1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Fast Userspace Mutexes (which I call "Futexes!").
4 * (C) Rusty Russell, IBM 2002
6 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
7 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
9 * Removed page pinning, fix privately mapped COW pages and other cleanups
10 * (C) Copyright 2003, 2004 Jamie Lokier
12 * Robust futex support started by Ingo Molnar
13 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
14 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
16 * PI-futex support started by Ingo Molnar and Thomas Gleixner
17 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
18 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
20 * PRIVATE futexes by Eric Dumazet
21 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
23 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
24 * Copyright (C) IBM Corporation, 2009
25 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
27 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
28 * enough at me, Linus for the original (flawed) idea, Matthew
29 * Kirkwood for proof-of-concept implementation.
31 * "The futexes are also cursed."
32 * "But they come in a choice of three flavours!"
34 #include <linux/compat.h>
35 #include <linux/slab.h>
36 #include <linux/poll.h>
38 #include <linux/file.h>
39 #include <linux/jhash.h>
40 #include <linux/init.h>
41 #include <linux/futex.h>
42 #include <linux/mount.h>
43 #include <linux/pagemap.h>
44 #include <linux/syscalls.h>
45 #include <linux/signal.h>
46 #include <linux/export.h>
47 #include <linux/magic.h>
48 #include <linux/pid.h>
49 #include <linux/nsproxy.h>
50 #include <linux/ptrace.h>
51 #include <linux/sched/rt.h>
52 #include <linux/sched/wake_q.h>
53 #include <linux/sched/mm.h>
54 #include <linux/hugetlb.h>
55 #include <linux/freezer.h>
56 #include <linux/memblock.h>
57 #include <linux/fault-inject.h>
58 #include <linux/refcount.h>
60 #include <asm/futex.h>
62 #include "locking/rtmutex_common.h"
65 * READ this before attempting to hack on futexes!
67 * Basic futex operation and ordering guarantees
68 * =============================================
70 * The waiter reads the futex value in user space and calls
71 * futex_wait(). This function computes the hash bucket and acquires
72 * the hash bucket lock. After that it reads the futex user space value
73 * again and verifies that the data has not changed. If it has not changed
74 * it enqueues itself into the hash bucket, releases the hash bucket lock
77 * The waker side modifies the user space value of the futex and calls
78 * futex_wake(). This function computes the hash bucket and acquires the
79 * hash bucket lock. Then it looks for waiters on that futex in the hash
80 * bucket and wakes them.
82 * In futex wake up scenarios where no tasks are blocked on a futex, taking
83 * the hb spinlock can be avoided and simply return. In order for this
84 * optimization to work, ordering guarantees must exist so that the waiter
85 * being added to the list is acknowledged when the list is concurrently being
86 * checked by the waker, avoiding scenarios like the following:
90 * sys_futex(WAIT, futex, val);
91 * futex_wait(futex, val);
94 * sys_futex(WAKE, futex);
99 * lock(hash_bucket(futex));
101 * unlock(hash_bucket(futex));
104 * This would cause the waiter on CPU 0 to wait forever because it
105 * missed the transition of the user space value from val to newval
106 * and the waker did not find the waiter in the hash bucket queue.
108 * The correct serialization ensures that a waiter either observes
109 * the changed user space value before blocking or is woken by a
114 * sys_futex(WAIT, futex, val);
115 * futex_wait(futex, val);
118 * smp_mb(); (A) <-- paired with -.
120 * lock(hash_bucket(futex)); |
124 * | sys_futex(WAKE, futex);
125 * | futex_wake(futex);
127 * `--------> smp_mb(); (B)
130 * unlock(hash_bucket(futex));
131 * schedule(); if (waiters)
132 * lock(hash_bucket(futex));
133 * else wake_waiters(futex);
134 * waiters--; (b) unlock(hash_bucket(futex));
136 * Where (A) orders the waiters increment and the futex value read through
137 * atomic operations (see hb_waiters_inc) and where (B) orders the write
138 * to futex and the waiters read -- this is done by the barriers for both
139 * shared and private futexes in get_futex_key_refs().
141 * This yields the following case (where X:=waiters, Y:=futex):
149 * Which guarantees that x==0 && y==0 is impossible; which translates back into
150 * the guarantee that we cannot both miss the futex variable change and the
153 * Note that a new waiter is accounted for in (a) even when it is possible that
154 * the wait call can return error, in which case we backtrack from it in (b).
155 * Refer to the comment in queue_lock().
157 * Similarly, in order to account for waiters being requeued on another
158 * address we always increment the waiters for the destination bucket before
159 * acquiring the lock. It then decrements them again after releasing it -
160 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
161 * will do the additional required waiter count housekeeping. This is done for
162 * double_lock_hb() and double_unlock_hb(), respectively.
165 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
166 #define futex_cmpxchg_enabled 1
168 static int __read_mostly futex_cmpxchg_enabled
;
172 * Futex flags used to encode options to functions and preserve them across
176 # define FLAGS_SHARED 0x01
179 * NOMMU does not have per process address space. Let the compiler optimize
182 # define FLAGS_SHARED 0x00
184 #define FLAGS_CLOCKRT 0x02
185 #define FLAGS_HAS_TIMEOUT 0x04
188 * Priority Inheritance state:
190 struct futex_pi_state
{
192 * list of 'owned' pi_state instances - these have to be
193 * cleaned up in do_exit() if the task exits prematurely:
195 struct list_head list
;
200 struct rt_mutex pi_mutex
;
202 struct task_struct
*owner
;
206 } __randomize_layout
;
209 * struct futex_q - The hashed futex queue entry, one per waiting task
210 * @list: priority-sorted list of tasks waiting on this futex
211 * @task: the task waiting on the futex
212 * @lock_ptr: the hash bucket lock
213 * @key: the key the futex is hashed on
214 * @pi_state: optional priority inheritance state
215 * @rt_waiter: rt_waiter storage for use with requeue_pi
216 * @requeue_pi_key: the requeue_pi target futex key
217 * @bitset: bitset for the optional bitmasked wakeup
219 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
220 * we can wake only the relevant ones (hashed queues may be shared).
222 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
223 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
224 * The order of wakeup is always to make the first condition true, then
227 * PI futexes are typically woken before they are removed from the hash list via
228 * the rt_mutex code. See unqueue_me_pi().
231 struct plist_node list
;
233 struct task_struct
*task
;
234 spinlock_t
*lock_ptr
;
236 struct futex_pi_state
*pi_state
;
237 struct rt_mutex_waiter
*rt_waiter
;
238 union futex_key
*requeue_pi_key
;
240 } __randomize_layout
;
242 static const struct futex_q futex_q_init
= {
243 /* list gets initialized in queue_me()*/
244 .key
= FUTEX_KEY_INIT
,
245 .bitset
= FUTEX_BITSET_MATCH_ANY
249 * Hash buckets are shared by all the futex_keys that hash to the same
250 * location. Each key may have multiple futex_q structures, one for each task
251 * waiting on a futex.
253 struct futex_hash_bucket
{
256 struct plist_head chain
;
257 } ____cacheline_aligned_in_smp
;
260 * The base of the bucket array and its size are always used together
261 * (after initialization only in hash_futex()), so ensure that they
262 * reside in the same cacheline.
265 struct futex_hash_bucket
*queues
;
266 unsigned long hashsize
;
267 } __futex_data __read_mostly
__aligned(2*sizeof(long));
268 #define futex_queues (__futex_data.queues)
269 #define futex_hashsize (__futex_data.hashsize)
273 * Fault injections for futexes.
275 #ifdef CONFIG_FAIL_FUTEX
278 struct fault_attr attr
;
282 .attr
= FAULT_ATTR_INITIALIZER
,
283 .ignore_private
= false,
286 static int __init
setup_fail_futex(char *str
)
288 return setup_fault_attr(&fail_futex
.attr
, str
);
290 __setup("fail_futex=", setup_fail_futex
);
292 static bool should_fail_futex(bool fshared
)
294 if (fail_futex
.ignore_private
&& !fshared
)
297 return should_fail(&fail_futex
.attr
, 1);
300 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
302 static int __init
fail_futex_debugfs(void)
304 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
307 dir
= fault_create_debugfs_attr("fail_futex", NULL
,
312 debugfs_create_bool("ignore-private", mode
, dir
,
313 &fail_futex
.ignore_private
);
317 late_initcall(fail_futex_debugfs
);
319 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
322 static inline bool should_fail_futex(bool fshared
)
326 #endif /* CONFIG_FAIL_FUTEX */
329 static void compat_exit_robust_list(struct task_struct
*curr
);
331 static inline void compat_exit_robust_list(struct task_struct
*curr
) { }
334 static inline void futex_get_mm(union futex_key
*key
)
336 mmgrab(key
->private.mm
);
338 * Ensure futex_get_mm() implies a full barrier such that
339 * get_futex_key() implies a full barrier. This is relied upon
340 * as smp_mb(); (B), see the ordering comment above.
342 smp_mb__after_atomic();
346 * Reflects a new waiter being added to the waitqueue.
348 static inline void hb_waiters_inc(struct futex_hash_bucket
*hb
)
351 atomic_inc(&hb
->waiters
);
353 * Full barrier (A), see the ordering comment above.
355 smp_mb__after_atomic();
360 * Reflects a waiter being removed from the waitqueue by wakeup
363 static inline void hb_waiters_dec(struct futex_hash_bucket
*hb
)
366 atomic_dec(&hb
->waiters
);
370 static inline int hb_waiters_pending(struct futex_hash_bucket
*hb
)
373 return atomic_read(&hb
->waiters
);
380 * hash_futex - Return the hash bucket in the global hash
381 * @key: Pointer to the futex key for which the hash is calculated
383 * We hash on the keys returned from get_futex_key (see below) and return the
384 * corresponding hash bucket in the global hash.
386 static struct futex_hash_bucket
*hash_futex(union futex_key
*key
)
388 u32 hash
= jhash2((u32
*)&key
->both
.word
,
389 (sizeof(key
->both
.word
)+sizeof(key
->both
.ptr
))/4,
391 return &futex_queues
[hash
& (futex_hashsize
- 1)];
396 * match_futex - Check whether two futex keys are equal
397 * @key1: Pointer to key1
398 * @key2: Pointer to key2
400 * Return 1 if two futex_keys are equal, 0 otherwise.
402 static inline int match_futex(union futex_key
*key1
, union futex_key
*key2
)
405 && key1
->both
.word
== key2
->both
.word
406 && key1
->both
.ptr
== key2
->both
.ptr
407 && key1
->both
.offset
== key2
->both
.offset
);
411 * Take a reference to the resource addressed by a key.
412 * Can be called while holding spinlocks.
415 static void get_futex_key_refs(union futex_key
*key
)
421 * On MMU less systems futexes are always "private" as there is no per
422 * process address space. We need the smp wmb nevertheless - yes,
423 * arch/blackfin has MMU less SMP ...
425 if (!IS_ENABLED(CONFIG_MMU
)) {
426 smp_mb(); /* explicit smp_mb(); (B) */
430 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
432 ihold(key
->shared
.inode
); /* implies smp_mb(); (B) */
434 case FUT_OFF_MMSHARED
:
435 futex_get_mm(key
); /* implies smp_mb(); (B) */
439 * Private futexes do not hold reference on an inode or
440 * mm, therefore the only purpose of calling get_futex_key_refs
441 * is because we need the barrier for the lockless waiter check.
443 smp_mb(); /* explicit smp_mb(); (B) */
448 * Drop a reference to the resource addressed by a key.
