
,
1438 struct futex_pi_state
**ps
, int set_waiters
)
1440 struct futex_q
*top_waiter
= NULL
;
1444 if (get_futex_value_locked(&curval
, pifutex
))
1448 * Find the top_waiter and determine if there are additional waiters.
1449 * If the caller intends to requeue more than 1 waiter to pifutex,
1450 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1451 * as we have means to handle the possible fault. If not, don't set
1452 * the bit unecessarily as it will force the subsequent unlock to enter
1455 top_waiter
= futex_top_waiter(hb1
, key1
);
1457 /* There are no waiters, nothing for us to do. */
1461 /* Ensure we requeue to the expected futex. */
1462 if (!match_futex(top_waiter
->requeue_pi_key
, key2
))
1466 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1467 * the contended case or if set_waiters is 1. The pi_state is returned
1468 * in ps in contended cases.
1470 vpid
= task_pid_vnr(top_waiter
->task
);
1471 ret
= futex_lock_pi_atomic(pifutex
, hb2
, key2
, ps
, top_waiter
->task
,
1474 requeue_pi_wake_futex(top_waiter
, key2
, hb2
);
1481 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1482 * @uaddr1: source futex user address
1483 * @flags: futex flags (FLAGS_SHARED, etc.)
1484 * @uaddr2: target futex user address
1485 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1486 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1487 * @cmpval: @uaddr1 expected value (or %NULL)
1488 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1489 * pi futex (pi to pi requeue is not supported)
1491 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1492 * uaddr2 atomically on behalf of the top waiter.
1495 * >=0 - on success, the number of tasks requeued or woken;
1498 static int futex_requeue(u32 __user
*uaddr1
, unsigned int flags
,
1499 u32 __user
*uaddr2
, int nr_wake
, int nr_requeue
,
1500 u32
*cmpval
, int requeue_pi
)
1502 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1503 int drop_count
= 0, task_count
= 0, ret
;
1504 struct futex_pi_state
*pi_state
= NULL
;
1505 struct futex_hash_bucket
*hb1
, *hb2
;
1506 struct futex_q
*this, *next
;
1510 * Requeue PI only works on two distinct uaddrs. This
1511 * check is only valid for private futexes. See below.
1513 if (uaddr1
== uaddr2
)
1517 * requeue_pi requires a pi_state, try to allocate it now
1518 * without any locks in case it fails.
1520 if (refill_pi_state_cache())
1523 * requeue_pi must wake as many tasks as it can, up to nr_wake
1524 * + nr_requeue, since it acquires the rt_mutex prior to
1525 * returning to userspace, so as to not leave the rt_mutex with
1526 * waiters and no owner. However, second and third wake-ups
1527 * cannot be predicted as they involve race conditions with the
1528 * first wake and a fault while looking up the pi_state. Both
1529 * pthread_cond_signal() and pthread_cond_broadcast() should
1537 if (pi_state
!= NULL
) {
1539 * We will have to lookup the pi_state again, so free this one
1540 * to keep the accounting correct.
1542 free_pi_state(pi_state
);
1546 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1547 if (unlikely(ret
!= 0))
1549 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
,
1550 requeue_pi
? VERIFY_WRITE
: VERIFY_READ
);
1551 if (unlikely(ret
!= 0))
1555 * The check above which compares uaddrs is not sufficient for
1556 * shared futexes. We need to compare the keys:
1558 if (requeue_pi
&& match_futex(&key1
, &key2
)) {
1563 hb1
= hash_futex(&key1
);
1564 hb2
= hash_futex(&key2
);
1567 hb_waiters_inc(hb2
);
1568 double_lock_hb(hb1
, hb2
);
1570 if (likely(cmpval
!= NULL
)) {
1573 ret
= get_futex_value_locked(&curval
, uaddr1
);
1575 if (unlikely(ret
)) {
1576 double_unlock_hb(hb1
, hb2
);
1577 hb_waiters_dec(hb2
);
1579 ret
= get_user(curval
, uaddr1
);
1583 if (!(flags
& FLAGS_SHARED
))
1586 put_futex_key(&key2
);
1587 put_futex_key(&key1
);
1590 if (curval
!= *cmpval
) {
1596 if (requeue_pi
&& (task_count
- nr_wake
< nr_requeue
)) {
1598 * Attempt to acquire uaddr2 and wake the top waiter. If we
1599 * intend to requeue waiters, force setting the FUTEX_WAITERS
1600 * bit. We force this here where we are able to easily handle
1601 * faults rather in the requeue loop below.
1603 ret
= futex_proxy_trylock_atomic(uaddr2
, hb1
, hb2
, &key1
,
1604 &key2
, &pi_state
, nr_requeue
);
1607 * At this point the top_waiter has either taken uaddr2 or is
1608 * waiting on it. If the former, then the pi_state will not
1609 * exist yet, look it up one more time to ensure we have a
1610 * reference to it. If the lock was taken, ret contains the
1611 * vpid of the top waiter task.
