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>
67 #include <linux/fault-inject.h>
69 #include <asm/futex.h>
71 #include "locking/rtmutex_common.h"
74 * READ this before attempting to hack on futexes!
76 * Basic futex operation and ordering guarantees
77 * =============================================
79 * The waiter reads the futex value in user space and calls
80 * futex_wait(). This function computes the hash bucket and acquires
81 * the hash bucket lock. After that it reads the futex user space value
82 * again and verifies that the data has not changed. If it has not changed
83 * it enqueues itself into the hash bucket, releases the hash bucket lock
86 * The waker side modifies the user space value of the futex and calls
87 * futex_wake(). This function computes the hash bucket and acquires the
88 * hash bucket lock. Then it looks for waiters on that futex in the hash
89 * bucket and wakes them.
91 * In futex wake up scenarios where no tasks are blocked on a futex, taking
92 * the hb spinlock can be avoided and simply return. In order for this
93 * optimization to work, ordering guarantees must exist so that the waiter
94 * being added to the list is acknowledged when the list is concurrently being
95 * checked by the waker, avoiding scenarios like the following:
99 * sys_futex(WAIT, futex, val);
100 * futex_wait(futex, val);
103 * sys_futex(WAKE, futex);
108 * lock(hash_bucket(futex));
110 * unlock(hash_bucket(futex));
113 * This would cause the waiter on CPU 0 to wait forever because it
114 * missed the transition of the user space value from val to newval
115 * and the waker did not find the waiter in the hash bucket queue.
117 * The correct serialization ensures that a waiter either observes
118 * the changed user space value before blocking or is woken by a
123 * sys_futex(WAIT, futex, val);
124 * futex_wait(futex, val);
127 * mb(); (A) <-- paired with -.
129 * lock(hash_bucket(futex)); |
133 * | sys_futex(WAKE, futex);
134 * | futex_wake(futex);
136 * `-------> mb(); (B)
139 * unlock(hash_bucket(futex));
140 * schedule(); if (waiters)
141 * lock(hash_bucket(futex));
142 * else wake_waiters(futex);
143 * waiters--; (b) unlock(hash_bucket(futex));
145 * Where (A) orders the waiters increment and the futex value read through
146 * atomic operations (see hb_waiters_inc) and where (B) orders the write
147 * to futex and the waiters read -- this is done by the barriers for both
148 * shared and private futexes in get_futex_key_refs().
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
;
259 * The base of the bucket array and its size are always used together
260 * (after initialization only in hash_futex()), so ensure that they
261 * reside in the same cacheline.
264 struct futex_hash_bucket
*queues
;
265 unsigned long hashsize
;
266 } __futex_data __read_mostly
__aligned(2*sizeof(long));
267 #define futex_queues (__futex_data.queues)
268 #define futex_hashsize (__futex_data.hashsize)
272 * Fault injections for futexes.
274 #ifdef CONFIG_FAIL_FUTEX
277 struct fault_attr attr
;
281 .attr
= FAULT_ATTR_INITIALIZER
,
282 .ignore_private
= false,
285 static int __init
setup_fail_futex(char *str
)
287 return setup_fault_attr(&fail_futex
.attr
, str
);
289 __setup("fail_futex=", setup_fail_futex
);
291 static bool should_fail_futex(bool fshared
)
293 if (fail_futex
.ignore_private
&& !fshared
)
296 return should_fail(&fail_futex
.attr
, 1);
299 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
301 static int __init
fail_futex_debugfs(void)
303 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
306 dir
= fault_create_debugfs_attr("fail_futex", NULL
,
311 if (!debugfs_create_bool("ignore-private", mode
, dir
,
312 &fail_futex
.ignore_private
)) {
313 debugfs_remove_recursive(dir
);
320 late_initcall(fail_futex_debugfs
);
322 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
325 static inline bool should_fail_futex(bool fshared
)
329 #endif /* CONFIG_FAIL_FUTEX */
331 static inline void futex_get_mm(union futex_key
*key
)
333 atomic_inc(&key
->private.mm
->mm_count
);
335 * Ensure futex_get_mm() implies a full barrier such that
336 * get_futex_key() implies a full barrier. This is relied upon
337 * as full barrier (B), see the ordering comment above.
339 smp_mb__after_atomic();
343 * Reflects a new waiter being added to the waitqueue.
345 static inline void hb_waiters_inc(struct futex_hash_bucket
*hb
)
348 atomic_inc(&hb
->waiters
);
350 * Full barrier (A), see the ordering comment above.
352 smp_mb__after_atomic();
357 * Reflects a waiter being removed from the waitqueue by wakeup
360 static inline void hb_waiters_dec(struct futex_hash_bucket
*hb
)
363 atomic_dec(&hb
->waiters
);
367 static inline int hb_waiters_pending(struct futex_hash_bucket
*hb
)
370 return atomic_read(&hb
->waiters
);
377 * We hash on the keys returned from get_futex_key (see below).
379 static struct futex_hash_bucket
*hash_futex(union futex_key
*key
)
381 u32 hash
= jhash2((u32
*)&key
->both
.word
,
382 (sizeof(key
->both
.word
)+sizeof(key
->both
.ptr
))/4,
384 return &futex_queues
[hash
& (futex_hashsize
- 1)];
388 * Return 1 if two futex_keys are equal, 0 otherwise.
390 static inline int match_futex(union futex_key
*key1
, union futex_key
*key2
)
393 && key1
->both
.word
== key2
->both
.word
394 && key1
->both
.ptr
== key2
->both
.ptr
395 && key1
->both
.offset
== key2
->both
.offset
);
399 * Take a reference to the resource addressed by a key.
400 * Can be called while holding spinlocks.
403 static void get_futex_key_refs(union futex_key
*key
)
408 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
410 ihold(key
->shared
.inode
); /* implies MB (B) */
412 case FUT_OFF_MMSHARED
:
413 futex_get_mm(key
); /* implies MB (B) */
417 * Private futexes do not hold reference on an inode or
418 * mm, therefore the only purpose of calling get_futex_key_refs
419 * is because we need the barrier for the lockless waiter check.
421 smp_mb(); /* explicit MB (B) */
426 * Drop a reference to the resource addressed by a key.
427 * The hash bucket spinlock must not be held. This is
428 * a no-op for private futexes, see comment in the get
431 static void drop_futex_key_refs(union futex_key
*key
)
433 if (!key
->both
.ptr
) {
434 /* If we're here then we tried to put a key we failed to get */
439 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
441 iput(key
->shared
.inode
);
443 case FUT_OFF_MMSHARED
:
444 mmdrop(key
->private.mm
);
450 * get_futex_key() - Get parameters which are the keys for a futex
451 * @uaddr: virtual address of the futex
452 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
453 * @key: address where result is stored.
454 * @rw: mapping needs to be read/write (values: VERIFY_READ,
457 * Return: a negative error code or 0
459 * The key words are stored in *key on success.
461 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
462 * offset_within_page). For private mappings, it's (uaddr, current->mm).
463 * We can usually work out the index without swapping in the page.
465 * lock_page() might sleep, the caller should not hold a spinlock.
468 get_futex_key(u32 __user
*uaddr
, int fshared
, union futex_key
*key
, int rw
)
470 unsigned long address
= (unsigned long)uaddr
;
471 struct mm_struct
*mm
= current
->mm
;
472 struct page
*page
, *page_head
;
476 * The futex address must be "naturally" aligned.
478 key
->both
.offset
= address
% PAGE_SIZE
;
479 if (unlikely((address
% sizeof(u32
)) != 0))
481 address
-= key
->both
.offset
;
483 if (unlikely(!access_ok(rw
, uaddr
, sizeof(u32
))))
486 if (unlikely(should_fail_futex(fshared
)))
490 * PROCESS_PRIVATE futexes are fast.
491 * As the mm cannot disappear under us and the 'key' only needs
492 * virtual address, we dont even have to find the underlying vma.
493 * Note : We do have to check 'uaddr' is a valid user address,
494 * but access_ok() should be faster than find_vma()
497 key
->private.mm
= mm
;
498 key
->private.address
= address
;
499 get_futex_key_refs(key
); /* implies MB (B) */
504 /* Ignore any VERIFY_READ mapping (futex common case) */
505 if (unlikely(should_fail_futex(fshared
)))
508 err
= get_user_pages_fast(address
, 1, 1, &page
);
510 * If write access is not required (eg. FUTEX_WAIT), try
511 * and get read-only access.
513 if (err
== -EFAULT
&& rw
== VERIFY_READ
) {
514 err
= get_user_pages_fast(address
, 1, 0, &page
);
522 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
524 if (unlikely(PageTail(page
))) {
526 /* serialize against __split_huge_page_splitting() */
528 if (likely(__get_user_pages_fast(address
, 1, !ro
, &page
) == 1)) {
529 page_head
= compound_head(page
);
531 * page_head is valid pointer but we must pin
532 * it before taking the PG_lock and/or
533 * PG_compound_lock. The moment we re-enable
534 * irqs __split_huge_page_splitting() can
535 * return and the head page can be freed from
536 * under us. We can't take the PG_lock and/or
537 * PG_compound_lock on a page that could be
538 * freed from under us.
540 if (page
!= page_head
) {
551 page_head
= compound_head(page
);
552 if (page
!= page_head
) {
558 lock_page(page_head
);
561 * If page_head->mapping is NULL, then it cannot be a PageAnon
562 * page; but it might be the ZERO_PAGE or in the gate area or
563 * in a special mapping (all cases which we are happy to fail);
564 * or it may have been a good file page when get_user_pages_fast
565 * found it, but truncated or holepunched or subjected to
566 * invalidate_complete_page2 before we got the page lock (also
567 * cases which we are happy to fail). And we hold a reference,
568 * so refcount care in invalidate_complete_page's remove_mapping
569 * prevents drop_caches from setting mapping to NULL beneath us.
