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/compat.h>
48 #include <linux/slab.h>
49 #include <linux/poll.h>
51 #include <linux/file.h>
52 #include <linux/jhash.h>
53 #include <linux/init.h>
54 #include <linux/futex.h>
55 #include <linux/mount.h>
56 #include <linux/pagemap.h>
57 #include <linux/syscalls.h>
58 #include <linux/signal.h>
59 #include <linux/export.h>
60 #include <linux/magic.h>
61 #include <linux/pid.h>
62 #include <linux/nsproxy.h>
63 #include <linux/ptrace.h>
64 #include <linux/sched/rt.h>
65 #include <linux/sched/wake_q.h>
66 #include <linux/sched/mm.h>
67 #include <linux/hugetlb.h>
68 #include <linux/freezer.h>
69 #include <linux/bootmem.h>
70 #include <linux/fault-inject.h>
72 #include <asm/futex.h>
74 #include "locking/rtmutex_common.h"
77 * READ this before attempting to hack on futexes!
79 * Basic futex operation and ordering guarantees
80 * =============================================
82 * The waiter reads the futex value in user space and calls
83 * futex_wait(). This function computes the hash bucket and acquires
84 * the hash bucket lock. After that it reads the futex user space value
85 * again and verifies that the data has not changed. If it has not changed
86 * it enqueues itself into the hash bucket, releases the hash bucket lock
89 * The waker side modifies the user space value of the futex and calls
90 * futex_wake(). This function computes the hash bucket and acquires the
91 * hash bucket lock. Then it looks for waiters on that futex in the hash
92 * bucket and wakes them.
94 * In futex wake up scenarios where no tasks are blocked on a futex, taking
95 * the hb spinlock can be avoided and simply return. In order for this
96 * optimization to work, ordering guarantees must exist so that the waiter
97 * being added to the list is acknowledged when the list is concurrently being
98 * checked by the waker, avoiding scenarios like the following:
102 * sys_futex(WAIT, futex, val);
103 * futex_wait(futex, val);
106 * sys_futex(WAKE, futex);
111 * lock(hash_bucket(futex));
113 * unlock(hash_bucket(futex));
116 * This would cause the waiter on CPU 0 to wait forever because it
117 * missed the transition of the user space value from val to newval
118 * and the waker did not find the waiter in the hash bucket queue.
120 * The correct serialization ensures that a waiter either observes
121 * the changed user space value before blocking or is woken by a
126 * sys_futex(WAIT, futex, val);
127 * futex_wait(futex, val);
130 * smp_mb(); (A) <-- paired with -.
132 * lock(hash_bucket(futex)); |
136 * | sys_futex(WAKE, futex);
137 * | futex_wake(futex);
139 * `--------> smp_mb(); (B)
142 * unlock(hash_bucket(futex));
143 * schedule(); if (waiters)
144 * lock(hash_bucket(futex));
145 * else wake_waiters(futex);
146 * waiters--; (b) unlock(hash_bucket(futex));
148 * Where (A) orders the waiters increment and the futex value read through
149 * atomic operations (see hb_waiters_inc) and where (B) orders the write
150 * to futex and the waiters read -- this is done by the barriers for both
151 * shared and private futexes in get_futex_key_refs().
153 * This yields the following case (where X:=waiters, Y:=futex):
161 * Which guarantees that x==0 && y==0 is impossible; which translates back into
162 * the guarantee that we cannot both miss the futex variable change and the
165 * Note that a new waiter is accounted for in (a) even when it is possible that
166 * the wait call can return error, in which case we backtrack from it in (b).
167 * Refer to the comment in queue_lock().
169 * Similarly, in order to account for waiters being requeued on another
170 * address we always increment the waiters for the destination bucket before
171 * acquiring the lock. It then decrements them again after releasing it -
172 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
173 * will do the additional required waiter count housekeeping. This is done for
174 * double_lock_hb() and double_unlock_hb(), respectively.
177 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
178 #define futex_cmpxchg_enabled 1
180 static int __read_mostly futex_cmpxchg_enabled
;
184 * Futex flags used to encode options to functions and preserve them across
188 # define FLAGS_SHARED 0x01
191 * NOMMU does not have per process address space. Let the compiler optimize
194 # define FLAGS_SHARED 0x00
196 #define FLAGS_CLOCKRT 0x02
197 #define FLAGS_HAS_TIMEOUT 0x04
200 * Priority Inheritance state:
202 struct futex_pi_state
{
204 * list of 'owned' pi_state instances - these have to be
205 * cleaned up in do_exit() if the task exits prematurely:
207 struct list_head list
;
212 struct rt_mutex pi_mutex
;
214 struct task_struct
*owner
;
218 } __randomize_layout
;
221 * struct futex_q - The hashed futex queue entry, one per waiting task
222 * @list: priority-sorted list of tasks waiting on this futex
223 * @task: the task waiting on the futex
224 * @lock_ptr: the hash bucket lock
225 * @key: the key the futex is hashed on
226 * @pi_state: optional priority inheritance state
227 * @rt_waiter: rt_waiter storage for use with requeue_pi
228 * @requeue_pi_key: the requeue_pi target futex key
229 * @bitset: bitset for the optional bitmasked wakeup
231 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
232 * we can wake only the relevant ones (hashed queues may be shared).
234 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
235 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
236 * The order of wakeup is always to make the first condition true, then
239 * PI futexes are typically woken before they are removed from the hash list via
240 * the rt_mutex code. See unqueue_me_pi().
243 struct plist_node list
;
245 struct task_struct
*task
;
246 spinlock_t
*lock_ptr
;
248 struct futex_pi_state
*pi_state
;
249 struct rt_mutex_waiter
*rt_waiter
;
250 union futex_key
*requeue_pi_key
;
252 } __randomize_layout
;
254 static const struct futex_q futex_q_init
= {
255 /* list gets initialized in queue_me()*/
256 .key
= FUTEX_KEY_INIT
,
257 .bitset
= FUTEX_BITSET_MATCH_ANY
261 * Hash buckets are shared by all the futex_keys that hash to the same
262 * location. Each key may have multiple futex_q structures, one for each task
263 * waiting on a futex.
265 struct futex_hash_bucket
{
268 struct plist_head chain
;
269 } ____cacheline_aligned_in_smp
;
272 * The base of the bucket array and its size are always used together
273 * (after initialization only in hash_futex()), so ensure that they
274 * reside in the same cacheline.
277 struct futex_hash_bucket
*queues
;
278 unsigned long hashsize
;
279 } __futex_data __read_mostly
__aligned(2*sizeof(long));
280 #define futex_queues (__futex_data.queues)
281 #define futex_hashsize (__futex_data.hashsize)
285 * Fault injections for futexes.
287 #ifdef CONFIG_FAIL_FUTEX
290 struct fault_attr attr
;
294 .attr
= FAULT_ATTR_INITIALIZER
,
295 .ignore_private
= false,
298 static int __init
setup_fail_futex(char *str
)
300 return setup_fault_attr(&fail_futex
.attr
, str
);
302 __setup("fail_futex=", setup_fail_futex
);
304 static bool should_fail_futex(bool fshared
)
306 if (fail_futex
.ignore_private
&& !fshared
)
309 return should_fail(&fail_futex
.attr
, 1);
312 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
314 static int __init
fail_futex_debugfs(void)
316 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
319 dir
= fault_create_debugfs_attr("fail_futex", NULL
,
324 if (!debugfs_create_bool("ignore-private", mode
, dir
,
325 &fail_futex
.ignore_private
)) {
326 debugfs_remove_recursive(dir
);
333 late_initcall(fail_futex_debugfs
);
335 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
338 static inline bool should_fail_futex(bool fshared
)
342 #endif /* CONFIG_FAIL_FUTEX */
345 static void compat_exit_robust_list(struct task_struct
*curr
);
347 static inline void compat_exit_robust_list(struct task_struct
*curr
) { }
350 static inline void futex_get_mm(union futex_key
*key
)
352 mmgrab(key
->private.mm
);
354 * Ensure futex_get_mm() implies a full barrier such that
355 * get_futex_key() implies a full barrier. This is relied upon
356 * as smp_mb(); (B), see the ordering comment above.
358 smp_mb__after_atomic();
362 * Reflects a new waiter being added to the waitqueue.
364 static inline void hb_waiters_inc(struct futex_hash_bucket
*hb
)
367 atomic_inc(&hb
->waiters
);
369 * Full barrier (A), see the ordering comment above.
371 smp_mb__after_atomic();
376 * Reflects a waiter being removed from the waitqueue by wakeup
379 static inline void hb_waiters_dec(struct futex_hash_bucket
*hb
)
382 atomic_dec(&hb
->waiters
);
386 static inline int hb_waiters_pending(struct futex_hash_bucket
*hb
)
389 return atomic_read(&hb
->waiters
);
396 * hash_futex - Return the hash bucket in the global hash
397 * @key: Pointer to the futex key for which the hash is calculated
399 * We hash on the keys returned from get_futex_key (see below) and return the
400 * corresponding hash bucket in the global hash.
402 static struct futex_hash_bucket
*hash_futex(union futex_key
*key
)
404 u32 hash
= jhash2((u32
*)&key
->both
.word
,
405 (sizeof(key
->both
.word
)+sizeof(key
->both
.ptr
))/4,
407 return &futex_queues
[hash
& (futex_hashsize
- 1)];
412 * match_futex - Check whether two futex keys are equal
413 * @key1: Pointer to key1
414 * @key2: Pointer to key2
416 * Return 1 if two futex_keys are equal, 0 otherwise.
418 static inline int match_futex(union futex_key
*key1
, union futex_key
*key2
)
421 && key1
->both
.word
== key2
->both
.word
422 && key1
->both
.ptr
== key2
->both
.ptr
423 && key1
->both
.offset
== key2
->both
.offset
);
427 * Take a reference to the resource addressed by a key.
428 * Can be called while holding spinlocks.
431 static void get_futex_key_refs(union futex_key
*key
)
437 * On MMU less systems futexes are always "private" as there is no per
438 * process address space. We need the smp wmb nevertheless - yes,
439 * arch/blackfin has MMU less SMP ...
441 if (!IS_ENABLED(CONFIG_MMU
)) {
442 smp_mb(); /* explicit smp_mb(); (B) */
446 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
448 ihold(key
->shared
.inode
); /* implies smp_mb(); (B) */
450 case FUT_OFF_MMSHARED
:
451 futex_get_mm(key
); /* implies smp_mb(); (B) */
455 * Private futexes do not hold reference on an inode or
456 * mm, therefore the only purpose of calling get_futex_key_refs
457 * is because we need the barrier for the lockless waiter check.
459 smp_mb(); /* explicit smp_mb(); (B) */
464 * Drop a reference to the resource addressed by a key.
465 * The hash bucket spinlock must not be held. This is
466 * a no-op for private futexes, see comment in the get
469 static void drop_futex_key_refs(union futex_key
*key
)
471 if (!key
->both
.ptr
) {
472 /* If we're here then we tried to put a key we failed to get */
477 if (!IS_ENABLED(CONFIG_MMU
))
480 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
482 iput(key
->shared
.inode
);
484 case FUT_OFF_MMSHARED
:
485 mmdrop(key
->private.mm
);
491 * get_futex_key() - Get parameters which are the keys for a futex
492 * @uaddr: virtual address of the futex
493 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
494 * @key: address where result is stored.
495 * @rw: mapping needs to be read/write (values: VERIFY_READ,
498 * Return: a negative error code or 0
500 * The key words are stored in @key on success.
502 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
503 * offset_within_page). For private mappings, it's (uaddr, current->mm).
504 * We can usually work out the index without swapping in the page.
506 * lock_page() might sleep, the caller should not hold a spinlock.
509 get_futex_key(u32 __user
*uaddr
, int fshared
, union futex_key
*key
, int rw
)
511 unsigned long address
= (unsigned long)uaddr
;
512 struct mm_struct
*mm
= current
->mm
;
513 struct page
*page
, *tail
;
514 struct address_space
*mapping
;
518 * The futex address must be "naturally" aligned.
520 key
->both
.offset
= address
% PAGE_SIZE
;
521 if (unlikely((address
% sizeof(u32
)) != 0))
523 address
-= key
->both
.offset
;
525 if (unlikely(!access_ok(rw
, uaddr
, sizeof(u32
))))
528 if (unlikely(should_fail_futex(fshared
)))
532 * PROCESS_PRIVATE futexes are fast.
533 * As the mm cannot disappear under us and the 'key' only needs
534 * virtual address, we dont even have to find the underlying vma.
535 * Note : We do have to check 'uaddr' is a valid user address,
536 * but access_ok() should be faster than find_vma()
539 key
->private.mm
= mm
;
540 key
->private.address
= address
;
541 get_futex_key_refs(key
); /* implies smp_mb(); (B) */
546 /* Ignore any VERIFY_READ mapping (futex common case) */
547 if (unlikely(should_fail_futex(fshared
)))
550 err
= get_user_pages_fast(address
, 1, 1, &page
);
552 * If write access is not required (eg. FUTEX_WAIT), try
553 * and get read-only access.
555 if (err
== -EFAULT
&& rw
== VERIFY_READ
) {
556 err
= get_user_pages_fast(address
, 1, 0, &page
);
565 * The treatment of mapping from this point on is critical. The page
566 * lock protects many things but in this context the page lock
567 * stabilizes mapping, prevents inode freeing in the shared
568 * file-backed region case and guards against movement to swap cache.