1618 * If we acquired the lock, then the user
1619 * space value of uaddr2 should be vpid. It
1620 * cannot be changed by the top waiter as it
1621 * is blocked on hb2 lock if it tries to do
1622 * so. If something fiddled with it behind our
1623 * back the pi state lookup might unearth
1624 * it. So we rather use the known value than
1625 * rereading and handing potential crap to
1628 ret
= lookup_pi_state(ret
, hb2
, &key2
, &pi_state
);
1635 double_unlock_hb(hb1
, hb2
);
1636 hb_waiters_dec(hb2
);
1637 put_futex_key(&key2
);
1638 put_futex_key(&key1
);
1639 ret
= fault_in_user_writeable(uaddr2
);
1644 /* The owner was exiting, try again. */
1645 double_unlock_hb(hb1
, hb2
);
1646 hb_waiters_dec(hb2
);
1647 put_futex_key(&key2
);
1648 put_futex_key(&key1
);
1656 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1657 if (task_count
- nr_wake
>= nr_requeue
)
1660 if (!match_futex(&this->key
, &key1
))
1664 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1665 * be paired with each other and no other futex ops.
1667 * We should never be requeueing a futex_q with a pi_state,
1668 * which is awaiting a futex_unlock_pi().
1670 if ((requeue_pi
&& !this->rt_waiter
) ||
1671 (!requeue_pi
&& this->rt_waiter
) ||
1678 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1679 * lock, we already woke the top_waiter. If not, it will be
1680 * woken by futex_unlock_pi().
1682 if (++task_count
<= nr_wake
&& !requeue_pi
) {
1687 /* Ensure we requeue to the expected futex for requeue_pi. */
1688 if (requeue_pi
&& !match_futex(this->requeue_pi_key
, &key2
)) {
1694 * Requeue nr_requeue waiters and possibly one more in the case
1695 * of requeue_pi if we couldn't acquire the lock atomically.
1698 /* Prepare the waiter to take the rt_mutex. */
1699 atomic_inc(&pi_state
->refcount
);
1700 this->pi_state
= pi_state
;
1701 ret
= rt_mutex_start_proxy_lock(&pi_state
->pi_mutex
,
1705 /* We got the lock. */
1706 requeue_pi_wake_futex(this, &key2
, hb2
);
1711 this->pi_state
= NULL
;
1712 free_pi_state(pi_state
);
1716 requeue_futex(this, hb1
, hb2
, &key2
);
1721 double_unlock_hb(hb1
, hb2
);
1722 hb_waiters_dec(hb2
);
1725 * drop_futex_key_refs() must be called outside the spinlocks. During
1726 * the requeue we moved futex_q's from the hash bucket at key1 to the
1727 * one at key2 and updated their key pointer. We no longer need to
1728 * hold the references to key1.
1730 while (--drop_count
>= 0)
1731 drop_futex_key_refs(&key1
);
1734 put_futex_key(&key2
);
1736 put_futex_key(&key1
);
1738 if (pi_state
!= NULL
)
1739 free_pi_state(pi_state
);
1740 return ret
? ret
: task_count
;
1743 /* The key must be already stored in q->key. */
1744 static inline struct futex_hash_bucket
*queue_lock(struct futex_q
*q
)
1745 __acquires(&hb
->lock
)
1747 struct futex_hash_bucket
*hb
;
1749 hb
= hash_futex(&q
->key
);
1752 * Increment the counter before taking the lock so that
1753 * a potential waker won't miss a to-be-slept task that is
1754 * waiting for the spinlock. This is safe as all queue_lock()
1755 * users end up calling queue_me(). Similarly, for housekeeping,
1756 * decrement the counter at queue_unlock() when some error has
1757 * occurred and we don't end up adding the task to the list.
1761 q
->lock_ptr
= &hb
->lock
;
1763 spin_lock(&hb
->lock
); /* implies MB (A) */
1768 queue_unlock(struct futex_hash_bucket
*hb
)
1769 __releases(&hb
->lock
)
1771 spin_unlock(&hb
->lock
);
1776 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1777 * @q: The futex_q to enqueue
1778 * @hb: The destination hash bucket
1780 * The hb->lock must be held by the caller, and is released here. A call to
1781 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1782 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1783 * or nothing if the unqueue is done as part of the wake process and the unqueue
1784 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1787 static inline void queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
1788 __releases(&hb
->lock
)
1793 * The priority used to register this element is
1794 * - either the real thread-priority for the real-time threads
1795 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1796 * - or MAX_RT_PRIO for non-RT threads.
1797 * Thus, all RT-threads are woken first in priority order, and
1798 * the others are woken last, in FIFO order.
1800 prio
= min(current
->normal_prio
, MAX_RT_PRIO
);
1802 plist_node_init(&q
->list
, prio
);
1803 plist_add(&q
->list
, &hb
->chain
);
1805 spin_unlock(&hb
->lock
);
1809 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1810 * @q: The futex_q to unqueue
1812 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1813 * be paired with exactly one earlier call to queue_me().
1816 * 1 - if the futex_q was still queued (and we removed unqueued it);
1817 * 0 - if the futex_q was already removed by the waking thread
1819 static int unqueue_me(struct futex_q
*q
)
1821 spinlock_t
*lock_ptr
;
1824 /* In the common case we don't take the spinlock, which is nice. */
1826 lock_ptr
= q
->lock_ptr
;
1828 if (lock_ptr
!= NULL
) {
1829 spin_lock(lock_ptr
);
1831 * q->lock_ptr can change between reading it and
1832 * spin_lock(), causing us to take the wrong lock. This
1833 * corrects the race condition.
1835 * Reasoning goes like this: if we have the wrong lock,
1836 * q->lock_ptr must have changed (maybe several times)
1837 * between reading it and the spin_lock(). It can
1838 * change again after the spin_lock() but only if it was
1839 * already changed before the spin_lock(). It cannot,
1840 * however, change back to the original value. Therefore
1841 * we can detect whether we acquired the correct lock.