571 * The case we do have to guard against is when memory pressure made
572 * shmem_writepage move it from filecache to swapcache beneath us:
573 * an unlikely race, but we do need to retry for page_head->mapping.
575 if (!page_head
->mapping
) {
576 int shmem_swizzled
= PageSwapCache(page_head
);
577 unlock_page(page_head
);
585 * Private mappings are handled in a simple way.
587 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
588 * it's a read-only handle, it's expected that futexes attach to
589 * the object not the particular process.
591 if (PageAnon(page_head
)) {
593 * A RO anonymous page will never change and thus doesn't make
594 * sense for futex operations.
596 if (unlikely(should_fail_futex(fshared
)) || ro
) {
601 key
->both
.offset
|= FUT_OFF_MMSHARED
; /* ref taken on mm */
602 key
->private.mm
= mm
;
603 key
->private.address
= address
;
605 key
->both
.offset
|= FUT_OFF_INODE
; /* inode-based key */
606 key
->shared
.inode
= page_head
->mapping
->host
;
607 key
->shared
.pgoff
= basepage_index(page
);
610 get_futex_key_refs(key
); /* implies MB (B) */
613 unlock_page(page_head
);
618 static inline void put_futex_key(union futex_key
*key
)
620 drop_futex_key_refs(key
);
624 * fault_in_user_writeable() - Fault in user address and verify RW access
625 * @uaddr: pointer to faulting user space address
627 * Slow path to fixup the fault we just took in the atomic write
630 * We have no generic implementation of a non-destructive write to the
631 * user address. We know that we faulted in the atomic pagefault
632 * disabled section so we can as well avoid the #PF overhead by
633 * calling get_user_pages() right away.
635 static int fault_in_user_writeable(u32 __user
*uaddr
)
637 struct mm_struct
*mm
= current
->mm
;
640 down_read(&mm
->mmap_sem
);
641 ret
= fixup_user_fault(current
, mm
, (unsigned long)uaddr
,
643 up_read(&mm
->mmap_sem
);
645 return ret
< 0 ? ret
: 0;
649 * futex_top_waiter() - Return the highest priority waiter on a futex
650 * @hb: the hash bucket the futex_q's reside in
651 * @key: the futex key (to distinguish it from other futex futex_q's)
653 * Must be called with the hb lock held.
655 static struct futex_q
*futex_top_waiter(struct futex_hash_bucket
*hb
,
656 union futex_key
*key
)
658 struct futex_q
*this;
660 plist_for_each_entry(this, &hb
->chain
, list
) {
661 if (match_futex(&this->key
, key
))
667 static int cmpxchg_futex_value_locked(u32
*curval
, u32 __user
*uaddr
,
668 u32 uval
, u32 newval
)
673 ret
= futex_atomic_cmpxchg_inatomic(curval
, uaddr
, uval
, newval
);
679 static int get_futex_value_locked(u32
*dest
, u32 __user
*from
)
684 ret
= __copy_from_user_inatomic(dest
, from
, sizeof(u32
));
687 return ret
? -EFAULT
: 0;
694 static int refill_pi_state_cache(void)
696 struct futex_pi_state
*pi_state
;
698 if (likely(current
->pi_state_cache
))
701 pi_state
= kzalloc(sizeof(*pi_state
), GFP_KERNEL
);
706 INIT_LIST_HEAD(&pi_state
->list
);
707 /* pi_mutex gets initialized later */
708 pi_state
->owner
= NULL
;
709 atomic_set(&pi_state
->refcount
, 1);
710 pi_state
->key
= FUTEX_KEY_INIT
;
712 current
->pi_state_cache
= pi_state
;
717 static struct futex_pi_state
* alloc_pi_state(void)
719 struct futex_pi_state
*pi_state
= current
->pi_state_cache
;
722 current
->pi_state_cache
= NULL
;
728 * Must be called with the hb lock held.
730 static void free_pi_state(struct futex_pi_state
*pi_state
)
735 if (!atomic_dec_and_test(&pi_state
->refcount
))
739 * If pi_state->owner is NULL, the owner is most probably dying
740 * and has cleaned up the pi_state already
742 if (pi_state
->owner
) {
743 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
744 list_del_init(&pi_state
->list
);
745 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
747 rt_mutex_proxy_unlock(&pi_state
->pi_mutex
, pi_state
->owner
);
750 if (current
->pi_state_cache
)
754 * pi_state->list is already empty.
755 * clear pi_state->owner.
756 * refcount is at 0 - put it back to 1.
758 pi_state
->owner
= NULL
;
759 atomic_set(&pi_state
->refcount
, 1);
760 current
->pi_state_cache
= pi_state
;
765 * Look up the task based on what TID userspace gave us.
768 static struct task_struct
* futex_find_get_task(pid_t pid
)
770 struct task_struct
*p
;
773 p
= find_task_by_vpid(pid
);
783 * This task is holding PI mutexes at exit time => bad.
784 * Kernel cleans up PI-state, but userspace is likely hosed.
785 * (Robust-futex cleanup is separate and might save the day for userspace.)
787 void exit_pi_state_list(struct task_struct
*curr
)
789 struct list_head
*next
, *head
= &curr
->pi_state_list
;
790 struct futex_pi_state
*pi_state
;
791 struct futex_hash_bucket
*hb
;
792 union futex_key key
= FUTEX_KEY_INIT
;
794 if (!futex_cmpxchg_enabled
)
797 * We are a ZOMBIE and nobody can enqueue itself on
798 * pi_state_list anymore, but we have to be careful
799 * versus waiters unqueueing themselves:
801 raw_spin_lock_irq(&curr
->pi_lock
);
802 while (!list_empty(head
)) {
805 pi_state
= list_entry(next
, struct futex_pi_state
, list
);
807 hb
= hash_futex(&key
);
808 raw_spin_unlock_irq(&curr
->pi_lock
);
810 spin_lock(&hb
->lock
);
812 raw_spin_lock_irq(&curr
->pi_lock
);
814 * We dropped the pi-lock, so re-check whether this
815 * task still owns the PI-state:
817 if (head
->next
!= next
) {
818 spin_unlock(&hb
->lock
);
822 WARN_ON(pi_state
->owner
!= curr
);
823 WARN_ON(list_empty(&pi_state
->list
));
824 list_del_init(&pi_state
->list
);
825 pi_state
->owner
= NULL
;
826 raw_spin_unlock_irq(&curr
->pi_lock
);
828 rt_mutex_unlock(&pi_state
->pi_mutex
);
830 spin_unlock(&hb
->lock
);
832 raw_spin_lock_irq(&curr
->pi_lock
);
834 raw_spin_unlock_irq(&curr
->pi_lock
);
838 * We need to check the following states:
840 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
842 * [1] NULL | --- | --- | 0 | 0/1 | Valid
843 * [2] NULL | --- | --- | >0 | 0/1 | Valid
845 * [3] Found | NULL | -- | Any | 0/1 | Invalid
847 * [4] Found | Found | NULL | 0 | 1 | Valid
848 * [5] Found | Found | NULL | >0 | 1 | Invalid
850 * [6] Found | Found | task | 0 | 1 | Valid
852 * [7] Found | Found | NULL | Any | 0 | Invalid
854 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
855 * [9] Found | Found | task | 0 | 0 | Invalid
856 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
858 * [1] Indicates that the kernel can acquire the futex atomically. We
859 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
861 * [2] Valid, if TID does not belong to a kernel thread. If no matching
862 * thread is found then it indicates that the owner TID has died.
864 * [3] Invalid. The waiter is queued on a non PI futex
866 * [4] Valid state after exit_robust_list(), which sets the user space
867 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
869 * [5] The user space value got manipulated between exit_robust_list()
870 * and exit_pi_state_list()
872 * [6] Valid state after exit_pi_state_list() which sets the new owner in
873 * the pi_state but cannot access the user space value.
875 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
877 * [8] Owner and user space value match
879 * [9] There is no transient state which sets the user space TID to 0
880 * except exit_robust_list(), but this is indicated by the
881 * FUTEX_OWNER_DIED bit. See [4]
883 * [10] There is no transient state which leaves owner and user space
888 * Validate that the existing waiter has a pi_state and sanity check
889 * the pi_state against the user space value. If correct, attach to
892 static int attach_to_pi_state(u32 uval
, struct futex_pi_state
*pi_state
,
893 struct futex_pi_state
**ps
)
895 pid_t pid
= uval
& FUTEX_TID_MASK
;
898 * Userspace might have messed up non-PI and PI futexes [3]
900 if (unlikely(!pi_state
))
903 WARN_ON(!atomic_read(&pi_state
->refcount
));
906 * Handle the owner died case:
908 if (uval
& FUTEX_OWNER_DIED
) {
910 * exit_pi_state_list sets owner to NULL and wakes the
911 * topmost waiter. The task which acquires the
912 * pi_state->rt_mutex will fixup owner.
914 if (!pi_state
->owner
) {
916 * No pi state owner, but the user space TID
917 * is not 0. Inconsistent state. [5]
922 * Take a ref on the state and return success. [4]
928 * If TID is 0, then either the dying owner has not
929 * yet executed exit_pi_state_list() or some waiter
930 * acquired the rtmutex in the pi state, but did not
931 * yet fixup the TID in user space.
933 * Take a ref on the state and return success. [6]
939 * If the owner died bit is not set, then the pi_state
940 * must have an owner. [7]
942 if (!pi_state
->owner
)
947 * Bail out if user space manipulated the futex value. If pi
948 * state exists then the owner TID must be the same as the
949 * user space TID. [9/10]
951 if (pid
!= task_pid_vnr(pi_state
->owner
))
954 atomic_inc(&pi_state
->refcount
);
960 * Lookup the task for the TID provided from user space and attach to
961 * it after doing proper sanity checks.