449 * The hash bucket spinlock must not be held. This is
450 * a no-op for private futexes, see comment in the get
453 static void drop_futex_key_refs(union futex_key
*key
)
455 if (!key
->both
.ptr
) {
456 /* If we're here then we tried to put a key we failed to get */
461 if (!IS_ENABLED(CONFIG_MMU
))
464 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
466 iput(key
->shared
.inode
);
468 case FUT_OFF_MMSHARED
:
469 mmdrop(key
->private.mm
);
480 * futex_setup_timer - set up the sleeping hrtimer.
481 * @time: ptr to the given timeout value
482 * @timeout: the hrtimer_sleeper structure to be set up
483 * @flags: futex flags
484 * @range_ns: optional range in ns
486 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
489 static inline struct hrtimer_sleeper
*
490 futex_setup_timer(ktime_t
*time
, struct hrtimer_sleeper
*timeout
,
491 int flags
, u64 range_ns
)
496 hrtimer_init_sleeper_on_stack(timeout
, (flags
& FLAGS_CLOCKRT
) ?
497 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
500 * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
501 * effectively the same as calling hrtimer_set_expires().
503 hrtimer_set_expires_range_ns(&timeout
->timer
, *time
, range_ns
);
509 * get_futex_key() - Get parameters which are the keys for a futex
510 * @uaddr: virtual address of the futex
511 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
512 * @key: address where result is stored.
513 * @rw: mapping needs to be read/write (values: FUTEX_READ,
516 * Return: a negative error code or 0
518 * The key words are stored in @key on success.
520 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
521 * offset_within_page). For private mappings, it's (uaddr, current->mm).
522 * We can usually work out the index without swapping in the page.
524 * lock_page() might sleep, the caller should not hold a spinlock.
527 get_futex_key(u32 __user
*uaddr
, int fshared
, union futex_key
*key
, enum futex_access rw
)
529 unsigned long address
= (unsigned long)uaddr
;
530 struct mm_struct
*mm
= current
->mm
;
531 struct page
*page
, *tail
;
532 struct address_space
*mapping
;
536 * The futex address must be "naturally" aligned.
538 key
->both
.offset
= address
% PAGE_SIZE
;
539 if (unlikely((address
% sizeof(u32
)) != 0))
541 address
-= key
->both
.offset
;
543 if (unlikely(!access_ok(uaddr
, sizeof(u32
))))
546 if (unlikely(should_fail_futex(fshared
)))
550 * PROCESS_PRIVATE futexes are fast.
551 * As the mm cannot disappear under us and the 'key' only needs
552 * virtual address, we dont even have to find the underlying vma.
553 * Note : We do have to check 'uaddr' is a valid user address,
554 * but access_ok() should be faster than find_vma()
557 key
->private.mm
= mm
;
558 key
->private.address
= address
;
559 get_futex_key_refs(key
); /* implies smp_mb(); (B) */
564 /* Ignore any VERIFY_READ mapping (futex common case) */
565 if (unlikely(should_fail_futex(fshared
)))
568 err
= get_user_pages_fast(address
, 1, FOLL_WRITE
, &page
);
570 * If write access is not required (eg. FUTEX_WAIT), try
571 * and get read-only access.
573 if (err
== -EFAULT
&& rw
== FUTEX_READ
) {
574 err
= get_user_pages_fast(address
, 1, 0, &page
);
583 * The treatment of mapping from this point on is critical. The page
584 * lock protects many things but in this context the page lock
585 * stabilizes mapping, prevents inode freeing in the shared
586 * file-backed region case and guards against movement to swap cache.
588 * Strictly speaking the page lock is not needed in all cases being
589 * considered here and page lock forces unnecessarily serialization
590 * From this point on, mapping will be re-verified if necessary and
591 * page lock will be acquired only if it is unavoidable
593 * Mapping checks require the head page for any compound page so the
594 * head page and mapping is looked up now. For anonymous pages, it
595 * does not matter if the page splits in the future as the key is
596 * based on the address. For filesystem-backed pages, the tail is
597 * required as the index of the page determines the key. For
598 * base pages, there is no tail page and tail == page.
601 page
= compound_head(page
);
602 mapping
= READ_ONCE(page
->mapping
);
605 * If page->mapping is NULL, then it cannot be a PageAnon
606 * page; but it might be the ZERO_PAGE or in the gate area or
607 * in a special mapping (all cases which we are happy to fail);
608 * or it may have been a good file page when get_user_pages_fast
609 * found it, but truncated or holepunched or subjected to
610 * invalidate_complete_page2 before we got the page lock (also
611 * cases which we are happy to fail). And we hold a reference,
612 * so refcount care in invalidate_complete_page's remove_mapping
613 * prevents drop_caches from setting mapping to NULL beneath us.
615 * The case we do have to guard against is when memory pressure made
616 * shmem_writepage move it from filecache to swapcache beneath us:
617 * an unlikely race, but we do need to retry for page->mapping.
619 if (unlikely(!mapping
)) {
623 * Page lock is required to identify which special case above
624 * applies. If this is really a shmem page then the page lock
625 * will prevent unexpected transitions.
628 shmem_swizzled
= PageSwapCache(page
) || page
->mapping
;
639 * Private mappings are handled in a simple way.
641 * If the futex key is stored on an anonymous page, then the associated
642 * object is the mm which is implicitly pinned by the calling process.
644 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
645 * it's a read-only handle, it's expected that futexes attach to
646 * the object not the particular process.
648 if (PageAnon(page
)) {
650 * A RO anonymous page will never change and thus doesn't make
651 * sense for futex operations.
653 if (unlikely(should_fail_futex(fshared
)) || ro
) {
658 key
->both
.offset
|= FUT_OFF_MMSHARED
; /* ref taken on mm */
659 key
->private.mm
= mm
;
660 key
->private.address
= address
;
662 get_futex_key_refs(key
); /* implies smp_mb(); (B) */
668 * The associated futex object in this case is the inode and
669 * the page->mapping must be traversed. Ordinarily this should
670 * be stabilised under page lock but it's not strictly
671 * necessary in this case as we just want to pin the inode, not
672 * update the radix tree or anything like that.
674 * The RCU read lock is taken as the inode is finally freed
675 * under RCU. If the mapping still matches expectations then the
676 * mapping->host can be safely accessed as being a valid inode.
680 if (READ_ONCE(page
->mapping
) != mapping
) {
687 inode
= READ_ONCE(mapping
->host
);
696 * Take a reference unless it is about to be freed. Previously
697 * this reference was taken by ihold under the page lock
698 * pinning the inode in place so i_lock was unnecessary. The
699 * only way for this check to fail is if the inode was
700 * truncated in parallel which is almost certainly an
701 * application bug. In such a case, just retry.
703 * We are not calling into get_futex_key_refs() in file-backed
704 * cases, therefore a successful atomic_inc return below will
705 * guarantee that get_futex_key() will still imply smp_mb(); (B).
707 if (!atomic_inc_not_zero(&inode
->i_count
)) {
714 /* Should be impossible but lets be paranoid for now */
715 if (WARN_ON_ONCE(inode
->i_mapping
!= mapping
)) {
723 key
->both
.offset
|= FUT_OFF_INODE
; /* inode-based key */
724 key
->shared
.inode
= inode
;
725 key
->shared
.pgoff
= basepage_index(tail
);
734 static inline void put_futex_key(union futex_key
*key
)
736 drop_futex_key_refs(key
);
740 * fault_in_user_writeable() - Fault in user address and verify RW access
741 * @uaddr: pointer to faulting user space address
743 * Slow path to fixup the fault we just took in the atomic write
746 * We have no generic implementation of a non-destructive write to the
747 * user address. We know that we faulted in the atomic pagefault
748 * disabled section so we can as well avoid the #PF overhead by
749 * calling get_user_pages() right away.
751 static int fault_in_user_writeable(u32 __user
*uaddr
)
753 struct mm_struct
*mm
= current
->mm
;
756 down_read(&mm
->mmap_sem
);
757 ret
= fixup_user_fault(current
, mm
, (unsigned long)uaddr
,
758 FAULT_FLAG_WRITE
, NULL
);
759 up_read(&mm
->mmap_sem
);
761 return ret
< 0 ? ret
: 0;
765 * futex_top_waiter() - Return the highest priority waiter on a futex
766 * @hb: the hash bucket the futex_q's reside in
767 * @key: the futex key (to distinguish it from other futex futex_q's)
769 * Must be called with the hb lock held.
771 static struct futex_q
*futex_top_waiter(struct futex_hash_bucket
*hb
,
772 union futex_key
*key
)
774 struct futex_q
*this;
776 plist_for_each_entry(this, &hb
->chain
, list
) {
777 if (match_futex(&this->key
, key
))
783 static int cmpxchg_futex_value_locked(u32
*curval
, u32 __user
*uaddr
,
784 u32 uval
, u32 newval
)
789 ret
= futex_atomic_cmpxchg_inatomic(curval
, uaddr
, uval
, newval
);
795 static int get_futex_value_locked(u32
*dest
, u32 __user
*from
)
800 ret
= __get_user(*dest
, from
);
803 return ret
? -EFAULT
: 0;
810 static int refill_pi_state_cache(void)
812 struct futex_pi_state
*pi_state
;
814 if (likely(current
->pi_state_cache
))
817 pi_state
= kzalloc(sizeof(*pi_state
), GFP_KERNEL
);
822 INIT_LIST_HEAD(&pi_state
->list
);
823 /* pi_mutex gets initialized later */
824 pi_state
->owner
= NULL
;
825 refcount_set(&pi_state
->refcount
, 1);
826 pi_state
->key
= FUTEX_KEY_INIT
;
828 current
->pi_state_cache
= pi_state
;
833 static struct futex_pi_state
*alloc_pi_state(void)
835 struct futex_pi_state
*pi_state
= current
->pi_state_cache
;
838 current
->pi_state_cache
= NULL
;
843 static void get_pi_state(struct futex_pi_state
*pi_state
)
845 WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state
->refcount
));
849 * Drops a reference to the pi_state object and frees or caches it
850 * when the last reference is gone.
852 static void put_pi_state(struct futex_pi_state
*pi_state
)
857 if (!refcount_dec_and_test(&pi_state
->refcount
))
861 * If pi_state->owner is NULL, the owner is most probably dying
862 * and has cleaned up the pi_state already
864 if (pi_state
->owner
) {
865 struct task_struct
*owner
;
867 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
868 owner
= pi_state
->owner
;
870 raw_spin_lock(&owner
->pi_lock
);
871 list_del_init(&pi_state
->list
);
872 raw_spin_unlock(&owner
->pi_lock
);
874 rt_mutex_proxy_unlock(&pi_state
->pi_mutex
, owner
);
875 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
878 if (current
->pi_state_cache
) {
882 * pi_state->list is already empty.
883 * clear pi_state->owner.
884 * refcount is at 0 - put it back to 1.
886 pi_state
->owner
= NULL
;
887 refcount_set(&pi_state
->refcount
, 1);
888 current
->pi_state_cache
= pi_state
;
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 (!refcount_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(!refcount_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
= find_get_task_by_vpid(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
) || WARN_ON(plist_node_empty(&q
->list
)))
1510 lockdep_assert_held(q
->lock_ptr
);
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
1545 wake_q_add_safe(wake_q
, p
);
1549 * Caller must hold a reference on @pi_state.