1843 if (unlikely(lock_ptr
!= q
->lock_ptr
)) {
1844 spin_unlock(lock_ptr
);
1849 BUG_ON(q
->pi_state
);
1851 spin_unlock(lock_ptr
);
1855 drop_futex_key_refs(&q
->key
);
1860 * PI futexes can not be requeued and must remove themself from the
1861 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1864 static void unqueue_me_pi(struct futex_q
*q
)
1865 __releases(q
->lock_ptr
)
1869 BUG_ON(!q
->pi_state
);
1870 free_pi_state(q
->pi_state
);
1873 spin_unlock(q
->lock_ptr
);
1877 * Fixup the pi_state owner with the new owner.
1879 * Must be called with hash bucket lock held and mm->sem held for non
1882 static int fixup_pi_state_owner(u32 __user
*uaddr
, struct futex_q
*q
,
1883 struct task_struct
*newowner
)
1885 u32 newtid
= task_pid_vnr(newowner
) | FUTEX_WAITERS
;
1886 struct futex_pi_state
*pi_state
= q
->pi_state
;
1887 struct task_struct
*oldowner
= pi_state
->owner
;
1888 u32 uval
, uninitialized_var(curval
), newval
;
1892 if (!pi_state
->owner
)
1893 newtid
|= FUTEX_OWNER_DIED
;
1896 * We are here either because we stole the rtmutex from the
1897 * previous highest priority waiter or we are the highest priority
1898 * waiter but failed to get the rtmutex the first time.
1899 * We have to replace the newowner TID in the user space variable.
1900 * This must be atomic as we have to preserve the owner died bit here.
1902 * Note: We write the user space value _before_ changing the pi_state
1903 * because we can fault here. Imagine swapped out pages or a fork
1904 * that marked all the anonymous memory readonly for cow.
1906 * Modifying pi_state _before_ the user space value would
1907 * leave the pi_state in an inconsistent state when we fault
1908 * here, because we need to drop the hash bucket lock to
1909 * handle the fault. This might be observed in the PID check
1910 * in lookup_pi_state.
1913 if (get_futex_value_locked(&uval
, uaddr
))
1917 newval
= (uval
& FUTEX_OWNER_DIED
) | newtid
;
1919 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
))
1927 * We fixed up user space. Now we need to fix the pi_state
1930 if (pi_state
->owner
!= NULL
) {
1931 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
1932 WARN_ON(list_empty(&pi_state
->list
));
1933 list_del_init(&pi_state
->list
);
1934 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
1937 pi_state
->owner
= newowner
;
1939 raw_spin_lock_irq(&newowner
->pi_lock
);
1940 WARN_ON(!list_empty(&pi_state
->list
));
1941 list_add(&pi_state
->list
, &newowner
->pi_state_list
);
1942 raw_spin_unlock_irq(&newowner
->pi_lock
);
1946 * To handle the page fault we need to drop the hash bucket
1947 * lock here. That gives the other task (either the highest priority
1948 * waiter itself or the task which stole the rtmutex) the
1949 * chance to try the fixup of the pi_state. So once we are
1950 * back from handling the fault we need to check the pi_state
1951 * after reacquiring the hash bucket lock and before trying to
1952 * do another fixup. When the fixup has been done already we
1956 spin_unlock(q
->lock_ptr
);
1958 ret
= fault_in_user_writeable(uaddr
);
1960 spin_lock(q
->lock_ptr
);
1963 * Check if someone else fixed it for us:
1965 if (pi_state
->owner
!= oldowner
)
1974 static long futex_wait_restart(struct restart_block
*restart
);
1977 * fixup_owner() - Post lock pi_state and corner case management
1978 * @uaddr: user address of the futex
1979 * @q: futex_q (contains pi_state and access to the rt_mutex)
1980 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1982 * After attempting to lock an rt_mutex, this function is called to cleanup
1983 * the pi_state owner as well as handle race conditions that may allow us to
1984 * acquire the lock. Must be called with the hb lock held.
1987 * 1 - success, lock taken;
1988 * 0 - success, lock not taken;
1989 * <0 - on error (-EFAULT)
1991 static int fixup_owner(u32 __user
*uaddr
, struct futex_q
*q
, int locked
)
1993 struct task_struct
*owner
;
1998 * Got the lock. We might not be the anticipated owner if we
1999 * did a lock-steal - fix up the PI-state in that case:
2001 if (q
->pi_state
->owner
!= current
)
2002 ret
= fixup_pi_state_owner(uaddr
, q
, current
);
2007 * Catch the rare case, where the lock was released when we were on the
2008 * way back before we locked the hash bucket.
2010 if (q
->pi_state
->owner
== current
) {
2012 * Try to get the rt_mutex now. This might fail as some other
2013 * task acquired the rt_mutex after we removed ourself from the
2014 * rt_mutex waiters list.
2016 if (rt_mutex_trylock(&q
->pi_state
->pi_mutex
)) {
2022 * pi_state is incorrect, some other task did a lock steal and
2023 * we returned due to timeout or signal without taking the
2024 * rt_mutex. Too late.
2026 raw_spin_lock(&q
->pi_state
->pi_mutex
.wait_lock
);
2027 owner
= rt_mutex_owner(&q
->pi_state
->pi_mutex
);
2029 owner
= rt_mutex_next_owner(&q
->pi_state
->pi_mutex
);
2030 raw_spin_unlock(&q
->pi_state
->pi_mutex
.wait_lock
);
2031 ret
= fixup_pi_state_owner(uaddr
, q
, owner
);
2036 * Paranoia check. If we did not take the lock, then we should not be
2037 * the owner of the rt_mutex.