963 static int attach_to_pi_owner(u32 uval
, union futex_key
*key
,
964 struct futex_pi_state
**ps
)
966 pid_t pid
= uval
& FUTEX_TID_MASK
;
967 struct futex_pi_state
*pi_state
;
968 struct task_struct
*p
;
971 * We are the first waiter - try to look up the real owner and attach
972 * the new pi_state to it, but bail out when TID = 0 [1]
976 p
= futex_find_get_task(pid
);
980 if (unlikely(p
->flags
& PF_KTHREAD
)) {
986 * We need to look at the task state flags to figure out,
987 * whether the task is exiting. To protect against the do_exit
988 * change of the task flags, we do this protected by
991 raw_spin_lock_irq(&p
->pi_lock
);
992 if (unlikely(p
->flags
& PF_EXITING
)) {
994 * The task is on the way out. When PF_EXITPIDONE is
995 * set, we know that the task has finished the
998 int ret
= (p
->flags
& PF_EXITPIDONE
) ? -ESRCH
: -EAGAIN
;
1000 raw_spin_unlock_irq(&p
->pi_lock
);
1006 * No existing pi state. First waiter. [2]
1008 pi_state
= alloc_pi_state();
1011 * Initialize the pi_mutex in locked state and make @p
1014 rt_mutex_init_proxy_locked(&pi_state
->pi_mutex
, p
);
1016 /* Store the key for possible exit cleanups: */
1017 pi_state
->key
= *key
;
1019 WARN_ON(!list_empty(&pi_state
->list
));
1020 list_add(&pi_state
->list
, &p
->pi_state_list
);
1021 pi_state
->owner
= p
;
1022 raw_spin_unlock_irq(&p
->pi_lock
);
1031 static int lookup_pi_state(u32 uval
, struct futex_hash_bucket
*hb
,
1032 union futex_key
*key
, struct futex_pi_state
**ps
)
1034 struct futex_q
*match
= futex_top_waiter(hb
, key
);
1037 * If there is a waiter on that futex, validate it and
1038 * attach to the pi_state when the validation succeeds.
1041 return attach_to_pi_state(uval
, match
->pi_state
, ps
);
1044 * We are the first waiter - try to look up the owner based on
1045 * @uval and attach to it.
1047 return attach_to_pi_owner(uval
, key
, ps
);
1050 static int lock_pi_update_atomic(u32 __user
*uaddr
, u32 uval
, u32 newval
)
1052 u32
uninitialized_var(curval
);
1054 if (unlikely(should_fail_futex(true)))
1057 if (unlikely(cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
)))
1060 /*If user space value changed, let the caller retry */
1061 return curval
!= uval
? -EAGAIN
: 0;
1065 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1066 * @uaddr: the pi futex user address
1067 * @hb: the pi futex hash bucket
1068 * @key: the futex key associated with uaddr and hb
1069 * @ps: the pi_state pointer where we store the result of the
1071 * @task: the task to perform the atomic lock work for. This will
1072 * be "current" except in the case of requeue pi.
1073 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1076 * 0 - ready to wait;
1077 * 1 - acquired the lock;
1080 * The hb->lock and futex_key refs shall be held by the caller.
1082 static int futex_lock_pi_atomic(u32 __user
*uaddr
, struct futex_hash_bucket
*hb
,
1083 union futex_key
*key
,
1084 struct futex_pi_state
**ps
,
1085 struct task_struct
*task
, int set_waiters
)
1087 u32 uval
, newval
, vpid
= task_pid_vnr(task
);
1088 struct futex_q
*match
;
1092 * Read the user space value first so we can validate a few
1093 * things before proceeding further.
1095 if (get_futex_value_locked(&uval
, uaddr
))
1098 if (unlikely(should_fail_futex(true)))
1104 if ((unlikely((uval
& FUTEX_TID_MASK
) == vpid
)))
1107 if ((unlikely(should_fail_futex(true))))
1111 * Lookup existing state first. If it exists, try to attach to
1114 match
= futex_top_waiter(hb
, key
);
1116 return attach_to_pi_state(uval
, match
->pi_state
, ps
);
1119 * No waiter and user TID is 0. We are here because the
1120 * waiters or the owner died bit is set or called from
1121 * requeue_cmp_pi or for whatever reason something took the
1124 if (!(uval
& FUTEX_TID_MASK
)) {
1126 * We take over the futex. No other waiters and the user space
1127 * TID is 0. We preserve the owner died bit.
1129 newval
= uval
& FUTEX_OWNER_DIED
;
1132 /* The futex requeue_pi code can enforce the waiters bit */
1134 newval
|= FUTEX_WAITERS
;
1136 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1137 /* If the take over worked, return 1 */
1138 return ret
< 0 ? ret
: 1;
1142 * First waiter. Set the waiters bit before attaching ourself to
1143 * the owner. If owner tries to unlock, it will be forced into
1144 * the kernel and blocked on hb->lock.
1146 newval
= uval
| FUTEX_WAITERS
;
1147 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1151 * If the update of the user space value succeeded, we try to
1152 * attach to the owner. If that fails, no harm done, we only
1153 * set the FUTEX_WAITERS bit in the user space variable.
1155 return attach_to_pi_owner(uval
, key
, ps
);
1159 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1160 * @q: The futex_q to unqueue
1162 * The q->lock_ptr must not be NULL and must be held by the caller.
1164 static void __unqueue_futex(struct futex_q
*q
)
1166 struct futex_hash_bucket
*hb
;
1168 if (WARN_ON_SMP(!q
->lock_ptr
|| !spin_is_locked(q
->lock_ptr
))
1169 || WARN_ON(plist_node_empty(&q
->list
)))
1172 hb
= container_of(q
->lock_ptr
, struct futex_hash_bucket
, lock
);
1173 plist_del(&q
->list
, &hb
->chain
);
1178 * The hash bucket lock must be held when this is called.
1179 * Afterwards, the futex_q must not be accessed. Callers
1180 * must ensure to later call wake_up_q() for the actual
1183 static void mark_wake_futex(struct wake_q_head
*wake_q
, struct futex_q
*q
)
1185 struct task_struct
*p
= q
->task
;
1187 if (WARN(q
->pi_state
|| q
->rt_waiter
, "refusing to wake PI futex\n"))
1191 * Queue the task for later wakeup for after we've released
1192 * the hb->lock. wake_q_add() grabs reference to p.
1194 wake_q_add(wake_q
, p
);
1197 * The waiting task can free the futex_q as soon as
1198 * q->lock_ptr = NULL is written, without taking any locks. A
1199 * memory barrier is required here to prevent the following
1200 * store to lock_ptr from getting ahead of the plist_del.
1206 static int wake_futex_pi(u32 __user
*uaddr
, u32 uval
, struct futex_q
*this,
1207 struct futex_hash_bucket
*hb
)
1209 struct task_struct
*new_owner
;
1210 struct futex_pi_state
*pi_state
= this->pi_state
;
1211 u32
uninitialized_var(curval
), newval
;
1220 * If current does not own the pi_state then the futex is
1221 * inconsistent and user space fiddled with the futex value.
1223 if (pi_state
->owner
!= current
)
1226 raw_spin_lock(&pi_state
->pi_mutex
.wait_lock
);
1227 new_owner
= rt_mutex_next_owner(&pi_state
->pi_mutex
);
1230 * It is possible that the next waiter (the one that brought
1231 * this owner to the kernel) timed out and is no longer
1232 * waiting on the lock.
1235 new_owner
= this->task
;
1238 * We pass it to the next owner. The WAITERS bit is always
1239 * kept enabled while there is PI state around. We cleanup the
1240 * owner died bit, because we are the owner.
1242 newval
= FUTEX_WAITERS
| task_pid_vnr(new_owner
);
1244 if (unlikely(should_fail_futex(true)))
1247 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
)) {
1249 } else if (curval
!= uval
) {
1251 * If a unconditional UNLOCK_PI operation (user space did not
1252 * try the TID->0 transition) raced with a waiter setting the
1253 * FUTEX_WAITERS flag between get_user() and locking the hash
1254 * bucket lock, retry the operation.
1256 if ((FUTEX_TID_MASK
& curval
) == uval
)
1262 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
1266 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
1267 WARN_ON(list_empty(&pi_state
->list
));
1268 list_del_init(&pi_state
->list
);
1269 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
1271 raw_spin_lock_irq(&new_owner
->pi_lock
);
1272 WARN_ON(!list_empty(&pi_state
->list
));
1273 list_add(&pi_state
->list
, &new_owner
->pi_state_list
);
1274 pi_state
->owner
= new_owner
;
1275 raw_spin_unlock_irq(&new_owner
->pi_lock
);
1277 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
1279 deboost
= rt_mutex_futex_unlock(&pi_state
->pi_mutex
, &wake_q
);
1282 * First unlock HB so the waiter does not spin on it once he got woken
1283 * up. Second wake up the waiter before the priority is adjusted. If we
1284 * deboost first (and lose our higher priority), then the task might get
1285 * scheduled away before the wake up can take place.
1287 spin_unlock(&hb
->lock
);
1290 rt_mutex_adjust_prio(current
);
1296 * Express the locking dependencies for lockdep:
1299 double_lock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1302 spin_lock(&hb1
->lock
);
1304 spin_lock_nested(&hb2
->lock
, SINGLE_DEPTH_NESTING
);
1305 } else { /* hb1 > hb2 */
1306 spin_lock(&hb2
->lock
);
1307 spin_lock_nested(&hb1
->lock
, SINGLE_DEPTH_NESTING
);
1312 double_unlock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1314 spin_unlock(&hb1
->lock
);
1316 spin_unlock(&hb2
->lock
);
1320 * Wake up waiters matching bitset queued on this futex (uaddr).