1551 static int wake_futex_pi(u32 __user
*uaddr
, u32 uval
, struct futex_pi_state
*pi_state
)
1553 u32
uninitialized_var(curval
), newval
;
1554 struct task_struct
*new_owner
;
1555 bool postunlock
= false;
1556 DEFINE_WAKE_Q(wake_q
);
1559 new_owner
= rt_mutex_next_owner(&pi_state
->pi_mutex
);
1560 if (WARN_ON_ONCE(!new_owner
)) {
1562 * As per the comment in futex_unlock_pi() this should not happen.
1564 * When this happens, give up our locks and try again, giving
1565 * the futex_lock_pi() instance time to complete, either by
1566 * waiting on the rtmutex or removing itself from the futex
1574 * We pass it to the next owner. The WAITERS bit is always kept
1575 * enabled while there is PI state around. We cleanup the owner
1576 * died bit, because we are the owner.
1578 newval
= FUTEX_WAITERS
| task_pid_vnr(new_owner
);
1580 if (unlikely(should_fail_futex(true)))
1583 ret
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
);
1584 if (!ret
&& (curval
!= uval
)) {
1586 * If a unconditional UNLOCK_PI operation (user space did not
1587 * try the TID->0 transition) raced with a waiter setting the
1588 * FUTEX_WAITERS flag between get_user() and locking the hash
1589 * bucket lock, retry the operation.
1591 if ((FUTEX_TID_MASK
& curval
) == uval
)
1601 * This is a point of no return; once we modify the uval there is no
1602 * going back and subsequent operations must not fail.
1605 raw_spin_lock(&pi_state
->owner
->pi_lock
);
1606 WARN_ON(list_empty(&pi_state
->list
));
1607 list_del_init(&pi_state
->list
);
1608 raw_spin_unlock(&pi_state
->owner
->pi_lock
);
1610 raw_spin_lock(&new_owner
->pi_lock
);
1611 WARN_ON(!list_empty(&pi_state
->list
));
1612 list_add(&pi_state
->list
, &new_owner
->pi_state_list
);
1613 pi_state
->owner
= new_owner
;
1614 raw_spin_unlock(&new_owner
->pi_lock
);
1616 postunlock
= __rt_mutex_futex_unlock(&pi_state
->pi_mutex
, &wake_q
);
1619 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1622 rt_mutex_postunlock(&wake_q
);
1628 * Express the locking dependencies for lockdep:
1631 double_lock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1634 spin_lock(&hb1
->lock
);
1636 spin_lock_nested(&hb2
->lock
, SINGLE_DEPTH_NESTING
);
1637 } else { /* hb1 > hb2 */
1638 spin_lock(&hb2
->lock
);
1639 spin_lock_nested(&hb1
->lock
, SINGLE_DEPTH_NESTING
);
1644 double_unlock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1646 spin_unlock(&hb1
->lock
);
1648 spin_unlock(&hb2
->lock
);
1652 * Wake up waiters matching bitset queued on this futex (uaddr).
1655 futex_wake(u32 __user
*uaddr
, unsigned int flags
, int nr_wake
, u32 bitset
)
1657 struct futex_hash_bucket
*hb
;
1658 struct futex_q
*this, *next
;
1659 union futex_key key
= FUTEX_KEY_INIT
;
1661 DEFINE_WAKE_Q(wake_q
);
1666 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, FUTEX_READ
);
1667 if (unlikely(ret
!= 0))
1670 hb
= hash_futex(&key
);
1672 /* Make sure we really have tasks to wakeup */
1673 if (!hb_waiters_pending(hb
))
1676 spin_lock(&hb
->lock
);
1678 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
1679 if (match_futex (&this->key
, &key
)) {
1680 if (this->pi_state
|| this->rt_waiter
) {
1685 /* Check if one of the bits is set in both bitsets */
1686 if (!(this->bitset
& bitset
))
1689 mark_wake_futex(&wake_q
, this);
1690 if (++ret
>= nr_wake
)
1695 spin_unlock(&hb
->lock
);
1698 put_futex_key(&key
);
1703 static int futex_atomic_op_inuser(unsigned int encoded_op
, u32 __user
*uaddr
)
1705 unsigned int op
= (encoded_op
& 0x70000000) >> 28;
1706 unsigned int cmp
= (encoded_op
& 0x0f000000) >> 24;
1707 int oparg
= sign_extend32((encoded_op
& 0x00fff000) >> 12, 11);
1708 int cmparg
= sign_extend32(encoded_op
& 0x00000fff, 11);
1711 if (encoded_op
& (FUTEX_OP_OPARG_SHIFT
<< 28)) {
1712 if (oparg
< 0 || oparg
> 31) {
1713 char comm
[sizeof(current
->comm
)];
1715 * kill this print and return -EINVAL when userspace
1718 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1719 get_task_comm(comm
, current
), oparg
);
1725 if (!access_ok(uaddr
, sizeof(u32
)))
1728 ret
= arch_futex_atomic_op_inuser(op
, oparg
, &oldval
, uaddr
);
1733 case FUTEX_OP_CMP_EQ
:
1734 return oldval
== cmparg
;
1735 case FUTEX_OP_CMP_NE
:
1736 return oldval
!= cmparg
;
1737 case FUTEX_OP_CMP_LT
:
1738 return oldval
< cmparg
;
1739 case FUTEX_OP_CMP_GE
:
1740 return oldval
>= cmparg
;
1741 case FUTEX_OP_CMP_LE
:
1742 return oldval
<= cmparg
;
1743 case FUTEX_OP_CMP_GT
:
1744 return oldval
> cmparg
;
1751 * Wake up all waiters hashed on the physical page that is mapped
1752 * to this virtual address:
1755 futex_wake_op(u32 __user
*uaddr1
, unsigned int flags
, u32 __user
*uaddr2
,
1756 int nr_wake
, int nr_wake2
, int op
)
1758 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1759 struct futex_hash_bucket
*hb1
, *hb2
;
1760 struct futex_q
*this, *next
;
1762 DEFINE_WAKE_Q(wake_q
);
1765 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, FUTEX_READ
);
1766 if (unlikely(ret
!= 0))
1768 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, FUTEX_WRITE
);
1769 if (unlikely(ret
!= 0))
1772 hb1
= hash_futex(&key1
);
1773 hb2
= hash_futex(&key2
);
1776 double_lock_hb(hb1
, hb2
);
1777 op_ret
= futex_atomic_op_inuser(op
, uaddr2
);
1778 if (unlikely(op_ret
< 0)) {
1779 double_unlock_hb(hb1
, hb2
);
1781 if (!IS_ENABLED(CONFIG_MMU
) ||
1782 unlikely(op_ret
!= -EFAULT
&& op_ret
!= -EAGAIN
)) {
1784 * we don't get EFAULT from MMU faults if we don't have
1785 * an MMU, but we might get them from range checking
1791 if (op_ret
== -EFAULT
) {
1792 ret
= fault_in_user_writeable(uaddr2
);
1797 if (!(flags
& FLAGS_SHARED
)) {
1802 put_futex_key(&key2
);
1803 put_futex_key(&key1
);
1808 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1809 if (match_futex (&this->key
, &key1
)) {
1810 if (this->pi_state
|| this->rt_waiter
) {
1814 mark_wake_futex(&wake_q
, this);
1815 if (++ret
>= nr_wake
)
1822 plist_for_each_entry_safe(this, next
, &hb2
->chain
, list
) {
1823 if (match_futex (&this->key
, &key2
)) {
1824 if (this->pi_state
|| this->rt_waiter
) {
1828 mark_wake_futex(&wake_q
, this);
1829 if (++op_ret
>= nr_wake2
)
1837 double_unlock_hb(hb1
, hb2
);
1840 put_futex_key(&key2
);
1842 put_futex_key(&key1
);
1848 * requeue_futex() - Requeue a futex_q from one hb to another
1849 * @q: the futex_q to requeue
1850 * @hb1: the source hash_bucket
1851 * @hb2: the target hash_bucket
1852 * @key2: the new key for the requeued futex_q
1855 void requeue_futex(struct futex_q
*q
, struct futex_hash_bucket
*hb1
,
1856 struct futex_hash_bucket
*hb2
, union futex_key
*key2
)
1860 * If key1 and key2 hash to the same bucket, no need to
1863 if (likely(&hb1
->chain
!= &hb2
->chain
)) {
1864 plist_del(&q
->list
, &hb1
->chain
);
1865 hb_waiters_dec(hb1
);
1866 hb_waiters_inc(hb2
);
1867 plist_add(&q
->list
, &hb2
->chain
);
1868 q
->lock_ptr
= &hb2
->lock
;
1870 get_futex_key_refs(key2
);
1875 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1877 * @key: the key of the requeue target futex
1878 * @hb: the hash_bucket of the requeue target futex
1880 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1881 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1882 * to the requeue target futex so the waiter can detect the wakeup on the right
1883 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1884 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1885 * to protect access to the pi_state to fixup the owner later. Must be called
1886 * with both q->lock_ptr and hb->lock held.
1889 void requeue_pi_wake_futex(struct futex_q
*q
, union futex_key
*key
,
1890 struct futex_hash_bucket
*hb
)
1892 get_futex_key_refs(key
);
1897 WARN_ON(!q
->rt_waiter
);
1898 q
->rt_waiter
= NULL
;
1900 q
->lock_ptr
= &hb
->lock
;
1902 wake_up_state(q
->task
, TASK_NORMAL
);
1906 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1907 * @pifutex: the user address of the to futex
1908 * @hb1: the from futex hash bucket, must be locked by the caller
1909 * @hb2: the to futex hash bucket, must be locked by the caller
1910 * @key1: the from futex key
1911 * @key2: the to futex key
1912 * @ps: address to store the pi_state pointer
1913 * @exiting: Pointer to store the task pointer of the owner task
1914 * which is in the middle of exiting
1915 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1917 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1918 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1919 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1920 * hb1 and hb2 must be held by the caller.
1922 * @exiting is only set when the return value is -EBUSY. If so, this holds
1923 * a refcount on the exiting task on return and the caller needs to drop it
1924 * after waiting for the exit to complete.
1927 * - 0 - failed to acquire the lock atomically;
1928 * - >0 - acquired the lock, return value is vpid of the top_waiter
1932 futex_proxy_trylock_atomic(u32 __user
*pifutex
, struct futex_hash_bucket
*hb1
,
1933 struct futex_hash_bucket
*hb2
, union futex_key
*key1
,
1934 union futex_key
*key2
, struct futex_pi_state
**ps
,
1935 struct task_struct
**exiting
, int set_waiters
)
1937 struct futex_q
*top_waiter
= NULL
;
1941 if (get_futex_value_locked(&curval
, pifutex
))
1944 if (unlikely(should_fail_futex(true)))
1948 * Find the top_waiter and determine if there are additional waiters.
1949 * If the caller intends to requeue more than 1 waiter to pifutex,
1950 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1951 * as we have means to handle the possible fault. If not, don't set
1952 * the bit unecessarily as it will force the subsequent unlock to enter
1955 top_waiter
= futex_top_waiter(hb1
, key1
);
1957 /* There are no waiters, nothing for us to do. */
1961 /* Ensure we requeue to the expected futex. */
1962 if (!match_futex(top_waiter
->requeue_pi_key
, key2
))
1966 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1967 * the contended case or if set_waiters is 1. The pi_state is returned
1968 * in ps in contended cases.
1970 vpid
= task_pid_vnr(top_waiter
->task
);
1971 ret
= futex_lock_pi_atomic(pifutex
, hb2
, key2
, ps
, top_waiter
->task
,
1972 exiting
, set_waiters
);
1974 requeue_pi_wake_futex(top_waiter
, key2
, hb2
);
1981 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1982 * @uaddr1: source futex user address
1983 * @flags: futex flags (FLAGS_SHARED, etc.)