2039 if (rt_mutex_owner(&q
->pi_state
->pi_mutex
) == current
)
2040 printk(KERN_ERR
"fixup_owner: ret = %d pi-mutex: %p "
2041 "pi-state %p\n", ret
,
2042 q
->pi_state
->pi_mutex
.owner
,
2043 q
->pi_state
->owner
);
2046 return ret
? ret
: locked
;
2050 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2051 * @hb: the futex hash bucket, must be locked by the caller
2052 * @q: the futex_q to queue up on
2053 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2055 static void futex_wait_queue_me(struct futex_hash_bucket
*hb
, struct futex_q
*q
,
2056 struct hrtimer_sleeper
*timeout
)
2059 * The task state is guaranteed to be set before another task can
2060 * wake it. set_current_state() is implemented using set_mb() and
2061 * queue_me() calls spin_unlock() upon completion, both serializing
2062 * access to the hash list and forcing another memory barrier.
2064 set_current_state(TASK_INTERRUPTIBLE
);
2069 hrtimer_start_expires(&timeout
->timer
, HRTIMER_MODE_ABS
);
2070 if (!hrtimer_active(&timeout
->timer
))
2071 timeout
->task
= NULL
;
2075 * If we have been removed from the hash list, then another task
2076 * has tried to wake us, and we can skip the call to schedule().
2078 if (likely(!plist_node_empty(&q
->list
))) {
2080 * If the timer has already expired, current will already be
2081 * flagged for rescheduling. Only call schedule if there
2082 * is no timeout, or if it has yet to expire.
2084 if (!timeout
|| timeout
->task
)
2085 freezable_schedule();
2087 __set_current_state(TASK_RUNNING
);
2091 * futex_wait_setup() - Prepare to wait on a futex
2092 * @uaddr: the futex userspace address
2093 * @val: the expected value
2094 * @flags: futex flags (FLAGS_SHARED, etc.)
2095 * @q: the associated futex_q
2096 * @hb: storage for hash_bucket pointer to be returned to caller
2098 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2099 * compare it with the expected value. Handle atomic faults internally.
2100 * Return with the hb lock held and a q.key reference on success, and unlocked
2101 * with no q.key reference on failure.
2104 * 0 - uaddr contains val and hb has been locked;
2105 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2107 static int futex_wait_setup(u32 __user
*uaddr
, u32 val
, unsigned int flags
,
2108 struct futex_q
*q
, struct futex_hash_bucket
**hb
)
2114 * Access the page AFTER the hash-bucket is locked.
2115 * Order is important:
2117 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2118 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2120 * The basic logical guarantee of a futex is that it blocks ONLY
2121 * if cond(var) is known to be true at the time of blocking, for
2122 * any cond. If we locked the hash-bucket after testing *uaddr, that
2123 * would open a race condition where we could block indefinitely with
2124 * cond(var) false, which would violate the guarantee.
2126 * On the other hand, we insert q and release the hash-bucket only
2127 * after testing *uaddr. This guarantees that futex_wait() will NOT
2128 * absorb a wakeup if *uaddr does not match the desired values
2129 * while the syscall executes.
2132 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
->key
, VERIFY_READ
);
2133 if (unlikely(ret
!= 0))
2137 *hb
= queue_lock(q
);
2139 ret
= get_futex_value_locked(&uval
, uaddr
);
2144 ret
= get_user(uval
, uaddr
);
2148 if (!(flags
& FLAGS_SHARED
))
2151 put_futex_key(&q
->key
);
2162 put_futex_key(&q
->key
);
2166 static int futex_wait(u32 __user
*uaddr
, unsigned int flags
, u32 val
,
2167 ktime_t
*abs_time
, u32 bitset
)
2169 struct hrtimer_sleeper timeout
, *to
= NULL
;
2170 struct restart_block
*restart
;
2171 struct futex_hash_bucket
*hb
;
2172 struct futex_q q
= futex_q_init
;
2182 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2183 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2185 hrtimer_init_sleeper(to
, current
);
2186 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2187 current
->timer_slack_ns
);
2192 * Prepare to wait on uaddr. On success, holds hb lock and increments
2195 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2199 /* queue_me and wait for wakeup, timeout, or a signal. */
2200 futex_wait_queue_me(hb
, &q
, to
);
2202 /* If we were woken (and unqueued), we succeeded, whatever. */
2204 /* unqueue_me() drops q.key ref */
2205 if (!unqueue_me(&q
))
2208 if (to
&& !to
->task
)
2212 * We expect signal_pending(current), but we might be the
2213 * victim of a spurious wakeup as well.