570 * Strictly speaking the page lock is not needed in all cases being
571 * considered here and page lock forces unnecessarily serialization
572 * From this point on, mapping will be re-verified if necessary and
573 * page lock will be acquired only if it is unavoidable
575 * Mapping checks require the head page for any compound page so the
576 * head page and mapping is looked up now. For anonymous pages, it
577 * does not matter if the page splits in the future as the key is
578 * based on the address. For filesystem-backed pages, the tail is
579 * required as the index of the page determines the key. For
580 * base pages, there is no tail page and tail == page.
583 page
= compound_head(page
);
584 mapping
= READ_ONCE(page
->mapping
);
587 * If page->mapping is NULL, then it cannot be a PageAnon
588 * page; but it might be the ZERO_PAGE or in the gate area or
589 * in a special mapping (all cases which we are happy to fail);
590 * or it may have been a good file page when get_user_pages_fast
591 * found it, but truncated or holepunched or subjected to
592 * invalidate_complete_page2 before we got the page lock (also
593 * cases which we are happy to fail). And we hold a reference,
594 * so refcount care in invalidate_complete_page's remove_mapping
595 * prevents drop_caches from setting mapping to NULL beneath us.
597 * The case we do have to guard against is when memory pressure made
598 * shmem_writepage move it from filecache to swapcache beneath us:
599 * an unlikely race, but we do need to retry for page->mapping.
601 if (unlikely(!mapping
)) {
605 * Page lock is required to identify which special case above
606 * applies. If this is really a shmem page then the page lock
607 * will prevent unexpected transitions.
610 shmem_swizzled
= PageSwapCache(page
) || page
->mapping
;
621 * Private mappings are handled in a simple way.
623 * If the futex key is stored on an anonymous page, then the associated
624 * object is the mm which is implicitly pinned by the calling process.
626 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
627 * it's a read-only handle, it's expected that futexes attach to
628 * the object not the particular process.
630 if (PageAnon(page
)) {
632 * A RO anonymous page will never change and thus doesn't make
633 * sense for futex operations.
635 if (unlikely(should_fail_futex(fshared
)) || ro
) {
640 key
->both
.offset
|= FUT_OFF_MMSHARED
; /* ref taken on mm */
641 key
->private.mm
= mm
;
642 key
->private.address
= address
;
644 get_futex_key_refs(key
); /* implies smp_mb(); (B) */
650 * The associated futex object in this case is the inode and
651 * the page->mapping must be traversed. Ordinarily this should
652 * be stabilised under page lock but it's not strictly
653 * necessary in this case as we just want to pin the inode, not
654 * update the radix tree or anything like that.
656 * The RCU read lock is taken as the inode is finally freed
657 * under RCU. If the mapping still matches expectations then the
658 * mapping->host can be safely accessed as being a valid inode.
662 if (READ_ONCE(page
->mapping
) != mapping
) {
669 inode
= READ_ONCE(mapping
->host
);
678 * Take a reference unless it is about to be freed. Previously
679 * this reference was taken by ihold under the page lock
680 * pinning the inode in place so i_lock was unnecessary. The
681 * only way for this check to fail is if the inode was
682 * truncated in parallel which is almost certainly an
683 * application bug. In such a case, just retry.
685 * We are not calling into get_futex_key_refs() in file-backed
686 * cases, therefore a successful atomic_inc return below will
687 * guarantee that get_futex_key() will still imply smp_mb(); (B).
689 if (!atomic_inc_not_zero(&inode
->i_count
)) {
696 /* Should be impossible but lets be paranoid for now */
697 if (WARN_ON_ONCE(inode
->i_mapping
!= mapping
)) {
705 key
->both
.offset
|= FUT_OFF_INODE
; /* inode-based key */
706 key
->shared
.inode
= inode
;
707 key
->shared
.pgoff
= basepage_index(tail
);
716 static inline void put_futex_key(union futex_key
*key
)
718 drop_futex_key_refs(key
);
722 * fault_in_user_writeable() - Fault in user address and verify RW access
723 * @uaddr: pointer to faulting user space address
725 * Slow path to fixup the fault we just took in the atomic write
728 * We have no generic implementation of a non-destructive write to the
729 * user address. We know that we faulted in the atomic pagefault
730 * disabled section so we can as well avoid the #PF overhead by
731 * calling get_user_pages() right away.
733 static int fault_in_user_writeable(u32 __user
*uaddr
)
735 struct mm_struct
*mm
= current
->mm
;
738 down_read(&mm
->mmap_sem
);
739 ret
= fixup_user_fault(current
, mm
, (unsigned long)uaddr
,
740 FAULT_FLAG_WRITE
, NULL
);
741 up_read(&mm
->mmap_sem
);
743 return ret
< 0 ? ret
: 0;
747 * futex_top_waiter() - Return the highest priority waiter on a futex
748 * @hb: the hash bucket the futex_q's reside in
749 * @key: the futex key (to distinguish it from other futex futex_q's)
751 * Must be called with the hb lock held.
753 static struct futex_q
*futex_top_waiter(struct futex_hash_bucket
*hb
,
754 union futex_key
*key
)
756 struct futex_q
*this;
758 plist_for_each_entry(this, &hb
->chain
, list
) {
759 if (match_futex(&this->key
, key
))
765 static int cmpxchg_futex_value_locked(u32
*curval
, u32 __user
*uaddr
,
766 u32 uval
, u32 newval
)
771 ret
= futex_atomic_cmpxchg_inatomic(curval
, uaddr
, uval
, newval
);
777 static int get_futex_value_locked(u32
*dest
, u32 __user
*from
)
782 ret
= __get_user(*dest
, from
);
785 return ret
? -EFAULT
: 0;
792 static int refill_pi_state_cache(void)
794 struct futex_pi_state
*pi_state
;
796 if (likely(current
->pi_state_cache
))
799 pi_state
= kzalloc(sizeof(*pi_state
), GFP_KERNEL
);
804 INIT_LIST_HEAD(&pi_state
->list
);
805 /* pi_mutex gets initialized later */
806 pi_state
->owner
= NULL
;
807 atomic_set(&pi_state
->refcount
, 1);
808 pi_state
->key
= FUTEX_KEY_INIT
;
810 current
->pi_state_cache
= pi_state
;
815 static struct futex_pi_state
*alloc_pi_state(void)
817 struct futex_pi_state
*pi_state
= current
->pi_state_cache
;
820 current
->pi_state_cache
= NULL
;
825 static void get_pi_state(struct futex_pi_state
*pi_state
)
827 WARN_ON_ONCE(!atomic_inc_not_zero(&pi_state
->refcount
));
831 * Drops a reference to the pi_state object and frees or caches it
832 * when the last reference is gone.
834 static void put_pi_state(struct futex_pi_state
*pi_state
)
839 if (!atomic_dec_and_test(&pi_state
->refcount
))
843 * If pi_state->owner is NULL, the owner is most probably dying
844 * and has cleaned up the pi_state already
846 if (pi_state
->owner
) {
847 struct task_struct
*owner
;
849 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
850 owner
= pi_state
->owner
;
852 raw_spin_lock(&owner
->pi_lock
);
853 list_del_init(&pi_state
->list
);
854 raw_spin_unlock(&owner
->pi_lock
);
856 rt_mutex_proxy_unlock(&pi_state
->pi_mutex
, owner
);
857 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
860 if (current
->pi_state_cache
) {
864 * pi_state->list is already empty.
865 * clear pi_state->owner.
866 * refcount is at 0 - put it back to 1.
868 pi_state
->owner
= NULL
;
869 atomic_set(&pi_state
->refcount
, 1);
870 current
->pi_state_cache
= pi_state
;
875 * Look up the task based on what TID userspace gave us.
878 static struct task_struct
*futex_find_get_task(pid_t pid
)
880 struct task_struct
*p
;
883 p
= find_task_by_vpid(pid
);
892 #ifdef CONFIG_FUTEX_PI
895 * This task is holding PI mutexes at exit time => bad.
896 * Kernel cleans up PI-state, but userspace is likely hosed.
897 * (Robust-futex cleanup is separate and might save the day for userspace.)
899 static void exit_pi_state_list(struct task_struct
*curr
)
901 struct list_head
*next
, *head
= &curr
->pi_state_list
;
902 struct futex_pi_state
*pi_state
;
903 struct futex_hash_bucket
*hb
;
904 union futex_key key
= FUTEX_KEY_INIT
;
906 if (!futex_cmpxchg_enabled
)
909 * We are a ZOMBIE and nobody can enqueue itself on
910 * pi_state_list anymore, but we have to be careful
911 * versus waiters unqueueing themselves:
913 raw_spin_lock_irq(&curr
->pi_lock
);
914 while (!list_empty(head
)) {
916 pi_state
= list_entry(next
, struct futex_pi_state
, list
);
918 hb
= hash_futex(&key
);
921 * We can race against put_pi_state() removing itself from the
922 * list (a waiter going away). put_pi_state() will first
923 * decrement the reference count and then modify the list, so
924 * its possible to see the list entry but fail this reference
927 * In that case; drop the locks to let put_pi_state() make
928 * progress and retry the loop.
930 if (!atomic_inc_not_zero(&pi_state
->refcount
)) {
931 raw_spin_unlock_irq(&curr
->pi_lock
);
933 raw_spin_lock_irq(&curr
->pi_lock
);
936 raw_spin_unlock_irq(&curr
->pi_lock
);
938 spin_lock(&hb
->lock
);
939 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
940 raw_spin_lock(&curr
->pi_lock
);
942 * We dropped the pi-lock, so re-check whether this
943 * task still owns the PI-state:
945 if (head
->next
!= next
) {
946 /* retain curr->pi_lock for the loop invariant */
947 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
948 spin_unlock(&hb
->lock
);
949 put_pi_state(pi_state
);
953 WARN_ON(pi_state
->owner
!= curr
);
954 WARN_ON(list_empty(&pi_state
->list
));
955 list_del_init(&pi_state
->list
);
956 pi_state
->owner
= NULL
;
958 raw_spin_unlock(&curr
->pi_lock
);
959 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
960 spin_unlock(&hb
->lock
);
962 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
963 put_pi_state(pi_state
);
965 raw_spin_lock_irq(&curr
->pi_lock
);
967 raw_spin_unlock_irq(&curr
->pi_lock
);
970 static inline void exit_pi_state_list(struct task_struct
*curr
) { }
974 * We need to check the following states:
976 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
978 * [1] NULL | --- | --- | 0 | 0/1 | Valid
979 * [2] NULL | --- | --- | >0 | 0/1 | Valid
981 * [3] Found | NULL | -- | Any | 0/1 | Invalid
983 * [4] Found | Found | NULL | 0 | 1 | Valid
984 * [5] Found | Found | NULL | >0 | 1 | Invalid
986 * [6] Found | Found | task | 0 | 1 | Valid
988 * [7] Found | Found | NULL | Any | 0 | Invalid
990 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
991 * [9] Found | Found | task | 0 | 0 | Invalid
992 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
994 * [1] Indicates that the kernel can acquire the futex atomically. We
995 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
997 * [2] Valid, if TID does not belong to a kernel thread. If no matching
998 * thread is found then it indicates that the owner TID has died.
1000 * [3] Invalid. The waiter is queued on a non PI futex
1002 * [4] Valid state after exit_robust_list(), which sets the user space
1003 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
1005 * [5] The user space value got manipulated between exit_robust_list()
1006 * and exit_pi_state_list()
1008 * [6] Valid state after exit_pi_state_list() which sets the new owner in
1009 * the pi_state but cannot access the user space value.
1011 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
1013 * [8] Owner and user space value match
1015 * [9] There is no transient state which sets the user space TID to 0
1016 * except exit_robust_list(), but this is indicated by the
1017 * FUTEX_OWNER_DIED bit. See [4]
1019 * [10] There is no transient state which leaves owner and user space
1023 * Serialization and lifetime rules:
1027 * hb -> futex_q, relation
1028 * futex_q -> pi_state, relation
1030 * (cannot be raw because hb can contain arbitrary amount
1033 * pi_mutex->wait_lock:
1037 * (and pi_mutex 'obviously')
1041 * p->pi_state_list -> pi_state->list, relation
1043 * pi_state->refcount:
1051 * pi_mutex->wait_lock
1057 * Validate that the existing waiter has a pi_state and sanity check
1058 * the pi_state against the user space value. If correct, attach to
1061 static int attach_to_pi_state(u32 __user
*uaddr
, u32 uval
,
1062 struct futex_pi_state
*pi_state
,
1063 struct futex_pi_state
**ps
)
1065 pid_t pid
= uval
& FUTEX_TID_MASK
;
1070 * Userspace might have messed up non-PI and PI futexes [3]
1072 if (unlikely(!pi_state
))
1076 * We get here with hb->lock held, and having found a
1077 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1078 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1079 * which in turn means that futex_lock_pi() still has a reference on
1082 * The waiter holding a reference on @pi_state also protects against
1083 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1084 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1085 * free pi_state before we can take a reference ourselves.
1087 WARN_ON(!atomic_read(&pi_state
->refcount
));
1090 * Now that we have a pi_state, we can acquire wait_lock
1091 * and do the state validation.