1984 * @uaddr2: target futex user address
1985 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1986 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1987 * @cmpval: @uaddr1 expected value (or %NULL)
1988 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1989 * pi futex (pi to pi requeue is not supported)
1991 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1992 * uaddr2 atomically on behalf of the top waiter.
1995 * - >=0 - on success, the number of tasks requeued or woken;
1998 static int futex_requeue(u32 __user
*uaddr1
, unsigned int flags
,
1999 u32 __user
*uaddr2
, int nr_wake
, int nr_requeue
,
2000 u32
*cmpval
, int requeue_pi
)
2002 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
2003 int drop_count
= 0, task_count
= 0, ret
;
2004 struct futex_pi_state
*pi_state
= NULL
;
2005 struct futex_hash_bucket
*hb1
, *hb2
;
2006 struct futex_q
*this, *next
;
2007 DEFINE_WAKE_Q(wake_q
);
2009 if (nr_wake
< 0 || nr_requeue
< 0)
2013 * When PI not supported: return -ENOSYS if requeue_pi is true,
2014 * consequently the compiler knows requeue_pi is always false past
2015 * this point which will optimize away all the conditional code
2018 if (!IS_ENABLED(CONFIG_FUTEX_PI
) && requeue_pi
)
2023 * Requeue PI only works on two distinct uaddrs. This
2024 * check is only valid for private futexes. See below.
2026 if (uaddr1
== uaddr2
)
2030 * requeue_pi requires a pi_state, try to allocate it now
2031 * without any locks in case it fails.
2033 if (refill_pi_state_cache())
2036 * requeue_pi must wake as many tasks as it can, up to nr_wake
2037 * + nr_requeue, since it acquires the rt_mutex prior to
2038 * returning to userspace, so as to not leave the rt_mutex with
2039 * waiters and no owner. However, second and third wake-ups
2040 * cannot be predicted as they involve race conditions with the
2041 * first wake and a fault while looking up the pi_state. Both
2042 * pthread_cond_signal() and pthread_cond_broadcast() should
2050 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, FUTEX_READ
);
2051 if (unlikely(ret
!= 0))
2053 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
,
2054 requeue_pi
? FUTEX_WRITE
: FUTEX_READ
);
2055 if (unlikely(ret
!= 0))
2059 * The check above which compares uaddrs is not sufficient for
2060 * shared futexes. We need to compare the keys:
2062 if (requeue_pi
&& match_futex(&key1
, &key2
)) {
2067 hb1
= hash_futex(&key1
);
2068 hb2
= hash_futex(&key2
);
2071 hb_waiters_inc(hb2
);
2072 double_lock_hb(hb1
, hb2
);
2074 if (likely(cmpval
!= NULL
)) {
2077 ret
= get_futex_value_locked(&curval
, uaddr1
);
2079 if (unlikely(ret
)) {
2080 double_unlock_hb(hb1
, hb2
);
2081 hb_waiters_dec(hb2
);
2083 ret
= get_user(curval
, uaddr1
);
2087 if (!(flags
& FLAGS_SHARED
))
2090 put_futex_key(&key2
);
2091 put_futex_key(&key1
);
2094 if (curval
!= *cmpval
) {
2100 if (requeue_pi
&& (task_count
- nr_wake
< nr_requeue
)) {
2101 struct task_struct
*exiting
= NULL
;
2104 * Attempt to acquire uaddr2 and wake the top waiter. If we
2105 * intend to requeue waiters, force setting the FUTEX_WAITERS
2106 * bit. We force this here where we are able to easily handle
2107 * faults rather in the requeue loop below.
2109 ret
= futex_proxy_trylock_atomic(uaddr2
, hb1
, hb2
, &key1
,
2111 &exiting
, nr_requeue
);
2114 * At this point the top_waiter has either taken uaddr2 or is
2115 * waiting on it. If the former, then the pi_state will not
2116 * exist yet, look it up one more time to ensure we have a
2117 * reference to it. If the lock was taken, ret contains the
2118 * vpid of the top waiter task.
2119 * If the lock was not taken, we have pi_state and an initial
2120 * refcount on it. In case of an error we have nothing.
2127 * If we acquired the lock, then the user space value
2128 * of uaddr2 should be vpid. It cannot be changed by
2129 * the top waiter as it is blocked on hb2 lock if it
2130 * tries to do so. If something fiddled with it behind
2131 * our back the pi state lookup might unearth it. So
2132 * we rather use the known value than rereading and
2133 * handing potential crap to lookup_pi_state.
2135 * If that call succeeds then we have pi_state and an
2136 * initial refcount on it.
2138 ret
= lookup_pi_state(uaddr2
, ret
, hb2
, &key2
,
2139 &pi_state
, &exiting
);
2144 /* We hold a reference on the pi state. */
2147 /* If the above failed, then pi_state is NULL */
2149 double_unlock_hb(hb1
, hb2
);
2150 hb_waiters_dec(hb2
);
2151 put_futex_key(&key2
);
2152 put_futex_key(&key1
);
2153 ret
= fault_in_user_writeable(uaddr2
);
2160 * Two reasons for this:
2161 * - EBUSY: Owner is exiting and we just wait for the
2163 * - EAGAIN: The user space value changed.
2165 double_unlock_hb(hb1
, hb2
);
2166 hb_waiters_dec(hb2
);
2167 put_futex_key(&key2
);
2168 put_futex_key(&key1
);
2170 * Handle the case where the owner is in the middle of
2171 * exiting. Wait for the exit to complete otherwise
2172 * this task might loop forever, aka. live lock.
2174 wait_for_owner_exiting(ret
, exiting
);
2182 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
2183 if (task_count
- nr_wake
>= nr_requeue
)
2186 if (!match_futex(&this->key
, &key1
))
2190 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2191 * be paired with each other and no other futex ops.
2193 * We should never be requeueing a futex_q with a pi_state,
2194 * which is awaiting a futex_unlock_pi().
2196 if ((requeue_pi
&& !this->rt_waiter
) ||
2197 (!requeue_pi
&& this->rt_waiter
) ||
2204 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2205 * lock, we already woke the top_waiter. If not, it will be
2206 * woken by futex_unlock_pi().
2208 if (++task_count
<= nr_wake
&& !requeue_pi
) {
2209 mark_wake_futex(&wake_q
, this);
2213 /* Ensure we requeue to the expected futex for requeue_pi. */
2214 if (requeue_pi
&& !match_futex(this->requeue_pi_key
, &key2
)) {
2220 * Requeue nr_requeue waiters and possibly one more in the case
2221 * of requeue_pi if we couldn't acquire the lock atomically.
2225 * Prepare the waiter to take the rt_mutex. Take a
2226 * refcount on the pi_state and store the pointer in
2227 * the futex_q object of the waiter.
2229 get_pi_state(pi_state
);
2230 this->pi_state
= pi_state
;
2231 ret
= rt_mutex_start_proxy_lock(&pi_state
->pi_mutex
,
2236 * We got the lock. We do neither drop the
2237 * refcount on pi_state nor clear
2238 * this->pi_state because the waiter needs the
2239 * pi_state for cleaning up the user space
2240 * value. It will drop the refcount after
2243 requeue_pi_wake_futex(this, &key2
, hb2
);
2248 * rt_mutex_start_proxy_lock() detected a
2249 * potential deadlock when we tried to queue
2250 * that waiter. Drop the pi_state reference
2251 * which we took above and remove the pointer
2252 * to the state from the waiters futex_q
2255 this->pi_state
= NULL
;
2256 put_pi_state(pi_state
);
2258 * We stop queueing more waiters and let user
2259 * space deal with the mess.
2264 requeue_futex(this, hb1
, hb2
, &key2
);
2269 * We took an extra initial reference to the pi_state either
2270 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2271 * need to drop it here again.
2273 put_pi_state(pi_state
);
2276 double_unlock_hb(hb1
, hb2
);
2278 hb_waiters_dec(hb2
);
2281 * drop_futex_key_refs() must be called outside the spinlocks. During
2282 * the requeue we moved futex_q's from the hash bucket at key1 to the
2283 * one at key2 and updated their key pointer. We no longer need to
2284 * hold the references to key1.
2286 while (--drop_count
>= 0)
2287 drop_futex_key_refs(&key1
);
2290 put_futex_key(&key2
);
2292 put_futex_key(&key1
);
2294 return ret
? ret
: task_count
;
2297 /* The key must be already stored in q->key. */
2298 static inline struct futex_hash_bucket
*queue_lock(struct futex_q
*q
)
2299 __acquires(&hb
->lock
)
2301 struct futex_hash_bucket
*hb
;
2303 hb
= hash_futex(&q
->key
);
2306 * Increment the counter before taking the lock so that
2307 * a potential waker won't miss a to-be-slept task that is
2308 * waiting for the spinlock. This is safe as all queue_lock()
2309 * users end up calling queue_me(). Similarly, for housekeeping,
2310 * decrement the counter at queue_unlock() when some error has
2311 * occurred and we don't end up adding the task to the list.
2313 hb_waiters_inc(hb
); /* implies smp_mb(); (A) */
2315 q
->lock_ptr
= &hb
->lock
;
2317 spin_lock(&hb
->lock
);
2322 queue_unlock(struct futex_hash_bucket
*hb
)
2323 __releases(&hb
->lock
)
2325 spin_unlock(&hb
->lock
);
2329 static inline void __queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2334 * The priority used to register this element is
2335 * - either the real thread-priority for the real-time threads
2336 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2337 * - or MAX_RT_PRIO for non-RT threads.
2338 * Thus, all RT-threads are woken first in priority order, and
2339 * the others are woken last, in FIFO order.
2341 prio
= min(current
->normal_prio
, MAX_RT_PRIO
);
2343 plist_node_init(&q
->list
, prio
);
2344 plist_add(&q
->list
, &hb
->chain
);
2349 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2350 * @q: The futex_q to enqueue
2351 * @hb: The destination hash bucket
2353 * The hb->lock must be held by the caller, and is released here. A call to
2354 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2355 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2356 * or nothing if the unqueue is done as part of the wake process and the unqueue
2357 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2360 static inline void queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2361 __releases(&hb
->lock
)
2364 spin_unlock(&hb
->lock
);
2368 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2369 * @q: The futex_q to unqueue
2371 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2372 * be paired with exactly one earlier call to queue_me().
2375 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2376 * - 0 - if the futex_q was already removed by the waking thread
2378 static int unqueue_me(struct futex_q
*q
)
2380 spinlock_t
*lock_ptr
;
2383 /* In the common case we don't take the spinlock, which is nice. */
2386 * q->lock_ptr can change between this read and the following spin_lock.
2387 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2388 * optimizing lock_ptr out of the logic below.
2390 lock_ptr
= READ_ONCE(q
->lock_ptr
);
2391 if (lock_ptr
!= NULL
) {
2392 spin_lock(lock_ptr
);
2394 * q->lock_ptr can change between reading it and
2395 * spin_lock(), causing us to take the wrong lock. This
2396 * corrects the race condition.
2398 * Reasoning goes like this: if we have the wrong lock,
2399 * q->lock_ptr must have changed (maybe several times)
2400 * between reading it and the spin_lock(). It can
2401 * change again after the spin_lock() but only if it was
2402 * already changed before the spin_lock(). It cannot,
2403 * however, change back to the original value. Therefore
2404 * we can detect whether we acquired the correct lock.