1323 futex_wake(u32 __user
*uaddr
, unsigned int flags
, int nr_wake
, u32 bitset
)
1325 struct futex_hash_bucket
*hb
;
1326 struct futex_q
*this, *next
;
1327 union futex_key key
= FUTEX_KEY_INIT
;
1334 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_READ
);
1335 if (unlikely(ret
!= 0))
1338 hb
= hash_futex(&key
);
1340 /* Make sure we really have tasks to wakeup */
1341 if (!hb_waiters_pending(hb
))
1344 spin_lock(&hb
->lock
);
1346 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
1347 if (match_futex (&this->key
, &key
)) {
1348 if (this->pi_state
|| this->rt_waiter
) {
1353 /* Check if one of the bits is set in both bitsets */
1354 if (!(this->bitset
& bitset
))
1357 mark_wake_futex(&wake_q
, this);
1358 if (++ret
>= nr_wake
)
1363 spin_unlock(&hb
->lock
);
1366 put_futex_key(&key
);
1372 * Wake up all waiters hashed on the physical page that is mapped
1373 * to this virtual address:
1376 futex_wake_op(u32 __user
*uaddr1
, unsigned int flags
, u32 __user
*uaddr2
,
1377 int nr_wake
, int nr_wake2
, int op
)
1379 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1380 struct futex_hash_bucket
*hb1
, *hb2
;
1381 struct futex_q
*this, *next
;
1386 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1387 if (unlikely(ret
!= 0))
1389 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
1390 if (unlikely(ret
!= 0))
1393 hb1
= hash_futex(&key1
);
1394 hb2
= hash_futex(&key2
);
1397 double_lock_hb(hb1
, hb2
);
1398 op_ret
= futex_atomic_op_inuser(op
, uaddr2
);
1399 if (unlikely(op_ret
< 0)) {
1401 double_unlock_hb(hb1
, hb2
);
1405 * we don't get EFAULT from MMU faults if we don't have an MMU,
1406 * but we might get them from range checking
1412 if (unlikely(op_ret
!= -EFAULT
)) {
1417 ret
= fault_in_user_writeable(uaddr2
);
1421 if (!(flags
& FLAGS_SHARED
))
1424 put_futex_key(&key2
);
1425 put_futex_key(&key1
);
1429 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1430 if (match_futex (&this->key
, &key1
)) {
1431 if (this->pi_state
|| this->rt_waiter
) {
1435 mark_wake_futex(&wake_q
, this);
1436 if (++ret
>= nr_wake
)
1443 plist_for_each_entry_safe(this, next
, &hb2
->chain
, list
) {
1444 if (match_futex (&this->key
, &key2
)) {
1445 if (this->pi_state
|| this->rt_waiter
) {
1449 mark_wake_futex(&wake_q
, this);
1450 if (++op_ret
>= nr_wake2
)
1458 double_unlock_hb(hb1
, hb2
);
1461 put_futex_key(&key2
);
1463 put_futex_key(&key1
);
1469 * requeue_futex() - Requeue a futex_q from one hb to another
1470 * @q: the futex_q to requeue
1471 * @hb1: the source hash_bucket
1472 * @hb2: the target hash_bucket
1473 * @key2: the new key for the requeued futex_q
1476 void requeue_futex(struct futex_q
*q
, struct futex_hash_bucket
*hb1
,
1477 struct futex_hash_bucket
*hb2
, union futex_key
*key2
)
1481 * If key1 and key2 hash to the same bucket, no need to
1484 if (likely(&hb1
->chain
!= &hb2
->chain
)) {
1485 plist_del(&q
->list
, &hb1
->chain
);
1486 hb_waiters_dec(hb1
);
1487 hb_waiters_inc(hb2
);
1488 plist_add(&q
->list
, &hb2
->chain
);
1489 q
->lock_ptr
= &hb2
->lock
;
1491 get_futex_key_refs(key2
);
1496 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1498 * @key: the key of the requeue target futex
1499 * @hb: the hash_bucket of the requeue target futex
1501 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1502 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1503 * to the requeue target futex so the waiter can detect the wakeup on the right
1504 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1505 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1506 * to protect access to the pi_state to fixup the owner later. Must be called
1507 * with both q->lock_ptr and hb->lock held.
1510 void requeue_pi_wake_futex(struct futex_q
*q
, union futex_key
*key
,
1511 struct futex_hash_bucket
*hb
)
1513 get_futex_key_refs(key
);
1518 WARN_ON(!q
->rt_waiter
);
1519 q
->rt_waiter
= NULL
;
1521 q
->lock_ptr
= &hb
->lock
;
1523 wake_up_state(q
->task
, TASK_NORMAL
);
1527 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1528 * @pifutex: the user address of the to futex
1529 * @hb1: the from futex hash bucket, must be locked by the caller
1530 * @hb2: the to futex hash bucket, must be locked by the caller
1531 * @key1: the from futex key
1532 * @key2: the to futex key
1533 * @ps: address to store the pi_state pointer
1534 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1536 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1537 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1538 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1539 * hb1 and hb2 must be held by the caller.
1542 * 0 - failed to acquire the lock atomically;
1543 * >0 - acquired the lock, return value is vpid of the top_waiter
1546 static int futex_proxy_trylock_atomic(u32 __user
*pifutex
,
1547 struct futex_hash_bucket
*hb1
,
1548 struct futex_hash_bucket
*hb2
,
1549 union futex_key
*key1
, union futex_key
*key2
,
1550 struct futex_pi_state
**ps
, int set_waiters
)
1552 struct futex_q
*top_waiter
= NULL
;
1556 if (get_futex_value_locked(&curval
, pifutex
))
1559 if (unlikely(should_fail_futex(true)))
1563 * Find the top_waiter and determine if there are additional waiters.
1564 * If the caller intends to requeue more than 1 waiter to pifutex,
1565 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1566 * as we have means to handle the possible fault. If not, don't set
1567 * the bit unecessarily as it will force the subsequent unlock to enter
1570 top_waiter
= futex_top_waiter(hb1
, key1
);
1572 /* There are no waiters, nothing for us to do. */
1576 /* Ensure we requeue to the expected futex. */
1577 if (!match_futex(top_waiter
->requeue_pi_key
, key2
))
1581 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1582 * the contended case or if set_waiters is 1. The pi_state is returned
1583 * in ps in contended cases.
1585 vpid
= task_pid_vnr(top_waiter
->task
);
1586 ret
= futex_lock_pi_atomic(pifutex
, hb2
, key2
, ps
, top_waiter
->task
,
1589 requeue_pi_wake_futex(top_waiter
, key2
, hb2
);
1596 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1597 * @uaddr1: source futex user address
1598 * @flags: futex flags (FLAGS_SHARED, etc.)
1599 * @uaddr2: target futex user address
1600 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1601 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1602 * @cmpval: @uaddr1 expected value (or %NULL)
1603 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1604 * pi futex (pi to pi requeue is not supported)
1606 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1607 * uaddr2 atomically on behalf of the top waiter.
1610 * >=0 - on success, the number of tasks requeued or woken;
1613 static int futex_requeue(u32 __user
*uaddr1
, unsigned int flags
,
1614 u32 __user
*uaddr2
, int nr_wake
, int nr_requeue
,
1615 u32
*cmpval
, int requeue_pi
)
1617 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1618 int drop_count
= 0, task_count
= 0, ret
;
1619 struct futex_pi_state
*pi_state
= NULL
;
1620 struct futex_hash_bucket
*hb1
, *hb2
;
1621 struct futex_q
*this, *next
;
1626 * Requeue PI only works on two distinct uaddrs. This
1627 * check is only valid for private futexes. See below.
1629 if (uaddr1
== uaddr2
)
1633 * requeue_pi requires a pi_state, try to allocate it now
1634 * without any locks in case it fails.
1636 if (refill_pi_state_cache())
1639 * requeue_pi must wake as many tasks as it can, up to nr_wake
1640 * + nr_requeue, since it acquires the rt_mutex prior to
1641 * returning to userspace, so as to not leave the rt_mutex with
1642 * waiters and no owner. However, second and third wake-ups
1643 * cannot be predicted as they involve race conditions with the
1644 * first wake and a fault while looking up the pi_state. Both
1645 * pthread_cond_signal() and pthread_cond_broadcast() should
1653 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1654 if (unlikely(ret
!= 0))
1656 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
,
1657 requeue_pi
? VERIFY_WRITE
: VERIFY_READ
);
1658 if (unlikely(ret
!= 0))
1662 * The check above which compares uaddrs is not sufficient for
1663 * shared futexes. We need to compare the keys:
1665 if (requeue_pi
&& match_futex(&key1
, &key2
)) {
1670 hb1
= hash_futex(&key1
);
1671 hb2
= hash_futex(&key2
);
1674 hb_waiters_inc(hb2
);
1675 double_lock_hb(hb1
, hb2
);
1677 if (likely(cmpval
!= NULL
)) {
1680 ret
= get_futex_value_locked(&curval
, uaddr1
);
1682 if (unlikely(ret
)) {
1683 double_unlock_hb(hb1
, hb2
);
1684 hb_waiters_dec(hb2
);
1686 ret
= get_user(curval
, uaddr1
);
1690 if (!(flags
& FLAGS_SHARED
))
1693 put_futex_key(&key2
);
1694 put_futex_key(&key1
);
1697 if (curval
!= *cmpval
) {
1703 if (requeue_pi
&& (task_count
- nr_wake
< nr_requeue
)) {
1705 * Attempt to acquire uaddr2 and wake the top waiter. If we
1706 * intend to requeue waiters, force setting the FUTEX_WAITERS
1707 * bit. We force this here where we are able to easily handle
1708 * faults rather in the requeue loop below.
1710 ret
= futex_proxy_trylock_atomic(uaddr2
, hb1
, hb2
, &key1
,
1711 &key2
, &pi_state
, nr_requeue
);
1714 * At this point the top_waiter has either taken uaddr2 or is
1715 * waiting on it. If the former, then the pi_state will not
1716 * exist yet, look it up one more time to ensure we have a
1717 * reference to it. If the lock was taken, ret contains the
1718 * vpid of the top waiter task.