1093 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
1096 * Since {uval, pi_state} is serialized by wait_lock, and our current
1097 * uval was read without holding it, it can have changed. Verify it
1098 * still is what we expect it to be, otherwise retry the entire
1101 if (get_futex_value_locked(&uval2
, uaddr
))
1108 * Handle the owner died case:
1110 if (uval
& FUTEX_OWNER_DIED
) {
1112 * exit_pi_state_list sets owner to NULL and wakes the
1113 * topmost waiter. The task which acquires the
1114 * pi_state->rt_mutex will fixup owner.
1116 if (!pi_state
->owner
) {
1118 * No pi state owner, but the user space TID
1119 * is not 0. Inconsistent state. [5]
1124 * Take a ref on the state and return success. [4]
1130 * If TID is 0, then either the dying owner has not
1131 * yet executed exit_pi_state_list() or some waiter
1132 * acquired the rtmutex in the pi state, but did not
1133 * yet fixup the TID in user space.
1135 * Take a ref on the state and return success. [6]
1141 * If the owner died bit is not set, then the pi_state
1142 * must have an owner. [7]
1144 if (!pi_state
->owner
)
1149 * Bail out if user space manipulated the futex value. If pi
1150 * state exists then the owner TID must be the same as the
1151 * user space TID. [9/10]
1153 if (pid
!= task_pid_vnr(pi_state
->owner
))
1157 get_pi_state(pi_state
);
1158 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1175 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1180 * wait_for_owner_exiting - Block until the owner has exited
1181 * @exiting: Pointer to the exiting task
1183 * Caller must hold a refcount on @exiting.
1185 static void wait_for_owner_exiting(int ret
, struct task_struct
*exiting
)
1187 if (ret
!= -EBUSY
) {
1188 WARN_ON_ONCE(exiting
);
1192 if (WARN_ON_ONCE(ret
== -EBUSY
&& !exiting
))
1195 mutex_lock(&exiting
->futex_exit_mutex
);
1197 * No point in doing state checking here. If the waiter got here
1198 * while the task was in exec()->exec_futex_release() then it can
1199 * have any FUTEX_STATE_* value when the waiter has acquired the
1200 * mutex. OK, if running, EXITING or DEAD if it reached exit()
1201 * already. Highly unlikely and not a problem. Just one more round
1202 * through the futex maze.
1204 mutex_unlock(&exiting
->futex_exit_mutex
);
1206 put_task_struct(exiting
);
1209 static int handle_exit_race(u32 __user
*uaddr
, u32 uval
,
1210 struct task_struct
*tsk
)
1215 * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1216 * caller that the alleged owner is busy.
1218 if (tsk
&& tsk
->futex_state
!= FUTEX_STATE_DEAD
)
1222 * Reread the user space value to handle the following situation:
1226 * sys_exit() sys_futex()
1227 * do_exit() futex_lock_pi()
1228 * futex_lock_pi_atomic()
1229 * exit_signals(tsk) No waiters:
1230 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1231 * mm_release(tsk) Set waiter bit
1232 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1233 * Set owner died attach_to_pi_owner() {
1234 * *uaddr = 0xC0000000; tsk = get_task(PID);
1235 * } if (!tsk->flags & PF_EXITING) {
1237 * tsk->futex_state = } else {
1238 * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1241 * return -ESRCH; <--- FAIL
1244 * Returning ESRCH unconditionally is wrong here because the
1245 * user space value has been changed by the exiting task.
1247 * The same logic applies to the case where the exiting task is
1250 if (get_futex_value_locked(&uval2
, uaddr
))
1253 /* If the user space value has changed, try again. */
1258 * The exiting task did not have a robust list, the robust list was
1259 * corrupted or the user space value in *uaddr is simply bogus.
1260 * Give up and tell user space.
1266 * Lookup the task for the TID provided from user space and attach to
1267 * it after doing proper sanity checks.
1269 static int attach_to_pi_owner(u32 __user
*uaddr
, u32 uval
, union futex_key
*key
,
1270 struct futex_pi_state
**ps
,
1271 struct task_struct
**exiting
)
1273 pid_t pid
= uval
& FUTEX_TID_MASK
;
1274 struct futex_pi_state
*pi_state
;
1275 struct task_struct
*p
;
1278 * We are the first waiter - try to look up the real owner and attach
1279 * the new pi_state to it, but bail out when TID = 0 [1]
1281 * The !pid check is paranoid. None of the call sites should end up
1282 * with pid == 0, but better safe than sorry. Let the caller retry
1286 p
= futex_find_get_task(pid
);
1288 return handle_exit_race(uaddr
, uval
, NULL
);
1290 if (unlikely(p
->flags
& PF_KTHREAD
)) {
1296 * We need to look at the task state to figure out, whether the
1297 * task is exiting. To protect against the change of the task state
1298 * in futex_exit_release(), we do this protected by p->pi_lock:
1300 raw_spin_lock_irq(&p
->pi_lock
);
1301 if (unlikely(p
->futex_state
!= FUTEX_STATE_OK
)) {
1303 * The task is on the way out. When the futex state is
1304 * FUTEX_STATE_DEAD, we know that the task has finished
1307 int ret
= handle_exit_race(uaddr
, uval
, p
);
1309 raw_spin_unlock_irq(&p
->pi_lock
);
1311 * If the owner task is between FUTEX_STATE_EXITING and
1312 * FUTEX_STATE_DEAD then store the task pointer and keep
1313 * the reference on the task struct. The calling code will
1314 * drop all locks, wait for the task to reach
1315 * FUTEX_STATE_DEAD and then drop the refcount. This is
1316 * required to prevent a live lock when the current task
1317 * preempted the exiting task between the two states.
1327 * No existing pi state. First waiter. [2]
1329 * This creates pi_state, we have hb->lock held, this means nothing can
1330 * observe this state, wait_lock is irrelevant.
1332 pi_state
= alloc_pi_state();
1335 * Initialize the pi_mutex in locked state and make @p
1338 rt_mutex_init_proxy_locked(&pi_state
->pi_mutex
, p
);
1340 /* Store the key for possible exit cleanups: */
1341 pi_state
->key
= *key
;
1343 WARN_ON(!list_empty(&pi_state
->list
));
1344 list_add(&pi_state
->list
, &p
->pi_state_list
);
1346 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1347 * because there is no concurrency as the object is not published yet.
1349 pi_state
->owner
= p
;
1350 raw_spin_unlock_irq(&p
->pi_lock
);
1359 static int lookup_pi_state(u32 __user
*uaddr
, u32 uval
,
1360 struct futex_hash_bucket
*hb
,
1361 union futex_key
*key
, struct futex_pi_state
**ps
,
1362 struct task_struct
**exiting
)
1364 struct futex_q
*top_waiter
= futex_top_waiter(hb
, key
);
1367 * If there is a waiter on that futex, validate it and
1368 * attach to the pi_state when the validation succeeds.
1371 return attach_to_pi_state(uaddr
, uval
, top_waiter
->pi_state
, ps
);
1374 * We are the first waiter - try to look up the owner based on
1375 * @uval and attach to it.
1377 return attach_to_pi_owner(uaddr
, uval
, key
, ps
, exiting
);
1380 static int lock_pi_update_atomic(u32 __user
*uaddr
, u32 uval
, u32 newval
)
1383 u32
uninitialized_var(curval
);
1385 if (unlikely(should_fail_futex(true)))
1388 err
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
);
1392 /* If user space value changed, let the caller retry */
1393 return curval
!= uval
? -EAGAIN
: 0;
1397 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1398 * @uaddr: the pi futex user address
1399 * @hb: the pi futex hash bucket
1400 * @key: the futex key associated with uaddr and hb
1401 * @ps: the pi_state pointer where we store the result of the
1403 * @task: the task to perform the atomic lock work for. This will
1404 * be "current" except in the case of requeue pi.
1405 * @exiting: Pointer to store the task pointer of the owner task
1406 * which is in the middle of exiting
1407 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1410 * - 0 - ready to wait;
1411 * - 1 - acquired the lock;
1414 * The hb->lock and futex_key refs shall be held by the caller.
1416 * @exiting is only set when the return value is -EBUSY. If so, this holds
1417 * a refcount on the exiting task on return and the caller needs to drop it
1418 * after waiting for the exit to complete.
1420 static int futex_lock_pi_atomic(u32 __user
*uaddr
, struct futex_hash_bucket
*hb
,
1421 union futex_key
*key
,
1422 struct futex_pi_state
**ps
,
1423 struct task_struct
*task
,
1424 struct task_struct
**exiting
,
1427 u32 uval
, newval
, vpid
= task_pid_vnr(task
);
1428 struct futex_q
*top_waiter
;
1432 * Read the user space value first so we can validate a few
1433 * things before proceeding further.
1435 if (get_futex_value_locked(&uval
, uaddr
))
1438 if (unlikely(should_fail_futex(true)))
1444 if ((unlikely((uval
& FUTEX_TID_MASK
) == vpid
)))
1447 if ((unlikely(should_fail_futex(true))))
1451 * Lookup existing state first. If it exists, try to attach to
1454 top_waiter
= futex_top_waiter(hb
, key
);
1456 return attach_to_pi_state(uaddr
, uval
, top_waiter
->pi_state
, ps
);
1459 * No waiter and user TID is 0. We are here because the
1460 * waiters or the owner died bit is set or called from
1461 * requeue_cmp_pi or for whatever reason something took the
1464 if (!(uval
& FUTEX_TID_MASK
)) {
1466 * We take over the futex. No other waiters and the user space
1467 * TID is 0. We preserve the owner died bit.
1469 newval
= uval
& FUTEX_OWNER_DIED
;
1472 /* The futex requeue_pi code can enforce the waiters bit */
1474 newval
|= FUTEX_WAITERS
;
1476 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1477 /* If the take over worked, return 1 */
1478 return ret
< 0 ? ret
: 1;
1482 * First waiter. Set the waiters bit before attaching ourself to
1483 * the owner. If owner tries to unlock, it will be forced into
1484 * the kernel and blocked on hb->lock.
1486 newval
= uval
| FUTEX_WAITERS
;
1487 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1491 * If the update of the user space value succeeded, we try to
1492 * attach to the owner. If that fails, no harm done, we only
1493 * set the FUTEX_WAITERS bit in the user space variable.
1495 return attach_to_pi_owner(uaddr
, newval
, key
, ps
, exiting
);
1499 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1500 * @q: The futex_q to unqueue
1502 * The q->lock_ptr must not be NULL and must be held by the caller.
1504 static void __unqueue_futex(struct futex_q
*q
)
1506 struct futex_hash_bucket
*hb
;
1508 if (WARN_ON_SMP(!q
->lock_ptr
|| !spin_is_locked(q
->lock_ptr
))
1509 || WARN_ON(plist_node_empty(&q
->list
)))
1512 hb
= container_of(q
->lock_ptr
, struct futex_hash_bucket
, lock
);
1513 plist_del(&q
->list
, &hb
->chain
);
1518 * The hash bucket lock must be held when this is called.
1519 * Afterwards, the futex_q must not be accessed. Callers
1520 * must ensure to later call wake_up_q() for the actual
1523 static void mark_wake_futex(struct wake_q_head
*wake_q
, struct futex_q
*q
)
1525 struct task_struct
*p
= q
->task
;
1527 if (WARN(q
->pi_state
|| q
->rt_waiter
, "refusing to wake PI futex\n"))
1533 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1534 * is written, without taking any locks. This is possible in the event
1535 * of a spurious wakeup, for example. A memory barrier is required here
1536 * to prevent the following store to lock_ptr from getting ahead of the
1537 * plist_del in __unqueue_futex().
1539 smp_store_release(&q
->lock_ptr
, NULL
);
1542 * Queue the task for later wakeup for after we've released
1543 * the hb->lock. wake_q_add() grabs reference to p.
1545 wake_q_add(wake_q
, p
);
1550 * Caller must hold a reference on @pi_state.
1552 static int wake_futex_pi(u32 __user
*uaddr
, u32 uval
, struct futex_pi_state
*pi_state
)
1554 u32
uninitialized_var(curval
), newval
;
1555 struct task_struct
*new_owner
;
1556 bool postunlock
= false;
1557 DEFINE_WAKE_Q(wake_q
);
1560 new_owner
= rt_mutex_next_owner(&pi_state
->pi_mutex
);
1561 if (WARN_ON_ONCE(!new_owner
)) {
1563 * As per the comment in futex_unlock_pi() this should not happen.
1565 * When this happens, give up our locks and try again, giving
1566 * the futex_lock_pi() instance time to complete, either by
1567 * waiting on the rtmutex or removing itself from the futex
1575 * We pass it to the next owner. The WAITERS bit is always kept
1576 * enabled while there is PI state around. We cleanup the owner
1577 * died bit, because we are the owner.
1579 newval
= FUTEX_WAITERS
| task_pid_vnr(new_owner
);
1581 if (unlikely(should_fail_futex(true)))
1584 ret
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
);
1585 if (!ret
&& (curval
!= uval
)) {
1587 * If a unconditional UNLOCK_PI operation (user space did not
1588 * try the TID->0 transition) raced with a waiter setting the
1589 * FUTEX_WAITERS flag between get_user() and locking the hash
1590 * bucket lock, retry the operation.
1592 if ((FUTEX_TID_MASK
& curval
) == uval
)
1602 * This is a point of no return; once we modify the uval there is no
1603 * going back and subsequent operations must not fail.