2215 if (!signal_pending(current
))
2222 restart
= ¤t_thread_info()->restart_block
;
2223 restart
->fn
= futex_wait_restart
;
2224 restart
->futex
.uaddr
= uaddr
;
2225 restart
->futex
.val
= val
;
2226 restart
->futex
.time
= abs_time
->tv64
;
2227 restart
->futex
.bitset
= bitset
;
2228 restart
->futex
.flags
= flags
| FLAGS_HAS_TIMEOUT
;
2230 ret
= -ERESTART_RESTARTBLOCK
;
2234 hrtimer_cancel(&to
->timer
);
2235 destroy_hrtimer_on_stack(&to
->timer
);
2241 static long futex_wait_restart(struct restart_block
*restart
)
2243 u32 __user
*uaddr
= restart
->futex
.uaddr
;
2244 ktime_t t
, *tp
= NULL
;
2246 if (restart
->futex
.flags
& FLAGS_HAS_TIMEOUT
) {
2247 t
.tv64
= restart
->futex
.time
;
2250 restart
->fn
= do_no_restart_syscall
;
2252 return (long)futex_wait(uaddr
, restart
->futex
.flags
,
2253 restart
->futex
.val
, tp
, restart
->futex
.bitset
);
2258 * Userspace tried a 0 -> TID atomic transition of the futex value
2259 * and failed. The kernel side here does the whole locking operation:
2260 * if there are waiters then it will block, it does PI, etc. (Due to
2261 * races the kernel might see a 0 value of the futex too.)
2263 static int futex_lock_pi(u32 __user
*uaddr
, unsigned int flags
, int detect
,
2264 ktime_t
*time
, int trylock
)
2266 struct hrtimer_sleeper timeout
, *to
= NULL
;
2267 struct futex_hash_bucket
*hb
;
2268 struct futex_q q
= futex_q_init
;
2271 if (refill_pi_state_cache())
2276 hrtimer_init_on_stack(&to
->timer
, CLOCK_REALTIME
,
2278 hrtimer_init_sleeper(to
, current
);
2279 hrtimer_set_expires(&to
->timer
, *time
);
2283 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
.key
, VERIFY_WRITE
);
2284 if (unlikely(ret
!= 0))
2288 hb
= queue_lock(&q
);
2290 ret
= futex_lock_pi_atomic(uaddr
, hb
, &q
.key
, &q
.pi_state
, current
, 0);
2291 if (unlikely(ret
)) {
2294 /* We got the lock. */
2296 goto out_unlock_put_key
;
2301 * Task is exiting and we just wait for the
2305 put_futex_key(&q
.key
);
2309 goto out_unlock_put_key
;
2314 * Only actually queue now that the atomic ops are done:
2318 WARN_ON(!q
.pi_state
);
2320 * Block on the PI mutex:
2323 ret
= rt_mutex_timed_futex_lock(&q
.pi_state
->pi_mutex
, to
);
2325 ret
= rt_mutex_trylock(&q
.pi_state
->pi_mutex
);
2326 /* Fixup the trylock return value: */
2327 ret
= ret
? 0 : -EWOULDBLOCK
;
2330 spin_lock(q
.lock_ptr
);
2332 * Fixup the pi_state owner and possibly acquire the lock if we
2335 res
= fixup_owner(uaddr
, &q
, !ret
);
2337 * If fixup_owner() returned an error, proprogate that. If it acquired
2338 * the lock, clear our -ETIMEDOUT or -EINTR.
2341 ret
= (res
< 0) ? res
: 0;
2344 * If fixup_owner() faulted and was unable to handle the fault, unlock
2345 * it and return the fault to userspace.
2347 if (ret
&& (rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
))
2348 rt_mutex_unlock(&q
.pi_state
->pi_mutex
);
2350 /* Unqueue and drop the lock */
2359 put_futex_key(&q
.key
);
2362 destroy_hrtimer_on_stack(&to
->timer
);
2363 return ret
!= -EINTR
? ret
: -ERESTARTNOINTR
;
2368 ret
= fault_in_user_writeable(uaddr
);
2372 if (!(flags
& FLAGS_SHARED
))
2375 put_futex_key(&q
.key
);
2380 * Userspace attempted a TID -> 0 atomic transition, and failed.
2381 * This is the in-kernel slowpath: we look up the PI state (if any),
2382 * and do the rt-mutex unlock.
2384 static int futex_unlock_pi(u32 __user
*uaddr
, unsigned int flags
)
2386 u32
uninitialized_var(curval
), uval
, vpid
= task_pid_vnr(current
);
2387 union futex_key key
= FUTEX_KEY_INIT
;
2388 struct futex_hash_bucket
*hb
;
2389 struct futex_q
*match
;
2393 if (get_user(uval
, uaddr
))
2396 * We release only a lock we actually own:
2398 if ((uval
& FUTEX_TID_MASK
) != vpid
)
2401 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_WRITE
);
2405 hb
= hash_futex(&key
);
2406 spin_lock(&hb
->lock
);
2409 * Check waiters first. We do not trust user space values at
2410 * all and we at least want to know if user space fiddled
2411 * with the futex value instead of blindly unlocking.
2413 match
= futex_top_waiter(hb
, &key
);
2415 ret
= wake_futex_pi(uaddr
, uval
, match
);
2417 * The atomic access to the futex value generated a
2418 * pagefault, so retry the user-access and the wakeup:
2426 * We have no kernel internal state, i.e. no waiters in the
2427 * kernel. Waiters which are about to queue themselves are stuck
2428 * on hb->lock. So we can safely ignore them. We do neither
2429 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2432 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, 0))
2436 * If uval has changed, let user space handle it.
2438 ret
= (curval
== uval
) ? 0 : -EAGAIN
;
2441 spin_unlock(&hb
->lock
);
2442 put_futex_key(&key
);
2446 spin_unlock(&hb
->lock
);
2447 put_futex_key(&key
);
2449 ret
= fault_in_user_writeable(uaddr
);
2457 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2458 * @hb: the hash_bucket futex_q was original enqueued on
2459 * @q: the futex_q woken while waiting to be requeued
2460 * @key2: the futex_key of the requeue target futex
2461 * @timeout: the timeout associated with the wait (NULL if none)
2463 * Detect if the task was woken on the initial futex as opposed to the requeue
2464 * target futex. If so, determine if it was a timeout or a signal that caused
2465 * the wakeup and return the appropriate error code to the caller. Must be
2466 * called with the hb lock held.