2406 if (unlikely(lock_ptr
!= q
->lock_ptr
)) {
2407 spin_unlock(lock_ptr
);
2412 BUG_ON(q
->pi_state
);
2414 spin_unlock(lock_ptr
);
2418 drop_futex_key_refs(&q
->key
);
2423 * PI futexes can not be requeued and must remove themself from the
2424 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2427 static void unqueue_me_pi(struct futex_q
*q
)
2428 __releases(q
->lock_ptr
)
2432 BUG_ON(!q
->pi_state
);
2433 put_pi_state(q
->pi_state
);
2436 spin_unlock(q
->lock_ptr
);
2439 static int fixup_pi_state_owner(u32 __user
*uaddr
, struct futex_q
*q
,
2440 struct task_struct
*argowner
)
2442 struct futex_pi_state
*pi_state
= q
->pi_state
;
2443 u32 uval
, uninitialized_var(curval
), newval
;
2444 struct task_struct
*oldowner
, *newowner
;
2448 lockdep_assert_held(q
->lock_ptr
);
2450 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2452 oldowner
= pi_state
->owner
;
2455 * We are here because either:
2457 * - we stole the lock and pi_state->owner needs updating to reflect
2458 * that (@argowner == current),
2462 * - someone stole our lock and we need to fix things to point to the
2463 * new owner (@argowner == NULL).
2465 * Either way, we have to replace the TID in the user space variable.
2466 * This must be atomic as we have to preserve the owner died bit here.
2468 * Note: We write the user space value _before_ changing the pi_state
2469 * because we can fault here. Imagine swapped out pages or a fork
2470 * that marked all the anonymous memory readonly for cow.
2472 * Modifying pi_state _before_ the user space value would leave the
2473 * pi_state in an inconsistent state when we fault here, because we
2474 * need to drop the locks to handle the fault. This might be observed
2475 * in the PID check in lookup_pi_state.
2479 if (oldowner
!= current
) {
2481 * We raced against a concurrent self; things are
2482 * already fixed up. Nothing to do.
2488 if (__rt_mutex_futex_trylock(&pi_state
->pi_mutex
)) {
2489 /* We got the lock after all, nothing to fix. */
2495 * Since we just failed the trylock; there must be an owner.
2497 newowner
= rt_mutex_owner(&pi_state
->pi_mutex
);
2500 WARN_ON_ONCE(argowner
!= current
);
2501 if (oldowner
== current
) {
2503 * We raced against a concurrent self; things are
2504 * already fixed up. Nothing to do.
2509 newowner
= argowner
;
2512 newtid
= task_pid_vnr(newowner
) | FUTEX_WAITERS
;
2514 if (!pi_state
->owner
)
2515 newtid
|= FUTEX_OWNER_DIED
;
2517 err
= get_futex_value_locked(&uval
, uaddr
);
2522 newval
= (uval
& FUTEX_OWNER_DIED
) | newtid
;
2524 err
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
);
2534 * We fixed up user space. Now we need to fix the pi_state
2537 if (pi_state
->owner
!= NULL
) {
2538 raw_spin_lock(&pi_state
->owner
->pi_lock
);
2539 WARN_ON(list_empty(&pi_state
->list
));
2540 list_del_init(&pi_state
->list
);
2541 raw_spin_unlock(&pi_state
->owner
->pi_lock
);
2544 pi_state
->owner
= newowner
;
2546 raw_spin_lock(&newowner
->pi_lock
);
2547 WARN_ON(!list_empty(&pi_state
->list
));
2548 list_add(&pi_state
->list
, &newowner
->pi_state_list
);
2549 raw_spin_unlock(&newowner
->pi_lock
);
2550 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2555 * In order to reschedule or handle a page fault, we need to drop the
2556 * locks here. In the case of a fault, this gives the other task
2557 * (either the highest priority waiter itself or the task which stole
2558 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2559 * are back from handling the fault we need to check the pi_state after
2560 * reacquiring the locks and before trying to do another fixup. When
2561 * the fixup has been done already we simply return.
2563 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2564 * drop hb->lock since the caller owns the hb -> futex_q relation.
2565 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2568 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2569 spin_unlock(q
->lock_ptr
);
2573 ret
= fault_in_user_writeable(uaddr
);
2587 spin_lock(q
->lock_ptr
);
2588 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2591 * Check if someone else fixed it for us:
2593 if (pi_state
->owner
!= oldowner
) {
2604 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2608 static long futex_wait_restart(struct restart_block
*restart
);
2611 * fixup_owner() - Post lock pi_state and corner case management
2612 * @uaddr: user address of the futex
2613 * @q: futex_q (contains pi_state and access to the rt_mutex)
2614 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2616 * After attempting to lock an rt_mutex, this function is called to cleanup
2617 * the pi_state owner as well as handle race conditions that may allow us to
2618 * acquire the lock. Must be called with the hb lock held.
2621 * - 1 - success, lock taken;
2622 * - 0 - success, lock not taken;
2623 * - <0 - on error (-EFAULT)
2625 static int fixup_owner(u32 __user
*uaddr
, struct futex_q
*q
, int locked
)
2631 * Got the lock. We might not be the anticipated owner if we
2632 * did a lock-steal - fix up the PI-state in that case:
2634 * Speculative pi_state->owner read (we don't hold wait_lock);
2635 * since we own the lock pi_state->owner == current is the
2636 * stable state, anything else needs more attention.
2638 if (q
->pi_state
->owner
!= current
)
2639 ret
= fixup_pi_state_owner(uaddr
, q
, current
);
2644 * If we didn't get the lock; check if anybody stole it from us. In
2645 * that case, we need to fix up the uval to point to them instead of
2646 * us, otherwise bad things happen. [10]
2648 * Another speculative read; pi_state->owner == current is unstable
2649 * but needs our attention.
2651 if (q
->pi_state
->owner
== current
) {
2652 ret
= fixup_pi_state_owner(uaddr
, q
, NULL
);
2657 * Paranoia check. If we did not take the lock, then we should not be
2658 * the owner of the rt_mutex.
2660 if (rt_mutex_owner(&q
->pi_state
->pi_mutex
) == current
) {
2661 printk(KERN_ERR
"fixup_owner: ret = %d pi-mutex: %p "
2662 "pi-state %p\n", ret
,
2663 q
->pi_state
->pi_mutex
.owner
,
2664 q
->pi_state
->owner
);
2668 return ret
? ret
: locked
;
2672 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2673 * @hb: the futex hash bucket, must be locked by the caller
2674 * @q: the futex_q to queue up on
2675 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2677 static void futex_wait_queue_me(struct futex_hash_bucket
*hb
, struct futex_q
*q
,
2678 struct hrtimer_sleeper
*timeout
)
2681 * The task state is guaranteed to be set before another task can
2682 * wake it. set_current_state() is implemented using smp_store_mb() and
2683 * queue_me() calls spin_unlock() upon completion, both serializing
2684 * access to the hash list and forcing another memory barrier.
2686 set_current_state(TASK_INTERRUPTIBLE
);
2691 hrtimer_sleeper_start_expires(timeout
, HRTIMER_MODE_ABS
);
2694 * If we have been removed from the hash list, then another task
2695 * has tried to wake us, and we can skip the call to schedule().
2697 if (likely(!plist_node_empty(&q
->list
))) {
2699 * If the timer has already expired, current will already be
2700 * flagged for rescheduling. Only call schedule if there
2701 * is no timeout, or if it has yet to expire.
2703 if (!timeout
|| timeout
->task
)
2704 freezable_schedule();
2706 __set_current_state(TASK_RUNNING
);
2710 * futex_wait_setup() - Prepare to wait on a futex
2711 * @uaddr: the futex userspace address
2712 * @val: the expected value
2713 * @flags: futex flags (FLAGS_SHARED, etc.)
2714 * @q: the associated futex_q
2715 * @hb: storage for hash_bucket pointer to be returned to caller
2717 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2718 * compare it with the expected value. Handle atomic faults internally.
2719 * Return with the hb lock held and a q.key reference on success, and unlocked
2720 * with no q.key reference on failure.
2723 * - 0 - uaddr contains val and hb has been locked;
2724 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2726 static int futex_wait_setup(u32 __user
*uaddr
, u32 val
, unsigned int flags
,
2727 struct futex_q
*q
, struct futex_hash_bucket
**hb
)
2733 * Access the page AFTER the hash-bucket is locked.
2734 * Order is important:
2736 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2737 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2739 * The basic logical guarantee of a futex is that it blocks ONLY
2740 * if cond(var) is known to be true at the time of blocking, for
2741 * any cond. If we locked the hash-bucket after testing *uaddr, that
2742 * would open a race condition where we could block indefinitely with
2743 * cond(var) false, which would violate the guarantee.
2745 * On the other hand, we insert q and release the hash-bucket only
2746 * after testing *uaddr. This guarantees that futex_wait() will NOT
2747 * absorb a wakeup if *uaddr does not match the desired values
2748 * while the syscall executes.
2751 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
->key
, FUTEX_READ
);
2752 if (unlikely(ret
!= 0))
2756 *hb
= queue_lock(q
);
2758 ret
= get_futex_value_locked(&uval
, uaddr
);
2763 ret
= get_user(uval
, uaddr
);
2767 if (!(flags
& FLAGS_SHARED
))
2770 put_futex_key(&q
->key
);
2781 put_futex_key(&q
->key
);
2785 static int futex_wait(u32 __user
*uaddr
, unsigned int flags
, u32 val
,
2786 ktime_t
*abs_time
, u32 bitset
)
2788 struct hrtimer_sleeper timeout
, *to
;
2789 struct restart_block
*restart
;
2790 struct futex_hash_bucket
*hb
;
2791 struct futex_q q
= futex_q_init
;
2798 to
= futex_setup_timer(abs_time
, &timeout
, flags
,
2799 current
->timer_slack_ns
);
2802 * Prepare to wait on uaddr. On success, holds hb lock and increments
2805 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2809 /* queue_me and wait for wakeup, timeout, or a signal. */
2810 futex_wait_queue_me(hb
, &q
, to
);
2812 /* If we were woken (and unqueued), we succeeded, whatever. */
2814 /* unqueue_me() drops q.key ref */
2815 if (!unqueue_me(&q
))
2818 if (to
&& !to
->task
)
2822 * We expect signal_pending(current), but we might be the
2823 * victim of a spurious wakeup as well.
2825 if (!signal_pending(current
))
2832 restart
= ¤t
->restart_block
;
2833 restart
->fn
= futex_wait_restart
;
2834 restart
->futex
.uaddr
= uaddr
;
2835 restart
->futex
.val
= val
;
2836 restart
->futex
.time
= *abs_time
;
2837 restart
->futex
.bitset
= bitset
;
2838 restart
->futex
.flags
= flags
| FLAGS_HAS_TIMEOUT
;
2840 ret
= -ERESTART_RESTARTBLOCK
;
2844 hrtimer_cancel(&to
->timer
);
2845 destroy_hrtimer_on_stack(&to
->timer
);
2851 static long futex_wait_restart(struct restart_block
*restart
)
2853 u32 __user
*uaddr
= restart
->futex
.uaddr
;
2854 ktime_t t
, *tp
= NULL
;
2856 if (restart
->futex
.flags
& FLAGS_HAS_TIMEOUT
) {
2857 t
= restart
->futex
.time
;
2860 restart
->fn
= do_no_restart_syscall
;
2862 return (long)futex_wait(uaddr
, restart
->futex
.flags
,
2863 restart
->futex
.val
, tp
, restart
->futex
.bitset
);
2868 * Userspace tried a 0 -> TID atomic transition of the futex value
2869 * and failed. The kernel side here does the whole locking operation:
2870 * if there are waiters then it will block as a consequence of relying
2871 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2872 * a 0 value of the futex too.).