1606 raw_spin_lock(&pi_state
->owner
->pi_lock
);
1607 WARN_ON(list_empty(&pi_state
->list
));
1608 list_del_init(&pi_state
->list
);
1609 raw_spin_unlock(&pi_state
->owner
->pi_lock
);
1611 raw_spin_lock(&new_owner
->pi_lock
);
1612 WARN_ON(!list_empty(&pi_state
->list
));
1613 list_add(&pi_state
->list
, &new_owner
->pi_state_list
);
1614 pi_state
->owner
= new_owner
;
1615 raw_spin_unlock(&new_owner
->pi_lock
);
1617 postunlock
= __rt_mutex_futex_unlock(&pi_state
->pi_mutex
, &wake_q
);
1620 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1623 rt_mutex_postunlock(&wake_q
);
1629 * Express the locking dependencies for lockdep:
1632 double_lock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1635 spin_lock(&hb1
->lock
);
1637 spin_lock_nested(&hb2
->lock
, SINGLE_DEPTH_NESTING
);
1638 } else { /* hb1 > hb2 */
1639 spin_lock(&hb2
->lock
);
1640 spin_lock_nested(&hb1
->lock
, SINGLE_DEPTH_NESTING
);
1645 double_unlock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1647 spin_unlock(&hb1
->lock
);
1649 spin_unlock(&hb2
->lock
);
1653 * Wake up waiters matching bitset queued on this futex (uaddr).
1656 futex_wake(u32 __user
*uaddr
, unsigned int flags
, int nr_wake
, u32 bitset
)
1658 struct futex_hash_bucket
*hb
;
1659 struct futex_q
*this, *next
;
1660 union futex_key key
= FUTEX_KEY_INIT
;
1662 DEFINE_WAKE_Q(wake_q
);
1667 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_READ
);
1668 if (unlikely(ret
!= 0))
1671 hb
= hash_futex(&key
);
1673 /* Make sure we really have tasks to wakeup */
1674 if (!hb_waiters_pending(hb
))
1677 spin_lock(&hb
->lock
);
1679 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
1680 if (match_futex (&this->key
, &key
)) {
1681 if (this->pi_state
|| this->rt_waiter
) {
1686 /* Check if one of the bits is set in both bitsets */
1687 if (!(this->bitset
& bitset
))
1690 mark_wake_futex(&wake_q
, this);
1691 if (++ret
>= nr_wake
)
1696 spin_unlock(&hb
->lock
);
1699 put_futex_key(&key
);
1704 static int futex_atomic_op_inuser(unsigned int encoded_op
, u32 __user
*uaddr
)
1706 unsigned int op
= (encoded_op
& 0x70000000) >> 28;
1707 unsigned int cmp
= (encoded_op
& 0x0f000000) >> 24;
1708 int oparg
= sign_extend32((encoded_op
& 0x00fff000) >> 12, 11);
1709 int cmparg
= sign_extend32(encoded_op
& 0x00000fff, 11);
1712 if (encoded_op
& (FUTEX_OP_OPARG_SHIFT
<< 28)) {
1713 if (oparg
< 0 || oparg
> 31) {
1714 char comm
[sizeof(current
->comm
)];
1716 * kill this print and return -EINVAL when userspace
1719 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1720 get_task_comm(comm
, current
), oparg
);
1726 if (!access_ok(VERIFY_WRITE
, uaddr
, sizeof(u32
)))
1729 ret
= arch_futex_atomic_op_inuser(op
, oparg
, &oldval
, uaddr
);
1734 case FUTEX_OP_CMP_EQ
:
1735 return oldval
== cmparg
;
1736 case FUTEX_OP_CMP_NE
:
1737 return oldval
!= cmparg
;
1738 case FUTEX_OP_CMP_LT
:
1739 return oldval
< cmparg
;
1740 case FUTEX_OP_CMP_GE
:
1741 return oldval
>= cmparg
;
1742 case FUTEX_OP_CMP_LE
:
1743 return oldval
<= cmparg
;
1744 case FUTEX_OP_CMP_GT
:
1745 return oldval
> cmparg
;
1752 * Wake up all waiters hashed on the physical page that is mapped
1753 * to this virtual address:
1756 futex_wake_op(u32 __user
*uaddr1
, unsigned int flags
, u32 __user
*uaddr2
,
1757 int nr_wake
, int nr_wake2
, int op
)
1759 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1760 struct futex_hash_bucket
*hb1
, *hb2
;
1761 struct futex_q
*this, *next
;
1763 DEFINE_WAKE_Q(wake_q
);
1766 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1767 if (unlikely(ret
!= 0))
1769 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
1770 if (unlikely(ret
!= 0))
1773 hb1
= hash_futex(&key1
);
1774 hb2
= hash_futex(&key2
);
1777 double_lock_hb(hb1
, hb2
);
1778 op_ret
= futex_atomic_op_inuser(op
, uaddr2
);
1779 if (unlikely(op_ret
< 0)) {
1780 double_unlock_hb(hb1
, hb2
);
1782 if (!IS_ENABLED(CONFIG_MMU
) ||
1783 unlikely(op_ret
!= -EFAULT
&& op_ret
!= -EAGAIN
)) {
1785 * we don't get EFAULT from MMU faults if we don't have
1786 * an MMU, but we might get them from range checking
1792 if (op_ret
== -EFAULT
) {
1793 ret
= fault_in_user_writeable(uaddr2
);
1798 if (!(flags
& FLAGS_SHARED
)) {
1803 put_futex_key(&key2
);
1804 put_futex_key(&key1
);
1809 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1810 if (match_futex (&this->key
, &key1
)) {
1811 if (this->pi_state
|| this->rt_waiter
) {
1815 mark_wake_futex(&wake_q
, this);
1816 if (++ret
>= nr_wake
)
1823 plist_for_each_entry_safe(this, next
, &hb2
->chain
, list
) {
1824 if (match_futex (&this->key
, &key2
)) {
1825 if (this->pi_state
|| this->rt_waiter
) {
1829 mark_wake_futex(&wake_q
, this);
1830 if (++op_ret
>= nr_wake2
)
1838 double_unlock_hb(hb1
, hb2
);
1841 put_futex_key(&key2
);
1843 put_futex_key(&key1
);
1849 * requeue_futex() - Requeue a futex_q from one hb to another
1850 * @q: the futex_q to requeue
1851 * @hb1: the source hash_bucket
1852 * @hb2: the target hash_bucket
1853 * @key2: the new key for the requeued futex_q
1856 void requeue_futex(struct futex_q
*q
, struct futex_hash_bucket
*hb1
,
1857 struct futex_hash_bucket
*hb2
, union futex_key
*key2
)
1861 * If key1 and key2 hash to the same bucket, no need to
1864 if (likely(&hb1
->chain
!= &hb2
->chain
)) {
1865 plist_del(&q
->list
, &hb1
->chain
);
1866 hb_waiters_dec(hb1
);
1867 hb_waiters_inc(hb2
);
1868 plist_add(&q
->list
, &hb2
->chain
);
1869 q
->lock_ptr
= &hb2
->lock
;
1871 get_futex_key_refs(key2
);
1876 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1878 * @key: the key of the requeue target futex
1879 * @hb: the hash_bucket of the requeue target futex
1881 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1882 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1883 * to the requeue target futex so the waiter can detect the wakeup on the right
1884 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1885 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1886 * to protect access to the pi_state to fixup the owner later. Must be called
1887 * with both q->lock_ptr and hb->lock held.
1890 void requeue_pi_wake_futex(struct futex_q
*q
, union futex_key
*key
,
1891 struct futex_hash_bucket
*hb
)
1893 get_futex_key_refs(key
);
1898 WARN_ON(!q
->rt_waiter
);
1899 q
->rt_waiter
= NULL
;
1901 q
->lock_ptr
= &hb
->lock
;
1903 wake_up_state(q
->task
, TASK_NORMAL
);
1907 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1908 * @pifutex: the user address of the to futex
1909 * @hb1: the from futex hash bucket, must be locked by the caller
1910 * @hb2: the to futex hash bucket, must be locked by the caller
1911 * @key1: the from futex key
1912 * @key2: the to futex key
1913 * @ps: address to store the pi_state pointer
1914 * @exiting: Pointer to store the task pointer of the owner task
1915 * which is in the middle of exiting
1916 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1918 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1919 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1920 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1921 * hb1 and hb2 must be held by the caller.
1923 * @exiting is only set when the return value is -EBUSY. If so, this holds
1924 * a refcount on the exiting task on return and the caller needs to drop it
1925 * after waiting for the exit to complete.
1928 * - 0 - failed to acquire the lock atomically;
1929 * - >0 - acquired the lock, return value is vpid of the top_waiter
1933 futex_proxy_trylock_atomic(u32 __user
*pifutex
, struct futex_hash_bucket
*hb1
,
1934 struct futex_hash_bucket
*hb2
, union futex_key
*key1
,
1935 union futex_key
*key2
, struct futex_pi_state
**ps
,
1936 struct task_struct
**exiting
, int set_waiters
)
1938 struct futex_q
*top_waiter
= NULL
;
1942 if (get_futex_value_locked(&curval
, pifutex
))
1945 if (unlikely(should_fail_futex(true)))
1949 * Find the top_waiter and determine if there are additional waiters.
1950 * If the caller intends to requeue more than 1 waiter to pifutex,
1951 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1952 * as we have means to handle the possible fault. If not, don't set
1953 * the bit unecessarily as it will force the subsequent unlock to enter
1956 top_waiter
= futex_top_waiter(hb1
, key1
);
1958 /* There are no waiters, nothing for us to do. */
1962 /* Ensure we requeue to the expected futex. */
1963 if (!match_futex(top_waiter
->requeue_pi_key
, key2
))
1967 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1968 * the contended case or if set_waiters is 1. The pi_state is returned
1969 * in ps in contended cases.
1971 vpid
= task_pid_vnr(top_waiter
->task
);
1972 ret
= futex_lock_pi_atomic(pifutex
, hb2
, key2
, ps
, top_waiter
->task
,
1973 exiting
, set_waiters
);
1975 requeue_pi_wake_futex(top_waiter
, key2
, hb2
);
1982 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1983 * @uaddr1: source futex user address
1984 * @flags: futex flags (FLAGS_SHARED, etc.)
1985 * @uaddr2: target futex user address
1986 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1987 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1988 * @cmpval: @uaddr1 expected value (or %NULL)
1989 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1990 * pi futex (pi to pi requeue is not supported)
1992 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1993 * uaddr2 atomically on behalf of the top waiter.
1996 * - >=0 - on success, the number of tasks requeued or woken;
1999 static int futex_requeue(u32 __user
*uaddr1
, unsigned int flags
,
2000 u32 __user
*uaddr2
, int nr_wake
, int nr_requeue
,
2001 u32
*cmpval
, int requeue_pi
)
2003 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
2004 int drop_count
= 0, task_count
= 0, ret
;
2005 struct futex_pi_state
*pi_state
= NULL
;
2006 struct futex_hash_bucket
*hb1
, *hb2
;
2007 struct futex_q
*this, *next
;
2008 DEFINE_WAKE_Q(wake_q
);
2010 if (nr_wake
< 0 || nr_requeue
< 0)
2014 * When PI not supported: return -ENOSYS if requeue_pi is true,
2015 * consequently the compiler knows requeue_pi is always false past
2016 * this point which will optimize away all the conditional code
2019 if (!IS_ENABLED(CONFIG_FUTEX_PI
) && requeue_pi
)
2024 * Requeue PI only works on two distinct uaddrs. This
2025 * check is only valid for private futexes. See below.
2027 if (uaddr1
== uaddr2
)
2031 * requeue_pi requires a pi_state, try to allocate it now
2032 * without any locks in case it fails.
2034 if (refill_pi_state_cache())
2037 * requeue_pi must wake as many tasks as it can, up to nr_wake
2038 * + nr_requeue, since it acquires the rt_mutex prior to
2039 * returning to userspace, so as to not leave the rt_mutex with
2040 * waiters and no owner. However, second and third wake-ups
2041 * cannot be predicted as they involve race conditions with the
2042 * first wake and a fault while looking up the pi_state. Both
2043 * pthread_cond_signal() and pthread_cond_broadcast() should
2051 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
2052 if (unlikely(ret
!= 0))
2054 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
,
2055 requeue_pi
? VERIFY_WRITE
: VERIFY_READ
);
2056 if (unlikely(ret
!= 0))
2060 * The check above which compares uaddrs is not sufficient for
2061 * shared futexes. We need to compare the keys:
2063 if (requeue_pi
&& match_futex(&key1
, &key2
)) {
2068 hb1
= hash_futex(&key1
);
2069 hb2
= hash_futex(&key2
);
2072 hb_waiters_inc(hb2
);
2073 double_lock_hb(hb1
, hb2
);
2075 if (likely(cmpval
!= NULL
)) {
2078 ret
= get_futex_value_locked(&curval
, uaddr1
);
2080 if (unlikely(ret
)) {
2081 double_unlock_hb(hb1
, hb2
);
2082 hb_waiters_dec(hb2
);
2084 ret
= get_user(curval
, uaddr1
);
2088 if (!(flags
& FLAGS_SHARED
))
2091 put_futex_key(&key2
);
2092 put_futex_key(&key1
);
2095 if (curval
!= *cmpval
) {
2101 if (requeue_pi
&& (task_count
- nr_wake
< nr_requeue
)) {
2102 struct task_struct
*exiting
= NULL
;
2105 * Attempt to acquire uaddr2 and wake the top waiter. If we
2106 * intend to requeue waiters, force setting the FUTEX_WAITERS
2107 * bit. We force this here where we are able to easily handle
2108 * faults rather in the requeue loop below.