2469 * 0 = no early wakeup detected;
2470 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2473 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket
*hb
,
2474 struct futex_q
*q
, union futex_key
*key2
,
2475 struct hrtimer_sleeper
*timeout
)
2480 * With the hb lock held, we avoid races while we process the wakeup.
2481 * We only need to hold hb (and not hb2) to ensure atomicity as the
2482 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2483 * It can't be requeued from uaddr2 to something else since we don't
2484 * support a PI aware source futex for requeue.
2486 if (!match_futex(&q
->key
, key2
)) {
2487 WARN_ON(q
->lock_ptr
&& (&hb
->lock
!= q
->lock_ptr
));
2489 * We were woken prior to requeue by a timeout or a signal.
2490 * Unqueue the futex_q and determine which it was.
2492 plist_del(&q
->list
, &hb
->chain
);
2495 /* Handle spurious wakeups gracefully */
2497 if (timeout
&& !timeout
->task
)
2499 else if (signal_pending(current
))
2500 ret
= -ERESTARTNOINTR
;
2506 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2507 * @uaddr: the futex we initially wait on (non-pi)
2508 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2509 * the same type, no requeueing from private to shared, etc.
2510 * @val: the expected value of uaddr
2511 * @abs_time: absolute timeout
2512 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2513 * @uaddr2: the pi futex we will take prior to returning to user-space
2515 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2516 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2517 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2518 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2519 * without one, the pi logic would not know which task to boost/deboost, if
2520 * there was a need to.
2522 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2523 * via the following--
2524 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2525 * 2) wakeup on uaddr2 after a requeue
2529 * If 3, cleanup and return -ERESTARTNOINTR.
2531 * If 2, we may then block on trying to take the rt_mutex and return via:
2532 * 5) successful lock
2535 * 8) other lock acquisition failure
2537 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2539 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2545 static int futex_wait_requeue_pi(u32 __user
*uaddr
, unsigned int flags
,
2546 u32 val
, ktime_t
*abs_time
, u32 bitset
,
2549 struct hrtimer_sleeper timeout
, *to
= NULL
;
2550 struct rt_mutex_waiter rt_waiter
;
2551 struct rt_mutex
*pi_mutex
= NULL
;
2552 struct futex_hash_bucket
*hb
;
2553 union futex_key key2
= FUTEX_KEY_INIT
;
2554 struct futex_q q
= futex_q_init
;
2557 if (uaddr
== uaddr2
)
2565 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2566 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2568 hrtimer_init_sleeper(to
, current
);
2569 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2570 current
->timer_slack_ns
);
2574 * The waiter is allocated on our stack, manipulated by the requeue
2575 * code while we sleep on uaddr.
2577 debug_rt_mutex_init_waiter(&rt_waiter
);
2578 RB_CLEAR_NODE(&rt_waiter
.pi_tree_entry
);
2579 RB_CLEAR_NODE(&rt_waiter
.tree_entry
);
2580 rt_waiter
.task
= NULL
;
2582 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
2583 if (unlikely(ret
!= 0))
2587 q
.rt_waiter
= &rt_waiter
;
2588 q
.requeue_pi_key
= &key2
;
2591 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2594 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2599 * The check above which compares uaddrs is not sufficient for
2600 * shared futexes. We need to compare the keys:
2602 if (match_futex(&q
.key
, &key2
)) {
2607 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2608 futex_wait_queue_me(hb
, &q
, to
);
2610 spin_lock(&hb
->lock
);
2611 ret
= handle_early_requeue_pi_wakeup(hb
, &q
, &key2
, to
);
2612 spin_unlock(&hb
->lock
);
2617 * In order for us to be here, we know our q.key == key2, and since
2618 * we took the hb->lock above, we also know that futex_requeue() has
2619 * completed and we no longer have to concern ourselves with a wakeup
2620 * race with the atomic proxy lock acquisition by the requeue code. The
2621 * futex_requeue dropped our key1 reference and incremented our key2
2625 /* Check if the requeue code acquired the second futex for us. */
2628 * Got the lock. We might not be the anticipated owner if we
2629 * did a lock-steal - fix up the PI-state in that case.
2631 if (q
.pi_state
&& (q
.pi_state
->owner
!= current
)) {
2632 spin_lock(q
.lock_ptr
);
2633 ret
= fixup_pi_state_owner(uaddr2
, &q
, current
);
2634 spin_unlock(q
.lock_ptr
);
2638 * We have been woken up by futex_unlock_pi(), a timeout, or a
2639 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2642 WARN_ON(!q
.pi_state
);
2643 pi_mutex
= &q
.pi_state
->pi_mutex
;
2644 ret
= rt_mutex_finish_proxy_lock(pi_mutex
, to
, &rt_waiter
);
2645 debug_rt_mutex_free_waiter(&rt_waiter
);
2647 spin_lock(q
.lock_ptr
);
2649 * Fixup the pi_state owner and possibly acquire the lock if we
2652 res
= fixup_owner(uaddr2
, &q
, !ret
);
2654 * If fixup_owner() returned an error, proprogate that. If it
2655 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2658 ret
= (res
< 0) ? res
: 0;
2660 /* Unqueue and drop the lock. */
2665 * If fixup_pi_state_owner() faulted and was unable to handle the
2666 * fault, unlock the rt_mutex and return the fault to userspace.