1725 * If we acquired the lock, then the user
1726 * space value of uaddr2 should be vpid. It
1727 * cannot be changed by the top waiter as it
1728 * is blocked on hb2 lock if it tries to do
1729 * so. If something fiddled with it behind our
1730 * back the pi state lookup might unearth
1731 * it. So we rather use the known value than
1732 * rereading and handing potential crap to
1735 ret
= lookup_pi_state(ret
, hb2
, &key2
, &pi_state
);
1742 free_pi_state(pi_state
);
1744 double_unlock_hb(hb1
, hb2
);
1745 hb_waiters_dec(hb2
);
1746 put_futex_key(&key2
);
1747 put_futex_key(&key1
);
1748 ret
= fault_in_user_writeable(uaddr2
);
1754 * Two reasons for this:
1755 * - Owner is exiting and we just wait for the
1757 * - The user space value changed.
1759 free_pi_state(pi_state
);
1761 double_unlock_hb(hb1
, hb2
);
1762 hb_waiters_dec(hb2
);
1763 put_futex_key(&key2
);
1764 put_futex_key(&key1
);
1772 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1773 if (task_count
- nr_wake
>= nr_requeue
)
1776 if (!match_futex(&this->key
, &key1
))
1780 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1781 * be paired with each other and no other futex ops.
1783 * We should never be requeueing a futex_q with a pi_state,
1784 * which is awaiting a futex_unlock_pi().
1786 if ((requeue_pi
&& !this->rt_waiter
) ||
1787 (!requeue_pi
&& this->rt_waiter
) ||
1794 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1795 * lock, we already woke the top_waiter. If not, it will be
1796 * woken by futex_unlock_pi().
1798 if (++task_count
<= nr_wake
&& !requeue_pi
) {
1799 mark_wake_futex(&wake_q
, this);
1803 /* Ensure we requeue to the expected futex for requeue_pi. */
1804 if (requeue_pi
&& !match_futex(this->requeue_pi_key
, &key2
)) {
1810 * Requeue nr_requeue waiters and possibly one more in the case
1811 * of requeue_pi if we couldn't acquire the lock atomically.
1814 /* Prepare the waiter to take the rt_mutex. */
1815 atomic_inc(&pi_state
->refcount
);
1816 this->pi_state
= pi_state
;
1817 ret
= rt_mutex_start_proxy_lock(&pi_state
->pi_mutex
,
1821 /* We got the lock. */
1822 requeue_pi_wake_futex(this, &key2
, hb2
);
1827 this->pi_state
= NULL
;
1828 free_pi_state(pi_state
);
1832 requeue_futex(this, hb1
, hb2
, &key2
);
1837 free_pi_state(pi_state
);
1838 double_unlock_hb(hb1
, hb2
);
1840 hb_waiters_dec(hb2
);
1843 * drop_futex_key_refs() must be called outside the spinlocks. During
1844 * the requeue we moved futex_q's from the hash bucket at key1 to the
1845 * one at key2 and updated their key pointer. We no longer need to
1846 * hold the references to key1.
1848 while (--drop_count
>= 0)
1849 drop_futex_key_refs(&key1
);
1852 put_futex_key(&key2
);
1854 put_futex_key(&key1
);
1856 return ret
? ret
: task_count
;
1859 /* The key must be already stored in q->key. */
1860 static inline struct futex_hash_bucket
*queue_lock(struct futex_q
*q
)
1861 __acquires(&hb
->lock
)
1863 struct futex_hash_bucket
*hb
;
1865 hb
= hash_futex(&q
->key
);
1868 * Increment the counter before taking the lock so that
1869 * a potential waker won't miss a to-be-slept task that is
1870 * waiting for the spinlock. This is safe as all queue_lock()
1871 * users end up calling queue_me(). Similarly, for housekeeping,
1872 * decrement the counter at queue_unlock() when some error has
1873 * occurred and we don't end up adding the task to the list.
1877 q
->lock_ptr
= &hb
->lock
;
1879 spin_lock(&hb
->lock
); /* implies MB (A) */
1884 queue_unlock(struct futex_hash_bucket
*hb
)
1885 __releases(&hb
->lock
)
1887 spin_unlock(&hb
->lock
);
1892 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1893 * @q: The futex_q to enqueue
1894 * @hb: The destination hash bucket
1896 * The hb->lock must be held by the caller, and is released here. A call to
1897 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1898 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1899 * or nothing if the unqueue is done as part of the wake process and the unqueue
1900 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1903 static inline void queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
1904 __releases(&hb
->lock
)
1909 * The priority used to register this element is
1910 * - either the real thread-priority for the real-time threads
1911 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1912 * - or MAX_RT_PRIO for non-RT threads.
1913 * Thus, all RT-threads are woken first in priority order, and
1914 * the others are woken last, in FIFO order.
1916 prio
= min(current
->normal_prio
, MAX_RT_PRIO
);
1918 plist_node_init(&q
->list
, prio
);
1919 plist_add(&q
->list
, &hb
->chain
);
1921 spin_unlock(&hb
->lock
);
1925 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1926 * @q: The futex_q to unqueue
1928 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1929 * be paired with exactly one earlier call to queue_me().
1932 * 1 - if the futex_q was still queued (and we removed unqueued it);
1933 * 0 - if the futex_q was already removed by the waking thread
1935 static int unqueue_me(struct futex_q
*q
)
1937 spinlock_t
*lock_ptr
;
1940 /* In the common case we don't take the spinlock, which is nice. */
1942 lock_ptr
= q
->lock_ptr
;
1944 if (lock_ptr
!= NULL
) {
1945 spin_lock(lock_ptr
);
1947 * q->lock_ptr can change between reading it and
1948 * spin_lock(), causing us to take the wrong lock. This
1949 * corrects the race condition.
1951 * Reasoning goes like this: if we have the wrong lock,
1952 * q->lock_ptr must have changed (maybe several times)
1953 * between reading it and the spin_lock(). It can
1954 * change again after the spin_lock() but only if it was
1955 * already changed before the spin_lock(). It cannot,
1956 * however, change back to the original value. Therefore
1957 * we can detect whether we acquired the correct lock.
1959 if (unlikely(lock_ptr
!= q
->lock_ptr
)) {
1960 spin_unlock(lock_ptr
);
1965 BUG_ON(q
->pi_state
);
1967 spin_unlock(lock_ptr
);
1971 drop_futex_key_refs(&q
->key
);
1976 * PI futexes can not be requeued and must remove themself from the
1977 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1980 static void unqueue_me_pi(struct futex_q
*q
)
1981 __releases(q
->lock_ptr
)
1985 BUG_ON(!q
->pi_state
);
1986 free_pi_state(q
->pi_state
);
1989 spin_unlock(q
->lock_ptr
);
1993 * Fixup the pi_state owner with the new owner.
1995 * Must be called with hash bucket lock held and mm->sem held for non
1998 static int fixup_pi_state_owner(u32 __user
*uaddr
, struct futex_q
*q
,
1999 struct task_struct
*newowner
)
2001 u32 newtid
= task_pid_vnr(newowner
) | FUTEX_WAITERS
;
2002 struct futex_pi_state
*pi_state
= q
->pi_state
;
2003 struct task_struct
*oldowner
= pi_state
->owner
;
2004 u32 uval
, uninitialized_var(curval
), newval
;
2008 if (!pi_state
->owner
)
2009 newtid
|= FUTEX_OWNER_DIED
;
2012 * We are here either because we stole the rtmutex from the
2013 * previous highest priority waiter or we are the highest priority
2014 * waiter but failed to get the rtmutex the first time.
2015 * We have to replace the newowner TID in the user space variable.
2016 * This must be atomic as we have to preserve the owner died bit here.
2018 * Note: We write the user space value _before_ changing the pi_state
2019 * because we can fault here. Imagine swapped out pages or a fork
2020 * that marked all the anonymous memory readonly for cow.
2022 * Modifying pi_state _before_ the user space value would
2023 * leave the pi_state in an inconsistent state when we fault
2024 * here, because we need to drop the hash bucket lock to
2025 * handle the fault. This might be observed in the PID check
2026 * in lookup_pi_state.
2029 if (get_futex_value_locked(&uval
, uaddr
))
2033 newval
= (uval
& FUTEX_OWNER_DIED
) | newtid
;
2035 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
))
2043 * We fixed up user space. Now we need to fix the pi_state
2046 if (pi_state
->owner
!= NULL
) {
2047 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
2048 WARN_ON(list_empty(&pi_state
->list
));
2049 list_del_init(&pi_state
->list
);
2050 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
2053 pi_state
->owner
= newowner
;
2055 raw_spin_lock_irq(&newowner
->pi_lock
);
2056 WARN_ON(!list_empty(&pi_state
->list
));
2057 list_add(&pi_state
->list
, &newowner
->pi_state_list
);
2058 raw_spin_unlock_irq(&newowner
->pi_lock
);
2062 * To handle the page fault we need to drop the hash bucket
2063 * lock here. That gives the other task (either the highest priority
2064 * waiter itself or the task which stole the rtmutex) the
2065 * chance to try the fixup of the pi_state. So once we are
2066 * back from handling the fault we need to check the pi_state
2067 * after reacquiring the hash bucket lock and before trying to
2068 * do another fixup. When the fixup has been done already we
2072 spin_unlock(q
->lock_ptr
);
2074 ret
= fault_in_user_writeable(uaddr
);
2076 spin_lock(q
->lock_ptr
);
2079 * Check if someone else fixed it for us:
2081 if (pi_state
->owner
!= oldowner
)
2090 static long futex_wait_restart(struct restart_block
*restart
);
2093 * fixup_owner() - Post lock pi_state and corner case management
2094 * @uaddr: user address of the futex
2095 * @q: futex_q (contains pi_state and access to the rt_mutex)
2096 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2098 * After attempting to lock an rt_mutex, this function is called to cleanup
2099 * the pi_state owner as well as handle race conditions that may allow us to
2100 * acquire the lock. Must be called with the hb lock held.