2874 * Also serves as futex trylock_pi()'ing, and due semantics.
2876 static int futex_lock_pi(u32 __user
*uaddr
, unsigned int flags
,
2877 ktime_t
*time
, int trylock
)
2879 struct hrtimer_sleeper timeout
, *to
;
2880 struct futex_pi_state
*pi_state
= NULL
;
2881 struct task_struct
*exiting
= NULL
;
2882 struct rt_mutex_waiter rt_waiter
;
2883 struct futex_hash_bucket
*hb
;
2884 struct futex_q q
= futex_q_init
;
2887 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
2890 if (refill_pi_state_cache())
2893 to
= futex_setup_timer(time
, &timeout
, FLAGS_CLOCKRT
, 0);
2896 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
.key
, FUTEX_WRITE
);
2897 if (unlikely(ret
!= 0))
2901 hb
= queue_lock(&q
);
2903 ret
= futex_lock_pi_atomic(uaddr
, hb
, &q
.key
, &q
.pi_state
, current
,
2905 if (unlikely(ret
)) {
2907 * Atomic work succeeded and we got the lock,
2908 * or failed. Either way, we do _not_ block.
2912 /* We got the lock. */
2914 goto out_unlock_put_key
;
2920 * Two reasons for this:
2921 * - EBUSY: Task is exiting and we just wait for the
2923 * - EAGAIN: The user space value changed.
2926 put_futex_key(&q
.key
);
2928 * Handle the case where the owner is in the middle of
2929 * exiting. Wait for the exit to complete otherwise
2930 * this task might loop forever, aka. live lock.
2932 wait_for_owner_exiting(ret
, exiting
);
2936 goto out_unlock_put_key
;
2940 WARN_ON(!q
.pi_state
);
2943 * Only actually queue now that the atomic ops are done:
2948 ret
= rt_mutex_futex_trylock(&q
.pi_state
->pi_mutex
);
2949 /* Fixup the trylock return value: */
2950 ret
= ret
? 0 : -EWOULDBLOCK
;
2954 rt_mutex_init_waiter(&rt_waiter
);
2957 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2958 * hold it while doing rt_mutex_start_proxy(), because then it will
2959 * include hb->lock in the blocking chain, even through we'll not in
2960 * fact hold it while blocking. This will lead it to report -EDEADLK
2961 * and BUG when futex_unlock_pi() interleaves with this.
2963 * Therefore acquire wait_lock while holding hb->lock, but drop the
2964 * latter before calling __rt_mutex_start_proxy_lock(). This
2965 * interleaves with futex_unlock_pi() -- which does a similar lock
2966 * handoff -- such that the latter can observe the futex_q::pi_state
2967 * before __rt_mutex_start_proxy_lock() is done.
2969 raw_spin_lock_irq(&q
.pi_state
->pi_mutex
.wait_lock
);
2970 spin_unlock(q
.lock_ptr
);
2972 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2973 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2974 * it sees the futex_q::pi_state.
2976 ret
= __rt_mutex_start_proxy_lock(&q
.pi_state
->pi_mutex
, &rt_waiter
, current
);
2977 raw_spin_unlock_irq(&q
.pi_state
->pi_mutex
.wait_lock
);
2986 hrtimer_sleeper_start_expires(to
, HRTIMER_MODE_ABS
);
2988 ret
= rt_mutex_wait_proxy_lock(&q
.pi_state
->pi_mutex
, to
, &rt_waiter
);
2991 spin_lock(q
.lock_ptr
);
2993 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2994 * first acquire the hb->lock before removing the lock from the
2995 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2998 * In particular; it is important that futex_unlock_pi() can not
2999 * observe this inconsistency.
3001 if (ret
&& !rt_mutex_cleanup_proxy_lock(&q
.pi_state
->pi_mutex
, &rt_waiter
))
3006 * Fixup the pi_state owner and possibly acquire the lock if we
3009 res
= fixup_owner(uaddr
, &q
, !ret
);
3011 * If fixup_owner() returned an error, proprogate that. If it acquired
3012 * the lock, clear our -ETIMEDOUT or -EINTR.
3015 ret
= (res
< 0) ? res
: 0;
3018 * If fixup_owner() faulted and was unable to handle the fault, unlock
3019 * it and return the fault to userspace.
3021 if (ret
&& (rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
)) {
3022 pi_state
= q
.pi_state
;
3023 get_pi_state(pi_state
);
3026 /* Unqueue and drop the lock */
3030 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
3031 put_pi_state(pi_state
);
3040 put_futex_key(&q
.key
);
3043 hrtimer_cancel(&to
->timer
);
3044 destroy_hrtimer_on_stack(&to
->timer
);
3046 return ret
!= -EINTR
? ret
: -ERESTARTNOINTR
;
3051 ret
= fault_in_user_writeable(uaddr
);
3055 if (!(flags
& FLAGS_SHARED
))
3058 put_futex_key(&q
.key
);
3063 * Userspace attempted a TID -> 0 atomic transition, and failed.
3064 * This is the in-kernel slowpath: we look up the PI state (if any),
3065 * and do the rt-mutex unlock.
3067 static int futex_unlock_pi(u32 __user
*uaddr
, unsigned int flags
)
3069 u32
uninitialized_var(curval
), uval
, vpid
= task_pid_vnr(current
);
3070 union futex_key key
= FUTEX_KEY_INIT
;
3071 struct futex_hash_bucket
*hb
;
3072 struct futex_q
*top_waiter
;
3075 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
3079 if (get_user(uval
, uaddr
))
3082 * We release only a lock we actually own:
3084 if ((uval
& FUTEX_TID_MASK
) != vpid
)
3087 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, FUTEX_WRITE
);
3091 hb
= hash_futex(&key
);
3092 spin_lock(&hb
->lock
);
3095 * Check waiters first. We do not trust user space values at
3096 * all and we at least want to know if user space fiddled
3097 * with the futex value instead of blindly unlocking.
3099 top_waiter
= futex_top_waiter(hb
, &key
);
3101 struct futex_pi_state
*pi_state
= top_waiter
->pi_state
;
3108 * If current does not own the pi_state then the futex is
3109 * inconsistent and user space fiddled with the futex value.
3111 if (pi_state
->owner
!= current
)
3114 get_pi_state(pi_state
);
3116 * By taking wait_lock while still holding hb->lock, we ensure
3117 * there is no point where we hold neither; and therefore
3118 * wake_futex_pi() must observe a state consistent with what we
3121 * In particular; this forces __rt_mutex_start_proxy() to
3122 * complete such that we're guaranteed to observe the
3123 * rt_waiter. Also see the WARN in wake_futex_pi().
3125 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
3126 spin_unlock(&hb
->lock
);
3128 /* drops pi_state->pi_mutex.wait_lock */
3129 ret
= wake_futex_pi(uaddr
, uval
, pi_state
);
3131 put_pi_state(pi_state
);
3134 * Success, we're done! No tricky corner cases.
3139 * The atomic access to the futex value generated a
3140 * pagefault, so retry the user-access and the wakeup:
3145 * A unconditional UNLOCK_PI op raced against a waiter
3146 * setting the FUTEX_WAITERS bit. Try again.
3151 * wake_futex_pi has detected invalid state. Tell user
3158 * We have no kernel internal state, i.e. no waiters in the
3159 * kernel. Waiters which are about to queue themselves are stuck
3160 * on hb->lock. So we can safely ignore them. We do neither
3161 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3164 if ((ret
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, 0))) {
3165 spin_unlock(&hb
->lock
);
3180 * If uval has changed, let user space handle it.
3182 ret
= (curval
== uval
) ? 0 : -EAGAIN
;
3185 spin_unlock(&hb
->lock
);
3187 put_futex_key(&key
);
3191 put_futex_key(&key
);
3196 put_futex_key(&key
);
3198 ret
= fault_in_user_writeable(uaddr
);
3206 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3207 * @hb: the hash_bucket futex_q was original enqueued on
3208 * @q: the futex_q woken while waiting to be requeued
3209 * @key2: the futex_key of the requeue target futex
3210 * @timeout: the timeout associated with the wait (NULL if none)
3212 * Detect if the task was woken on the initial futex as opposed to the requeue
3213 * target futex. If so, determine if it was a timeout or a signal that caused
3214 * the wakeup and return the appropriate error code to the caller. Must be
3215 * called with the hb lock held.
3218 * - 0 = no early wakeup detected;
3219 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3222 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket
*hb
,
3223 struct futex_q
*q
, union futex_key
*key2
,
3224 struct hrtimer_sleeper
*timeout
)
3229 * With the hb lock held, we avoid races while we process the wakeup.
3230 * We only need to hold hb (and not hb2) to ensure atomicity as the
3231 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3232 * It can't be requeued from uaddr2 to something else since we don't
3233 * support a PI aware source futex for requeue.
3235 if (!match_futex(&q
->key
, key2
)) {
3236 WARN_ON(q
->lock_ptr
&& (&hb
->lock
!= q
->lock_ptr
));
3238 * We were woken prior to requeue by a timeout or a signal.
3239 * Unqueue the futex_q and determine which it was.
3241 plist_del(&q
->list
, &hb
->chain
);
3244 /* Handle spurious wakeups gracefully */
3246 if (timeout
&& !timeout
->task
)
3248 else if (signal_pending(current
))
3249 ret
= -ERESTARTNOINTR
;
3255 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3256 * @uaddr: the futex we initially wait on (non-pi)
3257 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3258 * the same type, no requeueing from private to shared, etc.
3259 * @val: the expected value of uaddr
3260 * @abs_time: absolute timeout
3261 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3262 * @uaddr2: the pi futex we will take prior to returning to user-space
3264 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3265 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3266 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3267 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3268 * without one, the pi logic would not know which task to boost/deboost, if
3269 * there was a need to.
3271 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3272 * via the following--
3273 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3274 * 2) wakeup on uaddr2 after a requeue
3278 * If 3, cleanup and return -ERESTARTNOINTR.
3280 * If 2, we may then block on trying to take the rt_mutex and return via:
3281 * 5) successful lock
3284 * 8) other lock acquisition failure
3286 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3288 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3294 static int futex_wait_requeue_pi(u32 __user
*uaddr
, unsigned int flags
,
3295 u32 val
, ktime_t
*abs_time
, u32 bitset
,
3298 struct hrtimer_sleeper timeout
, *to
;
3299 struct futex_pi_state
*pi_state
= NULL
;
3300 struct rt_mutex_waiter rt_waiter
;
3301 struct futex_hash_bucket
*hb
;
3302 union futex_key key2
= FUTEX_KEY_INIT
;
3303 struct futex_q q
= futex_q_init
;
3306 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
3309 if (uaddr
== uaddr2
)
3315 to
= futex_setup_timer(abs_time
, &timeout
, flags
,
3316 current
->timer_slack_ns
);
3319 * The waiter is allocated on our stack, manipulated by the requeue
3320 * code while we sleep on uaddr.