2668 if (ret
== -EFAULT
) {
2669 if (pi_mutex
&& rt_mutex_owner(pi_mutex
) == current
)
2670 rt_mutex_unlock(pi_mutex
);
2671 } else if (ret
== -EINTR
) {
2673 * We've already been requeued, but cannot restart by calling
2674 * futex_lock_pi() directly. We could restart this syscall, but
2675 * it would detect that the user space "val" changed and return
2676 * -EWOULDBLOCK. Save the overhead of the restart and return
2677 * -EWOULDBLOCK directly.
2683 put_futex_key(&q
.key
);
2685 put_futex_key(&key2
);
2689 hrtimer_cancel(&to
->timer
);
2690 destroy_hrtimer_on_stack(&to
->timer
);
2696 * Support for robust futexes: the kernel cleans up held futexes at
2699 * Implementation: user-space maintains a per-thread list of locks it
2700 * is holding. Upon do_exit(), the kernel carefully walks this list,
2701 * and marks all locks that are owned by this thread with the
2702 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2703 * always manipulated with the lock held, so the list is private and
2704 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2705 * field, to allow the kernel to clean up if the thread dies after
2706 * acquiring the lock, but just before it could have added itself to
2707 * the list. There can only be one such pending lock.
2711 * sys_set_robust_list() - Set the robust-futex list head of a task
2712 * @head: pointer to the list-head
2713 * @len: length of the list-head, as userspace expects
2715 SYSCALL_DEFINE2(set_robust_list
, struct robust_list_head __user
*, head
,
2718 if (!futex_cmpxchg_enabled
)
2721 * The kernel knows only one size for now:
2723 if (unlikely(len
!= sizeof(*head
)))
2726 current
->robust_list
= head
;
2732 * sys_get_robust_list() - Get the robust-futex list head of a task
2733 * @pid: pid of the process [zero for current task]
2734 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2735 * @len_ptr: pointer to a length field, the kernel fills in the header size
2737 SYSCALL_DEFINE3(get_robust_list
, int, pid
,
2738 struct robust_list_head __user
* __user
*, head_ptr
,
2739 size_t __user
*, len_ptr
)
2741 struct robust_list_head __user
*head
;
2743 struct task_struct
*p
;
2745 if (!futex_cmpxchg_enabled
)
2754 p
= find_task_by_vpid(pid
);
2760 if (!ptrace_may_access(p
, PTRACE_MODE_READ
))
2763 head
= p
->robust_list
;
2766 if (put_user(sizeof(*head
), len_ptr
))
2768 return put_user(head
, head_ptr
);
2777 * Process a futex-list entry, check whether it's owned by the
2778 * dying task, and do notification if so:
2780 int handle_futex_death(u32 __user
*uaddr
, struct task_struct
*curr
, int pi
)
2782 u32 uval
, uninitialized_var(nval
), mval
;
2785 if (get_user(uval
, uaddr
))
2788 if ((uval
& FUTEX_TID_MASK
) == task_pid_vnr(curr
)) {
2790 * Ok, this dying thread is truly holding a futex
2791 * of interest. Set the OWNER_DIED bit atomically
2792 * via cmpxchg, and if the value had FUTEX_WAITERS
2793 * set, wake up a waiter (if any). (We have to do a
2794 * futex_wake() even if OWNER_DIED is already set -
2795 * to handle the rare but possible case of recursive
2796 * thread-death.) The rest of the cleanup is done in
2799 mval
= (uval
& FUTEX_WAITERS
) | FUTEX_OWNER_DIED
;
2801 * We are not holding a lock here, but we want to have
2802 * the pagefault_disable/enable() protection because
2803 * we want to handle the fault gracefully. If the
2804 * access fails we try to fault in the futex with R/W
2805 * verification via get_user_pages. get_user() above
2806 * does not guarantee R/W access. If that fails we
2807 * give up and leave the futex locked.
2809 if (cmpxchg_futex_value_locked(&nval
, uaddr
, uval
, mval
)) {
2810 if (fault_in_user_writeable(uaddr
))
2818 * Wake robust non-PI futexes here. The wakeup of
2819 * PI futexes happens in exit_pi_state():
2821 if (!pi
&& (uval
& FUTEX_WAITERS
))
2822 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
2828 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2830 static inline int fetch_robust_entry(struct robust_list __user
**entry
,
2831 struct robust_list __user
* __user
*head
,
2834 unsigned long uentry
;
2836 if (get_user(uentry
, (unsigned long __user
*)head
))
2839 *entry
= (void __user
*)(uentry
& ~1UL);
2846 * Walk curr->robust_list (very carefully, it's a userspace list!)
2847 * and mark any locks found there dead, and notify any waiters.
2849 * We silently return on any sign of list-walking problem.