2103 * 1 - success, lock taken;
2104 * 0 - success, lock not taken;
2105 * <0 - on error (-EFAULT)
2107 static int fixup_owner(u32 __user
*uaddr
, struct futex_q
*q
, int locked
)
2109 struct task_struct
*owner
;
2114 * Got the lock. We might not be the anticipated owner if we
2115 * did a lock-steal - fix up the PI-state in that case:
2117 if (q
->pi_state
->owner
!= current
)
2118 ret
= fixup_pi_state_owner(uaddr
, q
, current
);
2123 * Catch the rare case, where the lock was released when we were on the
2124 * way back before we locked the hash bucket.
2126 if (q
->pi_state
->owner
== current
) {
2128 * Try to get the rt_mutex now. This might fail as some other
2129 * task acquired the rt_mutex after we removed ourself from the
2130 * rt_mutex waiters list.
2132 if (rt_mutex_trylock(&q
->pi_state
->pi_mutex
)) {
2138 * pi_state is incorrect, some other task did a lock steal and
2139 * we returned due to timeout or signal without taking the
2140 * rt_mutex. Too late.
2142 raw_spin_lock(&q
->pi_state
->pi_mutex
.wait_lock
);
2143 owner
= rt_mutex_owner(&q
->pi_state
->pi_mutex
);
2145 owner
= rt_mutex_next_owner(&q
->pi_state
->pi_mutex
);
2146 raw_spin_unlock(&q
->pi_state
->pi_mutex
.wait_lock
);
2147 ret
= fixup_pi_state_owner(uaddr
, q
, owner
);
2152 * Paranoia check. If we did not take the lock, then we should not be
2153 * the owner of the rt_mutex.
2155 if (rt_mutex_owner(&q
->pi_state
->pi_mutex
) == current
)
2156 printk(KERN_ERR
"fixup_owner: ret = %d pi-mutex: %p "
2157 "pi-state %p\n", ret
,
2158 q
->pi_state
->pi_mutex
.owner
,
2159 q
->pi_state
->owner
);
2162 return ret
? ret
: locked
;
2166 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2167 * @hb: the futex hash bucket, must be locked by the caller
2168 * @q: the futex_q to queue up on
2169 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2171 static void futex_wait_queue_me(struct futex_hash_bucket
*hb
, struct futex_q
*q
,
2172 struct hrtimer_sleeper
*timeout
)
2175 * The task state is guaranteed to be set before another task can
2176 * wake it. set_current_state() is implemented using smp_store_mb() and
2177 * queue_me() calls spin_unlock() upon completion, both serializing
2178 * access to the hash list and forcing another memory barrier.
2180 set_current_state(TASK_INTERRUPTIBLE
);
2185 hrtimer_start_expires(&timeout
->timer
, HRTIMER_MODE_ABS
);
2188 * If we have been removed from the hash list, then another task
2189 * has tried to wake us, and we can skip the call to schedule().
2191 if (likely(!plist_node_empty(&q
->list
))) {
2193 * If the timer has already expired, current will already be
2194 * flagged for rescheduling. Only call schedule if there
2195 * is no timeout, or if it has yet to expire.
2197 if (!timeout
|| timeout
->task
)
2198 freezable_schedule();
2200 __set_current_state(TASK_RUNNING
);
2204 * futex_wait_setup() - Prepare to wait on a futex
2205 * @uaddr: the futex userspace address
2206 * @val: the expected value
2207 * @flags: futex flags (FLAGS_SHARED, etc.)
2208 * @q: the associated futex_q
2209 * @hb: storage for hash_bucket pointer to be returned to caller
2211 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2212 * compare it with the expected value. Handle atomic faults internally.
2213 * Return with the hb lock held and a q.key reference on success, and unlocked
2214 * with no q.key reference on failure.
2217 * 0 - uaddr contains val and hb has been locked;
2218 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2220 static int futex_wait_setup(u32 __user
*uaddr
, u32 val
, unsigned int flags
,
2221 struct futex_q
*q
, struct futex_hash_bucket
**hb
)
2227 * Access the page AFTER the hash-bucket is locked.
2228 * Order is important:
2230 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2231 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2233 * The basic logical guarantee of a futex is that it blocks ONLY
2234 * if cond(var) is known to be true at the time of blocking, for
2235 * any cond. If we locked the hash-bucket after testing *uaddr, that
2236 * would open a race condition where we could block indefinitely with
2237 * cond(var) false, which would violate the guarantee.
2239 * On the other hand, we insert q and release the hash-bucket only
2240 * after testing *uaddr. This guarantees that futex_wait() will NOT
2241 * absorb a wakeup if *uaddr does not match the desired values
2242 * while the syscall executes.
2245 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
->key
, VERIFY_READ
);
2246 if (unlikely(ret
!= 0))
2250 *hb
= queue_lock(q
);
2252 ret
= get_futex_value_locked(&uval
, uaddr
);
2257 ret
= get_user(uval
, uaddr
);
2261 if (!(flags
& FLAGS_SHARED
))
2264 put_futex_key(&q
->key
);
2275 put_futex_key(&q
->key
);
2279 static int futex_wait(u32 __user
*uaddr
, unsigned int flags
, u32 val
,
2280 ktime_t
*abs_time
, u32 bitset
)
2282 struct hrtimer_sleeper timeout
, *to
= NULL
;
2283 struct restart_block
*restart
;
2284 struct futex_hash_bucket
*hb
;
2285 struct futex_q q
= futex_q_init
;
2295 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2296 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2298 hrtimer_init_sleeper(to
, current
);
2299 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2300 current
->timer_slack_ns
);
2305 * Prepare to wait on uaddr. On success, holds hb lock and increments
2308 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2312 /* queue_me and wait for wakeup, timeout, or a signal. */
2313 futex_wait_queue_me(hb
, &q
, to
);
2315 /* If we were woken (and unqueued), we succeeded, whatever. */
2317 /* unqueue_me() drops q.key ref */
2318 if (!unqueue_me(&q
))
2321 if (to
&& !to
->task
)
2325 * We expect signal_pending(current), but we might be the
2326 * victim of a spurious wakeup as well.
2328 if (!signal_pending(current
))
2335 restart
= ¤t
->restart_block
;
2336 restart
->fn
= futex_wait_restart
;
2337 restart
->futex
.uaddr
= uaddr
;
2338 restart
->futex
.val
= val
;
2339 restart
->futex
.time
= abs_time
->tv64
;
2340 restart
->futex
.bitset
= bitset
;
2341 restart
->futex
.flags
= flags
| FLAGS_HAS_TIMEOUT
;
2343 ret
= -ERESTART_RESTARTBLOCK
;
2347 hrtimer_cancel(&to
->timer
);
2348 destroy_hrtimer_on_stack(&to
->timer
);
2354 static long futex_wait_restart(struct restart_block
*restart
)
2356 u32 __user
*uaddr
= restart
->futex
.uaddr
;
2357 ktime_t t
, *tp
= NULL
;
2359 if (restart
->futex
.flags
& FLAGS_HAS_TIMEOUT
) {
2360 t
.tv64
= restart
->futex
.time
;
2363 restart
->fn
= do_no_restart_syscall
;
2365 return (long)futex_wait(uaddr
, restart
->futex
.flags
,
2366 restart
->futex
.val
, tp
, restart
->futex
.bitset
);
2371 * Userspace tried a 0 -> TID atomic transition of the futex value
2372 * and failed. The kernel side here does the whole locking operation:
2373 * if there are waiters then it will block as a consequence of relying
2374 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2375 * a 0 value of the futex too.).
2377 * Also serves as futex trylock_pi()'ing, and due semantics.
2379 static int futex_lock_pi(u32 __user
*uaddr
, unsigned int flags
,
2380 ktime_t
*time
, int trylock
)
2382 struct hrtimer_sleeper timeout
, *to
= NULL
;
2383 struct futex_hash_bucket
*hb
;
2384 struct futex_q q
= futex_q_init
;
2387 if (refill_pi_state_cache())
2392 hrtimer_init_on_stack(&to
->timer
, CLOCK_REALTIME
,
2394 hrtimer_init_sleeper(to
, current
);
2395 hrtimer_set_expires(&to
->timer
, *time
);
2399 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
.key
, VERIFY_WRITE
);
2400 if (unlikely(ret
!= 0))
2404 hb
= queue_lock(&q
);
2406 ret
= futex_lock_pi_atomic(uaddr
, hb
, &q
.key
, &q
.pi_state
, current
, 0);
2407 if (unlikely(ret
)) {
2409 * Atomic work succeeded and we got the lock,
2410 * or failed. Either way, we do _not_ block.
2414 /* We got the lock. */
2416 goto out_unlock_put_key
;
2421 * Two reasons for this:
2422 * - Task is exiting and we just wait for the
2424 * - The user space value changed.
2427 put_futex_key(&q
.key
);
2431 goto out_unlock_put_key
;
2436 * Only actually queue now that the atomic ops are done:
2440 WARN_ON(!q
.pi_state
);
2442 * Block on the PI mutex:
2445 ret
= rt_mutex_timed_futex_lock(&q
.pi_state
->pi_mutex
, to
);
2447 ret
= rt_mutex_trylock(&q
.pi_state
->pi_mutex
);
2448 /* Fixup the trylock return value: */
2449 ret
= ret
? 0 : -EWOULDBLOCK
;
2452 spin_lock(q
.lock_ptr
);
2454 * Fixup the pi_state owner and possibly acquire the lock if we
2457 res
= fixup_owner(uaddr
, &q
, !ret
);
2459 * If fixup_owner() returned an error, proprogate that. If it acquired
2460 * the lock, clear our -ETIMEDOUT or -EINTR.