2110 ret
= futex_proxy_trylock_atomic(uaddr2
, hb1
, hb2
, &key1
,
2112 &exiting
, nr_requeue
);
2115 * At this point the top_waiter has either taken uaddr2 or is
2116 * waiting on it. If the former, then the pi_state will not
2117 * exist yet, look it up one more time to ensure we have a
2118 * reference to it. If the lock was taken, ret contains the
2119 * vpid of the top waiter task.
2120 * If the lock was not taken, we have pi_state and an initial
2121 * refcount on it. In case of an error we have nothing.
2128 * If we acquired the lock, then the user space value
2129 * of uaddr2 should be vpid. It cannot be changed by
2130 * the top waiter as it is blocked on hb2 lock if it
2131 * tries to do so. If something fiddled with it behind
2132 * our back the pi state lookup might unearth it. So
2133 * we rather use the known value than rereading and
2134 * handing potential crap to lookup_pi_state.
2136 * If that call succeeds then we have pi_state and an
2137 * initial refcount on it.
2139 ret
= lookup_pi_state(uaddr2
, ret
, hb2
, &key2
,
2140 &pi_state
, &exiting
);
2145 /* We hold a reference on the pi state. */
2148 /* If the above failed, then pi_state is NULL */
2150 double_unlock_hb(hb1
, hb2
);
2151 hb_waiters_dec(hb2
);
2152 put_futex_key(&key2
);
2153 put_futex_key(&key1
);
2154 ret
= fault_in_user_writeable(uaddr2
);
2161 * Two reasons for this:
2162 * - EBUSY: Owner is exiting and we just wait for the
2164 * - EAGAIN: The user space value changed.
2166 double_unlock_hb(hb1
, hb2
);
2167 hb_waiters_dec(hb2
);
2168 put_futex_key(&key2
);
2169 put_futex_key(&key1
);
2171 * Handle the case where the owner is in the middle of
2172 * exiting. Wait for the exit to complete otherwise
2173 * this task might loop forever, aka. live lock.
2175 wait_for_owner_exiting(ret
, exiting
);
2183 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
2184 if (task_count
- nr_wake
>= nr_requeue
)
2187 if (!match_futex(&this->key
, &key1
))
2191 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2192 * be paired with each other and no other futex ops.
2194 * We should never be requeueing a futex_q with a pi_state,
2195 * which is awaiting a futex_unlock_pi().
2197 if ((requeue_pi
&& !this->rt_waiter
) ||
2198 (!requeue_pi
&& this->rt_waiter
) ||
2205 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2206 * lock, we already woke the top_waiter. If not, it will be
2207 * woken by futex_unlock_pi().
2209 if (++task_count
<= nr_wake
&& !requeue_pi
) {
2210 mark_wake_futex(&wake_q
, this);
2214 /* Ensure we requeue to the expected futex for requeue_pi. */
2215 if (requeue_pi
&& !match_futex(this->requeue_pi_key
, &key2
)) {
2221 * Requeue nr_requeue waiters and possibly one more in the case
2222 * of requeue_pi if we couldn't acquire the lock atomically.
2226 * Prepare the waiter to take the rt_mutex. Take a
2227 * refcount on the pi_state and store the pointer in
2228 * the futex_q object of the waiter.
2230 get_pi_state(pi_state
);
2231 this->pi_state
= pi_state
;
2232 ret
= rt_mutex_start_proxy_lock(&pi_state
->pi_mutex
,
2237 * We got the lock. We do neither drop the
2238 * refcount on pi_state nor clear
2239 * this->pi_state because the waiter needs the
2240 * pi_state for cleaning up the user space
2241 * value. It will drop the refcount after
2244 requeue_pi_wake_futex(this, &key2
, hb2
);
2249 * rt_mutex_start_proxy_lock() detected a
2250 * potential deadlock when we tried to queue
2251 * that waiter. Drop the pi_state reference
2252 * which we took above and remove the pointer
2253 * to the state from the waiters futex_q
2256 this->pi_state
= NULL
;
2257 put_pi_state(pi_state
);
2259 * We stop queueing more waiters and let user
2260 * space deal with the mess.
2265 requeue_futex(this, hb1
, hb2
, &key2
);
2270 * We took an extra initial reference to the pi_state either
2271 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2272 * need to drop it here again.
2274 put_pi_state(pi_state
);
2277 double_unlock_hb(hb1
, hb2
);
2279 hb_waiters_dec(hb2
);
2282 * drop_futex_key_refs() must be called outside the spinlocks. During
2283 * the requeue we moved futex_q's from the hash bucket at key1 to the
2284 * one at key2 and updated their key pointer. We no longer need to
2285 * hold the references to key1.
2287 while (--drop_count
>= 0)
2288 drop_futex_key_refs(&key1
);
2291 put_futex_key(&key2
);
2293 put_futex_key(&key1
);
2295 return ret
? ret
: task_count
;
2298 /* The key must be already stored in q->key. */
2299 static inline struct futex_hash_bucket
*queue_lock(struct futex_q
*q
)
2300 __acquires(&hb
->lock
)
2302 struct futex_hash_bucket
*hb
;
2304 hb
= hash_futex(&q
->key
);
2307 * Increment the counter before taking the lock so that
2308 * a potential waker won't miss a to-be-slept task that is
2309 * waiting for the spinlock. This is safe as all queue_lock()
2310 * users end up calling queue_me(). Similarly, for housekeeping,
2311 * decrement the counter at queue_unlock() when some error has
2312 * occurred and we don't end up adding the task to the list.
2316 q
->lock_ptr
= &hb
->lock
;
2318 spin_lock(&hb
->lock
); /* implies smp_mb(); (A) */
2323 queue_unlock(struct futex_hash_bucket
*hb
)
2324 __releases(&hb
->lock
)
2326 spin_unlock(&hb
->lock
);
2330 static inline void __queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2335 * The priority used to register this element is
2336 * - either the real thread-priority for the real-time threads
2337 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2338 * - or MAX_RT_PRIO for non-RT threads.
2339 * Thus, all RT-threads are woken first in priority order, and
2340 * the others are woken last, in FIFO order.
2342 prio
= min(current
->normal_prio
, MAX_RT_PRIO
);
2344 plist_node_init(&q
->list
, prio
);
2345 plist_add(&q
->list
, &hb
->chain
);
2350 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2351 * @q: The futex_q to enqueue
2352 * @hb: The destination hash bucket
2354 * The hb->lock must be held by the caller, and is released here. A call to
2355 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2356 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2357 * or nothing if the unqueue is done as part of the wake process and the unqueue
2358 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2361 static inline void queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2362 __releases(&hb
->lock
)
2365 spin_unlock(&hb
->lock
);
2369 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2370 * @q: The futex_q to unqueue
2372 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2373 * be paired with exactly one earlier call to queue_me().
2376 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2377 * - 0 - if the futex_q was already removed by the waking thread
2379 static int unqueue_me(struct futex_q
*q
)
2381 spinlock_t
*lock_ptr
;
2384 /* In the common case we don't take the spinlock, which is nice. */
2387 * q->lock_ptr can change between this read and the following spin_lock.
2388 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2389 * optimizing lock_ptr out of the logic below.
2391 lock_ptr
= READ_ONCE(q
->lock_ptr
);
2392 if (lock_ptr
!= NULL
) {
2393 spin_lock(lock_ptr
);
2395 * q->lock_ptr can change between reading it and
2396 * spin_lock(), causing us to take the wrong lock. This
2397 * corrects the race condition.
2399 * Reasoning goes like this: if we have the wrong lock,
2400 * q->lock_ptr must have changed (maybe several times)
2401 * between reading it and the spin_lock(). It can
2402 * change again after the spin_lock() but only if it was
2403 * already changed before the spin_lock(). It cannot,
2404 * however, change back to the original value. Therefore
2405 * we can detect whether we acquired the correct lock.
2407 if (unlikely(lock_ptr
!= q
->lock_ptr
)) {
2408 spin_unlock(lock_ptr
);
2413 BUG_ON(q
->pi_state
);
2415 spin_unlock(lock_ptr
);
2419 drop_futex_key_refs(&q
->key
);
2424 * PI futexes can not be requeued and must remove themself from the
2425 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2428 static void unqueue_me_pi(struct futex_q
*q
)
2429 __releases(q
->lock_ptr
)
2433 BUG_ON(!q
->pi_state
);
2434 put_pi_state(q
->pi_state
);
2437 spin_unlock(q
->lock_ptr
);
2440 static int fixup_pi_state_owner(u32 __user
*uaddr
, struct futex_q
*q
,
2441 struct task_struct
*argowner
)
2443 struct futex_pi_state
*pi_state
= q
->pi_state
;
2444 u32 uval
, uninitialized_var(curval
), newval
;
2445 struct task_struct
*oldowner
, *newowner
;
2449 lockdep_assert_held(q
->lock_ptr
);
2451 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2453 oldowner
= pi_state
->owner
;
2456 * We are here because either:
2458 * - we stole the lock and pi_state->owner needs updating to reflect
2459 * that (@argowner == current),
2463 * - someone stole our lock and we need to fix things to point to the
2464 * new owner (@argowner == NULL).
2466 * Either way, we have to replace the TID in the user space variable.
2467 * This must be atomic as we have to preserve the owner died bit here.
2469 * Note: We write the user space value _before_ changing the pi_state
2470 * because we can fault here. Imagine swapped out pages or a fork
2471 * that marked all the anonymous memory readonly for cow.
2473 * Modifying pi_state _before_ the user space value would leave the
2474 * pi_state in an inconsistent state when we fault here, because we
2475 * need to drop the locks to handle the fault. This might be observed
2476 * in the PID check in lookup_pi_state.
2480 if (oldowner
!= current
) {
2482 * We raced against a concurrent self; things are
2483 * already fixed up. Nothing to do.
2489 if (__rt_mutex_futex_trylock(&pi_state
->pi_mutex
)) {
2490 /* We got the lock after all, nothing to fix. */
2496 * Since we just failed the trylock; there must be an owner.
2498 newowner
= rt_mutex_owner(&pi_state
->pi_mutex
);
2501 WARN_ON_ONCE(argowner
!= current
);
2502 if (oldowner
== current
) {
2504 * We raced against a concurrent self; things are
2505 * already fixed up. Nothing to do.
2510 newowner
= argowner
;
2513 newtid
= task_pid_vnr(newowner
) | FUTEX_WAITERS
;
2515 if (!pi_state
->owner
)
2516 newtid
|= FUTEX_OWNER_DIED
;
2518 err
= get_futex_value_locked(&uval
, uaddr
);
2523 newval
= (uval
& FUTEX_OWNER_DIED
) | newtid
;
2525 err
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
);
2535 * We fixed up user space. Now we need to fix the pi_state
2538 if (pi_state
->owner
!= NULL
) {
2539 raw_spin_lock(&pi_state
->owner
->pi_lock
);
2540 WARN_ON(list_empty(&pi_state
->list
));
2541 list_del_init(&pi_state
->list
);
2542 raw_spin_unlock(&pi_state
->owner
->pi_lock
);
2545 pi_state
->owner
= newowner
;
2547 raw_spin_lock(&newowner
->pi_lock
);
2548 WARN_ON(!list_empty(&pi_state
->list
));
2549 list_add(&pi_state
->list
, &newowner
->pi_state_list
);
2550 raw_spin_unlock(&newowner
->pi_lock
);
2551 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2556 * In order to reschedule or handle a page fault, we need to drop the
2557 * locks here. In the case of a fault, this gives the other task
2558 * (either the highest priority waiter itself or the task which stole
2559 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2560 * are back from handling the fault we need to check the pi_state after
2561 * reacquiring the locks and before trying to do another fixup. When
2562 * the fixup has been done already we simply return.
2564 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2565 * drop hb->lock since the caller owns the hb -> futex_q relation.
2566 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2569 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2570 spin_unlock(q
->lock_ptr
);
2574 ret
= fault_in_user_writeable(uaddr
);
2588 spin_lock(q
->lock_ptr
);
2589 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2592 * Check if someone else fixed it for us:
2594 if (pi_state
->owner
!= oldowner
) {
2605 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2609 static long futex_wait_restart(struct restart_block
*restart
);
2612 * fixup_owner() - Post lock pi_state and corner case management
2613 * @uaddr: user address of the futex
2614 * @q: futex_q (contains pi_state and access to the rt_mutex)
2615 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2617 * After attempting to lock an rt_mutex, this function is called to cleanup
2618 * the pi_state owner as well as handle race conditions that may allow us to
2619 * acquire the lock. Must be called with the hb lock held.
2622 * - 1 - success, lock taken;
2623 * - 0 - success, lock not taken;
2624 * - <0 - on error (-EFAULT)
2626 static int fixup_owner(u32 __user
*uaddr
, struct futex_q
*q
, int locked
)
2632 * Got the lock. We might not be the anticipated owner if we
2633 * did a lock-steal - fix up the PI-state in that case:
2635 * Speculative pi_state->owner read (we don't hold wait_lock);
2636 * since we own the lock pi_state->owner == current is the
2637 * stable state, anything else needs more attention.
2639 if (q
->pi_state
->owner
!= current
)
2640 ret
= fixup_pi_state_owner(uaddr
, q
, current
);
2645 * If we didn't get the lock; check if anybody stole it from us. In
2646 * that case, we need to fix up the uval to point to them instead of
2647 * us, otherwise bad things happen. [10]
2649 * Another speculative read; pi_state->owner == current is unstable
2650 * but needs our attention.