2851 void exit_robust_list(struct task_struct
*curr
)
2853 struct robust_list_head __user
*head
= curr
->robust_list
;
2854 struct robust_list __user
*entry
, *next_entry
, *pending
;
2855 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
2856 unsigned int uninitialized_var(next_pi
);
2857 unsigned long futex_offset
;
2860 if (!futex_cmpxchg_enabled
)
2864 * Fetch the list head (which was registered earlier, via
2865 * sys_set_robust_list()):
2867 if (fetch_robust_entry(&entry
, &head
->list
.next
, &pi
))
2870 * Fetch the relative futex offset:
2872 if (get_user(futex_offset
, &head
->futex_offset
))
2875 * Fetch any possibly pending lock-add first, and handle it
2878 if (fetch_robust_entry(&pending
, &head
->list_op_pending
, &pip
))
2881 next_entry
= NULL
; /* avoid warning with gcc */
2882 while (entry
!= &head
->list
) {
2884 * Fetch the next entry in the list before calling
2885 * handle_futex_death:
2887 rc
= fetch_robust_entry(&next_entry
, &entry
->next
, &next_pi
);
2889 * A pending lock might already be on the list, so
2890 * don't process it twice:
2892 if (entry
!= pending
)
2893 if (handle_futex_death((void __user
*)entry
+ futex_offset
,
2901 * Avoid excessively long or circular lists:
2910 handle_futex_death((void __user
*)pending
+ futex_offset
,
2914 long do_futex(u32 __user
*uaddr
, int op
, u32 val
, ktime_t
*timeout
,
2915 u32 __user
*uaddr2
, u32 val2
, u32 val3
)
2917 int cmd
= op
& FUTEX_CMD_MASK
;
2918 unsigned int flags
= 0;
2920 if (!(op
& FUTEX_PRIVATE_FLAG
))
2921 flags
|= FLAGS_SHARED
;
2923 if (op
& FUTEX_CLOCK_REALTIME
) {
2924 flags
|= FLAGS_CLOCKRT
;
2925 if (cmd
!= FUTEX_WAIT_BITSET
&& cmd
!= FUTEX_WAIT_REQUEUE_PI
)
2931 case FUTEX_UNLOCK_PI
:
2932 case FUTEX_TRYLOCK_PI
:
2933 case FUTEX_WAIT_REQUEUE_PI
:
2934 case FUTEX_CMP_REQUEUE_PI
:
2935 if (!futex_cmpxchg_enabled
)
2941 val3
= FUTEX_BITSET_MATCH_ANY
;
2942 case FUTEX_WAIT_BITSET
:
2943 return futex_wait(uaddr
, flags
, val
, timeout
, val3
);
2945 val3
= FUTEX_BITSET_MATCH_ANY
;
2946 case FUTEX_WAKE_BITSET
:
2947 return futex_wake(uaddr
, flags
, val
, val3
);
2949 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, NULL
, 0);
2950 case FUTEX_CMP_REQUEUE
:
2951 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 0);
2953 return futex_wake_op(uaddr
, flags
, uaddr2
, val
, val2
, val3
);
2955 return futex_lock_pi(uaddr
, flags
, val
, timeout
, 0);
2956 case FUTEX_UNLOCK_PI
:
2957 return futex_unlock_pi(uaddr
, flags
);
2958 case FUTEX_TRYLOCK_PI
:
2959 return futex_lock_pi(uaddr
, flags
, 0, timeout
, 1);
2960 case FUTEX_WAIT_REQUEUE_PI
:
2961 val3
= FUTEX_BITSET_MATCH_ANY
;
2962 return futex_wait_requeue_pi(uaddr
, flags
, val
, timeout
, val3
,
2964 case FUTEX_CMP_REQUEUE_PI
:
2965 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 1);
2971 SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
2972 struct timespec __user
*, utime
, u32 __user
*, uaddr2
,
2976 ktime_t t
, *tp
= NULL
;
2978 int cmd
= op
& FUTEX_CMD_MASK
;
2980 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
2981 cmd
== FUTEX_WAIT_BITSET
||
2982 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
2983 if (copy_from_user(&ts
, utime
, sizeof(ts
)) != 0)
2985 if (!timespec_valid(&ts
))
2988 t
= timespec_to_ktime(ts
);
2989 if (cmd
== FUTEX_WAIT
)
2990 t
= ktime_add_safe(ktime_get(), t
);
2994 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2995 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2997 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
2998 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
2999 val2
= (u32
) (unsigned long) utime
;
3001 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
3004 static void __init
futex_detect_cmpxchg(void)
3006 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3010 * This will fail and we want it. Some arch implementations do
3011 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3012 * functionality. We want to know that before we call in any
3013 * of the complex code paths. Also we want to prevent
3014 * registration of robust lists in that case. NULL is
3015 * guaranteed to fault and we get -EFAULT on functional
3016 * implementation, the non-functional ones will return
3019 if (cmpxchg_futex_value_locked(&curval
, NULL
, 0, 0) == -EFAULT
)
3020 futex_cmpxchg_enabled
= 1;
3024 static int __init
futex_init(void)
3026 unsigned int futex_shift
;
3029 #if CONFIG_BASE_SMALL
3030 futex_hashsize
= 16;
3032 futex_hashsize
= roundup_pow_of_two(256 * num_possible_cpus());
3035 futex_queues
= alloc_large_system_hash("futex", sizeof(*futex_queues
),
3037 futex_hashsize
< 256 ? HASH_SMALL
: 0,
3039 futex_hashsize
, futex_hashsize
);
3040 futex_hashsize
= 1UL << futex_shift
;
3042 futex_detect_cmpxchg();
3044 for (i
= 0; i
< futex_hashsize
; i
++) {
3045 atomic_set(&futex_queues
[i
].waiters
, 0);
3046 plist_head_init(&futex_queues
[i
].chain
);
3047 spin_lock_init(&futex_queues
[i
].lock
);
3052 __initcall(futex_init
);