2463 ret
= (res
< 0) ? res
: 0;
2466 * If fixup_owner() faulted and was unable to handle the fault, unlock
2467 * it and return the fault to userspace.
2469 if (ret
&& (rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
))
2470 rt_mutex_unlock(&q
.pi_state
->pi_mutex
);
2472 /* Unqueue and drop the lock */
2481 put_futex_key(&q
.key
);
2484 destroy_hrtimer_on_stack(&to
->timer
);
2485 return ret
!= -EINTR
? ret
: -ERESTARTNOINTR
;
2490 ret
= fault_in_user_writeable(uaddr
);
2494 if (!(flags
& FLAGS_SHARED
))
2497 put_futex_key(&q
.key
);
2502 * Userspace attempted a TID -> 0 atomic transition, and failed.
2503 * This is the in-kernel slowpath: we look up the PI state (if any),
2504 * and do the rt-mutex unlock.
2506 static int futex_unlock_pi(u32 __user
*uaddr
, unsigned int flags
)
2508 u32
uninitialized_var(curval
), uval
, vpid
= task_pid_vnr(current
);
2509 union futex_key key
= FUTEX_KEY_INIT
;
2510 struct futex_hash_bucket
*hb
;
2511 struct futex_q
*match
;
2515 if (get_user(uval
, uaddr
))
2518 * We release only a lock we actually own:
2520 if ((uval
& FUTEX_TID_MASK
) != vpid
)
2523 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_WRITE
);
2527 hb
= hash_futex(&key
);
2528 spin_lock(&hb
->lock
);
2531 * Check waiters first. We do not trust user space values at
2532 * all and we at least want to know if user space fiddled
2533 * with the futex value instead of blindly unlocking.
2535 match
= futex_top_waiter(hb
, &key
);
2537 ret
= wake_futex_pi(uaddr
, uval
, match
, hb
);
2539 * In case of success wake_futex_pi dropped the hash
2545 * The atomic access to the futex value generated a
2546 * pagefault, so retry the user-access and the wakeup:
2551 * A unconditional UNLOCK_PI op raced against a waiter
2552 * setting the FUTEX_WAITERS bit. Try again.
2554 if (ret
== -EAGAIN
) {
2555 spin_unlock(&hb
->lock
);
2556 put_futex_key(&key
);
2560 * wake_futex_pi has detected invalid state. Tell user
2567 * We have no kernel internal state, i.e. no waiters in the
2568 * kernel. Waiters which are about to queue themselves are stuck
2569 * on hb->lock. So we can safely ignore them. We do neither
2570 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2573 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, 0))
2577 * If uval has changed, let user space handle it.
2579 ret
= (curval
== uval
) ? 0 : -EAGAIN
;
2582 spin_unlock(&hb
->lock
);
2584 put_futex_key(&key
);
2588 spin_unlock(&hb
->lock
);
2589 put_futex_key(&key
);
2591 ret
= fault_in_user_writeable(uaddr
);
2599 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2600 * @hb: the hash_bucket futex_q was original enqueued on
2601 * @q: the futex_q woken while waiting to be requeued
2602 * @key2: the futex_key of the requeue target futex
2603 * @timeout: the timeout associated with the wait (NULL if none)
2605 * Detect if the task was woken on the initial futex as opposed to the requeue
2606 * target futex. If so, determine if it was a timeout or a signal that caused
2607 * the wakeup and return the appropriate error code to the caller. Must be
2608 * called with the hb lock held.
2611 * 0 = no early wakeup detected;
2612 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2615 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket
*hb
,
2616 struct futex_q
*q
, union futex_key
*key2
,
2617 struct hrtimer_sleeper
*timeout
)
2622 * With the hb lock held, we avoid races while we process the wakeup.
2623 * We only need to hold hb (and not hb2) to ensure atomicity as the
2624 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2625 * It can't be requeued from uaddr2 to something else since we don't
2626 * support a PI aware source futex for requeue.
2628 if (!match_futex(&q
->key
, key2
)) {
2629 WARN_ON(q
->lock_ptr
&& (&hb
->lock
!= q
->lock_ptr
));
2631 * We were woken prior to requeue by a timeout or a signal.
2632 * Unqueue the futex_q and determine which it was.
2634 plist_del(&q
->list
, &hb
->chain
);
2637 /* Handle spurious wakeups gracefully */
2639 if (timeout
&& !timeout
->task
)
2641 else if (signal_pending(current
))
2642 ret
= -ERESTARTNOINTR
;
2648 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2649 * @uaddr: the futex we initially wait on (non-pi)
2650 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2651 * the same type, no requeueing from private to shared, etc.
2652 * @val: the expected value of uaddr
2653 * @abs_time: absolute timeout
2654 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2655 * @uaddr2: the pi futex we will take prior to returning to user-space
2657 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2658 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2659 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2660 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2661 * without one, the pi logic would not know which task to boost/deboost, if
2662 * there was a need to.
2664 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2665 * via the following--
2666 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2667 * 2) wakeup on uaddr2 after a requeue
2671 * If 3, cleanup and return -ERESTARTNOINTR.
2673 * If 2, we may then block on trying to take the rt_mutex and return via:
2674 * 5) successful lock
2677 * 8) other lock acquisition failure
2679 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2681 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2687 static int futex_wait_requeue_pi(u32 __user
*uaddr
, unsigned int flags
,
2688 u32 val
, ktime_t
*abs_time
, u32 bitset
,
2691 struct hrtimer_sleeper timeout
, *to
= NULL
;
2692 struct rt_mutex_waiter rt_waiter
;
2693 struct futex_hash_bucket
*hb
;
2694 union futex_key key2
= FUTEX_KEY_INIT
;
2695 struct futex_q q
= futex_q_init
;
2698 if (uaddr
== uaddr2
)
2706 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2707 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2709 hrtimer_init_sleeper(to
, current
);
2710 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2711 current
->timer_slack_ns
);
2715 * The waiter is allocated on our stack, manipulated by the requeue
2716 * code while we sleep on uaddr.
2718 debug_rt_mutex_init_waiter(&rt_waiter
);
2719 RB_CLEAR_NODE(&rt_waiter
.pi_tree_entry
);
2720 RB_CLEAR_NODE(&rt_waiter
.tree_entry
);
2721 rt_waiter
.task
= NULL
;
2723 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
2724 if (unlikely(ret
!= 0))
2728 q
.rt_waiter
= &rt_waiter
;
2729 q
.requeue_pi_key
= &key2
;
2732 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2735 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2740 * The check above which compares uaddrs is not sufficient for
2741 * shared futexes. We need to compare the keys:
2743 if (match_futex(&q
.key
, &key2
)) {
2749 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2750 futex_wait_queue_me(hb
, &q
, to
);
2752 spin_lock(&hb
->lock
);
2753 ret
= handle_early_requeue_pi_wakeup(hb
, &q
, &key2
, to
);
2754 spin_unlock(&hb
->lock
);
2759 * In order for us to be here, we know our q.key == key2, and since
2760 * we took the hb->lock above, we also know that futex_requeue() has
2761 * completed and we no longer have to concern ourselves with a wakeup
2762 * race with the atomic proxy lock acquisition by the requeue code. The
2763 * futex_requeue dropped our key1 reference and incremented our key2
2767 /* Check if the requeue code acquired the second futex for us. */
2770 * Got the lock. We might not be the anticipated owner if we
2771 * did a lock-steal - fix up the PI-state in that case.
2773 if (q
.pi_state
&& (q
.pi_state
->owner
!= current
)) {
2774 spin_lock(q
.lock_ptr
);
2775 ret
= fixup_pi_state_owner(uaddr2
, &q
, current
);
2776 if (ret
&& rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
)
2777 rt_mutex_unlock(&q
.pi_state
->pi_mutex
);
2779 * Drop the reference to the pi state which
2780 * the requeue_pi() code acquired for us.
2782 free_pi_state(q
.pi_state
);
2783 spin_unlock(q
.lock_ptr
);
2786 struct rt_mutex
*pi_mutex
;
2789 * We have been woken up by futex_unlock_pi(), a timeout, or a
2790 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2793 WARN_ON(!q
.pi_state
);
2794 pi_mutex
= &q
.pi_state
->pi_mutex
;
2795 ret
= rt_mutex_finish_proxy_lock(pi_mutex
, to
, &rt_waiter
);
2796 debug_rt_mutex_free_waiter(&rt_waiter
);
2798 spin_lock(q
.lock_ptr
);
2800 * Fixup the pi_state owner and possibly acquire the lock if we
2803 res
= fixup_owner(uaddr2
, &q
, !ret
);
2805 * If fixup_owner() returned an error, proprogate that. If it
2806 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2809 ret
= (res
< 0) ? res
: 0;
2812 * If fixup_pi_state_owner() faulted and was unable to handle
2813 * the fault, unlock the rt_mutex and return the fault to
2816 if (ret
&& rt_mutex_owner(pi_mutex
) == current
)
2817 rt_mutex_unlock(pi_mutex
);
2819 /* Unqueue and drop the lock. */
2823 if (ret
== -EINTR
) {
2825 * We've already been requeued, but cannot restart by calling
2826 * futex_lock_pi() directly. We could restart this syscall, but
2827 * it would detect that the user space "val" changed and return
2828 * -EWOULDBLOCK. Save the overhead of the restart and return
2829 * -EWOULDBLOCK directly.