2652 if (q
->pi_state
->owner
== current
) {
2653 ret
= fixup_pi_state_owner(uaddr
, q
, NULL
);
2658 * Paranoia check. If we did not take the lock, then we should not be
2659 * the owner of the rt_mutex.
2661 if (rt_mutex_owner(&q
->pi_state
->pi_mutex
) == current
) {
2662 printk(KERN_ERR
"fixup_owner: ret = %d pi-mutex: %p "
2663 "pi-state %p\n", ret
,
2664 q
->pi_state
->pi_mutex
.owner
,
2665 q
->pi_state
->owner
);
2669 return ret
? ret
: locked
;
2673 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2674 * @hb: the futex hash bucket, must be locked by the caller
2675 * @q: the futex_q to queue up on
2676 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2678 static void futex_wait_queue_me(struct futex_hash_bucket
*hb
, struct futex_q
*q
,
2679 struct hrtimer_sleeper
*timeout
)
2682 * The task state is guaranteed to be set before another task can
2683 * wake it. set_current_state() is implemented using smp_store_mb() and
2684 * queue_me() calls spin_unlock() upon completion, both serializing
2685 * access to the hash list and forcing another memory barrier.
2687 set_current_state(TASK_INTERRUPTIBLE
);
2692 hrtimer_start_expires(&timeout
->timer
, HRTIMER_MODE_ABS
);
2695 * If we have been removed from the hash list, then another task
2696 * has tried to wake us, and we can skip the call to schedule().
2698 if (likely(!plist_node_empty(&q
->list
))) {
2700 * If the timer has already expired, current will already be
2701 * flagged for rescheduling. Only call schedule if there
2702 * is no timeout, or if it has yet to expire.
2704 if (!timeout
|| timeout
->task
)
2705 freezable_schedule();
2707 __set_current_state(TASK_RUNNING
);
2711 * futex_wait_setup() - Prepare to wait on a futex
2712 * @uaddr: the futex userspace address
2713 * @val: the expected value
2714 * @flags: futex flags (FLAGS_SHARED, etc.)
2715 * @q: the associated futex_q
2716 * @hb: storage for hash_bucket pointer to be returned to caller
2718 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2719 * compare it with the expected value. Handle atomic faults internally.
2720 * Return with the hb lock held and a q.key reference on success, and unlocked
2721 * with no q.key reference on failure.
2724 * - 0 - uaddr contains val and hb has been locked;
2725 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2727 static int futex_wait_setup(u32 __user
*uaddr
, u32 val
, unsigned int flags
,
2728 struct futex_q
*q
, struct futex_hash_bucket
**hb
)
2734 * Access the page AFTER the hash-bucket is locked.
2735 * Order is important:
2737 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2738 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2740 * The basic logical guarantee of a futex is that it blocks ONLY
2741 * if cond(var) is known to be true at the time of blocking, for
2742 * any cond. If we locked the hash-bucket after testing *uaddr, that
2743 * would open a race condition where we could block indefinitely with
2744 * cond(var) false, which would violate the guarantee.
2746 * On the other hand, we insert q and release the hash-bucket only
2747 * after testing *uaddr. This guarantees that futex_wait() will NOT
2748 * absorb a wakeup if *uaddr does not match the desired values
2749 * while the syscall executes.
2752 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
->key
, VERIFY_READ
);
2753 if (unlikely(ret
!= 0))
2757 *hb
= queue_lock(q
);
2759 ret
= get_futex_value_locked(&uval
, uaddr
);
2764 ret
= get_user(uval
, uaddr
);
2768 if (!(flags
& FLAGS_SHARED
))
2771 put_futex_key(&q
->key
);
2782 put_futex_key(&q
->key
);
2786 static int futex_wait(u32 __user
*uaddr
, unsigned int flags
, u32 val
,
2787 ktime_t
*abs_time
, u32 bitset
)
2789 struct hrtimer_sleeper timeout
, *to
= NULL
;
2790 struct restart_block
*restart
;
2791 struct futex_hash_bucket
*hb
;
2792 struct futex_q q
= futex_q_init
;
2802 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2803 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2805 hrtimer_init_sleeper(to
, current
);
2806 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2807 current
->timer_slack_ns
);
2812 * Prepare to wait on uaddr. On success, holds hb lock and increments
2815 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2819 /* queue_me and wait for wakeup, timeout, or a signal. */
2820 futex_wait_queue_me(hb
, &q
, to
);
2822 /* If we were woken (and unqueued), we succeeded, whatever. */
2824 /* unqueue_me() drops q.key ref */
2825 if (!unqueue_me(&q
))
2828 if (to
&& !to
->task
)
2832 * We expect signal_pending(current), but we might be the
2833 * victim of a spurious wakeup as well.
2835 if (!signal_pending(current
))
2842 restart
= ¤t
->restart_block
;
2843 restart
->fn
= futex_wait_restart
;
2844 restart
->futex
.uaddr
= uaddr
;
2845 restart
->futex
.val
= val
;
2846 restart
->futex
.time
= *abs_time
;
2847 restart
->futex
.bitset
= bitset
;
2848 restart
->futex
.flags
= flags
| FLAGS_HAS_TIMEOUT
;
2850 ret
= -ERESTART_RESTARTBLOCK
;
2854 hrtimer_cancel(&to
->timer
);
2855 destroy_hrtimer_on_stack(&to
->timer
);
2861 static long futex_wait_restart(struct restart_block
*restart
)
2863 u32 __user
*uaddr
= restart
->futex
.uaddr
;
2864 ktime_t t
, *tp
= NULL
;
2866 if (restart
->futex
.flags
& FLAGS_HAS_TIMEOUT
) {
2867 t
= restart
->futex
.time
;
2870 restart
->fn
= do_no_restart_syscall
;
2872 return (long)futex_wait(uaddr
, restart
->futex
.flags
,
2873 restart
->futex
.val
, tp
, restart
->futex
.bitset
);
2878 * Userspace tried a 0 -> TID atomic transition of the futex value
2879 * and failed. The kernel side here does the whole locking operation:
2880 * if there are waiters then it will block as a consequence of relying
2881 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2882 * a 0 value of the futex too.).
2884 * Also serves as futex trylock_pi()'ing, and due semantics.
2886 static int futex_lock_pi(u32 __user
*uaddr
, unsigned int flags
,
2887 ktime_t
*time
, int trylock
)
2889 struct hrtimer_sleeper timeout
, *to
= NULL
;
2890 struct futex_pi_state
*pi_state
= NULL
;
2891 struct task_struct
*exiting
= NULL
;
2892 struct rt_mutex_waiter rt_waiter
;
2893 struct futex_hash_bucket
*hb
;
2894 struct futex_q q
= futex_q_init
;
2897 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
2900 if (refill_pi_state_cache())
2905 hrtimer_init_on_stack(&to
->timer
, CLOCK_REALTIME
,
2907 hrtimer_init_sleeper(to
, current
);
2908 hrtimer_set_expires(&to
->timer
, *time
);
2912 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
.key
, VERIFY_WRITE
);
2913 if (unlikely(ret
!= 0))
2917 hb
= queue_lock(&q
);
2919 ret
= futex_lock_pi_atomic(uaddr
, hb
, &q
.key
, &q
.pi_state
, current
,
2921 if (unlikely(ret
)) {
2923 * Atomic work succeeded and we got the lock,
2924 * or failed. Either way, we do _not_ block.
2928 /* We got the lock. */
2930 goto out_unlock_put_key
;
2936 * Two reasons for this:
2937 * - EBUSY: Task is exiting and we just wait for the
2939 * - EAGAIN: The user space value changed.
2942 put_futex_key(&q
.key
);
2944 * Handle the case where the owner is in the middle of
2945 * exiting. Wait for the exit to complete otherwise
2946 * this task might loop forever, aka. live lock.
2948 wait_for_owner_exiting(ret
, exiting
);
2952 goto out_unlock_put_key
;
2956 WARN_ON(!q
.pi_state
);
2959 * Only actually queue now that the atomic ops are done:
2964 ret
= rt_mutex_futex_trylock(&q
.pi_state
->pi_mutex
);
2965 /* Fixup the trylock return value: */
2966 ret
= ret
? 0 : -EWOULDBLOCK
;
2970 rt_mutex_init_waiter(&rt_waiter
);
2973 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2974 * hold it while doing rt_mutex_start_proxy(), because then it will
2975 * include hb->lock in the blocking chain, even through we'll not in
2976 * fact hold it while blocking. This will lead it to report -EDEADLK
2977 * and BUG when futex_unlock_pi() interleaves with this.
2979 * Therefore acquire wait_lock while holding hb->lock, but drop the
2980 * latter before calling __rt_mutex_start_proxy_lock(). This
2981 * interleaves with futex_unlock_pi() -- which does a similar lock
2982 * handoff -- such that the latter can observe the futex_q::pi_state
2983 * before __rt_mutex_start_proxy_lock() is done.
2985 raw_spin_lock_irq(&q
.pi_state
->pi_mutex
.wait_lock
);
2986 spin_unlock(q
.lock_ptr
);
2988 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2989 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2990 * it sees the futex_q::pi_state.
2992 ret
= __rt_mutex_start_proxy_lock(&q
.pi_state
->pi_mutex
, &rt_waiter
, current
);
2993 raw_spin_unlock_irq(&q
.pi_state
->pi_mutex
.wait_lock
);
3002 hrtimer_start_expires(&to
->timer
, HRTIMER_MODE_ABS
);
3004 ret
= rt_mutex_wait_proxy_lock(&q
.pi_state
->pi_mutex
, to
, &rt_waiter
);
3007 spin_lock(q
.lock_ptr
);
3009 * If we failed to acquire the lock (deadlock/signal/timeout), we must
3010 * first acquire the hb->lock before removing the lock from the
3011 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
3014 * In particular; it is important that futex_unlock_pi() can not
3015 * observe this inconsistency.
3017 if (ret
&& !rt_mutex_cleanup_proxy_lock(&q
.pi_state
->pi_mutex
, &rt_waiter
))
3022 * Fixup the pi_state owner and possibly acquire the lock if we
3025 res
= fixup_owner(uaddr
, &q
, !ret
);
3027 * If fixup_owner() returned an error, proprogate that. If it acquired
3028 * the lock, clear our -ETIMEDOUT or -EINTR.
3031 ret
= (res
< 0) ? res
: 0;
3034 * If fixup_owner() faulted and was unable to handle the fault, unlock
3035 * it and return the fault to userspace.
3037 if (ret
&& (rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
)) {
3038 pi_state
= q
.pi_state
;
3039 get_pi_state(pi_state
);
3042 /* Unqueue and drop the lock */
3046 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
3047 put_pi_state(pi_state
);
3056 put_futex_key(&q
.key
);
3059 hrtimer_cancel(&to
->timer
);
3060 destroy_hrtimer_on_stack(&to
->timer
);
3062 return ret
!= -EINTR
? ret
: -ERESTARTNOINTR
;
3067 ret
= fault_in_user_writeable(uaddr
);
3071 if (!(flags
& FLAGS_SHARED
))
3074 put_futex_key(&q
.key
);
3079 * Userspace attempted a TID -> 0 atomic transition, and failed.
3080 * This is the in-kernel slowpath: we look up the PI state (if any),
3081 * and do the rt-mutex unlock.
3083 static int futex_unlock_pi(u32 __user
*uaddr
, unsigned int flags
)
3085 u32
uninitialized_var(curval
), uval
, vpid
= task_pid_vnr(current
);
3086 union futex_key key
= FUTEX_KEY_INIT
;
3087 struct futex_hash_bucket
*hb
;
3088 struct futex_q
*top_waiter
;
3091 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
3095 if (get_user(uval
, uaddr
))
3098 * We release only a lock we actually own:
3100 if ((uval
& FUTEX_TID_MASK
) != vpid
)
3103 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_WRITE
);
3107 hb
= hash_futex(&key
);
3108 spin_lock(&hb
->lock
);
3111 * Check waiters first. We do not trust user space values at
3112 * all and we at least want to know if user space fiddled
3113 * with the futex value instead of blindly unlocking.
3115 top_waiter
= futex_top_waiter(hb
, &key
);
3117 struct futex_pi_state
*pi_state
= top_waiter
->pi_state
;
3124 * If current does not own the pi_state then the futex is
3125 * inconsistent and user space fiddled with the futex value.
3127 if (pi_state
->owner
!= current
)
3130 get_pi_state(pi_state
);
3132 * By taking wait_lock while still holding hb->lock, we ensure
3133 * there is no point where we hold neither; and therefore
3134 * wake_futex_pi() must observe a state consistent with what we
3137 * In particular; this forces __rt_mutex_start_proxy() to
3138 * complete such that we're guaranteed to observe the
3139 * rt_waiter. Also see the WARN in wake_futex_pi().
3141 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
3142 spin_unlock(&hb
->lock
);
3144 /* drops pi_state->pi_mutex.wait_lock */
3145 ret
= wake_futex_pi(uaddr
, uval
, pi_state
);
3147 put_pi_state(pi_state
);
3150 * Success, we're done! No tricky corner cases.
3155 * The atomic access to the futex value generated a
3156 * pagefault, so retry the user-access and the wakeup:
3161 * A unconditional UNLOCK_PI op raced against a waiter
3162 * setting the FUTEX_WAITERS bit. Try again.