3322 rt_mutex_init_waiter(&rt_waiter
);
3324 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, FUTEX_WRITE
);
3325 if (unlikely(ret
!= 0))
3329 q
.rt_waiter
= &rt_waiter
;
3330 q
.requeue_pi_key
= &key2
;
3333 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3336 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
3341 * The check above which compares uaddrs is not sufficient for
3342 * shared futexes. We need to compare the keys:
3344 if (match_futex(&q
.key
, &key2
)) {
3350 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3351 futex_wait_queue_me(hb
, &q
, to
);
3353 spin_lock(&hb
->lock
);
3354 ret
= handle_early_requeue_pi_wakeup(hb
, &q
, &key2
, to
);
3355 spin_unlock(&hb
->lock
);
3360 * In order for us to be here, we know our q.key == key2, and since
3361 * we took the hb->lock above, we also know that futex_requeue() has
3362 * completed and we no longer have to concern ourselves with a wakeup
3363 * race with the atomic proxy lock acquisition by the requeue code. The
3364 * futex_requeue dropped our key1 reference and incremented our key2
3368 /* Check if the requeue code acquired the second futex for us. */
3371 * Got the lock. We might not be the anticipated owner if we
3372 * did a lock-steal - fix up the PI-state in that case.
3374 if (q
.pi_state
&& (q
.pi_state
->owner
!= current
)) {
3375 spin_lock(q
.lock_ptr
);
3376 ret
= fixup_pi_state_owner(uaddr2
, &q
, current
);
3377 if (ret
&& rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
) {
3378 pi_state
= q
.pi_state
;
3379 get_pi_state(pi_state
);
3382 * Drop the reference to the pi state which
3383 * the requeue_pi() code acquired for us.
3385 put_pi_state(q
.pi_state
);
3386 spin_unlock(q
.lock_ptr
);
3389 struct rt_mutex
*pi_mutex
;
3392 * We have been woken up by futex_unlock_pi(), a timeout, or a
3393 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3396 WARN_ON(!q
.pi_state
);
3397 pi_mutex
= &q
.pi_state
->pi_mutex
;
3398 ret
= rt_mutex_wait_proxy_lock(pi_mutex
, to
, &rt_waiter
);
3400 spin_lock(q
.lock_ptr
);
3401 if (ret
&& !rt_mutex_cleanup_proxy_lock(pi_mutex
, &rt_waiter
))
3404 debug_rt_mutex_free_waiter(&rt_waiter
);
3406 * Fixup the pi_state owner and possibly acquire the lock if we
3409 res
= fixup_owner(uaddr2
, &q
, !ret
);
3411 * If fixup_owner() returned an error, proprogate that. If it
3412 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3415 ret
= (res
< 0) ? res
: 0;
3418 * If fixup_pi_state_owner() faulted and was unable to handle
3419 * the fault, unlock the rt_mutex and return the fault to
3422 if (ret
&& rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
) {
3423 pi_state
= q
.pi_state
;
3424 get_pi_state(pi_state
);
3427 /* Unqueue and drop the lock. */
3432 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
3433 put_pi_state(pi_state
);
3436 if (ret
== -EINTR
) {
3438 * We've already been requeued, but cannot restart by calling
3439 * futex_lock_pi() directly. We could restart this syscall, but
3440 * it would detect that the user space "val" changed and return
3441 * -EWOULDBLOCK. Save the overhead of the restart and return
3442 * -EWOULDBLOCK directly.
3448 put_futex_key(&q
.key
);
3450 put_futex_key(&key2
);
3454 hrtimer_cancel(&to
->timer
);
3455 destroy_hrtimer_on_stack(&to
->timer
);
3461 * Support for robust futexes: the kernel cleans up held futexes at
3464 * Implementation: user-space maintains a per-thread list of locks it
3465 * is holding. Upon do_exit(), the kernel carefully walks this list,
3466 * and marks all locks that are owned by this thread with the
3467 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3468 * always manipulated with the lock held, so the list is private and
3469 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3470 * field, to allow the kernel to clean up if the thread dies after
3471 * acquiring the lock, but just before it could have added itself to
3472 * the list. There can only be one such pending lock.
3476 * sys_set_robust_list() - Set the robust-futex list head of a task
3477 * @head: pointer to the list-head
3478 * @len: length of the list-head, as userspace expects
3480 SYSCALL_DEFINE2(set_robust_list
, struct robust_list_head __user
*, head
,
3483 if (!futex_cmpxchg_enabled
)
3486 * The kernel knows only one size for now:
3488 if (unlikely(len
!= sizeof(*head
)))
3491 current
->robust_list
= head
;
3497 * sys_get_robust_list() - Get the robust-futex list head of a task
3498 * @pid: pid of the process [zero for current task]
3499 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3500 * @len_ptr: pointer to a length field, the kernel fills in the header size
3502 SYSCALL_DEFINE3(get_robust_list
, int, pid
,
3503 struct robust_list_head __user
* __user
*, head_ptr
,
3504 size_t __user
*, len_ptr
)
3506 struct robust_list_head __user
*head
;
3508 struct task_struct
*p
;
3510 if (!futex_cmpxchg_enabled
)
3519 p
= find_task_by_vpid(pid
);
3525 if (!ptrace_may_access(p
, PTRACE_MODE_READ_REALCREDS
))
3528 head
= p
->robust_list
;
3531 if (put_user(sizeof(*head
), len_ptr
))
3533 return put_user(head
, head_ptr
);
3541 /* Constants for the pending_op argument of handle_futex_death */
3542 #define HANDLE_DEATH_PENDING true
3543 #define HANDLE_DEATH_LIST false
3546 * Process a futex-list entry, check whether it's owned by the
3547 * dying task, and do notification if so:
3549 static int handle_futex_death(u32 __user
*uaddr
, struct task_struct
*curr
,
3550 bool pi
, bool pending_op
)
3552 u32 uval
, uninitialized_var(nval
), mval
;
3555 /* Futex address must be 32bit aligned */
3556 if ((((unsigned long)uaddr
) % sizeof(*uaddr
)) != 0)
3560 if (get_user(uval
, uaddr
))
3564 * Special case for regular (non PI) futexes. The unlock path in
3565 * user space has two race scenarios:
3567 * 1. The unlock path releases the user space futex value and
3568 * before it can execute the futex() syscall to wake up
3569 * waiters it is killed.
3571 * 2. A woken up waiter is killed before it can acquire the
3572 * futex in user space.
3574 * In both cases the TID validation below prevents a wakeup of
3575 * potential waiters which can cause these waiters to block
3578 * In both cases the following conditions are met:
3580 * 1) task->robust_list->list_op_pending != NULL
3581 * @pending_op == true
3582 * 2) User space futex value == 0
3583 * 3) Regular futex: @pi == false
3585 * If these conditions are met, it is safe to attempt waking up a
3586 * potential waiter without touching the user space futex value and
3587 * trying to set the OWNER_DIED bit. The user space futex value is
3588 * uncontended and the rest of the user space mutex state is
3589 * consistent, so a woken waiter will just take over the
3590 * uncontended futex. Setting the OWNER_DIED bit would create
3591 * inconsistent state and malfunction of the user space owner died
3594 if (pending_op
&& !pi
&& !uval
) {
3595 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
3599 if ((uval
& FUTEX_TID_MASK
) != task_pid_vnr(curr
))
3603 * Ok, this dying thread is truly holding a futex
3604 * of interest. Set the OWNER_DIED bit atomically
3605 * via cmpxchg, and if the value had FUTEX_WAITERS
3606 * set, wake up a waiter (if any). (We have to do a
3607 * futex_wake() even if OWNER_DIED is already set -
3608 * to handle the rare but possible case of recursive
3609 * thread-death.) The rest of the cleanup is done in
3612 mval
= (uval
& FUTEX_WAITERS
) | FUTEX_OWNER_DIED
;
3615 * We are not holding a lock here, but we want to have
3616 * the pagefault_disable/enable() protection because
3617 * we want to handle the fault gracefully. If the
3618 * access fails we try to fault in the futex with R/W
3619 * verification via get_user_pages. get_user() above
3620 * does not guarantee R/W access. If that fails we
3621 * give up and leave the futex locked.
3623 if ((err
= cmpxchg_futex_value_locked(&nval
, uaddr
, uval
, mval
))) {
3626 if (fault_in_user_writeable(uaddr
))
3644 * Wake robust non-PI futexes here. The wakeup of
3645 * PI futexes happens in exit_pi_state():
3647 if (!pi
&& (uval
& FUTEX_WAITERS
))
3648 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
3654 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3656 static inline int fetch_robust_entry(struct robust_list __user
**entry
,
3657 struct robust_list __user
* __user
*head
,
3660 unsigned long uentry
;
3662 if (get_user(uentry
, (unsigned long __user
*)head
))
3665 *entry
= (void __user
*)(uentry
& ~1UL);
3672 * Walk curr->robust_list (very carefully, it's a userspace list!)
3673 * and mark any locks found there dead, and notify any waiters.
3675 * We silently return on any sign of list-walking problem.
3677 static void exit_robust_list(struct task_struct
*curr
)
3679 struct robust_list_head __user
*head
= curr
->robust_list
;
3680 struct robust_list __user
*entry
, *next_entry
, *pending
;
3681 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
3682 unsigned int uninitialized_var(next_pi
);
3683 unsigned long futex_offset
;
3686 if (!futex_cmpxchg_enabled
)
3690 * Fetch the list head (which was registered earlier, via
3691 * sys_set_robust_list()):
3693 if (fetch_robust_entry(&entry
, &head
->list
.next
, &pi
))
3696 * Fetch the relative futex offset:
3698 if (get_user(futex_offset
, &head
->futex_offset
))
3701 * Fetch any possibly pending lock-add first, and handle it
3704 if (fetch_robust_entry(&pending
, &head
->list_op_pending
, &pip
))
3707 next_entry
= NULL
; /* avoid warning with gcc */
3708 while (entry
!= &head
->list
) {
3710 * Fetch the next entry in the list before calling
3711 * handle_futex_death:
3713 rc
= fetch_robust_entry(&next_entry
, &entry
->next
, &next_pi
);
3715 * A pending lock might already be on the list, so
3716 * don't process it twice:
3718 if (entry
!= pending
) {
3719 if (handle_futex_death((void __user
*)entry
+ futex_offset
,
3720 curr
, pi
, HANDLE_DEATH_LIST
))
3728 * Avoid excessively long or circular lists:
3737 handle_futex_death((void __user
*)pending
+ futex_offset
,
3738 curr
, pip
, HANDLE_DEATH_PENDING
);
3742 static void futex_cleanup(struct task_struct
*tsk
)
3744 if (unlikely(tsk
->robust_list
)) {
3745 exit_robust_list(tsk
);
3746 tsk
->robust_list
= NULL
;
3749 #ifdef CONFIG_COMPAT
3750 if (unlikely(tsk
->compat_robust_list
)) {
3751 compat_exit_robust_list(tsk
);
3752 tsk
->compat_robust_list
= NULL
;
3756 if (unlikely(!list_empty(&tsk
->pi_state_list
)))
3757 exit_pi_state_list(tsk
);
3761 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3762 * @tsk: task to set the state on
3764 * Set the futex exit state of the task lockless. The futex waiter code
3765 * observes that state when a task is exiting and loops until the task has
3766 * actually finished the futex cleanup. The worst case for this is that the
3767 * waiter runs through the wait loop until the state becomes visible.
3769 * This is called from the recursive fault handling path in do_exit().
3771 * This is best effort. Either the futex exit code has run already or
3772 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3773 * take it over. If not, the problem is pushed back to user space. If the
3774 * futex exit code did not run yet, then an already queued waiter might
3775 * block forever, but there is nothing which can be done about that.