2835 put_futex_key(&q
.key
);
2837 put_futex_key(&key2
);
2841 hrtimer_cancel(&to
->timer
);
2842 destroy_hrtimer_on_stack(&to
->timer
);
2848 * Support for robust futexes: the kernel cleans up held futexes at
2851 * Implementation: user-space maintains a per-thread list of locks it
2852 * is holding. Upon do_exit(), the kernel carefully walks this list,
2853 * and marks all locks that are owned by this thread with the
2854 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2855 * always manipulated with the lock held, so the list is private and
2856 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2857 * field, to allow the kernel to clean up if the thread dies after
2858 * acquiring the lock, but just before it could have added itself to
2859 * the list. There can only be one such pending lock.
2863 * sys_set_robust_list() - Set the robust-futex list head of a task
2864 * @head: pointer to the list-head
2865 * @len: length of the list-head, as userspace expects
2867 SYSCALL_DEFINE2(set_robust_list
, struct robust_list_head __user
*, head
,
2870 if (!futex_cmpxchg_enabled
)
2873 * The kernel knows only one size for now:
2875 if (unlikely(len
!= sizeof(*head
)))
2878 current
->robust_list
= head
;
2884 * sys_get_robust_list() - Get the robust-futex list head of a task
2885 * @pid: pid of the process [zero for current task]
2886 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2887 * @len_ptr: pointer to a length field, the kernel fills in the header size
2889 SYSCALL_DEFINE3(get_robust_list
, int, pid
,
2890 struct robust_list_head __user
* __user
*, head_ptr
,
2891 size_t __user
*, len_ptr
)
2893 struct robust_list_head __user
*head
;
2895 struct task_struct
*p
;
2897 if (!futex_cmpxchg_enabled
)
2906 p
= find_task_by_vpid(pid
);
2912 if (!ptrace_may_access(p
, PTRACE_MODE_READ_REALCREDS
))
2915 head
= p
->robust_list
;
2918 if (put_user(sizeof(*head
), len_ptr
))
2920 return put_user(head
, head_ptr
);
2929 * Process a futex-list entry, check whether it's owned by the
2930 * dying task, and do notification if so:
2932 int handle_futex_death(u32 __user
*uaddr
, struct task_struct
*curr
, int pi
)
2934 u32 uval
, uninitialized_var(nval
), mval
;
2937 if (get_user(uval
, uaddr
))
2940 if ((uval
& FUTEX_TID_MASK
) == task_pid_vnr(curr
)) {
2942 * Ok, this dying thread is truly holding a futex
2943 * of interest. Set the OWNER_DIED bit atomically
2944 * via cmpxchg, and if the value had FUTEX_WAITERS
2945 * set, wake up a waiter (if any). (We have to do a
2946 * futex_wake() even if OWNER_DIED is already set -
2947 * to handle the rare but possible case of recursive
2948 * thread-death.) The rest of the cleanup is done in
2951 mval
= (uval
& FUTEX_WAITERS
) | FUTEX_OWNER_DIED
;
2953 * We are not holding a lock here, but we want to have
2954 * the pagefault_disable/enable() protection because
2955 * we want to handle the fault gracefully. If the
2956 * access fails we try to fault in the futex with R/W
2957 * verification via get_user_pages. get_user() above
2958 * does not guarantee R/W access. If that fails we
2959 * give up and leave the futex locked.
2961 if (cmpxchg_futex_value_locked(&nval
, uaddr
, uval
, mval
)) {
2962 if (fault_in_user_writeable(uaddr
))
2970 * Wake robust non-PI futexes here. The wakeup of
2971 * PI futexes happens in exit_pi_state():
2973 if (!pi
&& (uval
& FUTEX_WAITERS
))
2974 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
2980 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2982 static inline int fetch_robust_entry(struct robust_list __user
**entry
,
2983 struct robust_list __user
* __user
*head
,
2986 unsigned long uentry
;
2988 if (get_user(uentry
, (unsigned long __user
*)head
))
2991 *entry
= (void __user
*)(uentry
& ~1UL);
2998 * Walk curr->robust_list (very carefully, it's a userspace list!)
2999 * and mark any locks found there dead, and notify any waiters.
3001 * We silently return on any sign of list-walking problem.
3003 void exit_robust_list(struct task_struct
*curr
)
3005 struct robust_list_head __user
*head
= curr
->robust_list
;
3006 struct robust_list __user
*entry
, *next_entry
, *pending
;
3007 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
3008 unsigned int uninitialized_var(next_pi
);
3009 unsigned long futex_offset
;
3012 if (!futex_cmpxchg_enabled
)
3016 * Fetch the list head (which was registered earlier, via
3017 * sys_set_robust_list()):
3019 if (fetch_robust_entry(&entry
, &head
->list
.next
, &pi
))
3022 * Fetch the relative futex offset:
3024 if (get_user(futex_offset
, &head
->futex_offset
))
3027 * Fetch any possibly pending lock-add first, and handle it
3030 if (fetch_robust_entry(&pending
, &head
->list_op_pending
, &pip
))
3033 next_entry
= NULL
; /* avoid warning with gcc */
3034 while (entry
!= &head
->list
) {
3036 * Fetch the next entry in the list before calling
3037 * handle_futex_death:
3039 rc
= fetch_robust_entry(&next_entry
, &entry
->next
, &next_pi
);
3041 * A pending lock might already be on the list, so
3042 * don't process it twice:
3044 if (entry
!= pending
)
3045 if (handle_futex_death((void __user
*)entry
+ futex_offset
,
3053 * Avoid excessively long or circular lists:
3062 handle_futex_death((void __user
*)pending
+ futex_offset
,
3066 long do_futex(u32 __user
*uaddr
, int op
, u32 val
, ktime_t
*timeout
,
3067 u32 __user
*uaddr2
, u32 val2
, u32 val3
)
3069 int cmd
= op
& FUTEX_CMD_MASK
;
3070 unsigned int flags
= 0;
3072 if (!(op
& FUTEX_PRIVATE_FLAG
))
3073 flags
|= FLAGS_SHARED
;
3075 if (op
& FUTEX_CLOCK_REALTIME
) {
3076 flags
|= FLAGS_CLOCKRT
;
3077 if (cmd
!= FUTEX_WAIT_BITSET
&& cmd
!= FUTEX_WAIT_REQUEUE_PI
)
3083 case FUTEX_UNLOCK_PI
:
3084 case FUTEX_TRYLOCK_PI
:
3085 case FUTEX_WAIT_REQUEUE_PI
:
3086 case FUTEX_CMP_REQUEUE_PI
:
3087 if (!futex_cmpxchg_enabled
)
3093 val3
= FUTEX_BITSET_MATCH_ANY
;
3094 case FUTEX_WAIT_BITSET
:
3095 return futex_wait(uaddr
, flags
, val
, timeout
, val3
);
3097 val3
= FUTEX_BITSET_MATCH_ANY
;
3098 case FUTEX_WAKE_BITSET
:
3099 return futex_wake(uaddr
, flags
, val
, val3
);
3101 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, NULL
, 0);
3102 case FUTEX_CMP_REQUEUE
:
3103 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 0);
3105 return futex_wake_op(uaddr
, flags
, uaddr2
, val
, val2
, val3
);
3107 return futex_lock_pi(uaddr
, flags
, timeout
, 0);
3108 case FUTEX_UNLOCK_PI
:
3109 return futex_unlock_pi(uaddr
, flags
);
3110 case FUTEX_TRYLOCK_PI
:
3111 return futex_lock_pi(uaddr
, flags
, NULL
, 1);
3112 case FUTEX_WAIT_REQUEUE_PI
:
3113 val3
= FUTEX_BITSET_MATCH_ANY
;
3114 return futex_wait_requeue_pi(uaddr
, flags
, val
, timeout
, val3
,
3116 case FUTEX_CMP_REQUEUE_PI
:
3117 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 1);
3123 SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
3124 struct timespec __user
*, utime
, u32 __user
*, uaddr2
,
3128 ktime_t t
, *tp
= NULL
;
3130 int cmd
= op
& FUTEX_CMD_MASK
;
3132 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
3133 cmd
== FUTEX_WAIT_BITSET
||
3134 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
3135 if (unlikely(should_fail_futex(!(op
& FUTEX_PRIVATE_FLAG
))))
3137 if (copy_from_user(&ts
, utime
, sizeof(ts
)) != 0)
3139 if (!timespec_valid(&ts
))
3142 t
= timespec_to_ktime(ts
);
3143 if (cmd
== FUTEX_WAIT
)
3144 t
= ktime_add_safe(ktime_get(), t
);
3148 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3149 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3151 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
3152 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
3153 val2
= (u32
) (unsigned long) utime
;
3155 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
3158 static void __init
futex_detect_cmpxchg(void)
3160 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3164 * This will fail and we want it. Some arch implementations do
3165 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3166 * functionality. We want to know that before we call in any
3167 * of the complex code paths. Also we want to prevent
3168 * registration of robust lists in that case. NULL is
3169 * guaranteed to fault and we get -EFAULT on functional
3170 * implementation, the non-functional ones will return
3173 if (cmpxchg_futex_value_locked(&curval
, NULL
, 0, 0) == -EFAULT
)
3174 futex_cmpxchg_enabled
= 1;
3178 static int __init
futex_init(void)
3180 unsigned int futex_shift
;
3183 #if CONFIG_BASE_SMALL
3184 futex_hashsize
= 16;
3186 futex_hashsize
= roundup_pow_of_two(256 * num_possible_cpus());
3189 futex_queues
= alloc_large_system_hash("futex", sizeof(*futex_queues
),
3191 futex_hashsize
< 256 ? HASH_SMALL
: 0,
3193 futex_hashsize
, futex_hashsize
);
3194 futex_hashsize
= 1UL << futex_shift
;
3196 futex_detect_cmpxchg();
3198 for (i
= 0; i
< futex_hashsize
; i
++) {
3199 atomic_set(&futex_queues
[i
].waiters
, 0);
3200 plist_head_init(&futex_queues
[i
].chain
);
3201 spin_lock_init(&futex_queues
[i
].lock
);
3206 core_initcall(futex_init
);