3167 * wake_futex_pi has detected invalid state. Tell user
3174 * We have no kernel internal state, i.e. no waiters in the
3175 * kernel. Waiters which are about to queue themselves are stuck
3176 * on hb->lock. So we can safely ignore them. We do neither
3177 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3180 if ((ret
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, 0))) {
3181 spin_unlock(&hb
->lock
);
3196 * If uval has changed, let user space handle it.
3198 ret
= (curval
== uval
) ? 0 : -EAGAIN
;
3201 spin_unlock(&hb
->lock
);
3203 put_futex_key(&key
);
3207 put_futex_key(&key
);
3212 put_futex_key(&key
);
3214 ret
= fault_in_user_writeable(uaddr
);
3222 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3223 * @hb: the hash_bucket futex_q was original enqueued on
3224 * @q: the futex_q woken while waiting to be requeued
3225 * @key2: the futex_key of the requeue target futex
3226 * @timeout: the timeout associated with the wait (NULL if none)
3228 * Detect if the task was woken on the initial futex as opposed to the requeue
3229 * target futex. If so, determine if it was a timeout or a signal that caused
3230 * the wakeup and return the appropriate error code to the caller. Must be
3231 * called with the hb lock held.
3234 * - 0 = no early wakeup detected;
3235 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3238 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket
*hb
,
3239 struct futex_q
*q
, union futex_key
*key2
,
3240 struct hrtimer_sleeper
*timeout
)
3245 * With the hb lock held, we avoid races while we process the wakeup.
3246 * We only need to hold hb (and not hb2) to ensure atomicity as the
3247 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3248 * It can't be requeued from uaddr2 to something else since we don't
3249 * support a PI aware source futex for requeue.
3251 if (!match_futex(&q
->key
, key2
)) {
3252 WARN_ON(q
->lock_ptr
&& (&hb
->lock
!= q
->lock_ptr
));
3254 * We were woken prior to requeue by a timeout or a signal.
3255 * Unqueue the futex_q and determine which it was.
3257 plist_del(&q
->list
, &hb
->chain
);
3260 /* Handle spurious wakeups gracefully */
3262 if (timeout
&& !timeout
->task
)
3264 else if (signal_pending(current
))
3265 ret
= -ERESTARTNOINTR
;
3271 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3272 * @uaddr: the futex we initially wait on (non-pi)
3273 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3274 * the same type, no requeueing from private to shared, etc.
3275 * @val: the expected value of uaddr
3276 * @abs_time: absolute timeout
3277 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3278 * @uaddr2: the pi futex we will take prior to returning to user-space
3280 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3281 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3282 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3283 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3284 * without one, the pi logic would not know which task to boost/deboost, if
3285 * there was a need to.
3287 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3288 * via the following--
3289 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3290 * 2) wakeup on uaddr2 after a requeue
3294 * If 3, cleanup and return -ERESTARTNOINTR.
3296 * If 2, we may then block on trying to take the rt_mutex and return via:
3297 * 5) successful lock
3300 * 8) other lock acquisition failure
3302 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3304 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3310 static int futex_wait_requeue_pi(u32 __user
*uaddr
, unsigned int flags
,
3311 u32 val
, ktime_t
*abs_time
, u32 bitset
,
3314 struct hrtimer_sleeper timeout
, *to
= NULL
;
3315 struct futex_pi_state
*pi_state
= NULL
;
3316 struct rt_mutex_waiter rt_waiter
;
3317 struct futex_hash_bucket
*hb
;
3318 union futex_key key2
= FUTEX_KEY_INIT
;
3319 struct futex_q q
= futex_q_init
;
3322 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
3325 if (uaddr
== uaddr2
)
3333 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
3334 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
3336 hrtimer_init_sleeper(to
, current
);
3337 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
3338 current
->timer_slack_ns
);
3342 * The waiter is allocated on our stack, manipulated by the requeue
3343 * code while we sleep on uaddr.
3345 rt_mutex_init_waiter(&rt_waiter
);
3347 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
3348 if (unlikely(ret
!= 0))
3352 q
.rt_waiter
= &rt_waiter
;
3353 q
.requeue_pi_key
= &key2
;
3356 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3359 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
3364 * The check above which compares uaddrs is not sufficient for
3365 * shared futexes. We need to compare the keys:
3367 if (match_futex(&q
.key
, &key2
)) {
3373 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3374 futex_wait_queue_me(hb
, &q
, to
);
3376 spin_lock(&hb
->lock
);
3377 ret
= handle_early_requeue_pi_wakeup(hb
, &q
, &key2
, to
);
3378 spin_unlock(&hb
->lock
);
3383 * In order for us to be here, we know our q.key == key2, and since
3384 * we took the hb->lock above, we also know that futex_requeue() has
3385 * completed and we no longer have to concern ourselves with a wakeup
3386 * race with the atomic proxy lock acquisition by the requeue code. The
3387 * futex_requeue dropped our key1 reference and incremented our key2
3391 /* Check if the requeue code acquired the second futex for us. */
3394 * Got the lock. We might not be the anticipated owner if we
3395 * did a lock-steal - fix up the PI-state in that case.
3397 if (q
.pi_state
&& (q
.pi_state
->owner
!= current
)) {
3398 spin_lock(q
.lock_ptr
);
3399 ret
= fixup_pi_state_owner(uaddr2
, &q
, current
);
3400 if (ret
&& rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
) {
3401 pi_state
= q
.pi_state
;
3402 get_pi_state(pi_state
);
3405 * Drop the reference to the pi state which
3406 * the requeue_pi() code acquired for us.
3408 put_pi_state(q
.pi_state
);
3409 spin_unlock(q
.lock_ptr
);
3412 struct rt_mutex
*pi_mutex
;
3415 * We have been woken up by futex_unlock_pi(), a timeout, or a
3416 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3419 WARN_ON(!q
.pi_state
);
3420 pi_mutex
= &q
.pi_state
->pi_mutex
;
3421 ret
= rt_mutex_wait_proxy_lock(pi_mutex
, to
, &rt_waiter
);
3423 spin_lock(q
.lock_ptr
);
3424 if (ret
&& !rt_mutex_cleanup_proxy_lock(pi_mutex
, &rt_waiter
))
3427 debug_rt_mutex_free_waiter(&rt_waiter
);
3429 * Fixup the pi_state owner and possibly acquire the lock if we
3432 res
= fixup_owner(uaddr2
, &q
, !ret
);
3434 * If fixup_owner() returned an error, proprogate that. If it
3435 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3438 ret
= (res
< 0) ? res
: 0;
3441 * If fixup_pi_state_owner() faulted and was unable to handle
3442 * the fault, unlock the rt_mutex and return the fault to
3445 if (ret
&& rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
) {
3446 pi_state
= q
.pi_state
;
3447 get_pi_state(pi_state
);
3450 /* Unqueue and drop the lock. */
3455 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
3456 put_pi_state(pi_state
);
3459 if (ret
== -EINTR
) {
3461 * We've already been requeued, but cannot restart by calling
3462 * futex_lock_pi() directly. We could restart this syscall, but
3463 * it would detect that the user space "val" changed and return
3464 * -EWOULDBLOCK. Save the overhead of the restart and return
3465 * -EWOULDBLOCK directly.
3471 put_futex_key(&q
.key
);
3473 put_futex_key(&key2
);
3477 hrtimer_cancel(&to
->timer
);
3478 destroy_hrtimer_on_stack(&to
->timer
);
3484 * Support for robust futexes: the kernel cleans up held futexes at
3487 * Implementation: user-space maintains a per-thread list of locks it
3488 * is holding. Upon do_exit(), the kernel carefully walks this list,
3489 * and marks all locks that are owned by this thread with the
3490 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3491 * always manipulated with the lock held, so the list is private and
3492 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3493 * field, to allow the kernel to clean up if the thread dies after
3494 * acquiring the lock, but just before it could have added itself to
3495 * the list. There can only be one such pending lock.
3499 * sys_set_robust_list() - Set the robust-futex list head of a task
3500 * @head: pointer to the list-head
3501 * @len: length of the list-head, as userspace expects
3503 SYSCALL_DEFINE2(set_robust_list
, struct robust_list_head __user
*, head
,
3506 if (!futex_cmpxchg_enabled
)
3509 * The kernel knows only one size for now:
3511 if (unlikely(len
!= sizeof(*head
)))
3514 current
->robust_list
= head
;
3520 * sys_get_robust_list() - Get the robust-futex list head of a task
3521 * @pid: pid of the process [zero for current task]
3522 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3523 * @len_ptr: pointer to a length field, the kernel fills in the header size
3525 SYSCALL_DEFINE3(get_robust_list
, int, pid
,
3526 struct robust_list_head __user
* __user
*, head_ptr
,
3527 size_t __user
*, len_ptr
)
3529 struct robust_list_head __user
*head
;
3531 struct task_struct
*p
;
3533 if (!futex_cmpxchg_enabled
)
3542 p
= find_task_by_vpid(pid
);
3548 if (!ptrace_may_access(p
, PTRACE_MODE_READ_REALCREDS
))
3551 head
= p
->robust_list
;
3554 if (put_user(sizeof(*head
), len_ptr
))
3556 return put_user(head
, head_ptr
);
3564 /* Constants for the pending_op argument of handle_futex_death */
3565 #define HANDLE_DEATH_PENDING true
3566 #define HANDLE_DEATH_LIST false
3569 * Process a futex-list entry, check whether it's owned by the
3570 * dying task, and do notification if so:
3572 static int handle_futex_death(u32 __user
*uaddr
, struct task_struct
*curr
,
3573 bool pi
, bool pending_op
)
3575 u32 uval
, uninitialized_var(nval
), mval
;
3578 /* Futex address must be 32bit aligned */
3579 if ((((unsigned long)uaddr
) % sizeof(*uaddr
)) != 0)
3583 if (get_user(uval
, uaddr
))
3587 * Special case for regular (non PI) futexes. The unlock path in
3588 * user space has two race scenarios:
3590 * 1. The unlock path releases the user space futex value and
3591 * before it can execute the futex() syscall to wake up
3592 * waiters it is killed.
3594 * 2. A woken up waiter is killed before it can acquire the
3595 * futex in user space.
3597 * In both cases the TID validation below prevents a wakeup of
3598 * potential waiters which can cause these waiters to block
3601 * In both cases the following conditions are met:
3603 * 1) task->robust_list->list_op_pending != NULL
3604 * @pending_op == true
3605 * 2) User space futex value == 0
3606 * 3) Regular futex: @pi == false
3608 * If these conditions are met, it is safe to attempt waking up a
3609 * potential waiter without touching the user space futex value and
3610 * trying to set the OWNER_DIED bit. The user space futex value is
3611 * uncontended and the rest of the user space mutex state is
3612 * consistent, so a woken waiter will just take over the
3613 * uncontended futex. Setting the OWNER_DIED bit would create
3614 * inconsistent state and malfunction of the user space owner died
3617 if (pending_op
&& !pi
&& !uval
) {
3618 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
3622 if ((uval
& FUTEX_TID_MASK
) != task_pid_vnr(curr
))
3626 * Ok, this dying thread is truly holding a futex
3627 * of interest. Set the OWNER_DIED bit atomically
3628 * via cmpxchg, and if the value had FUTEX_WAITERS
3629 * set, wake up a waiter (if any). (We have to do a
3630 * futex_wake() even if OWNER_DIED is already set -
3631 * to handle the rare but possible case of recursive
3632 * thread-death.) The rest of the cleanup is done in
3635 mval
= (uval
& FUTEX_WAITERS
) | FUTEX_OWNER_DIED
;
3638 * We are not holding a lock here, but we want to have
3639 * the pagefault_disable/enable() protection because
3640 * we want to handle the fault gracefully. If the
3641 * access fails we try to fault in the futex with R/W
3642 * verification via get_user_pages. get_user() above
3643 * does not guarantee R/W access. If that fails we
3644 * give up and leave the futex locked.
3646 if ((err
= cmpxchg_futex_value_locked(&nval
, uaddr
, uval
, mval
))) {
3649 if (fault_in_user_writeable(uaddr
))
3667 * Wake robust non-PI futexes here. The wakeup of
3668 * PI futexes happens in exit_pi_state():
3670 if (!pi
&& (uval
& FUTEX_WAITERS
))
3671 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
3677 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3679 static inline int fetch_robust_entry(struct robust_list __user
**entry
,
3680 struct robust_list __user
* __user
*head
,
3683 unsigned long uentry
;
3685 if (get_user(uentry
, (unsigned long __user
*)head
))
3688 *entry
= (void __user
*)(uentry
& ~1UL);
3695 * Walk curr->robust_list (very carefully, it's a userspace list!)
3696 * and mark any locks found there dead, and notify any waiters.
3698 * We silently return on any sign of list-walking problem.