3777 void futex_exit_recursive(struct task_struct
*tsk
)
3779 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3780 if (tsk
->futex_state
== FUTEX_STATE_EXITING
)
3781 mutex_unlock(&tsk
->futex_exit_mutex
);
3782 tsk
->futex_state
= FUTEX_STATE_DEAD
;
3785 static void futex_cleanup_begin(struct task_struct
*tsk
)
3788 * Prevent various race issues against a concurrent incoming waiter
3789 * including live locks by forcing the waiter to block on
3790 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3791 * attach_to_pi_owner().
3793 mutex_lock(&tsk
->futex_exit_mutex
);
3796 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3798 * This ensures that all subsequent checks of tsk->futex_state in
3799 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3800 * tsk->pi_lock held.
3802 * It guarantees also that a pi_state which was queued right before
3803 * the state change under tsk->pi_lock by a concurrent waiter must
3804 * be observed in exit_pi_state_list().
3806 raw_spin_lock_irq(&tsk
->pi_lock
);
3807 tsk
->futex_state
= FUTEX_STATE_EXITING
;
3808 raw_spin_unlock_irq(&tsk
->pi_lock
);
3811 static void futex_cleanup_end(struct task_struct
*tsk
, int state
)
3814 * Lockless store. The only side effect is that an observer might
3815 * take another loop until it becomes visible.
3817 tsk
->futex_state
= state
;
3819 * Drop the exit protection. This unblocks waiters which observed
3820 * FUTEX_STATE_EXITING to reevaluate the state.
3822 mutex_unlock(&tsk
->futex_exit_mutex
);
3825 void futex_exec_release(struct task_struct
*tsk
)
3828 * The state handling is done for consistency, but in the case of
3829 * exec() there is no way to prevent futher damage as the PID stays
3830 * the same. But for the unlikely and arguably buggy case that a
3831 * futex is held on exec(), this provides at least as much state
3832 * consistency protection which is possible.
3834 futex_cleanup_begin(tsk
);
3837 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3838 * exec a new binary.
3840 futex_cleanup_end(tsk
, FUTEX_STATE_OK
);
3843 void futex_exit_release(struct task_struct
*tsk
)
3845 futex_cleanup_begin(tsk
);
3847 futex_cleanup_end(tsk
, FUTEX_STATE_DEAD
);
3850 long do_futex(u32 __user
*uaddr
, int op
, u32 val
, ktime_t
*timeout
,
3851 u32 __user
*uaddr2
, u32 val2
, u32 val3
)
3853 int cmd
= op
& FUTEX_CMD_MASK
;
3854 unsigned int flags
= 0;
3856 if (!(op
& FUTEX_PRIVATE_FLAG
))
3857 flags
|= FLAGS_SHARED
;
3859 if (op
& FUTEX_CLOCK_REALTIME
) {
3860 flags
|= FLAGS_CLOCKRT
;
3861 if (cmd
!= FUTEX_WAIT
&& cmd
!= FUTEX_WAIT_BITSET
&& \
3862 cmd
!= FUTEX_WAIT_REQUEUE_PI
)
3868 case FUTEX_UNLOCK_PI
:
3869 case FUTEX_TRYLOCK_PI
:
3870 case FUTEX_WAIT_REQUEUE_PI
:
3871 case FUTEX_CMP_REQUEUE_PI
:
3872 if (!futex_cmpxchg_enabled
)
3878 val3
= FUTEX_BITSET_MATCH_ANY
;
3880 case FUTEX_WAIT_BITSET
:
3881 return futex_wait(uaddr
, flags
, val
, timeout
, val3
);
3883 val3
= FUTEX_BITSET_MATCH_ANY
;
3885 case FUTEX_WAKE_BITSET
:
3886 return futex_wake(uaddr
, flags
, val
, val3
);
3888 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, NULL
, 0);
3889 case FUTEX_CMP_REQUEUE
:
3890 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 0);
3892 return futex_wake_op(uaddr
, flags
, uaddr2
, val
, val2
, val3
);
3894 return futex_lock_pi(uaddr
, flags
, timeout
, 0);
3895 case FUTEX_UNLOCK_PI
:
3896 return futex_unlock_pi(uaddr
, flags
);
3897 case FUTEX_TRYLOCK_PI
:
3898 return futex_lock_pi(uaddr
, flags
, NULL
, 1);
3899 case FUTEX_WAIT_REQUEUE_PI
:
3900 val3
= FUTEX_BITSET_MATCH_ANY
;
3901 return futex_wait_requeue_pi(uaddr
, flags
, val
, timeout
, val3
,
3903 case FUTEX_CMP_REQUEUE_PI
:
3904 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 1);
3910 SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
3911 struct __kernel_timespec __user
*, utime
, u32 __user
*, uaddr2
,
3914 struct timespec64 ts
;
3915 ktime_t t
, *tp
= NULL
;
3917 int cmd
= op
& FUTEX_CMD_MASK
;
3919 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
3920 cmd
== FUTEX_WAIT_BITSET
||
3921 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
3922 if (unlikely(should_fail_futex(!(op
& FUTEX_PRIVATE_FLAG
))))
3924 if (get_timespec64(&ts
, utime
))
3926 if (!timespec64_valid(&ts
))
3929 t
= timespec64_to_ktime(ts
);
3930 if (cmd
== FUTEX_WAIT
)
3931 t
= ktime_add_safe(ktime_get(), t
);
3935 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3936 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3938 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
3939 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
3940 val2
= (u32
) (unsigned long) utime
;
3942 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
3945 #ifdef CONFIG_COMPAT
3947 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3950 compat_fetch_robust_entry(compat_uptr_t
*uentry
, struct robust_list __user
**entry
,
3951 compat_uptr_t __user
*head
, unsigned int *pi
)
3953 if (get_user(*uentry
, head
))
3956 *entry
= compat_ptr((*uentry
) & ~1);
3957 *pi
= (unsigned int)(*uentry
) & 1;
3962 static void __user
*futex_uaddr(struct robust_list __user
*entry
,
3963 compat_long_t futex_offset
)
3965 compat_uptr_t base
= ptr_to_compat(entry
);
3966 void __user
*uaddr
= compat_ptr(base
+ futex_offset
);
3972 * Walk curr->robust_list (very carefully, it's a userspace list!)
3973 * and mark any locks found there dead, and notify any waiters.
3975 * We silently return on any sign of list-walking problem.
3977 static void compat_exit_robust_list(struct task_struct
*curr
)
3979 struct compat_robust_list_head __user
*head
= curr
->compat_robust_list
;
3980 struct robust_list __user
*entry
, *next_entry
, *pending
;
3981 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
3982 unsigned int uninitialized_var(next_pi
);
3983 compat_uptr_t uentry
, next_uentry
, upending
;
3984 compat_long_t futex_offset
;
3987 if (!futex_cmpxchg_enabled
)
3991 * Fetch the list head (which was registered earlier, via
3992 * sys_set_robust_list()):
3994 if (compat_fetch_robust_entry(&uentry
, &entry
, &head
->list
.next
, &pi
))
3997 * Fetch the relative futex offset:
3999 if (get_user(futex_offset
, &head
->futex_offset
))
4002 * Fetch any possibly pending lock-add first, and handle it
4005 if (compat_fetch_robust_entry(&upending
, &pending
,
4006 &head
->list_op_pending
, &pip
))
4009 next_entry
= NULL
; /* avoid warning with gcc */
4010 while (entry
!= (struct robust_list __user
*) &head
->list
) {
4012 * Fetch the next entry in the list before calling
4013 * handle_futex_death:
4015 rc
= compat_fetch_robust_entry(&next_uentry
, &next_entry
,
4016 (compat_uptr_t __user
*)&entry
->next
, &next_pi
);
4018 * A pending lock might already be on the list, so
4019 * dont process it twice:
4021 if (entry
!= pending
) {
4022 void __user
*uaddr
= futex_uaddr(entry
, futex_offset
);
4024 if (handle_futex_death(uaddr
, curr
, pi
,
4030 uentry
= next_uentry
;
4034 * Avoid excessively long or circular lists:
4042 void __user
*uaddr
= futex_uaddr(pending
, futex_offset
);
4044 handle_futex_death(uaddr
, curr
, pip
, HANDLE_DEATH_PENDING
);
4048 COMPAT_SYSCALL_DEFINE2(set_robust_list
,
4049 struct compat_robust_list_head __user
*, head
,
4052 if (!futex_cmpxchg_enabled
)
4055 if (unlikely(len
!= sizeof(*head
)))
4058 current
->compat_robust_list
= head
;
4063 COMPAT_SYSCALL_DEFINE3(get_robust_list
, int, pid
,
4064 compat_uptr_t __user
*, head_ptr
,
4065 compat_size_t __user
*, len_ptr
)
4067 struct compat_robust_list_head __user
*head
;
4069 struct task_struct
*p
;
4071 if (!futex_cmpxchg_enabled
)
4080 p
= find_task_by_vpid(pid
);
4086 if (!ptrace_may_access(p
, PTRACE_MODE_READ_REALCREDS
))
4089 head
= p
->compat_robust_list
;
4092 if (put_user(sizeof(*head
), len_ptr
))
4094 return put_user(ptr_to_compat(head
), head_ptr
);
4101 #endif /* CONFIG_COMPAT */
4103 #ifdef CONFIG_COMPAT_32BIT_TIME
4104 SYSCALL_DEFINE6(futex_time32
, u32 __user
*, uaddr
, int, op
, u32
, val
,
4105 struct old_timespec32 __user
*, utime
, u32 __user
*, uaddr2
,
4108 struct timespec64 ts
;
4109 ktime_t t
, *tp
= NULL
;
4111 int cmd
= op
& FUTEX_CMD_MASK
;
4113 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
4114 cmd
== FUTEX_WAIT_BITSET
||
4115 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
4116 if (get_old_timespec32(&ts
, utime
))
4118 if (!timespec64_valid(&ts
))
4121 t
= timespec64_to_ktime(ts
);
4122 if (cmd
== FUTEX_WAIT
)
4123 t
= ktime_add_safe(ktime_get(), t
);
4126 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
4127 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
4128 val2
= (int) (unsigned long) utime
;
4130 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
4132 #endif /* CONFIG_COMPAT_32BIT_TIME */
4134 static void __init
futex_detect_cmpxchg(void)
4136 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
4140 * This will fail and we want it. Some arch implementations do
4141 * runtime detection of the futex_atomic_cmpxchg_inatomic()
4142 * functionality. We want to know that before we call in any
4143 * of the complex code paths. Also we want to prevent
4144 * registration of robust lists in that case. NULL is
4145 * guaranteed to fault and we get -EFAULT on functional
4146 * implementation, the non-functional ones will return
4149 if (cmpxchg_futex_value_locked(&curval
, NULL
, 0, 0) == -EFAULT
)
4150 futex_cmpxchg_enabled
= 1;
4154 static int __init
futex_init(void)
4156 unsigned int futex_shift
;
4159 #if CONFIG_BASE_SMALL
4160 futex_hashsize
= 16;
4162 futex_hashsize
= roundup_pow_of_two(256 * num_possible_cpus());
4165 futex_queues
= alloc_large_system_hash("futex", sizeof(*futex_queues
),
4167 futex_hashsize
< 256 ? HASH_SMALL
: 0,
4169 futex_hashsize
, futex_hashsize
);
4170 futex_hashsize
= 1UL << futex_shift
;
4172 futex_detect_cmpxchg();
4174 for (i
= 0; i
< futex_hashsize
; i
++) {
4175 atomic_set(&futex_queues
[i
].waiters
, 0);
4176 plist_head_init(&futex_queues
[i
].chain
);
4177 spin_lock_init(&futex_queues
[i
].lock
);
4182 core_initcall(futex_init
);