3700 static void exit_robust_list(struct task_struct
*curr
)
3702 struct robust_list_head __user
*head
= curr
->robust_list
;
3703 struct robust_list __user
*entry
, *next_entry
, *pending
;
3704 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
3705 unsigned int uninitialized_var(next_pi
);
3706 unsigned long futex_offset
;
3709 if (!futex_cmpxchg_enabled
)
3713 * Fetch the list head (which was registered earlier, via
3714 * sys_set_robust_list()):
3716 if (fetch_robust_entry(&entry
, &head
->list
.next
, &pi
))
3719 * Fetch the relative futex offset:
3721 if (get_user(futex_offset
, &head
->futex_offset
))
3724 * Fetch any possibly pending lock-add first, and handle it
3727 if (fetch_robust_entry(&pending
, &head
->list_op_pending
, &pip
))
3730 next_entry
= NULL
; /* avoid warning with gcc */
3731 while (entry
!= &head
->list
) {
3733 * Fetch the next entry in the list before calling
3734 * handle_futex_death:
3736 rc
= fetch_robust_entry(&next_entry
, &entry
->next
, &next_pi
);
3738 * A pending lock might already be on the list, so
3739 * don't process it twice:
3741 if (entry
!= pending
) {
3742 if (handle_futex_death((void __user
*)entry
+ futex_offset
,
3743 curr
, pi
, HANDLE_DEATH_LIST
))
3751 * Avoid excessively long or circular lists:
3760 handle_futex_death((void __user
*)pending
+ futex_offset
,
3761 curr
, pip
, HANDLE_DEATH_PENDING
);
3765 static void futex_cleanup(struct task_struct
*tsk
)
3767 if (unlikely(tsk
->robust_list
)) {
3768 exit_robust_list(tsk
);
3769 tsk
->robust_list
= NULL
;
3772 #ifdef CONFIG_COMPAT
3773 if (unlikely(tsk
->compat_robust_list
)) {
3774 compat_exit_robust_list(tsk
);
3775 tsk
->compat_robust_list
= NULL
;
3779 if (unlikely(!list_empty(&tsk
->pi_state_list
)))
3780 exit_pi_state_list(tsk
);
3784 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3785 * @tsk: task to set the state on
3787 * Set the futex exit state of the task lockless. The futex waiter code
3788 * observes that state when a task is exiting and loops until the task has
3789 * actually finished the futex cleanup. The worst case for this is that the
3790 * waiter runs through the wait loop until the state becomes visible.
3792 * This is called from the recursive fault handling path in do_exit().
3794 * This is best effort. Either the futex exit code has run already or
3795 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3796 * take it over. If not, the problem is pushed back to user space. If the
3797 * futex exit code did not run yet, then an already queued waiter might
3798 * block forever, but there is nothing which can be done about that.
3800 void futex_exit_recursive(struct task_struct
*tsk
)
3802 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3803 if (tsk
->futex_state
== FUTEX_STATE_EXITING
)
3804 mutex_unlock(&tsk
->futex_exit_mutex
);
3805 tsk
->futex_state
= FUTEX_STATE_DEAD
;
3808 static void futex_cleanup_begin(struct task_struct
*tsk
)
3811 * Prevent various race issues against a concurrent incoming waiter
3812 * including live locks by forcing the waiter to block on
3813 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3814 * attach_to_pi_owner().
3816 mutex_lock(&tsk
->futex_exit_mutex
);
3819 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3821 * This ensures that all subsequent checks of tsk->futex_state in
3822 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3823 * tsk->pi_lock held.
3825 * It guarantees also that a pi_state which was queued right before
3826 * the state change under tsk->pi_lock by a concurrent waiter must
3827 * be observed in exit_pi_state_list().
3829 raw_spin_lock_irq(&tsk
->pi_lock
);
3830 tsk
->futex_state
= FUTEX_STATE_EXITING
;
3831 raw_spin_unlock_irq(&tsk
->pi_lock
);
3834 static void futex_cleanup_end(struct task_struct
*tsk
, int state
)
3837 * Lockless store. The only side effect is that an observer might
3838 * take another loop until it becomes visible.
3840 tsk
->futex_state
= state
;
3842 * Drop the exit protection. This unblocks waiters which observed
3843 * FUTEX_STATE_EXITING to reevaluate the state.
3845 mutex_unlock(&tsk
->futex_exit_mutex
);
3848 void futex_exec_release(struct task_struct
*tsk
)
3851 * The state handling is done for consistency, but in the case of
3852 * exec() there is no way to prevent futher damage as the PID stays
3853 * the same. But for the unlikely and arguably buggy case that a
3854 * futex is held on exec(), this provides at least as much state
3855 * consistency protection which is possible.
3857 futex_cleanup_begin(tsk
);
3860 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3861 * exec a new binary.
3863 futex_cleanup_end(tsk
, FUTEX_STATE_OK
);
3866 void futex_exit_release(struct task_struct
*tsk
)
3868 futex_cleanup_begin(tsk
);
3870 futex_cleanup_end(tsk
, FUTEX_STATE_DEAD
);
3873 long do_futex(u32 __user
*uaddr
, int op
, u32 val
, ktime_t
*timeout
,
3874 u32 __user
*uaddr2
, u32 val2
, u32 val3
)
3876 int cmd
= op
& FUTEX_CMD_MASK
;
3877 unsigned int flags
= 0;
3879 if (!(op
& FUTEX_PRIVATE_FLAG
))
3880 flags
|= FLAGS_SHARED
;
3882 if (op
& FUTEX_CLOCK_REALTIME
) {
3883 flags
|= FLAGS_CLOCKRT
;
3884 if (cmd
!= FUTEX_WAIT
&& cmd
!= FUTEX_WAIT_BITSET
&& \
3885 cmd
!= FUTEX_WAIT_REQUEUE_PI
)
3891 case FUTEX_UNLOCK_PI
:
3892 case FUTEX_TRYLOCK_PI
:
3893 case FUTEX_WAIT_REQUEUE_PI
:
3894 case FUTEX_CMP_REQUEUE_PI
:
3895 if (!futex_cmpxchg_enabled
)
3901 val3
= FUTEX_BITSET_MATCH_ANY
;
3902 case FUTEX_WAIT_BITSET
:
3903 return futex_wait(uaddr
, flags
, val
, timeout
, val3
);
3905 val3
= FUTEX_BITSET_MATCH_ANY
;
3906 case FUTEX_WAKE_BITSET
:
3907 return futex_wake(uaddr
, flags
, val
, val3
);
3909 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, NULL
, 0);
3910 case FUTEX_CMP_REQUEUE
:
3911 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 0);
3913 return futex_wake_op(uaddr
, flags
, uaddr2
, val
, val2
, val3
);
3915 return futex_lock_pi(uaddr
, flags
, timeout
, 0);
3916 case FUTEX_UNLOCK_PI
:
3917 return futex_unlock_pi(uaddr
, flags
);
3918 case FUTEX_TRYLOCK_PI
:
3919 return futex_lock_pi(uaddr
, flags
, NULL
, 1);
3920 case FUTEX_WAIT_REQUEUE_PI
:
3921 val3
= FUTEX_BITSET_MATCH_ANY
;
3922 return futex_wait_requeue_pi(uaddr
, flags
, val
, timeout
, val3
,
3924 case FUTEX_CMP_REQUEUE_PI
:
3925 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 1);
3931 SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
3932 struct timespec __user
*, utime
, u32 __user
*, uaddr2
,
3936 ktime_t t
, *tp
= NULL
;
3938 int cmd
= op
& FUTEX_CMD_MASK
;
3940 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
3941 cmd
== FUTEX_WAIT_BITSET
||
3942 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
3943 if (unlikely(should_fail_futex(!(op
& FUTEX_PRIVATE_FLAG
))))
3945 if (copy_from_user(&ts
, utime
, sizeof(ts
)) != 0)
3947 if (!timespec_valid(&ts
))
3950 t
= timespec_to_ktime(ts
);
3951 if (cmd
== FUTEX_WAIT
)
3952 t
= ktime_add_safe(ktime_get(), t
);
3956 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3957 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3959 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
3960 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
3961 val2
= (u32
) (unsigned long) utime
;
3963 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
3966 #ifdef CONFIG_COMPAT
3968 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3971 compat_fetch_robust_entry(compat_uptr_t
*uentry
, struct robust_list __user
**entry
,
3972 compat_uptr_t __user
*head
, unsigned int *pi
)
3974 if (get_user(*uentry
, head
))
3977 *entry
= compat_ptr((*uentry
) & ~1);
3978 *pi
= (unsigned int)(*uentry
) & 1;
3983 static void __user
*futex_uaddr(struct robust_list __user
*entry
,
3984 compat_long_t futex_offset
)
3986 compat_uptr_t base
= ptr_to_compat(entry
);
3987 void __user
*uaddr
= compat_ptr(base
+ futex_offset
);
3993 * Walk curr->robust_list (very carefully, it's a userspace list!)
3994 * and mark any locks found there dead, and notify any waiters.
3996 * We silently return on any sign of list-walking problem.
3998 static void compat_exit_robust_list(struct task_struct
*curr
)
4000 struct compat_robust_list_head __user
*head
= curr
->compat_robust_list
;
4001 struct robust_list __user
*entry
, *next_entry
, *pending
;
4002 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
4003 unsigned int uninitialized_var(next_pi
);
4004 compat_uptr_t uentry
, next_uentry
, upending
;
4005 compat_long_t futex_offset
;
4008 if (!futex_cmpxchg_enabled
)
4012 * Fetch the list head (which was registered earlier, via
4013 * sys_set_robust_list()):
4015 if (compat_fetch_robust_entry(&uentry
, &entry
, &head
->list
.next
, &pi
))
4018 * Fetch the relative futex offset:
4020 if (get_user(futex_offset
, &head
->futex_offset
))
4023 * Fetch any possibly pending lock-add first, and handle it
4026 if (compat_fetch_robust_entry(&upending
, &pending
,
4027 &head
->list_op_pending
, &pip
))
4030 next_entry
= NULL
; /* avoid warning with gcc */
4031 while (entry
!= (struct robust_list __user
*) &head
->list
) {
4033 * Fetch the next entry in the list before calling
4034 * handle_futex_death:
4036 rc
= compat_fetch_robust_entry(&next_uentry
, &next_entry
,
4037 (compat_uptr_t __user
*)&entry
->next
, &next_pi
);
4039 * A pending lock might already be on the list, so
4040 * dont process it twice:
4042 if (entry
!= pending
) {
4043 void __user
*uaddr
= futex_uaddr(entry
, futex_offset
);
4045 if (handle_futex_death(uaddr
, curr
, pi
,
4051 uentry
= next_uentry
;
4055 * Avoid excessively long or circular lists:
4063 void __user
*uaddr
= futex_uaddr(pending
, futex_offset
);
4065 handle_futex_death(uaddr
, curr
, pip
, HANDLE_DEATH_PENDING
);
4069 COMPAT_SYSCALL_DEFINE2(set_robust_list
,
4070 struct compat_robust_list_head __user
*, head
,
4073 if (!futex_cmpxchg_enabled
)
4076 if (unlikely(len
!= sizeof(*head
)))
4079 current
->compat_robust_list
= head
;
4084 COMPAT_SYSCALL_DEFINE3(get_robust_list
, int, pid
,
4085 compat_uptr_t __user
*, head_ptr
,
4086 compat_size_t __user
*, len_ptr
)
4088 struct compat_robust_list_head __user
*head
;
4090 struct task_struct
*p
;
4092 if (!futex_cmpxchg_enabled
)
4101 p
= find_task_by_vpid(pid
);
4107 if (!ptrace_may_access(p
, PTRACE_MODE_READ_REALCREDS
))
4110 head
= p
->compat_robust_list
;
4113 if (put_user(sizeof(*head
), len_ptr
))
4115 return put_user(ptr_to_compat(head
), head_ptr
);
4123 COMPAT_SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
4124 struct compat_timespec __user
*, utime
, u32 __user
*, uaddr2
,
4128 ktime_t t
, *tp
= NULL
;
4130 int cmd
= op
& FUTEX_CMD_MASK
;
4132 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
4133 cmd
== FUTEX_WAIT_BITSET
||
4134 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
4135 if (compat_get_timespec(&ts
, utime
))
4137 if (!timespec_valid(&ts
))
4140 t
= timespec_to_ktime(ts
);
4141 if (cmd
== FUTEX_WAIT
)
4142 t
= ktime_add_safe(ktime_get(), t
);
4145 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
4146 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
4147 val2
= (int) (unsigned long) utime
;
4149 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
4151 #endif /* CONFIG_COMPAT */
4153 static void __init
futex_detect_cmpxchg(void)
4155 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
4159 * This will fail and we want it. Some arch implementations do
4160 * runtime detection of the futex_atomic_cmpxchg_inatomic()
4161 * functionality. We want to know that before we call in any
4162 * of the complex code paths. Also we want to prevent
4163 * registration of robust lists in that case. NULL is
4164 * guaranteed to fault and we get -EFAULT on functional
4165 * implementation, the non-functional ones will return
4168 if (cmpxchg_futex_value_locked(&curval
, NULL
, 0, 0) == -EFAULT
)
4169 futex_cmpxchg_enabled
= 1;
4173 static int __init
futex_init(void)
4175 unsigned int futex_shift
;
4178 #if CONFIG_BASE_SMALL
4179 futex_hashsize
= 16;
4181 futex_hashsize
= roundup_pow_of_two(256 * num_possible_cpus());
4184 futex_queues
= alloc_large_system_hash("futex", sizeof(*futex_queues
),
4186 futex_hashsize
< 256 ? HASH_SMALL
: 0,
4188 futex_hashsize
, futex_hashsize
);
4189 futex_hashsize
= 1UL << futex_shift
;
4191 futex_detect_cmpxchg();
4193 for (i
= 0; i
< futex_hashsize
; i
++) {
4194 atomic_set(&futex_queues
[i
].waiters
, 0);
4195 plist_head_init(&futex_queues
[i
].chain
);
4196 spin_lock_init(&futex_queues
[i
].lock
);
4201 core_initcall(futex_init
);